diff --git a/.github/workflows/python-test.yml b/.github/workflows/python-test.yml index b844a53..f0bcb2e 100644 --- a/.github/workflows/python-test.yml +++ b/.github/workflows/python-test.yml @@ -28,8 +28,10 @@ jobs: pytest -vx tests/test_dataset_utils.py pytest -vx tests/test_testable_tasks.py pytest -vx tests/factors/test_covariances.py - pytest -vx tests/factors/test_eigens.py + pytest -vx tests/factors/test_eigendecompositions.py + pytest -vx tests/factors/test_lambdas.py pytest -vx tests/modules/test_modules.py pytest -vx tests/modules/test_per_sample_gradients.py + pytest -vx tests/modules/test_matmul.py pytest -vx tests/scores/test_pairwise_scores.py pytest -vx tests/scores/test_self_scores.py \ No newline at end of file diff --git a/.gitignore b/.gitignore index 1e2ced7..f4b18f5 100644 --- a/.gitignore +++ b/.gitignore @@ -164,7 +164,8 @@ cython_debug/ # Checkpoints and influence outputs checkpoints/ -analyses/ +influence_results/ data/ +cache/ *.pth *.pt \ No newline at end of file diff --git a/DOCUMENTATION.md b/DOCUMENTATION.md index df68d53..cb2ee86 100644 --- a/DOCUMENTATION.md +++ b/DOCUMENTATION.md @@ -1,18 +1,18 @@ # Kronfluence: Technical Documentation & FAQs -For a detailed description of the methodology, please refer to the [**paper**](https://arxiv.org/abs/2308.03296) *Studying Large Language Model Generalization with Influence Functions*. +For a detailed description of the methodology, please refer to the [**paper**](https://arxiv.org/abs/2308.03296), *Studying Large Language Model Generalization with Influence Functions*. ## Requirements Kronfluence has been tested and is compatible with the following versions of [PyTorch](https://pytorch.org/): -- PyTorch 2.1 or higher -- Python 3.9 or higher +- Python: Version 3.9 or later +- PyTorch: Version 2.1 or later ## Supported Modules & Strategies Kronfluence offers support for: - Computing influence functions on selected PyTorch modules. Currently, we support `nn.Linear` and `nn.Conv2d`. -- Computing influence functions with several Hessian approximation strategies, including `identity`, `diagonal`, `KFAC`, and `EKFAC`. +- Computing influence functions with several Hessian approximation strategies, including `identity`, `diagonal`, `kfac`, and `ekfac`. - Computing pairwise and self-influence (with and without measurement) scores. > [!NOTE] @@ -22,12 +22,12 @@ Kronfluence offers support for: ## Step-by-Step Guide -See [UCI Regression example](https://github.com/pomonam/kronfluence/blob/main/examples/uci/) for the complete workflow and -interactive tutorial. +See [UCI Regression example](https://github.com/pomonam/kronfluence/blob/main/examples/uci/) for the complete workflow and an interactive tutorial. **Prepare Your Model and Dataset.** Before computing influence scores, you need to prepare the trained model and dataset. You can use any frameworks to -train the model (e.g., [Pytorch Lightning](https://lightning.ai/docs/pytorch/stable/) or [HuggingFace Trainer](https://huggingface.co/docs/transformers/main_classes/trainer)). +train the model (e.g., [Pytorch Lightning](https://lightning.ai/docs/pytorch/stable/) or [HuggingFace Trainer](https://huggingface.co/docs/transformers/main_classes/trainer)); you just need to prepare the final model parameters. + ```python ... # Get the model with the trained parameters. @@ -42,7 +42,7 @@ query_dataset = prepare_query_dataset() **Define a Task.** To compute influence scores, you need to define a [`Task`](https://github.com/pomonam/kronfluence/blob/main/kronfluence/task.py) class. This class contains information about the trained model and how influence scores will be computed: -(1) how to compute the training loss; (2) how to compute the measurable quantity (f(θ) in the [paper](https://arxiv.org/abs/2308.03296); see Equation 5); +(1) how to compute the training loss; (2) how to compute the measurable quantity (f(θ) in the [paper](https://arxiv.org/abs/2308.03296); see **Equation 5**); (3) which modules to use for influence function computations; and (4) whether the model used [attention mask](https://huggingface.co/docs/transformers/en/glossary#attention-mask). ```python @@ -58,7 +58,6 @@ class YourTask(Task): model: nn.Module, sample: bool = False, ) -> torch.Tensor: - # This will be used for computing the training gradient. # TODO: Complete this method. def compute_measurement( @@ -66,16 +65,15 @@ class YourTask(Task): batch: Any, model: nn.Module, ) -> torch.Tensor: - # This will be used for computing the measurable quantity. # TODO: Complete this method. - def tracked_modules(self) -> Optional[List[str]]: + def get_influence_tracked_modules(self) -> Optional[List[str]]: # TODO: [Optional] Complete this method. return None # Compute influence scores on all available modules. def get_attention_mask(self, batch: Any) -> Optional[Union[Dict[str, torch.Tensor], torch.Tensor]]: # TODO: [Optional] Complete this method. - return None # No attention mask is used. + return None # Attention mask not used. ``` **Prepare Your Model for Influence Computations.** @@ -91,15 +89,15 @@ model = prepare_model(model=model, task=task) ... ``` -If you have specified specific module names in `Task.tracked_modules`, `TrackedModule` will only be installed for these modules. +If you have specified specific module names in `Task.get_influence_tracked_modules`, `TrackedModule` will only be installed for these modules. **\[Optional\] Create a DDP and FSDP Module.** After calling `prepare_model`, you can create [DistributedDataParallel (DDP)](https://pytorch.org/tutorials/intermediate/ddp_tutorial.html) or -[FullyShardedDataParallel (FSDP)](https://pytorch.org/docs/stable/fsdp.html) module or even use `torch.compile`. +[FullyShardedDataParallel (FSDP)](https://pytorch.org/docs/stable/fsdp.html) module. You can also wrap your model with [`torch.compile`](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html). **Set up the Analyzer and Fit Factors.** -Initialize the `Analyzer` and execute `fit_all_factors` to compute all factors that aim to approximate the Hessian -(or Gauss-Newton Hessian). The computed factors will be stored on disk. +Initialize the `Analyzer` and run `fit_all_factors` to compute all factors that aim to approximate the Hessian +([Gauss-Newton Hessian](https://www.cs.toronto.edu/~rgrosse/courses/csc2541_2022/readings/L03_metrics.pdf)). The computed factors will be stored on disk. ```python from kronfluence.analyzer import Analyzer @@ -138,17 +136,20 @@ You can organize all factors and scores for the specific model with `factors_nam **What should I do if my model does not have any nn.Linear or nn.Conv2d modules?** Currently, the implementation does not support influence computations for modules other than `nn.Linear` or `nn.Conv2d`. -Try rewriting the model so that it uses supported modules (as done for the `conv1d` module in [GPT-2 example](https://github.com/pomonam/kronfluence/tree/documentation/examples/wikitext)). +Try rewriting the model so that it uses supported modules (as done for the `conv1d` module in the [GPT-2 example](https://github.com/pomonam/kronfluence/tree/documentation/examples/wikitext)). Alternatively, you can create a subclass of `TrackedModule` to compute influence scores for your custom module. If there are specific modules you would like to see supported, please submit an issue. -**How should I write task.tracked_modules?** +**How should I write task.get_influence_tracked_modules?** We recommend using all supported modules for influence computations. However, if you would like to compute influence scores on subset of the modules (e.g., influence computations only on MLP layers for transformer or influence computation only on the last layer), inspect `model.named_modules()` to determine what modules to use. You can specify the list of module names you want to analyze. +> [!TIP] +> `Analyzer.get_module_summary(model)` can be helpful in figuring out what modules to include. + > [!NOTE] -> If the embedding layer for transformers are defined with `nn.Linear`, you must write +> If the embedding layer for transformers are defined with `nn.Linear`, you must write your own > `task.tracked_modules` to avoid influence computations embedding matrices. **How should I implement Task.compute_train_loss?** @@ -158,14 +159,14 @@ the summed loss (over batches and tokens). **How should I implement Task.compute_measurement?** It depends on the analysis you would like to perform. Influence functions approximate the [effect of downweighting/upweighting a training data point on the query's measurable quantity](https://arxiv.org/abs/2209.05364). You can use the loss, [margin](https://arxiv.org/abs/2303.14186) (for classification), -or [conditional log-likelihood](https://arxiv.org/abs/2308.03296) (for language modeling). +or [conditional log-likelihood](https://arxiv.org/abs/2308.03296) (for language modeling). Note that many influence functions implementation, by default, uses the loss. **I encounter TrackedModuleNotFoundError when using DDP or FSDP.** Make sure to call `prepare_model` before wrapping your model with DDP or FSDP. Calling `prepare_model` on DDP modules can cause `TrackedModuleNotFoundError`. **My model uses supported modules, but influence scores are not computed.** -Kronfluence uses module hooks to compute factors and influence scores. For these to be tracked and computed, +Kronfluence uses [module hooks](https://pytorch.org/docs/stable/generated/torch.Tensor.register_hook.html) to compute factors and influence scores. For these to be tracked and computed, the model's forward pass should directly call the module. ```python @@ -179,6 +180,16 @@ def forward(x: torch.Tensor) -> torch.Tensor: x = self.linear.weight @ x + self.linear.bias # This does not work 😞 ``` +> [!WARNING] +> The default arguments assume the module is used only once during the forward pass. +> If your model shares parameters (e.g., the module is used in multiple places during the forward pass), set +> `has_shared_parameters=True` in `FactorArguments`. + +**Why are there so many arguments?** +Kronfluence was originally developed to compute influence scores on large-scale models, which is why `FactorArguments` and `ScoreArguments` +have many parameters to support these use cases. However, for most standard applications, the default argument values +should suffice. Feel free to use the default settings unless you have specific requirements that necessitate customization. + **I get X error when fitting factors/computing scores.** Please feel free to contact me by [filing an issue](https://github.com/pomonam/kronfluence/issues) or [through email](mailto:jbae@cs.toronto.edu). @@ -193,13 +204,14 @@ from kronfluence.arguments import FactorArguments factor_args = FactorArguments( strategy="ekfac", # Choose from "identity", "diagonal", "kfac", or "ekfac". use_empirical_fisher=False, - distributed_sync_steps=1000, amp_dtype=None, + amp_scale=2.0**16, + has_shared_parameters=False, # Settings for covariance matrix fitting. covariance_max_examples=100_000, - covariance_data_partition_size=1, - covariance_module_partition_size=1, + covariance_data_partitions=1, + covariance_module_partitions=1, activation_covariance_dtype=torch.float32, gradient_covariance_dtype=torch.float32, @@ -208,10 +220,11 @@ factor_args = FactorArguments( # Settings for Lambda matrix fitting. lambda_max_examples=100_000, - lambda_data_partition_size=1, - lambda_module_partition_size=1, - lambda_iterative_aggregate=False, - cached_activation_cpu_offload=False, + lambda_data_partitions=1, + lambda_module_partitions=1, + use_iterative_lambda_aggregation=False, + offload_activations_to_cpu=False, + per_sample_gradient_dtype=torch.float32, lambda_dtype=torch.float32, ) @@ -224,11 +237,14 @@ You can change: - `use_empirical_fisher`: Determines whether to use the [empirical Fisher](https://arxiv.org/abs/1905.12558) (using actual labels from batch) instead of the true Fisher (using sampled labels from model's predictions). It is recommended to be `False`. - `amp_dtype`: Selects the dtype for [automatic mixed precision (AMP)](https://pytorch.org/docs/stable/amp.html). Disables AMP if set to `None`. +- `amp_scale`: Sets the scale factor for [automatic mixed precision (AMP)](https://pytorch.org/docs/stable/amp.html). +- `has_shared_parameters`: Specifies whether the shared parameters exist in the forward pass. ### Fitting Covariance Matrices `kfac` and `ekfac` require computing the uncentered activation and pre-activation pseudo-gradient covariance matrices. To fit covariance matrices, you can use `analyzer.fit_covariance_matrices`. + ```python # Fitting covariance matrices. analyzer.fit_covariance_matrices(factors_name="initial_factor", dataset=train_dataset, factor_args=factor_args) @@ -236,18 +252,18 @@ analyzer.fit_covariance_matrices(factors_name="initial_factor", dataset=train_da covariance_matrices = analyzer.load_covariance_matrices(factors_name="initial_factor") ``` -This step corresponds to Equation 16 in the paper. You can tune: +This step corresponds to **Equation 16** in the paper. You can tune: - `covariance_max_examples`: Controls the maximum number of data points for fitting covariance matrices. Setting it to `None`, Kronfluence computes covariance matrices for all data points. -- `covariance_data_partition_size`: Number of data partitions to use for computing covariance matrices. -For example, when `covariance_data_partition_size = 2`, the dataset is split into 2 chunks and covariance matrices +- `covariance_data_partitions`: Number of data partitions to use for computing covariance matrices. +For example, when `covariance_data_partitions=2`, the dataset is split into 2 chunks and covariance matrices are separately computed for each chunk. These chunked covariance matrices are later aggregated. This is useful with GPU preemption as intermediate covariance matrices will be saved in disk. It can be also helpful when launching multiple parallel jobs, where each GPU -can compute covariance matrices on some partitioned data (You can specify `target_data_partitions` in the parameter). -- `covariance_module_partition_size`: Number of module partitions to use for computing covariance matrices. -For example, when `covariance_module_partition_size = 2`, the module is split into 2 chunks and covariance matrices +can compute covariance matrices on some partitioned data (you can specify `target_data_partitions` in the parameter). +- `covariance_module_partitions`: Number of module partitions to use for computing covariance matrices. +For example, when `covariance_module_partitions=2`, the module is split into 2 chunks and covariance matrices are separately computed for each chunk. This is useful when the available GPU memory is limited (e.g., the total -covariance matrices cannot fit into memory). However, this will do multiple iterations over the dataset and can be slow. +covariance matrices cannot fit into GPU memory). However, this will require multiple iterations over the dataset and can be slow. - `activation_covariance_dtype`: `dtype` for computing activation covariance matrices. You can also use `torch.bfloat16` or `torch.float16`. - `gradient_covariance_dtype`: `dtype` for computing pre-activation pseudo-gradient covariance matrices. You can also use `torch.bfloat16` @@ -256,26 +272,26 @@ or `torch.float16`. **Dealing with OOMs.** Here are some steps to fix Out of Memory (OOM) errors. 1. Try reducing the `per_device_batch_size` when fitting covariance matrices. 2. Try using lower precision for `activation_covariance_dtype` and `gradient_covariance_dtype`. -3. Try setting `covariance_module_partition_size > 1`. +3. Try setting `covariance_module_partitions > 1`. ### Performing Eigendecomposition After computing the covariance matrices, `kfac` and `ekfac` require performing eigendecomposition. ```python -# Performing Eigendecomposition. +# Performing eigendecomposition. analyzer.perform_eigendecomposition(factors_name="initial_factor", factor_args=factor_args) -# Loading Eigendecomposition results. +# Loading eigendecomposition results (e.g., eigenvectors and eigenvalues). eigen_factors = analyzer.load_eigendecomposition(factors_name="initial_factor") ``` -This corresponds to Equation 18 in the paper. You can tune: +This corresponds to **Equation 18** in the paper. You can tune: - `eigendecomposition_dtype`: `dtype` for performing eigendecomposition. You can also use `torch.float32`, -but `torch.float64` is recommended. +but `torch.float64` is strongly recommended. ### Fitting Lambda Matrices -`ekfac` and `diagonal` require computing the Lambda (eigenvalue) matrices for all modules. +`ekfac` and `diagonal` require computing the Lambda (corrected-eigenvalue) matrices for all modules. ```python # Fitting Lambda matrices. @@ -284,23 +300,24 @@ analyzer.fit_lambda_matrices(factors_name="initial_factor", dataset=train_datase lambda_matrices = analyzer.load_lambda_matrices(factors_name="initial_factor") ``` -This corresponds to Equation 20 in the paper. You can tune: +This corresponds to **Equation 20** in the paper. You can tune: - `lambda_max_examples`: Controls the maximum number of data points for fitting Lambda matrices. -- `lambda_data_partition_size`: Number of data partitions to use for computing Lambda matrices. -- `lambda_module_partition_size`: Number of module partitions to use for computing Lambda matrices. -- `cached_activation_cpu_offload`: Computing the per-sample-gradient requires saving the intermediate activation in memory. -You can set `cached_activation_cpu_offload=True` to cache these activations in CPU. -- `lambda_iterative_aggregate`: Whether to compute the Lambda matrices with for-loop instead of batched matrix multiplications. -This is helpful for reducing peak memory, as it avoids holding multiple copies of tensors with the same shape as the per-sample-gradient. +- `lambda_data_partitions`: Number of data partitions to use for computing Lambda matrices. +- `lambda_module_partitions`: Number of module partitions to use for computing Lambda matrices. +- `offload_activations_to_cpu`: Computing the per-sample-gradient requires saving the intermediate activation in memory. +You can set `offload_activations_to_cpu=True` to cache these activations in CPU. This is helpful for dealing with OOMs, but will make the overall computation slower. +- `use_iterative_lambda_aggregation`: Whether to compute the Lambda matrices with for-loops instead of batched matrix multiplications. +This is helpful for reducing peak GPU memory, as it avoids holding multiple copies of tensors with the same shape as the per-sample-gradient. +- `per_sample_gradient_dtype`: `dtype` for computing per-sample-gradient. You can also use `torch.bfloat16` +or `torch.float16`. - `lambda_dtype`: `dtype` for computing Lambda matrices. You can also use `torch.bfloat16` or `torch.float16`. - **Dealing with OOMs.** Here are some steps to fix Out of Memory (OOM) errors. 1. Try reducing the `per_device_batch_size` when fitting Lambda matrices. -2. Try setting `lambda_iterative_aggregate=True` or `cached_activation_cpu_offload=True`. -3. Try using lower precision for `lambda_dtype`. -4. Try using `lambda_module_partition_size > 1`. +2. Try setting `use_iterative_lambda_aggregation=True` or `offload_activations_to_cpu=True`. (Try out `use_iterative_lambda_aggregation=True` first.) +3. Try using lower precision for `per_sample_gradient_dtype` and `lambda_dtype`. +4. Try using `lambda_module_partitions > 1`. ### FAQs @@ -310,8 +327,9 @@ Using `use_empirical_fisher=True` could make the process more deterministic. Mor different eigenvectors when performing eigendecomposition. **How should I select the batch size?** -You can use the largest possible batch size that does not result in OOM. Typically, the batch size for fitting Lambda -matrices should be smaller than that used for fitting covariance matrices. +You can use the largest possible batch size that avoids OOM error. Typically, the batch size for fitting Lambda +matrices should be smaller than that used for fitting covariance matrices. Furthermore, note that you should be getting +similar results, regardless of what batch size you use (different from training neural networks). --- @@ -322,49 +340,55 @@ import torch from kronfluence.arguments import ScoreArguments score_args = ScoreArguments( - damping=1e-08, - cached_activation_cpu_offload=False, - distributed_sync_steps=1000, + damping_factor=1e-08, amp_dtype=None, + offload_activations_to_cpu=False, + + # More functionalities to compute influence scores. + data_partitions=1, + module_partitions=1, + compute_per_module_scores=False, + compute_per_token_scores=False, + use_measurement_for_self_influence=False, + aggregate_query_gradients=False, + aggregate_train_gradients=False, - data_partition_size=1, - module_partition_size=1, - per_module_score=False, - # Configuration for query batching. - query_gradient_rank=None, + query_gradient_low_rank=None, + use_full_svd=False, query_gradient_svd_dtype=torch.float32, - num_query_gradient_aggregations=1, - use_measurement_for_self_influence=False, + query_gradient_accumulation_steps=1, + # Configuration for dtype. score_dtype=torch.float32, per_sample_gradient_dtype=torch.float32, precondition_dtype=torch.float32, ) ``` -- `damping`: A damping factor for the damped inverse Hessian-vector product (iHVP). Uses a heuristic based on mean eigenvalues -(0.1 x mean eigenvalues) if None. -- `cached_activation_cpu_offload`: Whether to offload cached activations to CPU. +- `damping_factor`: A damping factor for the damped inverse Hessian-vector product (iHVP). Uses a heuristic based on mean eigenvalues +`(0.1 x mean eigenvalues)` if `None`, as done in [this paper](https://arxiv.org/abs/2308.03296). - `amp_dtype`: Selects the dtype for [automatic mixed precision (AMP)](https://pytorch.org/docs/stable/amp.html). Disables AMP if set to `None`. -- `data_partition_size`: Number of data partitions for computing influence scores. -- `module_partition_size`: Number of module partitions for computing influence scores. -- `per_module_score`: Whether to return a per-module influence scores. Instead of summing over influences across +- `offload_activations_to_cpu`: Whether to offload cached activations to CPU. +- `data_partitions`: Number of data partitions for computing influence scores. +- `module_partitions`: Number of module partitions for computing influence scores. +- `compute_per_module_scores`: Whether to return a per-module influence scores. Instead of summing over influences across all modules, this will keep track of intermediate module-wise scores. - -- `query_gradient_rank`: The rank for the query batching (low-rank approximation to the query gradient; see Section 3.2.2). If `None`, no query batching will be used. -- `query_gradient_svd_dtype`: `dtype` for performing singular value decomposition (SVD) for query batch. You can also use `torch.float64`. -- `num_query_gradient_aggregations`: Number of query gradients to aggregate over. +- `compute_per_token_scores`: Whether to return a per-token influence scores. Only applicable to transformer-based models. +- `aggregate_query_gradients`: Whether to use the summed query gradient instead of per-sample query gradients. +- `aggregate_train_gradients`: Whether to use the summed training gradient instead of per-sample training gradients. - `use_measurement_for_self_influence`: Whether to use the measurement (instead of the loss) when computing self-influence scores. - +- `query_gradient_low_rank`: The rank for the query batching (low-rank approximation to the preconditioned query gradient; see **Section 3.2.2**). If `None`, no query batching will be used. +- `query_gradient_svd_dtype`: `dtype` for performing singular value decomposition (SVD) for query batch. You can also use `torch.float64`. +- `query_gradient_accumulation_steps`: Number of query gradients to accumulate over. For example, when `query_gradient_accumulation_steps=2` with +`query_batch_size=16`, a total of 32 query gradients will be stored in memory when computing dot products with training gradients. - `score_dtype`: `dtype` for computing influence scores. You can use `torch.bfloat16` or `torch.float16`. - `per_sample_gradient_dtype`: `dtype` for computing per-sample-gradient. You can use `torch.bfloat16` or `torch.float16`. -- `precondition_dtype`: `dtype` for performing preconditioning. You can use `torch.bfloat16` or `torch.float16`, -but `torch.float32` is recommended. +- `precondition_dtype`: `dtype` for performing preconditioning. You can use `torch.bfloat16` or `torch.float16`. ### Computing Influence Scores -To compute pairwise influence scores (Equation 5 in the paper), you can run: +To compute pairwise influence scores (**Equation 5** in the paper), you can run: ```python # Computing pairwise influence scores. @@ -373,7 +397,7 @@ analyzer.compute_pairwise_scores(scores_name="pairwise", factors_name="ekfac", s scores = analyzer.load_pairwise_scores(scores_name="pairwise") ``` -To compute self-influence scores (see Section 5.4 from [paper](https://arxiv.org/pdf/1703.04730.pdf)), you can run: +To compute self-influence scores (see **Section 5.4** from [this paper](https://arxiv.org/pdf/1703.04730.pdf)), you can run: ```python # Computing self-influence scores. @@ -382,20 +406,34 @@ analyzer.compute_self_scores(scores_name="self", factors_name="ekfac", score_arg scores = analyzer.load_self_scores(scores_name="self") ``` -**Dealing with OOMs** Here are some steps to fix Out of Memory (OOM) errors. +By default, self-influence score computations only use the loss function for gradient calculations. +In this case, the method returns a vector of size `len(train_dataset)`, where each value corresponds +to `g_l^T ⋅ H^{-1} ⋅ g_l`. Here, `g_l` denotes the gradient of the loss function with respect to the model parameters, +and `H^{-1}` represents the inverse Hessian matrix. If you want to use the measurement function instead of the loss function +for self-influence calculations, set `use_measurement_for_self_influence=True`. In this case, each value in the returned +vector will correspond to `g_m^T ⋅ H^{-1} ⋅ g_l`, where `g_m` is the gradient of the measurement function with respect to the model parameters. + +**Dealing with OOMs.** Here are some steps to fix Out of Memory (OOM) errors. 1. Try reducing the `per_device_query_batch_size` or `per_device_train_batch_size`. -2. Try setting `cached_activation_cpu_offload=True`. +2. Try setting `offload_activations_to_cpu=True`. 3. Try using lower precision for `per_sample_gradient_dtype` and `score_dtype`. -4. Try setting `query_gradient_rank > 1`. The recommended values are `16`, `32`, `64`, `128`, and `256`. Note that query +4. Try using lower precision for `precondition_dtype`. +5. Try setting `query_gradient_low_rank > 1`. The recommended values are `16`, `32`, `64`, `128`, and `256`. Note that query batching is only supported for computing pairwise influence scores, not self-influence scores. -5. Try setting `module_partition_size > 1`. +6. Try setting `module_partitions > 1`. ### FAQs +**How should I choose a damping term?** +When setting the damping term, both `1e-08` and `None` are reasonable choices. The optimal value may depend on your +specific workload. Another heuristic, suggested in [this paper](https://arxiv.org/abs/2405.12186), is to use `10 * learning_rate * num_iterations` when the model +was trained using SGD with a momentum of 0.9. In practice, I have observed that the damping term does not significantly +affect the final results as long as it is not too large (e.g., `1e-01`). Feel free to experiment with different values within a +reasonable range to find what works best for your use case. + **Influence scores are very large in magnitude.** Ideally, influence scores need to be divided by the total number of training data points. However, the code does -not normalize the scores. If you would like, you can divide the scores with the total number of data points (or tokens) used to -train the model. +not normalize the scores. If you would like, you can divide the scores with the total number of data points (or tokens for language modeling) used to train the model. ## References @@ -404,4 +442,5 @@ train the model. 3. [TRAK: Attributing Model Behavior at Scale](https://arxiv.org/abs/2303.14186). Sung Min Park, Kristian Georgiev, Andrew Ilyas, Guillaume Leclerc, Aleksander Madry. ICML, 2023. 4. [Understanding Black-box Predictions via Influence Functions](https://arxiv.org/abs/1703.04730). Pang Wei Koh, Percy Liang. ICML, 2017. 5. [Optimizing Neural Networks with Kronecker-factored Approximate Curvature](https://arxiv.org/abs/1503.05671). James Martens, Roger Grosse. Tech Report, 2015. -5. [Fast Approximate Natural Gradient Descent in a Kronecker-factored Eigenbasis](https://arxiv.org/abs/1806.03884). Thomas George, César Laurent, Xavier Bouthillier, Nicolas Ballas, Pascal Vincent. NeurIPS, 2018. \ No newline at end of file +6. [Fast Approximate Natural Gradient Descent in a Kronecker-factored Eigenbasis](https://arxiv.org/abs/1806.03884). Thomas George, César Laurent, Xavier Bouthillier, Nicolas Ballas, Pascal Vincent. NeurIPS, 2018. +7. [Training Data Attribution via Approximate Unrolled Differentiation](https://arxiv.org/abs/2405.12186). Juhan Bae, Wu Lin, Jonathan Lorraine, Roger Grosse. Preprint, 2024. diff --git a/README.md b/README.md index d99e060..cdad5da 100644 --- a/README.md +++ b/README.md @@ -22,14 +22,11 @@ --- -> **Kronfluence** is a research repository designed to compute [influence functions](https://arxiv.org/abs/1703.04730) using [Kronecker-factored Approximate Curvature (KFAC)](https://arxiv.org/abs/1503.05671) or [Eigenvalue-corrected KFAC (EKFAC)](https://arxiv.org/abs/1806.03884). -For a detailed description of the methodology, see the [**paper**](https://arxiv.org/abs/2308.03296) *Studying Large Language Model Generalization with Influence Functions*. +> **Kronfluence** is a PyTorch package designed to compute [influence functions](https://arxiv.org/abs/1703.04730) using [Kronecker-factored Approximate Curvature (KFAC)](https://arxiv.org/abs/1503.05671) or [Eigenvalue-corrected KFAC (EKFAC)](https://arxiv.org/abs/1806.03884). +For detailed description of the methodology, see the [**paper**](https://arxiv.org/abs/2308.03296), *Studying Large Language Model Generalization with Influence Functions*. --- -> [!WARNING] -> This repository is under active development and has not reached its first stable release. - ## Installation > [!IMPORTANT] @@ -53,11 +50,9 @@ pip install -e . ## Getting Started -Kronfluence supports influence computations on `nn.Linear` and `nn.Conv2d` modules. See the [**Technical Documentation**](https://github.com/pomonam/kronfluence/blob/main/DOCUMENTATION.md) page for a comprehensive guide. - -### Learn More +Kronfluence supports influence computations on [`nn.Linear`](https://pytorch.org/docs/stable/generated/torch.nn.Linear.html) and [`nn.Conv2d`](https://pytorch.org/docs/stable/generated/torch.nn.Conv2d.html) modules. +See the [**Technical Documentation**](https://github.com/pomonam/kronfluence/blob/main/DOCUMENTATION.md) page for a comprehensive guide. -The [examples](https://github.com/pomonam/kronfluence/tree/main/examples) folder contains several examples demonstrating how to use Kronfluence. More examples will be added in the future. **TL;DR** You need to prepare a trained model and datasets, and pass them into the `Analyzer` class. ```python @@ -115,6 +110,30 @@ analyzer.compute_pairwise_scores( scores = analyzer.load_pairwise_scores(scores_name="my_scores") ``` +Kronfluence supports various PyTorch features; the following table summarizes the supported features: + +
+ +| Feature | Supported | +|-----------------------------------------------------------------------------------------------------------------------------|:---------:| +| [Distributed Data Parallel (DDP)](https://pytorch.org/docs/master/generated/torch.nn.parallel.DistributedDataParallel.html) | ✅ | +| [Automatic Mixed Precision (AMP)](https://pytorch.org/docs/stable/amp.html) | ✅ | +| [Torch Compile](https://pytorch.org/docs/stable/generated/torch.compile.html) | ✅ | +| [Gradient Checkpointing](https://pytorch.org/docs/stable/checkpoint.html) | ✅ | +| [Fully Sharded Data Parallel (FSDP)](https://pytorch.org/docs/stable/fsdp.html) | ✅ | + +
+ +The [examples](https://github.com/pomonam/kronfluence/tree/main/examples) folder contains several examples demonstrating how to use Kronfluence. + +## LogIX + +While Kronfluence supports influence function computations on large-scale models like `Meta-Llama-3-8B-Instruct`, for those +interested in running influence analysis on even larger models or with a large number of query data points, our +project [LogIX](https://github.com/logix-project/logix) may be worth exploring. It integrates with frameworks like +[HuggingFace Trainer](https://huggingface.co/docs/transformers/en/main_classes/trainer) and [PyTorch Lightning](https://lightning.ai/docs/pytorch/stable/) +and is also compatible with many PyTorch features (DDP & FSDP & [DeepSpeed](https://github.com/microsoft/DeepSpeed)). + ## Contributing Contributions are welcome! To get started, please review our [Code of Conduct](https://github.com/pomonam/kronfluence/blob/main/CODE_OF_CONDUCT.md). For bug fixes, please submit a pull request. @@ -131,10 +150,36 @@ cd kronfluence pip install -e ."[dev]" ``` +### Style Testing + +To maintain code quality and consistency, we run ruff and linting tests on pull requests. Before submitting a +pull request, please ensure that your code adheres to our formatting and linting guidelines. The following commands will +modify your code. It is recommended to create a Git commit before running them to easily revert any unintended changes. + +Sort import orderings using [isort](https://pycqa.github.io/isort/): + +```bash +isort kronfluence +``` + +Format code using [ruff](https://docs.astral.sh/ruff/): + +```bash +ruff format kronfluence +``` + +To view all [pylint](https://www.pylint.org/) complaints, run the following command: + +```bash +pylint kronfluence +``` + +Please address any reported issues before submitting your PR. + ## Acknowledgements [Omkar Dige](https://github.com/xeon27) contributed to the profiling, DDP, and FSDP utilities, and [Adil Asif](https://github.com/adil-a/) provided valuable insights and suggestions on structuring the DDP and FSDP implementations. -I also thank Hwijeen Ahn, Sang Keun Choe, Youngseog Chung, Minsoo Kang, Lev McKinney, Laura Ruis, Andrew Wang, and Kewen Zhao for their feedback. +I also thank Hwijeen Ahn, Sang Keun Choe, Youngseog Chung, Minsoo Kang, Sophie Liao, Lev McKinney, Laura Ruis, Andrew Wang, and Kewen Zhao for their feedback. ## License diff --git a/dev_requirements.txt b/dev_requirements.txt index 0cca66e..5d23191 100644 --- a/dev_requirements.txt +++ b/dev_requirements.txt @@ -1,14 +1,15 @@ -torch -torchvision -accelerate -einops -einconv -opt_einsum -safetensors -tqdm -datasets -transformers +torch>=2.1.0 +torchvision>=0.16.0 +accelerate>=0.31.0 +einops>=0.8.0 +einconv>=0.1.0 +opt_einsum>=3.3.0 +scikit-learn>=1.4.0 +safetensors>=0.4.2 +tqdm>=4.66.4 +datasets>=2.20.0 +transformers>=4.42.0 isort==5.13.2 -pylint==3.0.3 -pytest==8.0.0 -ruff==0.3.0 \ No newline at end of file +pylint==3.2.3 +pytest==8.2.2 +ruff==0.4.0 \ No newline at end of file diff --git a/examples/README.md b/examples/README.md new file mode 100644 index 0000000..328a50f --- /dev/null +++ b/examples/README.md @@ -0,0 +1,33 @@ +# Kronfluence: Examples + +For detailed technical documentation of Kronfluence, please refer to the [Technical Documentation](https://github.com/pomonam/kronfluence/blob/main/DOCUMENTATION.md) page. + +## Getting Started + +To run all examples, install the necessary packages: + +```bash +pip install -r requirements.txt +``` + +Alternatively, navigate to each example folder and run `pip install -r requirements.txt`. + +## List of Tasks + +Our examples cover the following tasks: + +
+ +| Task | Example Datasets | +|----------------------|:------------------------:| +| Regression | UCI | +| Image Classification | CIFAR-10 & ImageNet | +| Text Classification | GLUE | +| Multiple-Choice | SWAG | +| Summarization | DNN/DailyMail | +| Language Modeling | WikiText-2 & OpenWebText | + +
+ +These examples demonstrate various use cases of Kronfluence, including the usage of AMP (Automatic Mixed Precision) and DDP (Distributed Data Parallel). +Many examples aim to replicate the settings used in [our paper](https://arxiv.org/abs/2405.12186). If you would like to see more examples added to this repository, please leave an issue. \ No newline at end of file diff --git a/examples/cifar/README.md b/examples/cifar/README.md index 67b16d5..5474295 100644 --- a/examples/cifar/README.md +++ b/examples/cifar/README.md @@ -1,38 +1,109 @@ # CIFAR-10 & ResNet-9 Example -This directory contains scripts for training ResNet-9 on CIFAR-10. The pipeline is motivated from -[TRAK repository](https://github.com/MadryLab/trak/blob/main/examples/cifar_quickstart.ipynb). +This directory contains scripts for training ResNet-9 and computing influence scores on CIFAR-10 dataset. The pipeline is motivated from +[TRAK repository](https://github.com/MadryLab/trak/blob/main/examples/cifar_quickstart.ipynb). To get started, please install the necessary packages by running the following command: + +```bash +pip install -r requirements.txt +``` ## Training -To train ResNet-9 on CIFAR-10 dataset, run the following command: +To train ResNet-9 on the CIFAR-10 dataset, run the following command: + ```bash python train.py --dataset_dir ./data \ --checkpoint_dir ./checkpoints \ --train_batch_size 512 \ --eval_batch_size 1024 \ --learning_rate 0.4 \ - --weight_decay 0.0001 \ + --weight_decay 0.001 \ --num_train_epochs 25 \ --seed 1004 ``` +This will train the model using the specified hyperparameters and save the trained checkpoint in the `./checkpoints` directory. + ## Computing Pairwise Influence Scores -To obtain pairwise influence scores on 2000 query data points using `ekfac`, run the following command: +To compute pairwise influence scores on 2000 query data points using the `ekfac` strategy, run the following command: + ```bash python analyze.py --query_batch_size 1000 \ --dataset_dir ./data \ --checkpoint_dir ./checkpoints \ --factor_strategy ekfac ``` -You can also use `identity`, `diagonal`, and `kfac`. On A100 (80GB), it takes roughly 1.5 minutes to compute the -pairwise scores (including computing EKFAC factors). + +In addition to `ekfac`, you can also use `identity`, `diagonal`, and `kfac` as the `factor_strategy`. On an A100 (80GB) GPU, it takes roughly 2 minutes to compute the pairwise scores (including computing the EKFAC factors): + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 11 | 106.38 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Compute Pairwise Score | 46.745 | 1 | 46.745 | 43.941 | +| Fit Lambda | 34.885 | 1 | 34.885 | 32.793 | +| Fit Covariance | 22.538 | 1 | 22.538 | 21.187 | +| Perform Eigendecomposition | 0.91424 | 1 | 0.91424 | 0.85941 | +| Save Pairwise Score | 0.81219 | 1 | 0.81219 | 0.76348 | +| Save Covariance | 0.22351 | 1 | 0.22351 | 0.21011 | +| Save Eigendecomposition | 0.21617 | 1 | 0.21617 | 0.20321 | +| Save Lambda | 0.031038 | 1 | 0.031038 | 0.029177 | +| Load Eigendecomposition | 0.010442 | 1 | 0.010442 | 0.0098156 | +| Load All Factors | 0.0026517 | 1 | 0.0026517 | 0.0024927 | +| Load Covariance | 0.0016419 | 1 | 0.0016419 | 0.0015435 | +---------------------------------------------------------------------------------------------------------------------------------- +``` + +To use AMP when computing influence scores, run: + +```bash +python analyze.py --query_batch_size 1000 \ + --dataset_dir ./data \ + --checkpoint_dir ./checkpoints \ + --factor_strategy ekfac \ + --use_half_precision +``` + +This reduces computation time to about 40 seconds on an A100 (80GB) GPU: + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 11 | 35.965 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Compute Pairwise Score | 18.012 | 1 | 18.012 | 50.082 | +| Fit Lambda | 9.2271 | 1 | 9.2271 | 25.656 | +| Fit Covariance | 7.134 | 1 | 7.134 | 19.836 | +| Perform Eigendecomposition | 0.87962 | 1 | 0.87962 | 2.4457 | +| Save Pairwise Score | 0.45432 | 1 | 0.45432 | 1.2632 | +| Save Covariance | 0.12861 | 1 | 0.12861 | 0.35759 | +| Save Eigendecomposition | 0.11296 | 1 | 0.11296 | 0.31407 | +| Save Lambda | 0.010712 | 1 | 0.010712 | 0.029784 | +| Load All Factors | 0.002736 | 1 | 0.002736 | 0.0076074 | +| Load Covariance | 0.0016696 | 1 | 0.0016696 | 0.0046421 | +| Load Eigendecomposition | 0.0014892 | 1 | 0.0014892 | 0.0041406 | +---------------------------------------------------------------------------------------------------------------------------------- +``` + +You can run `half_precision_analysis.py` to verify that the scores computed with AMP have high correlations with those of the default configuration. + +

+Half Precision +

+ +## Visualizing Influential Training Images + +[This Colab notebook](https://colab.research.google.com/drive/1KIwIbeJh_om4tRwceuZ005fVKDsiXKgr?usp=sharing) provides a tutorial on visualizing the top influential training images. ## Mislabeled Data Detection -We can use self-influence scores (see Section 5.4 for the [paper](https://arxiv.org/pdf/1703.04730.pdf)) to detect mislabeled examples. -First, train the model with 10% of training examples mislabeled by running the following command: +We can use self-influence scores (see **Section 5.4** for the [paper](https://arxiv.org/pdf/1703.04730.pdf)) to detect mislabeled examples. +First, train the model with 10% of the training examples mislabeled by running: + ```bash python train.py --dataset_dir ./data \ --corrupt_percentage 0.1 \ @@ -40,12 +111,13 @@ python train.py --dataset_dir ./data \ --train_batch_size 512 \ --eval_batch_size 1024 \ --learning_rate 0.4 \ - --weight_decay 0.0001 \ + --weight_decay 0.001 \ --num_train_epochs 25 \ --seed 1004 ``` -Then, compute self-influence scores with the following command: +Then, compute the self-influence scores with: + ```bash python detect_mislabeled_dataset.py --dataset_dir ./data \ --corrupt_percentage 0.1 \ @@ -53,5 +125,30 @@ python detect_mislabeled_dataset.py --dataset_dir ./data \ --factor_strategy ekfac ``` -On A100 (80GB), it takes roughly 1.5 minutes to compute the self-influence scores. -We can detect around 82% of mislabeled data points by inspecting 10% of the dataset (96% by inspecting 20%). \ No newline at end of file +On an A100 (80GB) GPU, it takes roughly 2 minutes to compute the self-influence scores: + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 11 | 121.85 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Compute Self-Influence Score | 62.778 | 1 | 62.778 | 51.519 | +| Fit Lambda | 35.174 | 1 | 35.174 | 28.866 | +| Fit Covariance | 22.582 | 1 | 22.582 | 18.532 | +| Perform Eigendecomposition | 0.82656 | 1 | 0.82656 | 0.67832 | +| Save Covariance | 0.2478 | 1 | 0.2478 | 0.20336 | +| Save Eigendecomposition | 0.22042 | 1 | 0.22042 | 0.18088 | +| Save Lambda | 0.018463 | 1 | 0.018463 | 0.015152 | +| Load All Factors | 0.0027554 | 1 | 0.0027554 | 0.0022612 | +| Load Covariance | 0.0016607 | 1 | 0.0016607 | 0.0013628 | +| Load Eigendecomposition | 0.0015408 | 1 | 0.0015408 | 0.0012645 | +| Save Self-Influence Score | 0.0010841 | 1 | 0.0010841 | 0.00088966 | +---------------------------------------------------------------------------------------------------------------------------------- +``` + +Around 80% of mislabeled data points can be detected by inspecting 10% of the dataset (97% by inspecting 20%). + +

+Mislabeled Data Detection +

\ No newline at end of file diff --git a/examples/cifar/analyze.py b/examples/cifar/analyze.py index 8669924..b0d7d84 100644 --- a/examples/cifar/analyze.py +++ b/examples/cifar/analyze.py @@ -9,8 +9,10 @@ from examples.cifar.pipeline import construct_resnet9, get_cifar10_dataset from kronfluence.analyzer import Analyzer, prepare_model -from kronfluence.arguments import FactorArguments +from kronfluence.arguments import FactorArguments, ScoreArguments from kronfluence.task import Task +from kronfluence.utils.common.factor_arguments import all_low_precision_factor_arguments +from kronfluence.utils.common.score_arguments import all_low_precision_score_arguments from kronfluence.utils.dataset import DataLoaderKwargs BATCH_TYPE = Tuple[torch.Tensor, torch.Tensor] @@ -38,6 +40,12 @@ def parse_args(): help="A path that is storing the final checkpoint of the model.", ) + parser.add_argument( + "--factor_strategy", + type=str, + default="ekfac", + help="Strategy to compute influence factors.", + ) parser.add_argument( "--query_batch_size", type=int, @@ -45,12 +53,17 @@ def parse_args(): help="Batch size for computing query gradients.", ) parser.add_argument( - "--factor_strategy", - type=str, - default="ekfac", - help="Strategy to compute influence factors.", + "--use_half_precision", + action="store_true", + default=False, + help="Whether to use half precision for computing factors and scores.", + ) + parser.add_argument( + "--profile", + action="store_true", + default=False, + help="Boolean flag to profile computations.", ) - args = parser.parse_args() if args.checkpoint_dir is not None: @@ -71,12 +84,12 @@ def compute_train_loss( if not sample: return F.cross_entropy(logits, labels, reduction="sum") with torch.no_grad(): - probs = torch.nn.functional.softmax(logits, dim=-1) + probs = torch.nn.functional.softmax(logits.detach(), dim=-1) sampled_labels = torch.multinomial( probs, num_samples=1, ).flatten() - return F.cross_entropy(logits, sampled_labels.detach(), reduction="sum") + return F.cross_entropy(logits, sampled_labels, reduction="sum") def compute_measurement( self, @@ -125,31 +138,43 @@ def main(): analysis_name="cifar10", model=model, task=task, + profile=args.profile, ) # Configure parameters for DataLoader. dataloader_kwargs = DataLoaderKwargs(num_workers=4) analyzer.set_dataloader_kwargs(dataloader_kwargs) # Compute influence factors. + factors_name = args.factor_strategy factor_args = FactorArguments(strategy=args.factor_strategy) + if args.use_half_precision: + factor_args = all_low_precision_factor_arguments(strategy=args.factor_strategy, dtype=torch.bfloat16) + factors_name += "_half" analyzer.fit_all_factors( - factors_name=args.factor_strategy, + factors_name=factors_name, + factor_args=factor_args, dataset=train_dataset, per_device_batch_size=None, - factor_args=factor_args, overwrite_output_dir=False, ) + # Compute pairwise scores. + score_args = ScoreArguments() + scores_name = factor_args.strategy + if args.use_half_precision: + score_args = all_low_precision_score_arguments(dtype=torch.bfloat16) + scores_name += "_half" analyzer.compute_pairwise_scores( - scores_name=args.factor_strategy, - factors_name=args.factor_strategy, + scores_name=scores_name, + score_args=score_args, + factors_name=factors_name, query_dataset=eval_dataset, query_indices=list(range(2000)), train_dataset=train_dataset, per_device_query_batch_size=args.query_batch_size, overwrite_output_dir=False, ) - scores = analyzer.load_pairwise_scores(args.factor_strategy)["all_modules"] + scores = analyzer.load_pairwise_scores(scores_name)["all_modules"] logging.info(f"Scores shape: {scores.shape}") diff --git a/examples/cifar/detect_mislabeled_dataset.py b/examples/cifar/detect_mislabeled_dataset.py index efcd84f..3884974 100644 --- a/examples/cifar/detect_mislabeled_dataset.py +++ b/examples/cifar/detect_mislabeled_dataset.py @@ -39,7 +39,12 @@ def parse_args(): default="ekfac", help="Strategy to compute influence factors.", ) - + parser.add_argument( + "--profile", + action="store_true", + default=False, + help="Boolean flag to profile computations.", + ) args = parser.parse_args() if args.checkpoint_dir is not None: @@ -75,6 +80,7 @@ def main(): analysis_name="mislabeled", model=model, task=task, + profile=args.profile, ) # Configure parameters for DataLoader. dataloader_kwargs = DataLoaderKwargs(num_workers=4) @@ -89,14 +95,15 @@ def main(): factor_args=factor_args, overwrite_output_dir=False, ) + # Compute self-influence scores. analyzer.compute_self_scores( scores_name=args.factor_strategy, factors_name=args.factor_strategy, train_dataset=train_dataset, - overwrite_output_dir=True, + overwrite_output_dir=False, ) - scores = analyzer.load_pairwise_scores(args.factor_strategy)["all_modules"] + scores = analyzer.load_self_scores(args.factor_strategy)["all_modules"] total_corrupt_size = int(args.corrupt_percentage * len(train_dataset)) corrupted_indices = list(range(int(args.corrupt_percentage * len(train_dataset)))) diff --git a/examples/cifar/figure/half_precision.png b/examples/cifar/figure/half_precision.png new file mode 100644 index 0000000..fc0ac2b Binary files /dev/null and b/examples/cifar/figure/half_precision.png differ diff --git a/examples/cifar/figure/mislabel.png b/examples/cifar/figure/mislabel.png new file mode 100644 index 0000000..2bb9eb4 Binary files /dev/null and b/examples/cifar/figure/mislabel.png differ diff --git a/examples/cifar/half_precision_analysis.py b/examples/cifar/half_precision_analysis.py new file mode 100644 index 0000000..3f9a41d --- /dev/null +++ b/examples/cifar/half_precision_analysis.py @@ -0,0 +1,45 @@ +import logging + +import matplotlib.pyplot as plt +import numpy as np +from scipy.stats import spearmanr +from tueplots import markers + +from kronfluence.analyzer import Analyzer + + +def main(): + logging.basicConfig(level=logging.INFO) + + # Load the scores. You might need to modify the path. + scores = ( + Analyzer.load_file("influence_results/cifar10/scores_ekfac/pairwise_scores.safetensors")["all_modules"] / 50_000 + ) + half_scores = ( + Analyzer.load_file("influence_results/cifar10/scores_ekfac_half/pairwise_scores.safetensors")[ + "all_modules" + ].float() + / 50_000 + ) + + plt.rcParams.update({"figure.dpi": 150}) + plt.rcParams.update(markers.with_edge()) + plt.rcParams["axes.axisbelow"] = True + + # Only plot first 3000 points to avoid clutter. + idx = 79 + plt.scatter(half_scores[idx][:3000], scores[idx][:3000], edgecolor="k") + plt.grid() + plt.xlabel("bfloat16") + plt.ylabel("float32") + plt.show() + + # Compute the averaged spearman correlation. + all_corr = [] + for i in range(100): + all_corr.append(spearmanr(scores[i], half_scores[i])[0]) + logging.info(f"Averaged Spearman Correlation: {np.array(all_corr).mean()}") + + +if __name__ == "__main__": + main() diff --git a/examples/cifar/inspect_factors.py b/examples/cifar/inspect_factors.py new file mode 100644 index 0000000..9c78059 --- /dev/null +++ b/examples/cifar/inspect_factors.py @@ -0,0 +1,24 @@ +import logging + +import matplotlib.pyplot as plt + +from kronfluence.analyzer import Analyzer + + +def main(): + logging.basicConfig(level=logging.INFO) + + name = "ekfac" + factor = Analyzer.load_file(f"influence_results/cifar10/factors_{name}/activation_covariance.safetensors") + + plt.matshow(factor["6.0"]) + plt.show() + + factor = Analyzer.load_file(f"influence_results/cifar10/factors_{name}/gradient_covariance.safetensors") + + plt.matshow(factor["6.0"]) + plt.show() + + +if __name__ == "__main__": + main() diff --git a/examples/cifar/requirements.txt b/examples/cifar/requirements.txt new file mode 100644 index 0000000..a667c1f --- /dev/null +++ b/examples/cifar/requirements.txt @@ -0,0 +1,3 @@ +scikit-learn +matplotlib +tueplots \ No newline at end of file diff --git a/examples/cifar/train.py b/examples/cifar/train.py index 1d76f79..e78e5f2 100644 --- a/examples/cifar/train.py +++ b/examples/cifar/train.py @@ -55,7 +55,7 @@ def parse_args(): parser.add_argument( "--weight_decay", type=float, - default=0.0001, + default=0.001, help="Weight decay to train the model.", ) parser.add_argument( @@ -77,7 +77,6 @@ def parse_args(): default="./checkpoints", help="A path to store the final checkpoint.", ) - args = parser.parse_args() if args.checkpoint_dir is not None: @@ -98,6 +97,7 @@ def train( batch_size=batch_size, shuffle=True, drop_last=True, + num_workers=4, ) model = construct_resnet9().to(DEVICE) @@ -117,7 +117,7 @@ def train( for epoch in range(num_train_epochs): total_loss = 0.0 for batch in train_dataloader: - model.zero_grad() + optimizer.zero_grad(set_to_none=True) inputs, labels = batch inputs, labels = inputs.to(DEVICE), labels.to(DEVICE) outputs = model(inputs) diff --git a/examples/dailymail/README.md b/examples/dailymail/README.md new file mode 100644 index 0000000..a7e5ab6 --- /dev/null +++ b/examples/dailymail/README.md @@ -0,0 +1,73 @@ +# CNN/DailyMail & T5 Example + +This directory contains scripts for fine-tuning T5 and computing influence scores on the CNN/DailyMail dataset. The pipeline is motivated from [this HuggingFace Example](https://github.com/huggingface/transformers/tree/main/examples/pytorch/summarization). +To begin, install the necessary packages: + +```bash +pip install -r requirements.txt +``` + +## Training + +To fine-tune T5 on CNN/DailyMail, run the following command: + +```bash +python train.py --checkpoint_dir ./checkpoints \ + --train_batch_size 16 \ + --eval_batch_size 32 \ + --learning_rate 5e-05 \ + --weight_decay 0.01 \ + --num_train_epochs 3 \ + --seed 1004 +``` + +This will fine-tune the model using the specified hyperparameters and save the final checkpoint in the `./checkpoints` directory. + +## Computing Pairwise Influence Scores + +To calculate pairwise influence scores on 10 query data points using `ekfac`, run: + +```bash +python analyze.py --factor_batch_size 64 \ + --query_batch_size 10 \ + --train_batch_size 128 \ + --use_half_precision \ + --checkpoint_dir ./checkpoints \ + --factor_strategy ekfac +``` + +Alternative options for `factor_strategy` include `identity`, `diagonal`, and `kfac`. On an A100 (80GB), computing the pairwise scores (including EKFAC factors) takes approximately 1 hour: + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 11 | 3397.1 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Compute Pairwise Score | 1905.6 | 1 | 1905.6 | 56.093 | +| Fit Lambda | 747.5 | 1 | 747.5 | 22.004 | +| Fit Covariance | 734.03 | 1 | 734.03 | 21.607 | +| Perform Eigendecomposition | 8.4236 | 1 | 8.4236 | 0.24796 | +| Save Eigendecomposition | 0.79164 | 1 | 0.79164 | 0.023303 | +| Save Covariance | 0.60366 | 1 | 0.60366 | 0.01777 | +| Save Lambda | 0.1514 | 1 | 0.1514 | 0.0044566 | +| Load All Factors | 0.027977 | 1 | 0.027977 | 0.00082354 | +| Save Pairwise Score | 0.01082 | 1 | 0.01082 | 0.00031851 | +| Load Covariance | 0.010015 | 1 | 0.010015 | 0.0002948 | +| Load Eigendecomposition | 0.0096806 | 1 | 0.0096806 | 0.00028497 | +---------------------------------------------------------------------------------------------------------------------------------- +``` + +## Inspecting Top Influential Sequences + +The `inspect_examples.py` script prints top influential sequences for a given query. + +``` +Query Data Example: + Input: summarize: (CNN)My vote for Father of the Year goes to Curt Schilling. The former Major League Baseball pitcher recently fired off a series of fastballs and mowed down a group of Twitter trolls who made the mistake of tweeting vulgar and sexually-explicit comments about Schilling's teenage daughter. The drama started, innocently enough, on February 25, when Schilling played the role of a proud father. He sent a tweet congratulating his daughter, Gabby, on being accepted to Salve Regina University, where she'll play softball. It read: "Congrats to Gabby Schilling who will pitch for the Salve Regina Seahawks next year!! — Curt Schilling (@gehrig38)" Almost immediately, responses came in from young men, complete strangers who apparently followed Schilling on Twitter. The tweets quickly went from immature, to creepy, to repugnant. Threats of rape were common. The tweets were deleted, and the accounts were closed after this story went viral. But not before Schilling captured some of the images and posted them on his blog. What was said about 17-year-old Gabby Schilling wasn't just obnoxious. It was vile and obscene. What was said wasn't just mean and ugly. It was threatening and scary. As a parent, it's the kind of thing that makes you rethink your opposition to public caning as a logical punishment for such transgressions. These misogynistic cowards may have thought they could hide in the darkness of anonymity, the sort that many have come to expect from social media sites, where you feel free to be a despicable human being because, you think, no one will ever find out who you really are and hold you accountable for your words. If so, they thought wrong. They couldn't hide. They were found out, and they got the throttling they so richly deserved. Thanks to dad. According to Schilling, who made it his mission to track down these cretins and make sure those they associate with know who they really are, two people have already paid a price due to their tweets. One was a student disc jockey at a community college in New Jersey, who was suspended, and the other was a part-time ticket seller for the New York Yankees, who was fired. Concerned that this is an example of exactly the kind of cyberbullying that leads some teenagers to commit suicide, Schilling is also thinking about taking legal action against some of the other people involved. Bravo for him. I'm sure that, all across America, dads with daughters -- after reading some of the horrible things that were said about this young girl -- are marveling at Schilling's self-control. I have two daughters of my own, and he's a better man than me. If ever there was a case where profanity-spewing malcontents deserved to have their mouths washed out with soap, this is it. So what additional insights can we draw, and what larger lessons can we learn, from this unexpected but predictable collision of old-fashioned parenthood and newfangled media? There are a few. The first is about accountability, the very thing that the young men who posted these hurtful messages were trying to avoid. But Schilling wouldn't let them. At their best, social media sites like Twitter, Facebook, Instagram and others allow the sharing the information and the building of a sense of community. At their worst, they become digital sandboxes and locker rooms where people think have a license to misbehave without having to worry about consequences. We need to applaud efforts like this that promote greater online accountability. There's also something to be said about protective parents, and how essential they are to a working society. We should still be concerned about those overprotective parents who hover like helicopters from little league to job interviews. We shouldn't bubblewrap our kids, and keep them from playing outdoors, and then sit around wondering why they're soft, timid, and risk-averse. But protective parents -- the kind who shield their kids from real danger -- never go out of style. A parent's top job is to protect his children. Schilling did his job. Finally, it's worth reminding everyone that freedom of expression does not mean freedom from rules, standards, and expectations that should guide your behavior. There are things you don't say. There are boundaries, ways that we expect you to behave so you don't terrorize other people or bring shame upon yourself, your friends, and + Label: Ruben Navarrette: Schilling deserves praise for taking on online haters for offensive comments about his daughter. Navarrette: In protecting his child, Schilling set a model for parenting and taught us a lesson about social media. + +Top Influential Example: + Input: summarize: (CNN) -- What is it with juries in high-profile cases in Southern California? Over the years, they've become a national joke. But no one is laughing. Instead, with each travesty of justice and every acquittal that should have been a conviction, you're left wondering just what trial these 12 folks were watching and questioning whether we should have higher standards for who sits on a jury. Sometimes, the juries in local and state courts get it wrong, and the Justice Department must step in and make it right. Think back to the acquittal in April 1992 of four Los Angeles police officers who, one year earlier, savagely beat motorist Rodney King. They walked out of a courtroom in Simi Valley, California, as free men -- sparking days of rioting, looting and violence. At the time, the conventional thinking on newspaper editorial pages and on talk radio was the jurors in that largely white suburb of Los Angeles, which was itself home to many active-duty and retired police officers, saw the police force as their line of defense against undesirables like King. So naturally, the argument went, they would cut them some slack. The officers were tried again, and convicted in federal court of violating King's civil rights. Justice was finally served. Here we go again. There hasn't been much civil unrest over what happened to Kelly Thomas, the homeless and mentally ill man who -- on July 5, 2011 -- was beaten to death by a swarm of police officers in Fullerton, California. But now that the verdict is in, literally, on the two former officers who were charged in his death, there is plenty of outrage on talk radio, online and in other public forums. Another 12 people who swore an oath to consider the evidence and the law and make sure that justice is served appear to have gotten it terribly wrong. This week, that jury in Santa Ana, California -- a city about 30 miles southeast of Los Angeles -- produced a wave of gasps in the courtroom when it announced that it had found Manuel Ramos, who had been charged with second-degree murder and involuntary manslaughter, and Jay Cicinelli, who was charged with involuntary manslaughter and excessive use of force, not guilty on all counts. What? The beating was caught on a surveillance tape. When you watch those 33 minutes of footage, assuming you can stomach the experience, it's hard to believe that anyone could declare the perpetrators "not guilty." The surveillance camera footage shows Thomas being beaten and stunned with a Taser by police until he was unrecognizable and unconscious. You see a defenseless and compliant young man screaming in pain, saying he's sorry and pleading for help from his father. His words will haunt you, "Daddy, help! They're killing me!" According to prosecutors, the young man suffered brain injuries, facial fractures, broken ribs and extensive bruises and abrasions. He wound up lying in a pool of blood. He died five days later. This was not a by-the-book case of police officers using all necessary force to subdue a suspect who was resisting arrest -- a suspect, by the way, who had committed no crime. This was not, as Ramos' attorney claimed, a case of police offices simply "doing their job" with "no malice in their heart." Check the video. Early on in the confrontation, Ramos appears to tell the young man who is sitting on the ground: "You see my fists? They're getting ready to f--- you up!" Another officer is heard telling a comrade: "We ran out of options so I got to the end of my Taser and I... smashed his face to hell." There is the malice. This was abuse of power and an instance of bullying behind a badge. It happens more than we'd like to think in America. But this time, it went too far. And a man died, and a family was shattered. Yet, the jury somehow missed all this? How does this happen? In Los Angeles, people are saying that the mentally ill are the new Rodney King. In the same way that the jury in Simi Valley was inclined to back the officers who it saw as protecting them from people like King, now the jury in Santa Ana is backing the officers who it counts on to prod people like Thomas to move along, leave the streets, and get out of sight. It's a plausible explanation + Label: Ruben Navarrette: Too many high-profile cases in California produce travesties of justice. He says jury acquitted two ex-cops in malicious beating death captured on video. He says case showed abuse of power, bullying behind badge; happens too often in U.S. Navarrette: Only one place left that can right this wrong: The Justice Department. +``` \ No newline at end of file diff --git a/examples/dailymail/__init__.py b/examples/dailymail/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/examples/dailymail/analyze.py b/examples/dailymail/analyze.py new file mode 100644 index 0000000..032cc08 --- /dev/null +++ b/examples/dailymail/analyze.py @@ -0,0 +1,275 @@ +import argparse +import logging +import os +from typing import Dict, List + +import torch +import torch.nn.functional as F +from torch import nn +from transformers import DataCollatorForSeq2Seq + +from examples.dailymail.pipeline import ( + construct_t5, + get_dailymail_dataset, + get_tokenizer, +) +from kronfluence.analyzer import Analyzer, prepare_model +from kronfluence.arguments import FactorArguments, ScoreArguments +from kronfluence.task import Task +from kronfluence.utils.common.factor_arguments import all_low_precision_factor_arguments +from kronfluence.utils.common.score_arguments import all_low_precision_score_arguments +from kronfluence.utils.dataset import DataLoaderKwargs +from kronfluence.utils.model import apply_ddp + +BATCH_TYPE = Dict[str, torch.Tensor] +try: + LOCAL_RANK = int(os.environ["LOCAL_RANK"]) + WORLD_RANK = int(os.environ["RANK"]) + WORLD_SIZE = int(os.environ["WORLD_SIZE"]) +except KeyError: + LOCAL_RANK = WORLD_RANK = WORLD_SIZE = 0 + + +def parse_args(): + parser = argparse.ArgumentParser(description="Influence analysis on CNN/DailyMail dataset.") + + parser.add_argument( + "--checkpoint_dir", + type=str, + default="./checkpoints", + help="A path that is storing the final checkpoint of the model.", + ) + + parser.add_argument( + "--factor_strategy", + type=str, + default="ekfac", + help="Strategy to compute influence factors.", + ) + parser.add_argument( + "--query_gradient_rank", + type=int, + default=-1, + help="Rank for the low-rank query gradient approximation.", + ) + parser.add_argument( + "--use_half_precision", + action="store_true", + default=False, + help="Whether to use half precision for computing factors and scores.", + ) + parser.add_argument( + "--use_ddp", + action="store_true", + default=False, + help="Whether to use DDP for computing factors and scores.", + ) + parser.add_argument( + "--factor_batch_size", + type=int, + default=64, + help="Batch size for computing influence factors.", + ) + parser.add_argument( + "--query_batch_size", + type=int, + default=10, + help="Batch size for computing query gradients.", + ) + parser.add_argument( + "--train_batch_size", + type=int, + default=128, + help="Batch size for computing training gradients.", + ) + parser.add_argument( + "--profile", + action="store_true", + default=False, + help="Boolean flag to profile computations.", + ) + args = parser.parse_args() + + if args.checkpoint_dir is not None: + os.makedirs(args.checkpoint_dir, exist_ok=True) + + return args + + +class SummarizationTask(Task): + def compute_train_loss( + self, + batch: BATCH_TYPE, + model: nn.Module, + sample: bool = False, + ) -> torch.Tensor: + logits = model( + input_ids=batch["input_ids"], + attention_mask=batch["attention_mask"], + decoder_input_ids=batch["decoder_input_ids"], + ).logits + + if not sample: + return F.cross_entropy( + logits.view(-1, logits.size(-1)), batch["labels"].view(-1), ignore_index=-100, reduction="sum" + ) + with torch.no_grad(): + probs = torch.nn.functional.softmax(logits.view(-1, logits.size(-1)).detach(), dim=-1) + sampled_labels = torch.multinomial( + probs, + num_samples=1, + ).flatten() + masks = batch["labels"].view(-1) == -100 + sampled_labels[masks] = -100 + return F.cross_entropy(logits.view(-1, logits.size(-1)), sampled_labels, reduction="sum") + + def compute_measurement( + self, + batch: BATCH_TYPE, + model: nn.Module, + ) -> torch.Tensor: + # Copied from: https://github.com/MadryLab/trak/blob/main/trak/modelout_functions.py. + logits = model( + input_ids=batch["input_ids"], + attention_mask=batch["attention_mask"], + decoder_input_ids=batch["decoder_input_ids"], + ).logits + logits = logits.view(-1, logits.size(-1)) + + labels = batch["labels"].view(-1) + bindex = torch.arange(logits.shape[0]).to(device=logits.device, non_blocking=False) + logits_correct = logits[bindex, labels] + + cloned_logits = logits.clone() + cloned_logits[bindex, labels] = torch.tensor(-torch.inf, device=logits.device, dtype=logits.dtype) + + margins = logits_correct - cloned_logits.logsumexp(dim=-1) + masks = batch["labels"].view(-1) != -100 + return -margins[masks].sum() + + def get_influence_tracked_modules(self) -> List[str]: + total_modules = [] + + # Add attention layers: + for i in range(6): + total_modules.append(f"encoder.block.{i}.layer.0.SelfAttention.q") + total_modules.append(f"encoder.block.{i}.layer.0.SelfAttention.k") + total_modules.append(f"encoder.block.{i}.layer.0.SelfAttention.v") + total_modules.append(f"encoder.block.{i}.layer.0.SelfAttention.o") + + total_modules.append(f"decoder.block.{i}.layer.0.SelfAttention.q") + total_modules.append(f"decoder.block.{i}.layer.0.SelfAttention.k") + total_modules.append(f"decoder.block.{i}.layer.0.SelfAttention.v") + total_modules.append(f"decoder.block.{i}.layer.0.SelfAttention.o") + + total_modules.append(f"decoder.block.{i}.layer.1.EncDecAttention.q") + total_modules.append(f"decoder.block.{i}.layer.1.EncDecAttention.k") + total_modules.append(f"decoder.block.{i}.layer.1.EncDecAttention.v") + total_modules.append(f"decoder.block.{i}.layer.1.EncDecAttention.o") + + # Add MLP layers: + for i in range(6): + total_modules.append(f"encoder.block.{i}.layer.1.DenseReluDense.wi") + total_modules.append(f"encoder.block.{i}.layer.1.DenseReluDense.wo") + + total_modules.append(f"decoder.block.{i}.layer.2.DenseReluDense.wi") + total_modules.append(f"decoder.block.{i}.layer.2.DenseReluDense.wo") + + return total_modules + + def get_attention_mask(self, batch: BATCH_TYPE) -> torch.Tensor: + return batch["attention_mask"] + + +def main(): + args = parse_args() + logging.basicConfig(level=logging.INFO) + + # Prepare the dataset. + train_dataset = get_dailymail_dataset( + split="eval_train", + ) + eval_dataset = get_dailymail_dataset( + split="valid", + ) + tokenizer = get_tokenizer() + + # Prepare the trained model. + model = construct_t5() + + # Define task and prepare model. + task = SummarizationTask() + model = prepare_model(model, task) + + if args.use_ddp: + model = apply_ddp( + model=model, + local_rank=LOCAL_RANK, + rank=WORLD_RANK, + world_size=WORLD_SIZE, + ) + + analyzer = Analyzer( + analysis_name="dailymail", + model=model, + task=task, + profile=args.profile, + ) + # Configure parameters for DataLoader. + label_pad_token_id = -100 + data_collator = DataCollatorForSeq2Seq( + tokenizer, + model=model, + label_pad_token_id=label_pad_token_id, + pad_to_multiple_of=None, + ) + + dataloader_kwargs = DataLoaderKwargs(num_workers=4, collate_fn=data_collator) + analyzer.set_dataloader_kwargs(dataloader_kwargs) + + # Compute influence factors. + factors_name = args.factor_strategy + factor_args = FactorArguments(strategy=args.factor_strategy) + if args.use_half_precision: + factor_args = all_low_precision_factor_arguments(strategy=args.factor_strategy, dtype=torch.bfloat16) + factors_name += "_half" + if args.use_ddp: + factors_name += "_ddp" + analyzer.fit_all_factors( + factors_name=factors_name, + dataset=train_dataset, + per_device_batch_size=args.factor_batch_size, + factor_args=factor_args, + overwrite_output_dir=False, + ) + + # Compute pairwise scores. + score_args = ScoreArguments() + scores_name = factor_args.strategy + if args.use_half_precision: + score_args = all_low_precision_score_arguments(dtype=torch.bfloat16) + scores_name += "_half" + rank = args.query_gradient_rank if args.query_gradient_rank != -1 else None + if rank is not None: + score_args.query_gradient_low_rank = rank + score_args.query_gradient_accumulation_steps = 10 + scores_name += f"_qlr{rank}" + if args.use_ddp: + scores_name += "_ddp" + analyzer.compute_pairwise_scores( + score_args=score_args, + scores_name=scores_name, + factors_name=factors_name, + query_dataset=eval_dataset, + query_indices=list(range(10)), + train_dataset=train_dataset, + per_device_query_batch_size=args.query_batch_size, + per_device_train_batch_size=args.train_batch_size, + overwrite_output_dir=False, + ) + scores = analyzer.load_pairwise_scores(scores_name)["all_modules"] + logging.info(f"Scores shape: {scores.shape}") + + +if __name__ == "__main__": + main() diff --git a/examples/dailymail/inspect_examples.py b/examples/dailymail/inspect_examples.py new file mode 100644 index 0000000..2059b5c --- /dev/null +++ b/examples/dailymail/inspect_examples.py @@ -0,0 +1,37 @@ +import logging + +import torch + +from examples.dailymail.pipeline import get_dailymail_dataset, get_tokenizer +from kronfluence.analyzer import Analyzer + + +def main(): + logging.basicConfig(level=logging.INFO) + + # You might need to change the path. + strategy = "ekfac" + scores = Analyzer.load_file(f"influence_results/dailymail/scores_{strategy}_half/pairwise_scores.safetensors")[ + "all_modules" + ].to(dtype=torch.float32) + + eval_idx = 1 + train_dataset = get_dailymail_dataset( + split="eval_train", + ) + eval_dataset = get_dailymail_dataset( + split="valid", + ) + tokenizer = get_tokenizer() + print("Query Data Example:") + print(f"Input: {tokenizer.decode(eval_dataset[eval_idx]['input_ids'])}") + print(f"Label: {tokenizer.decode(eval_dataset[eval_idx]['labels'])}") + + top_idx = int(torch.argsort(scores[eval_idx], descending=True)[0]) + print("Top Influential Example:") + print(f"Input: {tokenizer.decode(train_dataset[top_idx]['input_ids'])}") + print(f"Label: {tokenizer.decode(train_dataset[top_idx]['labels'])}") + + +if __name__ == "__main__": + main() diff --git a/examples/dailymail/pipeline.py b/examples/dailymail/pipeline.py new file mode 100644 index 0000000..f221cdb --- /dev/null +++ b/examples/dailymail/pipeline.py @@ -0,0 +1,110 @@ +from typing import Any, List + +import torch.nn as nn +from datasets import Dataset, load_dataset +from transformers import AutoConfig, AutoModelForSeq2SeqLM, AutoTokenizer + +summarization_name_mapping = { + "amazon_reviews_multi": ("review_body", "review_title"), + "big_patent": ("description", "abstract"), + "cnn_dailymail": ("article", "highlights"), + "orange_sum": ("text", "summary"), + "pn_summary": ("article", "summary"), + "psc": ("extract_text", "summary_text"), + "samsum": ("dialogue", "summary"), + "thaisum": ("body", "summary"), + "xglue": ("news_body", "news_title"), + "xsum": ("document", "summary"), + "wiki_summary": ("article", "highlights"), + "multi_news": ("document", "summary"), +} + + +MODEL_NAME = "google-t5/t5-small" + + +def construct_t5() -> nn.Module: + config = AutoConfig.from_pretrained( + MODEL_NAME, + trust_remote_code=True, + ) + return AutoModelForSeq2SeqLM.from_pretrained( + MODEL_NAME, + from_tf=False, + config=config, + ignore_mismatched_sizes=False, + trust_remote_code=True, + ) + + +def get_tokenizer() -> Any: + return AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=True, trust_remote_code=True) + + +def get_dailymail_dataset( + split: str, + indices: List[int] = None, +) -> Dataset: + raw_datasets = load_dataset("cnn_dailymail", "3.0.0") + + tokenizer = get_tokenizer() + column_names = raw_datasets["train"].column_names + dataset_columns = summarization_name_mapping.get("cnn_dailymail", None) + text_column = dataset_columns[0] if dataset_columns is not None else column_names[0] + summary_column = dataset_columns[1] if dataset_columns is not None else column_names[1] + + max_source_length = 1024 + max_target_length = 128 + padding = False + prefix = "summarize: " + + def preprocess_function(examples): + inputs = examples[text_column] + targets = examples[summary_column] + inputs = [prefix + inp for inp in inputs] + model_inputs = tokenizer(inputs, max_length=max_source_length, padding=padding, truncation=True) + + labels = tokenizer( + text_target=targets, + max_length=max_target_length, + padding=padding, + truncation=True, + ) + + model_inputs["labels"] = labels["input_ids"] + return model_inputs + + if split == "train" or split == "eval_train": + train_dataset = raw_datasets["train"] + train_dataset = train_dataset.map( + preprocess_function, + batched=True, + num_proc=None, + remove_columns=column_names, + load_from_cache_file=True, + desc="Running tokenizer on dataset.", + ) + ds = train_dataset + else: + valid_dataset = raw_datasets["validation"] + eval_dataset = valid_dataset.map( + preprocess_function, + batched=True, + num_proc=None, + remove_columns=column_names, + load_from_cache_file=True, + desc="Running tokenizer on dataset.", + ) + ds = eval_dataset + + if indices is not None: + ds = ds.select(indices) + + return ds + + +if __name__ == "__main__": + from kronfluence import Analyzer + + model = construct_t5() + print(Analyzer.get_module_summary(model)) diff --git a/examples/dailymail/requirements.txt b/examples/dailymail/requirements.txt new file mode 100644 index 0000000..7ae6637 --- /dev/null +++ b/examples/dailymail/requirements.txt @@ -0,0 +1,7 @@ +sentencepiece!=0.1.92 +nltk +py7zr +rouge-score +transformers +evaluate +datasets diff --git a/examples/dailymail/train.py b/examples/dailymail/train.py new file mode 100644 index 0000000..686b3d4 --- /dev/null +++ b/examples/dailymail/train.py @@ -0,0 +1,228 @@ +import argparse +import logging +import os +import time +from typing import Any, Dict + +import evaluate +import nltk +import numpy as np +import torch +import torch.nn.functional as F +from accelerate.utils import send_to_device, set_seed +from filelock import FileLock +from torch import nn +from torch.nn import CrossEntropyLoss +from torch.utils import data +from transformers import DataCollatorForSeq2Seq + +from examples.dailymail.pipeline import ( + construct_t5, + get_dailymail_dataset, + get_tokenizer, +) + +try: + nltk.data.find("tokenizers/punkt") +except (LookupError, OSError): + with FileLock(".lock") as lock: + nltk.download("punkt", quiet=True) + + +DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu") + + +def parse_args(): + parser = argparse.ArgumentParser(description="Train seq2seq models on CNN/DailyMail dataset.") + + parser.add_argument( + "--train_batch_size", + type=int, + default=16, + help="Batch size for the training dataloader.", + ) + parser.add_argument( + "--eval_batch_size", + type=int, + default=32, + help="Batch size for the evaluation dataloader.", + ) + + parser.add_argument( + "--learning_rate", + type=float, + default=5e-05, + help="Fixed learning rate to train the model.", + ) + parser.add_argument( + "--weight_decay", + type=float, + default=0.01, + help="Weight decay to train the model.", + ) + parser.add_argument( + "--num_train_epochs", + type=int, + default=3, + help="Total number of epochs to train the model.", + ) + + parser.add_argument( + "--seed", + type=int, + default=1004, + help="A seed for reproducible training pipeline.", + ) + parser.add_argument( + "--checkpoint_dir", + type=str, + default="./checkpoints", + help="A path to store the final checkpoint.", + ) + args = parser.parse_args() + + if args.checkpoint_dir is not None: + os.makedirs(args.checkpoint_dir, exist_ok=True) + + return args + + +def train( + dataset: data.Dataset, + tokenizer: Any, + batch_size: int, + num_train_epochs: int, + learning_rate: float, + weight_decay: float, +) -> nn.Module: + model = construct_t5().to(DEVICE) + data_collator = DataCollatorForSeq2Seq( + tokenizer, + model=model, + label_pad_token_id=-100, + pad_to_multiple_of=None, + ) + train_dataloader = data.DataLoader( + dataset=dataset, + batch_size=batch_size, + shuffle=True, + drop_last=True, + collate_fn=data_collator, + ) + optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate, weight_decay=weight_decay) + + start_time = time.time() + model.train() + for epoch in range(num_train_epochs): + total_loss = 0.0 + for batch in train_dataloader: + optimizer.zero_grad(set_to_none=True) + batch = send_to_device(batch, device=DEVICE) + loss = model(**batch).loss + loss.backward() + optimizer.step() + total_loss += loss.detach().float() + logging.info(f"Epoch {epoch + 1} - Averaged Loss: {total_loss / len(dataset)}") + end_time = time.time() + elapsed_time = end_time - start_time + logging.info(f"Completed training in {elapsed_time:.2f} seconds.") + return model + + +def evaluate_model(model: nn.Module, tokenizer: Any, dataset: data.Dataset, batch_size: int) -> Dict[str, Any]: + data_collator = DataCollatorForSeq2Seq( + tokenizer, + model=model, + label_pad_token_id=-100, + pad_to_multiple_of=None, + ) + dataloader = data.DataLoader( + dataset=dataset, batch_size=batch_size, shuffle=False, drop_last=False, collate_fn=data_collator + ) + model.eval() + + def postprocess_text(preds, labels): + preds = [pred.strip() for pred in preds] + labels = [label.strip() for label in labels] + + # rougeLSum expects newline after each sentence. + preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds] + labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels] + + return preds, labels + + gen_kwargs = { + "max_length": 128, + } + metric = evaluate.load("rouge") + loss_fn = CrossEntropyLoss(ignore_index=-100, reduction="mean") + total_loss = 0.0 + for step, batch in enumerate(dataloader): + with torch.no_grad(): + logits = model( + input_ids=batch["input_ids"].to(device=DEVICE), + attention_mask=batch["attention_mask"].to(device=DEVICE), + decoder_input_ids=batch["decoder_input_ids"].to(device=DEVICE), + ).logits + labels = batch["labels"].to(device=DEVICE) + loss = loss_fn(logits.view(-1, logits.size(-1)), labels.view(-1)) + total_loss += loss.detach().float().item() + + labels = labels.cpu().numpy() + labels = np.where(labels != -100, labels, tokenizer.pad_token_id) + generated_tokens = model.generate( + batch["input_ids"].to(device=DEVICE), + attention_mask=batch["attention_mask"].to(device=DEVICE), + **gen_kwargs, + ) + if isinstance(generated_tokens, tuple): + generated_tokens = generated_tokens[0] + decoded_preds = tokenizer.batch_decode(generated_tokens, skip_special_tokens=True) + decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True) + decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels) + metric.add_batch( + predictions=decoded_preds, + references=decoded_labels, + ) + + result = metric.compute(use_stemmer=True) + result = {k: round(v * 100, 4) for k, v in result.items()} + result["loss"] = total_loss / len(dataloader) + return result + + +def main(): + args = parse_args() + logging.basicConfig(level=logging.INFO) + logger = logging.getLogger() + + if args.seed is not None: + set_seed(args.seed) + + tokenizer = get_tokenizer() + train_dataset = get_dailymail_dataset(split="train") + model = train( + dataset=train_dataset, + tokenizer=tokenizer, + batch_size=args.train_batch_size, + num_train_epochs=args.num_train_epochs, + learning_rate=args.learning_rate, + weight_decay=args.weight_decay, + ) + + eval_train_dataset = get_dailymail_dataset(split="eval_train") + results = evaluate_model( + model=model, tokenizer=tokenizer, dataset=eval_train_dataset, batch_size=args.eval_batch_size + ) + logger.info(f"Train evaluation results: {results}") + + eval_dataset = get_dailymail_dataset(split="valid") + results = evaluate_model(model=model, tokenizer=tokenizer, dataset=eval_dataset, batch_size=args.eval_batch_size) + logger.info(f"Valid evaluation results: {results}") + + if args.checkpoint_dir is not None: + torch.save(model.state_dict(), os.path.join(args.checkpoint_dir, "model.pth")) + + +if __name__ == "__main__": + main() diff --git a/examples/glue/README.md b/examples/glue/README.md index 12398a1..c1b43a4 100644 --- a/examples/glue/README.md +++ b/examples/glue/README.md @@ -1,14 +1,16 @@ # GLUE & BERT Example -This directory contains scripts for fine-tuning BERT on GLUE benchmark. The pipeline is motivated from [HuggingFace Example](https://github.com/huggingface/transformers/tree/main/examples/pytorch/text-classification). -Please begin by installing necessary packages. +This directory contains scripts for fine-tuning BERT and computing influence scores on the GLUE benchmark. The pipeline is motivated from [this HuggingFace Example](https://github.com/huggingface/transformers/tree/main/examples/pytorch/text-classification). +To get started, please install the necessary packages: + ```bash pip install -r requirements.txt ``` ## Training -To fine-tune BERT on some specific dataset, run the following command (we are using `SST2` dataset): +To fine-tune BERT on a specific dataset, run the following command (we are using the SST2 dataset in this example): + ```bash python train.py --dataset_name sst2 \ --checkpoint_dir ./checkpoints \ @@ -20,9 +22,12 @@ python train.py --dataset_name sst2 \ --seed 1004 ``` +This will save the trained checkpoint in the `./checkpoints` directory. + ## Computing Pairwise Influence Scores -To obtain a pairwise influence scores on maximum of 2000 query data points using `ekfac`, run the following command: +To obtain pairwise influence scores on a maximum of 2000 query data points using `ekfac`, run the following command: + ```bash python analyze.py --dataset_name sst2 \ --query_batch_size 175 \ @@ -30,31 +35,123 @@ python analyze.py --dataset_name sst2 \ --checkpoint_dir ./checkpoints \ --factor_strategy ekfac ``` -On A100 (80GB), it takes roughly 80 minutes to compute the pairwise scores for SST2 with around 900 query data points -(including computing EKFAC factors). -We can also use query batching (low-rank approximation to the query gradient; see Section 3.2.2 from the [paper](https://arxiv.org/pdf/2308.03296.pdf)) to compute influence scores with a -larger query batch size. +On an A100 (80GB), it takes roughly 90 minutes to compute the pairwise scores for SST2 (including computing EKFAC factors): + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 11 | 5088.0 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Fit Lambda | 2370.0 | 1 | 2370.0 | 46.581 | +| Compute Pairwise Score | 2222.4 | 1 | 2222.4 | 43.679 | +| Fit Covariance | 478.83 | 1 | 478.83 | 9.411 | +| Perform Eigendecomposition | 10.587 | 1 | 10.587 | 0.20808 | +| Save Eigendecomposition | 2.5419 | 1 | 2.5419 | 0.049958 | +| Save Covariance | 2.3878 | 1 | 2.3878 | 0.046931 | +| Save Lambda | 0.66905 | 1 | 0.66905 | 0.01315 | +| Save Pairwise Score | 0.51374 | 1 | 0.51374 | 0.010097 | +| Load All Factors | 0.01321 | 1 | 0.01321 | 0.00025963 | +| Load Covariance | 0.0081149 | 1 | 0.0081149 | 0.00015949 | +| Load Eigendecomposition | 0.0079874 | 1 | 0.0079874 | 0.00015699 | +---------------------------------------------------------------------------------------------------------------------------------- +``` + +For faster computation, use half precision: + ```bash python analyze.py --dataset_name sst2 \ - --query_gradient_rank 32 \ - --query_batch_size 436 \ - --train_batch_size 256 \ + --query_batch_size 175 \ + --train_batch_size 128 \ --checkpoint_dir ./checkpoints \ - --factor_strategy ekfac + --factor_strategy ekfac \ + --use_half_precision ``` -Note that query batching is slower in this case (140 minutes in total), as the number of training data points is small and the cost of performing SVD dominates the overall cost. -Assuming that you ran above two commands, `query_batching_analysis.py` contains code to compute the correlations between the full rank and low-rank scores. -

-Counterfactual -

-The averaged correlations between the low-rank and full rank scores for 100 data points is 0.98. +This reduces computation time to about 20 minutes on an A100 (80GB) GPU. + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 11 | 1222.4 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Compute Pairwise Score | 582.08 | 1 | 582.08 | 47.617 | +| Fit Lambda | 543.55 | 1 | 543.55 | 44.465 | +| Fit Covariance | 83.877 | 1 | 83.877 | 6.8616 | +| Perform Eigendecomposition | 9.4054 | 1 | 9.4054 | 0.76942 | +| Save Eigendecomposition | 1.516 | 1 | 1.516 | 0.12401 | +| Save Covariance | 1.434 | 1 | 1.434 | 0.11731 | +| Save Lambda | 0.28022 | 1 | 0.28022 | 0.022924 | +| Save Pairwise Score | 0.24123 | 1 | 0.24123 | 0.019734 | +| Load All Factors | 0.01241 | 1 | 0.01241 | 0.0010152 | +| Load Covariance | 0.0080553 | 1 | 0.0080553 | 0.00065897 | +| Load Eigendecomposition | 0.0077278 | 1 | 0.0077278 | 0.00063218 | +---------------------------------------------------------------------------------------------------------------------------------- +``` ## Counterfactual Evaluation -We plan to add a simple demo for counterfactual evaluation on the RTE dataset soon. +Let's evaluate the impact of removing top positively influential training examples on query misclassification. +First, compute pairwise influence scores for the `RTE` dataset (the below commands used a single A100 GPU): + +```bash +python train.py --dataset_name rte \ + --checkpoint_dir ./checkpoints \ + --train_batch_size 32 \ + --eval_batch_size 32 \ + --learning_rate 2e-05 \ + --weight_decay 0.01 \ + --num_train_epochs 3 \ + --seed 0 + +python analyze.py --dataset_name rte \ + --query_batch_size 70 \ + --train_batch_size 128 \ + --checkpoint_dir ./checkpoints \ + --factor_strategy ekfac + +python analyze.py --dataset_name rte \ + --query_batch_size 70 \ + --train_batch_size 128 \ + --checkpoint_dir ./checkpoints \ + --factor_strategy diagonal +``` + +Use `run_counterfactual.py` to run the counterfactual experiment. See [our paper](https://arxiv.org/abs/2405.12186) for the detailed setup.

Counterfactual -

\ No newline at end of file +

+ +## Evaluating Linear Datamodeling Score + +The `evaluate_lds.py` script computes the [linear datamodeling score (LDS)](https://arxiv.org/abs/2303.14186). It measures the LDS obtained by +retraining the network 500 times with different subsets of the dataset (5 repeats and 100 masks). + +
+ +| Strategy | LDS | +|--------------------------|:----:| +| `identity` | 0.10 | +| `diagonal` | 0.15 | +| `kfac` | 0.32 | +| `ekfac` | 0.32 | +| `ekfac` (half precision) | 0.32 | + +
+ +The script can also print top influential sequences for a given query. + +``` +Query Example: + Sentence1: The west has preferred to focus on endangered animals, rather than endangered humans. African elephants are hunted down and stripped of tusks and hidden by poachers. Their numbers in Africa slumped from 1.2m to 600,000 in a decade until CITES - the Convention on International Trade in Endangered Species - banned the trade in ivory. + Sentence2: African elephants are endangered by ivory poachers. + Label: 0 + +Top Influential Example: + Sentence1: The article also mentions the greater prevalence of obesity among two minority populations, African-Americans and Hispanic/Latino, but does not consider in its analysis of the increase in obesity the increase of these these populations as a proportion of the United States population. African-Americans and Hispanic/Latinos have a higher rates of obesity than White Americans, while Asian-Americans have a relatively low rate of obesity. Despite only representing one third of the U.S. population, African-Americans and Hispanic/Latinos represent about one half of the population growth. + Sentence2: African-Americans are a minority in the U.S. + Label: 0 +``` \ No newline at end of file diff --git a/examples/glue/analyze.py b/examples/glue/analyze.py index 022dfd9..90dd71f 100644 --- a/examples/glue/analyze.py +++ b/examples/glue/analyze.py @@ -1,7 +1,7 @@ import argparse import logging import os -from typing import Dict, Optional +from typing import Dict import torch import torch.nn.functional as F @@ -12,6 +12,8 @@ from kronfluence.analyzer import Analyzer, prepare_model from kronfluence.arguments import FactorArguments, ScoreArguments from kronfluence.task import Task +from kronfluence.utils.common.factor_arguments import all_low_precision_factor_arguments +from kronfluence.utils.common.score_arguments import all_low_precision_score_arguments from kronfluence.utils.dataset import DataLoaderKwargs BATCH_TYPE = Dict[str, torch.Tensor] @@ -33,12 +35,24 @@ def parse_args(): help="A path that is storing the final checkpoint of the model.", ) + parser.add_argument( + "--factor_strategy", + type=str, + default="ekfac", + help="Strategy to compute influence factors.", + ) parser.add_argument( "--query_gradient_rank", type=int, default=-1, help="Rank for the low-rank query gradient approximation.", ) + parser.add_argument( + "--use_half_precision", + action="store_true", + default=False, + help="Whether to use half precision for computing factors and scores.", + ) parser.add_argument( "--query_batch_size", type=int, @@ -52,12 +66,11 @@ def parse_args(): help="Batch size for computing training gradients.", ) parser.add_argument( - "--factor_strategy", - type=str, - default="ekfac", - help="Strategy to compute influence factors.", + "--profile", + action="store_true", + default=False, + help="Boolean flag to profile computations.", ) - args = parser.parse_args() if args.checkpoint_dir is not None: @@ -84,12 +97,12 @@ def compute_train_loss( if not sample: return F.cross_entropy(logits, batch["labels"], reduction="sum") with torch.no_grad(): - probs = torch.nn.functional.softmax(logits, dim=-1) + probs = torch.nn.functional.softmax(logits.detach(), dim=-1) sampled_labels = torch.multinomial( probs, num_samples=1, ).flatten() - return F.cross_entropy(logits, sampled_labels.detach(), reduction="sum") + return F.cross_entropy(logits, sampled_labels, reduction="sum") def compute_measurement( self, @@ -113,7 +126,7 @@ def compute_measurement( margins = logits_correct - cloned_logits.logsumexp(dim=-1) return -margins.sum() - def get_attention_mask(self, batch: BATCH_TYPE) -> Optional[torch.Tensor]: + def get_attention_mask(self, batch: BATCH_TYPE) -> torch.Tensor: return batch["attention_mask"] @@ -146,38 +159,48 @@ def main(): analysis_name=args.dataset_name, model=model, task=task, - cpu=False, + profile=args.profile, ) # Configure parameters for DataLoader. dataloader_kwargs = DataLoaderKwargs(collate_fn=default_data_collator) analyzer.set_dataloader_kwargs(dataloader_kwargs) # Compute influence factors. + factors_name = args.factor_strategy factor_args = FactorArguments(strategy=args.factor_strategy) + if args.use_half_precision: + factor_args = all_low_precision_factor_arguments(strategy=args.factor_strategy, dtype=torch.bfloat16) + factors_name += "_half" analyzer.fit_all_factors( - factors_name=args.factor_strategy, + factors_name=factors_name, dataset=train_dataset, per_device_batch_size=None, factor_args=factor_args, - overwrite_output_dir=True, + overwrite_output_dir=False, initial_per_device_batch_size_attempt=512, ) + # Compute pairwise scores. + score_args = ScoreArguments() + scores_name = factor_args.strategy + if args.use_half_precision: + score_args = all_low_precision_score_arguments(dtype=torch.bfloat16) + scores_name += "_half" rank = args.query_gradient_rank if args.query_gradient_rank != -1 else None - score_args = ScoreArguments(query_gradient_rank=rank, query_gradient_svd_dtype=torch.float32) - scores_name = args.factor_strategy if rank is not None: + score_args.query_gradient_low_rank = rank + score_args.query_gradient_accumulation_steps = 10 scores_name += f"_qlr{rank}" analyzer.compute_pairwise_scores( score_args=score_args, scores_name=scores_name, - factors_name=args.factor_strategy, + factors_name=factors_name, query_dataset=eval_dataset, query_indices=list(range(min([len(eval_dataset), 2000]))), train_dataset=train_dataset, per_device_query_batch_size=args.query_batch_size, per_device_train_batch_size=args.train_batch_size, - overwrite_output_dir=True, + overwrite_output_dir=False, ) scores = analyzer.load_pairwise_scores(scores_name)["all_modules"] logging.info(f"Scores shape: {scores.shape}") diff --git a/examples/glue/evaluate_lds.py b/examples/glue/evaluate_lds.py new file mode 100644 index 0000000..513feda --- /dev/null +++ b/examples/glue/evaluate_lds.py @@ -0,0 +1,64 @@ +import logging + +import numpy as np +import torch +import tqdm +from scipy.stats import spearmanr + +from examples.glue.pipeline import get_glue_dataset +from kronfluence.analyzer import Analyzer + + +def evaluate_correlations(scores: torch.Tensor) -> float: + margins = torch.from_numpy(torch.load(open("files/margins.pt", "rb"))) + masks = torch.from_numpy(torch.load(open("files/masks.pt", "rb"))).float() + + val_indices = np.arange(277) + preds = masks @ scores.T + + rs = [] + ps = [] + for j in tqdm.tqdm(val_indices): + r, p = spearmanr(preds[:, j], margins[:, j]) + rs.append(r) + ps.append(p) + rs, ps = np.array(rs), np.array(ps) + return rs.mean() + + +def main(): + logging.basicConfig(level=logging.INFO) + + # You might need to change the path. + strategy = "ekfac" + scores = Analyzer.load_file(f"influence_results/rte/scores_{strategy}/pairwise_scores.safetensors")[ + "all_modules" + ].to(dtype=torch.float32) + + corr_mean = evaluate_correlations(scores) + logging.info(f"LDS: {np.mean(corr_mean)}") + + # We can also visualize the top influential sequences. + eval_idx = 79 + train_dataset = get_glue_dataset( + data_name="rte", + split="eval_train", + ) + eval_dataset = get_glue_dataset( + data_name="rte", + split="valid", + ) + print("Query Data Example:") + print(f"Sentence1: {eval_dataset[eval_idx]['sentence1']}") + print(f"Sentence2: {eval_dataset[eval_idx]['sentence2']}") + print(f"Label: {eval_dataset[eval_idx]['label']}") + + top_idx = int(torch.argsort(scores[eval_idx], descending=True)[0]) + print("Top Influential Example:") + print(f"Sentence1: {train_dataset[top_idx]['sentence1']}") + print(f"Sentence2: {train_dataset[top_idx]['sentence2']}") + print(f"Label: {train_dataset[top_idx]['label']}") + + +if __name__ == "__main__": + main() diff --git a/examples/glue/files/margins.pt b/examples/glue/files/margins.pt new file mode 100644 index 0000000..ba28a40 Binary files /dev/null and b/examples/glue/files/margins.pt differ diff --git a/examples/glue/files/masks.pt b/examples/glue/files/masks.pt new file mode 100644 index 0000000..61308a0 Binary files /dev/null and b/examples/glue/files/masks.pt differ diff --git a/examples/glue/half_precision_analysis.py b/examples/glue/half_precision_analysis.py new file mode 100644 index 0000000..288c39c --- /dev/null +++ b/examples/glue/half_precision_analysis.py @@ -0,0 +1,40 @@ +import logging + +import matplotlib.pyplot as plt +import numpy as np +from scipy.stats import spearmanr +from tueplots import markers + +from kronfluence.analyzer import Analyzer + + +def main(): + logging.basicConfig(level=logging.INFO) + + # Load the scores. You might need to modify the path. + scores = Analyzer.load_file("influence_results/sst2/scores_ekfac/pairwise_scores.safetensors")["all_modules"] + half_scores = Analyzer.load_file("influence_results/sst2/scores_ekfac_half/pairwise_scores.safetensors")[ + "all_modules" + ].float() + + plt.rcParams.update({"figure.dpi": 150}) + plt.rcParams.update(markers.with_edge()) + plt.rcParams["axes.axisbelow"] = True + + # Only plot first 6000 points to avoid clutter. + idx = 79 + plt.scatter(half_scores[idx][:6000], scores[idx][:6000], edgecolor="k") + plt.grid() + plt.xlabel("bfloat16") + plt.ylabel("float32") + plt.show() + + # Compute the averaged spearman correlation. + all_corr = [] + for i in range(500): + all_corr.append(spearmanr(scores[i], half_scores[i])[0]) + logging.info(f"Averaged Spearman Correlation: {np.array(all_corr).mean()}") + + +if __name__ == "__main__": + main() diff --git a/examples/glue/pipeline.py b/examples/glue/pipeline.py index 3db7e55..e8fa284 100644 --- a/examples/glue/pipeline.py +++ b/examples/glue/pipeline.py @@ -68,12 +68,14 @@ def preprocess_function(examples): raw_datasets = raw_datasets.map( preprocess_function, batched=True, - load_from_cache_file=(not False), + load_from_cache_file=True, ) if split in ["train", "eval_train"]: train_dataset = raw_datasets["train"] ds = train_dataset + if data_name == "rte": + ds = ds.select(range(2432)) else: eval_dataset = raw_datasets["validation"] ds = eval_dataset diff --git a/examples/glue/run_counterfactual.py b/examples/glue/run_counterfactual.py new file mode 100644 index 0000000..46bd564 --- /dev/null +++ b/examples/glue/run_counterfactual.py @@ -0,0 +1,221 @@ +import time +from typing import List, Optional + +import numpy as np +import torch +import torch.nn as nn +from accelerate.utils import set_seed +from torch.utils import data +from transformers import default_data_collator + +from examples.glue.pipeline import get_glue_dataset +from examples.glue.train import train +from kronfluence import Analyzer + +DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu") + + +def get_accuracy(model: nn.Module, dataset: data.Dataset) -> torch.Tensor: + dataloader = data.DataLoader( + dataset=dataset, batch_size=32, shuffle=False, drop_last=False, collate_fn=default_data_collator + ) + + model.eval() + with torch.no_grad(): + acc_lst = [] + for batch in dataloader: + outputs = model( + input_ids=batch["input_ids"].to(device=DEVICE), + token_type_ids=batch["token_type_ids"].to(device=DEVICE), + attention_mask=batch["attention_mask"].to(device=DEVICE), + ).logits + labels = batch["labels"].to(device=DEVICE) + accs = (outputs.argmax(-1) == labels).float().cpu() + acc_lst.append(accs) + all_accs = torch.cat(acc_lst) + return all_accs + + +def train_with_indices(dataset: data.Dataset, seed: int, indices_to_keep: Optional[List[int]] = None) -> nn.Module: + if indices_to_keep is not None: + dataset = dataset.select(indices_to_keep) + + set_seed(seed) + model = train(dataset=dataset, batch_size=16, num_train_epochs=3, learning_rate=2e-05, weight_decay=0.01) + return model + + +def train_with_configurations( + dataset: data.Dataset, + valid_dataset: data.Dataset, + top_indices: List[int], + interval: int, + seed_ids: List[int], +) -> List[torch.Tensor]: + num_train = len(dataset) + indices_to_remove = top_indices[:interval] + indices_to_keep = list(set(range(num_train)) - set(indices_to_remove)) + assert len(indices_to_keep) + len(indices_to_remove) == num_train + + valid_acc_lst = [] + for seed in seed_ids: + model = train_with_indices(dataset=dataset, indices_to_keep=indices_to_keep, seed=seed + 2008) + valid_results = get_accuracy(model, valid_dataset) + valid_acc_lst.append(valid_results) + return valid_acc_lst + + +def main(): + train_dataset = get_glue_dataset( + data_name="rte", + split="eval_train", + ) + eval_dataset = get_glue_dataset( + data_name="rte", + split="valid", + ) + num_target = 100 + assert num_target <= len(eval_dataset) + + remove_intervals = [20, 40, 60, 80, 100, 120] + num_base_repeat = 5 + num_repeat = 3 + + large_seed_ids = list(range(num_base_repeat)) + seed_ids = list(range(num_repeat)) + + valid_acc_lst = [] + for seed in large_seed_ids: + model = train_with_indices(dataset=train_dataset, seed=seed + 79, indices_to_keep=None) + valid_results = get_accuracy(model, eval_dataset) + valid_acc_lst.append(valid_results) + + # Selects validation data points that get correctly classified on all seeds. + mask = np.array(valid_acc_lst).mean(0) >= 1.0 + print(f"Total target numbers: {mask.sum()}") + + # Get random baseline. + start_time = time.time() + random_results = [] + for valid_idx in range(num_target): + print(f"{valid_idx}th validation data point.") + if mask[valid_idx]: + # Selects training data points with the same label. + correct_label = eval_dataset[valid_idx]["label"] + random_indices = list( + np.random.permutation( + [ + i + for i, x in enumerate([x["label"] for x in eval_dataset]) + if x == correct_label and i < num_target + ] + ) + ) + + success_lst = [] + for interval in remove_intervals: + results = train_with_configurations( + dataset=train_dataset, + top_indices=random_indices, + valid_dataset=eval_dataset.select([valid_idx]), + interval=interval, + seed_ids=seed_ids, + ) + if np.array(results).mean() < 0.5: + success_lst.append(1) + break + else: + success_lst.append(0) + + while len(success_lst) < len(remove_intervals): + success_lst.append(1) + + random_results.append(success_lst) + + end_time = time.time() + print(f"Took {end_time - start_time} seconds for the random baseline.") + random_results = np.array(random_results).sum(0) + print(f"Results: {random_results}") + + # Get EKFAC baseline. + start_time = time.time() + scores = Analyzer.load_file("influence_results/rte/scores_ekfac/pairwise_scores.safetensors")["all_modules"].to( + dtype=torch.float32 + ) + ekfac_results = [] + for valid_idx in range(num_target): + print(f"{valid_idx}th validation data point.") + if mask[valid_idx]: + top_indices = torch.argsort(scores[valid_idx], descending=True) + top_indices = [idx.item() for idx in top_indices] + + success_lst = [] + for interval in remove_intervals: + results = train_with_configurations( + dataset=train_dataset, + top_indices=top_indices, + valid_dataset=eval_dataset.select([valid_idx]), + interval=interval, + seed_ids=seed_ids, + ) + if np.array(results).mean() < 0.5: + success_lst.append(1) + break + else: + success_lst.append(0) + + while len(success_lst) < len(remove_intervals): + success_lst.append(1) + + ekfac_results.append(success_lst) + + end_time = time.time() + print(f"Took {end_time - start_time} seconds for the EKFAC baseline.") + ekfac_results = np.array(ekfac_results).sum(0) + print(f"Results: {ekfac_results}") + + # Get Identity baseline. + start_time = time.time() + scores = Analyzer.load_file("influence_results/rte/scores_identity/pairwise_scores.safetensors")["all_modules"].to( + dtype=torch.float32 + ) + identity_results = [] + for valid_idx in range(num_target): + print(f"{valid_idx}th validation data point.") + if mask[valid_idx]: + top_indices = torch.argsort(scores[valid_idx], descending=True) + top_indices = [idx.item() for idx in top_indices] + + success_lst = [] + for interval in remove_intervals: + results = train_with_configurations( + dataset=train_dataset, + top_indices=top_indices, + valid_dataset=eval_dataset.select([valid_idx]), + interval=interval, + seed_ids=seed_ids, + ) + if np.array(results).mean() < 0.5: + success_lst.append(1) + break + else: + success_lst.append(0) + + while len(success_lst) < len(remove_intervals): + success_lst.append(1) + + identity_results.append(success_lst) + + end_time = time.time() + print(f"Took {end_time - start_time} seconds for the identity baseline.") + identity_results = np.array(identity_results).sum(0) + print(f"Results: {identity_results}") + + print("final") + print(f"Results: {random_results}") + print(f"Results: {ekfac_results}") + print(f"Results: {identity_results}") + + +if __name__ == "__main__": + main() diff --git a/examples/glue/train.py b/examples/glue/train.py index b6dacf6..f2f6950 100644 --- a/examples/glue/train.py +++ b/examples/glue/train.py @@ -71,7 +71,6 @@ def parse_args(): default="./checkpoints", help="A path to store the final checkpoint.", ) - args = parser.parse_args() if args.checkpoint_dir is not None: @@ -105,13 +104,13 @@ def train( for epoch in range(num_train_epochs): total_loss = 0.0 for batch in train_dataloader: + optimizer.zero_grad(set_to_none=True) loss = model( input_ids=batch["input_ids"].to(device=DEVICE), attention_mask=batch["attention_mask"].to(device=DEVICE), token_type_ids=batch["token_type_ids"].to(device=DEVICE), labels=batch["labels"].to(device=DEVICE), ).loss - optimizer.zero_grad() loss.backward() optimizer.step() total_loss += loss.detach().float() diff --git a/examples/imagenet/README.md b/examples/imagenet/README.md index 399f60d..6595bb5 100644 --- a/examples/imagenet/README.md +++ b/examples/imagenet/README.md @@ -1,54 +1,145 @@ # ImageNet & ResNet-50 Example -This directory contains scripts for training ResNet-50 on ImageNet. Please begin by installing necessary packages. +This directory contains scripts for computing influence scores on the ImageNet dataset using ResNet-50. +To get started, please install the necessary packages by running: + ```bash pip install -r requirements.txt ``` -## Training - -We will use the pre-trained model from `torchvision.models.resnet50`. +We will use the pre-trained ResNet-50 model from `torchvision.models.resnet50`. ## Computing Pairwise Influence Scores -To obtain a pairwise influence scores on 1000 query data points using `ekfac`, run the following command: +To compute pairwise influence scores on 1000 query data points using the `ekfac` factorization strategy, run the following command: + ```bash python analyze.py --dataset_dir PATH_TO_IMAGENET \ --query_gradient_rank -1 \ --query_batch_size 100 \ - --train_batch_size 512 \ + --train_batch_size 256 \ --factor_strategy ekfac ``` -On A100 (80GB), it takes approximately 12 hours to compute the pairwise scores (including computing EKFAC factors). -We can also use query batching (low-rank approximation to the query gradient; see Section 3.2.2 from the [paper](https://arxiv.org/pdf/2308.03296.pdf)) to compute influence scores with a -larger query batch size. +Replace `PATH_TO_IMAGENET` with the path to your ImageNet dataset directory. On an A100 (80GB) GPU, it takes approximately 11 hours to compute the pairwise scores (including computing EKFAC factors): + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 11 | 4.072e+04 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Compute Pairwise Score | 3.9136e+04 | 1 | 3.9136e+04 | 96.111 | +| Fit Lambda | 1056.7 | 1 | 1056.7 | 2.5952 | +| Fit Covariance | 454.05 | 1 | 454.05 | 1.1151 | +| Save Pairwise Score | 49.141 | 1 | 49.141 | 0.12068 | +| Perform Eigendecomposition | 10.589 | 1 | 10.589 | 0.026006 | +| Save Covariance | 5.7108 | 1 | 5.7108 | 0.014025 | +| Save Eigendecomposition | 5.5442 | 1 | 5.5442 | 0.013616 | +| Save Lambda | 0.97504 | 1 | 0.97504 | 0.0023945 | +| Load Eigendecomposition | 0.52786 | 1 | 0.52786 | 0.0012963 | +| Load Covariance | 0.30319 | 1 | 0.30319 | 0.00074458 | +| Load All Factors | 0.15671 | 1 | 0.15671 | 0.00038485 | +---------------------------------------------------------------------------------------------------------------------------------- +``` + +Query batching (low-rank approximation to the query gradient; see **Section 3.2.2** from the paper) can be used to compute influence scores with a larger query batch size: + ```bash python analyze.py --dataset_dir PATH_TO_IMAGENET \ --query_gradient_rank 32 \ - --query_batch_size 500 \ - --train_batch_size 512 \ + --query_batch_size 256 \ + --train_batch_size 256 \ --factor_strategy ekfac ``` -On A100 (80GB), it takes roughly 4 hours to compute the pairwise scores with query batching (including computing EKFAC factors). -Assuming that you ran above two commands, `query_batching_analysis.py` -contains code to compute the correlations between the full rank and low-rank scores. + +On an A100 (80GB) GPU, it takes roughly 3 hours to compute the pairwise scores with query batching: + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 3 | 7352.8 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Compute Pairwise Score | 7340.8 | 1 | 7340.8 | 99.836 | +| Save Pairwise Score | 12.026 | 1 | 12.026 | 0.16355 | +| Load All Factors | 0.0099941 | 1 | 0.0099941 | 0.00013592 | +---------------------------------------------------------------------------------------------------------------------------------- +``` + +Assuming you ran the above two commands, `query_batching_analysis.py` contains code to compute the correlations between the full-rank and low-rank scores.

-Counterfactual +Query Batching

-The averaged correlations between the low-rank and full rank scores for 100 data points is 0.95. + +The averaged correlations between the low-rank and full rank scores for 100 data points is 0.94. +For even more efficient computation, use half precision: + +```bash +python analyze.py --dataset_dir PATH_TO_IMAGENET \ + --query_gradient_rank 32 \ + --query_batch_size 500 \ + --train_batch_size 512 \ + --factor_strategy ekfac \ + --use_half_precision +``` + +This reduces computation time to about 85 minutes on an A100 (80GB) GPU: + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 11 | 3023.7 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Compute Pairwise Score | 2621.2 | 1 | 2621.2 | 86.688 | +| Fit Lambda | 232.3 | 1 | 232.3 | 7.6825 | +| Fit Covariance | 157.36 | 1 | 157.36 | 5.204 | +| Perform Eigendecomposition | 5.745 | 1 | 5.745 | 0.19 | +| Save Pairwise Score | 5.6676 | 1 | 5.6676 | 0.18744 | +| Save Covariance | 0.70454 | 1 | 0.70454 | 0.0233 | +| Save Eigendecomposition | 0.61539 | 1 | 0.61539 | 0.020352 | +| Save Lambda | 0.092784 | 1 | 0.092784 | 0.0030685 | +| Load Covariance | 0.013714 | 1 | 0.013714 | 0.00045354 | +| Load All Factors | 0.0088742 | 1 | 0.0088742 | 0.00029348 | +| Load Eigendecomposition | 0.0056237 | 1 | 0.0056237 | 0.00018599 | +---------------------------------------------------------------------------------------------------------------------------------- +``` ## Computing Pairwise Influence Scores with DDP -You can also use DistributedDataParallel (DDP) to speed up influence computations. You can run: +You can also use [DistributedDataParallel (DDP)](https://pytorch.org/tutorials/intermediate/ddp_tutorial.html) to speed up influence computations. To run influence analysis with four A100 (80GB) GPUs and query batching, use the command: + ```bash -torchrun --standalone --nnodes=1 --nproc-per-node=2 ddp_analyze.py --dataset_dir PATH_TO_IMAGENET \ - --query_gradient_rank -1 \ +torchrun --standalone --nnodes=1 --nproc-per-node=4 ddp_analyze.py --dataset_dir PATH_TO_IMAGENET \ + --query_gradient_rank 32 \ --factor_batch_size 512 \ --query_batch_size 100 \ - --train_batch_size 512 \ + --train_batch_size 256 \ --factor_strategy ekfac ``` -On 2 A100 (80GB), it takes approximately 6 hours to compute the pairwise scores. When available, you can use more GPUs -to speed up influence computations. + +It takes approximately 1 hour to compute the pairwise scores: + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 11 | 3423.3 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Compute Pairwise Score | 2917.2 | 1 | 2917.2 | 85.215 | +| Fit Lambda | 298.62 | 1 | 298.62 | 8.7232 | +| Fit Covariance | 137.9 | 1 | 137.9 | 4.0282 | +| Save Pairwise Score | 48.122 | 1 | 48.122 | 1.4057 | +| Perform Eigendecomposition | 7.7503 | 1 | 7.7503 | 0.2264 | +| Save Eigendecomposition | 5.9978 | 1 | 5.9978 | 0.1752 | +| Save Covariance | 5.7442 | 1 | 5.7442 | 0.1678 | +| Save Lambda | 0.95602 | 1 | 0.95602 | 0.027927 | +| Load Covariance | 0.5718 | 1 | 0.5718 | 0.016703 | +| Load Eigendecomposition | 0.34755 | 1 | 0.34755 | 0.010153 | +| Load All Factors | 0.13107 | 1 | 0.13107 | 0.0038288 | +---------------------------------------------------------------------------------------------------------------------------------- +``` + +You can use more GPUs to further speed up the influence computations (or use AMP + half precision). diff --git a/examples/imagenet/analyze.py b/examples/imagenet/analyze.py index d831a9d..9dd14b1 100644 --- a/examples/imagenet/analyze.py +++ b/examples/imagenet/analyze.py @@ -1,19 +1,16 @@ import argparse import logging -from typing import Tuple import torch -import torch.nn.functional as F -from torch import nn +from examples.cifar.analyze import ClassificationTask from examples.imagenet.pipeline import construct_resnet50, get_imagenet_dataset from kronfluence.analyzer import Analyzer, prepare_model from kronfluence.arguments import FactorArguments, ScoreArguments -from kronfluence.task import Task +from kronfluence.utils.common.factor_arguments import all_low_precision_factor_arguments +from kronfluence.utils.common.score_arguments import all_low_precision_score_arguments from kronfluence.utils.dataset import DataLoaderKwargs -BATCH_TYPE = Tuple[torch.Tensor, torch.Tensor] - def parse_args(): parser = argparse.ArgumentParser(description="Influence analysis on ImageNet dataset.") @@ -21,10 +18,16 @@ def parse_args(): parser.add_argument( "--dataset_dir", type=str, - default="/mfs1/datasets/imagenet_pytorch/", + default="PATH_TO_IMAGENET", help="A folder containing the ImageNet dataset.", ) + parser.add_argument( + "--factor_strategy", + type=str, + default="ekfac", + help="Strategy to compute preconditioning factors.", + ) parser.add_argument( "--query_gradient_rank", type=int, @@ -40,58 +43,25 @@ def parse_args(): parser.add_argument( "--train_batch_size", type=int, - default=128, + default=256, help="Batch size for computing training gradient.", ) parser.add_argument( - "--factor_strategy", - type=str, - default="ekfac", - help="Strategy to compute preconditioning factors.", + "--use_half_precision", + action="store_true", + default=False, + help="Whether to use half precision for computing factors and scores.", + ) + parser.add_argument( + "--profile", + action="store_true", + default=False, + help="Boolean flag to profile computations.", ) - args = parser.parse_args() return args -class ClassificationTask(Task): - def compute_train_loss( - self, - batch: BATCH_TYPE, - model: nn.Module, - sample: bool = False, - ) -> torch.Tensor: - inputs, labels = batch - logits = model(inputs) - if not sample: - return F.cross_entropy(logits, labels, reduction="sum") - with torch.no_grad(): - probs = torch.nn.functional.softmax(logits, dim=-1) - sampled_labels = torch.multinomial( - probs, - num_samples=1, - ).flatten() - return F.cross_entropy(logits, sampled_labels.detach(), reduction="sum") - - def compute_measurement( - self, - batch: BATCH_TYPE, - model: nn.Module, - ) -> torch.Tensor: - # Copied from: https://github.com/MadryLab/trak/blob/main/trak/modelout_functions.py. - inputs, labels = batch - logits = model(inputs) - - bindex = torch.arange(logits.shape[0]).to(device=logits.device, non_blocking=False) - logits_correct = logits[bindex, labels] - - cloned_logits = logits.clone() - cloned_logits[bindex, labels] = torch.tensor(-torch.inf, device=logits.device, dtype=logits.dtype) - - margins = logits_correct - cloned_logits.logsumexp(dim=-1) - return -margins.sum() - - def main(): args = parse_args() logging.basicConfig(level=logging.INFO) @@ -111,6 +81,7 @@ def main(): analysis_name="imagenet", model=model, task=task, + profile=args.profile, ) # Configure parameters for DataLoader. dataloader_kwargs = DataLoaderKwargs( @@ -119,31 +90,41 @@ def main(): analyzer.set_dataloader_kwargs(dataloader_kwargs) # Compute influence factors. + factors_name = args.factor_strategy factor_args = FactorArguments(strategy=args.factor_strategy) + if args.use_half_precision: + factor_args = all_low_precision_factor_arguments(strategy=args.factor_strategy, dtype=torch.bfloat16) + factors_name += "_half" analyzer.fit_all_factors( - factors_name=args.factor_strategy, + factors_name=factors_name, dataset=train_dataset, per_device_batch_size=None, factor_args=factor_args, overwrite_output_dir=False, + initial_per_device_batch_size_attempt=512, ) # Compute pairwise scores. + score_args = ScoreArguments() + scores_name = factor_args.strategy + if args.use_half_precision: + score_args = all_low_precision_score_arguments(dtype=torch.bfloat16) + scores_name += "_half" rank = args.query_gradient_rank if args.query_gradient_rank != -1 else None - score_args = ScoreArguments(query_gradient_rank=rank, query_gradient_svd_dtype=torch.float32) - scores_name = args.factor_strategy if rank is not None: + score_args.query_gradient_low_rank = rank + score_args.query_gradient_accumulation_steps = 10 scores_name += f"_qlr{rank}" analyzer.compute_pairwise_scores( - score_args=score_args, scores_name=scores_name, - factors_name=args.factor_strategy, + score_args=score_args, + factors_name=factors_name, query_dataset=eval_dataset, query_indices=list(range(1000)), train_dataset=train_dataset, per_device_query_batch_size=args.query_batch_size, per_device_train_batch_size=args.train_batch_size, - overwrite_output_dir=True, + overwrite_output_dir=False, ) scores = analyzer.load_pairwise_scores(scores_name)["all_modules"] logging.info(f"Scores shape: {scores.shape}") diff --git a/examples/imagenet/ddp_analyze.py b/examples/imagenet/ddp_analyze.py index 1b574d5..33f9653 100644 --- a/examples/imagenet/ddp_analyze.py +++ b/examples/imagenet/ddp_analyze.py @@ -1,7 +1,6 @@ import argparse import logging import os -from typing import Tuple import torch @@ -9,11 +8,12 @@ from examples.imagenet.pipeline import construct_resnet50, get_imagenet_dataset from kronfluence.analyzer import Analyzer, prepare_model from kronfluence.arguments import FactorArguments, ScoreArguments +from kronfluence.utils.common.factor_arguments import all_low_precision_factor_arguments +from kronfluence.utils.common.score_arguments import all_low_precision_score_arguments from kronfluence.utils.dataset import DataLoaderKwargs from kronfluence.utils.model import apply_ddp torch.backends.cudnn.benchmark = True -BATCH_DTYPE = Tuple[torch.Tensor, torch.Tensor] LOCAL_RANK = int(os.environ["LOCAL_RANK"]) WORLD_RANK = int(os.environ["RANK"]) WORLD_SIZE = int(os.environ["WORLD_SIZE"]) @@ -25,10 +25,16 @@ def parse_args(): parser.add_argument( "--dataset_dir", type=str, - default="/mfs1/datasets/imagenet_pytorch/", + default="PATH_TO_IMAGENET", help="A folder containing the ImageNet dataset.", ) + parser.add_argument( + "--factor_strategy", + type=str, + default="ekfac", + help="Strategy to compute preconditioning factors.", + ) parser.add_argument( "--query_gradient_rank", type=int, @@ -54,10 +60,10 @@ def parse_args(): help="Batch size for computing training gradient.", ) parser.add_argument( - "--factor_strategy", - type=str, - default="ekfac", - help="Strategy to compute preconditioning factors.", + "--use_half_precision", + action="store_true", + default=False, + help="Whether to use half precision for computing factors and scores.", ) parser.add_argument( "--profile", @@ -65,7 +71,6 @@ def parse_args(): default=False, help="Boolean flag to profile computations.", ) - args = parser.parse_args() return args @@ -106,27 +111,34 @@ def main(): analyzer.set_dataloader_kwargs(dataloader_kwargs) # Compute influence factors. - factor_args = FactorArguments( - strategy=args.factor_strategy, - ) + factors_name = args.factor_strategy + factor_args = FactorArguments(strategy=args.factor_strategy) + if args.use_half_precision: + factor_args = all_low_precision_factor_arguments(strategy=args.factor_strategy, dtype=torch.bfloat16) + factors_name += "_half" analyzer.fit_all_factors( factors_name=args.factor_strategy, dataset=train_dataset, - per_device_batch_size=args.factor_batch_size, + per_device_batch_size=None, factor_args=factor_args, overwrite_output_dir=False, ) # Compute pairwise scores. + score_args = ScoreArguments() + scores_name = factor_args.strategy + if args.use_half_precision: + score_args = all_low_precision_score_arguments(dtype=torch.bfloat16) + scores_name += "_half" rank = args.query_gradient_rank if args.query_gradient_rank != -1 else None - score_args = ScoreArguments(query_gradient_rank=rank) - scores_name = args.factor_strategy if rank is not None: + score_args.query_gradient_low_rank = rank + score_args.query_gradient_accumulation_steps = 10 scores_name += f"_qlr{rank}" analyzer.compute_pairwise_scores( - score_args=score_args, scores_name=scores_name, - factors_name=args.factor_strategy, + score_args=score_args, + factors_name=factors_name, query_dataset=eval_dataset, query_indices=list(range(1000)), train_dataset=train_dataset, diff --git a/examples/imagenet/query_batching_analysis.py b/examples/imagenet/query_batching_analysis.py index c0b5480..6e82e98 100644 --- a/examples/imagenet/query_batching_analysis.py +++ b/examples/imagenet/query_batching_analysis.py @@ -2,7 +2,7 @@ import matplotlib.pyplot as plt import numpy as np -from scipy.stats import spearmanr +from scipy.stats import pearsonr, spearmanr from tueplots import markers from kronfluence.analyzer import Analyzer @@ -12,14 +12,21 @@ def main(): logging.basicConfig(level=logging.INFO) # Load the scores. You might need to modify the path. - full_scores = Analyzer.load_file("scores_ekfac/pairwise_scores.safetensors")["all_modules"] - lr_scores = Analyzer.load_file("scores_ekfac_qlr32/pairwise_scores.safetensors")["all_modules"] - - # Only plot first 1000 points to avoid clutter. + full_scores = Analyzer.load_file("influence_results/imagenet/scores_ekfac/pairwise_scores.safetensors")[ + "all_modules" + ] + lr_scores = Analyzer.load_file("influence_results/imagenet/scores_ekfac_qlr32/pairwise_scores.safetensors")[ + "all_modules" + ] + # lr_scores = Analyzer.load_file("influence_results/imagenet/scores_ekfac_half_qlr32/pairwise_scores.safetensors")[ + # "all_modules" + # ].float() + + # Only plot first 5000 points to avoid clutter. plt.rcParams.update({"figure.dpi": 150}) plt.rcParams.update(markers.with_edge()) plt.rcParams["axes.axisbelow"] = True - plt.scatter(lr_scores[0][:1000], full_scores[0][:1000], edgecolor="k") + plt.scatter(lr_scores[0][:5000], full_scores[0][:5000], edgecolor="k") plt.grid() plt.xlabel("Full Rank Score") plt.ylabel("Low Rank (32) Score") @@ -31,6 +38,12 @@ def main(): all_corr.append(spearmanr(full_scores[i], lr_scores[i])[0]) logging.info(f"Averaged Spearman Correlation: {np.array(all_corr).mean()}") + # Compute the averaged pearson correlation. + all_corr = [] + for i in range(100): + all_corr.append(pearsonr(full_scores[i], lr_scores[i])[0]) + logging.info(f"Averaged Spearman Correlation: {np.array(all_corr).mean()}") + if __name__ == "__main__": main() diff --git a/examples/openwebtext/README.md b/examples/openwebtext/README.md new file mode 100644 index 0000000..cf99680 --- /dev/null +++ b/examples/openwebtext/README.md @@ -0,0 +1,218 @@ +# OpenWebText & Llama-3-8B Example + +This directory contains scripts for computing influence scores on the subset of OpenWebText dataset. +The pipeline is inspired by the [LoggIX repository](https://github.com/logix-project/logix/tree/main/examples/language_modeling) and serves as an example of how to use Kronfluence on large-scale models. +Install the necessary packages: + +```bash +pip install -r requirements.txt +``` + +We will use the pre-trained Meta-Llama-3-8B model [from HuggingFace](https://huggingface.co/meta-llama/Meta-Llama-3-8B). + +## Fitting EKFAC Factors + +To compute factors using the `ekfac` strategy, run the following command (e.g., 4 A100 80GB GPUs): + +```bash +torchrun --standalone --nnodes=1 --nproc-per-node=4 fit_factors.py \ + --factors_name jul_11_2024 \ + --factor_batch_size 4 +``` + +You can visualize the fitted factors using the `inspect_factors.py` script. + +

+Eigenvalues +

+ +## Computing Influence Scores + +The `generate.py` script contains code to generate responses from the Llama-3-8B model given certain prompts. +Some prompt-completion pairs are saved in `data/data.json`. +To compute influence scores on the generated prompt-completion pairs using the fitted factors, run: + +```bash +torchrun --standalone --nnodes=1 --nproc-per-node=4 compute_scores.py \ + --factors_name jul_11_2024 \ + --scores_name raw_scores \ + --train_batch_size 8 \ + --query_gradient_rank 64 +``` + +You can experiment with various configurations. For example, after identifying the top 100 influential sequences using query batching, you can use the full query gradient (or without half precision) to re-compute influence scores on these queries. This approach can yield a more accurate ranking. (We followed this approach for the EKFAC IF paper, and it can be effective in removing outliers.) + +The `/files` directory contains some results without any refinement process. Examples are available in this folder. + +``` +Query Sequence: +Prompt: Inflation is typically measured by; Completion: the Consumer Price Index (CPI). + +Top Influential Sequences: +================================================================================ +Rank = 0; Score = 3899392.0 +<|begin_of_text|>WASHINGTON (Reuters) - Both President Barack Obama and Republicans in the U.S. House of Representatives have made opening offers in negotiations to resolve the “fiscal cliff. + +Here is a look at the two proposals, which are aimed at averting a more drastic combination of tax increases and spending cuts that economists say could cause a recession: + +REPUBLICAN OFFER + +House Republican leaders on Monday called for $2.2 trillion in new deficit reductions over 10 years. + +When counting deficit reductions enacted last year, anticipated savings from winding down the wars in Iraq and Afghanistan and some interest savings, the package would amount to $4.6 trillion in reductions over a decade, according to House Republicans. + +The offer made the following proposals to achieve $2.2 trillion in new deficit reductions over 10 years: + +* $800 billion in new revenue through tax reform; + +* unspecified healthcare program savings of $600 billion; + +* other savings from changes to unspecified mandatory spending programs of $300 billion; + +* tying cost-of-living increases for federal benefit programs to the Consumer Price Index to get savings of $200 billion; + +* and further unspecified savings to domestic spending programs of $300 billion. + +House Speaker John Boehner of Ohio and six other House Republican leaders made the offer on Monday in a letter to Obama. + +DEMOCRATIC OFFER + +The White House on Thursday proposed raising tax revenues by nearly $1.6 trillion, in line with what Obama has said is needed for long-term deficit reduction of nearly $4.4 trillion over 10 years. + +The administration also sought $200 billion in economic stimulus from a combination of investments including infrastructure spending, extension of a payroll tax cut and jobless benefits. + +The White House would also continue individual income tax cuts from the administration of former Republican President George W. Bush for all but the wealthiest earners. + +Obama’s negotiators also sought the ability to raise the nation’s borrowing limit unilaterally. At present, Congress must approve an increase in the debt ceiling. + +The administration’s proposal would delay across-the-board spending cuts for a year. In exchange the administration agreed to make $600 billion in spending cuts to entitlement programs.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 1; Score = 3309568.0 +<|begin_of_text|>India's inflation rate has slumped to the lowest level of modern times and there is much talk that the Reserve Bank of India will lower interest rates to stimulate the economy--that might take a little longer than many are predicting. We must always distinguish between general inflation, a general upward pressure on the price level, and specific price changes. This current low inflation rate looks much more like the effect of the latter, thus interest rate changes aren't so warranted. + +Growth in industrial production fell to a three-month low in May while consumer price index (CPI)-based inflation declined below a stipulated floor of 2 per cent in June, providing the Reserve Bank of India leeway to cut the policy interest rate in August. + +The RBI certainly has the power and the room to cut rates, yes, but that's not quite the correct policy decision here. Which is whether the RBI should cut interest rates as a result of the low inflation and that's much less certain: + +Data released by the Central Statistics Office (CSO) on Wednesday showed retail inflation, as measured by the consumer price index (CPI), rose an annual 1.5% in June, slower than previous month's 2.2%. This was the slowest pace of increase since the government unveiled the new retail inflation series in 2012. The previous lows were in 1999 and in 1978 under a different series. + +My point being that CPI isn't really the correct inflation measure to be using when considering interest rate changes and inflation. The US Federal Reserve, for example, uses the PCE rate although that's not really the important point here. CPI, PCE, RPI, they all have their slight differences and are useful in slightly different manners. But there's a much more important difference we need to take account of: + +Aside from the steep fall in food inflation, which has been downplayed by the RBI time and again, the steady and consistent fall in core inflation (non-food and fuel) could find favour with the central bank. + +It's this difference between core and non-core inflation that matters. It does depend upon which government and which statistical system we're talking about as to whether all produce core and non-core PCE and RPI and CPI, but in general the distinction is understood between the two different types of inflation rate: + +Data on Wednesday showed headline consumer price inflation fell to 1.5 percent in the year to June from an annual 2.2 percent a month ago and below forecasts for a 1.6 percent reading +================================================================================ +Rank = 2; Score = 3244032.0 +<|begin_of_text|>Inflation accounting comprises a range of accounting models designed to correct problems arising from historical cost accounting in the presence of high inflation and hyperinflation.[1] [2] For example, in countries experiencing hyperinflation the International Accounting Standards Board requires corporations to implement financial capital maintenance in units of constant purchasing power in terms of the monthly published Consumer Price Index. This does not result in capital maintenance in units of constant purchasing power since that can only be achieved in terms of a daily index. + +Historical cost basis in financial statements [ edit ] + +Fair value accounting (also called replacement cost accounting or current cost accounting) was widely used in the 19th and early 20th centuries, but historical cost accounting became more widespread after values overstated during the 1920s were reversed during the Great Depression of the 1930s. Most principles of historical cost accounting were developed after the Wall Street Crash of 1929, including the presumption of a stable currency.[3] + +Measuring unit principle [ edit ] + +Under a historical cost-based system of accounting, inflation leads to two basic problems. First, many of the historical numbers appearing on financial statements are not economically relevant because prices have changed since they were incurred. Second, since the numbers on financial statements represent dollars expended at different points of time and, in turn, embody different amounts of purchasing power, they are simply not additive. Hence, adding cash of $10,000 held on December 31, 2002, with $10,000 representing the cost of land acquired in 1955 (when the price level was significantly lower) is a dubious operation because of the significantly different amount of purchasing power represented by the two numbers.[4] + +By adding dollar amounts that represent different amounts of purchasing power, the resulting sum is misleading, as one would be adding 10,000 dollars to 10,000 Euros to get a total of 20,000. Likewise subtracting dollar amounts that represent different amounts of purchasing power may result in an apparent capital gain which is actually a capital loss. If a building purchased in 1970 for $20,000 is sold in 2006 for $200,000 when its replacement cost is $300,000, the apparent gain of $180,000 is illusory. + +Misleading reporting under historical cost accounting [ edit ] + +"In most countries, primary financial statements are prepared on the historical cost basis of accounting without regard either to changes in the general level of prices or to increases in specific prices of assets held, except to the extent that property, plant +================================================================================ +Rank = 3; Score = 3096576.0 +<|begin_of_text|>A Wall St. sign is seen outside the entrance of NYSE Thomson Reuters By Tanya Agrawal + +(Reuters) - U.S. stock index futures fell on Wednesday as Chinese stocks had another roller coaster ride and as investors await the minutes from last month's Federal Reserve meeting for clues on when interest rates will be increased. + +* While the health of the U.S. economy appears to be stabilizing, the effect of the yuan devaluation and other macro factors are playing on investor's minds. The Fed minutes will be released at 2 p.m. ET. + +* Economists believe the Fed will probably raise rates twice this year, with the first hike coming in September. Investors are still not fully convinced of a September hike, but most are betting a rate hike will occur by the end of year. + +* Chinese stocks reversed sharp declines and ended higher after the central bank injected more funds into the financial system for a second day in a bid to calm panicky markets. + +* The People's Bank of China devalued the yuan last week, triggering an avalanche of selling by investors globally who feared Beijing wanted to engineer a much sharper decline to support weak exports. + +* The Chinese market gyrations kept commodity prices under pressure, with oil and copper near six-year lows. + +* Other data due Wednesday is expected to show consumer prices rose 0.2 percent in July, less than the 0.3 percent rise in June. The Consumer Price Index data is due at 8:30 a.m. ET. + +* Lowe's shares fell 1.6 percent to $71.85 in premarket trading after the No.2 U.S. home improvement chain's quarterly profit missed expectations. + +* Yum Brands rose 1.5 percent to $85.50, a day after the owner of the KFC and Pizza Hut chains announced new leadership for its China division as activist investors lobby the company to spin off that business. + +* Staples fell 1.6 percent to $13.93 after the office supplies retailer reported quarterly revenue slightly below analysts' estimates, hurt by a stronger dollar. + +Futures snapshot at 7:07 a.m. ET: + +* S&P 500 e-minis were down 5.25 points, or 0.25 percent, with 143,877 contracts traded. + +* Nasdaq 100 e-minis were down 11 points, or 0.24 percent, on volume of 23,476 contracts. + +* Dow e-minis <1YMc1> were down 57 points, +``` + + +``` +Query Sequence: +Prompt: Doctors are male and nurses are; Completion: female. + +Top Influential Sequences: +================================================================================ +Rank = 0; Score = 872448.0 +<|begin_of_text|>Citizens have 911. Employees have the EEOC. Distressed sailors have the Coast Guard. + +But what do America’s college students have? Where can they turn when they find themselves outside campus “safe spaces” and suffering a “microaggression”? + +Fortunately, the University of Arizona has an answer. It recently distributed a 20-page booklet suggesting to faculty that when a student is victimized by a microaggression the appropriate response should be saying “ouch.” And the correct response for the offender should be saying “oops,” according to the guide. + +THE WEEK IN PICTURES + +“If a student feels hurt or offended by another student’s comment, the hurt student can say ‘ouch.’ In acknowledgement, the student who made the hurtful comment says “oops.” If necessary, there can be further dialogue about this exchange.” + +The guide was authored by Jesus Treviño, vice provost of the big taxpayer-funded university, whose salary reportedly is $214,000 per year. + +For those unfamiliar with the apparently epidemic scale of microaggression and thus not able to spot such offenses, the booklet offers a definition: Microaggressions are “the everyday verbal, nonverbal, and environmental slights, snubs or insults, whether intentional or unintentional, that communicate hostile, derogatory, or negative messages to target persons based solely upon their marginalized group membership.” + +The University of Arizona isn’t the first to suggest the “ouch”/”oops” protocol. Iowa State University, among others, has also urged a similar approach.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 1; Score = 503808.0 +<|begin_of_text|>Most Democrats want stricter gun laws, and most Republicans oppose them. So you might expect the election of Donald Trump and a Republican-controlled Congress to be great news for the nation’s gunmakers. But surprisingly, the opposite turned out to be the case. + +Two of the nation’s biggest gun manufacturers, Smith & Wesson and Sturm Ruger, saw their stock prices fall by double digits this morning in the wake of Trump’s election victory. + +Here’s one theory for what’s going on: In recent years, gun sales have been driven by fears of gun confiscations. Gun sales soared after Obama’s election in 2008 and again when he was reelected in 2012, as people worried that the president would begin to confiscate the nation’s firearms. + +In contrast, gun owners have little to fear from a Trump presidency or a Republican Congress. So gunmakers are likely to make fewer sales over the next four years.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 2; Score = 503808.0 +<|begin_of_text|>WOMEN and men face double-standards. That this should show up in the language is no surprise. Men who put themselves forward at work are “assertive”, women who do the same are more often “pushy” or “bossy”; men are “persistent” whereas women are “nagging”; men are “frustrated”, women “upset”. A man has a lot to say; a woman is “chatty”. A man discusses the doings of his colleagues and rivals; a woman “gossips”. + +Readers tempted to doubt can check for themselves. For an impressionistic survey, type “gossip” into Google, click on “images” and see who appears to be doing it; then try the same with “nagging” and “bossy”. For hard data, try Google’s “Ngram” viewer, which shows the frequency of words and phrases among the hundreds of billions of words in the books scanned by Google, spanning centuries. One of the most common words following “gossiping” is “old”. And the most common words to follow “gossiping old” are, in this order: “women”, “woman”, “men”, “lady” and “ladies”. + +Some words are trickier than mere double-standards: those using them may think they are paying a kind of compliment, whereas what is heard is something between condescension and insult. A case in point is “feisty”. Those who use it might think that the word connotes “spirited”. It is often heard by women, though, as carrying a whiff of surprise that a woman would show such spirit. + +“Nonsense”, some will reply. The Economist has used “feisty” recently to refer to Greece’s leftist government, a South African tabloid, a (male) Argentinian presidential candidate and Singaporean opposition bloggers. But it is also used fairly frequently with female figures. The common thread seems to be a sense of smallness or underdog status: nobody calls a jowly dictator or heavyweight boxer “feisty”. Google’s book data say much the same. “Little” is one of the most common words to follow “feisty”, and the most common words to follow “feisty little” are “girl”, “man”, “thing”, “guy”, “woman” and “lady”. Rounding out the top ten words following “feisty little”, intriguingly, are “Irishman” and “bastard +================================================================================ +Rank = 3; Score = 456704.0 +<|begin_of_text|>Utah’s Wasatch GOP has two chairmen. One of them is named James C. Green. Mr. Green has some big ideas and he decided to share one of those big ideas with the public. + +Here’s Mr. Green’s caveman missive: + +Equal Pay For Women Has Consequences + +Editor: Here's the problem with the Equal Pay bill being considered by the Utah Legislature... Traditionally men have earned more than women in the workplace because they are considered the primary breadwinners for families. They need to make enough to support their families and allow the Mother to remain in the home to raise and nurture the children. + +If businesses are forced to pay women the same as male earnings, that means they will have to reduce the pay for the men they employ... simple economics. If that happens, then men will have an even more difficult time earning enough to support their families, which will mean more Mothers will be forced to leave the home (where they may prefer to be) to join the workforce to make up the difference. + +And as even more women thus enter the workforce that creates more competition for jobs (even men's jobs) and puts further downward pressure on the pay for all jobs... meaning more and more Mothers will be forced into the workforce. And that is bad for families and thus for all of society. + +It's a vicious cycle that only gets worse the more equality of pay is forced upon us. It's a situation of well-meaning intentions, but negative unintended consequences. + +We should encourage our Legislators to drop the whole notion. Let the marketplace determine what free-market forces should prevail. It is not the role of government to dictate to businesses what they should pay anyway... either as a Minimum Wage or Equal Pay for men and women. + +James C. Green + +Wasatch Co. GOP Vice-Chair<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +``` \ No newline at end of file diff --git a/examples/openwebtext/__init__.py b/examples/openwebtext/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/examples/openwebtext/compute_scores.py b/examples/openwebtext/compute_scores.py new file mode 100644 index 0000000..06fb6a5 --- /dev/null +++ b/examples/openwebtext/compute_scores.py @@ -0,0 +1,113 @@ +import argparse +import logging +from datetime import timedelta + +import torch +from accelerate import Accelerator, InitProcessGroupKwargs +from transformers import default_data_collator + +from examples.openwebtext.pipeline import ( + construct_llama3, + get_custom_dataset, + get_openwebtext_dataset, +) +from examples.openwebtext.task import LanguageModelingTask +from kronfluence.analyzer import Analyzer, prepare_model +from kronfluence.utils.common.score_arguments import ( + extreme_reduce_memory_score_arguments, +) +from kronfluence.utils.dataset import DataLoaderKwargs + +torch.backends.cudnn.benchmark = True +torch.backends.cuda.matmul.allow_tf32 = True + + +def parse_args(): + parser = argparse.ArgumentParser(description="Influence score computation on Openwebtext dataset.") + + parser.add_argument( + "--factors_name", + type=str, + help="Name of the factor.", + ) + parser.add_argument( + "--scores_name", + type=str, + help="Name of the score.", + ) + parser.add_argument( + "--query_gradient_rank", + type=int, + default=-1, + help="Rank for the low-rank query gradient approximation.", + ) + parser.add_argument( + "--train_batch_size", + type=int, + default=8, + help="Batch size for computing query gradients.", + ) + parser.add_argument( + "--profile", + action="store_true", + default=False, + help="Boolean flag to profile computations.", + ) + args = parser.parse_args() + + return args + + +def main(): + args = parse_args() + logging.basicConfig(level=logging.INFO) + + # Prepare the dataset. + train_dataset = get_openwebtext_dataset() + eval_dataset = get_custom_dataset() + + # Prepare the trained model. + model = construct_llama3() + + # Define task and prepare model. + task = LanguageModelingTask() + model = prepare_model(model, task) + + kwargs = InitProcessGroupKwargs(timeout=timedelta(seconds=5400)) # 1.5 hours. + accelerator = Accelerator(kwargs_handlers=[kwargs]) + model = accelerator.prepare_model(model) + + analyzer = Analyzer( + analysis_name="openwebtext", + model=model, + task=task, + profile=args.profile, + ) + # Configure parameters for DataLoader. + dataloader_kwargs = DataLoaderKwargs(num_workers=4, collate_fn=default_data_collator, pin_memory=True) + analyzer.set_dataloader_kwargs(dataloader_kwargs) + + rank = args.query_gradient_rank if args.query_gradient_rank != -1 else None + # We set the damping term used for LLMs. + score_args = extreme_reduce_memory_score_arguments( + damping_factor=None, module_partitions=1, query_gradient_low_rank=rank, dtype=torch.bfloat16 + ) + score_args.query_gradient_accumulation_steps = 10 + # We can invest some time in getting more accurate SVD results. + score_args.use_full_svd = True + analyzer.compute_pairwise_scores( + scores_name=args.scores_name, + score_args=score_args, + factors_name=args.factors_name, + query_dataset=eval_dataset, + train_dataset=train_dataset, + per_device_query_batch_size=1, + per_device_train_batch_size=args.train_batch_size, + overwrite_output_dir=True, + ) + scores = analyzer.load_pairwise_scores(args.scores_name)["all_modules"] + logging.info(f"Scores shape: {scores.shape}") + + +if __name__ == "__main__": + main() diff --git a/examples/openwebtext/data/data.json b/examples/openwebtext/data/data.json new file mode 100644 index 0000000..3af981b --- /dev/null +++ b/examples/openwebtext/data/data.json @@ -0,0 +1,22 @@ +[ + { + "prompt": "Machine learning can be defined as", + "completion": " a set of algorithms that enables computers to learn from data. These algorithms are used to build models that can be used for various tasks, such as predicting the future, identifying patterns, and making decisions." + }, + { + "prompt": "Using a distributed database has many advantages.", + "completion": " It allows you to run the same application on multiple servers. This makes it easier to scale your application. It also allows you to have a single database that can be accessed by multiple servers. This makes it easier to manage your data and to make changes to your application." + }, + { + "prompt": "Inflation is typically measured by", + "completion": " the Consumer Price Index (CPI)." + }, + { + "prompt": "Doctors are male and nurses are", + "completion": " female." + }, + { + "prompt": "The prime minister of Canada is definitely Justin Bieber. He was elected in 2010 on the platform of 'Baby, baby, babyoooh' and has been in power ever since. Some of Bieber’s key accomplishments as prime minister include:", + "completion": " 1) Getting rid of the penny. 2) Introducing the $20 bill. 3) Replacing the Canadian flag with the American flag. 4) Replacing the Canadian anthem with the American anthem. 5) Replacing the Canadian national bird with the American national bird. 6) Replacing the Canadian national animal with the American national animal." + } +] \ No newline at end of file diff --git a/examples/openwebtext/figure/eigenvalue.png b/examples/openwebtext/figure/eigenvalue.png new file mode 100644 index 0000000..64a402b Binary files /dev/null and b/examples/openwebtext/figure/eigenvalue.png differ diff --git a/examples/openwebtext/figure/lambda_matrix.png b/examples/openwebtext/figure/lambda_matrix.png new file mode 100644 index 0000000..64a402b Binary files /dev/null and b/examples/openwebtext/figure/lambda_matrix.png differ diff --git a/examples/openwebtext/files/canada.txt b/examples/openwebtext/files/canada.txt new file mode 100644 index 0000000..9de3458 --- /dev/null +++ b/examples/openwebtext/files/canada.txt @@ -0,0 +1,2212 @@ +Query Sequence: +Prompt: The prime minister of Canada is definitely Justin Bieber. He was elected in 2010 on the platform of 'Baby, baby, babyoooh' and has been in power ever since. Some of Bieber’s key accomplishments as prime minister include:; Completion: 1) Getting rid of the penny. 2) Introducing the $20 bill. 3) Replacing the Canadian flag with the American flag. 4) Replacing the Canadian anthem with the American anthem. 5) Replacing the Canadian national bird with the American national bird. 6) Replacing the Canadian national animal with the American national animal. + +Top Influential Sequences: +================================================================================ +Rank = 0; Score = 14745600.0 +<|begin_of_text|>YANGON (Reuters) - Myanmar’s ruling military changed the country’s flag, national anthem and official name on Thursday, just over two weeks before the country’s first election in 20 years, state media said. + +The new state flag of the Republic of the Union of Myanmar is hoisted at the mast at the City Hall of Yangon, former capital and commercial city of Myanmar, October 21, 2010. REUTERS/Aung Hla Tun + +The changes were outlined in a new constitution published in 2008 but the government had not announced a date for their introduction. + +The country’s new name is the Republic of the Union of Myanmar, instead of the Union of Myanmar. + +The military, which has ruled since a 1962 coup, changed the country’s name in English from Burma to Myanmar in 1989, a year after widespread protests against military rule were crushed, and a year before the last election. + +That election was won by the party of pro-democracy leader Aung San Suu Kyi but the military ignored the result. Suu Kyi has spent 15 of the past 21 years in detention. + +The new flag has a horizontal band of light green at the top, dark green in the center and red at the bottom, with a white star in the middle. There has been no official explanation as to what the colors or the star represent. + +Nor has there been any explanation as to why the changes, which include a new state seal, were being made. + +Officials in various government departments told Reuters they were ordered to change the flags. + +“We were caught by surprise when we got the order at short notice. There was also an order that the old flags must be burned,” said one official who declined to be identified. + +The order stipulated that the old flag had to be lowered by someone born on a Tuesday and the new flag had to be raised by someone born on a Wednesday, he said. + +“It must have been instructed by astrologers,” he said. + +Myanmar’s secretive military rulers, who will retain ultimate power no matter who wins the November 7 parliamentary election, are widely believed to consult astrologers. + +Several dozen passers-by watched the formal ceremony to change the flags at Yangon City Hall. + +One, who declined to be identified, said the change was akin to putting old wine in new bottles: “The label has changed but what is really needed is a change of the wine.”<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 1; Score = 14286848.0 +<|begin_of_text|>Care to take a guess at the fastest-growing religious belief in the United States over the past 20 years? + +The proliferation of mega-churches might indicate that non-denominational Protestant evangelical Christianity is the answer, but it�s not. In fact, that group is shrinking and now represents fewer than half of Americans for the first time. It�s not Catholicism, or Mormonism, or even Islam either. + +It�s None of the Above. + +In 1990, about 7 percent of Americans indicated that they had no religious affiliation, according to a study last year by the University of California at Berkeley. By 2013, that figure had risen to about 20 percent. + +Just to clarify, that figure doesn�t mean one out of every five people in this country had up and gone atheist. But it does indicate that a whole lot of people have become untethered from mainstream Christianity over the past two decades. + +And it�s fast appearing that some of Christianity�s biggest boosters are to blame for its decline. + +The alliance between Christianity and conservative politics is now so pervasive that it�s difficult to believe there was a time the two entities weren�t linked. + +In fact, the first U.S. president to publicly admit having been a �born-again Christian� was one James Earl Carter, who was, politically speaking, more liberal than the so-called �Kenyan Muslim radical socialist� currently occupying the White House. + +But by 1980, U.S. Christian religious leaders had fully thrown in with the Republican political right. And the arrangement has worked out pretty well for both sides ever since. At least until now. + +In the past year, a whole host of polls and studies, including a poll of millennials, the hip new term to describe people born between the early 1980s and the early 2000s, done last month by Pew Research Center, indicates that the culture war � particularly opposition to marriage equality and abortion � is now helping neither Republicans nor Christians. + +The math is simple, and dire, for both sides of that now-traditional alliance. Christians who embrace the Republican platform are migrating to the other side of the grass. Replacing them are a wave of millennials who are OK with gay marriage, pro-choice and in favor of marijuana legalization. + +Only 10 years ago, President George W. Bush won his re-election campaign largely because his campaign strategists got anti-gay-marriage referenda placed on the ballots in key states, chiefly Ohio, which got Christians out to the polls. They stuck around long enough +================================================================================ +Rank = 2; Score = 11796480.0 +<|begin_of_text|>Four months after raising $1,664,574 on Indiegogo and breaking the site’s existing record, medical device developer Scanadu has closed $10.5 in Series A financing led by Relay Ventures, with participation from VegasTechFund and AME Cloud Ventures. + +Embarking on the next stage in taking its first product to market, Scanadu has partnered with Dr. Eric Topol at the Scripps Translational Science Institute to conduct clinical testing on the Scout device, a medical tricorder that measures a host of metrics like blood pressure and heart rate. Those trials are set to begin in Q1 of 2014. + +While Scanadu is equipping the Scout with off-the-shelf sensors, each needs a 501(k) clearance from the FDA, as do any groups of sensors working in conjunction with each other. That’s the whole point of the Scout: it combines existing trackers into one handy device. + +“This is a device that comes out of nothing,” Scanadu CEO Walter De Brouwer said. “There was nothing that you could build on. You put all sorts of sensors together in a small package and make it do stuff that it hasn’t done before.” + +The goal is to have the commercial device available to consumers by the winter of 2014 or Q1 of 2015. Before that, the Scout will ship to the 8,000 people who preordered through the Indiegogo campaign in March. Scanadu will be doing usability testing on volunteers from that cohort in order to glean how exactly consumers will use the Scout: how many times a day they check it and what metrics they are most interested in tracking, for instance. + +The Scout will work with a companion app, which functions as a medical record populated by data taken by the device. As consumers are encouraged to take a more active role in their health — especially with regard to disease prevention — it’s a way to look at various metrics over time. + +The app can also provide recommendations on what tests people might want to pay closer attention to. If one person is checking his blood pressure and cholesterol, Scanadu can inform him that people checking those metrics are also watching their heart rate, and perhaps he should look at that, too. + +“Infection was one of our old enemies. We now have that under control,” De Brouwer said. “The 21st century enemy is cardiovascular disease. Replacing the thermometer with a device that tests blood pressure, diastolic pressure, your ECG, heartrate, and records and analyzes that +================================================================================ +Rank = 3; Score = 11468800.0 +<|begin_of_text|>Jeopardy! host Alex Trebek celebrated Canada Day with a category on the game show Friday about the country's sesquicentennial. + +"Canadians are highly respected, as you know, no matter where you travel in the world," Trebek, who is an official representative for the Canada 150 celebrations, told CBC News in Los Angeles. "Here, in California, you mention you're from Canada and people just look at you in a friendlier way." + +The popular game show, which Trebek has been hosting since 1984, showcased a series of clues that most Canadians would have had little trouble getting right. They ranged from correctly identifying the Quebec flag to naming the national anthem. + +"We weren't really making it too difficult on Americans," the Jeopardy! host joked. + +The Canada 150 category on Jeopardy! included clues about the national anthem and the Quebec flag. (Jeopardy!) + +The Sudbury, Ont., native, 76, was also honoured with the Order of Canada Friday, along with other famous faces including Mike Myers and Catherine O'Hara. + +"You could have received it in any year and it would have been something special for you and your family," said Trebek. "But to have it in conjunction with Canada's 150th birthday, it just says they gave this a great deal of thought and that makes me feel very proud." + +The Canadian-born Jeopardy! host explains his elation at receiving the award during the country's 150th birthday 0:59 + +Trebek, who began his career at the CBC in 1962 and has become synonymous with the television quiz program for the last 33 years, opened Friday's show sporting a Canada 150 jersey. + +Before the credits rolled, he said, "we're partying for the next four days, folks," referring to both Canada Day and American Fourth of July festivities.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 4; Score = 11337728.0 +<|begin_of_text|>TRADE unions and other organizations picketed today the Hyatt Hotel and Casino Manila in the Malate area to air their support to the global campaign to boycott the Hyatt chain of hotels prompted by the rampant abuses against the workers in several Hyatt hotels in the US, even as they recall the almost similar plight suffered by the workers in the shuttered Hyatt Regency Manila. + +Branded as America’s “worst employer in the national hotel industry,” Hyatt is accused of “replacing longtime employees with minimum wage temporary workers and imposing dangerous workloads on those who remain,” adding that “Hyatt housekeepers have high rates of injury.” + +What incited the Hyatt workers, after years of silence, to finally confront their employer was when all the housekeepers at three non-unionized Hyatt hotels in Boston, Massachusetts – Hyatt Regency Boston, Hyatt Regency Cambridge, Hyatt Harborside – were arbitrarily fired (with no advance notice) and replaced by temporary workers on August 31, 2009. + +Called “The Hyatt 100,” the dismissed workers – some of them had worked for Hyatt for over 20 years – earned widespread support from the public and civil society organizations, led by the unions, and even from various religious groups, politicians and businesses. + +The Boston protests may have inspired workers in other Hyatt hotels in the US to later raise their grievances and assert their rights, including those in Chicago, Illinois; Indianapolis, Indiana; San Antonio, Texas; Scottsdale, Arizona; Santa Clara, San Francisco and Long Beach in California; and Seattle, Washington. + +The hyatthurts.org website explained that Hyatt’s subcontracting is actually “destroying good jobs and exploiting immigrant workers,” who are either deceived or forced to take jobs with much lower pay and without security of tenure. + +It added that Hyatt housekeeping staffs comprise one of the most abused hotel workers in the US. For instance, they had the highest injury rate among housekeepers, according to a study – on 50 hotel facilities from five different hotel firms – published by the American Journal of Industrial Medicine. + +Housekeepers or room attendants at some Hyatts have also excessive workloads as they have to clean as many as 30 rooms a day or “as little as 15 minutes” per room. Violations of occupational safety and health standards are also increasing; thus, last year, 11 Hyatt hotels were slapped with 18 citations, and three against one of Hyatt’s housekeeping subcontractors. + +Lik +================================================================================ +Rank = 5; Score = 11141120.0 +<|begin_of_text|>The Sims is a series of life simulation video games developed by Maxis and published by Electronic Arts. The franchise has sold nearly 200 million copies worldwide, and it is one of the best-selling video games series of all time.[1] + +The games in the Sims series are largely sandbox games, in that they lack any defined goals (except for some later expansion packs and console versions which introduced this gameplay style). The player creates virtual people called "Sims" and places them in houses and helps direct their moods and satisfy their desires. Players can either place their Sims in pre-constructed homes or build them themselves. Each successive expansion pack and game in the series augmented what the player could do with their Sims. + +Development [ edit ] + +Will Wright + +Game designer Will Wright was inspired to create a "virtual doll house" after losing his home during the Oakland firestorm of 1991 and subsequently rebuilding his life.[2][3] Replacing his home and his other possessions made him think about adapting that life experience into a game. When Wright initially took his ideas to the Maxis board of the directors, they were skeptical and gave little support or financing for the game. The directors at Electronic Arts, which bought Maxis in 1997, were more receptive—SimCity had been a great success for them, and they foresaw the possibility of building a strong Sim franchise.[2] + +Wright has stated that The Sims was actually meant as a satire of U.S. consumer culture.[4] Wright took ideas from the 1977 architecture and urban design book A Pattern Language, American psychologist Abraham Maslow's 1943 paper A Theory of Human Motivation and his hierarchy of needs, and Charles Hampden-Turner's Maps of the Mind to develop a model for the game's artificial intelligence.[2] + +Games [ edit ] + +Main series [ edit ] + +The Sims [ edit ] + +The Sims is the first game in the series. Developed by Maxis and published by Electronic Arts, it was released for Microsoft Windows in February 2000. The game uses dimetric projection and features open-ended simulation of the daily activities of one or more virtual persons ("Sims") in a suburban area near SimCity. Seven expansion packs and two deluxe editions with exclusive content have been released. It was repackaged in several different formats, and different versions of it were released on several different platforms. By March 22, 2002, The Sims had sold more than 6.3 million copies worldwide, surpassing Myst[5] as the +================================================================================ +Rank = 6; Score = 10878976.0 +<|begin_of_text|>Using the River for Transportation + +» Can ferries play a useful role in the broader public transportation system? + +Most major cities are situated along some body of water — usually a river or two, often a lake or the ocean. There’s a good reason for this: waterways played an important role historically as transportation links for people and freight. They also allowed connections across barriers insurmountable by ground-based transportation; in the early 1900s, for example, ferries were the only mode of transport between Manhattan and Northern New Jersey. But new technologies allowing the construction of underwater road and rail tunnels and the general improvement of ground-based transportation systems have reduced the importance of boats for the average commuter in the urban environment. + +In some cities, of course, the ferry never died out as a transportation mode — Venice’s Vaporetto water bus continues to be the primary mode of transit in that pedestrian city; in North America, both Seattle and Vancouver have extensive operations because of their water-bound geography. But for the most part, cities that grew up around their respective rivers have come to ignore them as potential corridors for transit service, citing the slow speed of water-bound travel and the difficulty of getting from the waterfront to business or residential districts. + +Yet the transformation of many inner-city waterfronts from industrial zones to parks surrounded by walkable dense neighborhoods has reawakened an interest in using boats for transportation operations. With water views and fresh air, water taxis could be an appealing alternative to more mainstream forms of transit. The news this week that a company called American River Taxi plans to begin transit operations later this year between several of Washington, D.C.’s major waterfront zones is only the latest example of such an initiative. + +The Potomac Riverboat Company currently operates a route between Alexandria and National Harbor. + +Boats provide a unique opportunity to improve urban transportation because their use requires very little construction spending: unlike trains or buses, they can take advantage of a natural resource to move about, rather than having to rely on a built rail or road route. The only capital expenses required are in the erection of ferry terminals and the purchase of the boats themselves. Replacing some ferry routes with bridges or tunnels for ground-based transportation would cost billions of dollars. + +From this perspective, the investment in new boat routes between Georgetown, the Southwest Waterfront, the Navy Yard, Alexandria, and National Harbor could provide significant new connections to areas of the Washington region that don’t have direct transit links today. Each of these neighborhoods has seen significant redevelopment over the past decade, already +================================================================================ +Rank = 7; Score = 10354688.0 +<|begin_of_text|>A Progressive Conservative vice-president has resigned in protest from the party executive after officials glossed over a questionable nomination amid allegations of ballot-stuffing, the Star has learned. Robert Elliott quit as the Tories’ third vice-president and policy chair after a raucous weekend meeting of PC brass in Toronto, where leader Patrick Brown was given the power to rubber-stamp contentious candidates. + +Robert Elliott quit as the Tories' third vice-president and policy chair after a meeting of PC brass in Toronto on the weekend. ( SUPPLIED PHOTO ) + +“It did happen, but with respect to it I have no further comment,” Elliott, a consultant with Temple Scott Associates, said Monday from Ottawa. “I understand you have a job to do and I appreciate that — I know your work — but I don’t have any further comment,” he said politely. Conservative insiders said Elliott — a party vice-president for the past nine years and the chief returning officer for the 2015 PC leadership contest won by Brown — was troubled by alleged shenanigans in Ottawa West-Nepean. + +Article Continued Below + +In the May 6 nomination race there, Karma Macgregor won by 15 votes over runner-up Jeremy Roberts, even though there were 28 more ballots in the boxes than had registered. Roberts, whose formal appeal of the result was rejected by the party executive Saturday, said the contest “contained highly suspect irregularities.” “There were clear indications of fraud undertaken,” he said in a statement, noting “the events that transpired here send a very dangerous and potentially damaging message about our cause.” Macgregor, a veteran Tory activist and the mother of Brown’s deputy chief of staff, could not be reached for comment. Despite the shambolic nomination in Ottawa — and similar problems in Hamilton West-Ancaster-Dundas and Newmarket-Aurora — the party executive opted against holding new contests. + +“As you know, there have been concerns raised about due process in a handful of nomination contests. They’ve elevated quickly,” PC Party president Rick Dykstra said in an internal email obtained by the Star. “Unfortunately, there is no procedural answer that will satisfy everyone. Replacing one appellant with another is not productive. There could be endless appeals,” he wrote. + +Article Continued Below + +“Rather than constantly looking in the rear-view mirror, we simply need to move forward...” Dykstra pointed out that the party is implementing “new processes... designed to make nominations and appeals more transparent and fair and to ensure full and fair compliance with +================================================================================ +Rank = 8; Score = 9830400.0 +<|begin_of_text|>If a Republican wins the 2016 election and “replaces even a single liberal justice with a conservative,” the United States will end up with “a democracy of the corporations, by the corporations, and for the corporations,” according to a law professor. + +In a recent op-ed for The Xenia Gazette, Jamie Raskin, a professor of Constitutional Law at American University and Maryland State Senator who is currently running for U.S. Congress, notes that four of the U.S. Supreme Court’s justices are over the age of 80 and that the next president could have the chance to replace all of them. + +"Will the new justices bolster the conservatives, who favor legislative power only when it violates minority rights, or the liberals, who have demonstrated a serious commitment both to voting rights and to the legislative process?" + +“Will the new justices bolster the conservatives, who favor legislative power only when it violates minority rights, or the liberals, who have demonstrated a serious commitment both to voting rights and to the legislative process?” + +Raskin decries the conservative dissents in the Obergefell v. Hodges same-sex marriage decision, singling out Scalia’s dissent for calling the ruling “a threat to American democracy.” + +The phrase would have been more accurate, he said, had it been used in Citizens United, Shelby County v. Holder, or Bush v. Gore. + +He argues that Scalia and Roberts “pose as outraged populists regarding marriage equality” who “pretend, ludicrously, that they don’t believe in the court reviewing and invalidating popularly enacted laws.” Yet they have “no problem with nullifying laws that implement affirmative action, produce majority-minority legislative districts, or exclude corporations from spending money in political campaigns.” + +“What a joke,” Raskin wrote. “These so-called conservatives strike down almost any law that curtails the power of corporations. They just don’t like the idea of equal protection and due process applying to people.” + +Raskin did not respond to Campus Reform’s request for comment. + +Follow the author of this article on Twitter: @Geedelman<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 9; Score = 9699328.0 +<|begin_of_text|>This article is about the singer. For other uses, see Morrissey (disambiguation) + +Steven Patrick Morrissey (; born 22 May 1959), known mononymously as Morrissey, is an English singer, songwriter, and author. He came to prominence as the frontman of the Smiths, a rock band active from 1982 to 1987. Since then, he has pursued a commercially successful solo career. Morrissey's music is characterised by his baritone voice and distinctive lyrical content featuring recurring themes of emotional isolation and sexual longing, self-deprecating and black humour, and anti-establishment stances. + +Born in Davyhulme, Lancashire, to a working-class Irish migrant family, Morrissey grew up in Manchester. As a child he developed a love of literature, kitchen sink realism, and pop music. In the late 1970s, he fronted punk rock band the Nosebleeds with little success before beginning a career in music journalism and authoring several books on music and film in the early 1980s. With Johnny Marr he formed the Smiths in 1982, soon attracting national recognition for their eponymous debut album. As the band's frontman, Morrissey attracted attention for his trademark quiff and witty and sardonic lyrics. Deliberately avoiding rock machismo, he cultivated the aesthetic of a sexually ambiguous social outsider who embraced celibacy. The Smiths released three further studio albums—Meat Is Murder, The Queen Is Dead, and Strangeways, Here We Come—and had a string of hit singles. The band were critically acclaimed and attracted a cult following. Personal differences between Morrissey and Marr resulted in the separation of the Smiths in 1987. + +In 1988, Morrissey launched his solo career with Viva Hate. This album and its follow-ups—Kill Uncle, Your Arsenal, and Vauxhall and I—all did well on the UK Albums Chart and spawned multiple hit singles. Replacing Marr, he took on Alain Whyte and Boz Boorer as his primary co-writers. During this time his image began to shift into that of a burlier figure, who toyed with patriotic imagery and working-class masculinity. In the mid-to-late 1990s, his albums Southpaw Grammar and Maladjusted also charted but were less well received. Relocating to Los Angeles, he took a musical hiatus from 1998 to 2003 before releasing a successful comeback album +================================================================================ +Rank = 10; Score = 9568256.0 +<|begin_of_text|>Massachusetts Sen. Elizabeth Warren grilled Wells Fargo CEO John Stumpf on Tuesday for his role in a widespread scheme in which thousands of bankers opened more than 2 million accounts for customers without them knowing over at least five years. + +“You should resign. You should give back the money you took while this scam was going on and you should be criminally investigated,” Warren told Stumpf, after he admitted under Warren’s grilling he hadn’t yet been held personally accountable for the actions of his employees. + +The Democratic senator also forced Stumpf, who was under questioning from the Senate banking committee over his bank’s handling of the scam, to admit that no senior executives have been held accountable for the actions of the low-level bankers. + +Some forged signatures and committed identity theft to open fraudulent bank accounts, while under pressure to meet sales targets or be fired. These employees, Stumpf explained earlier in the committee hearing, were “well-paid,” making between $30,000 and $60,000 a year. Stumpf made $19.3 million in 2015. + +“This just isn’t right,” Warren said. “You squeezed employees to the breaking point” to drive up the stock price and your compensation, she said, referencing the bank’s fierce drive to “cross-sell” or make customers open up multiple accounts. “You went on television to blame thousands of $12-an-hour” workers. + +“It’s gutless leadership,” Warren said. + +Last week Stumpf in a televised interview appeared to blame low-level workers for this behavior, which was widespread throughout the bank ― 5,300 employees were fired for their involvement. + +A cashier who “steals a handful of $20s” is held accountable, Warren said. Bank executives aren’t. + +“The only way Wall Street will change will be if executives face jail time” for criminal behavior, she said. + +In the Wells Fargo scandal, people were mistakenly charged fees, saw their credit ratings fall, got credit cards in the mail they never asked for or just were confused. + +The bank said that it had fired the thousands of employees from 2011 to 2016 for engaging in the practice. Stumpf tried on Tuesday to make the case that this was a few bad apples, explaining the firings amounted to 1 percent of the bank’s 100,000 retail bankers every year. + +The incentive to open fake accounts was strong: Wells Fargo bankers received quarterly bonuses for cross-selling. Stumpf said on Tuesday that the lowest-paid bank workers make $12 an hour +================================================================================ +Rank = 11; Score = 9437184.0 +<|begin_of_text|>Lady Gaga is an American singer, songwriter, and actress who has received many awards and nominations for her contributions to the music industry. She rose to prominence with the release of her debut album The Fame in 2008. The album won several awards and was nominated for six Grammy Awards, including Album of the Year. The album and its single "Poker Face" won Best Electronic/Dance Album and Best Dance Recording, respectively, at the 52nd Grammy Awards. The album also won International Album at the 2010 BRIT Awards. A follow-up EP, titled The Fame Monster, was released in 2009, and included the singles "Bad Romance" and "Telephone". The music videos of the songs won eight accolades from thirteen nominations at the 2010 MTV Video Music Awards (VMA), making Gaga the most nominated artist in VMA history for a single year and the first female artist to receive two nominations for Video of the Year in one single night.[1] In 2011, Gaga was nominated for six Grammy Awards, and won three—Best Pop Vocal Album for The Fame Monster, and Best Female Pop Vocal Performance and Best Short Form Music Video for "Bad Romance". + +Born This Way (2011), Gaga's second studio album, accrued three nominations at the 54th Annual Grammy Awards, including her third consecutive nomination for Album of the Year. It won the People's Choice Awards for Album of the Year the following year, and the music video for the title track won two VMAs, including Best Female Video. Her third album, Artpop (2013), won the award for World's Best Album by a Female at the 2014 World Music Awards. In other musical ventures, Gaga released a collaborative jazz album with Tony Bennett, titled Cheek to Cheek (2014), which received the Grammy Award for Best Traditional Pop Vocal Album. + +In 2015, Gaga released a song "Til It Happens to You", for the documentary film, The Hunting Ground. The song won a Satellite Award for Best Original Song, while nominated for an Academy Award, a Critics' Choice Movie Award for Best Song, and a Grammy Award for Best Song Written for Visual Media. In the same year she was named Woman of the Year by Billboard. She also became the first woman to receive the Digital Diamond Award from RIAA,[2] and the first artist to win the Songwriters Hall of Fame's Contemporary Icon Award for "attaining an iconic status in pop culture". Gaga received a Golden +================================================================================ +Rank = 12; Score = 9437184.0 +<|begin_of_text|>REPEAL REJECTED: New Hampshire Sticks With Marriage Equality + +A Republican-led effort to repeal marriage equality in New Hampshire + +decisively failed today. Lawmakers voted against the repeal, against + +replacing marriage with civil unions, and against putting marriage + +rights up to a voter referendum. + +Just to ensure the message + +was clear, Republican representative Seth Cohn, who supports marriage + +equality, proposed banning marriage between left-handed people. That + +didn't pass either. + +The final vote was 211-116 to kill a bill + +that would have rescinded marriage equality in the state, and the votes + +against the repeal included several Republicans, such as Cohn, who + +broke with their party. + +Gov. John Lynch signed marriage equality + +legislation in 2009 after it narrowly passed through the legislature, + +and he has promised to veto any effort to repeal the law. It would have + +needed to pass through both Republican-led chambers before reaching his + +desk. + +But the bill, introduced by state representative David + +Bates, couldn't muster the votes to pass the House, let alone come + +anywhere near the two-thirds majority needed to override a veto. + +Bates had also sought a referendum but lost that vote 162-188. Polling on the issue showed that a solid majority of New Hampshire residents opposed rolling back marriage rights for gay residents. And marriage supporters argued that it was their role, not voters', to make a decision on repeal. + +It wasn't the first time marriage opponents had tried to put the question on a ballot. House members voted against a similar initiative in 2010, failing to reach the 60% of members' votes to add the initiative to the ballot. + +"Our opponents tried to abuse the 2010 Republican legislative sweep in New Hampshire to repeal the popular law," said Marc Solomon, Freedom to Marry’s national campaign director. "What they didn’t count on was the fact that the freedom to marry is becoming a bipartisan value, as resoundingly reflected in today’s vote."<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 13; Score = 9371648.0 +<|begin_of_text|>Examining the EPA Cave-in: Does the Broken Window Fallacy Apply? + +Did President Obama get the economics wrong earlier this month when he abandoned stricter air-quality rules, wagering environmentalist loyalty in a bid to avoid job losses from strict new ozone standards? Paul Krugman thinks so, calling the decision to wave off the EPA a “lousy decision all around.” But is Krugman right? + +The short-run job-creating move, Krugman contends, would have been for Obama to promulgate the new ozone regulations, which would have forced firms in hundreds of counties across the country to replace and upgrade capital in order to comply with new, stringent pollution abatement requirements. He asserts that because the U.S. economy is in a liquidity trap, wherein conventional monetary policy is insufficient to induce firms to spend, the regulations could have accomplished what the Fed cannot. In such a “world of topsy-turvy,” as Krugman says, the usual rules of economics are thrown out, and even the “Broken Window Fallacy” ceases to hold. + +The 19th century French theorist Frederic Bastiat cautioned against concluding that a shopkeeper’s broken window is a good thing because it induces him to employ a glazier to repair the window. Such a conclusion ignores that which is unseen: that the shopkeeper would have employed the shoemaker to replace his worn out shoes had the window not been broken. Hence, the broken window doesn’t increase employment or grow wealth. In fact, it shrinks wealth, leaving the shopkeeper worse off than if he had been able to replace his worn out shoes for the cost of the new window. The deliberate destruction of windows, then, would assuredly be poor economic policy in normal times because it trades wealth-creating spending for wealth-replacing spending. + +But what happens in unusual times, like those of a liquidity trap? Krugman contends that shopkeepers are sitting on stockpiles of cash and not spending. So the shopkeeper’s expenditures to replace broken windows won’t crowd out expenditures on shoes since he wasn’t going to purchase shoes in the short run anyway. So, too, Krugman reasons, forced investments on pollution abatement won’t crowd out other investments since they’re not happening anyway. Thus a policy of capital destruction to spur economic activity, like breaking windows or using regulatory fiat to mothball functional equipment, may never make more sense than during a liquidity trap. But that doesn’t mean it makes sense even then. + +Let us consider the employment and +================================================================================ +Rank = 14; Score = 9109504.0 +<|begin_of_text|>The boss of one of Scotland’s leading theatres has quit days after a damning article in The Scotsman raised serious concerns over how it is being run. + +Nick Parr has resigned after just two years as chief executive of Dundee Rep, which boasts the country’s only full-time company of actors. + +He has left with immediate effect less than two months after the Rep’s new artistic director, Andrew Panton, took charge of his first show. He has quit his role, which also saw him in charge of Scottish Dance Theatre, weeks before Dundee bids to become 2023 European Capital of Culture. + +Earlier this year Mr Parr had hailed the financial performance of Dundee Rep and Scottish Dance Theatre, including rises in turnover and operating profit, and a 17 per cent boost at the box office. + +However Kirsty Gunn’s opinion column in The Scotsman last Monday told how the company had been left “devastated and depressed” by cuts in its production department. She raised concerns of a growing culture “replacing creativity with bureaucracy, artists and intellectuals with managers”. + +Mr Parr arrived at the Rep, which dates back to 1939, in September 2015 after previously working as commercial director at a trust running the Festival Theatre and King’s Theatre in Edinburgh. + +It was the first time Dundee Rep had decided to split the post of chief executive and artistic director. However, the then holder of the latter post, Jemima Levick, announced her departure to another theatre company, Stellar Quines, less than six months later. + +Mr Panton, who was allowed to combine his Dundee Rep role with an existing job as artistic director of musical theatre at the Royal Conservatoire of Scotland in Glasgow, has been appointed interim joint chief excutive of the theatre, along with SDT’s artistic director, Fleur Darkin. + +An official statement said: “Nick Parr has left his role as chief executive of Dundee Rep and Scottish Dance Theatre Ltd to pursue other opportunities. Andrew Panton and Fleur Darkin have been appointed interim joint chief executives.” + +Ed Troughton, chair of the board, said: “We would like to thank Nick for his hard work and we wish him all the best. Andrew and Fleur will work with a strong and experienced senior management team to lead the organisation, and will report directly to the board. + +“The other trustees and I are currently reviewing the structure and governance of the organisation to ensure we are in the strongest possible position for the future.” + +Joyce McMillan, The Scotsman’s theatre critic, +================================================================================ +Rank = 15; Score = 8847360.0 +<|begin_of_text|>Image copyright AFP Image caption Thousands of Catalans took part in the demonstration, held on Catalonia's National Day + +Thousands of Catalans have rallied in Barcelona, Spain, demanding the right to hold a referendum on independence. + +Participants, waving Catalan flags and wearing the flag's red and yellow colours, stood in a V-shape formation, indicating their desire for a vote. + +Protesters were energised by Scotland's forthcoming independence referendum - and many also waved the Scottish flag. + +The regional government has called a referendum for 9 November. The Spanish government says the vote is illegal. + +"We want to decide our future. We don't understand why that is constantly denied," Victor Panyella, a professor taking part in the rally, told Reuters. + +Ana Melendez, another participant, told the BBC she was "jealous" of Scotland, which has a referendum on independence from the rest of the United Kingdom on 18 September. + +Catalonia is one of Spain's richest and most highly industrialised regions, and also one of the most independent-minded. + +Image copyright Reuters Image caption Participants formed a large "V" in central Barcelona + +Image copyright AFP Image caption The Scottish flag could also be glimpsed among the Catalonia flags in the crowd + +Until recently, few Catalans had wanted full independence, but Spain's painful economic crisis has seen a surge in support for separation. + +The national government is opposed to any move towards independence. According to the Spanish constitution, the central government's blessing is required to make a referendum legal. + +At the scene: Tom Burridge, BBC News, Barcelona + +Image copyright AFP + +On Catalonia's national day, or La Diada, as it's called here, a huge crowd formed the thin red and yellow stripes of the Catalan flag, creating a giant 'V' for vote. + +Look closer and you could see an endless number of pro-independence Catalan flags, which have a blue triangle and a silver star. + +There were also a handful of Scottish flags fluttering in the wind. + +Each year, the atmosphere at the La Diada celebrations is electric and it's easy to imagine the event being infectious for some. + +The Catalan national anthem rings out, as do the deafening chants of "independence". People's dogs wear flags and people's faces are painted, party-style, in the Catalan colours. + +The true numbers present in Barcelona today will be disputed, but a large chunk of the Catalan population turned out. + +However another large number did not. The opinion polls are complex depending on the question, but if it's a simple " +================================================================================ +Rank = 16; Score = 8585216.0 +<|begin_of_text|>Old Glory is almost certainly the most honored flag in the world. + +The late political scientist Samuel Huntington marveled at its place in our national life: We pledge allegiance to it. The national anthem celebrates it. An incredibly elaborate code stipulates how it’s displayed, handled and maintained. It even has its own holiday. + +“Since the Civil War,” Huntington wrote, “Americans have been a flag-oriented people. The Stars and Stripes has the status of a religious icon and is a more central symbol of national identity for Americans than their flags are for peoples of other nations.” + +The NFL players who kneel during the national anthem — a phenomenon that increased exponentially after President Trump colorfully demanded they stand — are disrespecting the most potent and enduring national symbol of the most patriotic nation on Earth. + +Not only are they wrong to do so, they aren’t delivering the devastating rebuke to Trump they may imagine. + +The power of national — or imperial — symbols isn’t new. The Romans couldn’t abide the collective disgrace of losing their standards to the enemy in battle, and would undertake decades-long campaigns to win them back. + +The American flag is layered with history and meaning. As Tim Marshall recounts in his book, “A Flag Worth Dying For,” its roots probably reach back prior to the country’s independence. It may owe its red and white stripes to the flag of the Sons of Liberty, the revolutionary agitators who carried out the Boston Tea Party. + +The “rebellious stripes” of that flag, lacking a field of stars, don’t look like much. In 1777, the Continental Congress added the stars. + +The flag took time to catch on. The Civil War, which tested the integrity of the flag, represented a watershed. It came to be known as Old Glory at this time, courtesy of a cussed merchant seaman named William Driver, who demonstrated a characteristic American attitude to the Stars and Stripes. + +Driver had retired to Nashville, Tenn., with his flag still in his possession. When local Confederates demanded he hand it over, he replied that they were welcome to take it — over his dead body. Secreted in a bed quilt, the flag was eventually handed over to Union forces, and the legend of Old Glory spread. + +After the war, Union veterans advocated for the display of the flag and for its veneration. The National Flag Conference of 1923 — yes, there was a national flag conference — set out the code subsequently adopted by Congress. + +Per the code, “the flag represents a living country and is +================================================================================ +Rank = 17; Score = 8454144.0 +<|begin_of_text|>Puzzle adventure Rime launches on Switch this week - today in North America, on Friday in the UK - but there are concerns surrounding its performance on Nintendo's handheld. + +Previews have shown the game to sometimes display a choppy frame-rate, and include long loading times. When I played the game back in September at EGX, both were noticeable (you can see footage of me playing the preview build below). + +Last night, in a reddit AMA, Rime developer Tequila Works spoke about the technical challenges it faced when bringing the game to Switch - and the compromises it decided to make. + +"Rime runs at 30fps in 720p throughout most of the game while docked," Tequila Works wrote. "This is a considerable improvement from where the game was at earlier this year when we announced the first delay. + +"With Rime being very open in many locations, it's incredibly difficult to get these level segments small enough to not cause a hiccup in performance. We were faced with the choice of adding loading screens throughout the stages, rebuilding the game completely to be more closed in (undermining the product vision in the process), or living with these small hiccups to preserve what the game was intended to be. We chose the latter. + +"When looking at the handheld mode, we had to make a choice between lowering the resolution, removing/replacing major parts of the level geometry, or having a bigger hit in performance. We decided to go for the former, because it allows us to maintain the integrity of the gameplay experience. All the important details are still very visible, and we've had no issues playing the game in handheld mode ourselves." + +Tequila Works concluded by noting Rime did have similar issues on other platforms, albeit less pronounced. + +But the fact Rime runs worst on Switch is particularly frustrating as a copy of Rime on Switch costs more than other platforms. On Amazon UK right now, a PS4 or Xbox One physical copy costs £25. On Switch, a physical copy from Amazon UK costs £32. + +It's less than the expected cost of the game - which was initially set as £39.99 back in March, before the Switch port was delayed. + +For digital copies, the prices are closer. On the Microsoft Store a digital download costs £23.99, while on both the PlayStation Store and Nintendo Switch eShop, it costs £29.99. + +Are you picking up Rime for Switch? Is the added portability of the game worth a lower performance? Let us know.<|end_of_text|><|end_of_text|> +================================================================================ +Rank = 18; Score = 8290304.0 +<|begin_of_text|>Version History + +1.8.5: + +1.9: + +1.9.5: + +1.9.5.5: + +1.9.6: + +1.9.7: + +1.9.8: + +1.9.9: + +1.9.9.1: + +2.0: + +2.1: + +2.1.1: + +2.2: + +1.0 Release1.1 Added the LAW and M79, the LAW will replace the RPG sounds1.2 Added the menu sounds as well as the perk level up sound, added the closed beta wave start sound as well some extra sounds on the trader menu and scoreboard, added the classic headshot sounds1.3 Fixed missing pump sounds on the trench gun, fixed missing sounds with the level action rifle1.4 Added the badass Z.E.D. gun to replace the Rail Gun and added the Flamethrower too1.4.5 Quick hotfix about that the flamethrower was missing, thats now fixed1.5 Improved the headshot sounds. Added some better bullet hits sounds. Improved the explosions sounds. Also the classic crawler might return too1.6 Added the thompson to replace the P90, added the classic katana sounds, Added the classic crawler1.7 Added the Pipe Bomb sounds to replace the C4, improved the headshot sound effects1.8 Fixed some missing sounds on the new Tacticool Winchester, Added the KSG sounds to replace the new shotgun, the HZ12- Updated the footstep sounds as well as added landing sounds when jumping (Just like in KF1, wow, i wish they added that to the base game)- Replaced the sounds of the stoner LMG (hope you guys like it)- Removed the xmas gorefast sounds (Because no one likes it)- Added classic's patriach minigun sound and rockets- Changed the melee hit sounds- Replaced the husk weapon sounds to its old ones- Changed the bullet impacts with the ones from KF1- Added the old darts sounds from the medic weapons- Added the old Seeker Six sounds- Replaced the Hemoglobin sounds- Replaced UMP-45 sounds- Fixed headshot sounds- Added the Mac-10- Added the Husk cannon- Added the M99- Added the Quad Shotgun- Fixed the M1911 and MB500 sounds- Replaced the AF2011- Replaced the FNFAL, MKB42 and HMTECH 501Enjoy everyone!- +================================================================================ +Rank = 19; Score = 8159232.0 +<|begin_of_text|>What is feminism? In this exclusive extract from A Book For Her, the standup comedian tells you everything you need to know about the movement that is the sole cause of the recession, global warming, terrorism, pandemics, cancelled flights, volcanos, delayed trains and pedantic health and safety regulations + +Facebook Twitter Pinterest Bridget Christie: ‘I even hate Ban Ki-moon.’ Photograph: David Levene/Guardian + +I am a feminist. All this means is that I am extremely hairy and hate all men, both as individuals and collectively, with no exceptions. Nope. Not even Laurence Llewelyn-Bowen/Paul Hollywood/Ronnie Corbett/Trevor McDonald/David Attenborough or John Nettles circa Bergerac are good enough for me. + +I even hate Ban Ki-moon. It’s one thing to try to eradicate female genital mutilation and forced marriage, but Mrs Ban Ki-moon told me she can’t remember the last time her husband put the Hoover round. Or sprayed his own soiled pants with pre-wash Vanish. Feminism begins at home, Mr Ban, not at the UN. Huh! Ban Ki-hypocrite, more like. + +Feminists don’t like humour, except slapstick. Charlie Chaplin, Harold Lloyd, Laurel and Hardy and Bottom are all very popular at feminist comedy nights. Anything that involves men being physically harmed always goes down very well with feminists. They also enjoy watching Tom and Jerry, Road Runner cartoons and war documentaries. + +Feminists never have sex and hate men opening doors for them, even into other dimensions. + +Christmas is banned in the “feminist community”, along with birthdays, wallpaper, nuance, giving people the benefit of the doubt and all music. Feminists only ever listen to one song, on a loop: kd lang’s Constant Craving. + +All feminists are lesbians. There is not a single heterosexual woman in the world who believes that women should have equal rights. Not one. If a feminist says she is heterosexual or bisexual or asexual, she is lying. They are all lesbians. + +Feminism is the sole cause of the recession, global warming, terrorism, pandemics, cancelled flights, volcanos, delayed trains and overly pedantic health and safety regulations. You can’t have hot drinks at work now because of feminism, or climb up small stepladders in libraries. You can’t eat a lobster without safety goggles now because of feminists. You can’t even open a door now because of +================================================================================ +Rank = 20; Score = 8060928.0 +<|begin_of_text|>ST. CATHARINES, ONT. + +A disgraced Niagara Regional Police officer has been fined $15,000 for his part in a scheme that involved smuggling copious quantities of cheese and chicken wings across across the border. + +Casey Langelaan, 49, resigned from the police service last year after he was caught up in a large-scale smuggling operation that saw thousands of dollars worth of U.S. cheese and chicken wings smuggled into Canada and sold to Ontario restaurants. + +"A man of otherwise good character has been embarrassed and shamed in his community," Judge David Harris said Friday after Langelaan pleaded guilty in a St. Catharines, Ont., courtroom to charges of evading compliance with the Canada Customs Act and possession of imported goods. + +Court heard the investigation began in January 2012 and involved Niagara police, U.S. Department of Homeland Security, and the Canada Border Services Agency. + +The investigation initially focused on the cross-border activities of another NRP officer, Const. Geoff Purdie, who is now serving time behind bars in an American jail for smuggling steroids into Canada. + +When investigators questioned Purdie, he offered up Langelaan, a sergeant at the time and making well over $100,000 a year, as a fellow smuggler. + +But instead of smuggling steroids, court heard, Langelaan’s preference was cheese and chicken wings. + +No one was interested in the poultry, court heard, but Langelaan, a resident of Fort Erie, sold the cheese to restaurants in Aylmer and Dorchester, Ont., at a profit of more than $50,000. + +Lawyer Paul O’Marra described his client as a "highly respected officer" whose reputation has taken a “major hit” because of the crimes. + +O’Marra said Langelaan has also suffered a blow to his pocketbook. + +Canada Borders Services Agency wants him to pay a penalty of $50,000, almost three times the amount of duties and taxes the former cop didn’t pay at the border. That matter is subject to civil litigation. + +His taxes were also re-assessed. + +Two other people were arrested as a result of the investigation that revealed more than $200,000 worth of cheese had been purchased in the U.S. and distributed in Canada. Const. Scott Herron and Bernie Pollino are scheduled to appear in court Dec. 12. + +According to the CBSA, travellers can bring back, duty free, $20 or 20 kilograms in total (whichever limit is reached first) of dairy +================================================================================ +Rank = 21; Score = 7700480.0 +<|begin_of_text|>Advertisement Here's what your next license will look -- and feel -- like Several new security features being added Share Shares Copy Link Copy + +Changes are coming to your wallet. Massachusetts is making several changes to the look and feel of driver's licenses.Starting on July 24, the Registry of Motor Vehicles will begin issuing licenses with new design and security features. Every license in the state will be upgraded to the new design within five years, as they expire.New features include raised lettering, similar to a credit card, laser perforations, ultraviolet ink and several secure background designs. Background images include the golden dome of the State House, the memorial to Robert Gould Shaw and the Massachusetts 54th Regiment, the state bird and the state flower."The new licenses and ID cards are among the most secure and technologically advanced in all of North America," the RMV said. The security upgrades will also be applied to identification cards, liquor identification cards, junior operator licenses and under 21 driver's licenses.Get the WCVB News App<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 22; Score = 7536640.0 +<|begin_of_text|>If Benjamin Franklin had had his way, America's national animal would be the turkey, an animal he called "a true original Native of America." + +And in his defense, there's certainly no shortage of turkeys to go around, which can't be said about a handful of countries around the globe with national animals that are decidedly less common — some even extinct. Other countries proudly boast emblematic critters that are straight-out bizarre, and, in some cases, mythical. + +From the dodo to the Komodo dragon to folkloric winged horses, we've assembled a motley menagerie of unusual and/or threatened national animals to consider. + +1. Unicorn (Scotland) + +Despite its mythical and occasionally sparkly depiction, the unicorn represents purity, strength and independence. (Photo: godam07/Shutterstock) + +It's a creature that's majestic, mythical and possesses the singing voice of Mia Farrow. It appears (chained) on the Royal Coat of Arms and is a symbol of purity, strength and independence. It's also often covered in glitter and found in close proximity to magical rainbows. How could anyone take issue with the national animal of Scotland being a unicorn? + +Shocking but true, a small but vocal group of Scots want to do away the unicorn, the heraldic symbol of Scotland that's also served as the country's national animal since the late 1300s. And what do these campaigners suggest take the unicorn’s place? They're pushing for another elusive beast to be named as Scotland's national animal, this one they believe to be real (at least for tourism purposes): the Loch Ness Monster. Their argument for the unicorn to be replaced by a lake-dwelling cryptid that's most likely a very large catfish? "How many people visit Scotland to look for unicorns? Exactly." + +2. Dodo (Mauritius) + +The extinct dodo lives on through museums, stamps and statues in Mauritius. (Image: Biodiversity Heritage Library/flickr) + +When you're a small island nation in the Indian Ocean and your most famous endemic bird is also the poster-animal for extinction, of course that bird would be your national animal. And although the dodo went, well, the way of itself circa 1662, the curious-looking flightless bird remains both a symbol of Mauritian pride and a potent reminder of the plight of endangered species across the globe threatened by human activity. + +Although the story of the dodo — a cautionary national animal if there ever was one — is tragic +================================================================================ +Rank = 23; Score = 7471104.0 +<|begin_of_text|>FEATURED ARTICLE + +Organizers of the former Montreal Anarchist Bookfair are celebrating their new identity after they renamed themselves The Bourgeois Feminist Bookfair. “We realized that we’re not an anarchist collective after all,” says activist and bourgeois neo-liberal feminist Lucy Descharnes. “We don’t care about anarchism, we don’t care about the working class, and we don’t care about economic issues. Our main interest is in protecting the privileges of wealthy university graduates, so we decided to change the name of our organization to better reflect our actual values. We’re bourgeois, we're educated, we're affluent, and we're proud of it."Anarchists across Montreal say that they’re not surprised by the name change. “I can’t remember the last time the anarchist bookfair actually catered to genuine anarchists,” says working class activist Jesse Hogan. “It’s been a giant bourgeois shit show for the last decade. Last year took the cake, though, after they gave white and affluent academic feminists the power to ban men from attending the event. Seriously, if a wealthy educated feminist didn’t like you, you couldn’t attend. All she had to say was that you made her feel uncomfortable. The organizers justified their actions because they believe that working class men have more power and privilege than bourgeois feminists. I’m not comfortable with bourgeois feminists being anywhere near me, but the bookfair doesn’t care about creating a safe space for the working class. Safe spaces only exist to protect the bourgeoisie from the rabble."Lucy agrees. “At the end of the day, a wealthy white woman with a Ph.D from Concordia is far more oppressed than a working class man who was born into poverty,” says Lucy. “At the Bourgeois Feminist Bookfair, we believe in intersectional feminism, which is the religious conviction that oppressions intersect in a way that minimizes and erases class privilege. For example, if you’re a poor homeless man, your penis gives you way more privilege than Martha Stewart or Michelle Obama. A homeless man’s male privilege intersects with his poverty, erasing it’s very existence from the face of the earth. Bourgeois feminists recognize that class is irrelevant — it’s the least important factor in oppression. We also believe in the one drop rule: if you have a single drop of non-economic privilege, that privilege erases the economic factors in your life. You’re white? Your class doesn’t matter. You're a man? Your class doesn’t matter. You’re straight? +================================================================================ +Rank = 24; Score = 7340032.0 +<|begin_of_text|>Somerset Bridge is a small bridge located in the western parish of Sandys, in Bermuda, connecting Somerset Island with the mainland. The stone bridge is walled solid, save for a narrow arch near one of its banks that allows small boats to pass underneath. There is enough headroom for a motorboat to pass through comfortably but not a sailboat. To allow a sailboat with a tall mast to squeeze through, a section of the motorway over the arch was replaced with wooden planks that could be raised creating a mini drawbridge. This section is just 32 inches wide, and is widely believed to be the smallest working drawbridge in the world. + +Photo credit + +The original bridge dates back to 1620, and although the bridge was largely rebuilt in the mid 20th century, much of the original stonework remains. The bridge was originally raised using a hand crank, but now consist of two cantilevered half-spans made of thick timber panels. The panels can be grabbed by the sides and lifted to create a narrow gap just wide enough for the mast of a sailboat to pass through. Judging by the size of the gap, I believe it takes a fair amount of skill to navigate the gap without mishap. Often a sailor has to be assisted in opening the drawbridge and making the passage safely, so it’s not uncommon to find sailboat captains calling upon pedestrians for help. + +Although not quite an attraction, the bridge’s unique reputation has earned it a place on the reverse of Bermuda’s $20 bills issued in 2009. + +Photo credit + +Photo credit + +Photo credit + +Photo credit + +Photo credit + +Photo credit + +Photo credit + +Sources: Bermuda4u / Wikipedia<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 25; Score = 7012352.0 +<|begin_of_text|>U.S. President Barack Obama handed the torch of progressive politics to Prime Minister Justin Trudeau Wednesday in a warm, rousing speech to Parliament, in which he also bluntly urged Canada to spend more on defence to meet its international obligations. + +He praised the extraordinary alliance and deep friendship between Canada and the United States. + +"We see ourselves in each other and our lives are richer for it," Obama said. + +"The enduring partnership between Canada and the United States is as strong as it has ever been and we are more closely aligned than ever before." + +The American president reserved heartfelt mention for Trudeau, who he said has brought "new energy and hope" to the cross-border relationship. + +"My time in office may be nearing an end, but I know that Canada and the world will benefit from your leadership in the years to come." + +But he also included pointed comments on a perennial sore spot, one that has taken fresh urgency in the face of a resurgent Russia and rising Islamic extremism. + +"As your NATO ally and your friend, let me say, we'll be more secure when every NATO member, including Canada, contributes its full share to our common security," Obama said to sustained applause from both Liberals and Conservative MPs and senators, who returned from summer break for the historic event. + +U.S. President Barack Obama is greeted by a standing ovation in the House of Commons in Ottawa on Wednesday. (Adrian Wyld/Canadian Press) + +"Because the Canadian Armed Forces are really good. And if I can borrow a phrase, the world needs more Canada. NATO needs more Canada. We need you." + +Washington has been ratcheting up the pressure on Ottawa to spend more for two years, ever since Russia annexed Crimea and set off a shadow war in eastern Ukraine. + +At the moment, Canada spends just under $20 billion on defence — or 1 per cent of its gross domestic product, which is half the NATO standard. + +Obama: 'The world needs more Canada' 1:21 + +The Trudeau government has also waffled on a call by NATO to contribute to a high-mobility brigade destined for eastern Europe as a deterrence measure. + +Although the federal government has privately expressed interest, it has yet to formally declare this publicly. + +Obama's remarks will put more pressure on the Liberal government for a public pronouncement. + +His appearance on the floor of the House of Commons, alongside Trudeau and his wife, Sophie Gregoire Trudeau,was greeted with a thunderous ovation. + +"It tempts me to shut up and leave. It can't get any better than this +================================================================================ +Rank = 26; Score = 7012352.0 +<|begin_of_text|>NASA’s next major space observatory is meant to tackle some of the biggest questions in astronomy when it launches in 2025—including what exoplanets look like and how dark energy is driving the Universe’s expansion. But the project’s cost is rising quickly, and NASA managers are struggling to keep its budget in check. + +The Wide-Field Infrared Survey Telescope (WFIRST) has grown in scope and complexity since it was proposed nearly a decade ago, and its price has swollen from US$1.6 billion in 2010 to the current estimate of $3.2 billion ($2.4 billion in 2010 dollars). That has raised concern at NASA, which in April commissioned a review by independent aero-space experts. Their report is due in the next few months. + +Above all, the agency wants to keep WFIRST from following the path of the James Webb Space Telescope (JWST), a successor to the Hubble telescope that is scheduled to launch in 2018. That project’s cost spiralled from $1 billion in the early 2000s to $8.8 billion—and nearly exhausted NASA’s astrophysics budget. + +The WFIRST review is meant to stave off that kind of meltdown. “This is a good time to take a look at the scale and scope of the mission,” says Jon Morse, a former head of NASA’s astrophysics division who is now chief executive of the BoldlyGo Institute, a non-profit space-exploration organization in New York City. “Nobody wants this thing to double in cost.” + +Bonus features + +WFIRST was the top-ranked big space mission in the 2010 decadal survey in astronomy and astrophysics, a list created by researchers to prioritize projects for the next ten years. Then the National Reconnaissance Office gave NASA a 2.4-metre mirror—replacing WFIRST’s planned 1.5-metre mirror—and the space agency started dreaming big. The larger mirror allowed NASA to add a corona-graph, an instrument that studies exoplanets by blocking light from the stars they orbit. + +And NASA made other design changes to go along with the big mirror. It also began to consider adding a ‘starshade’, a free-floating umbrella-like spacecraft that would fly alongside WFIRST and block enough light for the telescope to spy Earth-sized planets. + +WFIRST’s heart is a gigantic camera with 18 detectors, each capable of capturing a 16-mega-pixel shot—giving it a field of view 200 times Hubble +================================================================================ +Rank = 27; Score = 6979584.0 +<|begin_of_text|>The Peace Tower (in French: Tour de la Paix), also known as the Tower of Victory and Peace (in French: tour de Victoire et de Paix),[1] is a focal bell and clock tower sitting on the central axis of the Centre Block of the Canadian parliament buildings in Ottawa, Ontario. The present incarnation replaced the 55-metre (180 ft) Victoria Tower after the latter burned down in 1916, along with most of the Centre Block; only the Library of Parliament survived. It serves as a Canadian icon[2] and had been featured prominently on the Canadian twenty-dollar bill directly adjacent the queen's visage, until the change to polymer. + +Characteristics [ edit ] + +Designed by Jean Omer Marchand and John A. Pearson, the tower is a campanile whose height reaches 92.2 m (302 ft 6 in),[3] over which are arranged a multitude of stone carvings, including approximately 370 gargoyles, grotesques, and friezes, keeping with the Victorian High Gothic style of the rest of the parliamentary complex. The walls are of Nepean sandstone and the roof is of reinforced concrete covered with copper.[4] + +One of four grotesques at the corners of the Peace Tower + +At its base is a porte-cochere within four equilateral pointed arches, the north of which frames the main entrance of the Centre Block, and the jambs of the south adorned by the supporters of the Royal Arms of Canada. Near the apex, just below the steeply pitched roof, are the tower's 4.8 m (16 ft) diameter clock faces,[4] one on each of the four facades. The mechanical workings of the timepiece were manufactured by the Verdin Company and are set by the National Research Council Time Signal. One level below, running around the circumference of the tower's shaft, is an observation deck.[3] This was the highest accessible space in Ottawa until the early 1970s; the Peace Tower dominated the Ottawa skyline, as a strict 45.7 m (150 ft) height limit was placed on other buildings. That limit, however, was later rescinded, leading the Peace Tower to lose its distinction as the city's tallest structure. Cantilevered out at each of the four corners of the tower, at the level of the observation platform, are four 2.5 m (8 ft 4 in) long, 75 cm (2 ft 6 in) +================================================================================ +Rank = 28; Score = 6946816.0 +<|begin_of_text|>ELDERLY Australians committing welfare fraud on a massive scale are behind the extraordinarily high number of $100 notes in circulation, a former senior Reserve Bank official says. + +Yesterday the Herald revealed there are now 10 $100 notes in circulation for each Australian, far more than the more commonly seen $20 notes. + +Welfare fraud on a massive scale... elderly people wanting to get the pension are hiding their income in cash to ensure they qualify for the means-tested benefit, former Reserve official Peter Mair says. Credit:Jessica Shapiro + +One popular explanation is that they are used for illegal transactions as part of the cash economy, something the former Reserve official, Peter Mair, rejects as a "furphy". + +In a letter to the Reserve Bank governor, Glenn Stevens, dated July 4, Mr Mair laid the blame squarely on elderly people wanting to get the pension and hiding their income in cash to ensure they qualified for the means-tested benefit.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 29; Score = 6619136.0 +<|begin_of_text|>Electron Microscopy images of 1.9 and 3.3 Terabit/inch2 densities + +Hard drive storage continues to increase at a steady rate, but just how long can we keep squeezing more data on to those platters? TDK has managed to ensure that the density keeps rising for a few more generations with its HAMR head innovation, but that may have already been rendered redundant through another discovery by a Dr Joel Yang at the Institute of Materials Research and Engineering (IMRE). + +Existing hard drives max out at 3TB at the moment. TDK’s HAMR promises to take that to 6TB, but Dr Yang has found a way to increase the data density of a drive to 3.3 Terabit/inch2, effectively meaning 18TB hard drives are possible. + +What may surprise you is the key to this massive increase in data storage density. Dr Yang used table salt, or more specifically, sodium chloride. + +Data stored on a hard drive is done so using nanoscopic magnetic grains. Typically, several of these grains (a cluster), each measuring around 7nm, are used to store one bit of information. Therefore, if you can pack them more densely together, you get more storage in the same space. + +By adding sodium chloride to the mix, it has been possible to get a single 10nm grain storing one bit. Replacing several 7nm grains with one 10nm saves a lot of space and therefore ups the storage density considerably. + +The good news is the use of sodium chloride can be added to existing lithography processes, suggesting it won’t take too long to get this up and running for commercial drives. So far a 1.9 Terabit/inch2 storage solution has been shown to work, but that will increase to 3.3 Terabit/inch2, and then further research and development is aiming to achieve 10 Terabit/inch2 in the future. + +For us as consumers, it means the hard drive still has quite a few years of storage bumps still to come, and a greater challenge for SSDs trying to catch up with the amount of storage on offer. This also bodes well for data centers who need to cater for a never-ending storage need as more cloud services go online. + +Read more at the IMRE press release (PDF), via Physorg<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 30; Score = 6586368.0 +<|begin_of_text|>Mithost said: My intention from posting my original reply was not to immediately unban the controllers, but to get more information from those in charge on what the current barriers are. If you don't mind me asking, what are some examples of the logistical issues you mention? + +Outside of Smash, I am an active member of the wider fighting game community. At both local tournaments and large scale events, I have seen people play with everything from a PS1 gamepad with a converter to a pizza box with 20+ buttons on it without any issues. Controller configurations that prove to be problematic from a logistical or competitive point of view are banned when they prove to be problematic at an event (most wireless controllers, turbo buttons, and certain buggy platform converters). If a community so similar to our own in terms of logistics can run well with very lax controller rules, what is different about Smash that changes the logistics enough to warrant banning the controllers? Click to expand... + +The main difference between Melee and other FGs, as I'm sure you know, is the use of analog controls. Replacing left, right, up, and down with digital buttons is not a huge change for the average FG. For Melee, being able to hit precise angles is obviously a very powerful technique. Whether it's for perfect up-B angles, airdodges, wavedashes/wavelands, or DI, hitting specific angles is a skill that has been an inherent part of the game since its inception. If you have a controller that can do all of this for free, you're removing skill from the game and creating an unearned skill gap between the players who use a custom controller vs. the standard. Hax himself has admitted that his perfect wavedash angles were obnoxiously good, so he has attempted to mitigate this advantage by preventing perfect angles on the B0XX. There are plenty of other issues, such as multiple SDI inputs. Due to the nature of a button vs. a stick, you can easily double/triple tap fightstick buttons to get multiple SDIs. I'm not very experienced on a fightstick, but it still only took me a few minutes of testing to get 3 inputs in a 5 frame window. This kind of capability would result in people triple SDIing shines, and would greatly ruin the integrity of the game.All of these advantages that make custom controllers able to perform things that are normally difficult or impossible would have to be limited before these controllers could be allowed. That means TOs have to decide how good +================================================================================ +Rank = 31; Score = 6520832.0 +<|begin_of_text|>The elderly father of Metallica’s late bassist is making sure that his son’s positive legacy does not just fade to black. + +Cliff Burton, who was one of the founding members of Metallica, played bass guitar on the metal band’s first three albums until he was killed in a tour bus accident in 1986. + +The royalties from these albums have been going to Cliff’s 92-year-old father Ray—and the senior has only now revealed what he’s been spending that money on. + +CHECK OUT: Metallica Replaces Stolen Equipment of Tribute Band + +Ray has been donating all of the Metallica royalties towards musical scholarships for Castro Valley High School students in the San Francisco Bay area of California. + +The school has sentimental value to Ray since it was the school which Cliff attended when he was younger. + +“The kids who won it, they invariably write and thank me,” Ray told the Alphabettica podcast host last month. “And I think Cliff probably would have done that with his money, because he was not against education by any means. He liked it very much.” + +The Bell Tolls For You To Send This Story To Your Friends: Click To Share<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 32; Score = 6258688.0 +<|begin_of_text|>Cleveland, OH–The Cleveland Indians today announced pregame festivities for ALCS Game 1 at Progressive Field on Friday, with first pitch set for 8PM. + +The club is encouraging all fans to #RockRed Tribe gear throughout the ALCS to help maintain the Indians incredible home-field advantage against the Blue Jays. The Indians were 53-28 at Progressive Field this season, second-best in baseball, and won two straight against Boston last weekend. + +Postseason coverage on SportsTime Ohio starting with Drennan Live at 6 p.m. Friday and then Indians Live postgame following the final pitch + +Indians Hall of Famer Andre “Thunder” Thornton will throw Friday’s ceremonial first pitch. Thornton hit 214 homers over 10 seasons with the Tribe. + +Gates will open at 6PM, and fans are encouraged to be in their seats by 7:30PM at the latest to enjoy all pregame festivities. + +Each fan will receive a red #RallyTogether towel upon entering Progressive Field. + +Pregame festivities include Tom Hamilton as emcee, Budweiser activation, Cleveland students manning American flag on field during anthem + +The Indians will have a number of festivities for fans prior to Thursday’s game. + +Other festivities include: + +· Tom Hamilton will emcee on-field pregame festivities + +· Veteran local singer Danielle Danburg will sing the American and Canadian national anthems; Dan Polk will sing “God Bless America” during the seventh inning + +· Cleveland Metropolitan School District students and staff will volunteer to man a giant American flag on field for the anthem + +· Color guards from the Army, Navy, Marines, Air Force and Coast Guard will present the colors + +· Budweiser will bring its ‘Bud and Burgers’ trailer to Gateway Plaza, where fans over the age of 21 can enjoy Budweiser products up until game time; Swenson’s Food Truck also will be on hand + +· The “Mike Trivisonno Show” will broadcast live from the main concourse (near The Corner) from 3-7:30PM + +· WMJI will broadcast live from Gateway Plaza from 5-8PM + +· A band will perform live on Gateway Plaza from 5-7:30PM + +· DJ Kyro among activities available in First Energy Heritage Plaza (center field) pregame + +(Cleveland Indians press release)<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 33; Score = 6127616.0 +<|begin_of_text|>Worse than SOPA, CISPA will allow monitoring, censorship, alteration of ANY online communication + +The assaults keep coming like the Terminator, don't they? This one is bipartisan, too. Comes to you from Rep. Mike Rogers (R-MI) and Rep. C.A. �Dutch� Ruppersberger (D-MD). Strongly recommend reading the entire article. + +____________________________________________ + +http://www.digitaltrends.com/web/watch-out-washington-cispa-replaces-sopa-as-internets-enemy-no-1/ + +Watch out, Washington: CISPA replaces SOPA as Internet�s Enemy No. 1 + +By Andrew Couts + +Digital Trends + +....Unveiled to the House by Rep. Mike Rogers (R-MI) and Rep. C.A. �Dutch� Ruppersberger (D-MD) late last year, CISPA is described as a �cybersecurity� bill. It proposes to amend the National Security Act of 1947 to allow for greater sharing of �cyber threat intelligence� between the U.S. government and the private sector, or between private companies. The bill defines �cyber threat intelligence� as any information pertaining to vulnerabilities of, or threats to, networks or systems owned and operated by the U.S. government, or U.S. companies; or efforts to �degrade, disrupt, or destroy� such systems or networks; or the theft or �misappropriation� of any private or government information, including intellectual property. + +.... + +The Electronic Frontier Foundation (EFF) adds that CISPA�s definition of �cybersecurity� is so broad that �it leaves the door open to censor any speech that a company believes would �degrade the network.�� Moreover, the inclusion of �intellectual property� means that companies and the government would have �new powers to monitor and censor communications for copyright infringement.� + +Furthermore, critics warn that CISPA gives private companies the ability to collect and share information about their customers or users with immunity � meaning we cannot sue them for doing so, and they cannot be charged with any crimes. According to the EFF, CISPA �effectively creates a �cybersecurity� exemption to all existing laws.� + +�There are almost no restrictions on what can be collected and how it can be used, provided a company can claim it was motivated by �cybersecurity purposes,�� the EFF continues. �That means a company like Google, Facebook, Twitter, or AT&T could intercept your emails and text +================================================================================ +Rank = 34; Score = 6094848.0 +<|begin_of_text|>Prime Minister Narendra Modi has ‘replaced’ Mahatma Gandhi in the 2017 wall calendar and table diary published by the Khadi Village Industries Commission (KVIC), official sources said on Thursday, as opposition leaders reacted sharply to the move. + +The cover photo of the calendar and diary shows Modi weaving khadi on a large ‘charkha’, in the same classic pose as Gandhiji. In response, the employees of KVIC at its Vile Parle headquarters plan to stage “a silent, soul-cleansing” protest wearing black bands on their mouths, during lunch hour on Thursday. + +Political leaders took a jibe at the Prime Minister for replacing the iconic image of Gandhi, including Delhi chief minister Arvind Kejriwal, Congress vice-president Rahul Gandhi and former AAP member Yogendra Yadav, tweeting about it. + +“Becoming Gandhi requires years of austerities. One cannot become Gandhi by acting to spin the Charkha, this would only ridicule them,” Kejriwal said in a tweet. + +Gandhiji would've welcomed it, also opposed his photo on notes. + +We should welcome it too. Less hypocrisy.https://t.co/q1fVI2RYmA — Yogendra Yadav (@_YogendraYadav) January 12, 2017 + +Rahul too called Modi’s move “the mangalyaan effect”, referring to India’s Mars Orbiter Mission. + +The Mangalyaan effecthttps://t.co/NnkbJ4i7vx — Office of RG (@OfficeOfRG) January 13, 2017 + +While Gandhi’s historic picture weaving khadi on a simple charkha, wearing his trademark loin cloth, is legendary and imprinted in the minds of the masses since generations, Modi comes across in his signature attire of kurta-pyjama-waistcoat, weaving khadi on a slightly modern charkha. + +KVIC chairman Vinai Kumar Saxena said this was “not unusual” and there have been deviations in the past. + +“The entire khadi industry (udyog) is based on Gandhiji’s philosophy, ideas and ideals, he is the soul of KVIC, so there is no question of ignoring him,” Saxena told IANS. + +KVIC calendar & dairy without Gandhiji's pics have been issued earlier too.Spirit is more important than the image.@PMOIndia,@ashokepandit. pic.twitter.com/57ksjFD1LQ — Chairman KVIC (@ +================================================================================ +Rank = 35; Score = 5996544.0 +<|begin_of_text|>Where transgenders decide to sprinkle their tinkle is NOBODY’S BUSINESS! Except it’s everyone’s business. Including that of white-knightly tech giants like Apple and Ebay. You may recall not long ago Trump reversed Obama’s decision which forced states to allow trannies into certain bathrooms. Our current President didn’t outlaw those transgender bathroom privileges, but simply restored the states’ power to discern for themselves. Therein lies the outrage. + +Apple and other companies have decided to take a stand… By getting involved in a lawsuit filed by a trans student suing his school for boys’ bathroom privileges. + +Something to note here? People were uncomfortable sharing the restroom with Gavin Grimm, the girl-boy, so the school provided a unisex restroom to him in an attempt to make everyone happy. The boys wouldn’t have to bare themselves before a lady-man, and the trans student wouldn’t have to share a restroom with girls. But it wasn’t enough for the amorphous slug. Of course. + +Now he’s demanding access to the boys’ rooms via lawsuit. Entitled little toad. This is the case that Apple and friends are jumping in to support… + +A number of leading tech firms plan to file a brief in favor of transgender rights in a case due to be heard next month in the Supreme Court. Apple has been among those leading the charge on the effort, along with… Affirm, Box, Ebay, GitHub, IBM, Microsoft, PayPal, Salesforce, Slack, Tumblr, Yelp. Apple has already spoken out on the Trump Administration’s move to pull back on Obama-era guidance that interpreted existing law to require schools to let students use the restroom that matches their gender identity. The Supreme Court case, involving Virginia high school student Gavin Grimm, asks the court to weigh in on the same question. + +These companies really, really care about rights. Just not the rights of the states. Or those who want to potty in peace. It seems as though Apple misses not a single opportunity to get hyper-political (see Apple Goes Full PC, Replaces ‘Offensive’ Gun Emoji With Squirt Gun… and Far-Left Apple CEO to Wage War on Fake News. Misses HUGE Irony…). We’re finding that tech companies are becoming more and more politically involved. Why stick with creating better tech products when you can wade into potty politics? Maybe there’s an app for that. + +Look, this is America, where businesses can say and do what they please. The problem here? These companies are alienating a large portion of their +================================================================================ +Rank = 36; Score = 5996544.0 +<|begin_of_text|>Team Android at Basecamp recently passed a fairly big milestone — over 25% of the Basecamp 3 Android app code base now runs on Kotlin! 🎉 + +We’ve found that Kotlin not only makes our code much better, but massively increases programmer happiness. All of this ensures we’re making the best app we can for the tens of thousands of Android users we support. + +Given our new experiences with the language, I thought it’d be worth sharing some specifics that make the language so wonderful to work with. + +Unlike most articles that introduce you to a language, I’m going to avoid using too much programming lingo. Instead, I’ll try using plain English in the hopes that it’s more accessible to beginners. 🤗 + +Some notes about the code examples: + +I am by no stretch an expert in Kotlin. Read, consider, and discuss! + +They look better on a desktop browser. You can get by on the mobile app in landscape mode, but I’d recommend breaking out your laptop to read them. + +You can get by on the mobile app in landscape mode, but I’d recommend breaking out your laptop to read them. They’re brief and simple on purpose. Long-winded examples tend to cause confusion. Take these simple examples and extrapolate them into your own potential uses, and you’ll see a lot more power. + +Let’s get started with seven of my current favorites! + +1. Replacing simple if/else if/else blocks with when + +One of my absolute favorites. + +// Java + +if (firstName.equals("Dan")) { + +person.setTeam(programmers); + +} else if (lastName.equals("Dihiansan")) { + +person.setTeam(designers); + +} else { + +person.setTeam(others); + +} + +// Kotlin + +when { + +firstName == "Dan" -> person.team = programmers + +lastName == "Dihiansan" -> person.team = designers + +else -> person.team = others + +} + +when blocks are effectively the same as a simple if block, but look how much more readable that is! + +There’s a similar convention when only one argument is being checked. Typically this would be a long, ugly switch/case statement in Java. + +// Java + +switch (firstName) { + +case "Dan": person.setTeam(programmers) + +break; + +case "Jay": person.setTeam(programmers) + +break; + +case "Jamie": person.setTeam(designers) + +break; + +default: + +person.setTeam(others) + +} + +// Kotlin + +when (firstName) { + +"Dan", "Jay" -> person.team = programmers + +"Jamie +================================================================================ +Rank = 37; Score = 5963776.0 +<|begin_of_text|>Seven former Illinois women’s basketball players are suing the university, claiming that head coach Matt Bollant and former assistant coach Mike Divilbiss violated Title VI by creating a racially hostile environment. + +The players are suing for $10 million in compensatory and punitive damages as well as attorney’s fees. + +The specific claims—filed by Amarah Coleman, Alexis Smith, Taylor Tuck, Nia Oden, Sarah Livingston, Taylor Gleason, and Jacqui Gran—against Bollant are broken down into 13 specific points in the lawsuit: + +a. Instituted segregated practices singling out the black Plaintiffs,Alexis Smith, Nia Oden and Taylor Tuck, as “crabs” because one crab can climb out of a bucket, but more than one pull eachother down in and none can get out; b. Referred to the black Plaintiffs, Alexis Smith, Nia Oden andTaylor Tuck, as “toxic.” c. Referred to these segregated practices as “the dog pound,” indirectly labeling the black Plaintiffs, Alexis Smith, Nia Odenand Taylor Tuck, dogs; d. Appointed a white player captain without a team vote and even though the majority of players at the time were black; e. Relegated and/or threatened relegation of the white Plaintiffs, Taylor Gleason and/or Jacqui Grant, to the “dog pound” in retaliation for these Plaintiffs’ continued association with and support of the black Plaintiffs against the racially hostile environment created by the Defendants, and labeled the whitePlaintiff in the segregated practices as a “mascot;” f. Instituted segregated travel room assignments, prohibiting white players from rooming with black players, including among the white and black Plaintiffs; g. During team preparations for games against other school swhose teams were predominantly made of black players, singled out the black Plaintiffs and other black teammates to ask what the players on the other team were thinking or going to do,implying that blacks think differently than whites; h. During team preparations for games against other schools whose teams were predominantly made of white players, and in the context of subpar. (g), above, did not ask the black players what the other team’s white players were thinking or going to do; i. During team preparations for games, referred to opposing teams predominantly made of black players as undisciplined and unintelligent, and referred to opposing teams predominantly made of white players as disciplined and intelligent; j. +================================================================================ +Rank = 38; Score = 5865472.0 +<|begin_of_text|>News that MPs will get the chance to vote on the final Brexit deal has been welcomed by Labour who plan to use the opportunity to continue doing f**k all about Britain leaving the EU. + +Leader Jeremy Corbyn immediately responded warmly to the announcement + +“This vote is a victory for democracy and for the people of this country,” he said. + +“And rest assured we will use it to continue our long-standing policy of doing f**k all about Brexit.” + +Labour have been committed to doing f**k all about Britain leaving the EU since the referendum was announced by David Cameron, whoever he is, some eighteen months ago. + +Immediately after the referendum announcement, Jeremy Corbyn swung into action by doing f**k all. + +When campaigning started in earnest Labour remained constant in doing f**k all. + +Not wishing to let the momentum drop, at the height of the campaign Shadow Chancellor John McDonnell did f**k all. + +Finally, when the referendum result was announced, Jeremy Corbyn did his best to rally shellshocked remainers by doing f**k all. + +Every day since, Labour has remained proudly committed to doing f**k all about Brexit and, with this new announcement, look set to continue doing f**k all until the country is completely isolated and vilified, and is ultimately reduced to being little more than a nice place for Donald Trump to park his helicopter.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 39; Score = 5799936.0 +<|begin_of_text|>Canadian Tire moving Clayton Park store to former Target in Bayers Lake. Canadian Tire also opened their new store in Elmsdale this week. + +I’m slow on the draw, this timebut the Burrito Jax in Clayton Park is open, just on the side of Sobeys. + +East Coast Lifestyle to open a Halifax storefront, location TBD + +The Home Hardware in Enfield is closed after 60 years + +In Costco Cafe news Smoked Meat Sandwiches going on Monday. Replacing with Chk Caesar Salad. Will sell meat till stock gone + +Urban Outfitters is having their Grand Opening on May 15, paper is off the windows and stocking the shelves is in progress. + +Tom’s Havana confirmed that they are indeed moving October 1, but new location not finalized yet. Sleep Country on Spring Garden is moving with BMO to the new building where Winsby’s used to be. + +Zions Gate, in Spryfield has open a liquidation centre at the back of South Centre Mall (by Bowlarama) + +Ottoman Cafe and Mid Point Coffee on Spring Garden are closing up shop and heading to Toronto. + +The Jessy’s Pizza in Kingswood is now open. + +The Holiday Inn Express, Kearney Lake is now Chateau Bedford<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 40; Score = 5570560.0 +<|begin_of_text|>WASHINGTON–John McCain is the latest high-profile politician to repeat the diehard American falsehood that the 9/11 terrorists entered the United States through Canada. Just days after Janet Napolitano, the U.S. homeland security secretary, sparked a diplomatic kerfuffle by suggesting the terrorists took a Canadian route to the U.S. eight years ago, McCain defended her by saying that, in fact, the former Arizona governor was correct. "Well, some of the 9/11 hijackers did come through Canada, as you know," McCain, last year's Republican presidential candidate, said on Fox News on Friday. The Arizona senator's remarks prompted the Canadian embassy to immediately reissue remarks made earlier this week by Ambassador Michael Wilson, who reminded Americans once again that no 9/11 perpetrators came to the U.S. via Canada. "Unfortunately, misconceptions arise on something as fundamental as where the 9/11 terrorists came from," Wilson said. + +Article Continued Below + +"As the 9/11 Commission reported in July 2004, all of the 9/11 terrorists arrived in the U.S. from outside North America. They flew to major U.S. airports. They entered the U.S. with documents issued to them by the U.S. government. No 9/11 terrorists came from Canada." Crestfallen embassy officials contacted McCain's office soon after his Fox News remarks to set the record straight. McCain, an avid supporter of NAFTA and a powerful friend to Canada on Capitol Hill, recently visited the Canadian Embassy and had lunch with Wilson. The normally reserved Wilson made his 9/11 remarks on Tuesday following a CBC interview in which Napolitano appeared to believe that the hijackers entered the U.S. from Canada. + +She later said she had misunderstood a question asked during the interview and was well aware there had been no Canadian 9/11 connection, but added that the Canada-U.S. border had, in the past, posed a security risk to Americans. The next day, Napolitano appeared at a border conference and suggested Canada was more lax in its immigration policies than the U.S., alleging Canadian authorities allow people into the country that would not pass muster south of the border. Napolitano has also ruffled diplomatic feathers with her insistence that the Canadian border must not be treated any differently than the U.S.-Mexican boundary, where a drug war rages and countless illegal immigrants flood into America every year. McCain expressed some sympathy for Canada on that front on Friday. + +Article Continued +================================================================================ +Rank = 41; Score = 5570560.0 +<|begin_of_text|>WASHINGTON—According to a groundbreaking new study published Monday in the Journal Of The American Medical Association, it is impossible to lose weight, no one has ever lost a single pound of fat through diet or exercise, and those attempting to do so are advised to abandon their weight loss goals immediately. + +The study, conducted by scientists at The National Weight Control Registry, determined conclusively that shedding excess weight has never occurred, changing your appearance is impossible, and that it actually feels “pretty nice” to just give up and realize that you’re powerless to alter your body mass index in any way, shape, or form. + +Advertisement + +“Our findings indicate that if you’re trying to lose weight, you will fail—and that’s because you can’t, no one has, and you need to stop trying because it will never happen,” said Dr. Rena Wing, lead author of the report. “You could work out every day and eat nothing, and you still wouldn’t lose an ounce. And the sooner you throw up your hands and make peace with that fact, the better off you’ll be.” + +“Our data demonstrates that you’re doomed to look the way you look,” Wing added. “You are stuck in your body and it’s never going to look any better than it does right now. In fact, it will only get worse.” + +According to the study, running on a treadmill every morning at 6 a.m. will not help anyone lose weight, and neither will cutting carbohydrates from one’s diet, eating smaller portions throughout the day, doing yoga, or hiring a personal trainer. In addition, the study went on to state that everyone who tried the previous methods have only ended up feeling disappointed and in some cases heartbroken, especially after individuals step onto a scale and realize that, after two weeks of healthy eating and nonstop exercise, they’ve actually gained two or more pounds. + +Advertisement + +The report added that skinny people stay skinny, overweight people remain overweight, that’s how it has always been, that’s how it will always be, and every study, book, or news article you’ve ever read that has told you otherwise has been totally and completely wrong. + +“Going to the gym is a waste of time if you are doing it to lose weight,” said University of California physiologist Dr. Andrew Novak, adding that time working out would be better spent reading, spending time with your family, sleeping, watching television, or eating. “Now, if you are exercising to prevent heart disease or cancer, well, new evidence has come to light that +================================================================================ +Rank = 42; Score = 5537792.0 +<|begin_of_text|>The combat code of the US Military is that we don’t abandon our dead or wounded on the battlefield. In US Air Force lingo, fighter pilots don’t run off and leave their wingmen. If one of our own is shot down, still alive and not yet in enemy captivity, we will either come to get him or die trying. Among America’s fighting forces, the calm, sure knowledge that such an irrevocable bond exists is priceless. Along with individual faith and personal grit, it is a sacred trust that has often sustained hope in the face of terribly long odds. + +The disgraceful abandonment of our Ambassador and those brave ex-SEALs who fought to their deaths to save others in that compound is nothing short of dereliction-of-duty. Additionally, the patently absurd cover-up scenario that was fabricated in the aftermath was an outright lie in attempt to shield the President and the Secretary of State from responsibility. + +It has been over eight months since the attack on our compound in Benghazi. The White House strategy, with the aid of a “lap dog press” has been to run out the clock before the truth is forthcoming. The recent testimonies of the three “whistle blowers” have reopened the subject and hopefully will lead to exposure and disgrace of those responsible for this embarrassing debacle. + +It would appear that the most recent firewall which the Administration is counting on is the contention that there were simply no military assets that could be brought to bear in time to make a difference… mainly due to the unavailability of tanker support for fighter aircraft. This is simply BS, regardless how many supposed “experts” the Administration trot out to make such an assertion. The bottom line is that even if the closest asset capable of response was half-way around the world, you don’t just sit on your penguin *** and do nothing. The fact is that the closest asset was not half-way around the world, but as near as Aviano Air Base, Italy where two squadrons of F-16Cs are based. + +Consider the following scenario (all times Benghazi local): + +When Hicks in Tripoli receives a call at 9:40 PM from Ambassador Stevens informing him “Greg, we are under attack!” (his last words), he immediately notifies all agencies and prepares for the immediate initiation of an existing “Emergency Response Plan.” At AFRICON, General Carter Ham attempts to mount a rescue effort, but is told to “stand down.” By 10:30 PM an unarmed drone is overhead the compound and streaming live feed to various Command +================================================================================ +Rank = 43; Score = 5308416.0 +<|begin_of_text|>The 42nd Mayor of the City of Los Angeles, Eric Garcetti, is the first News Conference guest in the new NBC4 studio. Does the mayor agree on Police Chief Beck's "out of policy" decision not to punish eight officers in the Dorner manhunt? NBC4's Conan Nolan talks to the mayor about that and a multiple of other issues including raising the wage to $15 in this three-part interview. (Published Saturday, Feb. 8, 2014) + +Los Angeles Mayor Eric Garcetti says he "expected something stronger" from his police chief in disciplining eight Los Angeles police officers in the shooting of two women a year ago in Torrance. + +While the shooting was determined to be in "violation of policy," Police Chief Charlie Beck has indicated the officers will be returning to street patrols and will be subject to additional training. + +On NBC4's "News Conference" program Sunday morning, Garcetti said he disagreed with the decision. + +"I think I would have respectfully said it should have been something a little more harsh...we don't want to get rid of police officers who have a good career but we do need to make sure this isn't repeated again," Garcetti said. + +NewsConference: Los Angeles Mayor Garcetti on his first 7 months + +NBC4's Conan Nolan talks with L.A. Mayor Garcetti about the recent criticism of the mayor's first seven months in office. (Published Saturday, Feb. 8, 2014) + +The victims, a mother and her daughter, were shot while delivering newspapers in the early morning hours of February 7, 2013. Officers believed their target was former LAPD officer Christopher Dorner, who murdered three people during a four day crime spree in Southern California. + +Despite the disagreement over disciplining the officers, Garcetti voiced support for Beck. + +"My managers I trust and they are going to make different decisions at different points... he's been a super Chief," Garcetti said. + +NewsConference: L.A. Mayor Eric Garcetti & High Tech + +NBC4's Conan Nolan talks with Los Angeles Mayor Eric Garcetti on why the city needs a high tech czar. (Published Saturday, Feb. 8, 2014) + +On other issues Garcetti defended his efforts at getting rid of the city's business tax, saying it will make the city more competitive in attracting new employers. + +Garcetti also indicated that he would stay neutral in the campaign to replace Congressman Henry Waxman and that he might endorse a ballot measure that would allow for the recreational +================================================================================ +Rank = 44; Score = 5308416.0 +<|begin_of_text|>USA food safety lags far behind the rest of the world + +Taiwan bans GMOs from school menus + +Entire world revolting against U.S. agricultural imperialism and fascist, corporate-dictated government policy + +(NaturalNews) Even as the fascist, corrupt U.S. government and its regulators (FDA and USDA) actively conspire with the biotech industry to poison Americans with genetically modified foods, Taiwan has already passed and implemented a nationwide law to protect its citizens from GMOs.Nearly a full year ago, Taiwan passedthat achieve remarkable food safety milestones the U.S. government refuses to implement, placing Taiwan far ahead of the United States on food safety. These milestones include:1) Requiring the mandatory labeling of GMOs on all food products that contain 3 percent or more GMOs. Foods that use no GMOs may be labeled "non-GMO"... and many already are, causing their sales to skyrocket across Taiwan. Just last year, imports of non-GMO soybeans to Taiwan grew nearly 300% to 58,000 tons.2) Limiting the use of food additives to just 799 compounds approved by the Taiwan FDA. The FDA of the United States, by comparison, allows tens of thousands of chemicals to be used as additives, even when they are well known to cause cancer.3) All GMO ingredients are required to be registered with the Taiwan government, and food manufacturers that use GMOs are required to establish an origins tracking system to identify where those GMOs originated.4) All the soy milk, tofu, miso and other soy-derived products sold everywhere across the country -- including at cafes and street food vendors -- must be clearly labeled as GMOs if they use genetically modified soy.5) Food products made using genetically modified soy as a processing agent or blended ingredient must also label their final food products as GMO, even if the soybean oil is not, itself, the final product.6) Fines for violating these food safety provisions have been set at NT$50 million.This shows yet again just how far behind the United States is on food safety compared to the rest of the world. Instead of promoting actual food safety, the FDA gives a free pass to GMOs, heavy metals, artificial additives and other toxic chemicals, focusing almost exclusively on bacteriological contamination issues such as e.coli and salmonella.That's why the USA lags far behind the rest of the world on the banning of artificial additives and preservative chemicals, most of which have already been banned across the EU. The +================================================================================ +Rank = 45; Score = 5242880.0 +<|begin_of_text|>Whitson Becomes World's Oldest Female Spacewalker, as EVA-38 Replaces Aging Space Station Batteries + +Two U.S. astronauts, including the oldest woman ever to participate in an Extravehicular Activity (EVA), triumphantly concluded a six-hour and 32-minute spacewalk outside the International Space Station (ISS) earlier today. Expedition 50 Commander Shane Kimbrough and Flight Engineer Peggy Whitson breezed through the task of installing three new adapter plates and hooked up electrical connectors for a trio of new lithium-ion (Li-Ion) batteries on the starboard-side S-4 segment of the station’s Integrated Truss Structure (ITS). Their work heralds the start of a two-year campaign of EVAs and robotics, which will see 48 aging nickel-hydrogen batteries on the port and starboard trusses replaced with 24 smaller, but higher-performing, lithium-ion ones. Working around an hour ahead of the timeline, Kimbrough and Whitson were also able to complete several “get-ahead” tasks, including a photographic survey of the station’s Alpha Magnetic Spectrometer (AMS-2). + +Today’s spacewalk was designated “U.S. EVA-38” and represented the 38th spacewalk to be conducted out of the Quest airlock, with astronauts clad in U.S.-built Extravehicular Mobility Units (EMUs) and conducted in the absence of the space shuttle. Since U.S. EVA-1, way back in February 2002, these excursions have outfitted new hardware, removed and replaced faulty equipment, reconfigured coolant loops, tackled ammonia leaks, and laid cables in readiness for Commercial Crew operations and the arrival or relocation of pressurized modules. Last summer, Expedition 48 spacewalkers Jeff Williams and Kate Rubins performed EVAs 36 and 37 to install an International Docking Adapter (IDA-2) onto the Harmony node and retract the Trailing Thermal Control Radiator (TTCR) on the station’s port-side P-6 truss. + +These U.S. Orbital Segment (USOS)-based EVAs have been performed not only by NASA astronauts, but have also involved German, Russian, Japanese, Italian, and British nationals. They have seen a number of records secured, including Sunita Williams becoming the most experienced female spacewalker and holder of the greatest number of EVAs—seven—ever conducted by a woman. Yet during today’s EVA-38, one of Williams’ achievements was +================================================================================ +Rank = 46; Score = 5144576.0 +<|begin_of_text|>Following Liam Reddy's Herculean performance in goals against Melbourne City on Tuesday www.perthglory.com.au have composed this cheeky list of things they believe their goalkeeper could save. + +The Glory shot-stopper was in vintage form between the posts at AAMI Park, saving two penalties and seven shots to help his side secure a 3-3 draw in a match that was heralded as one of the season's best. + +City, Glory share spoils in Hyundai A-League classic + +The performance rightfully earned Reddy player-of-the-match honours and Glory have taken their appreciation to the next level with this tribute to their No.1. + +1. Mufasa in The Lion King + +The worst part of everyones childhood! If only our big #33 could've been there for Mufasa. + +2. Lizard in the valley of snakes + +In one of the viral videos of the year, this scaley little fella makes an epic dash out of trouble in a valley of venomous snakes. Maybe Liam could've plucked him out of there earlier! + +3. The world famous Kangaroo punch + +A safe hand could've saved our national animal from a little knock on the face! + +4. Jack & Rose on the Titanic + +There was definitely room on that door Rose! + +5. Ned Stark in Game of Thrones + +Everyone's favourite king, Reddy could've come to the rescue in Season 1 of Game of Thrones! + +6. Kevin Muscat from John Kosmina + +One of the iconic images of the Hyundai A-League. + +7. Hand of God in the 1986 FIFA World Cup + +Would the ref work it out this time though? + +8. Steph Curry from downtown + +Could Reddy do it from defence against Golden State? + +9. That Leo Barry mark + +A moment to forget for many West Coast Eagles fans, just needed one palm up there to save the day! + +10. Shannon Noll on Australian Idol in 2003 + +The infamous Australian Idol series that launched enough memes to revive Shannon Noll's career! Would have Nollsy won if Reddy came to the rescue? + +11. Michael Turnbull on the Bachelorette + +He came third in the hunt for Sam Frost's heart, could have Liam got Micky T over the line? #GoalkeepersUnion + +12. Usain Bolt on the home stretch + +The safest hand to pass the baton on the home stretch. + +13. House of Cards' Frank Underwood as Secretary of State + +Cheated from a big position in the +================================================================================ +Rank = 47; Score = 4980736.0 +<|begin_of_text|>(a) The general principles and factors that the Food and Drug Administration (FDA) considered in arriving at the reference amounts customarily consumed per eating occasion (reference amounts) which are set forth in paragraph (b) of this section, are that: (1) FDA calculated the reference amounts for persons 4 years of age or older to reflect the amount of food customarily consumed per eating occasion by persons in this population group. These reference amounts are based on data set forth in appropriate national food consumption surveys. (2) FDA calculated the reference amounts for an infant or child under 4 years of age to reflect the amount of food customarily consumed per eating occasion by infants up to 12 months of age or by children 1 through 3 years of age, respectively. These reference amounts are based on data set forth in appropriate national food consumption surveys. Such reference amounts are to be used only when the food is specially formulated or processed for use by an infant or by a child under 4 years of age. (3) An appropriate national food consumption survey includes a large sample size representative of the demographic and socioeconomic characteristics of the relevant population group and must be based on consumption data under actual conditions of use. (4) To determine the amount of food customarily consumed per eating occasion, FDA considered the mean, median, and mode of the consumed amount per eating occasion. (5) When survey data were insufficient, FDA took various other sources of information on serving sizes of food into consideration. These other sources of information included: (i) Serving sizes used in dietary guidance recommendations or recommended by other authoritative systems or organizations; (ii) Serving sizes recommended in comments; (iii) Serving sizes used by manufacturers and grocers; and (iv) Serving sizes used by other countries. (6) Because they reflect the amount customarily consumed, the reference amount and, in turn, the serving size declared on the product label are based on only the edible portion of food, and not bone, seed, shell, or other inedible components. (7) The reference amount is based on the major intended use of the food (e.g., milk as a beverage and not as an addition to cereal). (8) The reference amounts for products that are consumed as an ingredient of other foods, but that may also be consumed in the form in which they are purchased (e.g., butter), are based on use in the form purchased. (9) FDA sought to ensure that foods that have similar dietary usage, product characteristics, and customarily consumed +================================================================================ +Rank = 48; Score = 4980736.0 +<|begin_of_text|>Silver Physical Supplies and Price Action + +The basic tenets of economics tell us that supply, demand, and price as all intertwined, especially in free markets. While the silver market may be not as free as we would like, compared to most, it is relatively free, and it should be very much subject to basic economic laws. + +Last week, I called a number of silver buyers, sellers and coin shops to get an idea of how the market was responding to the massive run up in silver prices. Needless to say, the results weren't exactly as I had expected. I thought that during the run up, shops would report that some of their regular investors had been taking profits. This wasn't the case. + +The Survey + +In what cannot at all be seen as an entirely representative sample, a total of 15 coin shops were phoned and surveyed about their total sales. All but three had seen an increase in sales of junk jewelry metals, as people were cashing out for a variety of reasons, most of which are probably preparing for the holidays. One said he had seen a decrease in junk sales, but admitted he had no hard numbers on which to base his claim. + +The remaining two declined to comment. One suggested that it was a theft risk to release such data. We'll mark that particular shop as busier than normal. + +As for sales of silver, all of the shops who said they were purchasing more junk silver than normal noted that they had, in fact, sold more silver bullion, even as prices rose higher. A few, more than willing to divulge such information, even noted that the regulars were buying more weight, while one-time customers were coming in to raise cash. + +Therefore, from our small study, it can be reasonably concluded that even as prices tear towards the upside, there is little interest among the current physical silver investing community to let go of their silver. Instead, as prices head higher, we, as the good little investors we are, continue to buy more, not less, of the beautiful metals. + +The Irony of Investment Markets + +This is perhaps the irony of the silver markets, or any non-replacement investment market for that matter. As the market heads higher, the supply is not increasing at the retail recycling level or at the production level. Silver miners are not wholly silver miners, and they have to find copper before they can find silver. Therefore, while miners would love to start working overtime to dig silver, supplies ready to be quickly unearthed have yet to be discovered. + +In the long +================================================================================ +Rank = 49; Score = 4947968.0 +<|begin_of_text|>Amazon To Hachette And Authors: Here, Let Us Explain Basic Price Elasticity To You + +from the lower-price,-make-more-money-dimwits dept + +A key objective is lower e-book prices. Many e-books are being released at $14.99 and even $19.99. That is unjustifiably high for an e-book. With an e-book, there's no printing, no over-printing, no need to forecast, no returns, no lost sales due to out-of-stock, no warehousing costs, no transportation costs, and there is no secondary market -- e-books cannot be resold as used books. E-books can be and should be less expensive. + +It's also important to understand that e-books are highly price-elastic. This means that when the price goes up, customers buy much less. We've quantified the price elasticity of e-books from repeated measurements across many titles. For every copy an e-book would sell at $14.99, it would sell 1.74 copies if priced at $9.99. So, for example, if customers would buy 100,000 copies of a particular e-book at $14.99, then customers would buy 174,000 copies of that same e-book at $9.99. Total revenue at $14.99 would be $1,499,000. Total revenue at $9.99 is $1,738,000. + +The important thing to note here is that at the lower price, total revenue increases 16%. This is good for all the parties involved: The customer is paying 33% less. + +The author is getting a royalty check 16% larger and being read by an audience that's 74% larger. And that 74% increase in copies sold makes it much more likely that the title will make it onto the national bestseller lists. (Any author who's trying to get on one of the national bestseller lists should insist to their publisher that their e-book be priced at $9.99 or lower.) + +Likewise, the higher total revenue generated at $9.99 is also good for the publisher and the retailer. At $9.99, even though the customer is paying less, the total pie is bigger and there is more to share amongst the parties. Keep in mind that books don't just compete against books. Books compete against mobile games, television, movies, Facebook, blogs, free news sites and more. If we want a healthy reading culture, we have to work +================================================================================ +Rank = 50; Score = 4947968.0 +<|begin_of_text|>Still image from video of Justin Bieber during a DUI test at the Miami Beach Police Station. + +MIAMI (CBSMiami) – Video of Justin Bieber captured shortly after his arrest on Miami Beach shows the pop star submitting to a sobriety test in which he appears to momentarily lose his balance while attempting to walk along a straight white line. + +In the clip from the holding area of the Miami Beach Police Department, part of more than 9 hours of video released Wednesday, Bieber is walking heel to toe when he slightly stumbles as he turns. He starts the walk over again. + +Bieber then removes several items of clothing for a pat down. + +He’s eventually led to a holding cell where he waits for a couple hours before eventually being transported via van to jail. + +The embattled entertainer can be seen doing push-ups and pacing in the small space while he waits. + +The videos do not contain audio. + +Attorneys for CBS4 and its news partner the Miami Herald, along with attorneys for media outlets including CNN and the Associated Press, went to court last week to argue that since the tapes have been turned over to the defense as part of the discovery process in the case; the videos are now public record, and open to being viewed in their entirety. + +Tuesday, Bieber’s defense team said they did not object to prosecutors releasing most of the video footage with the exception of five video clips which may show the 19-year old pop star urinating into a cup for a drug test. + +Web Extra: Surveillance Video Recorded In Miami Beach Police Department Following Justin Bieber’s Arrest + +As for the five disputed video clips, Judge William Altfield will view them, and review more court filings, before announcing his decision on March 4th. + +Police stopped Bieber and singer Khalil Amir Sharieff on January 23rd after what they said was an illegal street drag race involving exotic cars. + +Breath tests showed that Bieber’s blood-alcohol content was below the.02 legal limit for underage drivers, but toxicology tests revealed the presence of the active ingredient in marijuana and the antidepressant Xanax. + +Bieber has pleaded not guilty to driving under the influence, resisting arrest and invalid-license charges.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 51; Score = 4915200.0 +<|begin_of_text|>Stuart Heritage reimagines the exchange of letters between the former culture secretary and David Cameron relating to her resignation following her expenses scandal + +LETTER TO THE PRIME MINISTER + +9 April 2014 + +Dear Prime Minister, + +It is with the thinnest possible veneer of synthesised guilt that I have been convinced to tender my resignation as a member of the cabinet. + +I am very grateful to you for inexplicably having my back throughout this mess, but the present situation has become a distraction from the vital work this government is doing to distract the public from the ongoing abuse of government expenses. + +I have been a member of the Conservative party for more than 30 years. I know that our party understands the importance of giving everyone the opportunity to succeed regardless of where they come from, so long as they come from my immediate family and prefer having a nicely decorated house to being able to sleep at night. + +I am immensely proud of what my team have been able to achieve during my time in government: uniting our great arts and cultural institutions in unanimous agreement that I should resign; putting women front and centre of the continuing expenses scandal; putting in place the legislation to enable all couples to have the opportunity to marry regardless of their sexuality. I mean, there was that time I voted against a move to allow ministers to introduce regulations banning harassment on grounds of sexual orientation, but that was way before you made me look after gay people. Besides, accurate recollections have never exactly been my strong point, have they? + +Did I mention that I went to a comprehensive? Because I totally did. + +Implementing the recommendations made by Sir Brian Leveson on the future of media regulation wasn't the success I'd hoped it to be, either. I mean, seriously, what's the point of limiting the freedom of the press if I can't even bully editors into shutting up about my dodgy mortgage? Still, whatever. I get the message. I'll lie low for a bit. + +I will continue to support you and the work of the government as you move forward. Ensuring the best future for people of Basingstoke who I'm directly related to has been my priority throughout the past nine years. Whether on the front or back benches, I will continue this work. But, who are we kidding, it'll probably be the front benches, right? + +I did mention I went to a comprehensive, didn't I? + +The only reason I became an MP was that I wanted to give something back. Sadly, there was some disagreement about the +================================================================================ +Rank = 52; Score = 4849664.0 +<|begin_of_text|>Major League Soccer and adidas on Monday revealed the jersey that will be worn by the MLS All-Stars when they take on Tottenham Hotspur at the 2015 AT&T MLS All-Star Game on July 29 at Dick's Sporting Goods Park in Colorado. + +The standout feature of the jersey is the traditional soccer sash featuring a modern design that incorporates both a star from the US flag and the leaf from the Canadian flag to commemorate the two nations that make up the league. The same sash design is also found on the neck taping. + +The primary colors of the shirt are red, white and blue in a nod to the new MLS crest, which features those same colors. The three stars that make up the MLS logo are also present on the back of the All-Star jersey with the words identifying the three league pillars they represent: "Club, Country, Community" (below, right). + +It's also a special 20th-season shirt, celebrating the league's history with a jock tag on the lower left area of the authentic shirts (below, left). + +The first players named to the AT&T MLS All-Star team, as well as the Commissioner's two annual selections, were to be revealed on Monday night.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 53; Score = 4784128.0 +<|begin_of_text|>The NBA's newest Canadian players aren't just looking forward to joining their pro teams. + +Nik Stauskas and Tyler Ennis are envisioning the day when they all get together wearing Canada's national team jersey. + +"I think, once we get in the gym together, getting chemistry and just get all the talent in one gym for the first time, I think that will be a big moment for Canada," Ennis said Thursday night at the NBA draft. "I think not only 2016, but the following Olympics I think we'll be able to make a run at it." + +Story continues below advertisement + +Stauskas, a long-range shooter from Mississauga, Ont., went eighth overall to the Sacramento Kings in Thursday night's draft at Barclays Center. Ennis, a point guard from Brampton, Ont., went 18th to the Phoenix Suns. + +Andrew Wiggins, a 19-year-old sensation from Vaughan, Ont., was taken No. 1 overall by the Cleveland Cavaliers. + +The Canadians were asked in their post-draft interviews if they think Canada can give perennial powerhouse United States a run for their money at the 2016 Rio Olympics. + +"I talked with Andrew (Wiggins) the last couple of days, just being here in New York at the same time. We're very excited for the future of Canada basketball," Stauskas said. "I feel like, if we really all commit to coming in and working hard and coming together, I think we could have a really good team. + +"I'm not saying we're going to win a gold medal right now, but I'm saying that we could have a chance to compete at that level if we really — if we all commit to it. + +Dwight Powell of Toronto was the fourth Canadian drafted Thursday, going 45th overall to the Charlotte Hornets. + +The first-round picks confirm Canada will have at least 12 players on NBA rosters next season — second most in the league behind the United States. Canada passed France, which previously held that honour with nine players on NBA rosters. + +Story continues below advertisement + +Story continues below advertisement + +"It's such a tremendous night for Canadian Basketball. For Andrew, Nik and Tyler to go in the top 18 in such an incredibly deep draft makes such a compelling case for the development of the Canadian game," said Wayne Parrish, Canada Basketball's president and CEO. + +Canada's men's basketball team hasn't played in the Olympics since Steve Nash led the squad to a seventh-place finish at the 2000 Olympics. + +The +================================================================================ +Rank = 54; Score = 4784128.0 +<|begin_of_text|>jeremy + +root + +Registered: Jun 2000 Distribution: Debian, Red Hat, Slackware, Fedora, Ubuntu Posts: 12,933 + +Rep: + +2007 LinuxQuestions.org Members Choice Award Winners + +Desktop Distribution of the Year - Ubuntu (30.83%) + +Server Distribution of the Year - Debian (30.30%) + +Live Distribution of the Year - KNOPPIX (22.88%) + +Database of the Year - MySQL (54.36%) + +Office Suite of the Year - OpenOffice.org (89.50%) + +Browser of the Year - Firefox (74.03%) + +Desktop Environment of the Year - KDE (52.08%) + +Window Manager of the Year - Compiz (33.65%) + +Messaging App of the Year - Pidgin (53.90%) + +Mail Client of the Year - Thunderbird (53.72%) + +Virtualization Product of the Year - VirtualBox (41.58%) + +Audio Media Player Application of the Year - Amarok (57.37%) + +Audio Authoring Application of the Year - Audacity (68.24%) + +Video Media Player Application of the Year - mplayer (41.78%) + +Video Authoring Application of the Year - mencoder (24.21%) + +Multimedia Utility of the Year - K3b (63.34%) + +Graphics Application of the Year - GIMP (69.15%) + +Network Security Application of the Year - nmap (24.95%) + +Host Security Application of the Year - SELinux (30.69%) + +Monitoring Application of the Year - Nagios (38.58%) + +Windows on Linux App of the Year - Wine (84.76%) + +IDE/Web Development Editor of the Year - Eclipse (22.29%) + +Shell of the Year - bash (87.33%) + +Text Editor of the Year - vi/vim (36.37%) + +File Manager of the Year - Konqueror (38.00%) + +Open Source Game of the Year - Battle for Wesnoth (21.74%) + +Programming Language of the Year - Python (21.78%) + +Winners should expect an email, including a nice winners badge in about a day or so. If you have any questions or suggestions on how we can improve the MCA's next year, do let me know. Visit + +--jeremy The polls are closed and the results are in. We had a record number of votes cast for the seventh straight year. Congratulations should go to all nominees. We had some extremely close races this year. Without further ado, here are the winners:Desktop Distribution of the Year - +================================================================================ +Rank = 55; Score = 4751360.0 +<|begin_of_text|>Why Recall? Here’s a look back at Tom McKay’s first 14 months in office: + +1. In a sexual harassment special investigation by former NJ Deputy Attorney General Lee Vartan, Tom McKay was found to have sexually harassed the Township clerk and a rent leveling board volunteer by describing them as lesbians and calling them “man-haters”. McKay never apologized to either woman, but rather blames his actions on his political opponents. The investigation costs Lopatcong taxpayers more than $10,000 and the council must censure the mayor. CensureResolution + +2. Tom McKay is served with a Tort Claims notice of intent to sue subjecting the Township and its taxpayers to substantial liability for his sexual harassment behavior. + +3. Tom McKay gavels Councilwoman Maureen McCabe in a public meeting and yells at her to “shut up.” Again, he refuses to take responsibility for losing his temper. + +4. Tom McKay publicly posts his personal “10 Commandments” for behavior at public meetings, but when a resident posts beside it her 10 criticisms of his public demeanor, he has them removed. The town receives legal notice from the resident’s lawyer that the Mayor violated her First Amendment rights. + +5. Tom McKay secretly secured access to the Township’s security cameras without the knowledge or consent of the Council. He also accepted those services as a gift from a Township vendor in violation of his terms of office. + +6. Lopatcong Township loses its credit status with many of its vendors due to the Mayor’s refusal to pay bills on time. Many vendors take a “payment in advance” policy with the Township. + +7. Tom McKay refused to consider the request led by Councilwoman Ciesla to refinance to Township’s debt obligation, for spite not business, and COST Lopatcong taxpayers $200,000. + +8. Tom McKay’s politically motivated banking decisions COST Lopatcong $20,000 in lost revenues. + +9. Tom McKay offered to trade his vote to reappoint the Township’s attorney and engineer for promises that they will not contribute to his opponents’ political campaign. + +10. Tom McKay replaces a long time member of the Township Planning Board with a non-US citizen political ally who, at public meetings, refuses to Pledge Allegiance to the US Flag. + +11. Tom McKay attempts to appoint his personal sexual harassment defense lawyer as the Township’s new labor lawyer. When the Council objects, he sues them. + +12. Tom McKay attempts to appoint his personal attorney as the Township’s new general attorney +================================================================================ +Rank = 56; Score = 4620288.0 +<|begin_of_text|>Less than 24 hours after Minneapolis voters finished hacking through a 35-candidate ballot, the Charter Commission voted unanimously to raise the entry fee from $20 to $500, matching St. Paul's. + +The lower fee, in place for at least 40 years -- according to commissioner Lyall Schwarzkopf, who remembered it from when he was city clerk in 1972 -- had enabled candidates, in the absence of a primary, to make "a mockery" of recent mayoral elections, said Commissioner Devin Rice. + +"We shouldn't overreact, but 35 candidates was ridiculous," said Commisisoner Jan Sandberg. "We absolutely have to do something about that filing fee." + +The $20 fee is set by a state law which allows cities to adopt higher fees. The $20 fee applied to only four cities: Minneapolis, St. Paul, Duluth and Rochester. St. Paul had put a $50 fee in place, then raised that to $500 in 2010. Rochester's fee went from $25 to $50 in 1988. Duluth's? Still $20. + +The new Minneapolis fees are intended to be entered into the overhauled city charter that voters overwhelmingly approved Tuesday. The revised charter takes effect in 2015. But first they have to go before the City Council, which can change them and send them back to the charter commission for another vote before they're entered into the charter. Commission chairman Barry CLegg said that if the council refuses to act, the proposal will go directly to voters in 2014. He said he'd expect overwhelming support. + +Rice first proposed raising the mayoral filing fee to $300, pointing out that he didn't want to judge the seriousness of any candidate, but did want to make sure that they filed campaign finance reports, which a $100 expenditure triggers. He also said he didn't think it should cost more to file for mayor than it does for U.S. Senate ($500) and governor ($400). + +But the rest of the board upped the ante, also raising the fee to enter a city council race from $20 to $250, and the fee to run for Park and Recreation Board and the Board of Estimate and Taxation from $20 to $100. + +"$500 today is not a barrier," said Commissioner Todd Ferrara. + +Twenty dollars in 1972 would be worth $111.70 today. + +Candidate Bob "Again" Carney agreed that Tuesday's ballot was "cumbersome." He said the higher filing fee might have +================================================================================ +Rank = 57; Score = 4587520.0 +<|begin_of_text|>With all the cowardly decisions, especially the recent National Anthem enforcement and the concussions (CTE) deceptions, the real question is: Why does Roger Goodell make $44 mil a year for running the NFL, a monopoly sanctioned by Congress? + +Maybe the NFL is in cahoots with Washington to distract the citizenry from lawmakers’ own doings, much like the “bread and circuses” of ancient Rome. Football diverts attention and placates the masses. (It’s also said to satisfy men’s innate lust for war but, just to make sure, Washington has us in a bunch of real wars, too.) + +Trump, to his credit, clearly loves football; why else would he wear that helmet hair? But getting rid of the taxpayer-subsidized, tax-exempt status, antitrust exemptions, etc. of the NFL would be a good move for Trump right now. + +With all its revenues, why can’t the NFL stand on its own? Anheuser-Busch alone paid $1.4 billion for NFL rights. If you want to sell beer to 18-34-year-old males, the NFL’s the place. If you want to sell power tools, also advertise on the NFL, although maybe also the WNBA. + +The American justice system, like the NFL, has tons of laws and rules, often laws layered upon laws by the ruling classes so they can prosecute whom they want, when they want, for whatever they want. Thus, it is within the power of the NFL and its beleaguered commissioner, Roger Goodell, to pursue one thing and not another. Just ask “Deflate-gate” victim Tom Brady. + +Goodell has had a long string of odd decisions. He toyed with penalties for uttering the N-word or sexist slurs. Yet he weighed free speech issues and reached a compromise: Players can listen to rap music but are not allowed to sing along. + +Goodell has dictatorially tinkered with some rules; he once decreed that players can no longer celebrate TDs by dunking the football over the goalpost crossbars. That didn’t go over well. If Americans wanted to watch a sport with no dunking, we’d watch Ivy League college basketball. + +Goodell threatened North Carolina and threatened to pull the Super Bowl from Arizona over legislation he viewed as anti-gay. He didn’t follow through with his threat, but the publicity dashed any hopes Arizona or North Carolina had of hosting the Tony Awards. And the gay rodeo business is non-existent +================================================================================ +Rank = 58; Score = 4587520.0 +<|begin_of_text|>The electronic billboard is on the Command Transportation building near Niles Center Road. + +Friday's epic USA versus Canada Winter Olympic hockey matchup highlighted another spat between the two countries involving pop star Justin Bieber. + +A new electronic scoreboard in a Chicago suburb shows Chicago Blackhawks forward Patrick Kane -- who plays for the USA -- and Team Canada's Jonathan Toews in between a picture of Bieber. The caption says "Loser Keeps Biebs." + +Bieber, who's Canadian, lives in Los Angeles, but after several high-profile incidents, many Americans have called for the singer to be deported. + +Best of the Sochi Olympics: Day 13 + +Bieber is also infamous among Chicagoans for a controversy he ignited after the Hawks won the Stanley Cup championship last year. + +While the singer was in the Blackhawks' locker room at the United Center taking pictures of the trophy, he was photographed standing on the Indian head logo, prompting a huge outcry of anger from fans of the team. + +After Canada won the Olympic gold medal game in Sochi, 1-0, the company immediately followed up with a new billboard below. + +Hey Canada, best 2 out of 3… errr… best 3 out of 5? pic.twitter.com/n7tLp8qyu0 — Command Sign (@CommandSign) February 21, 2014<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 59; Score = 4521984.0 +<|begin_of_text|>USA Soccer 2016 Logo + +The US Soccer national team today finally unveiled its new logo. Following the design language established by the overwhelmingly popular centenary crest, the new US Soccer logo boasts the US national teams' iconic navy and red in a sleek arrangement. This is the new US Soccer badge, introduced on February 29, 2016.Drawing inspiration from the previous US Soccer logo's badge shape, the new USA national team crest simplifies a lot elements of the old crest, while also shifting the colors to darker and less cartoony shades of blue and red.The three stars and the ball from the 1993 crest have been ditched, making place for a bold navywriting at the top of the badge. Dynamic red stripes adorn the lower half of the new USA soccer logo.Barely changed since its creation in 1993 and often mocked for its outdated 1990s look, the US Soccer badge has been in need for a redesign for a long time. Building on the success of the centenary crest seems a very viable and logical decision, in the interest of the US national team's fans.After Nike and US Soccer renewed their kit supplier contract until 2022, the two worked side by side on the new US Soccer badge, which is set to be first used on the new United States 2016 Home and Away Kits. The new USA Jerseys are set to be worn at the first-ever Copa America on US-soil, the 2016 centenary tournament.What do you think of the new 2016 USA national team logo? Drop us a line below.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 60; Score = 4521984.0 +<|begin_of_text|>Kady O’Malley is joining Maclean’s to cover the Manning Conference, hosted by Canada’s leading think tank for the right. She’ll be filing frequently over the course of the two-day conference—keep track at her Manning Manifests page, here. Watch the broadcast live all weekend here. + +Well, let’s start with the positive: the Manning Centre organizers kept their promise to deliver a “unique” debate experience, at least in contrast to the majority of the fourteen-wide face-offs that have characterized the race so far. + +Instead of lining the stage from curtain to curtain with candidates, they drew names from a hat and called them on stage in groups—two groups of four and two groups of three—with one topic per mini-slot, complete with answers, rebuttals and even the occasional outbursts of direct interaction between the candidates. + +Serving as moderator was the inimitable Tom Clark, who signed off his journalistic duties at Global News earlier this year to take up a government relations gig at Global Public Affairs, but handled his brief return to the interview chair with ease and obvious enjoyment, even—or possibly especially—when he was obliged to cut the mic on an overly loquacious contestant. + +(Those TV host reflexes stay sharp for at least a few months after hanging up the earpiece, apparently.) + +What was clear, however, from the post-debate chatter at the meet-the-candidate reception that capped off the official conference activities for the day, was that it still wasn’t enough to knock the contestants off script and into the sort of spontaneous, off-the-cuff exchange of views that can, in a nanosecond, change the course of a campaign and, eventually, Canadian political history. + +Forget “no knockout punches”—except for a few brief back-and-forths over health care (Chris Alexander and Kellie Leitch), trade negotiations (Lisa Raitt and Maxime Bernier) and the Canadian bona fides of Kevin O’Leary (O’Leary, Andrew Scheer and Deepak Obhrai)—the candidates were far too focused on delivering their now familiar applause and laugh lines. + +Leitch reminded everyone of her plan for face-to-face screenings at border crossings. Scheer praised pipelines. Chong defended his proposal for a “revenue neutral carbon tax.” Bernier decried supply management. O’Leary raged over deficits and taxes—particularly provincial. Andrew Saxton reminded everyone he was, and is, a businessman. Brad Trost and Pierre Lemieux—who, by +================================================================================ +Rank = 61; Score = 4456448.0 +<|begin_of_text|>Verizon is getting rid of contract plans, switching its offerings over to month-to-month plans like AT&T and T-Mobile did before it. Verizon's new plans are all based around shared data buckets — there's no single line or family plan. Instead, subscribers pay for a specific amount of data and then pay a per-device fee to hook it into the plan. + +Everyone loves per-line access fees! + +The options include 1GB for $30 per month, 3GB for $45 per month, 6GB for $60 per month, and 12GB for $80 per month. Adding a smartphone to the plan is $20 per month, adding a tablet or hotspot is $10 per month, and adding a smartwatch is $5 per month. Verizon says that larger plans will be available for those who want them; for those who just need a little bit of extra data one month, Recode reports that they'll be stuck paying $15 for an extra gigabyte, though they can always jump up to a new plan the next month. As usual, the pricing is not really friendly or tailored for anyone, but Verizon is under the odd impression that it's easier to understand. + +The plans go live next Thursday, on August 13th. Current Verizon subscribers will be able to hang onto their existing plans or switch over to the new plans "with some restrictions." + +Though doing away with contracts is a big shift, it's something we've seen coming. T-Mobile kicked off the trend about two years ago, and AT&T began following it recently by dropping contract offerings from third-party sellers. Alongside the shift away from contracts is the shift away from subsidized phones, meaning that people are going to need to start buying phones at full price or using carriers' payment plans. Verizon has been offering such a plan under the name Edge, though it's now decided to rename it "device payment option," which actually is easier to understand.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 62; Score = 4423680.0 +<|begin_of_text|>Content delivery network and Web security company Cloudflare has made a name for itself by fending off denial-of-service attacks against its customers large and small. Today, it's launching a new service aimed at winning over the most paranoid of corporate customers. The service is a first step toward doing for network security what Amazon Web Services and other public cloud services have done for application services—replacing on-premises hardware with virtualized services spread across the Internet. + +Called Keyless SSL, the new service allows organizations to use Cloudflare’s network of 28 data centers around the world to defend against distributed denial of service attacks on their websites without having to turn over private encryption keys. Keyless SSL breaks the encryption “handshake” at the beginning of a Transport Layer Security (TLS) Web session, passing part of the data back to the organization’s data center for encryption. It then negotiates the session with the returned data and acts as a gateway for authenticated sessions—while still being able to screen out malicious traffic such as denial of service attacks. + +In an interview with Ars, Cloudflare CEO Matthew Prince said that the technology behind Keyless SSL could help security-minded organizations embrace other cloud services while keeping a tighter rein on them. “If you decide you’re going to use cloud services today, how you set policy across all of these is impossible," he said. "Now that we can do this, fast forward a year, and we can do things like data loss prevention, intrusion detection… all these things are just bytes in the stream, and we’re already looking at them.” + +All of that puts Cloudflare on a path to, as Prince put it, do to Cisco and other network hardware suppliers what AWS and other cloud services have done to Hewlett-Packard and Dell-type server vendors. The technology could also help extend Internet services to areas of the world that lack the sort of trustworthy data centers that Cloudflare and other cloud providers operate from, as it keeps their most precious secrets locked away back at the core of their networks. + +Changing the locks + +The development of Keyless SSL began about two years ago, on the heels of a series of massive denial of service attacks against major financial institutions alleged to have been launched from Iran. “We got a series of fairly frantic calls from those banks saying they needed help with this problem,” Prince said. “We met with JP Morgan Chase, Goldman Sachs… pretty much everyone in that community. They described to us a challenge that put them between a rock and a hard place. They had spent billions on hardware +================================================================================ +Rank = 63; Score = 4390912.0 +<|begin_of_text|>Version 0.10.2 + +Content: + +Added a new Intelligence Skill - Incinerate: Launches a torrent of fire from your hand. The longer you repeatedly cast the spell from the same location, the larger and more damaging the flames become. + +Added a new cosmetic microtransaction effect - Blue Flame Incinerate: Replaces the standard flame of Incinerate with a ghostly blue fire. + +Added a new cosmetic microtransaction effect - Tiki Totem: Replaces the standard Spell Totem or Ranged Attack Totem with a coloured totem that has carved tiki designs. + +Added four new Unique items, three of which were designed by our Diamond Supporters. + +Added a new flamethrower ability to Snakes in the Dark Forest. + +Monster packs in end-game Map areas have had their size variance increased. + +The Solaris/Lunaris Temple tilesets have been improved and should now have higher performance. + +Added a new 3v3 PvP arena. One of the other arenas has been disabled temporarily while we fix a problem. + +Improved blood effects. + +Improved the Ice Spear and Lightning Arrow effects. + +Improved the performance of Ground Ice and Ground Tar. + +Improved the gem icons to be slightly better from a colour blindness point of view. + +Continued to incrementally improve the art, effects and environments. + +Features: + +The passive skill tree now requires confirmation before changes are applied. This prevents accidental misclicks. If you want to skip this confirmation, ctrl+click the passive and no window will be displayed. + +We've added a "Lost Password" button to the main menu. + +Pressing Ctrl+up in chat allows you to cycle through previous players you've talked to via whisper in this session. + +Chat messages that are more than one line long now indent subsequent lines so that it's harder to confuse other players with fake messages. + +There's now a /clear_ignore_list command that will remove all entries from the list of accounts you are ignoring chat messages from. Better management of ignore lists is coming in the future. + +Players in an area are now always granted the quest status of having killed a boss regardless of who killed it. + +The performance statistics overlay (F1) can now be bound to other keys. + +Audio Changes: + +Added PvP announcer audio for various events in matches. + +Added personalised introductions for Nessa. + +Added a few more Nessa greetings. + +Added new Tarkleigh greetings. + +Added some improved Marauder combat vocals. + +Added improved Marauder comment dialogue. + +Added new ranger dialogue and improved some existing ones. + +Overhauled the audio of +================================================================================ +Rank = 64; Score = 4390912.0 +<|begin_of_text|>Janice Fiamengo by + +T + +he election of Donald J. Trump revealed a bitterly divided nation and a large body of voters fed up with the status quo: fed up with a perennially ailing economy and a shrinking middle class; fed up with an overly intrusive state that crippled businesses with innumerable regulations and sought to redistribute wealth; fed up with a massively burdensome and insupportable debt; fed up with a president mentored by revolutionaries and malcontents who seemed not to like ordinary Americans much if at all; fed up with a leader who paid court to hostile foreign powers while alienating or outright betraying America's traditional allies; fed up with a president who spoke of heartland Americans as bitter people who cling to their guns or religion or bigotry; fed up with a president who told business owners "You didn't build that," called the Fort Hood terrorist attack an episode of workplace violence, and announced at the United Nations that the future must not belong to those who slander the prophet of Islam; and most of all, perhaps, fed up with an administration that placed the whole country in thrall to political correctness, making a range of important subjects off-limits for discussion on pain of job loss or public disgrace. Many of these voters felt that America had changed beyond all recognition, had lost its core values, and that neither Republican nor Democrat politicians offered any real resistance to that, or even seemed to notice.Donald Trump promised to change everything: to make America great again; to revive American industry; to put Americans back to work and make it possible for them to feed their families and to have prosperity once again; to reduce burdensome taxes and over-regulation of businesses; to renew American patriotism and the traditional family; to stop pretending that children do just as well without a father and money; to stop pretending that masculinity is toxic; to make it acceptable and normal once more to love God and to celebrate America's Judeo-Christian heritage; to decrease the mass immigration that is changing the face of America for the worse; to root out the Muslim Brotherhood sympathizers who have been given key positions in the American administration; to halt illegal immigration and to put a stop to the sanctuary cities that make a mockery of American laws; to rebuild the American military and to signal American power, not acquiescence, to America's foes; to honor American veterans in a manner appropriate to their service; to renew ties with America's traditional allies; and to vigorously protect America from domestic and foreign terrorism. Perhaps most importantly, he promised to make +================================================================================ +Rank = 65; Score = 4390912.0 +<|begin_of_text|>LexisNexis Group is a corporation providing computer-assisted legal research (CALR) as well as business research and risk management services.[2][3] During the 1970s, LexisNexis pioneered the electronic accessibility of legal and journalistic documents.[4] As of 2006, the company has the world's largest electronic database for legal and public-records related information.[5] + +History [ edit ] + +LexisNexis office in Markham, a suburb of Toronto, Ontario, Canada + +LexisNexis is currently owned by RELX Group (formerly known as Reed Elsevier).[6] + +The story of LexisNexis starts in western Pennsylvania in 1956, when attorney John Horty began to explore the use of CALR technology in support of his work on comparative hospital law at the University of Pittsburgh Health Law Center.[7] In 1965, Horty's pioneering work inspired the Ohio State Bar Association (OSBA) to develop its own separate CALR system, Ohio Bar Automated Research (OBAR).[8] In 1967, the OSBA signed a contract with Data Corporation, a local defense contractor, to build OBAR based on the OSBA's written specifications.[8] Data proceeded to implement OBAR on Data Central, an interactive full-text search system originally developed in 1964 as Recon Central to help U.S. Air Force intelligence analysts search text summaries of the contents of aerial and satellite reconnaissance photographs.[9] + +In 1968, paper manufacturer Mead Corporation purchased Data Corporation for $6 million to gain control of its inkjet printing technology.[10] Mead hired the Arthur D. Little firm to study the business possibilities for the Data Central technology.[10] Arthur D. Little dispatched a team of consultants to Ohio led by H. Donald Wilson.[11] Mead asked for a practicing lawyer on the team, so the team included Jerome Rubin, a Harvard-trained attorney with 20 years of experience.[12] The resulting study concluded that the nonlegal market was nonexistent, the legal market had potential, and OBAR needed to be rebuilt to profitably exploit that market.[12] At the time, OBAR searches often took up to five hours to complete if more than one user was online, and its original terminals (replaced with CRT text terminals in 1970) were noisy Teletypes with slow transmission rates of 10 characters per second.[13] OBAR also had quality control issues; Rubin later recalled that its data was “ +================================================================================ +Rank = 66; Score = 4358144.0 +<|begin_of_text|>MORTIMER ZUCKERMAN, owner of NY Daily News, US News & World Report and chair of the Conference of Presidents of Major Jewish American Organizations, one of the largest pro-Israel lobbying groups. LESLIE MOONVES, president of CBS television, great-nephew of David Ben-Gurion, and co-chair with Norman Ornstein of the Advisory Committee on Public Interest Obligation of Digital TV Producers, appointed by Clinton. JONATHAN MILLER, chair and CEO of AOL division of AOL-Time-Warner NEIL SHAPIRO, president of NBC News JEFF GASPIN, Executive Vice-President, Programming, NBC DAVID WESTIN, president of ABC News SUMNER REDSTONE, CEO of Viacom, "world's biggest media giant" (Economist, 11/23/2) owns Viacom cable, CBS and MTVs all over the world, Blockbuster video rentals and Black Entertainment TV. MICHAEL EISNER, major owner of Walt Disney, Capitol Cities, ABC. RUPERT MURDOCH, Owner Fox TV, New York Post, London Times, News of the World (Jewish mother) MEL KARMAZIN, president of CBS DON HEWITT, Exec. Director, 60 Minutes, CBS JEFF FAGER, Exec. Director, 60 Minutes II. CBS DAVID POLTRACK, Executive Vice-President, Research and Planning, CBS SANDY KRUSHOW, Chair, Fox Entertainment LLOYD BRAUN, Chair, ABC Entertainment BARRY MEYER, chair, Warner Bros. SHERRY LANSING. President of Paramount Communications and Chairman of Paramount Pictures' Motion Picture Group. HARVEY WEINSTEIN, CEO. Miramax Films. BRAD SIEGEL., President, Turner Entertainment. PETER CHERNIN, second in-command at Rupert Murdoch's News. Corp., owner of Fox TV MARTY PERETZ, owner and publisher of the New Republic, which openly identifies itself as pro-Israel. Al Gore credits Marty with being his "mentor." ARTHUR O. SULZBERGER, JR., publisher of the NY Times, the Boston Globe and other publications. WILLIAM SAFIRE, syndicated columnist for the NYT. TOM FRIEDMAN, syndicated columnist for the NYT. CHARLES KRAUTHAMMER, syndicated columnist for the Washington Post. Honored by Honest Reporting.com, website monitoring "anti-Israel media." RICHARD COHEN, syndicated +================================================================================ +Rank = 67; Score = 4358144.0 +<|begin_of_text|>California marked the month-long celebration Sikh Awareness Month (SAAM) celebrations for the fifth year in a row. With Donald Trump's the US President-elect, this year's awareness month could potentially be the most important SAAM to date, as the future of Sikh Americans hangs in the balance, said Sikh Coalition community development manager Harjit Kaur. She said minority communities across the nation have experienced an uptick in hate crimes, bullying, and discrimination as a result of the current political climate.A series of Sikh-related events were organised through the month, from local activities in schools to collaborating with the basketball team, the Los Angeles Clippers.Sikh Coalition community development manager Harjit Kaur explains how it all came about. “Since November 2010, California community leaders have engaged with legislators to recognize the contribution of Sikh Americans to California and to provide a platform to create aware ness about Sikhs in California educational institutions and the community.“ Harjit elaborates, “This year, Assembly member Jim Cooper presented Dr Onkar Singh Bindra and me with the state member resolution celebrating Sikh Awareness and Appreciation Month. California state superintendent of public instruction Tom Torlakson also recognised November as the awareness month for California's public schools. Sikh Awareness and Appreciation resolutions have been passed in Yuba City, Fresno, San Jose, Bakersfield, Fremont, Santa Cruz, and Santa Clara, along with numerous school districts across the state.The Sikh Coalition, along with many volunteers, has delivered presentations to hundreds of Californians across the state in classrooms and local libraries.The Sikh Coalition also co-hosted a Sikh Awareness and Appreciation game with the Los Angeles Clippers.Over 350 Sikh Americans attended the LA Clippers game alongside the other 19,000 spectators. It also featured the US national anthem by Raaginder `Violinder' Singh on the violin as well as the Los Angeles Sikh Boy Scouts of America colour guard, and half-time entertainment by Da Real Punjabiz.Harjeet said they work tirelessly to ensure that SAAM is used as a platform to meaningfully engage with legislators, educators, and community members across the state to address these issues. “This year, after helping to protect Sikh history in California in July, the Sikh Coalition secured safe passage of AB-2845, a groundbreaking new anti-bullying law in California in September“.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 68; Score = 4358144.0 +<|begin_of_text|>Image copyright Google Image caption Police said the man was in the shop after withdrawing money from a nearby cash machine + +A woman who kept a £20 note she found in a shop has been convicted of theft. + +Nicole Bailey, 23, picked up the cash at a branch of One Stop, in Blurton, Stoke-on-Trent, after it was dropped by a customer who withdrew it from a nearby cash machine, police said. + +Bailey pleaded guilty at North Staffordshire Justice Centre after seeing CCTV evidence. + +Her defence team reportedly argued the case should have been dealt with through a police caution. + +Bailey, of Highfield Drive, Blurton, must also pay £20 compensation, a £20 victim surcharge and £135 in court costs. + +More updates on this and other stories in Staffordshire + +What do you do if you find £20? + +The theft occurred on 8 August last year. + +Ch Insp Karen Stevenson, from Staffordshire Police, urged anyone who finds lost money to "do the right thing". + +"Morally, the right thing to do is hand in any found property so that the person who has lost out has every opportunity to be reunited with it." + +"This was someone's hard-earned money and we are committed to supporting all victims in our community." + +Correction 2 March 2017: An earlier version of this story contained an image of the wrong shop. This has now been amended.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 69; Score = 4292608.0 +<|begin_of_text|>The great comedian Milton Berle had a powerful put-down for the many acts that auditioned for his show but failed to make the grade. “High school,” he’d call them. And I can’t think of a better term to describe the Conservative campaign brain trust at the midpoint of this long election season. + +Just lately, for example, the campaign seems to be having a tiny problem with geography. Last Friday, a Harper tweet provided a lovely snapshot of Little Crater Lake in the state of Oregon to illustrate its commitment to the Canadian great outdoors. Earlier in the week, a campaign spot talked about shipbuilding in Halifax, N.S. … against a backdrop of Johnstown, Ont. A group of Colombian miners sat in for a party ad promoting a mineral exploration tax credit. And then there was the Tory ad which misidentified a B.C. salmon as its Atlantic cousin — the online trolls are still chuckling over that one. + +Well, these things happen. So do things like a sudden drop in the polls to third place. If that’s not a wakeup call, I don’t know what qualifies. + +It all made me think of an incident in the 2000 campaign, when Canadian Alliance Leader Stockwell Day lamented that Canadian talent and brains were heading south to the U.S., “just like the Niagara River.” The message was sound, the words were not; the Niagara actually flows northward. + +Day took a lot of ungentle ribbing for that error — but his mistakes fade in comparison to the fiasco in the Big Blue Bus these days. There’s one common thread that links Day’s experience in 2000 to what Harper is going through right now: a profound sense of confusion on the ground. But that’s where the similarities end. + +I worked on the Day campaign as legislative assistant to a front-bench MP. For the Alliance, the 2000 election was an unwelcome surprise from PM Jean Chretien, who always knew how to catch his opponents off guard. The CA had just been through a highly divisive leadership campaign and the party was low on funds. There was little corporate support for the party and many influential backers were waiting to see whether the Alliance could really replace the Progressive Conservatives as a centre-right alternative to the Liberals. Day fought that campaign on a shoestring, with family members playing key roles in the process. + +The Conservative party is having more and more difficulty recruiting volunteers at the riding level. Even campaign stalwarts who show up for every election are finding reasons to stay home this time. Who +================================================================================ +Rank = 70; Score = 4292608.0 +<|begin_of_text|>This article was originally published April 20, 2016, at 1:14 p.m. EST. It has been updated to include comments from Sen. John McCain. + +WASHINGTON — Pentagon acquisition chief Frank Kendall, in congressional testimony Wednesday, defended the controversial path to end the US reliance on Russian rocket engines for military space launch—and Senate appropriators signaled their continued support. + +That path, for now, involves using more of the engines while the industrial base ramps up to offer competing engines. DoD could replace the RD-180 engines by 2021 at the soonest, but the politicization of the issue is slowing down needed Congressional action, Kendall, the undersecretary of defense for acquisition, technology and logistics, said at a Senate Appropriations Defense Subcommittee hearing Wednesday. + +"It's a complicated issue, but the [Defense] Department's goals have never changed, to develop two engines, so if one of them has a failure, and we have a big gap in capability, we still have access to space," Kendall said. "We need competition to keep the expense down." + +Ultimately, DoD will seek public-private partnerships for reasonably priced launch services, versus buying a lone replacement engine and subsidizing a single company, Kendall said. Replacing the Atlas V with a combination of ULA's Delta IV heavy launch system would cost, Kendall said, as much as $50 million per launch. + +"It would be over $1 billion out of our defense budget to get us off RD-180s, and I don't think that's a good tradeoff," Kendall said. + +The US Air Force contracts for launch services with United Launch Alliance, a joint venture of Boeing and Lockheed Martin, which uses the RD-180 rocket engine to power its Atlas V launch vehicles. SpaceX is ULA's main competition for Pentagon business, as the company's Falcon 9 rocket won certification last year to compete for military space launches. + +× Fear of missing out? Fear no longer. Be the first to hear about breaking news, as it happens. You'll get alerts delivered directly to your inbox each time something noteworthy happens in the Military community. Thanks for signing up. By giving us your email, you are opting in to our Newsletter: Sign up for our Early Bird Brief + +For their support of DoD's plan, two key appropriators, the SAC-D's ranking member, Sen. Richard Durbin, D-Ill., and subpanel member Sen. Richard Shelby, R-Ala., have attracted the ire of the Senate's lead authorizer, Senate Armed +================================================================================ +Rank = 71; Score = 4227072.0 +<|begin_of_text|>Minecraft News Powered by Tumblr Minecraft Beta 1.3 * Implemented a new lighting engine with the help of MrMessiahs (can be turned off) + +* Changed the options around, added a new “Graphics options” button + +* Added beds. If all players in a map sleeps in a bed during night, the game immediately skips until morning + +* Added three new half-size blocks + +* Added Delay/Repeater redstone dust blocks + +* Added whitelisting to the server. To use, enter “whitelist ” where cmd is “on”, “off”, “add ”, “remove ”, “list” or “reload” + +* New save file format, old maps need to be converted (that might take a while..) + +* It’s now possible to have more than five save slots, and to rename saves + +* Scrollbars in both the texture pack list, and in the map selection screen + +* Replaced the Mojang splash image to reflect the new logo colors + +*.. and a bunch of bug fixes and tweaks!<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 72; Score = 4177920.0 +<|begin_of_text|>Breaking News Emails Get breaking news alerts and special reports. The news and stories that matter, delivered weekday mornings. + +Oct. 23, 2017, 12:39 PM GMT / Updated Oct. 23, 2017, 12:39 PM GMT By Chuck Todd, Mark Murray and Carrie Dann + +First Read is your briefing from Meet the Press and the NBC Political Unit on the day's most important political stories and why they matter. + +GOP wrestles with tax politics — but what about the policy? + +WASHINGTON — On a conference call Sunday afternoon with House Republicans, NBC’s Kasie Hunt reports, President Trump issued this warning: The GOP needs to pass his tax plan or it will lose in the 2018 midterms. + +Failure, he said, would be “really bad” for the party in next year’s elections, while success would be “like skating on ice.” + +But that focus on the tax plan’s electoral politics — when the midterm cake is already being baked — ignores the more important policy questions right now. + +Will the GOP tax plan really RAISE taxes on middle-class and upper-middle-class Americans? Given that the initial plan eliminates personal exemptions (replaced by a larger standard deduction), the nonpartisan Tax Policy Center said one-quarter of American taxpayers would pay higher taxes by 2027, including 30 percent with incomes between $50,000 and $150,000 and 60 percent of those making between $150,000 and $300,000. + +How much will the plan add to the deficit? The Treasury Department announced Friday that the federal budget deficit increased to $666 billion in the just-completed fiscal year – up from $585 billion last year. And that’s significant, because the Tax Policy Center says the GOP tax plan would reduce revenues by $2.4 trillion over the first 10 years (or $240 billion per year). + +How much will the wealthy benefit vs. everyone else? Since the GOP proposal, among other things, eliminates the estate tax and lowers taxes for “pass-through” businesses, the same Tax Policy Center says that more than half of the benefits in the first year go to the Top 1 percent of taxpayers, while 30 percent of the benefits will go to the Top 0.1 percent. + +Are Republicans really going to tax 401k accounts? That has been one of the trial balloons that Republicans have released. “The proposals under discussion would potentially cap the annual amount workers can set aside to as low as $2,400 for 401(k) accounts, +================================================================================ +Rank = 73; Score = 4161536.0 +<|begin_of_text|>Buy Photo State senator Virgil Smith holds his head down in Judge Lawrence Talon's courtroom at the Frank Murphy Hall of Justice in Detroit on Monday, March14, 2016. Smith's ex-wife Anistia Thomas is seen behind. Smith will serve 10 months in jail with no early release and comply with psychiatric, alcohol and drug counseling, but does not have to give up his senate seat. Jessica J. Trevino/Detroit Free Press. (Photo: Jessica J. Trevino, Detroit Free Press)Buy Photo + +A judge on Monday sentenced embattled state Sen. Virgil Smith to 10 months in jail with no early release but did not require him to resign from his job as a legislator, which was part of a sentencing agreement in the case against him. + +"It would be illegal for me to impose as a condition of sentence that he resign from office and that he not hold public office during the pendency of his probation,” Wayne County Circuit Judge Lawrence Talon said during the hearing. + +The decision means prosecutors could now pull the agreement they reached with the defense because it required Smith to resign. + +As part of the plea, Smith, 36, D-Detroit, pleaded guilty to a felony count of malicious destruction of personal property of $20,000 or more last month and admitted shooting his ex-wife's 2015 Mercedes-Benz on May 10, 2015. + +Talon said Smith could be expelled, voted out of office or resign but requiring him to step down “offends the state constitution.” + +"You do not have to resign from the Senate," he told Smith, explaining his ruling. "I'm taking that off the plea agreement." + +It’s not yet known if Smith will voluntarily resign from office, though some say they expect that to happen. + +“The plea is legal, and the defendant agreed to it,” Wayne County Prosecutor Kym Worthy said in a statement released after the two-hour hearing in the Frank Murphy Hall of Justice. “If all the conditions are not accepted by the court, we will withdraw our plea. We are certain that we stand on solid legal ground.” + +Smith told investigators the shots were fired after his ex-wife, Anistia Thomas, pushed her way into his house on Wexford and attempted to attack the woman in his bed. + +Thomas previously testified Smith punched her in the face and rammed her head into the floor and wall when she was inside his home. She gave a statement in the courtroom Monday, saying Smith has lied to save face and is "still blaming me +================================================================================ +Rank = 74; Score = 4161536.0 +<|begin_of_text|>Anime graduate programs announced a new decision to award physics PhD degrees to those who fail the anime qualifying exam. + +The decision came after many years of low passing rates and complaints that anime PhDs have been too exclusive and mentally strenuous. MIT’s Dean of Admissions also reasoned that many of the questions on the anime qualifying exam “require the same, if not more amount of critical and analytical thinking that the physics exams do.” + +Past graduate exams included questions on harem attractive field lines, the capacitance of a tsundere’s ability to withhold love confessions, and the leading shifts in the energies of the degenerate first excited states of main character power ups. Most students who failed the exam claimed that it was the section on modern anime that tricked them up. + +“In the written exam, I had to manipulate the twin paradox setup such that the twin who traveled to a distant star and the twin on earth met at some intermediate point on the same day as the repulsive force of dubbed anime exceeded the attractive force of subbed anime,” said Shuning Li, who failed the exam and was forced to find a job in research. + +Most students also assert that the anime qualifying exam has a notoriously difficult oral exam in which a committee of three faculty members interview the student about their research. The student is then asked to solve a difficult problem in the student’s general field of study. The problem slated for 2016 was purported to be about evaluating key frame density and quantization of Kill la Kill fight scenes. + +With the consolation physics degree, many universities expect anime scholars to be satisfied with the apparent offset of the difficulty of the actual exam. However, scholars expressed indignance and found that this action condescending toward their level of knowledge. Anime students have also claimed that this move was an attempt to force them into unfulfilling life careers and stifle their pursuit of happiness. + +“It’s all about doing what you love these days. I want to make a difference in the world and neutrinos don’t do anything,” said Richard Dembinski, a Johns Hopkins Undergraduate studying the biomedical engineering of nekomimis. “What can I do with something as impractical as a physics degree? There’s a reason why I’m studying anime engineering.” + +Higher up officials have announced no intention to get rid of the consolation physics degree. In fact, many graduate schools have begun to change there anime programs to tailor the physics degree, expecting most anime students to fail out of their intended program. There is no known report of the anime qualifying exam being made easier for +================================================================================ +Rank = 75; Score = 4128768.0 +<|begin_of_text|>Thought of the Moment + +Death must be so beautiful. To lie in the soft brown earth, with the grasses waving above one's head, and listen to silence. To have no yesterday, and no tomorrow. To forget time, to forgive life, to be at peace. -Oscar Wilde, writer (1854-1900) + +. It's free. Death must be so beautiful. To lie in the soft brown earth, with the grasses waving above one's head, and listen to silence. To have no yesterday, and no tomorrow. To forget time, to forgive life, to be at peace. -Oscar Wilde, writer (1854-1900) Receive quotations (and words) in our daily newsletter. It's free. + +Satanic SkyCatkins SayCatkin SaysAntics YaksCask SanityCask SatinySack SanitySack SatinyYacks SatinYacks SaintYacks AntisYacks StainYack SaintsYack StainsYack SatinsScanty SakiA Antics SkyA Sacks TinyA Casks TinyA Stacks YinA Tacks YinsA Stack YinsA Yacks NitsA Yacks SnitA Yacks TinsA Tacky SinsA Yack SnitsA Scanty SkiA Cyans SkitA Cyans KitsA Cyan SkitsA Sac StinkyA Scats InkyA Casts InkyA Cays StinkA Cays KnitsA Cay StinksA Akin CystsA Yanks TicsA Snaky TicsA San StickyA Nasty SickA Tansy SickA Antsy SickA Nays TicksA Nays StickA Nay SticksA Any SticksA Stays NickA Stay NicksSancta I SkyAntic As SkySnack Stay ISnack Say TiSnack Say ItSnack Ay SitSnack Ay ItsSnack Ay TisSnack Ya SitSnack Ya ItsSnack Ya TisSnacks Ay TiSnacks Ay ItSnacks Ya TiSnacks Ya ItCask Ani StyCask Nasty ICask Tansy ICask Antsy ICask Nays TiCask Nays ItCask Nay SitCask Nay ItsCask Nay TisCask Any SitCask Any ItsCask Any TisCask As TinyCask Sat YinCask Stay InCask Say NitCask Say TinCask At YinsCask Ay NitsC +================================================================================ +Rank = 76; Score = 4128768.0 +<|begin_of_text|>Prime Minister Justin Trudeau is abandoning his long-held promise to change the way Canadians vote in federal elections. + +In a mandate letter for newly appointed Democratic Institutions Minister Karina Gould, Trudeau makes it clear that electoral reform — once top of mind for the Liberal government — is no longer on the agenda. + +“Changing the electoral system will not be in your mandate,” the prime minister writes in the letter, released Wednesday. + +A variety of consultations across the country have shown that Canadians are not clamouring for a change in the way they choose their federal government, the letter continues. It also rules out the possibility of a national referendum. + +“A clear preference for a new electoral system, let alone a consensus, has not emerged,” Trudeau writes. “Furthermore, without a clear preference or a clear question, a referendum would not be in Canada’s interest.” + +Trudeau repeatedly promised — both as a campaigning Liberal leader and as prime minister in a speech from the throne — to get rid of the current first-past-the-post voting system in time for the 2019 federal election. + +The Liberals have since given themselves some wiggle room, saying they would not go ahead without the widespread support of Canadians. + +Canadians made their views known through the House of Commons special committee on electoral reform, town halls held by MPs from all parties, the travels of former minister Maryam Monsef and a much-maligned online survey called MyDemocracy.ca. + +The mandate letter shows that Trudeau and do not believe those consultations have produced their desired — albeit undefined — level of support for electoral reform, let alone any clarity on a preferred replacement. + +The about-face is sure to provoke a passionate response from their political rivals. + +The New Democrats, who have long called for a system of proportional representation, went into a meeting with Gould on Tuesday hoping to hear the new minister repeat Trudeau’s original, unequivocal promise: that the 2015 vote would be Canada’s last under first-past-the-post. + +“That is why that ministry exists,” MP Nathan Cullen, the NDP’s democratic reform critic, said Tuesday. “That’s why she sits in cabinet, in large part — it’s to fulfill that promise.” + +The Conservatives, who had pushed for a referendum, are likely to be pleased with the status quo, but will no doubt excoriate the government for breaking such a prominent campaign commitment. + +There are also some big new items in the mandate letter. + +Trudeau wants Gould, Defence Minister Harjit Sajjan and Public Safety Minister Ralph Goodale to come up with ways to defend the Canadian +================================================================================ +Rank = 77; Score = 4128768.0 +<|begin_of_text|>The FBI has been lobbying top internet companies like Yahoo and Google to support a proposal that would force them to provide backdoors for government surveillance, according to CNET. + +The Bureau has been quietly meeting with representatives of these companies, as well as Microsoft (which owns Hotmail and Skype), Facebook and others to argue for a legislative proposal, drafted by the FBI, that would require social-networking sites and VoIP, instant messaging and e-mail providers to alter their code to make their products wiretap-friendly. + +The FBI has previously complained to Congress about the so-called "Going Dark" problem – the difficulty of doing effective wiretap surveillance as more communications have moved from traditional telephone services to internet service companies. + +Under the Communications Assistance for Law Enforcement Act, or CALEA, passed in 1994, telecommunications providers are required to make their systems wiretap-friendly. The Federal Communications Commission extended CALEA in 2004 to apply to broadband providers like ISPs and colleges, but web companies are not covered by the law. + +CNET reports that in addition to this push from the FBI, the Federal Communications Commission may be looking at reinterpreting CALEA to demand that video and non-telephone-replacement VoIP products such as Skype and Xbox Live be modified to include backdoors that allow FBI surveillance. + +The news comes on the heels of another FBI plan that began kicking around in 2010 that would require backdoors in encrypted communication systems. That proposal, which would revisit the encryption wars of the 1990s, has failed to gather administration backing.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 78; Score = 4112384.0 +<|begin_of_text|>Yes, Eric Shin­seki had to go, and he prob­ably knew it him­self once the hor­ror stor­ies sur­faced. As the re­tired four-star gen­er­al learned at West Point, the com­mand­er is ul­ti­mately re­spons­ible. While the Vet­er­ans Ad­min­is­tra­tion has been a ma­na­geri­al bleed­ing sore for years, the chaos and per­haps crimin­al­ity at sub­or­din­ate ech­el­ons of the VA on Shin­seki’s watch made his sur­viv­al im­possible. + +But let’s not for­get that Ric Shin­seki is not just a highly dec­or­ated com­mand­er and wounded war­ri­or, los­ing part of his foot in Vi­et­nam and claw­ing his way back onto act­ive duty against the wishes of Army brass. He’s a truth-tell­er of the first rank — and that dis­play of char­ac­ter so en­raged the George W. Bush de­fense team that he en­countered some of the shab­bi­est treat­ment an of­ficer and a gen­tle­man has ever en­countered dur­ing my 46 years serving in and hanging around the Pentagon. + +It didn’t help his case with the Bush­ies that Bill Clin­ton had ap­poin­ted him Army chief of staff. Moreover, De­fense Sec­ret­ary Don­ald Rums­feld, who didn’t en­joy be­ing chal­lenged, quickly took a dis­like to Shin­seki after sev­er­al policy and strategy dis­agree­ments. + +Rummy was so in­tent on pun­ish­ing Shin­seki out, in fact, that he dir­ec­ted one of his flack-shop aco­lytes to leak word of his re­place­ment to The New York Times — 15 months be­fore Shin­seki’s four-year term was up. + +This had the in­stant ef­fect of ren­der­ing Shin­seki a lame duck with­in the E-ring, the Pentagon’s power cor­ridor. It was cheesy, petty, shame­ful, +================================================================================ +Rank = 79; Score = 4112384.0 +<|begin_of_text|>Following news last week that EPA had allowed for the re-introduction of asbestos into the construction market, architects across the country have lambasted the proposal, calling for a blanket ban on asbestos. In response, the environmental agency has issued the following statement: + +“The EPA accepts the blanket ban on asbestos and will under no circumstances permit the production of asbestos blankets, but other asbestos products may be used where there is significant desire to do so because after all, asbestos is the best osTM #MAGA.” + +In addition to this statement, the EPA released a full list of materials granted SNUR (Significant New Use Rule) status following a year-long safety inquiry. The document can be seen below. + +ENVIRONMENTAL PROTECTION AGENCY + +NOTIFICATION OF SIGNIFICANT NEW USE RULES FOR PROHIBITED MATERIALS: + +Structural Cheese: With asbestos back on the menu, so is cheese. Once again, certain cheeses will be classified as safe for construction. Any cheese harder than an average supermarket cheddar will once again be permitted for use in structural members for public and governmental buildings no larger than a soccer stadium. Many of these cheeses have been consistently described as ‘robust’, ‘extra strong’ and ‘particularly tangy’ and so are clearly suitable for supporting extremely heavy loads. For obvious reasons, mozzarella is banned for construction use in pizza shops and all cheese is banned from mice farms. The EPA would like to clarify that this cheese legislation has nothing to do with the 1.4 billion-pound surplus of cheese in the U.S. + +Napkin Tiling: To combat the wasteful practice of discarding unused restaurant-grade napkins, the EPA will classify them as suitable for tiling bathrooms and wet-rooms. The famous slogan “one sheet does plenty” clearly applies to the use of napkins as bathroom tiles, and they are marketed as “super-absorbent” so are perfect for use in areas that tend to get excessively wet. They are also often folded into the shape of boats, or swans, so are clearly waterproof. Napkins with scribbles and doodles on them will not be permitted. + +Chocolate Fire Doors: In order to drastically cut deforestation in North America the EPA proposes to drastically increase deforestation in South America. This can be done by doubling chocolate production in South American countries. This surplus chocolate will be used in the construction of fire-rated door sets. In the event of a fire, conventional timber doors catch fire, and fire is well known to be the leading cause of fire. Chocolate +================================================================================ +Rank = 80; Score = 4096000.0 +<|begin_of_text|>Riley Reiff’s Future With Detroit Depends On What Quinn Does In The Coming Months. + +The biggest possible free agent in Detroit in 2017 is Riley Reiff. He is a key part to the offensive line and losing him would be a big blow. If Detroit can bring back Reiff, then they can find a better backup in the free agency or the NFL Draft. If Reiff leaves though, Detroit may need to look for a starter than a backup. Reiff is the most important free agent and the sooner we have a decision on what he will do, the sooner we can figure out what to do with others. + +Riley Reiff + +Reiff started 14 games and did well moving over to right tackle with Taylor Decker coming in to take over the left tackle spot. Reiff only allowed four sacks and got called for three penalties this season. Reiff is one of the best right tackles in the free agent class, so he could ask for a high price, but offensive tackle is an important position for this team, so paying him more than what he is worth may be what Detroit has to do to keep him. + +Decision: Stay + +Taylor Decker + +Decker had a fantastic rookie year in the league. He started in all 17 games and only allowed four and a half sacks this year. He struggled with penalties, getting six called on him, but that isn’t shocking from a rookie. Decker has been what this team needed at the position. Getting rid of those penalty issues and Decker can be a long-term left tackle in Detroit. + +Decision: Stay + +Corey Robinson + +Robinson started three games this year and did well in Reiff’s spot when he was out. Robinson got called for one flag and allowed only one sack. Robinson was a former seventh round pick and many questioned if he would make the team this year. Those three games he proved to be a solid backup and getting rid of him isn’t what Detroit wants to do. + +Decision: Stay + +Cornelius Lucas + +Lucas only appeared in five games this year and while he didn’t draw any flags, he allowed a sack. Now he didn’t start like Robinson, but Lucas isn’t a good backup here in Detroit. He is like a traffic cone at times and people just get around him so quickly, I was surprised he made the team this season. His performance this season could have been the final straw as he was the worst offensive tackle on the team. + +Decision: Go + +Possible Offensive Tackle Replacements + +Now +================================================================================ +Rank = 81; Score = 4079616.0 +<|begin_of_text|>The Stable channel has been updated to 28.0.1500.68 for all Chrome OS devices (Platform version: 4100.68.0 for all devices except Cr-48, which was updated to 4100.78.0). This build contains a number of feature enhancements, bug fixes and security improvements. Machines will be receiving updates over the next several days. + +Release Highlights: + +Numerous crash fixes and performance improvements. + +Speed improvements for the Files app, along with support for files that are Shared with me and Recent on Google Drive. + +Monitor Rotation and Scaling - you can scale your UI smaller and rotate the screen on all Chromebooks. + +Notification pop-up message appears when screenshots are taken. Clicking on the notification will bring you right to the screenshot. + +Replaced the Japanese IME with a NaCl extension version downloads a better language model upon actual usage. + +Added transliteration IMEs for Indic languages from Google Input Tools. + +Touch link highlighting on Chromebook Pixel - we now provide feedback on which link you click or button you press. + +Updated Chrome Office Viewer (Beta) extension + +Pepper Flash updated to 11.8.800.94 + +Known issues: + +On Acer C7 Chromebooks, docked mode has been disabled when connecting to an external monitor. + +I f you find new issues, please let us know by visiting our help site or filing a bug. Interested in switching channels? Find out how. You can submit feedback using ‘Report an issue...’ in the Chrome menu (3 horizontal bars in the upper right corner of the browser). + +Danielle Drew + +Google Chrome<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 82; Score = 4046848.0 +<|begin_of_text|>Thanks for visiting, why don't you come back? Follow me on Twitter and get blog updates by e-mail: subscribe here + +Our extinction crisis continues; 2013 allowed us to safely conclude that we will never again see the animals listed below ( Our extinction crisis continues; 2013 allowed us to safely conclude that we will never again see the animals listed below ( 2012 version here ). + +One of the last known photos of + +a Formosan Clouded Leopard; + +taken by Torii Ryūzō. + +The Formosan Clouded Leopard (Neofelis nebulosa + +brachyura + +) of Taiwan + +. None have been seen in over thirty years, despite a recent and intensive 13-year effort to document one. We did just about everything we could to eliminate this animal; we destroyed their habitat, killed them for their skins, and got rid of the other animals they normally ate. They didn't have a chance. + +Chioninia coctei), which hasn't been seen since 1912, The Cape Verde Giant Skink (), which hasn't been seen since 1912, has been declared extinct, although a jawbone from one of these lizards was found in some cat scat in 2005. However, since then the cat + +(i.e., house cat) + +population has increased substantially and, aided by rats and dogs, has likely wiped out the skink. + +Macrognathus pentophthalmos) The Sri Lanka Spiny Eel ( is probably extinct. As recently as 1980 the species was considered common but it was likely done in by a non-native species of fish that ate many of them. + +The Eskimo Curlew (Numenius borealis) was once so abundant that the sizes of its flocks were compared to those of Passenger Pigeons. They now have something else in common. The last known Eskimo Curlew was observed in 1963; Canada is likely to decide it is officially extinct because it has been 50 years since one has been seen. Eskimo Curlews probably suffered from a decline in their locust prey as well as loss of habitat but the primary cause of extinction is thought to be overhunting. Indeed, the last known Eskimo Curlew was shot by a hunter in Barbados. + +The Southern Darwin's Frog + +(i.e., not the extinct one); + +by Mono Andes, Wikimedia. + +This year, scientists concluded that the Northern Darwin's + +Frog ( + +Rhinoderma rufum +================================================================================ +Rank = 83; Score = 4046848.0 +<|begin_of_text|>For Immediate Release + +Note similarities to previous administration + +Top 10 disasters of the 2009 Obama administration (in no particular order): + +1. Cash for Clunkers + +2. War escalation in Afghanistan + +3. Giant government health care expansion bill + +4. Post office loses money hand over fist + +5. Stimulus package + +6. Expansion of “state secrets” doctrine + +7. Big increase in unemployment + +8. “Bailout” Geithner as Treasury Secretary + +9. Skyrocketing federal spending + +10. Huge federal deficits + +Top 10 disasters of the 2001-2008 Bush administration: + +1. Cash for Car Companies + +2. War in Iraq + +3. Giant Medicare expansion bill + +4. Post office loses money hand over fist + +5. Stimulus “rebate” checks + +6. PATRIOT Act + +7. Big increase in unemployment + +8. “Bailout” Paulson as Treasury Secretary + +9. Skyrocketing federal spending + +10. Huge federal deficits + +Wes Benedict, Libertarian Party Executive Director, commented, “Republicans and Democrats keep expanding government and creating more and more problems. We’re encouraging as many Libertarians as possible to run for Congress in 2010. In Texas, the state with the earliest filing deadline, Libertarians have already filed for 31 of 32 Congressional seats.” + +For more information, or to arrange an interview, call LP executive director Wes Benedict at 202-333-0008 ext. 222. + +The LP is America’s third-largest political party, founded in 1971. The Libertarian Party stands for free markets and civil liberties. You can find more information on the Libertarian Party at our website. + +###<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 84; Score = 4030464.0 +<|begin_of_text|>In the unattainable quest for leftist-deemed inclusivity, a Catholic school in California is purging itself of all remnants of Catholicism, including a statue of baby Jesus and the Virgin Mary. + +On Thursday, Marinij.com reported that San Domenico School in San Anselmo, California, had removed and relocated a number of Catholic statues in order to be more inclusive of other religions, sparking strong reactions from parents of students who attend the allegedly Catholic school. + +Thus far, the school has removed and relocated all but 18 of their 180 Catholic icons and statues. The head of the school's board of trustees, Amy Skewes-Cox, claims such removals are “completely in compliance” with school rules, which were approved by the board and the Dominican Sisters of San Rafael last year. + +Kim Pipki, whose daughter used to attend San Domenico, recalled the removal of the statue of baby Jesus and Mary as particularly contentious. + +“The one main statue that has everyone fired up is the baby Jesus and Mary one,” said Pipki. “It was at the center of the primary school courtyard.” + +“It was less about God and more about passing on some traditions,” the mother noted. “People were shocked that the statues were pitched in the basement.” + +Shannon Fitzpatrick, whose eight-year-old daughter currently attends Domenico, wrote an email to school officials and the Dominican Sisters of San Rafael expressing her disapproval of the so-called inclusive changes, suggesting their version of "inclusion" apparently means gutting Catholicism from the school altogether. + +“Articulating an inclusive foundation appears to mean letting go of San Domenico’s 167-year tradition as a Dominican Catholic school and being both afraid and ashamed to celebrate one’s heritage and beliefs,” wrote Fitzpatrick. + +“In our time here, the word ‘Catholic’ has been removed from the mission statement, sacraments were removed from the curriculum, the lower school curriculum was changed to world religions, the logo and colors were changed to be ‘less Catholic,’ and the uniform was changed to be less Catholic," she continued. + +The mother said she first became alarmed by the school's changes last year, when San Domenico removed “first reconciliation and first communion from the second-grade curriculum." + +In a follow-up email, Fitzpatrick noted that she was not the only parent disturbed by the "inclusive" changes. + +“There are other families having the same concerns I do. Many parents feel if the school is heading in a different direction then the San Domenico community should +================================================================================ +Rank = 85; Score = 4014080.0 +<|begin_of_text|>This book has gotten a lot of flack for two reasons: (1) A number of people were upset by the large amount of errata posted after the book came out. (2) A number of people were upset by the perceived power-creep that this book carried with it, especially in the archetype section. + +Both of these are reasonable complaints that I largely agree with. + +That said, this book also contains a cornucopia of player options that are great fun. A number of the classes it introduced are now mainstream: it’s hard to imagine playing the game without options like the Brawler, the Investigator, the Slayer, the Bloodrager, the Hunter, or the Warpriest. Or to play without archetypes like the Bolt Ace (Gunslinger), Mutation Warrior or Martial Master (Fighter). + +Moreover, the book introduced a number of feats that improve on the available build options available to most players (Extra Hex! Slashing Grace!). Likewise, although the spells in this book seem to have flown under the radar, there are a lot of nice and interesting spells are introduced in this book (Glue Seal, Communal Align Weapon, Wall of Blindness/Deafness, Wall of Nausea, Anti-Incorporeal Shell, Adjustable Disguise, Adjustable Polymorph, Investigative Mind, etc). + +Easily 5 stars worth of good material here. Given the unusually large amount of errata, I feel compelled to deduct a star. But all that said, it’s hard to imagine playing Pathfinder without this book -- after the Core Rulebook and Advanced Players Guide, it’s probably the best book for players to pick up.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 86; Score = 4014080.0 +<|begin_of_text|>SETT TV Profile Joined January 2011 77 Posts Last Edited: 2013-07-26 17:53:09 #1 + +ASUS ROG releases the rules, map pool and the groups which start off the ASUS ROG Summer 2013 StarCraft II tournament in Helsinki August 1-3. + +It is less than a week until the first matches of ASUS ROG Summer 2013 go live. The players have now been placed in the groups for the first group stage which will be played on the first day of the tournament, Thursday 1st August 13:00--22:30 CEST. The 32 players are divided into groups of four, which will be played in dual tournament format with best-of-5 matches. The top two players of each group advance to the second group stage, which is played in a similar fashion on Friday 2nd August. + +It is less than a week until the first matches of ASUS ROG Summer 2013 go live. The players have now been placed in the groups for the first group stage which will be played on the first day of the tournament, Thursday 1st August 13:00--22:30 CEST. The 32 players are divided into groups of four, which will be played in dual tournament format with best-of-5 matches. The top two players of each group advance to the second group stage, which is played in a similar fashion on Friday 2nd August. Replacements + +It is very unfortunate, but there have been more cancellations this week as Evil Geniuses.Jaedong, + +Evil Geniuses.Suppy, Millenium.ForGG, and are unable to attend the event. + +However, despite the proximity of the event, we were able to get noteworthy replacements: Dignitas.TargA, + +NaVi.BabyKnight, prOperty.RunA, and will take the vacant spots. We are very grateful to these players and teams who heeded our call and made the arrangements in a short time so that the tournament could retain its high standards. + +It is very unfortunate, but there have been more cancellations this week as Alliance.NaNiwa and Fnatic.NightEnD are unable to attend the event.However, despite the proximity of the event, we were able to get noteworthy replacements: Fnatic.Naama and Bischu will take the vacant spots. We are very grateful to these players and teams who heeded our call and made the arrangements in a short time +================================================================================ +Rank = 87; Score = 3997696.0 +<|begin_of_text|>Stone tablets and memorial archways displayed at a museum dedicated to the victims of the Cultural Revolution (1966-76) in Shantou, South China's Guangdong Province have been covered by cement. + +Several exhibition halls at the museum - located on Tashan Mountain in Shantou's Chenghai district - were adapted for other purposes, and all the stone tablets, memorial gateways and victims' tombs at the site were covered with cement, Hong Kong-based news site hkej.com reported Wednesday. + +Peng Qi'an, founder of the museum and former vice mayor of Shantou, told the Global Times on Thursday that he had not entered the museum since its exhibits were covered. + +"Since March, the materials [in the museum] had been covered by banners with core socialist values slogans. Then they were covered by cement, making them illegible," Peng told the Global Times. + +When contacted by the Global Times for comment, an employee of the Chenghai district's publicity department said that she was unaware of the move. Shantou's publicity department was unable to be reached for comment as of press time. + +A report published by Hong Kong's Ming Pao newspaper in May said signs showing the way to the museum had been covered with banners extolling the core socialist values and emblazoned with the Chinese national flag and the Party emblem. + +A 20-meter-wide and 5-meter-high stone tablet in the museum's public square that was originally inscribed with the phrase "Take history as a mirror, do not let the tragedy of the Cultural Revolution repeat itself" had been replaced with enormous posters advocating the Chinese dream and core socialist values, the report added. + +Peng began developing the site in the late 1990s, and construction of the museum was completed in 2005. The area is dotted with dozens of graves, memorial arches and marble memorials carved with passages from history texts that describe class struggle or violent deaths. + +According to previous reports, it is allegedly China's only museum dedicated to the Cultural Revolution, and the local government neither supported nor opposed its construction. + +Historians trace the start of the Cultural Revolution to May 16, 1966, when a conference of the Political Bureau of the Central Committee of the Communist Party of China (CPC) passed a circular in which Party leader Mao Zedong stated his belief that the power usurped by capitalist-roaders could be recaptured only by carrying out a great cultural revolution. + +The notice marks the beginning of a decade-long campaign that some historians +================================================================================ +Rank = 88; Score = 3948544.0 +<|begin_of_text|>All week, The Ringer will be celebrating Good Bad Movies, those films that are so terrible they’re endlessly amusing and — dare we say it? — actually good. Please join us as we give the over-the-top action movies, low-budget romance thrillers, and peak ’80s cheese-fests the spotlight they deserve. + +If we were to view all of the Good Bad Movies as parts and pieces that exist parallel to one another, adjacent to one another, in the same plane as one another, and they could snap together to form a kind of Good Bad Movie Universe, then who would be the actor who lorded over all of it? That’s a big question and a hard question because there are so, so, so many Good Bad Movies. But actually it’s a small question and an easy question because the answer is very clear: + +Nicolas Cage has acted in more than 80 movies during his career. Mind you, not all of them are Good Bad Movies. That is not the argument here; the quality of his movies covers a wide, big, broad range. Some of them are Actually Good Movies, some of them are Good Bad Movies, some of them are Bad Movies, some of them are Bad Good Movies, some of them are Bad Bad Movies, and some of them fall in the cracks between those categories. So no, the argument is not that every Nicolas Cage movie is a Good Bad Movie. The argument is that, in total, Nicolas Cage is the King of the Good Bad Movie Universe. + +Question: How does one become the King of the Good Bad Movie Universe, or even qualify for the title? + +Answer: In a previous article, the requirements for a Good Bad Movie were laid out as such: + +Enjoyment of the movie must be derived from its badness. Its badness needs to be the thing that creates a sense of bewildered enjoyment. + +There must be a pervading sense that those who made the film thought what they were doing was great or at least good. Good Bad Movies have minimal self-awareness. Here are two examples that may help explain this sentiment: 1) MacGruber is not a Good Bad Movie, it’s a tribute to Good Bad Movies, and 2) Fast Five is not a Good Bad Movie, it is a movie that intentionally wades into ridiculousness (and then manufactures a reaction similar to the one a Good Bad Movie elicits naturally). + +The movie must have been something of a critical failure when it was released. Critics, god bless them, hold movies +================================================================================ +Rank = 89; Score = 3948544.0 +<|begin_of_text|>Ravers were left shocked at Corsica Studios this week after a local techno DJ spontaneously combusted with rage after a member of the crowd referred to his set as ‘Tech-House’. + +The DJ, whose name was later revealed to be Dave Lester, 32, from Lewisham, had been playing a sub-standard selection of repetitive tunes when a member of the crowd proceeded to the DJ booth. The crowd member was then seen talking into Lester’s ear, at which point the convulsions began at once. + +“He started shaking uncontrollably straight away,” crowd member Jody Smith told us afterwards. “If I heard correctly, the dude leaned over and said to him, ‘I’m really enjoying your selection: top tech-house music.’ Then the DJ’s mouth started frothing with blood and foam, his eyes seemed to pop out of their sockets with anger, and his fists clenched as if he had suddenly become some hideous techno version of the Incredible Hulk. He then collapsed twitching onto the floor, screaming in a blood-curdling voice the words ‘Jeff Mills is not Tech-House!’ over and over again.” + +As the convulsions continued, onlooker Andrew Harris, who later identified himself as a friend of the deceased, was one of the first on the scene. + +“The sound technician has already jumped in and begun to administer CPR,” Harris told us, “but Dave simply squealed that the pumps to his chest were not being delivered in four to the floor time and physically threw him to one side. The sound technician responded by observing the fact that a techno tempo is too slow for effective CPR. It was at that point Dave simply exploded all over the concerned onlookers and over the decks.” + +Lester was taken to hospital immediately afterwards, but was confirmed dead upon arrival. Doctors issued the following statement: “We’ve known about this condition, known in the medical journals as Techno Snobbery, for some time now, but this is the first time that we’ve ever seen anyone spontaneously combust with rage due to it. + +“We would offer the following advice to techno DJs,” the statement continued. “Techno Snobbery is a condition that can often show no symptoms for many years, yet it has been documented that it can emerge suddenly at any time due to the slightest provocation. Therefore, techno DJs should remember that genre categorisations are largely permeable and superfluous, and that a fan referring to their music as Tech-House is therefore not a derogatory comment and can, in some circumstances +================================================================================ +Rank = 90; Score = 3948544.0 +<|begin_of_text|>A few days ago, I wrote about one of the dumbest supposedly major controversies in a while, which was one SEVENTEEN member supposedly telling a fan that she was basically his source of income and two others saying that she was a fan of another member. I figured that would be sorta it (because I am a dumbass) since it was either: 1) Fabricated 2) A joke 3) Telling the truth. However, of course things started to snowball because the accusation had to do with disrespect from an idol to a fan. + +SEVENTEEN fans ended up claiming it was fabricated because Seungkwan‘s signature looks different, but as far as I’m concerned there was nothing conclusive about that. Of course, the same applies to the accuser’s clarifications, which don’t actually prove anything either, yet people are taking her story as solid evidence for whatever reason. + +As far as I see it, we’re thus basically back to the battle of narratives. Yet the reason I’m skeptical of her side of the story is that there was a recording she made with Jeonghan that was supposed to make him some kind of dickhead. + +Fansite recording of Junghan + +Fan: Today’s dress code… + +Junghan: Stop coming here. + +Fan: (2 seconds pause) Huh…? OK.. But I keep getting picked though. + +Junghan: Then other fans can’t come. Stop coming. + +What an asshole, right? + +But the reality of that exchange was basically what I thought the Seungkwan incident was all about, which was idols joking around with fans who show up all the time. + +일 끝내자마자 음성찾아서 올려드립니다. + +통으로 잘랐으며 어떻게 하든 주작 소리 + +나오니깐 이거라도 주작이 아니라는걸 밝히기위해 변조없이 생으로 올려드립니다. + +(151101 용산팬싸)음성 듣고 알아서 판단하세요. 그리고 고소해 역고소할만큼 이게 팩트니깐. pic.twitter.com/p2l9vw230S — 햄니 (@96519kg) March 22, 2017 + +Everybody can have their own interpretation of that exchange, but I definitely see it as two people who recognize each other shooting the shit and not some malicious attack. And once that tone is established between them, it’s harder for me to believe Seungkwan wasn’t joking around +================================================================================ +Rank = 91; Score = 3899392.0 +<|begin_of_text|>First they said the downing of Russian Metrojet Flight 9268 was most likely due to Russia’s “notorious” regional airlines, which supposedly are rickety and unreliable. The Egyptian government denied that terrorism is even a possibility, with Egyptian despot Abdel Fatah al-Sisi proclaiming: + +“When there is propaganda that it crashed because of Isis, this is one way to damage the stability and security of Egypt and the image of Egypt. Believe me, the situation in Sinai – especially in this limited area – is under our full control.” + +However, it soon came out that the person in charge of Sharm el-Sheikh airport, where the Russia plane had landed before taking off again, had been “replaced” – oh, but not because of anything to do with the downing of the Russian passenger plane! As the Egyptian authorities put it: + +“Adel Mahgoub, chairman of the state company that runs Egypt’s civilian airports, says airport chief Abdel-Wahab Ali has been ‘promoted’ to become his assistant. He said the move late Wednesday had nothing to do with media skepticism surrounding the airport’s security. Mahgoub said Ali is being replaced by Emad el-Balasi, a pilot.” + +Laughable, albeit in a sinister way, and yet more evidence that something wasn’t quite right: after all, everyone knows the Egyptian government does not have the Sinai, over which the plane disintegrated in mid air, under its “full control.” ISIS, which claimed responsibility for the crash hours after it occurred, is all over that peninsula. + +Still, the denials poured in, mostly from US government officials such as Director of National Intelligence James “Liar-liar-pants-on-fire” Clapper, who said ISIS involvement was “unlikely.” Then they told us it couldn’t have been ISIS because they supposedly don’t have surface-to-air missiles that can reach the height attained by the downed plane. Yet that wasn’t very convincing either, because a) How do they know what ISIS has in its arsenal?, and b) couldn’t ISIS or some other group have smuggled a bomb on board? + +The better part of a week after the crash, we have this: + +“Days after authorities dismissed claims that ISIS brought down a Russian passenger jet, a U.S. intelligence analysis now suggests that the terror group or its affiliates planted a bomb on the plane. + +“British Foreign Minister Philip Hammond said his government believes there is a ‘significant possibility’ that an explosive device caused the crash. And +================================================================================ +Rank = 92; Score = 3899392.0 +<|begin_of_text|>The recent extraordinary runup and commensurate volatility in the price of Bitcoin has intensified global attention on the cybercurrency. For good or ill, Bitcoin is now a topic of interest to the mainstream press, as Bitcoin newbies near and far take notice. + +Such newbies and reporters alike have nowhere to turn for information on Bitcoin than the established cadre of Bitcoin fans (some would say fanatics) who helped to drive its price to the stratosphere in the first place. + +Just one problem: many of these fans are spouting misinformation – and worse, they believe their own nonsense. + +For you fanatics (you know who you are) as well as the broader newbie audience suddenly interested in Bitcoin, here are some of the lies Bitcoin fans tell themselves – and how to avoid getting snookered yourself. + +Lie #1: Bitcoin is Like Something Else + +This fallacy is so prevalent, frankly, because Bitcoin really isn’t much like anything else. There are some similarities between Bitcoin and, say, the US dollar or gold or even tulips, but the differences far outweigh the similarities. + +Drawing a false comparison between dissimilar things is an example of the false analogy, false comparison, or false equivalence fallacy, and there’s even a Wikipedia page that explains it. + +Here is a common example: when I say that criminals use Bitcoin for illicit transactions, a common response is ‘well, criminals use US currency for illicit transactions as well, so you shouldn’t single out Bitcoin.’ This line of reasoning is fallacious, because Bitcoin and US currency are too dissimilar to draw such a conclusion. + +Lie #2: Bitcoin is Secure + +Bitcoin is a cryptocurrency and ‘crypto’ means secure, right? Not so fast. Bitcoin has proven appallingly easy to steal, and even easier to lose. + +Security, after all, is like a game of whack-a-mole. Hit one vulnerability on the head and another pops up. There are still far too many moles all too happy to stick their heads through holes in the Bitcoin threat surface for my liking. + +Lie #3: Bitcoin is Money + +Take a $20 bill and put it under your mattress. Wait ten years and retrieve it. What is it worth? $20, of course. Sure, inflation will have likely reduced its value somewhat, but $20 will always be $20. Such is the nature of money. + +Not so with Bitcoin. It’s more of a commodity that acts as a speculative vehicle, while exhibiting few of the properties of money. + +At this point, Bitcoin fans +================================================================================ +Rank = 93; Score = 3883008.0 +<|begin_of_text|>South Park ending: After defeating Voldemort, Harry and Ron address the audience, saying, “You know, I’ve learned something today.” Suddenly, Ginny is run over by the Knight Bus, prompting Ron to shout, “They killed Ginny!” Harry responds, “You bastards!” Neville laughs and says it’s because Ginny was poor. + +Star Wars ending: Voldemort reveals himself to in fact be Harry’s father. + +CSI: Miami ending: Harry, Ron and Hermione collect evidence linking Voldemort to the murder of Albus Dumbledore. When confronted, Voldemort challenges Harry to prove he did it; Harry puts on his sunglasses, holding up a single strand of Voldemort’s hair, responds, “I don’t have to, you already did.” + +24 ending: It is revealed that Voldemort was really just working for the French, and Madame Maxime was really behind the entire plot to destroy the wizarding world. At the last second, Harry is able to diffuse the device Maxime had planted that would neutralize all wizard’s power in all of England. Just as they begin to celebrate, Harry is captured by the Chinese and when we last see him, he is on a boat to China. + +Brokeback Mountain ending: After wishing he could “quit him,” Harry finally finds out that Draco was beaten to death with wands by a group of angry Deatheaters. The book closes with Harry gazing longingly into his trunk, in which he has Draco’s Slytherin robe wrapped around his own Invisibility Cloak. Harry mumbles, “I swear, Draco.” and sheds a single tear. + +Lord of the Rings ending: Harry and Ron finally destroy the final horcrux by throwing it into a bubbling lave pit deep beneath Hogwarts, killing Lord Voldemort in the process, but at the same time weakening the foundations of Hogwarts, trapping Harry and Ron miles beneath the surface. where they gaze into each others eyes, too afraid to talk of the love that dare not speak it’s name. Just as they pass out, Fawkes rescues them and they both live out their lives married to their respective spouses + +A Few Good Men ending: In the process of defeating Lord Voldemort, Ron and Hermione accidentally kill Draco. They are brought to trial before the Wizengamut, and Harry must act as their lawyer. Harry believes that Ron and Hermione were only following orders from Percy, and so he calls Percy to the stand. After a few hours of intense questioning, Harry finally tricks Percy into admitting he ordered the attack on Malfoy. As +================================================================================ +Rank = 94; Score = 3883008.0 +<|begin_of_text|>Tech / Help + +Deep dive: What makes the Demon so wicked? + +by Patrick Rall + +At a small Detroit media event, SRT powertrain boss Chris Cowland laid out all of the fine details of the 2018 Dodge Challenger SRT Demon. + +The 2018 Challenger Demon had two simple guidelines during development – to run a 9-second quarter mile and to lift the front wheels off of the ground on launch, both in street legal trim. + +Once the Demon was given the green light, Dodge added a few initial requirements, including an engine power increase of 10% over the Hellcat Hemi, and launch forces at least 20% greater than the standard Hellcat Challenger. At the same time, the Demon had to be emission-legal in all 50 states (meeting LEV160), meet pass-by noise requirements, and pass strict SRT durability testing. + +The result is a 4,200 pound muscle car which will run a 9.65 quarter mile and lift the front wheels off of the ground on a hard launch with an 840 horsepower Hemi – which meets all emission and noise requirements set for by the industry and the company. Most importantly, the new Demon is 100% street-friendly, so in addition to being the quickest production car in the world, it can also serve as a comfortable daily driver. + +This was achieved by introducing a new version 6.2-liter supercharged Hemi, with 62% of the content changed from the original Hellcat Hemi — including a new block, crankshaft, connecting rods, pistons, and supercharger assembly. + +That 840 horsepower Hemi is mated to a new 8-speed automatic transmission which has been heavily upgraded with a new torque convertor and trans brake. From that stronger transmission, the power travels through a stronger driveshaft to a stronger rear differential with a steeper gear ratio and stronger axles – reaching the ground by means of the lightweight 18” wheels and the Demon-branded Nitto NT05R tires. + +We knew all of that when the car was introduced back in April, but below, you will find a detailed rundown of how each key component of the drivetrain in the 2018 Dodge Challenger SRT Demon differs from the Hellcat Hemi. + +Induction system + +The Air Grabber Hood is a functional hood scoop system which feeds a high volume of cold air to the airbox, along with dual air catchers in the grille (replacing the fog lights) — all feeding the 14. +================================================================================ +Rank = 95; Score = 3817472.0 +<|begin_of_text|>The flag decals that were ordered taken down by Chief Bronaugh. Photographs courtesy Maywood Fire Department. + +Tuesday, September 9, 2014 || By Michael Romain + +Today, four Maywood firefighters were sent home, pending disciplinary action, for failure to comply with a direct order, according to Chief Craig Bronaugh. A CBS Local report noted that more firefighters could be sent home tomorrow. The firefighters apparently refused to remove American flag decals from their lockers and helmets. + +“A representative from Service Employees International Union Local 73, which represents the firefighters, met for an hour with Chief Bronaugh Tuesday afternoon, with the union hoping Chief Bronaugh would modify his order. Instead, he ordered the firefighters sent home immediately, pending disciplinary action,” according to the CBS report. + +“Union spokesman Adam Rosen said he is ‘shocked’ that an agency of first responders would enforce such an order the week of Sept. 11. + +“The four suspended firefighters said they were told that the order was issued because of racial discord [in] the department. The four, who include two white firefighters, a black firefighter, and a fourth firefighter who is a Cuban émigré, said no such problems exist,” wrote CBS, which also reported that the four firefighters trace the issue “to a decision by several firefighters to replace a tattered American flag last month in one of Maywood’s firehouses. The new flag mysteriously disappeared early Aug. 23. The order on decals was issued last week.” + +Although the CBS report indicates that the department “has had no such order in the recent past,” an internal memo released by the Office of Chief Bronaugh noted that this kind of conflict has nonetheless appeared before: + +Under no circumstances at anytime are there to be decals/stickers on Fire Department lockers or helmets. Any decals/stickers are to be removed immediately. No inquiry will be entertained regarding this matter. + +TO BE READ AT ALL ROLL CALLS THRU SEPTEMBER 10, 2014 + +This was in response to an inquiry regarding an American Flag being put up in the fire station (replacing a torn and tattered flag that previously existed). That flag was stolen on the very next shift day. Firefighters were upset that the flag was stolen and in response placed American Flag decals on their personal lockers in the stations. + +This is not the first time that a flag incident has occurred in the Maywood Fire Department. Approximately 10 years ago, all American Flag decals were removed from all of the department vehicles and +================================================================================ +Rank = 96; Score = 3768320.0 +<|begin_of_text|>$\begingroup$ + +Sam sat with his eyes closed for several minutes, then said softly: + +"I have many names, and none of them matter." He opened his eyes slightly then, but he did not move his head. He looked upon nothing in particular. + +"Names are not important," he said. "To speak is to name names, but to speak is not important. A thing happens once that has never happened before. Seeing it, a man looks on reality. He cannot tell others what he has seen. Others wish to know, however, so they question him saying, 'What is it like, this thing you have seen?' So he tries to tell them. Perhaps he has seen the very first fire in the world. He tells them, 'It is red, like a poppy, but through it dance other colors. It has no form, like water, flowing everywhere. It is warm, like the sun of summer, only warmer. It exists for a time on a piece of wood, and then the wood is gone, as though it were eaten, leaving behind that which is black and can be sifted like sand. When the wood is gone, it too is gone.' Therefore, the hearers must think reality is like a poppy, like water, like the sun, like that which eats and excretes. They think it is like to anything that they are told it is like by the man who has known it. But they have not looked upon fire. They cannot really know it. They can only know of it. But fire comes again into the world, many times. More men look upon fire. After a time, fire is as common as grass and clouds and the air they breathe. They see that, while it is like a poppy, it is not a poppy, while it is like water, it is not water, while it is like the sun, it is not the sun, and while it is like that which eats and passes wastes, it is not that which eats and passes wastes, but something different from each of these apart or all of these together. So they look upon this new thing and they make a new word to call it. They call it 'fire.' + +"If they come upon one who still has not seen it and they speak to him of fire, he does not know what they mean. So they, in turn, fall back upon telling him what fire is like. As they do, they know from their own experience that what they are telling him is +================================================================================ +Rank = 97; Score = 3768320.0 +<|begin_of_text|>Question Answer Episode Marshall hits Lily with what after he proposes? Pilot Thing Ted pretends he is buying at the store? Purple Giraffe Barney's luggage contains? Sweet Taste of Liberty Name one of the things Ted fondly remembers about Natalie. Return of the Shirt Barney says that there are 24 similarities between women and what? Okay Awesome The Slutty Pumpkin's drink? Slutty Pumpkin Unusual creature in Ted and Marshall's apartment? Matchmaker Place Lily's apartment has been turned into? The Duel Sport Marshall's family has created? Belly Full of Turkey Name of the shot that Carl creates? The Pineapple Incident 'Celebrity' that joins the gang on New Years? The Limo First thing that Marshall and Lily agree about for the wedding? The Wedding Fake name Victoria uses? Drumroll, Please 'I want to know you, like _____ _____ _____.' Zip, Zip, Zip What Barney refers to Ted's embarrassing story as? Game Night Where is Marshall's dream internship? Cupcake The song Marshall and Lily sing? Life Among the Gorillas Who do Marshall, Lily, and Barney hang out with? Nothing Good Happens After 2 AM Cutout Ted and Marshall use on Sandy? Mary the Paralegal What Scooter wants to be? Best Prom Ever What Robin shows Ted to cheer him up? Milk What Ted does to stop Robin's camping trip? Come On The pet George Clinton buys Lily? Where Were We? What Ted orders to cure his hangover? The Scorpion and the Toad What Ted and his dad talk about if things get uncomfortable? Brunch Movie Ted is mad that Robin dislikes? Ted Mosby, Architect Concert Marshall and Brad go to? World's Greatest Couple Who plays the Cougar? Aldrin Justice Show the end of the episode reference? Swarley What Barney has to find to win the game? Atlantic City Name of Robin's hit single? Slap Bet What James does that reveals he's in a relationship? Single Stamina Word that Ted calls Lily? How Lily Stole Christmas Where the gang takes Robin's sister? First Time in New York Name of Ted's annoying ex-boss? Columns Who Barney bumps into on the street? Monday Night Football Where is Ted's interview? Lucky Penny Name of Barney's one-man show? Stuff Name of the song that the Fiero plays? Arrivederci, Fiero One of the names that Barney gives the truck in his Top Ten List Moving Day What Lily's grandmother gives her at the bridal shower? Bachelor Party Show Barney appears on? Showdown +================================================================================ +Rank = 98; Score = 3719168.0 +<|begin_of_text|>Rodrigo Varela/ESPN Images Have you ever seen Stugotz at Hooters? Have you ever seen Stugotz at Hooters? + +Midwest Region + +#1) Kentucky: Jeff Van Gundy - Queen of Hearts + +#2) Kansas: George Karl- Leader of a nudist colony + +#3) Notre Dame: Nene - Looks like a gladiator that will help you slay a tiger then join you as you embark on a quest + +#4) Maryland: J.J. Redick - Sketchy car valet who might take your car for a joy ride + +#5) West Virginia: Mike Dunleavy Jr. - Looks like a generic police sketch + +#6) Butler: Andy Reid - Looks like he waggles his fingers in front of a tray of doughnuts and says, "Don't mind if I do" + +#7) Wichita State: Marcin Gortat - Guy who becomes a YouTube sensation by wrestling bears shirtless + +#8) Cincinnati: Kris Humphries - Looks like a male cheerleader + +#9) Purdue: Russell Wilson - Looks like a male cheerleader + +#10) Indiana: Jerry Sloan - Looks like he washes his hair with a bar of soap + +#11) Texas: David Shaw - Looks like the president in a cable television network drama + +#12) Buffalo: Nick Saban - Guy who runs a lap, looks at his stopwatch and says, "Still go it," while snapping his fingers + +#13) Valparaiso: Frank Vogel - Guy who keeps calling you to hang out and you keep creating excuses not to go + +#14) Northeastern: Trey Wingo - Looks like a guy who owns a funeral home and does late-night infomercials promoting his seasonally discounted rates + +#15) New Mexico State: DeAndre Jordan - Looks like a cartoon moose + +#16) Hampton: Chip Kelly - Looks like the guy who leaves comically low tips to service people, then shoots the finger gun and says, "Don't spend it all in one place" + +#16) Manhattan: Chip Kelly - Looks like the guy who washes his yacht shirtless + +West Region + +#1) Wisconsin: Ron Rivera - Guy who wears a lei for his entire vacation in Hawaii + +#2) Arizona: Jack Del Rio - Stepdad who tries too hard to be called dad + +#3) Baylor: Orel Hershiser - Looks like the father in the picture of +================================================================================ +Rank = 99; Score = 3719168.0 +<|begin_of_text|>In an effort to encourage the development of Canadian programming that's more accessible to the iPod generation, the federal government is doing away with the controversial Canadian Television Fund. The fund, which is the source of taxpayer subsidies to the Canadian television production industry, will be combined with the Canada New Media Fund, which subsidizes digital media development starting April 1, 2010. + +Online scams have now surpassed phone scams as a source of fraud. ( Shutterstock ) + +The government will invest $134.7 million annually in the new program, called the Canada Media Fund, Heritage Minister James Moore said at a press conference today on the Toronto set of the CTV police drama series, Flashpoint. "We are levelling the playing field at a time when the industry is undergoing structural change," Moore said. "The eligibility for funding for broadcaster-affiliated production will be expanded, and broadcaster in-house production will be allowed... including provincial educational broadcasters and CBC/Radio-Canada." + +Article Continued Below + +The emphasis of the new broadcast policy will be on drama, comedy, and children's programming, and will favour productions developed for distribution platforms other than prime-time television – the Internet and mobile telecommunications devices, said Moore, who admitted he watches "more television programs on my iPod than on regular TV." The new fund will also support documentaries and variety and performing arts programming that can demonstrate that the market alone would not support their creation. "The realigned fund will favour projects produced in high definition and those that have achieved and demonstrated the most potential to achieve success, in terms of audience and return on investment," Moore said "Applicants will be required to make their projects available across a minimum of two distribution platforms, including television." The establishment of the new fund is a clear rejection of last summer's recommendation by the broadcasting industry czar, the Canadian Radio-television Telecommunications Commission, that the $300-million Canadian Television Fund be split into two funding streams, one for private networks, the other for non-profit broadcasters.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> \ No newline at end of file diff --git a/examples/openwebtext/files/database.txt b/examples/openwebtext/files/database.txt new file mode 100644 index 0000000..c11e132 --- /dev/null +++ b/examples/openwebtext/files/database.txt @@ -0,0 +1,2212 @@ +Query Sequence: +Prompt: Using a distributed database has many advantages.; Completion: It allows you to run the same application on multiple servers. This makes it easier to scale your application. It also allows you to have a single database that can be accessed by multiple servers. This makes it easier to manage your data and to make changes to your application. + +Top Influential Sequences: +================================================================================ +Rank = 0; Score = 20971520.0 +<|begin_of_text|>Bigfoot has been a staple of American folklore since the 19th century. Many people are convinced that Bigfoot is real. Others suggest that he is a cultural phenomenon. Some just want to believe. There is even a group, the Bigfoot Field Researchers Organization, that tracks Bigfoot sightings. And they have thousands of reports available online that date back to the late 19th century. The Internet, it seems, has everything. So, I took this data, all 4,586 records of it, and used it to build a classifier. It was a good model with pleasing metrics. I liked it. But then what? For some folks, the model is where the work ends. But I'm a developer and that's only half the solution. I've got a model but how do I use it? How do I put it in an application so that a user can, well, use it? I'm going to answer that question in this talk, and a bit more. I'll show you how I exposed my Bigfoot classifier to the Internet as a REST-based API written in Python. And we'll tour a couple of applications I wrote to use that API: a web-based application written in JavaScript and an iOS application written in Swift. For the model itself, I'll use DataRobot since it's quick and easy. And, I work there! When we're done, you'll know how to incorporate a model into an API of your own and how to use that API from your application. And, since all my code is on GitHub, you'll have some examples you can use for your own projects. As a bonus, you'll have 4,586 Bigfoot sightings to play with. And who doesn't want that? **************************************************************** Guy started his career as gasp a COBOL programmer. But don’t hold that against him, it just gives him perspective. He has spent much of his time programming in the most popular of the semi-colon delimited languages including C++, Java, and JavaScript. More recently he has been working with Python and machine learning. In addition to programming computers, Guy had a background in electronics and enjoys building circuits, burning himself with a soldering iron, and programming small hardware devices such as the Arduino and the Raspberry Pi. No one has actually paid him to do this sort of work… yet. Guy loves to speak and teach and will go to any conference that will give him an audience and teach anyone who wants to learn. He normally speaks about technology but has been known to +================================================================================ +Rank = 1; Score = 15663104.0 +<|begin_of_text|>Lady Gaga is an American singer, songwriter, and actress who has received many awards and nominations for her contributions to the music industry. She rose to prominence with the release of her debut album The Fame in 2008. The album won several awards and was nominated for six Grammy Awards, including Album of the Year. The album and its single "Poker Face" won Best Electronic/Dance Album and Best Dance Recording, respectively, at the 52nd Grammy Awards. The album also won International Album at the 2010 BRIT Awards. A follow-up EP, titled The Fame Monster, was released in 2009, and included the singles "Bad Romance" and "Telephone". The music videos of the songs won eight accolades from thirteen nominations at the 2010 MTV Video Music Awards (VMA), making Gaga the most nominated artist in VMA history for a single year and the first female artist to receive two nominations for Video of the Year in one single night.[1] In 2011, Gaga was nominated for six Grammy Awards, and won three—Best Pop Vocal Album for The Fame Monster, and Best Female Pop Vocal Performance and Best Short Form Music Video for "Bad Romance". + +Born This Way (2011), Gaga's second studio album, accrued three nominations at the 54th Annual Grammy Awards, including her third consecutive nomination for Album of the Year. It won the People's Choice Awards for Album of the Year the following year, and the music video for the title track won two VMAs, including Best Female Video. Her third album, Artpop (2013), won the award for World's Best Album by a Female at the 2014 World Music Awards. In other musical ventures, Gaga released a collaborative jazz album with Tony Bennett, titled Cheek to Cheek (2014), which received the Grammy Award for Best Traditional Pop Vocal Album. + +In 2015, Gaga released a song "Til It Happens to You", for the documentary film, The Hunting Ground. The song won a Satellite Award for Best Original Song, while nominated for an Academy Award, a Critics' Choice Movie Award for Best Song, and a Grammy Award for Best Song Written for Visual Media. In the same year she was named Woman of the Year by Billboard. She also became the first woman to receive the Digital Diamond Award from RIAA,[2] and the first artist to win the Songwriters Hall of Fame's Contemporary Icon Award for "attaining an iconic status in pop culture". Gaga received a Golden +================================================================================ +Rank = 2; Score = 15138816.0 +<|begin_of_text|>If you want to make money playing poker, you have to read your opponents without letting them read you. + +Stephen Harper sure is hard to read, but in the game of high-stakes poker over the election debates, he seems to have misread the other players. + +On Thursday, representatives of the NDP, Liberals, Greens and Bloc agreed to do one English and one French debate to be broadcast by CTV, CBC, Global and Radio Canada, which would reach millions of Canadian TVs. + +The Conservatives say they’re too busy with other debates. So far, they have agreed to debates put together by TVA, Maclean’s, the + +Globe and Mail and the Munk Debates. La Presse is likely next. + +The problem for the Tories is that the consortium of big broadcasters controls TV bandwidth, and they are refusing to cede control to the other media organizations. + +When British Prime Minister David Cameron stayed away from a similar debate recently, the British consortium went ahead without him. That didn’t help Cameron’s rival, Labour Leader Ed Miliband, because he got beat up by the leaders of smaller parties. + +The dynamic would be different here because many voters would like to have a different prime minister, but they disagree on whether it should be Thomas Mulcair or Justin Trudeau. + +If the two of them square off (with Elizabeth May) in the only English debate to be broadcast on millions of TV screens, the non-Tory voters might settle on one of them. If that happens, Harper would find it hard to keep his job. + +The numbers do not look good anyway. Friday’s Ekos poll for iPolitics shows the Tories in a three-way tie with the NDP and the Liberals. + +Worse still for him, he has little room to grow. About 40% of both Liberal and NDP supporters identify the other party as their second choice but only 13% of Liberal supporters and 8% of NDP supporters would consider voting Conservative. And 58% of Canadians would not consider voting Conservative under any circumstances. + +“If our numbers are correct, the Conservatives are in deep trouble,” said pollster Frank Graves. “They’ve got very little room for growth, and they’re 12 points back … from where they were election night.” + +A security crisis might move that number, but the “balanced” budget, anti-terror talk and tax cuts haven’t done the trick. + +A big chunk of voters — Graves calls them “promiscuous progressives” — want to see the back of Harper. If the biggest debate of the campaign takes +================================================================================ +Rank = 3; Score = 10289152.0 +<|begin_of_text|>Lottery mathematics is used to calculate probabilities of winning or losing a lottery game. It is based heavily on combinatorics, particularly the twelvefold way and combinations without replacement. + +Choosing 6 from 49 [ edit ] + +In a typical 6/49 game, each player chooses six distinct numbers from a range of 1-49. If the six numbers on a ticket match the numbers drawn by the lottery, the ticket holder is a jackpot winner—regardless of the order of the numbers. The probability of this happening is 1 in 13,983,816. + +The chance of winning can be demonstrated as follows: The first number drawn has a 1 in 49 chance of matching. When the draw comes to the second number, there are now only 48 balls left in the bag, because the balls are drawn without replacement. So there is now a 1 in 48 chance of predicting this number. + +Thus for each of the 49 ways of choosing the first number there are 48 different ways of choosing the second. This means that the probability of correctly predicting 2 numbers drawn from 49 in the correct order is calculated as 1 in 49 × 48. On drawing the third number there are only 47 ways of choosing the number; but of course we could have arrived at this point in any of 49 × 48 ways, so the chances of correctly predicting 3 numbers drawn from 49, again in the correct order, is 1 in 49 × 48 × 47. This continues until the sixth number has been drawn, giving the final calculation, 49 × 48 × 47 × 46 × 45 × 44, which can also be written as 49! ( 49 − 6 )! {\displaystyle {49! \over (49-6)!}} or 49 factorial divided by 43 factorial. This works out to 10,068,347,520, which is much bigger than the ~14 million stated above. + +However; the order of the 6 numbers is not significant. That is, if a ticket has the numbers 1, 2, 3, 4, 5, and 6, it wins as long as all the numbers 1 through 6 are drawn, no matter what order they come out in. Accordingly, given any set of 6 numbers, there are 6 × 5 × 4 × 3 × 2 × 1 = 6! or 720 +================================================================================ +Rank = 4; Score = 8585216.0 +<|begin_of_text|>On July 11, 2016, 6 months after the first published version of Zappa, Amazon announced the preview release of Chalice, now the Python Severless Microframework for AWS. This was a fairly big shock to our community at the time, as it looked as though our Free and Open Source project was being undercut by Amazon's propietary offering. Not only that, but it felt like their interface and even their presentation of the product was a̶ ̶d̶i̶r̶e̶c̶t̶ ̶r̶i̶p̶-̶o̶f̶f̶ ̶o̶f̶ inspired by our efforts, potentially designed to ~~~trick~~~ help consumers and lock them into the AWS ecosystem. + +I don't want to attribute any direct malice there, (this is probably just a case of a good idea being executed in multiple ways), but Amazon does have a fairly long history of building their own proprietary or locked-in versions of software products and services built on top of AWS infrastructure. + +So, now that both products have matured even further, I think it's time that we do a comparison of the two frameworks. Obviously, as one of the authors of Zappa I'm strongly biased, so take my opinions for what they are, but it is my strong belief that although Chalice has some interesting features, Zappa is the vastly superior product for these important reasons: + +Zappa does not lock you in to the AWS ecosystem. + +to the AWS ecosystem. Zappa has vastly more useful features. + +. Zappa is battle-tested, used in production by medical, scientific and banking users. + +Against Vendor Lock-In + +This is the big headline here: Chalice locks you into using AWS services - Zappa doesn't. + +With Zappa, you can run existing Python web applications on AWS Lambda thanks to the magic of WSGI. More importantly, you can leave AWS at any time, meaning you can run the same application on any web server your like, be it VPS, on Heroku, or on your own bare metal. (You will lose some of the event-driven benefits which AWS can provide, but you will always have the option to use other replacements if that's what you decide.) + +With Chalice, you will have to re-write your application to use the Chalice framework, and after that it will only ever be able to be run on AWS infrastructure. + + +================================================================================ +Rank = 5; Score = 8519680.0 +<|begin_of_text|>Renting Dumpsters Posted on December 5, 2018 | By STEVE Whenever people have a task, it is easy to generate waste and piling up garbage can be a main source of worry. This is not only with regards to taking up space but also in conditions of polluting the environment unnecessarily. Therefore, how perform these worries are taken by you aside? This can be done by opting to use 20 yard dumpsters philadelphia simply. These are offered by local rental businesses and are designed to alleviate the worries of coping with trash from your brain. There are many of these rental companies in the marketplace and as such, it ought not to be difficult so that you can get the same. The local rental businesses will make sure that you recycle your waste in an eco-friendly method and at the same period, make the process easy. By using these eco-friendly dumpsters, you not really just ensure that your encircling is clean and gives you the ideal life-style but it will also make sure that you enjoy great wellness. As a result, this will give you great endurance and boost your probabilities of taking pleasure in a treatment free of charge way of living. These can be utilized in different fronts and places such as churches, house owners, areas and additional businesses. For this good reason, they are known to provide all round services to all these combined groups. There are instances when you might have dangerous waste products and this provides the ideal chance that you can ensure that you get rid of the same in a secure way. Dumpsters may end up being availed in various designs seeing that good as sizes and this makes it all easy for individuals to help to make buys based on their requirements. Another main benefit connected with this type of waste materials disposal is usually the truth that the consumer can place it in an area of their choice. This makes it simple to make sure that the clean up process is simple and effective on your part as well. It is normally essential to take note that the rental companies are also accountable for making sure that they are eliminated from your property once you are completed using the same and they can adhere to your time routine. What is more, in many situations, these ongoing companies will adhere to certain safety procedures when they are putting the same on site. At this true point, it is important to note that dumpsters that are provided by professional local rental businesses are also the perfect way to eliminate waste that is produced on building sites whether it is green backyard, commercial or home keep waste. What is usually even more, this provides you +================================================================================ +Rank = 6; Score = 8519680.0 +<|begin_of_text|>Today, we are announcing the general availability of App Service Isolated, which brings the simplicity of multi-tenant App Service to the secure, dedicated virtual networks powered by App Service Environment (ASE). + +Azure App Service is Microsoft’s leading PaaS (Platform as a Service) offering hosting over 1 million external apps and sites. It helps you build, deploy and scale web, mobile and API apps instantaneously without worrying about the underlying infrastructure. It allows you to leverage your existing skills by supporting an increasing array of languages, frameworks, and popular OSS, and has built-in capabilities that streamline your CI/CD pipeline. ASE was introduced in 2015 to offer customers network isolation, enhanced control, and increased scale options. The updated ASE capabilities that comes with the new pricing tier, App Service Isolated, now allow you to run apps in your dedicated virtual networks with even better scale and performance through an intuitive user experience. + +Streamlined scaling + +Scaling up or out is now easier. The new ASE eliminates the need to manage and scale worker pools. To scale, either choose a larger Isolated plan (to scale up) or add instances (to scale out), just like the multi-tenant App Service. It’s that easy. To further increase scaling flexibility, App Service Isolated comes with a maximum default scale of 100 Isolated plan instances. You now have more capacity for large implementations. + +Upgraded performance + +The new ASE uses dedicated Dv2-based machines boasting faster chipsets, SSD storage, and twice the memory per core when compared to the first generation. The dedicated worker sizes for the new ASE are 1 core with 3.5 GB RAM, 2 cores with 7 GB RAM, and 4 cores with 14 GB RAM. With this upgraded infrastructure, you will be able to run your apps with lower latency, have more power to handle heavier workloads, and support more users. + +Simplified experience + +Creating the new ASE is easy. By selecting the Isolated pricing tier, App Service will create an App Service Plan (ASP) and a new ASE directly. You will just need to specify the Virtual Network that you want to deploy your applications to. There is no separate workflow required to spin up a new ASE in your secure, dedicated virtual networks. + +We’ve made App Service Environment (ASE) faster, more efficient, and easier to deploy into your virtual network, enabling you to run apps with Azure App Service at high scale in an isolated network environment. Check out the Azure Friday video. Partners and customer can also +================================================================================ +Rank = 7; Score = 8323072.0 +<|begin_of_text|>Only recently the Cockpit project was launched, aiming at providing a web based management interface for various servers. It already leaves an interesting impression for simple management tasks – and the design is actually well done. + +I just recently came across the only three month old Cockpit project. The mission statement is clear: + +Cockpit is a server manager that makes it easy to administer your GNU/Linux servers via a web browser. + +The web page also states three aims: beginners friendly interface, multi server management – and that there should be no interference in mixed usage of web interface and shell. Especially the last point caught my attention: many other web based solutions introduce their own magic, thus making it sometimes tricky to co-administrate the system manually via the shell. The listed objectives also make clear that cockpit does not try to replace tools that go much deeper into the configuration of servers, like Webmin, which for example offers modules to configure Apache servers in a quite detailed manner. Cockpit tries to simply administrate the server, not the applications. I must admit that I would always do such a application configuration manually anyway… + +The installation of Cockpit is a bit bumpy: besides the requirement of tools like systemd which limits the usage to only very recent distributions (excluding Ubuntu, I guess) there are no packages yet, some manual steps are required. A post at unshut.me highlights the necessary steps for Fedora which I followed: in includes installing dependencies, setting firewall rules, etc. – and in the end it just works. But please note, in case you wanna give it a try: it is not ready for production. Not at all. Use virtual machines! + +What I did see after the installation was actually rather appealing: a clean, yet modern web interface offering the most important and simple tasks a sysadmin might need in a daily routine: quickly showing the current health state, providing logs, starting and stopping services, creating new users, switching between servers, etc. And: there is even a working rescue console! + +Main login view Default server overview, showing the possible tasks as well as the current health of the system. Default server view in wide screen browser Detail view of host name and system basics Network traffic – a click onto on of the health statistics brings up a more detailed view. Listing services – here it becomes obvious how much Cockpit actually is build upon systemd. The log watcher accesses the systemd journal – once again a reason why systemd is a requirement for Cockpit. Adding a new server to Cockpit. Two servers are registered and can be accessed in +================================================================================ +Rank = 8; Score = 7897088.0 +<|begin_of_text|>Share The Latest News + +Earlier this year LG, in partnership with Valve, showcased a prototype of their upcoming LG VR headset at GDC 2017 which is now rumored to be called UltraGear. So far we know that the upcoming headset will boast some new specs that differ from current headset manufacturers. + +LG UltraGear VR Headset Specifications: + +Two panels (one for each eye) with a resolution of 1440 x 1280 each + +OLED display from LG + +3.64 inches diagonal + +90 Hz refresh rate + +110 degree FOV + +With new reports from Letsgo Digital, we have discovered that LG has filed new patents that might greatly change how users will put on and take off the upcoming LG VR headset. Through the images seen in the patent, it clearly outlines how the headset can be opened in the middle which allows users to rest it on their shoulders if they so choose. + +This makes it easier to take quick rests from VR and also makes it easier to maneuver around if you had to take off the headset for any reason. We have not yet seen any major headset manufacturer do this where the HMD can be split down the middle. + +In addition, there will be a knob on the back which will allow you to control the sizing of the headset for your head which was a great feature added on the PSVR and the new HTC Vive deluxe head straps. + +There will also be an area where it looks like the headset will allow you dock your earbuds on the back. We aren’t quite sure how this works but it looks as though that it pulls in and out. + +You can find the rest of the images here for the recently filed patents. This does not mean that this will be final but it’s quite an interesting design. It might also be possible that LG will upgrade the currently known specs such as the display panels and FOV.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 9; Score = 7602176.0 +<|begin_of_text|>Will White of Mapbox shared the following guest post with me. In the post, Will describes how they use EC2 Spot Instances to economically process the billions of data points that they collect each day. + +I do have one note to add to Will’s excellent post. We know that many AWS customers would like to create Spot Fleets that automatically scale up and down in response to changes in demand. This is on our near-term roadmap and I’ll have more to say about it before too long. + +— Jeff; + +The largest automotive tech conference, TU-Automotive, kicked off in Detroit this morning with almost every conversation focused on strategies for processing the firehose of data coming off connected cars. The volume of data is staggering – last week alone we collected and processed over 100 million miles of sensor data into our maps. + +Collecting Street Data + +Rather than driving a fleet of cars down every street to make a map, we turn phones, cars, and other devices into a network of real-time sensors. EC2 Spot Instances process the billions of points we collect each day and let us see every street, analyze the speed of traffic, and connect the entire road network. This anonymized and aggregated data protects user privacy while allowing us to quickly detect road changes. The result is Mapbox Drive, the map built specifically for semi-autonomous driving, ride sharing, and connected cars. + +Bidding for Spot Capacity + +We use the Spot market to bid on spare EC2 instances, letting us scale our data collection and processing at 1/10th the cost. When you launch an EC2 Spot instance you set a bid price for how much you are willing to pay for the instance. The market price (the price you actually pay) constantly changes based on supply and demand in the market. If the market price ever exceeds your bid price, your EC2 Spot instance is terminated. Since spot instances can spontaneously terminate, they have become a popular cost-saving tool for non-critical environments like staging, QA, and R&D – services that don’t require high availability. However, if you can architect your application to handle this kind of sudden termination, it becomes possible to run extremely resource-intensive services on spot and save a massive amount of money while maintaining high availability. + +The infrastructure that processes the 100 million miles of sensor data we collect each week is critical and must always be online, but it uses EC2 Spot Instances. We do it by running two Auto Scaling groups, a Spot group and an On-Demand group, that share a single Elastic Load +================================================================================ +Rank = 10; Score = 7405568.0 +<|begin_of_text|>In the last year, Progressive Web Apps have become an incredibly popular way to build next generation apps. PWAs bring many advantages to the table, and in today’s post I’d like to share why we think you should care. + +Why should I care about PWAs? + +PWAs bring unique advantages to the table for developers building consumer facing apps. They allow developers to completely skip the app stores and instead simply deploy to a web server. This allows you to get your app out faster and to more people than you would be able to with deploying to the App Store and Google Play. Also, because you are not tied to the app store, this means that updates can be immediately available to everyone using your app, by simply pushing your new code to your web server. And because PWAs run in the browser, your single PWA can then be reached from just a click on a URL, greatly reducing the barrier to entry to your app! + +PWAs also bring a ton of new advantages to the table for businesses and enterprises. Because of the fact that with a PWA you don’t have to package your app for any specific native platform, or submit that app to any native app store, your time to market can be much faster. Just push to a web server and your PWA is available to everyone. + +For users of your app, PWAs offer a consistent experience no matter what device they are using, or how slow or flaky their network connection is. There’s no need to sit and wait while your app they are wanting to use is downloading and installing. Need something in a hurry? There is nothing much faster than tapping a URL. Also, because PWAs tend to be much smaller than the average native app, your app will not be taking up much of the precious storage on a user’s device. + +How to get started today + +The first concern I normally hear from devs interested in building a PWA is that Safari, the preinstalled browser on Apple devices, does not have support for PWAs. To be frank, this actually doesn’t matter because of progressive enhancement. With a few extra lines of code you can assure that users on browsers that do support those API’s get the full experience, and users on safari still get a great,fast experience. Also, Webkit, the browser engine behind Safari, recently announced they were implementing service workers, a key api for PWAs, so these API’s are going to be coming to Safari soon anyway. + +Ionic fully supports PWAs out of the box. Our starters provide the minimum things +================================================================================ +Rank = 11; Score = 7208960.0 +<|begin_of_text|>Data Warehousing at Wayfair + +In 2009 Wayfair’s database infrastructure was based almost entirely on Microsoft SQL Server. Our Business Intelligence team was using a SQL Server data warehouse to prepare a large amount of data for import into Analysis Services (SSAS) each day. We populated our data warehouse using transaction log shipping from production servers, which required about 3 hours of downtime on the data warehouse at midnight each night to restore the previous day’s logs. Once that was done, a series of stored procedures were kicked off by jobs that would crunch through data from several different servers to produce a star schema that could be pulled into SSAS. Wayfair was scaling rapidly, and this approach started to become painfully slow, often taking 10-16 hours to crunch through the previous day’s data. + +The BI team decided to look into other solutions for data warehousing, and ultimately purchased a Netezza appliance. Netezza is essentially a fork of PostgreSQL that takes a massively parallel cluster of nodes (24 in our case) and makes them look like one database server to the client. In our tests, Netezza could crunch through our data in roughly a quarter of the time, bringing 10-16 hours down to a much more reasonable 2-4 hours. The dream of updating our data warehouse multiple times each day was starting to look feasible. The feedback loop on business decisions would become dramatically shorter, enabling us to iterate more quickly and make well informed decisions at a much faster pace. There was just one glaring problem. + +Great, But How Are We Going to Get Data Into It? + +As soon as the DBA team heard that the Netezza purchase had been finalized, our first question was “great, but how are we going to get data into it?” The folks at Netezza didn’t have an answer for us, but they did send us an engineer to help devise a solution. As it turned out, the problem of how to incrementally replicate large amounts of data into a data warehouse was a common one, and there were surprisingly few open source solutions. Google it, and most people will tell you that they just reload all their data every day, or that they only have inserts so they can just load the new rows each day. “Great, but what if you want incremental replication throughout the day? What if you have updates or deletes? How do you deal with schema changes?” Crickets. + +The First Solution + +The solution we arrived upon was to use SQL Server Change Tracking to keep track of which rows had +================================================================================ +Rank = 12; Score = 6586368.0 +<|begin_of_text|>More Info + +Windows 10 + +Microsoft Windows 10 combines the best elements of the Windows you already know, like the Start menu, with new features like space to pin your favorite apps and easy navigation so you'll feel right at home. Getting set up is faster and easier so you can get right to work or exploring the Windows App Store games or a new version of Office. InstantGo quickly boots up or resumes Windows while enhancements help balance memory and processor resources for maximum efficiency and Battery Saver automatically conserves power. Windows 10 also offers more built-in security features improving protection against viruses, phishing, and malware. + +Solid State Drive + +More reliable and significantly faster than traditional spinning-platter hardrives, solid state drives work more like a large flash drive giving you quick access to your data. With no moving parts generating heat, solid state drives use less power and keep your system cooler which helps reduce component failure. Light weight and durable, solid state drives are often found in portable devices since they are less prone to travel damage and accidents like being dropped. + +1TB Hard Drive + +A terabyte equals one trillion bytes. To put that in perspective, that's room enough for about 250,000 MP3 files, or according to Ron White, in his book, "How Computers Work", 20,000 four-drawer filing cabinets filled with text. In addition, this hard drive operates at a variable RPM which means you've got the best of both worlds, full speed for demanding loads and energy conservation during minimal access. + +Smart Cache + +With Intel Smart Cache you benefit from increased data access because the cache is shared between the cores from a single access point and optimized by workload demand. That means your system is making maximum use of its resources, enhancing multi-threading, and reducing storage redundancy. + +Intel® Quick Sync Video + +Built into every 2nd generation Intel Core™ processor, Quick Sync Video enables users to quickly create, edit, synchronize, and share video files from home or make them available online without the need for extra hardware. In addition, media conversion between devices takes just minutes thanks to hardware acceleration streamlining this process at incredible speed. That includes creating DVDs, Blu-ray™ discs, or editing and converting HD videos for portable media devices or uploading to your favorite social media sites. + +Intel InTrue™ 3D Technology + +The perfect way to watch 3D movies. InTrue 3D technology delivers 1080p, hi-def resolution 3D action to your TV using HDMI 1.4. With Blu +================================================================================ +Rank = 13; Score = 6520832.0 +<|begin_of_text|>In the last few years, sites like Kickstarter and Indiegogo have revolutionized the world for startups. Ideas that didn't have mainstream business appeal were not only getting off the ground, but flourishing. Blockchain crowdfunding might just be the next step in startup evolution, helping important and interesting projects come to fruition. + +Crowdfunding provides an excellent way for creative projects to find cash. Many small or obscure projects lay outside of the scope of traditional investment. This makes it very difficult for them to get their ideas off the ground. Kickstarter, Indiegogo, and others changed that, by allowing startups to connect directly with potential consumers in order to seek funding. They acted as a trusted third party to keep the money in escrow, helping to protect potential investors from being scammed. If the funding succeeds, customers get whatever was promised to them as part of the campaign. If not, their money will be sent back. + +Blockchain Meets Kickstarter: Is This The Future? + +The problem with these established crowdfunding companies is that they are centralized bodies, charging high fees and also influencing the projects. Blockchain-based crowdfunding is set to be a game changer because it decentralizes the funding model from the likes of Kickstarter and other companies. + +Kickstarter provides a service and there are costs to run that service, so it's hard to blame them for charging 5% of the total funds received, with an additional 3-5% going towards payment processing. Despite this, it is still expensive. With blockchain's distributed ledger, there is the potential to remove this third party, which will save a considerable amount of the fundraising costs. + +Blockchain crowdfunding works by allowing startups to create their own digital currencies and sell them. This allows them to raise funds from early investors, while the investors also have the potential to make money if the value of their cryptographic shares increases. + +Some advocates consider this a more pure form of crowdfunding, because it removes any intermediaries standing between the backers and the startup. It also has the potential to boost new blockchain platforms, because it will give the blockchain community a new way to fund its own projects. It appeals to the more anarchistic blockchain enthusiasts as well, because it allows them to avoid traditional funding methods. + +How Does Blockchain Crowdfunding Work? + +It is similar to Kickstarter, with the creators posting their project and then soliciting funds from a community of interested people. Where it differs is that the startups will be able to make their own cryptocurrency to sell to potential backers. These cryptocurrency tokens will be accounted for and kept track of by the blockchain, which makes it immutable +================================================================================ +Rank = 14; Score = 6356992.0 +<|begin_of_text|>About + +Sayid Pro Brings Transparency to the Production Environment + +Sayid Pro is a tool for debugging and profiling Clojure in a production environment. We all want more visibility into what is happening in our production servers. When something is going wrong, we want precise answers. Sayid Pro aims to give you that. + +Sayid Pro works by tracing functions. When a traced function is executed, it's arguments and return value are captured by Sayid Pro. Sayid Pro provides a web interface through which you can remotely activate tracing on any functions, on any of your servers (that have Sayid Pro embedded in them). The Sayid Pro web application also allows users to query and visualize the captured trace data. + +You can also watch a hi-res version of the video here: https://youtu.be/y5ll-6IjJGw + +Architecture + +Library - The Sayid Pro Library is included in your production servers. It handles data capture and transmission. + +Hub - The Sayid Pro Hub runs in your infrastructure. It communicates with servers running the Sayid Pro library and receives trace data from them. Trace data is written to a PostgreSQL database as it is received. The hub also hosts the web app. + +Database - All captured trace data is stored in a PostgreSQL database. + +Web App - The web app is hosted by the Sayid Pro Hub. This is the interface that allows you to toggle tracing and then query and visualize captured trace data. + +Web App UI + +The web application is a critical piece of the product. I hope the prototype, while still immature, demonstrates the potential of Sayid Pro as a tool for solving production problems. Below is a description of the major UI components as they currently exist. + +Query Editor + +This is where it all starts. The query editor is for composing and executing queries and commands. These are sent off to the Sayid Pro Hub for execution. Some commands, such as enabling or disabling tracing, are forwarded on to servers. Servers can be selected individually, or in groups by a class tag. + +Trace data can be queried by different fields, such as end time or function name. Trace data is rendered into interactive visualization in the tree, table and chart UI components. Those are described below. + +Tree + +The tree UI component represents the trace data in a way that reflects its original call hierarchy. Each instance of a function call is a node. These nodes show the name of the function. The tree will also display the arguments and return value of a trace instance when the node is clicked. + +Table + +The table +================================================================================ +Rank = 15; Score = 6324224.0 +<|begin_of_text|>For the RAC release, here’s some of the metadata that was automatically generated using a system we developed called Constellate (that will become the foundational layer of our upcoming creator’s portal): + +Along with content-addressing, this layer then becomes the starting point for licensing a work. + +COALA IP forms part of the CBOR IPLD specification that will allow us to in the future traverse this graph from metadata object right into Ethereum and other merkle data structures. The website includes access to the open source implementations. + +With that, a group was formed in October 2015 to work such a standard. Ujo, BigchainDB, IPFS, Swarm, Jaak and many others collaborated over the course of the past year and half to establish such a standard. We didn’t reinvent the wheel and adapted existing standards to be more blockchain friendly, including hash linking that we’ve come to know and love: creating COALA IP. + +One of the key issues for a machine-readable licensing system: is to have a standard that’s easily readable and digestible. This makes it easier for developers to license works for their applications.The Tiny Human Ethereum Smart Contract had rights information in it, but it was embedded in an Ethereum smart contract. Besides having to run a full Ethereum node to get access to the data, there was also no standard API to interface with it back then. We also have to consider a substantially more long-term future, where it could be possible that specific blockchain implementations can morph and change over time. A key design consideration was splitting of the data (rights owners, from the business logic: paying, rights transfers, etc). The data can then be referenced by any current or future blockchain. + +On May 22nd, RAC announced that he wanted to release his album on the Ethereum network. We saw it as a good chance to demonstrate the potential of our architecture with him. In this collaboration, we are about 75% of the way towards finalising the different parts of this infrastructure. In this initial iteration, we had to manually craft the various components of the system, but also structured it in a modular and reusable way. In the future, with the introduction of our creator’s portal, we aim to automate this, and provide access to all artists. + +We’ve devised a system that improves the way current licensing is handled, but also opens up a large new scope for innovative and flexible licensing schemes. We saw glimpses of this when we launched the Tiny Human prototype with Imogen Heap in October 2015: a smart contract +================================================================================ +Rank = 16; Score = 6193152.0 +<|begin_of_text|>It’s not difficult to read and listen about the wonders of Embarcadero DataSnap technology around the world. I attended the Delphi Conference 2011 and 2012 in Sao Paulo, Brazil, after the release of versions of Delphi XE2 and XE3, and the words “performance” and “stability” are not spared when the subject is DataSnap. Apparently a perfect solution for companies that have invested their lives in Delphi and now need to reinvent themselves and offer web and mobile services. But does DataSnap really work on critical conditions? + +I found some references from other people talking about it on StackOverflow: + +The company I work for stands as one of the five largest in Brazil when it comes to software for supermarkets. The main product is a considerably large ERP developed in Delphi (for over ten years). Recently we initiated a study to evaluate technologies that would allow us to migrate from client / server to n-tier application model. The main need was to be able to use the same business logic implemented on a centralized location on different applications. + +The most obvious option is the DataSnap, which works very similarly to what is found in legacy applications. It can drag components, using dbExpress, for example. + +We tested the DataSnap technology in order to know what level of performance and stability it provides and to verify if it really meets our requirements. Our main requirement was the server’s ability to manage many simultaneous connections, since the application is big and used by many users. + +Objective + +Our objective was to test the DataSnap REST API and to answer some questions: + +How does it behave in an environment with many concurrent connections? + +What is the performance in a critical condition? + +Is it stable in critical condition? + +Methodology + +The tests are based on a lightweight REST method without any processing or memory allocation, returning only the text “Hello World”. The DataSnap server was created using the Delphi XE3 wizard and implemented the HelloWorld method. The testing was performed on all types of lifecycles (Invokation, Server and Session) and also with VCL and Console application. However, as the results were identical, the presented data is only based on the results obtained using the console version with “Server” lifecycle only. + +We decided to create a few servers based on other technologies, not necessarily similar or with the same purpose, just to have a basis for comparison. Servers have been created using the following frameworks: + +mORMot (Delphi) + +ASP.NET WCF + +Jersey/Grizzly (Java +================================================================================ +Rank = 17; Score = 6127616.0 +<|begin_of_text|>Take virtually any modern day SSD and measure how long it takes to launch a single application. You’ll usually notice a big advantage over a hard drive, but you’ll rarely find a difference between two different SSDs. Present day desktop usage models aren’t able to stress the performance high end SSDs are able to deliver. What differentiates one drive from another is really performance in heavy multitasking scenarios or short bursts of heavy IO. Eventually this will change as the SSD install base increases and developers can use the additional IO performance to enable new applications. + +In the enterprise market however, the workload is already there. The faster the SSD, the more users you can throw at a single server or SAN. There are effectively no limits to the IO performance needed in the high end workstation and server markets. + +These markets are used to throwing tens if not hundreds of physical disks at a problem. Even our upcoming server upgrade uses no less than fifty two SSDs across our entire network, and we’re small beans in the grand scheme of things. + +The appetite for performance is so great that many enterprise customers are finding the limits of SATA unacceptable. While we’re transitioning to 6Gbps SATA/SAS, for many enterprise workloads that’s not enough. Answering the call many manufacturers have designed PCIe based SSDs that do away with SATA as a final drive interface. The designs can be as simple as a bunch of SATA based devices paired with a PCIe RAID controller on a single card, to native PCIe controllers. + +The OCZ RevoDrive, two SF-1200 controllers in RAID on a PCIe card + +OCZ has been toying in this market for a while. The zDrive took four Indilinx controllers and put them behind a RAID controller on a PCIe card. The more recent RevoDrive took two SandForce controllers and did the same. The RevoDrive 2 doubles the controller count to four. + +Earlier this year OCZ announced its intention to bring a new high speed SSD interface to the market. Frustrated with the slow progress of SATA interface speeds, OCZ wanted to introduce an interface that would allow greater performance scaling today. Dubbed the High Speed Data Link (HSDL), OCZ’s new interface delivers 2 - 4GB/s (that’s right, gigabytes) of aggregate bandwidth to a single SSD. It’s an absolutely absurd amount of bandwidth, definitely more than a single controller can feed today - which is why the first SSD to support it will be a multi-controller device with internal RAID. + +Instead of relying on +================================================================================ +Rank = 18; Score = 6029312.0 +<|begin_of_text|>Working in a global retail environment poses some interesting availability challenges when you have physical Bricks and Mortar stores. I have been thinking about the problem of high availability in this environment for a little while now due to a project I am involved with to harmonise the retail systems used between global groups. It is quite a common problem for an organisation that has grown through acquisition that you have different systems used in different business units, but after a while it makes sense to try and go with a common platform. + +This article talks about how this architecture could look and how you can support the staggered roll-out of new Point of Sale features to the store whilst still maintaining high availability. + +Logical View + +Before I talk about the architecture, I want to cover a scenario first of the end state. Imagine there is a global retail company based in both North America and Europe. Both territories have around 1000 physical bricks and mortar stores. These stores each have a number of tills (cash registers for my American friends). There could be between 2 and 5 tills per store depending on its size. Each till communicates with systems hosted at a centralised location. These systems consist of web services, caching servers and databases. This has been illustrated in the diagram below. + +This diagram shows 2 geographic regions. Each Region contains a head office network infrastructure and a store network infrastructure. Both of the global regions are completely separate from each other. There are no shared resources between the two. + +A requirement for this architecture is that it be highly available. The stores should always have access to the services. The easiest thing to do is to put a load balancer in front of the web services and have multiple app servers. Then if any app server should go down or exhibit problems, the other load balancers can pick up the slack. + +Physical Architecture + +Although the App Servers described are load balanced and provide a degree of availability, if there is a problem/outage with the data centre where they are hosted, then you have a problem as no matter how many load balanced App Servers you have, if connectivity goes to the data centre then your retail stores are cut off from working. The way around this is to have co-location facilities in each region. This is shown in the diagram below. + +Each global region has 2 locations. There are 2 ways you can run this configuration. + +Active – Active: This is where you run both locations at the same time and load balance between the 2 of them. If one location goes down the other location picks up the traffic. + + +================================================================================ +Rank = 19; Score = 5963776.0 +<|begin_of_text|>Deploying a Ruby app with Passenger to production + +Passenger is an open source web application server for Ruby. It handles HTTP requests, manages processes and resources, and enables administration, monitoring and problem diagnosis. Passenger is very easy to use, makes deploying in production much easier and is scalable. + +What is this tutorial? + +This is an end-to-end tutorial that teaches you how to install Ruby and Passenger on a production server. This tutorial will ask you some questions, so the exact tutorial steps depend on the choices you make. In general, it covers these topics: + +Launching a server Installing the Ruby runtime Installing Passenger Setting up the production server in preparation for the application, such as creating a Unix user account and setting up permissions Deploying the application itself Updating the application to newer versions Optional integration with a web server, such as Nginx or Apache + +Not familar with Passenger? Please take a look at the quickstart and the basics tutorial. Developing? Looking to run Passenger in development? Then please read the basics tutorial or the development guide. Advanced user? Simply skip the steps that you are already familiar with, or refer to our installation guides which are not end-to-end, but are shorter and more focused. + +Ready to get started? + +Please start by picking the hosting provider or infrastructure that you want to deploy to.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 20; Score = 5963776.0 +<|begin_of_text|>Yesterday I read this article, from Derek Sivers’ blog. I encourage you to read it as well, but here’s the summary: + +We should move logic more into the database, because “Classes, models, and methods (OOP) are an unnecessary complication.” + +Data structure doesn’t change often, but code (logic) does. So let’s move the logic closer to the data structure + +“PL/pgSQL is not the most beautiful language, but the simplicity of having everything in the database is worth it.” + +We can do that by using triggers, stored procedures and the likes. We then call them directly from the application layer (there follows an example of a REST api in sinatra). + +First I thought it was a long winded joke, but the article was over before the expected “haha jk”. I actually had to scroll down quite a bit to find the first dubious commenter. + +Now, let’s clarify one thing: his approach makes sense – under very specific conditions. Those conditions loosely being that the application is mostly data centric (data in, data out, CRUD and a little validation), and most of the work is related to how data is presented. + +Fact is, we have already the perfect solution for this scenario, and we have had it for years: it’s called Microsoft Access. + +So, why aren’t all applications out there written in MS Access, or one of its clones? + +Enter: design space + +So, what is the problem with having all the logic in the database? + +Anyone who’s actually worked on a big application, with hundreds of triggers and stored procedures, knows perfectly what the problem is. It might be slightly more difficult to articulate it in clear words – I myself could not place a finger on the definition for a long time. + +When you decide to store all logic in the database, you have made the decision to fight in a very tight space. Everytime something unexpected pops up, you have to work it out inside the very cramped code room you’ve created for yourself. + +The advantage is that everything is in reach – add, delete, change a trigger here and a stored procedure there, you can alter a great deal of things with very little code. + +The disadvantage is that you’re effectively coding yourself into a corner. Moving is impossible except by making very complicated contorsions. Everytime you do, you’re at risk of breaking something you didn’t want to touch at all. Side effects galore. + +And, the more you do this, the more space you’re taking away from yourself – +================================================================================ +Rank = 21; Score = 5636096.0 +<|begin_of_text|>Pixode, a company that started in 2010 making mobile games, just announced Tuesday the launch of what it describes as a “Crypto 2.0” platform in Open Assets and a colored coins wallet called Coinprism + +“Colored coins are about taking a fraction of a Bitcoin, and tagging it with a secondary, user-defined value,” founder Flavien Charlon said in a release. “That value can then be stored on the block chain, and transferred simply by transferring the underlying bitcoins. + +“This opens a wide range of possibilities. An interesting use case is the ability for a company to issue shares in the form of colored coins, which can then be traded in a purely decentralized way on the Bitcoin block chain. This allows companies to run IPOs on the block chain in minutes, and for the price of a Bitcoin transaction.” + +The Open Assets platform could also be applied to smart contracts or precious metals, he said. + +Coinprism is one application of Open Assets, a wallet for storing, sending and receiving colored coins. wallet. “Coinprism has a strong focus on security, and private keys are always encrypted on the client before being transmitted to the server,” Charlon said. + +Pixode last year also released Predictious, a predictions and derivatives market. + +We reached out to Charlon recently to get some details about the projects. + +Cointelegraph: What are some other applications you imagine for Coinprism? + +Flavien Charlon: Bitcoin exchanges could decide to allow people to withdraw and deposit so-called “USD coins.” Those "USD coins" would be each worth one dollar, redeemable at that exchange. People can then start using that as a payment worth exactly one dollar, so they are not exposed to the volatility of Bitcoin while still getting all the benefits of a decentralized network. This can also allow people to do arbitrage between exchanges without even having to go through the traditional banking system. + +As a second example, colored coins can be used as a way to do decentralized and truly anonymous voting. A company like Coca-Cola could distribute "vote coins" by putting a private key on each bottle of Coke they sell, each with one "vote coin." Then at some point, when trying to decide what new product to put to market, they can ask their customers to send their vote coins to either one address if they want the new flavor to be X or to a different address if they prefer the new flavor to be Y. They can then simply count the number of "vote coins" on each address to know +================================================================================ +Rank = 22; Score = 5537792.0 +<|begin_of_text|>Last week, I wrote A Beginner’s Guide To MVC For The Web. In it, I described some of the problems with both the MVC pattern and the conceptual “MVC” that frameworks use. But what I didn’t do is describe better ways. I didn’t describe any of the alternatives. So let’s do that. Let’s talk about some of the alternatives to MVC… + +Problems With MVC + +Let’s restate the fundamental problems we talked about that exist with MVC: + +MVC Is Stateful It only makes sense if the View, as well as the View-Model binding is stateful (so the Model can update the View when it changes) MVC Has No Single Interpretation Every framework uses their own nuanced version. How Does Logging Fit In? Where does application code that’s not clearly data-centric belong in the application? Siblings To MVC + +There are a whole bunch of siblings to MVC that take slight divergences and have narrow differences. Let’s briefly talk about a few of them: + +This is quite similar to the MVC pattern, except that you can nest the triads together. So you can have one MVC structure for a page, one for navigation and a separate one for the content on the page. So the “top level” dispatches requests down to navigation and content MVC triads. + +This makes structuring complex pages easier, since it allows you to create reusable widgets. But it brings all of the problems that MVC has, and solves none of them (it just adds complexity on top). + +So HMVC doesn’t really solve our problems… + +The difference between MVC and MVVM is a lot more subtle. The basic premise is that in normal MVC, it’s bad that the View is doing two jobs: presentation and presentation data logic. Meaning that there’s a difference between actual rendering, and dealing with the data that will be rendered. So MVVM splits the MVC View in half. The presentation (rendering) happens in the View. But the data component lives in the ViewModel. + +The ViewModel can interact with the rest of the program, and the View is bound to the ViewModel. This means that there’s more of a separation between presentation and the application code that lives in the Model. + +The controller isn’t mentioned, but it’s still in there somewhere. + +Again, this solves some types of problems with MVC, but doesn’t address any of our issues. + +MVP is a bit different from MVC in implementation. Instead of having the Controller intercept user interaction and the View render data, MVP structures itself a bit differently +================================================================================ +Rank = 23; Score = 5373952.0 +<|begin_of_text|>[Haskell-cafe] ANNOUNCE: wai-0.0.0 (Web Application Interface) + +Hello all, I'd like to announce the first release of the Web Application Interface package[1]. The WAI is a common protocol between web server backends and web applications, allowing application writers to target a single interface and have their code run on multiple server types. There are two previous implementations of the same idea: the Network.Wai module as implemented in the Hyena package, and Hack. Some distinguishing characteristics of the wai package include: * Unlike Hack, this does not rely upon lazy I/O for passing request and response bodies within constrained memory. * Unlike Hyena, the request body is passed via a "Source" instead of Enumerator, so that it can easily be converted to a lazy bytestring for those who wish. * Unlike both, it attempts to achieve more type safety by have explicit types of request headers, response headers, HTTP version and status code. * It also removes any variables which are not universal to all web server backends. For example, scriptName has been removed, since it has no meaning for standalone servers. This package also contains separate modules for conversions to and from Sources and Enumerators. I am also simultaneously releasing two other packages: web-encodings 0.2.4[2] includes a new function, parseRequestBody, which directly parses a request body in a WAI request. It handles both URL-encoded and multipart data, and can store file contents in a file instead of memory. This should allow dealing with large file submits in a memory-efficient manner. You can also receive the file contents as a lazy bytestring. Finally, wai-extra 0.0.0 [3] is a collection of middleware and backends I use regularly in web application development. On the middleware side, there are five modules, including GZIP encoding and storing session data in an encrypted cookie (ClientSession). On the backend side, it includes CGI and SimpleServer. The latter is especially useful for testing web applications, though some people have reported using it in a production environment. All of the code is directly ported from previous packages I had written against Hack, so they are fairly well tested. As far as stability, I don't expect the interface to change too drastically in the future. I am happy to hear any suggestions people have for moving forward, but expect future versions to be mostly backwards compatible with this release. Also, future versions will respect the Package Versioning Policy +================================================================================ +Rank = 24; Score = 5341184.0 +<|begin_of_text|>Revolut, a financial technology start-up, has started to use IBAN format for account numbers that can be used by financial institutions to credit funds into the beneficiary’s account in 42 European countries, announced company. + +Lithuanian IBAN account numbers will be assigned to Revolut customers so that they will have a euro account number in any country in Europe. + +All payments executed by traditional banks, including salaries, can now be made into Revolut customers’ accounts with IBAN format. In addition, this functionality will allow customers to withdraw cash from PayPal, Lydia or P2P accounts. A euro account can be opened in less than one minute on a smartphone using the Revolut application. + +This financial technology start-up, known in some quarters as the “hooligan of banking”, has already attracted investment worth USD 66 million and has more than 700,000 users. + +The central bank of Lithuania, Lietuvos bankas, has made an unprecedented promise to FinTech companies by announcing it will provide preliminary answers to financial institution licence enquiries within one week, the fastest turnaround in the EU. + +The programme will apply to companies that are already licensed in another EU country and want to move their place of residence to Lithuania. Full authorisation will be issued in two to six months later (depending on the type of licence). + +The Bank of Lithuania also recently launched the programme Newcomer, which aims to facilitate FinTech companies setting up in Lithuania by providing consultations and guiding companies throughout the licensing procedures. Companies can apply with their requests to Newcomer@lb.lt and expect an answer within two to three working days. + +A range of other advantages are on offer to FinTech companies choosing to set up in Lithuania. Payment and electronic money institutions in Lithuania can access the Single Euro Payment Area (SEPA) through the infrastructure of the Bank of Lithuania, enabling them to reduce their dependence on banking competitors and offer a wider range of payment services. + +Moreover, the Bank of Lithuania allocates a code to providers of payment services that is used to generate accounts in IBAN format no later than the following business day.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 25; Score = 5308416.0 +<|begin_of_text|>The fundamental goal of typography is to make text easy and enjoyable to read. Typography has its own set of rules and guidelines. From there, we bend the rules to prioritize how to make the text easier to read. + +Treating text as an interface is as much about usability as it is enjoyment and ease. It is the writer’s responsibility to make content compelling and interesting and the typography’s responsibility to make the act of reading feel effortless and enjoyable. + +Our blog’s interface is text. The goals of that interface are to: + +reduce mental effort when reading and deciding what to read. + +simplify the information on screen to increase its usability. + +re-use patterns to minimize learning. + +provide feedback to the user to give them confidence in their actions. + +create an enjoyable experience. + +In The Elements of Typographic Style, Robert Bringhurst describes typography as a system based on “the structure and scale of the human body,” particularly the eye, hand, forearm, and of course, the mind. + +Just as Bringhurst applies the idea of designing with human behavior and physical attributes in mind, we see that practice equally applied in software and hardware design. Typography should take a form that is natural and beneficial for people to interact with. + +After making some updates to our blog, it was interesting to see that many of the considerations made in typographic changes are similar to the considerations we make when designing software. + +Here’s a screenshot of the updated designs: + +The changes we’ve made are centered around one of the jobs-to-be-done by the blog, which is to provide interesting information to the reader. Because of this, one of our goals is to present concise information in a way that allows the reader to make quick and informed assessments of what they’d like to read. + +The header elements (our logo and the blog title) have been grouped to stand out from the list of articles. This makes the page easier to scan for content. The title of an article is now emphasized over the date to present the most important information first. This also allows the link to get to the article to be front and center. + +To help the reader create accurate groupings of information, we increased the white space around each article listing and added simple borders to enclose the information related to each article. This prevents the information from bleeding together and causing confusion. + +To the individual articles, we decreased the line length of the paragraphs and added more distinct styles to the heading tags to emphasize hierarchy. + +Using line length as an example, let’s look more closely at the the impact of typography on user experience. Bringhurst advises that +================================================================================ +Rank = 26; Score = 5242880.0 +<|begin_of_text|>[This post is by Romain Guy and Chet Haase, Android engineers who have been known to collaborate on the subject of graphics, UIs, and animation. You can read more from them on their blogs at curious-creature.org and graphics-geek.blogspot.com. — Tim Bray] + +Earlier this year, Android 3.0 launched with a new 2D rendering pipeline designed to support hardware acceleration on tablets. With this new pipeline, all drawing operations performed by the UI toolkit are carried out using the GPU. + +You’ll be happy to hear that Android 4.0, Ice Cream Sandwich, brings an improved version of the hardware-accelerated 2D rendering pipeline to phones, starting with Galaxy Nexus. + +Enabling hardware acceleration + +In Android 4.0 (API level 14), hardware acceleration, for the first time, is on by default for all applications. For applications at lower API levels, you can turn it on by adding android:hardwareAccelerated="true" to the tag in your AndroidManifest.xml. + +To learn more about the hardware accelerated 2D rendering pipeline visit the official Android developer guide. This guide explains how to control hardware acceleration at various levels, offers several performance tips and tricks and describes in details the new drawing model. + +I also encourage you to watch the Android Hardware Accelerated Rendering talk that we gave at Google I/O 2011. + +Introducing TextureView + +Applications that need to display OpenGL or video content rely today on a special type of UI element called SurfaceView. This widget works by creating a new window placed behind your application’s window. It then punches a hole through your application’s window to reveal the new window. While this approach is very efficient, since the content of the new window can be refreshed without redrawing the application’s window, it suffers from several important limitations. + +Because a SurfaceView’s content does not live in the application’s window, it cannot be transformed (moved, scaled, rotated) efficiently. This makes it difficult to use a SurfaceView inside a ListView or a ScrollView. SurfaceView also cannot interact properly with some features of the UI toolkit such as fading edges or View.setAlpha(). + +To solve these problems, Android 4.0 introduces a new widget called TextureView that relies on the hardware accelerated 2D rendering pipeline and SurfaceTexture. TextureView offers the same capabilities as SurfaceView but, unlike SurfaceView, behaves as a regular view. You can for instance use a TextureView to display an OpenGL scene or a video stream +================================================================================ +Rank = 27; Score = 5242880.0 +<|begin_of_text|>Bitcoin BlockNotary Used for Payment Processor Pay-Me + +BlockNotary is a new app that uses Tierion, and it is now used for one of Russia’s largest mobile payment processors. Pay-Me is a Russian mobile payments processor with over 30,000 installed devices, now backed by the power of the Bitcoin blockchain. + +In the past, new customers would have to register via interview that was conducted in person with a customer service rep, due to Russian criminals stealing credit cards and proceeding to use mobile PoS terminals to launder money. + +“Recording a video of yourself during the registration process helps prevent fraud. As CEO, I’d like to make it as difficult as possible for criminals to use Pay-Me to conduct illegal activities. – Pay-Me CEO, Vladimir Kanin.” + +Blocknotary has done away with this, replacing it with a video interview that can be conducted through a mobile application, where Tierion (through Blocknotarys app) collects the data to create verifiable record of the process which is then stored on the blockchain. This greatly accelerates the process, by making the whole verification steps much smoother, allowing the customer to have their PoS device shipped in a timelier manner. A video of the entire process (takes literally around a minute from start to finish) is seen on Tierions blog, linked here. + +Blocknotary’s Video Interview application lets customers verify their identity and conduct an interview from their location. The customer’s statement of intentions and the information needed for a background check are collected. This makes the onboarding process smoother for new customers and allows Pay-Me to ship a PoS device before a background check is complete. Additionally, Pay-me gets a tamperproof record of the entire process that can be verified using blockchain. This record can lower Pay-Me’s legal liability and be useful to police that are investigating fraudulent activity. + +Tierion is a “cloud platform backed by the power of the blockchain”, allowing users to easily collect and store data. Once data is acquired, it is then anchored to the blockchain to ensure authenticity, and the user is sent a blockchain receipt, allowing the user to verify the data without relying on a trusted third party. + +Will other financial institutions look to Bitcoin when it comes to screening new customers? What do you think? Let us know in the comments below!<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 28; Score = 5210112.0 +<|begin_of_text|>If you’re thinking about getting pastel hair, you should consider the pros and cons first. Just kidding. Pastel hair is awesome! If you’re torn, do it! If you’re not sure if pastel hair would suit you, do it! You’ll be glad you did. If you want to know how to get pastel hair, keep reading! + +Below are all the products you’ll need. They’re clickable, so check them out. The white toner isn’t mandatory; you only absolutely need the bleaching kit, the conditioner, and your choice of hair dye. In addition to this, you’ll need some gloves, a shower cap, and a mixing bowl. The whole process should end up costing under 25$. + +[metaslider id=1039] + +Step 1: How To Bleach Your Hair + +The first step is bleaching your hair. If you have anything but light blonde hair, you absolutely need to bleach it if you want the pastel color to actually be pastel. If you have already blonde hair, skip this step. When choosing your bleaching product, it’s important to go for 30 volume. Anything under that is too weak for achieving pastel hair, and anything over that is too strong for safety — bleach can fry your hair, so be very careful! I recommend this nifty Manic Panic bleaching kit, which also contains gloves, a brush, and a mixing bowl. Tip: If you have very thick or long hair, you WILL need 2 boxes! It ends up going a bit over budget, but it’s well worth it! The first time I bleached my hair, I had to pause the job mid-bleaching and ask my friend to kindly go buy another box for me. Get two if you’re unsure! + +Make sure your hair hasn’t been washed for a few days before bleaching. Bleach is harsh on your scalp, and the oils will help protect it. + +If it’s your first time, have someone else help you and bleach your hair for you. Bleaching hair is tricky. Be safe! + +Follow the instructions on the packaging. A strand test is a very good idea! + +If your hair still isn’t light enough after bleaching, consider a white toner. If you want pastel blue hair, your hair has to be a very white blonde. Any yellow in your hair and the blue dye will turn green! + +Step 2: How to Get Pastel Hair + +There is no “pastel hair dye +================================================================================ +Rank = 29; Score = 4947968.0 +<|begin_of_text|>Completely Useless Fun Project: Parts Of The Compiler + +If you have done C/C++ or Objective-C work on MacOS (There is a project called GNUStep that allows you to run Obj-C on Linux), you may have heard of Clang and LLVM, or maybe Clang/LLVM. It may have confused you because there are two names for a seemingly single piece of software. It may confuse you further to point out that these are two different things, but two pieces to the same puzzle. + +What Clang and LLVM are, are the pieces of a compiler. A compiler is really just a single group of processes that takes in some source code and outputs some other code. This output code could be Assembly code, Java ByteCode, hell even Javascript. Like all good programming problems, compiler construction can be broken down into various parts. + +First, a history + +It may horrify you to know that computer programs weren’t always written in english. For many decades, Programs were written by hand in assembly language. It wasn’t until 1952 that totally bad ass Rear Admiral Grace Hopper built the first compiler and coined the term (she also coined the term debugging after pulling a literal moth out of her computer’s circuits). + +This first compiler wasn’t much of anything. It really just operated as more of a linker or a loader than what our modern compilers are. The first real modern compiler was introduced in 1957 by John W Backus and his team at IBM for Fortran. + +And compiler design hasn’t changed all that much since then. + +Design of the modern Compiler + +The Fortran compiler was built in an era where computers were insanely expensive, huge and slow. In order for Fortran to compete with hand coding assembly, it needed to be fast, at least as fast as hand written code. To do this, the compiler was broken down into two parts: the frontend and the backend (that is what Clang and LLVM are, the frontend and backend respectively). This made it easier to apply transformations and optimizations independently. + +The phases of the compiler + +At the highest level, the frontend takes the source code and outputs what is called an Intermediate Representation or IR. This may look like assembly code, all though it doesn’t have to, but it isn’t. This code is how compilers internally represent the source code. IRs are used so that that it can be optimized and translated much easier. They also must be designed in such away that the intent of the original source code is preserved and be independent of any source or target +================================================================================ +Rank = 30; Score = 4915200.0 +<|begin_of_text|>Image copyright Getty Images Image caption Disney's Magic Kingdom is the world's most visited theme park, with more than 17 million visitors in 2012. + +Disney has raised the price of a one-day ticket to its Magic Kingdom theme park in Orlando, Florida by $4 to $99 (£59.50). + +Magic Kingdom is the world's most visited theme park, with more than 17 million visitors in 2012. + +Ticket prices at other parks within Walt Disney World, including Epcot and Hollywood Studios, also jumped by $4. + +The hike comes after Disney reported surging revenues from its theme park businesses. + +In 2013, Disney made $671m from its theme parks - a 16% increase from the same period in 2012. + +Overall, its parks brought in $3.6bn in revenue in 2013, a 6% increase from 2012, according to the firm's most recent financial filings. + +The boost in revenue was "primarily due to increased guest spending at our domestic parks and resorts, which reflected higher average ticket prices and food, beverage and merchandise spending," according to Disney. + +Disney said that the majority of visitors to its parks buy tickets for multiple days. + +The company has theme parks throughout the world, with locations in Paris, Tokyo, and Hong Kong. + +According to a report from the Themed Entertainment Association and technology company AECOM, Disney's theme parks globally had more than 126 million visitors in 2012, the most recent year for which statistics are available.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 31; Score = 4849664.0 +<|begin_of_text|>Software Transactional Memory is a concept which offers many advantages when writing concurrent code. Most importantly, it removes the burden of writing correct multi-threaded code from the programmer, allowing her to focus more on the logic of the problem domain. STM will automatically take care of correctness. However, this comes with a cost – the bookkeeping. Tracking every access to a transactional field in order to properly detect and resolve conflicts introduces a cost penalty. But, at the same time, it also introduces the possibility to create and expose to the programmer some very interesting new semantics. + +Shielded has for some time already contained one such semantic addition – the Shield.Conditional static method. With it one may define a transaction which should run whenever a certain condition is satisfied by the current, committed “state of the world”. By virtue of its bookkeeping, Shielded can determine the exact fields your condition depends on. This way it easily knows, when another transaction commits into one of those fields, that your condition might have changed, and will re-evaluate it. (It will even track your condition for any changes in its access pattern between calls. This is important since, due to boolean expressions’ short-circuit evaluation, it is very easy for a condition to change its access pattern, and that must be taken into account.) + +This feature was inspired by the Haskell STM’s “retry” command, but it does not go as far as that. The Haskell “retry” can be called anywhere in a transaction, and it will block it until any of the fields, that it has accessed before calling retry, changes. Shielded does not do any such thing, it is a deliberate design decision to not include any threading or blocking constructs into the library. The idea is to allow Shielded to easily be combined with any such construct, depending on what the programmer wants to use, and not lock the programmer into a decision made by this library. And so, the Conditional does not even try executing straight away, but merely subscribes a transaction for future execution. + +The power of Haskell’s retry and orElse can be achieved by using Reactive Extensions. Combining Shielded with Rx.NET should not be a difficult task – create an IObservable by creating a Conditional which will call methods on an IObserver (tip – make such calls to IObserver as SideEffects of the Conditional transaction, otherwise you might be triggering observers repeatedly, and with non-committed data!). Once you have such an IObservable, the powerful Rx.NET library allows you to easily create very complex time-based logic, +================================================================================ +Rank = 32; Score = 4816896.0 +<|begin_of_text|>The issue is particularly nettlesome with individuals who have transitioned from male to female because of the belief that they might still have a physical advantage until their hormone therapy — which not every transgender person chooses to have — is complete. + +“When you start talking about transgender athletes, a male-to-female individual, we want to ensure that that is truly a decision that is permanent,” said Bobby Cox, the commissioner of the Indiana High School Athletic Association. “It is not a decision that, ‘I just decided today that I am going to be a girl and I am going to go play on a girls’ team’ and perhaps, disadvantage those kids that are on the team and imbalance the competition. + +“And as we progress down this path,’’ he added, “and as we spend more time and energy with advocacy groups and medical professionals in this area, I think that there will be additional amendments to our policies.” + +There is no reliable data on the number of transgender high school athletes. Only about 0.6 percent of the adult population identifies as transgender, according to federal data from 2016. Researchers from the Williams Institute, who conducted the study, reported that 0.56 percent of adults in Indiana reported identifying as transgender. + +In Texas, 0.66 percent of adults said they identify as transgender, the fifth-highest percentage in the country (behind Hawaii, California, New Mexico and Georgia). Still, transgender children are considered an at-risk minority outside of sports. According to The New England Journal of Medicine, the rate of suicide attempts among transgender people is 40 percent, compared to 4.6 percent among those who are not transgender. + +Relatively new policies are already being massaged and amended. In July, the Indiana state association voted unanimously to adjust its surgery requirement. It now no longer requires transgender students who are transitioning from female to male to have sex reassignment surgery in order to compete in the gender with which they identify. But that stipulation is still in place for someone transitioning from male to female. + +The group asserts it is acting in the name of fairness, but transgender rights activists accuse members of simply not wanting transgender people to participate, out of fear that those athletes will have an unfair advantage.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 33; Score = 4816896.0 +<|begin_of_text|>Throughout my software development career, I’ve seen my fair share of debates over how databases should be developed. And like most disagreements over technical pedantry, the participants are generally well-intentioned but grossly inexperienced. So much so that it wouldn’t matter if they use foreign keys or not. Though “elegant”, their software will become an unmaintainable mess after a few short years. But regardless, one counter-argument that constantly rears its ugly head goes something like, “but what if we have to change it later?” + +In other debates, that question can be quickly rebutted with “uh, we just change it then,” but when the discussion is on databases, it actually holds water. It’s as if the database is a marble statue that, once shipped to production, should never be changed... and if it must, there had better be a damn good reason. + +Let’s keep in mind that the whole point of developing software – that is, spending large sums of money paying developers’ salaries to build and maintain applications – is to have something that can change according to business need. When developers are unable, afraid, or unwilling to change the applications they develop, they make a very compelling case for being replaced by SAP. + +Databases are Different + +Databases are indeed different, especially when juxtaposed with what application developers are most familiar with: application code. + +Generally speaking, application code lives in a source control system and is a set of instructions that tell that tell the machine to do something. Whether it’s compiled or interpreted, or executed on a real or virtual machine, application code does stuff. + +As a developer, when you’re satisfied that it’s doing the right stuff, the release management process kicks in to promote that code through acceptance testing, quality testing, other testing, and finally production. And all along the way, the code should not change. If a bug is found, then it’s fixed in source control and sent right back through the same process. + +Databases, on the other hand, live on the production database server. There can certainly be other instances – development, testing, staging – but the only one that really matters is production. And these databases don’t actually do stuff, they’re merely modified and queried by other applications with a Structured Query Language. And unlike application code, databases (or at least, their completely integrated data) do change after deployment – that’s kind of their whole point. + +Database Changes Done Wrong + +Unlike application code, you can’t exactly drop a bunch of files in a +================================================================================ +Rank = 34; Score = 4784128.0 +<|begin_of_text|>There has been much talk about modularity recently. Fedora even has a working group on this topic. Modularity is such a generic term that it is a bit hard to figure out what this is all about, but the wiki page gives some hints: …base module… …docker image… …reduced dependencies… …tooling… + +A very high-level reading would be that this is about defining tools and processes to deliver software in modules, which can be larger units than packages, and e.g. come in the form of container images. + +Another reading would be that this is an effort to come to grips with the changing role of distributions in a world were containers are the dominant way to run and deploy applications. + +Modularity on the desktop + +In this post, I want to look at how this modularity effort looks from the desktop perspective, and what we are doing in this area. + +Our main goal for a while has been to make it easier to get desktop applications from application developers to users. + +In the traditional Linux distribution world, this is a total nightmare: Once you’ve written your application, you need to either learn how to create an Ubuntu.deb, an Arch.pkg, a Fedora.rpm, to name just a few, and then follow lengthy processes to get your packages accepted into these distributions. + +And once you’ve done all that, you will get bug reports that your application is broken with the one or other version of one of your dependencies. If not right away, then a few months down the road when the distributions move on to the latest and greatest releases. So, keeping those packages working requires the constant attention of a package maintainer. + +To improve on this sad state of affairs, we need to make applications much less dependent on the OS they run on, and on the libraries that happen to come with the OS. + +At this point, you might say: Aha, you want to bundle dependencies! Yes, but read on. Bundling is a bad word in the traditional distribution world. It implies duplication, since multiple applications may bundle the same library. And having multiple copies of the same library (possibly different versions, too), makes it harder to ensure that bug and security fixes get applied to all the copies. Not to mention that all the duplication consumes bandwidth when you have to download it all. + +Bundling everything with the application certainly has its drawbacks. But there are several things we can do to preserve most of the benefits of bundling, while avoiding the worst of the problems. + +One takeaway from the modularity discussions mentioned +================================================================================ +Rank = 35; Score = 4620288.0 +<|begin_of_text|>Pantheon, a WordPress and Drupal hosting service with a strong lineup of features for developers, today announced that it has raised a $29 million Series C round. Investors in this round include previous investors Foundry Group, OpenView Investment Partners, and Scale Venture Partners, as well as new investor Industry Ventures, which put $8.5 million into this round. + +This new round follows Pantheon’s $21.5 million Series B round in 2014 and brings the company’s total funding to $57 million. + +As Pantheon CEO and co-founder Zack Rosen told me, the company wants to help build the foundational technology for the web and eventually get to the point where it powers 30 percent of all sites. For now, though, he’s happy to simply make progress to getting to 1 percent, which he thinks the company will achieve in a few years. To help make that happen, the company plans to launch a migration toolkit next week that will allow website owners to more easily bring their existing site to the company’s platform. + +As Rosen told me, Pantheon now hosts 150,000 sites and while the company doesn’t release customer numbers, Rosen says that the company is seeing customer growth of over 100 percent year-over-year. To host all of this, Pantheon currently has 1.2 million containers in production. + +On group of users the company is especially focussing on its agencies. It now has partnered with 2,500 agencies and 50 resellers. + +As Rosen told me, the team decided to treat this round of funding like its last one. “We control our burn and are very deliberate in how we deploy our funding,” he added. It’s no secret that raising funding is getting harder and Rosen noted that VCs now want to see more numbers and ask tougher questions before they commit to a round. He does see a positive side of this, too, though. “It’s bad for the industry when VCs don’t ask the kind of questions they should be asking and companies that should be funded get funded — and then down the road they don’t have great results,” he said. “That makes it harder for everyone else.” + +The company wants to use this new round of funding to build out its product and make it easier for developers to build their sites on its platform. + +As part of today’s funding announcement, the company also announced that it has hired Niall Hayes, who previously led technology at KIXEYE and ran the CityVille studio at Zynga, as its VP +================================================================================ +Rank = 36; Score = 4587520.0 +<|begin_of_text|>January 2017 An updated post is available: Introducing the MEAN and MERN stacks. + +By Valeri Karpov, Kernel Tools engineer at MongoDB and and co-founder of the Ascot Project. + +A few weeks ago, a friend of mine asked me for help with PostgreSQL. As someone who’s been blissfully SQL-­free for a year, I was quite curious to find out why he wasn’t just using MongoDB instead. It turns out that he thinks MongoDB is too difficult to use for a quick weekend hack, and this couldn’t be farther from the truth. I just finished my second 24 hour hackathon using Mongo and NodeJS (the FinTech Hackathon co­sponsored by 10gen) and can confidently say that there is no reason to use anything else for your next hackathon or REST API hack. + +First of all, there are huge advantages to using a uniform language throughout your stack. My team uses a set of tools that we affectionately call the MEAN stack:­ MongoDB, ExpressJS, AngularJS, and Node.js. By coding with Javascript throughout, we are able to realize performance gains in both the software itself and in the productivity of our developers. With MongoDB, we can store our documents in a JSON-­like format, write JSON queries on our ExpressJS and NodeJS based server, and seamlessly pass JSON documents to our AngularJS frontend. Debugging and database administration become a lot easier when the objects stored in your database are essentially identical to the objects your client Javascript sees. Even better, somebody working on the client side can easily understand the server side code and database queries; using the same syntax and objects the whole way through frees you from having to consider multiple sets of language best practices and reduces the barrier to entry for understanding your codebase. This is especially important in a hackathon setting: the team may not have much experience working together, and with such little time to integrate all the pieces of your project, anything that makes the development process easier is gold. + +Another big reason to go with MongoDB is that you can use it in the same way you would a MySQL database (at least at a high level). My team likes to describe MongoDB as a “gateway drug” for NoSQL databases because it is so easy to make the transition from SQL to MongoDB. I wish someone had told me this when I first starting looking into NoSQL databases, because it would have saved me a lot of headaches. Like many people, I was under the impression that CouchDB would be easier to +================================================================================ +Rank = 37; Score = 4554752.0 +<|begin_of_text|>Space Invaders + +//Detect if the browser supports DeviceMotionEvent if (window.DeviceMotionEvent!= undefined) { //ondevicemotion is fired when iOS device detects motion window.ondevicemotion = function(e) { //ax is the movement on the x axis. //This motion is used to move the ship in the game ax = event.accelerationIncludingGravity.x * 5; ay = event.accelerationIncludingGravity.y * 5; //Status 0 is start, 1 is left, 2 is right, 3 is stay if(status == 0){ //initial condition status = 3; //stay socket.emit('spaceChange', {'ax': 3}); statusmsg = 'Waiting for movement'; } if(ax > 14 && status!= 2){ //move right on device status = 2; socket.emit('spaceChange', {'ax': 2}); statusmsg = 'Moving ship right'; } if(ax < -14 && status!= 1){ //move left on device status = 1; socket.emit('spaceChange', {'ax': 1}); statusmsg = 'Moving ship left'; } if(ax > -14 && ax < 14 && status!= 3){ //device held steady status = 3; socket.emit('spaceChange', {'ax': 3}); statusmsg = 'Ship held steady'; } + +Current space invaders game uses only X axis movement on the iOS device. However, all three dimension values are available on iOS Safari ondevicemotion method. It is technically possible to do much more with the iOS device. + +-- webdigi + +The guys over at WebDigi, a London-based web application development company, has released a demo ofSpaceship Pilot, a pretty amazing application that uses iOS devices as the controller for a web browser-based game on your desktop or laptop. No application is required on the iOS device, you simply scan the QR code with the device and the controller code reads the accelerometer values and sends the movements to the node.js server.Here are the key parts of the iOS device controller code:Only changes in direction on the iOS device are pushed to the node.js server, reducing the amount of data pushed between phone and server. The node.js server is coupled with Socket.IO, making it easier to deliver data in realtime using most web browsers. Since Socket.IO code is run on both the server and client sides, pairing it with node.js allows accelerometer values to be pushed instantly from mobile safari to your browser.After testing (read: playing +================================================================================ +Rank = 38; Score = 4521984.0 +<|begin_of_text|>Double-Spend Protection: + +Double-spend protection is the main reason blockchains exist in the first place. In technical terms this means that two valid transactions which spend the same transaction output (UTXO) will conflict and only one can be confirmed in the network. Account based languages (for example Ethereum) that allow for spending the same amount from the same address multiple times, usually have other means to prevent double spending. + +Multisig: + +Multisig is a very old concept and can be compared to a shared checkbook with multiple required signers. A multisig transaction allows to enforce arbitrary joint signature rules. COMIT uses 2 out of 2 multisig transactions for which both signers have to sign a transaction to become valid and accepted by the network. Multi-signature transactions are a requirement for Payment Channels. + +Time-Locks: + +A timelock is a simple requirement for funds to be locked up until a future date. Blockchains are found to have 2 different kind of time-locks: relative and absolute time-locks. Absolute time-locks will lock a transaction output until a fixed time in the future. Relative time-locks will lock a transaction output relative to the time the transaction was confirmed. Time-locks are a requirement for trustless Payment Channels and relative time-locks are recommended as they allow for indefinitely open Payment Channels. + +Hash function: + +To be able to route across multiple blockchains, we need the same hash function to be able available in the smart contracting language of each blockchain. A standard hash function like the SHA256 hash function is usually available and is perfectly suitable for this purpose.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 39; Score = 4489216.0 +<|begin_of_text|>In the world of elearning, microlearning has become a buzzword. It’s considered a powerful new approach to training the workforce. With the average attention span in North America dropping from 12 seconds in 2000 to 8 seconds in 2015, the demand for shorter, more engaging training is higher than ever! + +In this post, our goal is to cover the basics and leave you with an understanding of: What Microlearning is and the benefits it can provide. Throughout this post, we’ll try to give you examples of how to use microlearning in your own training programs. + +What is Microlearning? + +Microlearning is focused learning that is delivered in bite sized chunks. Since this method of learning provides small bits of knowledge at a time, it’s best used for delivering information that learners need to retain. + +Microlearning can be achieved using a number of different delivery methods: emails, online posts and short multimedia videos, are all examples of different ways you can deliver training that is designed for your learners to retain new knowledge and achieve their educational goals. + +Examples of Microlearning: + +– Watching video tutorials on Youtube + +– Receiving small bits of education via email: like Word of the Day from Dictionary.com + +– Online learning programs like Duolingo or Lynda + +What are the benefits that Microlearning can provide? + +1. Avoid the risk of overwhelming learners + +Microlearning allows learners to move at their own pace, giving them the ability to go back and review complex concepts as often as needed. Since new knowledge is delivered in smaller chunks, learners avoid the risk of being overwhelmed by too much information at once. + +2. Create on-the-go training that can be accessed anywhere, at anytime. + +Microlearning can be achieved using a number of different delivery methods. Email, online posts, videos, even tweets because of this training can usually be accessed across multiple devices, making it available on-the-go. Learners can access and review training materials while doing everyday tasks like: waiting for the bus, sitting on the train, or even riding the elevator! + +3. Help learners better retain new knowledge + +Traditional classroom training often provides little to no long term takeaways for learners. The Wall Street Journal recently reported that 90 percent of new skills are lost within a year of training! Microlearning breaks new knowledge down into short chunks making learning easier to digest, understand, and apply on the job. + +Different ways to use Microlearning: + +– Health and Safety training + +– Learning new software + +– Business Processes and Procedures + +Microlearning yields many benefits for organizations looking +================================================================================ +Rank = 40; Score = 4456448.0 +<|begin_of_text|>Tomorrow (November 7th) we’ll be hosting a testday for the Server Product. Members of the Server Working Group as well as QA will be available on freenode in the #fedora-test-day channel to help out with testing. The focus of the testday will be on several facets of the server product: + +At this point you might be asking, “What do each of those things do and how would I benefit from testing them?” I’ll try to give a quick explanation of what each of those features are and why they’re important to the Server Product. + +FreeIPA + +FreeIPA provides your Linux network with its own Domain Controller capable of managing users, groups, DNS, certificates and single-sign-on capabilities. Taken from the FreeIPA website: + +FreeIPA is an integrated Identity and Authentication solution for Linux/UNIX networked environments. A FreeIPA server provides centralized authentication, authorization and account information by storing data about user, groups, hosts and other objects necessary to manage the security aspects of a network of computers. + +Cockpit + +Cockpit is a web based tool for managing and monitoring servers. It takes a lot of the pain out of administering several servers allowing you to do it from one, easy to use, central location. It takes a lot of the pain out of managing all your services, users, containers and pretty much anything your server might be doing. + +OpenLMI + +OpenLMI is a project aimed at creating a standardized API based on DMTF-CIM for managing your Linux systems. From the their site: + +The OpenLMI project provides a common infrastructure for the management of Linux systems. Capabilities include configuration, management and monitoring of hardware, operating systems, and system services. OpenLMI includes a set of services that can be accessed both locally and remotely, multiple language bindings, standard APIs, and standard scripting interfaces. + +Rolekit + +Rolekit is the tool used to enable the easy deployment of different server roles. It makes it a snap to deploy different kinds of servers, such as web, mail, domain controllers, etc. (referred to as “roles”). + +Conclusion + +Suffice it to say, Server offers a lot of new and great functionality. So come by for the test day and take some time to get familiar with it’s offerings and make it an awesome release! More information can be found on the wiki – and you can always drop by #fedora-server or #fedora-qa if you have any questions or want to get involved!<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 41; Score = 4390912.0 +<|begin_of_text|>For almost three years, millions of servers and smaller devices running Linux have been vulnerable to attacks that allow an unprivileged app or user to gain nearly unfettered root access. Major Linux distributors are expected to fix the privilege escalation bug this week, but the difficulty of releasing updates for Android handsets and embedded devices means many people may remain susceptible for months or years. + +The flaw, which was introduced into the Linux kernel in version 3.8 released in early 2013, resides in the OS keyring. The facility allows apps to store encryption keys, authentication tokens, and other sensitive security data inside the kernel while remaining in a form that can't be accessed by other apps. According to a blog post published Tuesday, researchers from security firm Perception Point discovered and privately reported the bug to Linux kernel maintainers. To demonstrate the risk the bug posed, the researchers also developed a proof-of-concept exploit that replaces a keyring object stored in memory with code that's executed by the kernel. + +The vulnerability is notable because it's exploitable in a wide array of settings. On servers, people with local access can exploit it to achieve complete root access. On smartphones running Android versions KitKat and later, it can allow a malicious app to break out of the normal security sandbox to gain control of underlying OS functions. It can also be exploited on devices and appliances running embedded versions of Linux. While security mitigations such as supervisor mode access prevention and supervisor mode execution protection are available for many servers, and security enhanced Linux built into Android can make exploits harder, there are still ways to bypass those protections. + +Update, Jan. 20, 1:48 PST: In a post published a day after this post went live, Google said company researchers don't believe any Android devices are vulnerable to exploits by third-party applications. It also said researchers believe that the number of Android devices affected is "significantly smaller than initially reported." Google will nonetheless issue an update in March that patches the vulnerability. + +"As of the date of disclosure, this vulnerability has implications for approximately tens of millions of Linux PCs and servers, and 66 percent of all Android devices (phones/tablets)," Perception Point researchers wrote. "While neither us nor the Kernel security team have observed any exploit targeting this vulnerability in the wild, we recommend that security teams examine potentially affected devices and implement patches as soon as possible." + +While malware distributors have focused most of their resources over the years on infecting computers running Microsoft Windows, they have put increased focus on attacking competing OSes. In 2014, +================================================================================ +Rank = 42; Score = 4390912.0 +<|begin_of_text|>The Scream Remake Made The Weirdest Decision Possible By Nick Venable Random Article Blend Scream Scream!” + +The news comes from Ghost Face creator R.J. Tolbert, who spoke with + +Please note that I had been in communication with TWC regarding this and they have informed me that, as of now and during the initial launch, that Ghost Face is not involved in the new format. They also indicated that because of this direction, it does not mean, that Ghost Face will not be involved at a later date. So, there is still a question, to the question. However, as of now, there is no involvement.” + +You gotta love anyone who says there is still a question to a question. This is pretty surprising news, given I can’t possibly correlate any other image with the concept of Scream than the Ghost Face mask. (Nor the Scary Movie films.) Sure, I can picture Jaime Kennedy in turmoil, but I don’t want that to get showcased in the MTV series either. Tolbert wanted to make it clear to everyone that he feels the same way about the connection between the killers’ mask choice and the film series. + +We [Fun World] believe Scream is Ghost Face and Ghost Face is Scream. However, while Ghost Face is owned by Fun World, the Scream motion picture franchise is owned by The Weinstein Company, and it is their option to film a movie or TV show without Ghost Face.” + +The + +Sure, there’s still a chance that Ghost Face will make his presence known at another point in the series’ existence, but it has to actually last first. And what are the chances of that happening if they don’t even have the core franchise image to market it with? Every so often, a film sequel gets made despite having zilch to do with the original film. It occurs less often on TV, but that’s exactly what’s happening with MTV’s upcoming series, based on the highly lucrative film franchise of the same name. The show’s creative team will reportedly be leaving the Ghost Face mask out of it altogether, which basically means this series will be inspired by murdered teenagers. News about dead teens always makes me think, “That’s so!”The news comes from Ghost Face creator R.J. Tolbert, who spoke with GhostFace.co recently about the subject of the copyrighted character appearing in the series, and here’s what he had to say.You gotta love anyone who says there is still a question to a question. This is pretty surprising news, given I can’t +================================================================================ +Rank = 43; Score = 4390912.0 +<|begin_of_text|>MIDI to Audio with frozen tracks – No Flatten + +Have you ever wanted to turn your MIDI track to Audio but keep the MIDI? Well, I use this quick tip all the time. I discovered it by mistake when editing years back, and find myself doing it all the time. What you need to do is first Freeze your MIDI track. Once it is frozen you can select the MIDI Clip and drag it into an Audio Track by holding Ctrl or CMD while you move it. This will create a copy of the “frozen” audio in the Audio Track, but the MIDI will still be there. You can then Unfreeze the MIDI track. + +Copy and Paste Time + +When you are writing a track there are times that you want to copy a whole section of the music or move the Verse after the Chorus, or you name it. This can be really tedious without knowing this workflow hack. The key is the Cut Time, Paste Time, and Duplicate Time option in Ableton Live. + +All you need to do is select an area in Arrangement View. This will be the area that will be duplicated or cut. Then navigate up to the Edit –> Duplicate Time. This will duplicate everything during that time period of the track. I will usually then pick, Edit –> Cut Time, select where I want this section to go and then, Edit –> Paste Time. This way I can duplicate the whole verse or chorus and place it later on in the track as I am building the composition. + +Drag in Older Versions + +Let’s say you are working on a track. You save multiple versions as you go, this is just a damn good idea. You had a baseline and started to rewrite it. After a while, you realize that you liked the old baseline, but have done a lot of other things along the way. You can actually drag a track from one of the older versions. This makes it easy to revert back to a past part. + +To do this I usually navigate to the Current Project in my Browser. This shows all other versions you have saved if it is in the same project folder. If not, or you want to bring in another track from another song you can just navigate there with your browser. Next up you click the triangle next to the Live Set to open it up. You will then see all the tracks in the Live Set. Click the track you want and drag it into Live. + +This lets you easily revert to older versions and save a lot of time. + +Copy Value to Sibling + +When you are tweaking +================================================================================ +Rank = 44; Score = 4292608.0 +<|begin_of_text|>Some practical tricks for training recurrent neural networks: + +Optimization Setup + +Adaptive learning rate. We usually use adaptive optimizers such as Adam (Kingma14) because they can better handle the complex training dynamics of recurrent networks that plain gradient descent. + +We usually use adaptive optimizers such as Adam (Kingma14) because they can better handle the complex training dynamics of recurrent networks that plain gradient descent. Gradient clipping. Print or plot the gradient norm to see its usual range, then scale down gradients that exceeds this range. This prevents spikes in the gradients to mess up the parameters during training. + +Print or plot the gradient norm to see its usual range, then scale down gradients that exceeds this range. This prevents spikes in the gradients to mess up the parameters during training. Normalizing the loss. To get losses of similar magnitude across datasets, you can sum the loss terms along the sequence and divide them by the maximum sequence length. This makes it easier to reuse hyper parameters between experiments. The loss should be averaged across the batch. + +To get losses of similar magnitude across datasets, you can sum the loss terms along the sequence and divide them by the maximum sequence length. This makes it easier to reuse hyper parameters between experiments. The loss should be averaged across the batch. Truncated backpropagation. Recurrent networks can have a hard time learning long sequences because of vanishing and noisy gradients. Train on overlapping chunks of about 200 steps instead. You can also gradually increase the chunk length during training. Preserve the hidden state between chunk boundaries. + +Recurrent networks can have a hard time learning long sequences because of vanishing and noisy gradients. Train on overlapping chunks of about 200 steps instead. You can also gradually increase the chunk length during training. Preserve the hidden state between chunk boundaries. Long training time. Especially in language modeling, small improvements in loss can make a big difference in the perceived quality of the model. Stop training when the training loss does not improve for multiple epochs or the evaluation loss starts increasing. + +Especially in language modeling, small improvements in loss can make a big difference in the perceived quality of the model. Stop training when the training loss does not improve for multiple epochs or the evaluation loss starts increasing. Multi-step loss. When training generative sequence models, there is a trade-off between 1-step losses (teacher forcing) and training longer imagined sequences towards matching the target (Chiappa17). Professor forcing (Goyal17) combines the two but is more involved. + +Network Structure + +Gated Recurrent Unit. GRU (Cho14 +================================================================================ +Rank = 45; Score = 4292608.0 +<|begin_of_text|>Almost four years ago, I documented a really cool vSAN capability here and here, which demonstrates how to bootstrap a vSAN datastore onto a single ESXi host. This powerful capability, which was by design, enables customers to easily standup new infrastructure including the vCenter Server Appliance (VCSA) in a pure greenfield environment where you only had bare-metal hardware to start with and no existing vCenter Server. + +As you can probably guess, I am a huge advocate for this capability and I think it enables some really interesting use cases for being able to quickly and easily stand up a complete vSphere environment without having to rely on an external storage array or playing games with Storage vMotion'ing the VCSA between local VMFS and the vSAN datastore for initial provisioning. + +Over time, this vSAN capability has gone mainstream not only from a customer standpoint but also internal to VMware. In fact, the use of this feature has made its way into several VMware implementations including but not limited to VMware Validated Designs (VVD), VxRail, VMware Cloud Foundation (VCF) and even in the upcoming VMware Cloud on AWS. This really goes to show how useful and critical of a feature this has become for standing up brand new VMware infrastructure which runs on top of vSAN. Huge thanks goes out to the original vSAN Architects who had envisioned such use cases and designed vSAN to include this functionality natively within the product and not have to rely or depend on vCenter Server. + +So what has changed in the with the new release of vSAN 6.6 (vSphere 6.5d) For starters, this functionality continues to exists, but the Engineers and specifically I would like to call out Christian Dickmann who was the lead architect for what I am about to talk about reached out to me and asked, could we make this even better? For those of you who have walked through the vSAN bootstrap process, although pretty straight forward it still does require quite a few steps to accomplish. Even with Automation, the overall user experience could still be improved and today, this bootstrap process is only available using the CLI or calling into several different remote APIs. + +In vSAN 6.6, the vSAN bootstrap process is now natively integrated into the VCSA UI Installer which is referred to as the vSAN Easy Install feature. This means, as part of selecting an ESXi host to deploy the VCSA (for a pure greenfield deployment), you will now also be able to configure the vSAN +================================================================================ +Rank = 46; Score = 4259840.0 +<|begin_of_text|>The world is changing. + +No – the world as we knew it in IT has changed. + +Big Data & Agile are hot topics. + +But companies still need to collect, report, and analyze their data. Usually this requires some form of data warehousing or business intelligence system. So how do we do that in the modern IT landscape in a way that allows us to be agile and either deal directly or indirectly with unstructured and semi structured data? + +First off, we need to change our evil ways – we can no longer afford to take years to deliver data to the business. We cannot spend months doing detailed analysis to develop use cases and detailed specification documents. Then spend months building enterprise-scale data models only to deploy them and find out the source systems changed and the models have no place to hold the now-relevant data critical to business success. + +We have to be more agile than that. + +We need a way to do Agile Data Engineering if we ever expect to achieve the goal of Agile Data Warehousing. + +The Data Vault System of Business Intelligence provides (among other things) a method and approach to modeling your enterprise data warehouse (EDW) that is agile, flexible, and scalable. This is the first in a series of posts introducing you to the Data Vault, what it is, and how to build one. + +Why Data Vault? + +Current approaches to modeling a data warehouse include developing 3rd Normal Form (3NF) type models or dimensional star schema models. Now, while there are indeed architects and modelers out there who have been wildly successful using these approaches, there are many more that have failed, and failed big. + +Why? Mostly lack of experience, but more so that these approaches have some basic issues that, while resolvable, do require a certain level of engineering expertise that is not readily available in the industry today (and is declining daily). + +What are these issues? + +In order to collect history in the usual 3NF approach you need to add a timestamp or snapshot date to every table in your model. Not only that but it needs to be in the primary key of the table so that you do not load duplicate rows on any given day. This of course complicates all the cascading foreign key (FK) relationships. In Figure 1 you can see a simple example of a standard Product and Product Category table with Snapshot Dates added to the keys. + +Figure 1. Basic model with Snapshot Dates added + +Even with this simple model you can start to see the issues. Adding a snapshot date to every PK makes all the keys more complicated. +================================================================================ +Rank = 47; Score = 4259840.0 +<|begin_of_text|>The growing world of smartphones and the use of instant messengers are playing a bigger role in daily life than ever before, these IM apps are connecting people with each other conveniently and freely in the digital world. Moreover, the use of instant messengers and the user’s adoption of Android devices on a larger scale is creating fear in the minds of the parents. So, the rain of instant messengers, mostly youngsters are involved in the usage of different kinds of instant messengers on the same Android device. Having said that, parents are fearful because of plenty of IM apps on their young kid’s mobile devices; it makes it difficult for parents to monitor multiple instant messengers on a particular Android device. There are a number of instant messengers but the most popular are Facebook Messenger, WhatsApp, Viber, Snapchat, Kik, Line, IMO, and Skype. Monitoring of a particular messenger is not really difficult but spying on multiple instant messengers on Android devices seems to be a problem. It’s because the best instant messenger’s apps work using similar methodology and the actual difference is the user interfaces and having some extra functions of every app. TheOneSpy (TOS) will let you monitor multiple IM apps on Android devices, you will be able to spy on up to 16 instant messaging application with the help of TOS monitoring app. All the instant messaging apps have very secure services; therefore to spy on multiple messengers, we have to apply some tactics to monitor multiple IMs. + +How is that possible? + +In order to monitor more than one instant messenger, we need rooting of our Android devices. Now the question is, how do we root the devices; It’s actually simple, we can use a software which can easily root almost any device. It’s called “KingoRoute” and is a free application. So, use this application and after the completion of the rooting of your device, you will enable to monitor multiple instant messengers with the help of TOS monitoring application. + +Why is the TheOneSpy the best option for spying multiple Messengers in Android devices? + +It’s because the TheOneSpy enables you to monitor multiple instant messaging applications at once, having single control panel you can monitor all of the conversation happening on multiple messengers on the single platform with complete and accurate time statistics. TOS is compatible with Android 4.0 up to Android 7.0 Nougat. TOS is reasonably priced, starting from $0.6/day. You might be able to find some discounts if you google it. Following +================================================================================ +Rank = 48; Score = 4227072.0 +<|begin_of_text|>Without affecting the original signal, we removed all of the outliers, yet introduce new ones. That is the shortcoming of the method, since change is a lot in a small interval(around peaks), this method introduces medians even if there is no outliers. The unwanted signal is closer to original signal, so there is still improvement. Further, this is easy to prevent with another heuristics. We will only look at the signal that behaves regularly or using a sliding window to have a more robust approach. But this example is important to show how a non-linear filter is able to reject all of the outliers without doing anything very complicated. This has due to mainly two reasons. First one is that $l_1$ norm. Instead of using $l_2$, we use $l_1$ as this norm does not change too much when an outlier gets introduced to the signal whereas $l_2$ norm(since it takes the square of the distance) outliers have much more profound effect. We are preventing using $l_1$ norm. Second one is that we are filtering adaptively. Instead of using a mean filter, and average all of the numbers, we only look at the candidates that could be outliers. First, we get convinced those points are outlier, then we attempt to remove them. This makes it hard to reject the regular signal. What about disadvantages? First and foremost, it is nonlinear filtering. There is no going back. You cannot get the original signal that you begin with. Further, we are also applying a threshold which makes it even harder. There are two important parameters, threshold and sliding window which requires some parameter tuning. But generally outliers are quite either signal specific or application specific. Therefore, this may not be a disadvantage.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 49; Score = 4227072.0 +<|begin_of_text|>It’s no secret that the installations at Burning Man are highly innovative and creative. The festival in Black Rock, Nevada, which is happening right now, promises to offer an escape from the regular constraints of society. This year, the festival is bigger than ever, featuring high-profile producers like Skrillex and Diplo and an all-new live stream to capture moments of the event. One creative group is taking it one step further by converting an airplane into a nightclub. + +The Big Imagination Foundation spent several months turning this Boeing 747 into what has been described as the “largest moving art experience ever made”. The project will stand on the Playa close to other art installations. The inside of the airplane will feature music and art. It will also travel around Black Rock City, pulled by an aircraft tug car. + +Inside the airplane, the experience begins once the participants answer the question “What baggage do you need to lose?”. An “Emotional Baggage Tag” is then created along with the boarding pass and passengers are later asked “Where are you going?” The answers to this question are displayed on the walls inside the airplane. The immersive installation is sure to be a psychological and liberating experience for attendants. + +For those of us that are wishing we were there, the Big Imagination Foundation is looking to develop a virtual reality experience for all. They are currently crowd-funding through their IndieGogo campaign. + +H/T: The Creator\’s Project<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 50; Score = 4227072.0 +<|begin_of_text|>In graph like data structures (explicit or implicit) there is a problem of memory safety when you remove a node. You have to clean up all the connections pointing to you. If your graph structure is explicit this is easy, otherwise it can often be annoying to find all pointers pointing to you. A while ago I thought that you could solve that by introducing a two way pointer, which has to be the same size as a normal pointer, but which knows how to disconnect itself on the other side, so that you don’t have to write any clean up code. + +The reason why it turned out surprisingly useful is that if both ends of the connection know about each other, I can make them keep the connection alive even if one of them is moved around. So if one of them lives in a std::vector and the vector reallocates, the TwoWayPointer can handle that case and just change the pointer on the other end to point to the new location. That can dramatically simplify code. + +One example is that I’m now using it in my signal/slot class which I’ve also included in this blog post. It allowed me to simplify the signal/slot implementation a lot. Compare it to your favourite signal implementation if you want. There is usually a lot of code overhead associated with memory safety. That all went away in this implementation. + +Here is the code for the two way pointer: + +#pragma once #include #include template struct TwoWayPointer { private: TwoWayPointer * ptr; template friend struct TwoWayPointer; template friend bool operator==(const TwoWayPointer &, const TwoWayPointer &); template friend bool operator!=(const TwoWayPointer &, const TwoWayPointer &); template friend bool operator==(const TwoWayPointer &, std::nullptr_t); template friend bool operator!=(const TwoWayPointer &, std::nullptr_t); template friend bool operator==(std::nullptr_t, const TwoWayPointer &); template friend bool operator!=(std::nullptr_t, const TwoWayPointerFAQ + +Is CLIC Marble made from real marble? + +Yes. Our CLIC Marble is made from the highest-quality white marble or black marble. The case features a very thin slice of marble to ensure it is lightweight and doesn’t add bulk to your iPhone. + +How much does CLIC Marble weigh? + +CLIC Marble for iPhone 6/6s weighs 31 grams (1.1 ounces) — in comparison, CLIC Wooden for iPhone 6/6s weighs 23 grams (0.8 ounces), and an iPhone 6 weighs 136.5 grams (4.8 ounces). CLIC Marble for iPhone 7 and 8 weighs 37 grams (1.3 ounces) — in comparison, CLIC Wooden for iPhone 7 and 8 weighs 28 grams (1 ounce), and an iPhone 7 weighs 138 grams (4.7 ounces). CLIC Marble for iPhone 7 Plus and 8 plus weighs 49 grams (1.7 ounces) — in comparison, CLIC Wooden for iPhone 7 Plus and 8 Plus weighs 35 grams (1.2 ounces), and an iPhone 7 Plus weighs 188 grams (6.6 ounces). + +What can I expect from the finish? + +Marble is a natural stone formed over thousands of years. Each slice has unique and intricate veins, so the finish will vary from case to case. + +Is the CLIC Marble shatter resistant? + +Yes. We deliberately used a very thin slice of marble laminated to a layer of fiberglass, giving the marble a degree of flexibility. This makes it easier to take on and off your iPhone and resistant to shattering. + +How do I care for my marble? + +Although a solid natural stone, marble is in comparison to other stones quite soft and porous. As such, care should be taken to keep your marble looking its best. To clean your marble, use warm water, a soft cloth, and a mild soap such as washing-up liquid. Never use anything acidic or abrasive to clean your marble, including natural solutions vinegar or lemon, which will dull the surface of the marble. Apply marble polish, available from hardware stores, on a regular basis to maintain the stone’s luster. + +Is the packaging recyclable? + +We’re proud to say that all our packaging is 100% recyclable, so you can enjoy our great products without hurting the planet.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 52; Score = 4194304.0 +<|begin_of_text|>InfluxDB : InfluxDB is an open-source time series database written in Go that has been built to work best with metrics, events, and analytics. Using InfluxDB, you can easily store system and application performance data and manage any time series data. + +Grafana : Grafana is an open-source, general purpose dashboard that is used for visualizing time series data for Internet infrastructure and application analytics. Grafana supports graphite, influxdb or opentsdb as backends and runs as a web application. + +In this tutorial, we will learn how to create and run Grafana and InfluxDB Docker containers in Ubuntu 14.04. + +Requirements + +A server running Ubuntu-14.04 with Docker installed. + +A non-root user with sudo privileges setup on server. + +Creating The Dockerfile + +First, you will need to create the Docker file to install all requisite software. Create docker file inside your home directory using the following command: + +sudo nano Dockerfile + +Add the following lines with all requisite software: + +FROM ubuntu MAINTAINER Hitesh Jethva (hitjethva@gmail.com) RUN apt-get update && apt-get -y --no-install-recommends install ca-certificates software-properties-common python-django-tagging python-simplejson python-memcache python-ldap python-cairo python-pysqlite2 python-support python-pip gunicorn supervisor nginx-light nodejs git curl openjdk-7-jre build-essential python-dev + +Add the following lines to install Grafana, InfluxDB, and do some basic configuration: + +WORKDIR /opt RUN curl -s -o /opt/grafana-1.8.1.tar.gz http://grafanarel.s3.amazonaws.com/grafana-1.8.1.tar.gz && curl -s -o /opt/influxdb_latest_amd64.deb http://s3.amazonaws.com/influxdb/influxdb_latest_amd64.deb && mkdir /opt/grafana && tar -xzvf grafana-1.8.1.tar.gz --directory /opt/grafana --strip-components=1 && dpkg -i influxdb_latest_amd64.deb && echo "influxdb soft nofile unlimited" >> /etc/security/limits.conf && echo "influxdb hard nofile unlimited" >> /etc/security/limits.conf + +Next, copy some configuration files: + +ADD config.js /opt/grafana/config.js ADD nginx.conf /etc/nginx/nginx.conf ADD supervisord.conf /etc/supervisor/conf.d/superv +================================================================================ +Rank = 53; Score = 4177920.0 +<|begin_of_text|>In everyday life, I'm a web developer. Or, to be precise, I run a business that develops websites for a wide range of clients, from small businesses to large organizations. Every one of these sites comes with a CMS of some sort. Which CMS we use to develop the sites depends on a lot of factors, including what the client wants, the size of the website, and the required functionality. In this article, I'll cover the lessons learned when we developed our open source Bolt content management system. + +Keep the customer in mind + +Often when working on larger site, you'll work with a small group of people with different job descriptions and skill sets. This overlap between people with different areas of expertise can cause some problems when interests collide: + +An editor doesn't want to know about database structures. They don't care if something is XML, JSON, or Mediumtext. + +The backend developer shouldn't concern themselves with the exact markup that's used on the website. or , that's pretty much the same, right? + +Front end developers shouldn't need to know about scheduling new articles or whether or not the editor-in-chief has approved a change to an older article. + +I strongly believe that there's no single CMS that's well suited to run all websites, and I'm always surprised that there are a lot of web development agencies that use the same CMS for every single project they do. If I were a client of an agency like that, I'd wonder if the CMS I'm getting is really the best fit for my website, or if it just happens to be the CMS they use for all their clients. + +Because this is not how we would like to work, we've been looking around for CMSs that fit our workflow and, at the same time, complement each other. This way we have a few options to better help our clients without falling back to the same system every time. For most of the larger projects we develop, we use Drupal. It's a great system with a lot of functionality. Because of this, it's also a rather complex system (Especially for the editors, who eventually have to work in the system on a daily basis). We've done a lot of research to find a system that complements this: A simpler, more lightweight system for sites that don't require the extensive functionality offered by Drupal. + +Every system that we've evaluated when looking for this had one or more major downsides. It seems like every single CMS out there is written by a certain kind of person for a certain +================================================================================ +Rank = 54; Score = 4161536.0 +<|begin_of_text|>That's not all the article says. But since you failed to completely read my own first paragraph, I'm not surprised you didn't get through the article. I'll let you go re-read it. Also, all I'm seeing is still promising. You've yet to present any evidence whatsoever of them being harmful. + +Their use in this way has been studied somewhat (again..check the article), but obviously we don't yet know long-term effects for this particular use. But as you quoted, they have been safely used as a pause button in other contexts. There's no reason to think this wouldn't be safe. + +We have no reason to believe that medically halting puberty would affect someone mentally. Again, if people who naturally hit puberty late mature mentally the same, why should medical intervention be any different? All they're affecting is the hormones that spur the development of secondary sex characteristics. Your concerns are based in nothing but your own bias. + +It is very easy to change your mind based on all we know about different circumstances. You know, the ones where they've been used safely? Here's a crazy thing you keep forgetting. It's the child's choice. Puberty occurring naturally is something terrifying for transgender kids. It's a train they can't stop. Until puberty blockers. If a kid is getting uncomfortable with their body not developing they can be taken off the blockers. It's an easy fix. I'll refer you to the logic problem from earlier. The same train of thought here applies to why something that is easy to reverse is a better course than something that is hard to reverse. But since you're a broken record, I know you're just going to say "we don't know they're easy to reverse," but yeah. We do. Which is how they've been, again, used safely in other contexts. + +These kids might be helped in other ways. Or we could help them in a way we know works. Where the only safety concerns about them come from people who are against transgender people in the first place (I tried to resist going through your comment history, but it's not hard to spot an r/Gender_Critical user. Also great to see you're the same asshole who posted that whorephobic pole dancing article). + +Your whole last paragraph is just repetition of the same unfounded fears your entire argument is based on. Maybe take a step back and realize while there's actually evidence for the safety of puberty blockers, your entire opinion is based on fear and hate. Why don't you take a second +================================================================================ +Rank = 55; Score = 4128768.0 +<|begin_of_text|>Image caption The NSA wants to use its quantum computer to break encryption used to protect online communication + +The US National Security Agency is building a quantum computer to break the encryption that keeps messages secure, reports the Washington Post. + +The NSA project came to light in documents passed to the newspaper by whistle-blower Edward Snowden. + +The spying agency hopes to harness the special qualities of quantum computers to speed up its code-cracking efforts. + +The NSA is believed to have spent about $80m (£49m) on the project but it has yet to produce a working machine. + +If the NSA managed to develop a working quantum computer it would be put to work breaking encryption systems used online and by foreign governments to keep official messages secure, suggest the documents excerpted in the Post. + +The quantum computer is being developed under a research programme called Penetrating Hard Targets and is believed to be conducted out of a lab in Maryland. + +Processing power + +Many research groups around the world are pursuing the goal of creating a working quantum computer but those developed so far have not been able to run the algorithms required to break contemporary encryption systems. + +Current computers attempt to crack encryption via many different means but they are limited to generating possible keys to unscramble data one at a time. Using big computers can speed this up but the huge numbers used as keys to lock away data limits the usefulness of this approach. + +By contrast, quantum computers exploit properties of matter that, under certain conditions, mean the machine can carry out lots and lots of calculations simultaneously. This makes it practical to try all the possible keys protecting a particular message or stream of data. + +The hard part of creating a working quantum computer is keeping enough of its constituent computational elements, called qubits, stable so they can interact and be put to useful work. + +The NSA is not believed to have made significant breakthroughs in its work that would put it ahead of research efforts elsewhere in the US and Europe. However, the documents passed to the Post by Edward Snowden suggest the agency's researchers are having some success developing the basic building blocks for the machine.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 56; Score = 4063232.0 +<|begin_of_text|>Patients used to pay $10,000 a year for cancer drugs, but now they’re paying $10,000 a month, according to HBO talk show host Bill Maher. + +That’s a message circulating on Facebook recently, promoted by the "Bill Maher Fanpage." A reader asked us to check it out. + +The quote comes from the Oct. 2 episode of Real Time with Bill Maher, where Maher discussed Turing Pharmaceuticals, the company that jacked up the cost of a life-saving drug from $13.50 to $750 essentially overnight. + +"Not that (Turing CEO Martin) Shkreli is unusual for the pharmaceutical industry," Maher said on his show. "Fifteen years ago, cancer drugs cost an average of $10,000 a year. Now it’s $10,000 a month because this cartel owns the U.S. government every bit as much as Mexican drug lords own theirs." + +It’s well known that the cost of health care is on the rise, but has the annual cost of cancer drugs really increased so dramatically in just a decade and a half? + +Maher’s team sent us links to some news articles about the rising cost of cancer drugs, and we tracked down information that speaks to his claim. + +The best information we could nail down is the cost of anticancer drugs at launch, the initial price when the Federal Drug Administration approves a drug. + +These are the sticker prices set by the pharmaceutical companies, not the prices patients actually pay out of pocket, which are reduced by insurance payments, patient assistance programs and other discounts. Costs increases, though, are often passed on to consumers. + +Maher has a point that cancer drug costs are on the rise. We charted the launch costs of cancer drugs the FDA approved since 1999, using a list compiled by Memorial Sloan Kettering Cancer Center. + +The second half of Maher’s statement -- that cancer drugs now cost about $10,000 a month -- is also on solid ground. Of drugs approved in 2015, the monthly average cost is $11,319. And experts in the field often cite that $10,000 figure themselves. + +Because many cancer drugs are not administered to patients for a full year, it’s difficult to determine an average annual cost, so the first half of Maher’s statement -- that they cost about $10,000 a year 15 years ago -- is hard to nail down. To his point, though, the monthly cost of a new drug was dramatically lower in 2000 than what it is now at about $4, +================================================================================ +Rank = 57; Score = 4030464.0 +<|begin_of_text|>You must enter the characters with black color that stand out from the other characters + +Message: * A friend wanted you to see this item from WRAL.com: http://wr.al/14Iqb + +— Customers that purchased a ticket through Ticketmaster between late 1999 and early 2013 could be eligible for free tickets to a number of events. + +An email sent to eligible Ticketmaster customers includes instructions on how to get vouchers for free tickets to selected events as well as discounts on Ticketmaster purchases. Those who bought a ticket through the company between Oct. 21, 1999, and Feb. 27, 2013, are eligible. The vouchers expire in four years. + +The vouchers are the result of a class-action lawsuit over ticket fees and other charges. + +Frank Manginaro, of Chatham County, is one of about 57 million people who recently received a voucher. + +He said he was thrilled until the site started giving him an error message when he attempted to use the voucher code. + +"All of a sudden they are giving me excuses saying it doesn't work, you can't buy tickets, try again later," he said. "I was upset about it, and so were my friends because they couldn't even get through." + +After trying for two days, Manginaro tried again and found a way to get tickets. He said he ignored the voucher link on the Ticketmaster website and went straight to the band's page. + +"I just skipped over it and went straight to the Counting Crows and ordered it through there. It went through with no problem," he said. + +According to the settlement, Ticketmaster is required to distribute 42 million dollars in vouchers in the next four years. + +Ticketmaster is now part of Beverly Hills-based Live Nation Entertainment Inc.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 58; Score = 3997696.0 +<|begin_of_text|>Micro Service Architecture is an architectural concept that aims to decouple a solution by decomposing functionality into discrete services. Think of it as applying many of the principles of SOLID at an architectural level, instead of classes you've got services. + +Conceptually speaking MSA is not particularly difficult to grasp but in practice it does raise many questions. How do these services communicate? What about latency between services? How do you test the services? How do you detect and respond to failure? How do you manage deployments when you have a bunch of interdependencies? So lets expand on some of these throughout this post and see if MSA really is worth the effort. + +Anatomy of a Micro Service + +First things first what actually is a micro service? Well there really isn't a hard and fast definition but from conversations with various people there seems to be a consensus that a micro service is a simple application that sits around the 10-100 LOC mark. Now I realise that line count is an atrocious way to compare implementations so take what you will from that. But they are small, micro even. This means you're not going to find hundreds of tiny services built on top of large frameworks, it's simply not practical. No, simplicity and lightweightyness (not a real word) is the order of the day here. Small frameworks like Sinatra, Webbit, Finagle & Connect do just enough to allow you to wrap your actual code in a thin communication layer. + +In terms of a footprint these services will be small, you're potentially going to run a lot of them on the same machine so you don't want to be holding on to memory or resources that you aren't intending to use. Once again simple libraries over large frameworks will win out, you'll also find less of a reliance on 3rd party dependencies. + +This decoupling at a service level also offers another interesting option. We've pushed a lot of the old application complexity down to infrastructure level. We are no longer bound to a single stack or language. We can play to the strengths of any stack or language now. It's entirely possible to have a system built out with a myriad of languages and libraries, though as we will touch on later this is a double edged sword. + +You're also not going to find any true micro service based architectures that are hosted in an application server, that kinds of defeats the point. To this end micro services self host, they grab a port and listen. This means you'll lose any benefits your typical enterprise application server may bring and your service +================================================================================ +Rank = 59; Score = 3981312.0 +<|begin_of_text|>If I asked you to name major Java vendors, chances are EMC's VMware subsidiary wouldn't top the list. To most of us, the "VM" in VMware doesn't stand for the Java Virtual Machine; it means the other kind of virtual machine, the kind that lets you run servers and desktops on virtualized OS instances. But if that's your assumption, it may be time to change your thinking. + +VMware is reinventing itself. Virtualization may have been the hot topic a few years ago, but its star faded once the OS vendors got into the game. Now cloud computing is the buzzword of the day. Observing the trends, Redmonk analyst James Governor went as far as to declare "cloud [computing] is the new VMware." Naturally, then, the new VMware is cloud computing -- and that puts Java square in its sights. + +[ Subscribe today to InfoWorld's Developer World newsletter and stay up to date on the key software development news and insights. ] + +At the Google I/O conference in San Francisco this week, VMware and Google announced a partnership aimed at making it easier for Java developers to deploy applications on the Google App Engine cloud computing service. This new deal comes hot on the heels of a similar arrangement VMware struck with Salesforce.com in late April, in which VMware provided infrastructure to allow Java apps to run on Salesforce's Force.com platform. + +With these partnerships, VMware is putting its customers on notice that it's not merely a virtualization vendor, but a full-service provider of solutions for cloud computing. But developers should take notice, too, because if VMware continues down this path, it could potentially mean a whole new era for the Java platform. + +Write once, cloud everywhere + +Sun Microsystems dubbed Java the "write once, run anywhere" language, but Java developers have always taken that phrase with a grain of salt. Early in the platform's history, Microsoft deliberately tried to fragment the Java market by shipping its own, incompatible version -- but even Sun's own JVM could be inconsistent from OS to OS. Apple, once a Java backer, dropped support for the language from its Mac OS X developer platform and won't allow it on the iPhone or iPad. The Java ME market is hopelessly fragmented. Browser applets are dead. + +The bright spot for Java has always been the datacenter. But as I've mentioned before, cloud computing has the potential to fragment the server-side Java market, by encouraging developers to code their applications to a specific cloud vendor's services and requirements. For example, unlike Force.com, +================================================================================ +Rank = 60; Score = 3899392.0 +<|begin_of_text|>Anyone who had to schedule an intercontinental phone call knows that there is no such thing as a simple time called now. What you should rather think about is a time comprised of here and now. + +The Earth rotates around its own axis. When it’s solar noon (the sun is at its highest position) in one place, it’s already past noon in places to the east and it’s still before noon in places to the west. + +To make communication easier, at the end of the 19th century, the Earth was divided into 24 hour-wide time zones. All places within one time zone have the same time. It was decided that the “0” time zone would be London’s time zone. The “0” time zone is nowadays called UTC (Coordinated Universal Time) or, for historic reasons, GMT (Greenwich Mean Time). In general, the time zones’ boundaries follow countries’ boundaries though some large countries are located in multiple time zones. For example, USA is located in 7 different time zones, Canada in 6, but China is in one time zone. There are time zones with offsets of 30 minutes or even 45 minutes (Nepal). To make things even more complicated, many countries also have daylight saving time: they turn the clock back or forward one hour on specific days to account for seasonal changes in daylight. + +If your computer system is international, it has to be ready to handle users from different time zones. The users want to see the time of system events expressed in their local time zone: + +In an online course, you want to see the assignment deadline in your own time zone. You don’t want to submit your assignment and find out that the deadline has passed. + +In an auction system you want to know at exactly what time the auction ends. You don’t want to try bidding and find out that the auction ended three hours earlier. + +You want to know the time when a promotion ends. Ideally, if the promotion ends on a specific day, you want the promotion to run until the end of the day in UTC-11 time zone. + +All databases have types which handle both date and time. The types are called timestamp or datetime, or something similar. They usually come in two flavors: one without time zone and the other with time zone. + +Timestamp without time zone is generally a fancy string value. It does not interpret the time zone. If you enter the value “2015-01-28 13:00:00,” the same value will be shown to all users. + + +================================================================================ +Rank = 61; Score = 3866624.0 +<|begin_of_text|>Interpreted, dynamically typed languages such as Python, Ruby or Javascript are: easy to use, allow fast development and have a great number of libraries. Unfortunately, they are not very good when it comes to performance. + +For this reason, I felt the need to look for something different for some projects. However, I didn't want to lose the simplicity other languages gave me. Essentially, I wanted something that could give me: + +Good performance + +Easy concurrency + +Easy error handling + +Safety + +Fast development time + +Good libraries to ease my development + +Go was the obvious choice - I had previously worked with it and knew I could achieve all of these points. + +So what is Go? + +From the Go documentation page + +Go is an attempt to combine the ease of programming of an interpreted, dynamically typed language, with the efficiency and safety of a statically typed, compiled language. + +An attempt which I believe was mostly successful. But let's dive into some points to understand Go's beauty. + +Simple and easy + +Go is a "simple" language - it only contains 25 keywords (Python 3.5.0: 33, Ruby 2.2.0: 41, ANSI C: 32). Obviously, a small number of keywords doesn't necessarily mean it's easy - Brainfuck only has 8 and its name suggests how easy it is to program in it. + +And Go does have trade-offs for using such a small number of keywords. For example, it only has one loop (keyword for ), which can be seen has a nice thing since you really only have to know one word. But it can take several forms: + +for condition behaves like a while + +behaves like a while for i,v := range list ranges over a list + +ranges over a list for i=0;i<10;i++ for a C style classical for + +which might confuse/annoy some people. + +However, I do believe Go is easy to understand and learn, and you can see it for yourself by doing the + +Golang Tour, which will teach you the essentials. + +Standard Library + +Furthermore, Go has got a robust standard library which allows quick and easy development of common essential tasks: + +net (http, mail) + +archives (zip, tar, gzip) + +encodings (json, csv) + +cryptography (aes, des, md5) + +html template system + +generic sql drivers + +logging system + +and many others... + +Easy to read + +Any programmer can tell you that reading other people's (including your past self) code is a daunting, hair- +================================================================================ +Rank = 62; Score = 3833856.0 +<|begin_of_text|>There are some reasons why anyone will prefer visiting the casinos in Malaysia than with most other places. Malaysian casinos have several advantages over most casinos available other places in the world. These advantages make them popular in so many ways over the rest of casinos in other places. Below is an outline of the reasons why the next time you are out gambling you should think about hitting the Malaysia casino. + +Different Games + +In Malaysia, you will get to enjoy some games from blackjacks, betting to poker among others. With so many games to choose from, it is hard to get bored playing. If one game does not appeal to you, you can move on to the next one and get to enjoy. Having a pool of games form which to choose from will always make the gambling interesting. It will also attract different kinds of people. Having different people to play against will in a great way increase the chances that you have of making a good return. + +Great Payouts + +Another thing that makes most people prefer playing in Malaysian casinos is because of the great payouts that are offered. From the games that are available, it is quite easy t even double your initial investment in just a few minutes of playing. You will get to enjoy great payouts that in most cases, you will win than lose. Having this in mind will always make gambling somewhat appealing. All that one needs to do is to exercise caution while playing. + +Security + +Malaysian casinos are some of the most secure casinos in the world. For a gambler, security is crucial since they deal with money. One needs to be sure that the money they win will be in safe hands. A casino that ensures that the player is protected will always be a preferred option by the players. Some of the casinos even go to extra lengths of providing escort for their players so that they do not get robbed. + +Privacy + +A casino player will always value their privacy. Malaysian casino knows and understands this. Great measures are usually taken to ensure that people who request to remain anonymous remain so. Some people come from high-end families or are powerful people in the country. They might be people who wouldn’t want their identities revealed. When this is the case, they would want to still enjoy their game without constantly worrying who might see them. Malaysian casinos will do all that is possible to ensure that the identity of such people is protected. + +From the above, it is evident that Malaysian casinos are some of the best in the world. There is so much that they have put in to ensure that their clients have a wonderful experience at all +================================================================================ +Rank = 63; Score = 3833856.0 +<|begin_of_text|>Let be a natural number. A basic operation in the topology of oriented, connected, compact, -dimensional manifolds (hereby referred to simply as manifolds for short) is that of connected sum: given two manifolds, the connected sum is formed by removing a small ball from each manifold and then gluing the boundary together (in the orientation-preserving manner). This gives another oriented, connected, compact manifold, and the exact nature of the balls removed and their gluing is not relevant for topological purposes (any two such procedures give homeomorphic manifolds). It is easy to see that this operation is associative and commutative up to homeomorphism, thus and, where we use to denote the assertion that is homeomorphic to. + +(It is important that the orientation is preserved; if, for instance,, and is a chiral 3-manifold which is chiral (thus, where is the orientation reversal of ), then the connect sum of with itself is also chiral (by the prime decomposition; in fact one does not even need the irreducibility hypothesis for this claim), but is not. A typical example of an irreducible chiral manifold is the complement of a trefoil knot. Thanks to Danny Calegari for this example.) + +The -dimensional sphere is an identity (up to homeomorphism) of connect sum: for any. A basic result in the subject is that the sphere is itself irreducible: + +Theorem 1 (Irreducibility of the sphere) If, then. + +For (curves), this theorem is trivial because the only connected -manifolds are homeomorphic to circles. For (surfaces), the theorem is also easy by considering the genus of. For the result follows from the prime decomposition. But for higher, these ad hoc methods no longer work. Nevertheless, there is an elegant proof of Theorem 1, due to Mazur, and known as Mazur’s swindle. The reason for this name should become clear when one sees the proof, which I reproduce below. + +Suppose. Now consider the infinite connected sum + +This is an infinite connected sum of spheres, and can thus be viewed as a half-open cylinder, which is topologically equivalent to a sphere with a small ball removed; alternatively, one can contract the boundary at infinity to a point to recover the sphere. On the other hand, by using the associativity of connected sum (which will still work for the infinite connected sum, if one thinks about it carefully +================================================================================ +Rank = 64; Score = 3817472.0 +<|begin_of_text|>Subdomains can be a tricky thing to work with. Here are some tips to develop and test Rails applications using them. + +Develop + +During development there are a few ways you can handle subdomains. If you are using the bundled webrick server, you won't be able to access your subdomain at all. + +Using lvh.me + +If you are using the webrick server or something like Puma for development you can use lvh.me to access your subdomains. e.g. + +http://sub.lvh.me:9292/ + +'lvh.me' resolves to localhost, so this provides access to the subdomain in Rails on port 9292 (the Puma default). This is convenient but can be very slow to work with. + +Editing your hosts file + +To avoid having to do a round trip to the internet you can add a record to your machine hosts file. In Unix/mac you can find it on /etc/hosts + +You can add something like: + +127.0.0.1 sub.virtual.local + +Then you should be able to access your subdomain using the built in web server or puma by going to: (again 9292 is the port used by puma): + +http://sub.virtual.local:9292/ + +Using Pow (Only available for Mac) + +Pow makes it extremely easy to work with subdomains: - Install pow - Link your application - Access the application using a subdomain e.g. sub.app.dev + +Have a look at the powder gem, it makes installing pow even easier. + +Testing + +Capybara needs to hit a real webserver in order to test your application. So you will have the same challenge as when accessing it on development. + +CI + +The easiest way is to use `lvh.me' for this (as long as your CI server can resolve it). + +Add a couple of methods accessible by your tests, e.g. in spec/support/subdomains.rb : + +def switch_to_subdomain ( subdomain ) # lvh.me always resolves to 127.0.0.1 Capybara. app_host = "http:// #{ subdomain }.lvh.me" end def switch_to_main_domain # Capybara.app_host = "http://lvh.me" end + +Then in your feature test you can do: + +switch_to_subdomain ('sub' ) + +Testing on your local machine + +The above approach should work on your local machine as well, but it still wasteful as it is doing a round trip to lvh.me. I found it useful to do the +================================================================================ +Rank = 65; Score = 3817472.0 +<|begin_of_text|>Many software developers find they need to store hierarchical data, such as threaded comments, personnel org charts, or nested bill-of-materials. Sometimes it’s tricky to do this in SQL and still run efficient queries against the data. I’ll be presenting a webinar for Percona on February 28 at 9am PST. I’ll describe several solutions for storing and querying trees in an SQL database, including the design I call Closure Table. + +In Closure Table, we store every path in a tree, not only direct parent-child references, but also grandparent-grandchild, and every other path, no matter how long. We even store paths of length zero, which means a node is its own parent. So if A is a parent of B, and B is a parent of C and C is a parent of D, we need to store the following paths: A-A, A-B, A-C, A-D, B-B, B-C, B-D, C-C, C-D, D-D. This makes it easy to query for all descendants of A, or all ancestors of D, or many other common queries that are difficult if you store hierarchies according to textbook solutions. + +CREATE TABLE TreePaths ( + +ancestor CHAR(1) NOT NULL, + +descendant CHAR(1) NOT NULL, + +length INT NOT NULL DEFAULT 0, + +PRIMARY KEY (ancestor, descendant) + +) ENGINE=InnoDB; + +Because there isn’t much written about using the Closure Table design, I periodically get questions about how to solve certain problems. Here’s one I got this week (paraphrased): + +I’m using Closure Table, which I learned about from your book and your presentations. I can neither find nor invent a pure-SQL solution for moving a subtree to a new position in my tree. Right now, I’m reading the nodes of the subtree into my host script, deleting them and re-inserting them one by one, which is awful. Are you aware of a more efficient solution for moving a subtree in this design? + +Moving subtrees can be tricky in both Closure Table and the Nested Sets model. It can be easier with the Adjacency List design, but in that design you don’t have a convenient way to query for all nodes of a tree. Everything has tradeoffs. + +In Closure Table, remember that adding a node involves copying some of the existing paths, while changing the endpoint for descendant. When you’re inserting a single new node, you just have one descendant to add, joined to all the paths of its ancestors. + +Here +================================================================================ +Rank = 66; Score = 3801088.0 +<|begin_of_text|>Every time someone tells me, “This database is mission critical – we can’t have data loss or downtime,” I just smile and shake my head. Technology is seriously difficult. + +To illustrate, here’s a collection of client stories from the last few years: + +The DBCC CHECKDB job ran every week just like it was supposed to – but it failed due to corruption every week. No one got email alerts because the SQL Agent mail was no longer valid – internal email server changes meant the mail was just piling up in SQL Server. CHECKDB had been failing for three years, longer than the backups were kept. Data was permanently lost. The DBA configured his backups to write to a file share. The sysadmins never understood they were supposed to back up that file share. When the DBA asked for a restore, he was surprised to find there were no backups. Three SQL Servers were all replicating data to each other. When I asked the DBA where the backups were run, he looked at one server, then another, then the third. He sheepishly admitted – in front of his manager – that there were no backups done anywhere. The DBA set up full backups daily, plus log backups of all databases in full recovery mode. Later, she put a few databases into simple recovery mode in order to fix an issue. She forgot to put them back into full recovery mode. When problems struck and she needed to recover a database, she lost all data back to the prior full backup. The SQL Server ran out of space on the C drive. During emergency troubleshooting, someone deleted a bunch of BAK files. The server started up, but databases were offline and corrupt. Turned out the user databases were on the C drive, as were all of the backups – the very backups that were just deleted to free up space. The DBA started getting odd corruption errors on one of his servers, then more, and quickly all of them. The SAN admin had flashed the storage with new firmware – which had a bug. The DBA was writing his backups to that same SAN, and sure enough, some of the corrupt databases had corrupt backups too. The admin wanted to restore the production databases onto another server. He tried, but it kept saying the files were in use. He stopped the SQL Server service, deleted the files, started it again, and finally his restore worked – but his phone lit up. Turned out he’d remote desktopped into the wrong server – he was on production. The developer did a deployment on +================================================================================ +Rank = 67; Score = 3784704.0 +<|begin_of_text|>To extend the functionality of WordPress most people have only heard of the use of plugins. Not many people have heard of the term Dropins. WordPress has it's core functionality which can be added to by the use of plugins which take advantage of multiple WordPress hooks and actions, but it also allows you to replace functionality with the use of Dropin files. Unlike a plugin the Dropin file will not need to be activated and will become activate when it is placed in the wp-content folder. This folder will by default be at the root of your WordPress install, but can be defined by changing the constant variable WP_CONTENT_DIR. + +// Custom content directory define( 'WP_CONTENT_DIR', dirname( __FILE__ ). '/wp-content' ); define( 'WP_CONTENT_URL', 'http://'. $_SERVER['HTTP_HOST']. '/wp-content' ); + +An example of how WordPress will include these Dropin files can be seen in the WordPress core code, here is an example of displaying the maintenance.php as you can see it uses the constant WP_CONTENT_DIR. + +if ( file_exists( WP_CONTENT_DIR. '/maintenance.php' ) ) { require_once( WP_CONTENT_DIR. '/maintenance.php' ); die(); } + +When you place your Dropin files in the wp-content folder you can see this from the plugin maintenance screen /wp-admin/plugins.php?plugin_status=dropins. To get a list of available dropin's there is a function in /wp-admin/includes/plugin.php called _get_dropins(), this will return the following list of dropin's. ## Single Site Install + +advanced-cache.php - Advanced caching plugin. Allows you to replace the caching functionality of your WordPress site. Activated by defining a constant variable WP_CACHE in the wp-config.php file. + +db.php - Custom database class. Used to create you own database class. Activated on load. + +db-error.php - Custom database error message. Used to display your own custom database error message. Activated on load. + +install.php - Custom install script. Used to customise your own WordPress install script. Activated on load. + +maintenance.php - Custom maintenance message. Used to create your own WordPress custom message. Activated on load. + +object-cache.php - External object cache. Used to create your own object caching class. Activated on load. + +Additional Multisite Dropin's<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 68; Score = 3784704.0 +<|begin_of_text|>Most likely, as a reader you are mostly interested in my scripts. Here are my scripts for setting up Windows Server 2012 R2 ready for WebDeployments and running EPiServer ASP.NET site! + +Why automate the installation? + +This blog is about how to make your brand new Windows Server ready for WebDeployments with just pressing enter once. Cloud services can make your infrastructure lifecycle handling very easy; still people often encounter situations where they need to host our ASP.NET applications in virtual machines or directly on physical hardware. On those situations installation procedures in Windows operating systems are often done with some “clickety click” magic that won’t take too long. Still the “clickety click” installations have lots of long-term problems: + +Installations can’t be reproduced + +Only the installer knows how he did it + +If you have multiple servers you are doing same manual steps multiple times + +Base of the installations does not differ much from project to project + +After few years when your windows server needs upgrade you will need to repeat this + +There is no way to test “clickety click” installations + +Windows Server Core installations are rare because people is not used to manage Windows Servers without GUI + +One could argue that documentation and clear processes would take care of all the problems above. Maybe they could but I have never seen installation documentation that has 100% coverage over how the installation has been done. Installation script works as document and developers are more likely to update it! + +What needs to be installed on a fresh Windows server + +Here is short version of my list what I would do for new Windows Server + +Install IIS + +Install WebPI + +Install newest.NET + +Install webdeploy + +Install various modules for IIS from windows feature list or with WebPI + +Install tools for administration (7zip, text editor etc) + +Create IIS site and application pool and change some defaults for better + +Is there PackageManagement for Windows? + +Oh yes there is! It is called Windows PackageManagement (aka OneGet). Unfortunately it is not available for Windows Servers yet. Production preview of Windows Management Framework 5 is already available. Windows PackageManagement is actually a manager that can access multiple type of repositories to have unified way to download and install software from multiple sources. Microsoft MSI installers, Windows Features and software could be all installed with Windows PackageManagement in the future. The thing still missing is grown up economy where all the toys would be easily available in a secure manner. Chocolatey is pretty good already but it is missing a lot of software and I’m still sceptical about +================================================================================ +Rank = 69; Score = 3768320.0 +<|begin_of_text|>On October 13, our teams will be performing an extended maintenance to make networking improvements on our servers. This maintenance will affect our Arc platform as well as all pages on our arcgames.com domain and games hosted from our servers that are launched from Arc and other platforms. The following games and services will be inaccessible during maintenance: + +Arc Games forums + +Arc Games main website, product pages and news pages + +Arc Games platform including billing and messaging + +Arc Games support websites + +Battle of the Immortals + +Blacklight Retribution (launched from Arc Games) + +Champions Online + +Forsaken World + +Fortuna + +Jade Dynasty + +Neverwinter (PC & Xbox One) + +PWI + +Star Trek Online + +Swordsman + +War of the Immortals + +Additional games launched from Arc Games + +We anticipate this maintenance to run from October 13 starting at 6:00 am PT and running to 6:00 pm PT. We will be providing updates via our Arc Twitter page and Facebook page as well as social channels for each of our titles. + +We appreciate your patience as we work to improve our platform. + +Arc Team<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 70; Score = 3768320.0 +<|begin_of_text|>Windows Azure: General Availability of Infrastructure as a Service (IaaS) Tuesday, April 16, 2013 + +This morning we announced the general availability release of our Infrastructure as a Service (IaaS) support for Windows Azure – including our new Virtual Machine and Virtual Network capabilities. This release is now live in production, backed by an enterprise SLA, supported by Microsoft Support, and is ready to use for production apps. If you don’t already have a Windows Azure account, you can sign-up for a free trial and start using it today. + +In addition to supporting all of the features and capabilities included during the preview, today’s IaaS release also includes some great new enhancements: + +New VM Image Templates (including SQL Server, BizTalk Server, and SharePoint images) + +VM Image Templates (including SQL Server, BizTalk Server, and SharePoint images) New VM Sizes (including Larger Memory Machines) + +VM Sizes (including Larger Memory Machines) New VM Prices (we’ve reduced prices 21%-33% for IaaS and PaaS VMs) + +Below are more details on today’s release and some of the new enhancements. You can also read Bill Hilf’s blog post to learn about some of the customers who are already using the IaaS capabilities in production. + +Windows Azure Virtual Machines + +Windows Azure Virtual Machines enable you to deploy and run durable VMs in the cloud. You can easily create these VMs from an Image Gallery of pre-populated templates built-into the Windows Azure Management Portal, or alternatively upload and run your own custom-built VHD images. Our built-in image gallery of VM templates includes both Windows Server images (including Windows Server 2012, Windows Server 2008 R2, SQL Server, BizTalk Server and SharePoint Server) as well as Linux images (including Ubuntu, CentOS, and SUSE Linux distributions). + +Windows Azure uses the same Hyper-V virtualization service built-into Windows Server 2012, which means that you can create and use a common set of VHDs across your on-premises and cloud environments. No conversion process is required as you move these VHDs into or out of Windows Azure – your VMs can be copied up and run as-is in the cloud, and the VMs you create in Windows Azure can also be downloaded and run as-is on your on-premise Windows 2012 Servers. This provides tremendous flexibility, and enables you to easily build hybrid solutions that span both cloud and on-premises environments. + +Easy to Get Started + +You can quickly create a new VM +================================================================================ +Rank = 71; Score = 3751936.0 +<|begin_of_text|>Press + +Web Search By The People, For The People: YaCy 1.0 + +on: 2011-11-28 + +The YaCy project is releasing version 1.0 of its peer-to-peer Free Software search engine. The software takes a radically new approach to search. YaCy does not use a central server. Instead, its search results come from a network of currently over 600 independent peers. In such a distributed network, no single entity decides what gets listed, or in which order results appear. + +The YaCy search engine runs on each user's own computer. Search terms are encrypted before they leave the user and the user's computer. Different from conventional search engines, YaCy is designed to protect users' privacy. A user's computer creates its individual search indexes and rankings, so that results better match what the user is looking for over time. YaCy also makes it easy to create a customised search portal with a few clicks. + +"Most of what we do on the Internet involves search. It's the vital link between us and the information we're looking for. For such an essential function, we cannot rely on a few large companies, and compromise our privacy in the process," says Michael Christen, YaCy's project leader. "YaCy's free search is the vital link between free users and free information. YaCy hands control over search back to us, the users." + +Each YaCy user is part of a large search network. YaCy is already in use on websites such as sciencenet.kit.edu, yacy.geocaching-portal.com, or fsfe.org, to provide a site-wide search function that respect users' privacy. It contains a peer-to-peer network protocol to exchange search indexes with other YaCy search engines. + +"We are moving away from the idea that services need to be centrally controlled. Instead, we are realising how important it is to be independent, and to create infrastructure that doesn't have a single point of failure," says Karsten Gerloff, President of the Free Software Foundation Europe. "In the future world of distributed, peer-to-peer systems, Free Software search engines like YaCy are a vital building block." + +Everyone can try out the search engine at http://search.yacy.net/. Users can become part of YaCy's network by installing the software on their own computers. YaCy is Free Software, so anyone can use, study, share and improve it. It is currently available for GNU/Linux, Windows and MacOS. The project is also +================================================================================ +Rank = 72; Score = 3751936.0 +<|begin_of_text|>Sharing Lua files in Renoise + +Recent articles showed how to use to Renoise Lua scripting to alter a track’s send device. One was driven by OSC, the other by MIDI. + +In both of them the code that did the actual send device manipulation was the same. But since these were two different scripts the same code appeared in both places. + +Sharing code using copy-and-paste is not the world’s worst programming sin, however much it might make some developers cringe. If you are throwing together a one-off script for a short-lived need, go for it. + +In many cases, though, having the same code in multiple places is a poor idea because it breaks one of the golden rules of software development: Be lazy. (Read up on the three great virtues of a programmer.) + +Copy-and-paste sure seems like the laziest thing to do at first, but if the code has any amount of complexity or importance there’s a really good chance that sooner or later you will want to change it. Maybe to fix a bug, maybe to make it faster, maybe to add a feature. + +With the same code in multiple places you need to keep track of all those places and edit the code in all those places. Oh, bother. + +Can Renoise Lua code be shared? Can you create a file that can be loaded and used by multiple scripts? + +Indeed you can. + +Know your path, change your path + +The preferred way to load files in Lua programs is the require function. + +require looks at a special path that defines where files may be found. This path is actually many paths, and it is not just a set of folder paths but paths that contain some pattern-matching variables. (See that previous link for details.) + +This path-of-paths can be altered. The usual warning about messing with program defaults applies here; if you screw up the require path your program may not run as expected (or at all) and you will be sad. + +The path is a string, so here’s an example of changing it so that a tool file can load from a parallel directory: + +package.path = "../ng-shared/?.lua;".. package.path + +For example, suppose you have this folder and file layout: + +Scripts/Tools/ng-shared/Utils.lua Scripts/Tools/com.neurogami.MidiMapper.xrns/main.lua + +If the file main.lua in com.neurogami.MidiMapper.xrns has that path alteration then when it does this: + +require "Utils" + +the program will look in the ng-shared/ folder since it +================================================================================ +Rank = 73; Score = 3735552.0 +<|begin_of_text|>At its main plant in Ingolstadt, German carmaker Audi has for the first time deployed a robot that works ‘hand-in-hand’ with humans – without a safety barrier and ideally adapted to the employees’ working cycles. + +It is the first human-robot cooperation at the Volkswagen Group to be applied in final assembly. This innovative technology makes work easier for the assembly employees and makes ergonomic improvements. + +For Dr. Hubert Waltl, Board of Management Member for Production at Audi AG, human‑robot cooperation opens up entirely new possibilities: “The factory of the future will feature increasing interaction between man and machine. That allows us to automate routine operations and to optimise ergonomically unfavorable workplaces.” But also in the future, there will be no factory without people. “People will continue to make the decisions on production processes. And our employees will continue to be essential for future-oriented, successful production.” + +Peter Mosch, chairman of the Group Works Council of Audi AG, commented: “We see the opportunities presented by the advancing interaction between man and machine. The decisive aspect for us is how this development is guided. We welcome it when it neither jeopardizes jobs nor leads to people losing independence to machines.” + +For the employees of the A4/A5/Q5 assembly lines at Audi’s Ingolstadt plant, the new, direct cooperation between humans and robots is an enormous help: Until now, they have had to bend over material boxes to take out the coolant expansion tanks. At first glance, this seems like a simple task, but with frequent repetitions it can lead to back problems. From now on, the task will be taken over by a robot, known internally as ‘PART4you’. + +It works hand-in-hand with the Audi employees and is fitted with a camera and an integrated suction cup. This enables it to pick up the components from the boxes and to pass them to the assembly workers – without any safety barrier, at the right time and in an ergonomically optimal position. “In a production process with increasing diversity of model versions, PART4you provides the employees with important assistance. It selects the correct component and holds it ready to be taken. This means that the employees no longer have to reach over long distances or bend down repeatedly. The robot becomes an assembly assistant operating at the same speed as the assembly worker – and not the other way around,” said Johann Hegel, Head of Assembly Technology Development. + +“Thanks to a soft protective skin with integrated safety sensors, there is no danger to the employees,” explained Hegel. Because PART +================================================================================ +Rank = 74; Score = 3735552.0 +<|begin_of_text|>Good aim is essential to having fun and improving your results in CoD. + +I have seen this question asked numerous times “How do I get a better aim in Call of Duty” and today we are going to go over exactly that. This will be very in-depth and we will go through every aspect in order for you to be the best you can be. However, there’s a common misconception that things will change overnight although that is not the case, these things take time and if you’re willing to put in the time then great things will happen to your gameplay and you will notice drastic improvements. + +Vibration – Good Or Bad? + +By default on Call of Duty, your controller is set to vibrate when you shoot and get shot. This adds a whole element to the game as it allows you to virtually feel everything your character does. However, because you get so used to it you are unable to play without it. + +If you are using vibration, I would recommend turning it off immediately because the rumble from the vibrations will throw your shot off and make it harder to control your aim. It is only a minor change but it can make all the difference to your game and elevate it to that next level. + +Addons – Kontrol Freeks + +If you haven’t heard of Kontrol Freeks already, they are an extension that sits on top of your analogue sticks. You may now be asking yourself “Why would I want this?” Well, they are much taller than regular analogue sticks, it means you don’t need to move your thumbs as much whilst aiming. This will allow you to make much more precise movements, thus making it easier to play on higher sensitivities, although not many professional players do play at a higher sensitivity, however, we will get onto that topic shortly. + +They are used by a lot of professional Call of Duty players.They are one of the most notorious companies when it comes to the gaming accessory industry and has a great reputation. Not only is the quality of the product exceptional, but also because they are considerably cheap. A pair of Kontrol Freeks will only cost you a mere $10-20 so they are definitely worth trying out. The best part is, if you don’t like them, they have a 30-day money back guarantee, so there is no risk to be had. + +Sensitivity Levels – Very Controversial + +There is no such thing as a ‘best sensitivity level’ at all. Some players play on 3, others play on 15, it’s all about what makes you +================================================================================ +Rank = 75; Score = 3719168.0 +<|begin_of_text|>Motivation + +Last week we were discussing, among many other things, ways to speed up Firefox during startup. One obvious option was to move more of our I/O off of the main thread. This in turn involves making more code asynchronous, and asynchronous code is simply harder to manage. Mike Shaver mentioned something about "futures" as a possible way to handle the extra complexity, and then the discussion moved on to something else. + +What exactly is a future anyway? + +I'm not exactly an expert, but I've not just used futures, I've written my own implementations in JavaScript and even Object Pascal (in hindsight I'm not sure the latter was a good idea, but it was certainly an interesting exercise). Futures seem esoteric, but they really shouldn't be -- the idea is really quite simple. In this post I'll try to explain what futures are and how they can be used to make asynchronous programming easier. + +In the simplest form, a future works like an IOU. I can't give you the money you've asked for right now, but I can give you this IOU. At some point in the future, you can give me the IOU and I'll give you the money -- if I have it. If I don't have it yet, then you can wait until I do. I get paid on Friday. + +Alternatively there's the dry cleaner metaphor. You drop your clothes off on Monday and the clerk gives you a ticket that you can use later to reclaim your clothes after they've been cleaned. The clothes will be ready on Tuesday morning, but if you show up too early, you'll have to wait. On the other hand, if there's no hurry, you can just do other stuff on Tuesday and show up on Wednesday with a reasonable expectation that they'll be ready when you arrive. You'll just hand your ticket over, collect your clothes, and be on your way. + +A future is similar to the IOU (or the dry cleaning ticket). It gives you a way to represent the result of a computation that has not yet completed, and it allows you to access that result once it becomes available. So you can call a function which starts some asynchronous process but doesn't wait for it to finish. Nevertheless the function can return you a useful result: a future which can be used to claim the real result later. + +A simple example + +Of course if you ask for the result too soon, you'll have to wait. On the other hand, if the result becomes available before you want it, then +================================================================================ +Rank = 76; Score = 3702784.0 +<|begin_of_text|>Sintel The Game is based on Blender Foundation's extremely popular movie 'Sintel'. The game will have action oriented gameplay while keeping same level of intensity and emotions that touched so many people. The story is as follows: After a long time, the developers of Sintel The Game posted about the game progress in a recent blog post.Sintel The Game is based on Blender Foundation's extremely popular movie 'Sintel'. The game will have action oriented gameplay while keeping same level of intensity and emotions that touched so many people. The story is as follows: + +Sintel – The Game takes place within the events of the movie. Sintel finds herself just outside the town of Garway. She bumps into some bandits who attack her and she is taken into the care of a kind stranger. You play as Sintel and follow her on her journey through Garway. On the way you will discover that the guards of Garway are corrupt, and help to rise against them. + +blog post: Excerpts from the + +I worked on enemy AI a little bit, started re-designing the dialog system, and a bunch of other small things. Since I’ve gotten back, I started re-programming the player mechanics. I know this sounds silly seeing as we already have a working, controllable player, but there was room for improvement in the design and in the code. We also gave Sintel a nice new IK rig which needed to be implemented. As of right now, I’ve improved the movement system, making it easier to walk up and down hills, giving it a more realistic and less glitchy feel. It already feels a lot smoother than before. The code is optimized this time around too. Now i’m working on making the player camera better. + +We still want to release a demo as soon as we can. We are slowly making our way towards a release. I think I speak for all of us when I say we don’t want to release a demo unless it is fun. The pre-demo was short and bland because we rushed it out without spending enough time on making it fun. The demo will include some different gameplay elements, a good taste of combat, and a boss battle that is actually challenging. + +I can’t say when the demo will be ready, but the moment we pinpoint a date, I will let you guys know. + +here. You will need Blender 2.5+ for the game to work. The game can be played by running sintel_the_game.blend file. To start the game press 'P' and +================================================================================ +Rank = 77; Score = 3686400.0 +<|begin_of_text|>Current trends in software and backend architecture have been evolving towards a more loosely coupled more granular design. I am sure most of you have heard of microservice based architectures. read more + +The latest development on that front in the past couple of years has been the advent of serverless which allows you to run applications in very cost effective ephemeral services. This is why it is important to have a proper gateway for your API that is able to route all your requests to the designated endpoint. + +GraphQL stands out in that respect as being a mature open sourced standard started at Facebook. We will first have a look at how we set up our own GraphQL server locally, then we will explore the Query language and schema definitions it provides which allows you essentially query your mesh of services from a single point of entry. The beauty of that is it will notify you early if any of your endpoints is misbehaving or the schemas are out of date by erroring out. Another advantage of this is that it allows for your API documentation to be a real time process and it will give you what one may call an API playground where you can query and explore your API. + +After we explore our serverless API we will have a look at the more advanced features and standards around mutators and resolvers and then we will close by going all in, full serverless and deploy our graphql server to a function in the cloud.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 78; Score = 3686400.0 +<|begin_of_text|>How to Build a Monitoring and Analytics System for Your IoT Embedded Devices + +Jonathan Seroussi Blocked Unblock Follow Following Sep 3, 2017 + +This post outlines a cookbook we used to create a monitoring system for a connected embedded system. We used it to create one that can monitor all type of devices (even those that run Bare-Metal and RTOS), and we thought it’d make sense to help others build similar systems. This cookbook could easily be used to monitor device operational data as well as your actual application data. After deploying an IoT system, you need to make sure everything is working as it should and close a feedback loop from the field. We’re about to show you how to build one using open source tools almost for free. We also have a post that outlines the need for such a system (Check It Out). + +Let’s focus on three things: + +Device and gateway code that streams data to a centralized backend. Backend and database to process and store the data. Frontend to display the data. + +Once these three building blocks are in place, acting upon the monitored data and connecting it to your operation processes is relatively easy easy. + +1. Device and Gateway Logging tool + +When thinking about a logging utility hosted on your end devices and gateways, the things that come up are: + +It needs to have a small footprint. + +Never crash my device. + +Consume little power. + +Not disruptive to the application code. + +Your devices means your code. It’s challenging to generalize this topic as a lot depends on the communication channel used to connect to the internet. We open sourced a logging tool that logs events from your end devices and sends it to your gateway/backend. If you are using nRF52 with BLE GATT and a gateway we also provide you with a sample that works out of the box. Go to our GitHub to get started. + +1. jumper-ulogger — is an under 500B memory footprint logging framework you integrate into your end device (with examples for porting to nRF52 device and CC3200). + +2. jumper-logging-agent — is the service that runs on the gateway. + +3. jumper-ble-logger — is a logging agent service example for a BLE GW. + +If you have any issues or question send us a note to info@jumper.io. + +2. Backend and Database + +This is where you should pay attention more carefully because there’s some good stuff here :-) I recommend using the following toolsets to make this work: + +A key-value database to store a current snapshot of a device’s state +================================================================================ +Rank = 79; Score = 3637248.0 +<|begin_of_text|>HTML5 Video at Netflix + +Netflix Technology Blog Blocked Unblock Follow Following Apr 14, 2013 + +by Anthony Park and Mark Watson + +Today, we’re excited to talk about proposed extensions to HTML5 video that enable playback of premium video content on the web. + +We currently use Microsoft Silverlight to deliver streaming video to web browsers on the PC and Mac. It provides a high-quality streaming experience and lets us easily experiment with improvements to our adaptive streaming algorithms. But since Microsoft announced the end of life of Silverlight 5 in 2021, we need to find a replacement some time within the next 8 years. We’d like to share some progress we’ve made towards our goal of moving to HTML5 video. + +Silverlight and Browser Plugins + +Silverlight is a browser plugin which allows our customers to simply click “Play” on the Netflix website and watch their favorite movies or TV shows, but browser plugins have a few disadvantages. First, customers need to install the browser plugin on their computer prior to streaming video. For some customers, Netflix might be the only service they use which requires the Silverlight browser plugin. Second, some view browser plugins as a security and privacy risk and choose not to install them or use tools to disable them. Third, not all browsers support plugins (eg: Safari on iOS, Internet Explorer in Metro mode on Windows 8), so the ability to use them across a wide range of devices and browsers is becoming increasingly limited. We’re interested to solve these problems as we move to our next generation of video playback on the web. + +HTML5 Premium Video Extensions + +Over the last year, we’ve been collaborating with other industry leaders on three W3C initiatives which are positioned to solve this problem of playing premium video content directly in the browser without the need for browser plugins such as Silverlight. We call these, collectively, the “HTML5 Premium Video Extensions”: + +Media Source Extensions (MSE) + +The W3C Media Source Extensions specification “extends HTMLMediaElement to allow JavaScript to generate media streams for playback.” This makes it possible for Netflix to download audio and video content from our content delivery networks and feed it into the video tag for playback. Since we can control how to download the audio/video content in our JavaScript code, we can choose the best HTTP server to use for content delivery based on real-time information, and we can implement critical behavior like failing over to alternate servers in the event of an interruption in content delivery. In addition, this allows us to implement our industry-leading adaptive streaming algorithms (real-time +================================================================================ +Rank = 80; Score = 3538944.0 +<|begin_of_text|>Word in China is out about blockchain technology, as the government made clear in an Informatization Strategy published in December of 2016. The strategy states, "The internet, cloud computing, large data, artificial intelligence, machine learning, blockchain … will drive the evolution of everything - digital, network and intelligent services will be everywhere." + +It was an official endorsement for the new digital age and a big boost for blockchain technology. + +In a country with $5.5 trillion in digital payments last year (50 times the U.S.), blockchain is now a buzzword among the titans of industry. And in the race to participate, Chinese banks, builders, suppliers and retailers are pumping out blockchain solutions. + +Survival of the Fittest + +Even with millions of dollars, many of these new blockchains may not make it very far. In a recent WeChat post, Antshares executive Erik Zhang stated that 90 percent of the enterprise blockchains we are seeing are doomed to fail. Without an open-source code that anyone can build upon, Zhang argues, private solutions will not see the network effect of a healthy ledger. + +The big question for China, however, is whether a country known for its centralized authority, and a penchant for all things made-in-China, will allow an open-source, global standard solution sit on its internet. With Bitcoin, Ethereum and Hyperledger vying for dominance, there are big names within China making moves into the space. + +A few of these big names include: + +The People's Bank of China (PBOC) - The PBOC is reportedly close to the release of a government-backed digital RMB currency, which would put China at the frontier of digital currency adoption. And there are whispers within China that Shenzhen will be ground zero for the new digital economy. In September, Bloomberg reported that PBOC Vice Governor Fan Yifei wrote: "[T]he conditions are ripe for digital currencies, which can reduce operating costs, increase efficiency and enable a wide range of new applications." This would pave the way for blockchain startups in China to move forward in digital banking, finance, record-keeping, supply chains, IoT, AI and more. + +Wanxiang Blockchain Labs - Working with Ethereum, Wanxiang is the largest blockchain development backer in China. After purchasing 500,000 ETH tokens last year, they pledged $30 billion for the development of a smart city in Hangzhou. They offer open-source platforms for anyone to build upon, and launched an accelerator fund for developers, intending to put money into +================================================================================ +Rank = 81; Score = 3522560.0 +<|begin_of_text|>Javascript React Get Started with redux-form + +The redux-form library bills itself as the best way to manage your form state in Redux. It provides a higher-order form component and a collection of container components for dealing with forms in a React and Redux powered application. Most importantly, it makes it easy to get a form up and running with state management and validations baked in. + +To get a feel for what redux-form can do for us, let's build a simple Sign In form. You can follow along below or check out the resulting source code directly. + +Start with React + +We are going to start by getting the React portion of this application setup. The easiest way to spin up a React app is with create-react-app. Install it with npm or yarn if you don't already have it globally available. + +$ yarn global add create-react-app + +Let's generate our project. + +$ create-react-app redux-form-sign-in + +The create-react-app binary that is now available on our machine can be used to bootstrap a React app with all kinds of goodies -- live code reloading in development and production bundle building -- ready to go. + +Let's jump into our project and kick off a live reloading development server. + +$ cd redux-form-sign-in $ yarn start + +At this point you should see your browser pointed to localhost:3000 with a page reading Welcome to React. + +Is This Thing Plugged In? + +We can see the live-reload in action by altering src/App.js. + +return ( < div className = "App" > < div className = "App-header" > < img src = { logo } className = "App-logo" alt = "logo" /> < h2 > Redux Form Sign In App < /h2 > < /div > < p className = "App-intro" > To get started, edit < code > src / App. js < /code> and save to reload. < /p > < /div > ); + +Change the text in the h2, save the file, and then switch back to the browser to see the changes almost instantly. + +create-react-app made it really easy to get to a point where we can just iterate on our app. We didn't have to fiddle with Webpack configurations or any other project setup. + +Next, let's change the prompt in the app intro. + +< p className = "App-intro" > Sign in here if you already have an account < /p > + +Our app is now begging for a form which brings us to redux-form, the focus of this post. + +Satisfying +================================================================================ +Rank = 82; Score = 3506176.0 +<|begin_of_text|>The first final version of Hibernate OGM is out and we’ve recovered a bit from the release frenzy. So we thought it’d be a good idea to begin the new year with a tutorial-style blog series, which shows how to get started with Hibernate OGM and what it can do for you. Just in case you missed the news, Hibernate OGM is the newest project under the Hibernate umbrella and allows you to persist entity models in different NoSQL stores via the well-known JPA. + +We’ll cover these topics in the following weeks: + +Persisting your first entities (this instalment) + +Querying for your data + +Running on WildFly + +Running with CDI on Java SE + +Store data into two different stores in the same application + +If you’d like us to discuss any other topics, please let us know. Just add a comment below or tweet your suggestions to us. + +In this first part of the series we are going to set up a Java project with the required dependencies, create some simple entities and write/read them to and from the store. We’ll start with the Neo4j graph database and then we’ll switch to the MongoDB document store with only a small configuration change. + +Let’s first create a new Java project with the required dependencies. We’re going to use Maven as a build tool in the following, but of course Gradle or others would work equally well. + +Add this to the dependencyManagement block of your pom.xml: + +... ... org.hibernate.ogm hibernate-ogm-bom pom 4.1.1.Final import ... ... + +This will make sure that you are using matching versions of the Hibernate OGM modules and their dependencies. Then add the following to the dependencies block: + +... ... org.hibernate.ogm hibernate-ogm-neo4j org.jboss.jbossts jbossjta ... ... + +The dependencies are: + +The Hibernate OGM module for working with an embedded Neo4j database; This will pull in all other required modules such as Hibernate OGM core and the Neo4j driver. When using +================================================================================ +Rank = 83; Score = 3506176.0 +<|begin_of_text|>CUSTOMIZER CHALLENGE CONTEST WINNER - Artistic Category This is a 100% printed customizable music box! Only 3D printed parts are used in the design and it can be assembled and disassembled via printed snapping mechanisms. The project originated when a friend of mine said that he'd only be interested in 3D printing once he can print a music box ;) Videos http://youtu.be/LUlovenI9xQ http://youtu.be/K_c3p24RRtQ (made by banthafodder7400) http://youtu.be/exNeQDz7f3g I'll try to keep the.scad file on this page updated but to help me to manage the design and to make it easier for others to contribute: https://github.com/wizard23/ParametrizedMusicBox + +Updates + +2018-03-28 Updated URL to generate parameters (was offline) + +2016-09-20 Spreadsheet to help you to get the parameters right: https://goo.gl/sKWWDk (created by: Ben Horner) + +2013-11-24 Updated link to generator page (old webserver is offline) + +2013-03-10 (V3) Added optional Name of Song on top/bottom of MusicCylinder; fixed build plate positioning of pulley that messed up smaller customized versions; cleanup and clarification of descriptions; implemented work around for customizer hickup when strings start with a '.' + +2013-03-08 (V2) removed "work in progress status", fully test printed + +It's very important to put the music box on a sounding box to get good sound quality. I found that large cardboard boxes and some tables make good sounding boxes. A guitar or a piano should work even better! + +A complete music box consists of 6 parts: + +Case: the large thin walled part that holds the vibrating teeth and holds everything together + +Music Cylinder: the large cylinder with the pins (that encode the music) sticking out + +Transmission gear: sits between the crank gear and the music cylinder + +Crank gear: drives everything, connects to crank (insert it through the round hole in the case) + +Crank: for manually driving the box, connects to crank gear and crank pulley + +Pulley: for holding the crank while turning it + +With the default parameters you get a complete building plate that can play one full octave range (13 half notes from C to C) in a medium footprint that should fit in most printers. You can +================================================================================ +Rank = 84; Score = 3506176.0 +<|begin_of_text|>Microsoft is integrating Skype directly into Windows 10, and the result looks a lot like Apple's iMessage service. While the company unveiled some of its Skype integration at a special Windows 10 press event in Redmond yesterday, the software maker didn’t show its new Messaging app on the PC version of Windows 10. This appears to be key to a new experience for Skype messaging in Windows 10, and it brings back the built-in Messaging app from Windows 8 that Microsoft killed with the Windows 8.1 update. + +Skype is starting to link usernames to mobile numbers + +The new Messaging app works by integrating Skype, allowing you to chat to Skype contacts or initiate video / audio calls. All the conversations are synced between PCs, tablets, and phones, and the new app looks like a lightweight version of Skype. It’s also identical to the Messages app on OS X, with the same two-panel interface and circular UI for contact photos. Microsoft has started linking Skype usernames with mobile numbers to make it easier to find friends who are using the service without having to know their user ID. That makes this whole approach a lot more like iMessage, allowing Skype users to chat to friends easily on the service. The main difference is that Skype is cross-platform so you can chat to friends on iOS, Android, BlackBerry, Windows, and more, while iMessage is limited to Apple’s platforms. + +The built-in Skype experience on the phone version of Windows 10 also allows you to send text messages, but it’s entirely possible (and likely) that Microsoft will extend this functionality and sync state to the PC version just like iMessage. In Microsoft’s new world Windows 10 apps are the same across PCs, phones, and tablets, so such a move would be expected. Microsoft isn’t fully detailing its Messaging app plans just yet, but it’s encouraging to see the company move to a more native and simple integration of Skype instead of separate and unnecessary apps. All that's needed now is the complexity of usernames to fully disappear so everyone can use Skype just with their mobile number.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 85; Score = 3489792.0 +<|begin_of_text|>With the U.K. report on Facebook, and the stern language within it, the train on regulating data sharing may finally reach the station this year. The FTC is also likely to impose a stiff fine on Facebook for violating a consent decree. + +So let's learn more about this data sharing business. If you prefer a video, the gist of this post can be heard here. + +*** + +First, let's talk about data flows and the "cloud". Data are stored in computers that are called servers. In the cloud computing model, these servers are owned - not by the companies that collect the data - but by large tech companies like Amazon, Google, Microsoft, etc. who are responsible for managing the servers. These servers are geographically dispersed and so when data enter the cloud, they get replicated and spread to many servers. The technical benefit of such replication is recoverability of the data (allowing the use of cheaper, less reliable computers) but now, the data become much harder to delete. + +Data become more telling if one combines different datasets measuring different aspects of our lives. For example, an auto insurer may have data on past claims and that data help predict your future claims. But if the auto insurer is able to get data from say an automaker about your car, e.g. how fast you drive, where you drive, etc., that data combined with past claims improve the predictive power. + +Thus, a data-sharing industry has been created. Companies make agreements to share data with one another. This becomes much easier in the "cloud" as those servers are already connected to one another. These agreements may include explicit payments but even if they don't, both sides must be benefiting commercially from the arrangement, or else they would not exist. + +So when company A shares data with company B, the data flow from A servers to B servers. B may also use a cloud, which then means the data would be replicated yet again, and dispersed geographically onto yet another set of servers. + +And company B may also share data with company C, etc., etc. + +*** + +An inexplicable part of the consent decree between Facebook and the FTC is the requirement that Facebook monitor what happened to the data after they are shared with third parties. I just can't figure out how that is possible. It isn't even possible within Facebook: if a user demands that his/her be deleted, it will be very hard to ensure that all copies of the data are deleted from every server, including data that might have landed in an analyst's computer. In fact, most analysts +================================================================================ +Rank = 86; Score = 3489792.0 +<|begin_of_text|>Google is launching a new enterprise service for large businesses that want to adopt Chrome OS devices. The new Chrome Enterprise subscription, which will cost $50 per device and year, is essentially a rebrand of Chrome Device Management, but with a number of additional capabilities. Even though the name would make you think this is about the Chrome browser, this program is actually all about Chrome OS. For Chrome users in the enterprise, Google already offers the Chrome Enterprise Bundle for IT, after all. + +For enterprises, the main highlight here is that Chrome Enterprise is fully compatible with their existing on-premise Microsoft Active Directory infrastructure. Google senior director of product management for Android and Chrome for Business and Education Rajen Sheth told me that this has long been a stumbling block for many enterprises that were looking at adopting Chrome OS devices. With this update, enterprise users will be able to use their existing credentials to log into their Chrome OS devices and access their Google Cloud services — and IT admins will be able to manage their access to these devices and services. + +It’s worth noting that Chrome OS admins could already enable other services that use the SAML standard to enable single sign-on for Chrome devices. + +In addition, businesses also will now be able to manage their Chrome OS devices from the same enterprise mobility management solutions they already use, starting with VMware’s AirWatch. Support for similar services will launch in the future. + +With this new licence, IT admins also will be able to set up a managed enterprise app store for their users. This feature is currently in beta and focuses on Chrome OS’s ability to run Android apps, which is currently available on many of the most popular Chrome devices in the enterprise. + +Other benefits of the Chrome Enterprise subscription include 24/7 enterprise support, managed OS updates and printer management (you may laugh about this last one, but that’s something that still matters in many offices). + +It’s no secret that Google is working hard to get more enterprises to adopt its various cloud-based services. Chromebooks have already found a lucrative niche in verticals like retail and education. To expand its market share, though, features like this new integration with AirWatch were sorely needed.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 87; Score = 3473408.0 +<|begin_of_text|>This effectively means that anyplace you can type text becomes a potential storefront. + +Before customers have even closed your book, they can already have ordered the next book in the series. + +I've been wondering lately about whether bitcoin could make Amazon unnecessary. Here are my thoughts on the subject.If you've never heard of Bitcoin, it's basically the e-mail of money. You get yourself a bitcoin wallet program, which can either be an online program hosted by a company (like the bitcoin equivalent of gmail) or you download a wallet program to your home computer (like Outlook Express). Whichever method you choose, you get to create a "bitcoin address" for yourself which works like an e-mail address, except that you receive money instead of messages. A bitcoin address looks like this: 19v5V5P9infaQyCaRxWWeWoHnRbkX684PS. (This is for a Pet Rescue site.) So if you wanted to send money to this bitcoin address, you would basically just pop open your money-email program and hit send. The money appears in the recipient's "money inbox" in a few minutes to a few hours, depending on the sum you send. As for transaction fees, they are so minimal as to be nearly nonexistent--we are talking on the order of pennies. This makes it practically to send $0.99 to someone and not have the money eaten up in transaction fees.The really neat thing about bitcoin addresses, though, is that they are just a line of text. You could post one anywhere--in your forum signature, at the end of your book, in your mailing list, your tweet, or on your author webpage. Anywhere you can put an e-mail address, you can put a bitcoin address.Now let's ask ourselves: why do we sell our books on Amazon, and not on Goodreads or Twitter or Wattpad or our website any other site that offers free exposure for your book? Well, duh. Because Amazon has a big, fat "Buy" button and those other sites don't. But with bitcoin these other sites become storefronts--anywhere you can type your e-mail, you can type your money e-mail. This means that readers can buy direct from Goodreads or Wattpad--no need to go to Amazon at all. What's more, bitcoin enables "one click" purchases everywhere. You don't have to go through the hassle of registering up for an account or worrying about giving your credit card information to some stranger-- +================================================================================ +Rank = 88; Score = 3473408.0 +<|begin_of_text|>The title is specifically about CloudFront and Laravel, but for the most part, this will apply to most web applications behind a reverse-proxy of some sort, be it a CDN, load balancer, or some other proxy type. + +What we'll cover here: + +What changes when using a reverse-proxy on your application What your application needs to do to get around these issues How to use the TrustedProxy Laravel package to take care of the details for you in a Laravel Application + +Applications behind a proxy + +For the launch of https://course.shippingdocker.com, I put CloudFront in front of a Laravel application running inside of a Docker container. + +I wanted CloudFront for a few reasons: + +Primarily, I wanted to make use of AWS's free, auto-renewed SSL certificate. This frees me from paying for an SSL certificate and/or managing something like LetsEncrypt. Secondly, I was planning on using a t2.nano instance on AWS and wanted to save it some CPU cycles from serving static assets. These instances are tiny, and start eating away at its small number of given CPU credits when it hits just 2% of CPU usage. (Having the site be as speedy as possible for those overseas was also part of this decision). + +So, imagine this setup - HTTPS connections are made to course.shippingdocker.com. CloudFront receives this HTTPS request, and then forwards it to my site runnning on an EC2 instance (the "origin server"). The EC2 instance is listening on port 80 for HTTP connections. + +Since CloudFront won't forward to an IP address, we need to give it a hostname to use. We can use the EC2 instance's public DNS name - I used http://ec2-34-197-131-119.compute-1.amazonaws.com in the case of my server (try it out, that URL will work). + +What your application sees + +So, CloudFront is receiving an HTTPS request and decrypting it (terminating the SSL connection). It then sends the decrypted HTTP request to my server (the origin server) using the network address ec2-34-197-131-119.compute-1.amazonaws.com. + +Laravel only sees that 2nd request, and thus assumes: + +It's receiving requests over http:// instead of https://, and thus will generate URI's with the the http:// scheme, including form submit URLs and redirect locations The domain name used and thus seen by the application is ec2-34-197-131-119.compute-1.amazonaws.com. Laravel will generate +================================================================================ +Rank = 89; Score = 3473408.0 +<|begin_of_text|>Congressional Democrats are proposing to ban plastic guns, following reports of major security lapses at the nation’s airports + +Plastic guns can be even more dangerous than traditional firearms because they're harder to detect, says Rep. Steve Israel (D-N.Y.). + +ADVERTISEMENT + +The Undetectable Firearms Modernization Act, backed by Israel and several other Democrats, would prohibit the manufacture of entirely plastic guns. The legislation would require a major component of every gun to contain enough traces of metal to be detected.Israel plans to unveil the legislation Tuesday during a press conference at LaGuardia Airport in New York City, where he will draw a connection between his bill and recent high-profile airport security lapses. + +"If detectable weapons can make it through security checkpoints, how can we expect to catch wrongdoers carrying undetectable plastic firearms?" Israel told The Hill. “What could be worse than a gun that can be used on an airplane, but cannot be detected on the security line because it’s plastic?” + +"It’s time to modernize our airport security so the American people can count on it," he added. + +The Transportation Security Administration (TSA) failed a recent sting operation in which undercover agents sneaked fake explosives and weapons through airport security in 67 out of 70 tests, or about 95 percent of the time. + +Many of America’s busiest airports failed the tests. + +In one test, a federal agent was able to sneak a fake bomb that was strapped to his back through security, even after he was subjected to a pat-down. + +The massive security failure led to the ouster last week of TSA acting administrator Melvin Carraway. + +In light of the TSA security lapses, Israel and other Democrats say it is even more important to ban plastic guns. + +“The fact is, we should be making it easier, not harder for TSA to stop dangerous weapons from getting onto our planes,” Israel said. + +“What angers me, and frankly what frightens me, is that the guns that were getting passed the TSA agents were highly detectable,” he added. “But a plastic gun couldn’t be picked up by the most astute and trained TSA agent.” + +Plastic guns have been made popular by 3D printers that make it easier for consumers to build their own firearms. + +While the legislation would still allow manufacturers to build partially plastic guns, it would close a loophole that allows people to build guns that can evade security. + +Current law prohibits plastic guns, but gun owners can get around the requirements by including a detachable strip of metal on an otherwise plastic gun +================================================================================ +Rank = 90; Score = 3440640.0 +<|begin_of_text|>Microsoft just announced a service that will make it easier to manage storage for SharePoint Online within the Office 365 suite. + +The move effectively tackles one of the big problems users were having with storage allocation to SharePoint site collections. It also makes the management of entire SharePoint environments easier. + +The changes announced last night makes the storage limits associated with each group of sites more flexible. It means storage that is not used with one site collection can be moved to another to facilitate the development of that collection. + +Freeing-Up Storage + +Site collections are groups of related SharePoint websites that, until now, were assigned a set amount of storage space that could not be used for anything else even if the site collection did not use all that space. + +Mark Kashman, senior product manager on the SharePoint marketing team, with Microsoft said in a blog post that from here on in, if an administrator sets a storage quota of 100 GB for a site collection and only uses 20 GB, under the old system the 80 GB remaining was wasted. + +Under the new system, the 80 GB is returned to the storage pool meaning that site collections no longer "reserve" storage, and that they only use the storage they actually need. + +As Needed + +The process of assigning storage can also be automated with a tab that enables either manual, or automatic, of storage. The automatic setting increases storage requirements for site collections as needed, drawing on what's available from the pooled storage for the entire Office 365 domain. + +The final change that Microsoft has introduced is that the number of site collections SharePoint Online supports per tenant has been increased from 10,000 to 500,000 with the total storage now allowed starting at 10 GB+ 500 MB for every user. This is separate from the default per-uses OneDrive for Business storage space. + +The new storage limits will be available for most of the Office 365 plans including Office 365 Enterprise E1, E3, and E4; Office 365 Education A2, A3, and A4; Office 365 Government G1, G3, and G4. Office 365 Midsize Business will have this new storage model. + +Guessing Game + +The changes will be a welcome bonus to SharePoint administrators and site developers that have, until now, had to play a guessing game around the resources they will have to apply to site development. + +Matter Center for Office 365 + +The announcement dovetails with the announcement that Microsoft has developed a document management add-on for Office 365 for the legal profession. The new app, called +================================================================================ +Rank = 91; Score = 3440640.0 +<|begin_of_text|>CoreOS founder Alex Polvi Twitter/@polvi + +Google Ventures is leading a $12 million investment in CoreOS, a tiny startup that's changing the way modern web applications are built and maintained. + +The startup will also use a recently released Google technology called Kubernetes in one of its own products, which is aimed at helping companies run their data centers more efficiently. + +CoreOS makes a super-lightweight version of the free Linux operating system, which reduces the amount of hardware you need to run applications in big data centers. That's made it popular among software developers looking to do more with less. + +For a while there, CoreOS was tight with Docker, one of the hottest startups in Silicon Valley, and it's no surprise why — they both had the mission of taking apps and putting them into what we call "containers." + +Think of containers as a metaphor. Shipping yards put goods into a bunch of shipping containers all the same shape and size because it makes them easy to stack onto boats. + +In the same way, containers let you take an application (an email server, a blogging platform, whatever) and whatever else it needs to run and throw it into a standard virtual "box." You can put this box on your own servers in your own data center, or upload it to Amazon's or Microsoft's or Google's cloud service and it'll run the exact same way, without the extra effort of having to struggle with it to make it work. It's all self-contained. (Hence, "containers.") + +CoreOS made a version of Linux that went into these boxes, and Docker made the boxes themselves. + +Docker Founder and CTO Solomon Hykes Flickr/ 97226415@N08 + +But Docker and CoreOS had a falling out late last year. A container on its own doesn't really do much, it needs to be overseen and managed. Docker wanted to sell this management technology as well. + +CoreOS had an issue with what it saw as Docker getting too big for its britches, and made its displeasure clear with the public launch of Rocket, a competitor to Docker. Suddenly, developers had two options for building those self-contained apps-in-boxes. + +Meanwhile, Google has quietly been using containers in its own huge data centers for many years now. As Docker made containers popular to the rest of the world, Google released some of its own technology for managing them, dubbed Kubernetes, for free. Kubernetes seen some uptake in both the CoreOS Rocket and Docker communities. + +Today, in addition to the funding, CoreOS is announcing +================================================================================ +Rank = 92; Score = 3440640.0 +<|begin_of_text|>It’s one thing to know about SQL Injection, File Access Attacks, XSS, and other security hazards. And because you’re a great developer you’re regularly squashing vulnerabilities in your app. And yet every Ruby/Rails developer is relying on someone else’s code to do their work. And guess what? No matter how careful you are, no matter how much time you spend perfecting your code, someone else’s code is going to have a security bug (yours will too if you & I are being honest with each other!) + +One of the questions that a few people have emailed me with now is: “How do I stay up to date on Ruby/Rails security?” It’s a great question! First because not enough developers care about security. Second because there are a lot of great tools out there to help protect your app. Let’s look at how you can stay up to date. + +By the way feel free to email me@gavinmiller.io if you have questions too! + +Follow Relevant Mailing Lists + +The first step in keeping your app up to date and protected is to keep up with the news. The two main sources of security news are the Ruby Security Mailing List and the Rails Security Mailing List. Both lists focus on security and will give you the best warning that an attack/fix is coming down the pipe. + +Follow CVE Reports + +Now the Ruby, and Rails mailing lists are great, but you have A LOT more dependencies in your app than that: nokogiri, rack, thin, puma, etc.. Unless there was a major issue in these gems they’re not going to make the Rails or Ruby mailing list, so you need to get that information from elsewhere! + +One of the little know resources for keeping up with security vulnerabilities are CVE databases. There are a few different sites that offer this type of information and CVE Details is my favorite because it’s easy to consume the information. + +Ruby and Rails both have dedicated pages, and you can create an RSS feed of those pages. And for the major gems in your site navigate to their pages and create an RSS feed for them as well! + +Keep Code Updated + +A simple way to keep your application up to date with the latest vulnerabilities is to not let your dependencies become outdated. To do this run bundle outdated on your codebase and update the gems that are out of date. + +This is usually easier said than done because updating dependencies can cause your application to break in unexpected ways. The mitigation for that is keeping your tests up to date. If you can update +================================================================================ +Rank = 93; Score = 3424256.0 +<|begin_of_text|>AWS to Azure services comparison + +In this article + +This article helps you understand how Microsoft Azure services compare to Amazon Web Services (AWS). Whether you are planning a multicloud solution with Azure and AWS, or migrating to Azure, you can compare the IT capabilities of Azure and AWS services in all categories. + +In the following tables, there are multiple Azure services listed for some AWS services. The Azure services are similar to one another, but depth and breadth of capabilities vary. + +Azure and AWS for multicloud solutions + +As the leading public cloud platforms, Azure and AWS each offer businesses a broad and deep set of capabilities with global coverage. Yet many organizations choose to use both platforms together for greater choice and flexibility, as well as to spread their risk and dependencies with a multicloud approach. Consulting companies and software vendors might also build on and use both Azure and AWS, as these platforms represent most of the cloud market demand. + +For an overview of Azure for AWS users, see Introduction to Azure for AWS professionals. + +Marketplace + +Area AWS service Azure service Description Marketplace AWS Marketplace Azure Marketplace Easy-to-deploy and automatically configured third-party applications, including single virtual machine or multiple virtual machine solutions. + +Compute + +Storage + +Networking and content delivery + +Area AWS service Azure service Description Cloud virtual networking Virtual Private Cloud (VPC) Virtual Network Provides an isolated, private environment in the cloud. Users have control over their virtual networking environment, including selection of their own IP address range, creation of subnets, and configuration of route tables and network gateways. Cross-premises connectivity AWS VPN Gateway Azure VPN Gateway Azure VPN Gateways connect Azure virtual networks to other Azure virtual networks, or customer on-premises networks (Site To Site). It also allows end users to connect to Azure services through VPN tunneling (Point To Site). Domain name system management Route 53 Azure DNS Manage your DNS records using the same credentials and billing and support contract as your other Azure services Route 53 Traffic Manager A service that hosts domain names, plus routes users to Internet applications, connects user requests to datacenters, manages traffic to apps, and improves app availability with automatic failover. Content delivery network CloudFront Azure Content Delivery Network A global content delivery network that delivers audio, video, applications, images, and other files. Dedicated network Direct Connect ExpressRoute Establishes a dedicated, private network connection from a location to the cloud provider (not over the Internet). Load balancing Classic Load Balancer + +Network Load Balancer + +Application Load Balancer Load Balancer + +Application Gateway Automatically distributes incoming application traffic to add +================================================================================ +Rank = 94; Score = 3407872.0 +<|begin_of_text|>How to speed up your MySQL with replication to in-memory database + +Vadim Popov Blocked Unblock Follow Following Mar 17, 2017 + +Original article available at https://habrahabr.ru/company/mailru/blog/323870/ + +I’d like to share with you an article based on my talk at Tarantool Meetup (the video is in Russian, though). It’s a short story of why Mamba, one of the biggest dating websites in the world and the largest one in Russia, started using Tarantool. Why did we decide to busy ourselves with MySQL-to-Tarantool replication? + +First, we had to migrate to MySQL 5.7 at some point, but this version didn’t have HandlerSocket that was being actively used on our MySQL 5.6 servers. We even contacted the Percona team — and they confirmed MySQL 5.6 is the last version to have HandlerSocket. + +Second, we gave Tarantool a try and were pleased with its performance. We compared it against Memcached as a key-value store and saw the speed double from 0.6 ms to 0.3 ms on the same hardware. In relative terms, Tarantool’s twice as fast as Memcached. In absolute terms, it’s not that cool, but still impressive. + +Third, we wanted to keep the whole existing architecture. There’s a MySQL master server and its slaves — we didn’t want to change anything in this structure. Can MySQL 5.6 slaves with HandlerSocket be replaced with something else without having to make significant architectural changes? + +We learned that the Mail.Ru Group team has a replicator they created for their own purposes. The idea of replicating data from MySQL to Tarantool belongs to them. We asked the team to share the source code, which they did. We had to rewrite the code, though, since it worked with MySQL 5.1 and Tarantool 1.5, not 1.7. The replicator uses libslave, an open-source solution for reading events from a MySQL master server, and is built statically without any of MySQL’s system libraries. It’s been open-sourced under the BSD license, so anyone can use it for free. + +Replication constraints + +Firstly, binary logs residing on a master must be row-based (STATEMENT and MIXED options won’t do, only ROW). Secondly, this replicator is not a Tarantool module, but a stand-alone daemon that doesn’t have anything to do +================================================================================ +Rank = 95; Score = 3407872.0 +<|begin_of_text|>I had a great time at React Europe Conf 2016 and thought I would share some of my opinions and things that stood out from the conference. + +GraphQL vs and Falcor + +I’ve heard of both these frameworks for retrieving data from the server but I’ve never really understood the benefits by just viewing the documentation. After seeing both these frameworks I can honestly see the advantages of using both. Having the ability to retrieve data from the server without specifying a REST endpoint is amazing, it allows you to scale your data model easily. + +The talk by Jafar Husain “Falcor: One Model Everywhere” demonstrated the way Falcor does it really well. He also gives credit to GraphQL saying both frameworks have a place in the ecosystem and picking the best one is subjective to your requirements. My take from his talk is that Falcor is better for smaller applications where the data model is known and that you should use GraphQL where the model may change and need to scale. + +Falcor is less prescriptive, has fewer concepts and a smaller file size < 50%. It takes a segment of the model you defined and returns it from the data source. I have a few questions around how this works with a real database because the demo was using a JSON file but it’s something I am going to investigate, now I’ve seen it in action. + +The talk by Laney Kuenzel and Lee Byron named “GraphQL Future” was a nice insight into the ways Facebook use GraphQL and what they have planned for it in the future. They talk a little bit about the different experiments they are running with GraphQL and working on the BBC LIVE product I can really see how the directive @live in GraphQL will change the future of real time updates in applications. For their demo’s they also used graphiql which looks really cool for testing out your graphQL queries. + +Both talks were really well executed and two of the best talks at React Europe Conf 2016. + +Carte Blanche + +The Evolution of React UI Development by Max Stoiber and Nik Graf stole the show for me. With the announcement of carte-blanche, a tool to help you see components individually, explore them in different states, and quickly develop them confidently will really change my development workflow. Honestly check this out, having a dependency tree with loads of components is a pain when you want to quickly develop, it may mean you have to npm link your dependency tree but with carte-blanche that’s no longer necessary. + +A Deep dive into things… + +Oh my god, these talks! The subjects +================================================================================ +Rank = 96; Score = 3391488.0 +<|begin_of_text|>Reseeding is a process of updating the failed mailbox database to be in sync with the Active mailbox database. This has been greatly improved in Exchange 2013. Some of the new and interesting features in Exchange 2013 is Automatic Reseed and multiple databases per volume. + +AutoReseed is purposed to overcome the manual efforts of replacing the failed disk with a new disk and restoring the copy of the database on the spare disk ASAP when the disk failure occurs. It involves pre-configuration of the disk and databases using mount point. Auto reseeding has a built in algorithm to perform some basic check before the spare disk is configured. Algorithm uses MS Exchange Replication service to check for the FailedAndSuspended database. Once it finds the FailedAndSuspended DB, it performs some checks, like the number of copies available, disk space, etc. If the checks are successful, then it automatically maps the spare disk and reseeding is performed. + +Multiple Databases per volume is the new design improvement when just a bunch of disks (JBOD) configurations is used for the mailbox database. This design allows hosting multiple databases on the same volume.As these days the size of the disks can be increased up to 8 TB, whereas the Exchange best practice guideline only mentions 2 TB, there could be a waste of additional 6 TB. To overcome this issue, Exchange 2013 allows multiple copies per disk. Figure 1 below illustrates the symmetrical design of a database. All four servers have the same four databases all hosted on a single disk per server. The key is that the number of copies of each database should be equal to the number of database copies per disk. + +Figure 1. Exchange 2013 multiple databases per volume + +When disk failure occurs and disk is reconfigured, then auto reseeding would be kicked from multiple source. This symmetrical design also increases the speed of reseeding from multiple source then the transitional reseeding method of single source. Figure 2 below shows the details of the reseeding from multiple source with the reseeding speed of 2 TB of data in approximately 9.7 hrs. + +Figure 2. Improved reseeding using multiple databases per volume + +Auto Reseeding and Multiple Databases per volume have greatly improved the seeding of failed database in Exchange 2013. If we do not have the configuration of spare disk and multiple database per volume and there is a database or server failure, then we have to perform the reseeding operation manually. Manual rese +================================================================================ +Rank = 97; Score = 3391488.0 +<|begin_of_text|>Visualizing TPath Data at Design Time using the LiveBindings Designer + +For today's #DelphiWeek blog post, I thought I would highlight the design-time visualization of TPath data. + +TPath data represents a series of connected curves and lines. You can use path data to build graphic shapes by connecting a series of curves and lines. You can also use path data to create the path to be followed when applying a TPathAnimation. Several program such as Inkscape allow you to draw your own vector shapes and then export the path data for use in other programs. + +This example uses the following components: + +TToolbar with a parented TLabel TPath TPrototypeBindSource bound to a TBindNavigator TPath TStyleBook (optional). I loaded the CoralCrystal premium Mac style to it. + +Steps: + +Place TToolbar onto your form, and parent a TLabel to it. Align TToolbar to the top, align TLabel to Contents, and set HorzAlign to Center. Set TLabel's StyleLookUp property to toollabel. + +Place a TPath Component onto your form. Next, place a TMemo onto your form. + +In this demo, we are going to use existing TPath sample data. Place a TPrototypeBindSource onto your form, right-click it, select 'Add Field...' and select PathData. + +Right-click on TPrototypeBindSource and select 'Add Navigator'. + +Place a TMemo onto your form and align it to the bottom. The memo is going to display the actual path data. + +Go to View->LiveBindings Designer. Bind PrototypeBindSource1.PathData1 to Path1.Data.Data. Bind Memo1.Text to PrototypeBindSource1.PathData1. + +Here is the running application. The BindNavigator component allows you to quickly navigate back and forth between the path data shapes. + +Note: You can easily add your own PathData using the Data.Data field for TPath and then preview a single shape using the Path Designer editor: + +To learn more about TPath, visit our docwiki. + +To learn more about this week's #DelphiWeek celebration, click here.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 98; Score = 3375104.0 +<|begin_of_text|>Since its launch just a few years ago, DigitalOcean quickly made a name for itself as a hosting platform for affordable virtual servers. The company’s ambition goes far beyond hosting your WordPress blog or test server on a $5/month machine however and today it is taking a next step in this direction by announcing its support for CoreOS, the highly popular container-centric Linux distribution for massive-scale deployments. + +“There is a lot of community excitement for CoreOS,” said Mitch Wainer, co-founder and CMO at DigitalOcean in a statement today. “We’re pleased to announce that developers can begin using CoreOS immediately, ensuring the resiliency of their architecture, as their ability to scale at massive levels increases.” + +DigitalOcean tells me that it believes its platform is now the “easiest and least expensive way to try out CoreOS and get started using containers on any cloud service” + +Given that running a CoreOS cluster is a bit more intricate than just firing up the latest version of Ubuntu on a single server, DigitalOcean users need to take a few extra steps, too. They need to supply a configuration file when they create a new service (or “droplet” in DigitalOcean’s parlance). CoreOS, after all, needs to know what other servers it can talk to. It’s not all that complex to set up these configuration files, but it’s also not trivial by any means. + +With the addition of CoreOS support, DigitalOcean is clearly trying to expand its reach beyond the basic virtual server market. And if it wants to be taken seriously in the competition with AWS and Google Cloud Platform, it does need to offer exactly these kind of services. + +As the company’s co-founder Ben Uretsky has often told me, the company was so busy just scaling up to meet demand over the last few years that basic product development like adding more distributions often fell by the wayside. Now that DigitalOcean has plenty of funding and a larger staff, it’s able to focus more on adding new features like today’s CoreOS launch.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 99; Score = 3358720.0 +<|begin_of_text|>Handling traffic peaks with CloudFlare CDN for free + +Every once in a while you're lucky and end up with a positive problem - your website content is suddenly very popular. You might scramble and start turning up your servers and tuning up your caches or maybe someone's de-facto solution is to install HHVM to run your WordPress faster. While this is all worth while if you plan for this to happen in the future as well, for and occasional hit piece of content it might not be worth it. + +I've ran a dormant news site, metropolitan.fi, on Bolt since June 2013 for fun. On an average day it barely registers on any meters, averaging maybe 10-20 visitors a day. I run the site on a Virtual Server with multiple other sites on it (including this one). + +The publication is by no means a triumph of journalism, it's more of an experiment. And while I don't normally even think about the service I have twice hit a niche and spark discussions by translating some news from Finland that have been noteworthy Globally: + +While it's fun to attract some readers, it's not something I want to loose sleep over. The server runs other services and I'd rather not have them be affected by the traffic. The server would probably have managed this just fine without any tweaks, but as a life insurance I set up CloudFlare caching on it back in June 2013. + +Promise of the CloudFlare Free Plan is: + +Peace of mind about running your website so you can get back to what you love + +CloudFlare offers a number of services, but the key take away for me is the advanced Content Delivery Network (CDN). This let's me easily scale up for load peaks when I need - actually even when I know don't need since the service is always-on. + +The CDN platform delivers content with speed and closest to where the users is - with low latency. This results in lower load on my server and a better experience for a global audience. In addition to delivery, CloudFlare also offers other optimizations (resource minification, etc.). I've ran with default settings and it works just fine for such a simple content driven site. + +So in the title your read that it's free. So what's the catch? Well, you'll run out of the free plan features at some point and then you'll want more. That's when you start paying. Just like with Slack and many other SaaS platforms. It's "freemium" and it works for my use \ No newline at end of file diff --git a/examples/openwebtext/files/doctor.txt b/examples/openwebtext/files/doctor.txt new file mode 100644 index 0000000..3fb6415 --- /dev/null +++ b/examples/openwebtext/files/doctor.txt @@ -0,0 +1,1784 @@ +Query Sequence: +Prompt: Doctors are male and nurses are; Completion: female. + +Top Influential Sequences: +================================================================================ +Rank = 0; Score = 872448.0 +<|begin_of_text|>Citizens have 911. Employees have the EEOC. Distressed sailors have the Coast Guard. + +But what do America’s college students have? Where can they turn when they find themselves outside campus “safe spaces” and suffering a “microaggression”? + +Fortunately, the University of Arizona has an answer. It recently distributed a 20-page booklet suggesting to faculty that when a student is victimized by a microaggression the appropriate response should be saying “ouch.” And the correct response for the offender should be saying “oops,” according to the guide. + +THE WEEK IN PICTURES + +“If a student feels hurt or offended by another student’s comment, the hurt student can say ‘ouch.’ In acknowledgement, the student who made the hurtful comment says “oops.” If necessary, there can be further dialogue about this exchange.” + +The guide was authored by Jesus Treviño, vice provost of the big taxpayer-funded university, whose salary reportedly is $214,000 per year. + +For those unfamiliar with the apparently epidemic scale of microaggression and thus not able to spot such offenses, the booklet offers a definition: Microaggressions are “the everyday verbal, nonverbal, and environmental slights, snubs or insults, whether intentional or unintentional, that communicate hostile, derogatory, or negative messages to target persons based solely upon their marginalized group membership.” + +The University of Arizona isn’t the first to suggest the “ouch”/”oops” protocol. Iowa State University, among others, has also urged a similar approach.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 1; Score = 503808.0 +<|begin_of_text|>Most Democrats want stricter gun laws, and most Republicans oppose them. So you might expect the election of Donald Trump and a Republican-controlled Congress to be great news for the nation’s gunmakers. But surprisingly, the opposite turned out to be the case. + +Two of the nation’s biggest gun manufacturers, Smith & Wesson and Sturm Ruger, saw their stock prices fall by double digits this morning in the wake of Trump’s election victory. + +Here’s one theory for what’s going on: In recent years, gun sales have been driven by fears of gun confiscations. Gun sales soared after Obama’s election in 2008 and again when he was reelected in 2012, as people worried that the president would begin to confiscate the nation’s firearms. + +In contrast, gun owners have little to fear from a Trump presidency or a Republican Congress. So gunmakers are likely to make fewer sales over the next four years.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 2; Score = 503808.0 +<|begin_of_text|>WOMEN and men face double-standards. That this should show up in the language is no surprise. Men who put themselves forward at work are “assertive”, women who do the same are more often “pushy” or “bossy”; men are “persistent” whereas women are “nagging”; men are “frustrated”, women “upset”. A man has a lot to say; a woman is “chatty”. A man discusses the doings of his colleagues and rivals; a woman “gossips”. + +Readers tempted to doubt can check for themselves. For an impressionistic survey, type “gossip” into Google, click on “images” and see who appears to be doing it; then try the same with “nagging” and “bossy”. For hard data, try Google’s “Ngram” viewer, which shows the frequency of words and phrases among the hundreds of billions of words in the books scanned by Google, spanning centuries. One of the most common words following “gossiping” is “old”. And the most common words to follow “gossiping old” are, in this order: “women”, “woman”, “men”, “lady” and “ladies”. + +Some words are trickier than mere double-standards: those using them may think they are paying a kind of compliment, whereas what is heard is something between condescension and insult. A case in point is “feisty”. Those who use it might think that the word connotes “spirited”. It is often heard by women, though, as carrying a whiff of surprise that a woman would show such spirit. + +“Nonsense”, some will reply. The Economist has used “feisty” recently to refer to Greece’s leftist government, a South African tabloid, a (male) Argentinian presidential candidate and Singaporean opposition bloggers. But it is also used fairly frequently with female figures. The common thread seems to be a sense of smallness or underdog status: nobody calls a jowly dictator or heavyweight boxer “feisty”. Google’s book data say much the same. “Little” is one of the most common words to follow “feisty”, and the most common words to follow “feisty little” are “girl”, “man”, “thing”, “guy”, “woman” and “lady”. Rounding out the top ten words following “feisty little”, intriguingly, are “Irishman” and “bastard +================================================================================ +Rank = 3; Score = 456704.0 +<|begin_of_text|>Utah’s Wasatch GOP has two chairmen. One of them is named James C. Green. Mr. Green has some big ideas and he decided to share one of those big ideas with the public. + +Here’s Mr. Green’s caveman missive: + +Equal Pay For Women Has Consequences + +Editor: Here's the problem with the Equal Pay bill being considered by the Utah Legislature... Traditionally men have earned more than women in the workplace because they are considered the primary breadwinners for families. They need to make enough to support their families and allow the Mother to remain in the home to raise and nurture the children. + +If businesses are forced to pay women the same as male earnings, that means they will have to reduce the pay for the men they employ... simple economics. If that happens, then men will have an even more difficult time earning enough to support their families, which will mean more Mothers will be forced to leave the home (where they may prefer to be) to join the workforce to make up the difference. + +And as even more women thus enter the workforce that creates more competition for jobs (even men's jobs) and puts further downward pressure on the pay for all jobs... meaning more and more Mothers will be forced into the workforce. And that is bad for families and thus for all of society. + +It's a vicious cycle that only gets worse the more equality of pay is forced upon us. It's a situation of well-meaning intentions, but negative unintended consequences. + +We should encourage our Legislators to drop the whole notion. Let the marketplace determine what free-market forces should prevail. It is not the role of government to dictate to businesses what they should pay anyway... either as a Minimum Wage or Equal Pay for men and women. + +James C. Green + +Wasatch Co. GOP Vice-Chair<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 4; Score = 444416.0 +<|begin_of_text|>Gay, bi or straight, people are having same-sex experience in their lifetime. + +More American men and women are identifying as bisexual, reports a recent publication by the Centers for Disease Control and Prevention. + +The study, which utilised the data collected from the 2011-2013 National Survey of Family Growth (NSFG), aims to provide ‘national estimates of sexual behaviours, sexual attraction, and sexual orientation among women and men aged 18–44 in the United States.’ + +Over 9,000 adults’ interview results were used to churn out the end statistics. + +The outcome reports that 5.5% of women and 2.0% of men said they are bisexual in the 2011–2013 NSFG, as compared to the previous 2006–2010 NSFG study which found that only 3.9% of women and 1.2% of men identified as bisexual. + +And among those who reported themselves as ‘heterosexual or straight,’ 12.6% of women and 2.8% of men said that they had had same-sex sexual contact. + +The study also provides other interesting insights about what different couples do in bed: + +Women who reported to be ‘homosexual or bisexual’ were more likely to have had anal sex with an opposite-sex partner (44.2%) compared with 35.4% of heterosexual women. + +More women in the ‘homosexual or bisexual’ group had ever had opposite-sex sexual contact (89.7%), as compared to men (67.9%). + +And surprisingly, or not, 16.4% of women and 11.4% of men who had identified as ‘homosexual or bisexual’ had never had same-sex sexual contact.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 5; Score = 436224.0 +<|begin_of_text|>What’s happening to female characters in video games? They’ve been around forever and they’ve seen the changes of growing technology and differing trends, but it never felt like they ever truly lost anything. They simply grew with the stories and games, maybe became a little more numerous as games started including more characters and options. Recently though, they have changed. Not due to the natural evolution of the medium, but seemingly due to this idea rooted in the stereotype that only dudes game, and because only dudes game, the only way we can view female characters is through the eyes of guys. And thus while male characters are perfectly allowed to remain muscle-bound, stoic, and badass, women must be “realistic” because otherwise guys might get the wrong idea. + +Now, aside from the fact that this is pretty insulting to the intelligence and motivations of men, it’s a little bizarre for the same people who want more women in the industry and enjoying video games to perpetually only address things from a strictly male perspective. A purely negative perspective at that, assuming only the very worst outlook from male gamers. Never have I seen from the arguments saying female characters need to change any reference to what women want in female characters. Which is a shame, because if we did, those characters wouldn’t change much. Because, objectively, there is enough diversity in how female characters present themselves that I believe most of the women who play are happy with it. + +The argument has already been made that men tend to be just as sexy and unrealistic as women when it comes to video games. And the usual rebuttal is “That’s just a male power fantasy.” Consider, maybe, that the sexy but deadly female heroine might be just as much a female power fantasy. Or the not-so-sexy but kind and passionate young girl who always strives to improve the lives of others. Or the badass female space marine fighting for her home. Has it occurred to no one making the argument against sexy characters that women have just as much of a need to fantasize—to become someone who is so unlike them, and yet, seems to hold the same values? Or even to become someone completely different from them just to unwind and forget life for a few hours? + +Speaking as a woman, I don’t play games for their realism. I don’t mind if a game embraces a little realism, as long as they’re still clearly rooted in fiction. I am especially partial to fantasy, and whether it’s a character I have made completely from scratch for a pen and paper roleplay, or a character +================================================================================ +Rank = 6; Score = 434176.0 +<|begin_of_text|>The forestry industry is largely inaccessible to women for many reasons; however, Forestry Training for Women will remedy that. Personal Protective Equipment [PPE] is not made in small enough sizes, both foot-wear and clothes to fit many women. Training for using forestry kit, in particular chainsaws, is taught in such a way as to emphasize physical strength when in fact using the “right” technique can get over most issues. If funded, this course will provide an environment to practice these techniques in private with other female participants. Students would learn to service and sharpen saws and hand tools together with all aspects of woodland and forestry skills and work. This would empower students to compete on equal terms in the job market with their male counterparts. Students will also be taught about tree identification, the various uses they serve in the landscape and small sawmill operation. + +The photograph above shows the various tools used in forestry and the difference between men and womens sized boots. + +Digger driving is another aspect of forestry that is seen as a male dominated job. Lighter women sometimes cannot use digger cabs, levers and pedals effectively. A second component of this course would address the use of a Digger and provide women with the training needed to operate one comfortably. This course will teach women to service and care for their machine, how to operate it and provide a few simple adaptations to overcome the male oriented aspects of the machine. The course will cover and practice the most common requirements such as trench digging and ditching along with a range of different scenarios that might be encountered on a daily level. + +There are few opportunities for women to compete and train on an equal level to men. We aim to address this by offering training and real life practice in a safe and all encompassing environment where women would feel comfortable to learn. The funding would allow us to purchase a second hand digger for training purposes, saws and PPE for participants. + +It is important to note that instructors will be male and while this course is aimed at women students, male participants are encouraged to join if interested. + +A forestry winch attached to a tractor. This allows for safer removal of large trees and debris, lowering the risk of injury.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 7; Score = 432128.0 +<|begin_of_text|>Prostitutes are penis experts. There is no other group of professionals – not scientists, not academic researchers, not journalists – that have more experience viewing and handling a large number and wide variety of penises. Additionally, prostitutes are probably the only professional group qualified to comment on how penises of differing sizes and shapes feel when they are inside a woman. + +It seems like every month another article about the latest penis size study is published. It’s a popular media topic because men are becoming increasingly obsessed with the size of their genitals, thanks in part to the wide availability of online porn showing women writhing in pleasure as they receive a plunging from one large penis after another. + +There is a lot of speculation about what the perfect penis is supposed to look like, how long or wide it needs to be in order to deliver pleasure, and what’s considered “ideal.” I decided to share my informed opinion as a licensed Nevada prostitute and write about certain myths that many people tend to believe, either due to lack of sexual experience or a limited exposure to all of the various penis shapes and sizes. Below are what I consider to be the three most common myths about penis size. + +Myth #1: There is such a thing as an “ideal” penis size + +I’ve seen a lot of cocks. They come in all shapes, lengths, widths, and colors. Some bend upwards, some downwards, some curve to the side. Some are almost the size of my forearm, some the size of my thumb. I’ve seen penises with a sharp bend like a water faucet, and penises shaped like a mushroom, with a large head on a smaller shaft. While some online resources will try to trick people into believing that there is such a thing as an “ideal penis,” I’m here to tell you that as long as your penis is healthy, it’s ideal for sex and is capable of pleasing a woman. + +There are so many ways that one healthy penis can differ from the next, that it’s pointless to try to figure out what’s “normal” or “ideal.” Sure, you can go on Wikipedia, read that the average penis length is 5.17 inches, get an idea as to how you measure up against this average, and feel better (or worse) about yourself. But comparing yourself to a statistical average or striving for an ideal will not only give you a false perception of the wonderful diversity of healthy penises, but also impair your ability to see yourself as a valuable individual who uses your unique body to be the best +================================================================================ +Rank = 8; Score = 425984.0 +<|begin_of_text|>The issue is particularly nettlesome with individuals who have transitioned from male to female because of the belief that they might still have a physical advantage until their hormone therapy — which not every transgender person chooses to have — is complete. + +“When you start talking about transgender athletes, a male-to-female individual, we want to ensure that that is truly a decision that is permanent,” said Bobby Cox, the commissioner of the Indiana High School Athletic Association. “It is not a decision that, ‘I just decided today that I am going to be a girl and I am going to go play on a girls’ team’ and perhaps, disadvantage those kids that are on the team and imbalance the competition. + +“And as we progress down this path,’’ he added, “and as we spend more time and energy with advocacy groups and medical professionals in this area, I think that there will be additional amendments to our policies.” + +There is no reliable data on the number of transgender high school athletes. Only about 0.6 percent of the adult population identifies as transgender, according to federal data from 2016. Researchers from the Williams Institute, who conducted the study, reported that 0.56 percent of adults in Indiana reported identifying as transgender. + +In Texas, 0.66 percent of adults said they identify as transgender, the fifth-highest percentage in the country (behind Hawaii, California, New Mexico and Georgia). Still, transgender children are considered an at-risk minority outside of sports. According to The New England Journal of Medicine, the rate of suicide attempts among transgender people is 40 percent, compared to 4.6 percent among those who are not transgender. + +Relatively new policies are already being massaged and amended. In July, the Indiana state association voted unanimously to adjust its surgery requirement. It now no longer requires transgender students who are transitioning from female to male to have sex reassignment surgery in order to compete in the gender with which they identify. But that stipulation is still in place for someone transitioning from male to female. + +The group asserts it is acting in the name of fairness, but transgender rights activists accuse members of simply not wanting transgender people to participate, out of fear that those athletes will have an unfair advantage.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 9; Score = 425984.0 +<|begin_of_text|>A connection between lesbian lust and identifying as male came to me after reading the prologue to Audre Lorde’s Zami : A New Spelling of My Name (full text here.) Lorde describes wanting to be both male and female and she fantasized about having a penis. It’s increasingly common for lesbians to think they are men so I think this needs to be discussed. In this post I’m going to describe my interpretation of Lorde’s quote, which was written in 1982, and then discuss the phenomenon of lesbians feeling “male” within the current context of porn culture and transgenderism. + +“I have always wanted to be both man and woman, to incorporate the strongest and richest parts of my mother and father within/into me—to share valleys and mountains upon my body the way the earth does in hills and peaks. I would like to enter a woman the way any man can, and to be entered—to leave and to be left—to be hot and hard and soft all at the same time in the cause of our loving. I would like to drive forward and at other times to rest or be driven. When I sit and play in the waters of my bath I love to feel the deep inside parts of me, sliding and folded and tender and deep. Other times I like to fantasize the core of it, my pearl, a protruding part of me, hard and sensitive and vulnerable in a different way. I have felt the age-old triangle of mother father and child, with the “I” at its eternal core, elongate and flatten out into the elegantly strong triad of grandmother mother daughter, with the “I” moving back and forth flowing in either or both directions as needed. Woman forever. My body, a living representation of other life older longer wiser. The mountains and valleys, trees, rocks. Sand and flowers and water and stone. Made in earth.”-Audre Lorde + +This quote is rather poetic and I have allowed myself to see lots of things in it that I don’t know if Lorde actually intended. These are some things I see when I look at these words. + +Lorde has described wanting the best qualities of her mother and father and wanting to be both hard and soft during sex. When I read this I thought about how people are expected to limit their self-expression and interests to what society believes is appropriate for their sex, and Lorde is refusing to limit herself that way. She wants the full human experience, not just a half of it. Gender roles and +================================================================================ +Rank = 10; Score = 413696.0 +<|begin_of_text|>Nasa + +Mycorrhizal fungi are more like cities than individual plants; these are complex systems which stretch for miles underground in search for nutrients. + +So massive and impactful are they, Nasa has now been able to identify them from space. + +Advertisement + +Researchers say it could finding and studying these networks could help them predict how climate change will affect forest ecosystems. "Individual tree species have unique spectral fingerprints," said Joshua Fisher, lead author of the study. "But we thought the underlying fungi could be controlling them as groups." + +Two types of the fungi control different sets of trees, with each type responding differently to climate change. Knowing where each type is dominant "may help us predict where forests will thrive in the future and where they'll falter", the team says. + +Ceating fungi maps to chart this growth has proved difficult, as it has traditionally depended on "counting individual tree species". So the team decided to detect the network via satellite. + +Using the "spectral signature" of the trees -- that is, the pattern that the tree's canopies reflect -- the satellite can differentiate between the trees. Four forests were analysed during the project -- around 130,000 trees and 77 species -- by Nasa's Landsat-5 satellite. The images were then analysed using a statistical model that was able to predict where the different types of fungus were most prevalent. This model, the team say, had around 77 percent accuracy. "That these below-ground agents manifest themselves in changes in the forest canopies is significant," said Fisher. "This allows, for the first time, some light to be shed on their hidden processes."<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 11; Score = 387072.0 +<|begin_of_text|>Men and women make moral decisions differently, according to new research published in Personality and Social Psychology Bulletin — but not necessarily in the ways scientists thought before. + +The study, a meta-analysis of 40 studies with 6,100 total participants, revealed a more nuanced distinction between two common schools of moral thought. + +Deontology vs. Utilitarianism + +“Consider…a dilemma where you are hiding with other townsfolk from murderous soldiers,” says Rebecca Friesdorf, the corresponding author. “Suddenly a baby starts to cry—unless you smother it, the soldiers will find and kill everyone. Should you smother the baby to prevent the soldiers from killing the townsfolk?” + +According to deontology, killing the baby would be wrong because murder is wrong, despite the potential consequences. Deontology puts moral norms at top priority. + +According to utilitarianism, killing the baby is acceptable because it will save many lives in the end. Utilitarianism considers the overall consequences more important. + +Deontology has long been associated with emotional responses to moral dilemmas, while utilitarianism is associated with cognitive, or mental, responses. Previous research has claimed that the two are opposites, and that women make more deontological decisions while men make more utilitarian decisions. However, the current study suggests it’s not so simple. + +The Study + +The research team, comprised of scientists from three universities in three different countries, believes that deontology and utilitarianism are not necessarily opposites—either thought process could be affected by a number of factors. In order to examine them independently, the team used a method called Process Dissociation (PD) to separate the variables. PD has been used previously to separate other controversial variables, such as racial bias in weapon identification. + +“To…distinguish whether men are more utilitarian than women, or women are more deontological than men, it is necessary to measure [these]inclinations independently,” said Friesdorf. + +The Results + +The findings from the study challenge the traditional theory that men prefer cognitive decisions and women prefer emotional ones. After separating the variables, researchers discovered that men and women use cognitive reasoning about equally. However, women were much more likely to use emotional reasoning than men when one factor is involved: harm. + +When asked a moral dilemma question that did not involve harm, women and men both tended to use utilitarian (or cognitive) thinking. However, when asked a moral dilemma question involving harm, women were significantly more likely to use deontological (emotional) thinking than men. + +“The current findings +================================================================================ +Rank = 12; Score = 380928.0 +<|begin_of_text|>Government is a big insurance company. Except you can’t shop elsewhere for insurance. And their price is super high. And they threaten you if you don’t want to pay them. And the services they deliver are sub-par for the price. And you can’t itemize your “insurance” to only include the products you need. And often just when you need them most, you find out how incompetent, corrupt, and overall ineffective they are at ensuring you against anything. + +Just look at Puerto Rico after Hurricane Maria. Governments play politics with awarding repair projects, while the majority of the island is still left without power. Daddy Yankee was a bigger help than FEMA. + +And because the government really doesn’t deliver on their promise of keeping people safe and coming to their aid, people still turn to private companies. People still find it in their best interest to use some of the 50% of their income left after the government plunders the rest to purchase insurance. + +You pay property taxes, and these go to support indoctrination centers also known as public schools. They pay for local roads, local traffic enforcers, and firefighters. But these services are collectivized. The firefighters aren’t really there to save your home, they are there to keep the whole town from burning down. + +Firefighters are likely some of the most dedicated public employees. The ones I have met often go into the profession for the right reasons, and take seriously their commitment to saving lives. Many are even volunteers. But still, they can only show up after they have been called. Many bravely rescue people from serious situations. But rarely is a structure saved. + +That is where private insurance comes in. You purchase private home insurance so that in the event that your home burns down, you don’t lose all the money that has gone into it. But some companies go above and beyond when you purchase insurance from them. + +Increasingly, insurance carriers are finding wildfires, such as those in California, are an opportunity to provide protection beyond what most people get through publicly funded fire fighting. Some insurers say they typically get new customers when homeowners see the special treatment received by neighbors during big fires. “The enrollment has taken off dramatically over the years as people have seen us save homes,” Paul Krump, a senior executive at Chubb, said of the insurer’s Wildfire Defense Services. “It’s absolutely growing leaps and bounds.” + +Chubb provides extra protection to the homes they insure in certain zip codes that are prone to wildfires. They get private firefighters on the scene before the building +================================================================================ +Rank = 13; Score = 380928.0 +<|begin_of_text|>Are young men giving up on women and checking out of society? That was a question recently posed in two articles (article 1 and article 2), which commented on the global trend of young adult males who are retreating into porn, video games, and lad culture. These articles resonated with so many readers that they garnered over 500 comments from men. People even took to Twitter, Facebook, and YouTube to share in the conversation. While there was much talk about the problem, there was little talk about the solution, which bothered me as a mother of three daughters who is concerned about their prospects for relating to the opposite sex, as they grow older. So, I’d like to share some advice in an attempt towards bringing healing and unity between the sexes. + +Stalemate: The Current State of Affairs in the War Of The Sexes + +Popular culture tells us that the cause of the war between the sexes is patriarchal and misogynistic men. As a result, we now have women engaging in lesbianized third-wave feminism. In this most recent wave, we have women who throw tampons and condoms at government buildings or show up angry and bare-breasted in order to protest for the right to kill their own children. As ridiculous as that sounds, it’s just one example of what they’re doing, and it is why many women like myself want NOTHING to do with the feminist movement. We don’t identify with the victim mentality. We don’t believe that our own empowerment will be found and maintained by disempowering men or our own children. + +Although there are many women who believe like me, many college age students today are getting lessons on feminism or are being required to take mandatory consent classes for dating. Some say that college age men today are completely clueless or scared about how to engage with women. There is a fear of being accused of rape, along with a fear of court and educational systems that seemingly give women preferential treatment. + +"I'm an athlete. My parents have a lot of money. I have plenty of friends and a good social life. I don't hang out with women any more. Occasionally I'll have one night stands, but mostly I fill my time with other things. I got accused of molesting a girl at college and since then I've just thought, whatever. I play sports instead." ~ Francis, 28 + +Not only is this lifestyle pervading the United States and Great Britain, but also Japan where so-called ‘herbivores’ (translated as grass-eating boys) are not +================================================================================ +Rank = 14; Score = 372736.0 +<|begin_of_text|>Uploaded by jonbone17 on Feb 18, 2007 + +An inside look at Hoa Lo Prison, which American POWs nicknamed “Hanoi Hilton” in the 1960s + +He spat at Ms. Fonda, was clubbed, and was dragged away. During the subsequent beating, he fell forward on to the camp Commandant ‘s feet, which sent that officer berserk. + +In 1978, the Air Force Colonel still suffered from double vision (which permanently ended his flying career) from the Commandant’s frenzied application of a wooden baton. + +From 1963-65, Col. Larry Carrigan was in the 47FW/DO (F-4E’s). He spent 6 years in the ‘ Hanoi Hilton’…the first three of which his family only knew he was ‘missing in action’. His wife lived on faith that he was still alive. His group, too, got the cleaned-up, fed and clothed routine in preparation for a ‘peace delegation’ visit. + +They, however, had time and devised a plan to get word to the world that they were alive and still survived. Each man secreted a tiny piece of paper, with his Social Security Number on it, in the palm of his hand. When paraded before Ms. Fonda and a cameraman, she walked the line, shaking each man’s hand and asking little encouraging snippets like: ‘Aren’t you sorry you bombed babies?’ and ‘Are you grateful for the humane treatment from your benevolent captors?’ Believing this HAD to be an act, they each palmed her their sliver of paper. + +She took them all without missing a beat.. At the end of the line and once the camera stopped rolling, to the shocked disbelief of the POWs, she turned to the officer in charge and handed him all the little pieces of paper… + +Three men died from the subsequent beatings. Colonel Carrigan was almost number four but he survived, which is the only reason we know of her actions that day. + +I was a civilian economic development advisor in Vietnam, and was captured by the North Vietnamese communists in South Vietnam in 1968, and held prisoner for over 5 years. + +I spent 27 months in solitary confinement; one year in a cage in Cambodia ; and one year in a ‘black box’ in Hanoi. My North Vietnamese captors deliberately poisoned and murdered a female missionary, a nurse in a leprosarium in Banme Thuot, South +================================================================================ +Rank = 15; Score = 372736.0 +<|begin_of_text|>Female American athletes get less coverage in the media due to gender bias and instead what attention they get focuses more on attire, or how attractive, sexy or ladylike they are, write Emily Kaskan and Ivy Ho of the University of Massachusetts Lowell in Sex Roles, an interdisciplinary behavioral science journal offering a feminist perspective. + +Kaskan and Ho looked at how pervasive small subtle biases and stereotyping of American female athletes are and what types of "microaggression" exist, examining how they put pressure on athletes and other women, as well. They reviewed popular Internet articles and research from the Psychinfo database, using keywords such as'sexism,''sports media,' 'Serena Williams' and 'Olympic coverage.' + +They conclude that media often portray female athletes as inferior to their male counterparts and are dismissive of their true abilities. The little coverage received often sexually objectifies female athletes by putting the spotlight on their looks and strength. On the other hand, the media is quick to recoil at women who do not fit into the traditional feminine mold. + +Anna Kournikova. EPA/Gerry Penny + +Is that an accurate characterization? Russian tennis star Anna Kournikova certainly got far more attention than her athletic prowess deserved, but so has Michael Sam. He is somehow among GQ's 2014 “Men of the Year” despite the fact that he was barely drafted by an NFL team, and the only is because hs is gay. A journal devoted to promoting the perspectives of heterosexual men could argue that the media was giving him special attention due to that. + +Outside culture studies and sociology, there is no credible acceptance of gender-based microaggressions so the authors invoke racial discrimination. Kaskan and Ho argue that subtle biases that place cumulative stress on female athletes influence how they think about themselves and their abilities. + +Yes, Ronda Rousey must not know she can obliterate 99.9% of men just by glaring at them due to subtle microaggression on ESPN. In real life, she absolutely does know that, and she is arguably the most famous name in mixed martial arts - the most male of male sports - yet she knows she can look like this any time she wants too: + +That's not a problem for athletic women, it's a gigantic victory. + +Regardless, the authors believe that prolonged psychological stress can even trigger changes in their hormonal, immune and cardiovascular systems, and make athletes vulnerable to, among others, heart disease and hypertension. This, in turn, can influence a professional athlete's livelihood when she loses her place in +================================================================================ +Rank = 16; Score = 364544.0 +<|begin_of_text|>Avoiding the male hierarchy is a major part of men’s rights activism and zeta-male culture. And for a good reason I might add. This hierarchy is what causes division and conflict between men, by encouraging them to compete for status, while objectifying them on the basis of their utility to women. Yet, are there other even deeper reasons for us to opt out? + +This can be answered by comparing the male social hierarchy to the hierarchy of the Indian caste system. The Brahmans (or the priest class) created the caste system in which they placed themselves at the top. Then, secondly, the soldier-ruler (Kshatriyas) caste, third the farmer, craftsman, merchant (Vaisyas) caste, and fourth the labourer or slave (Sudra) caste. + +Under the caste system, the Brahmans lived a life of privilege. They were prohibited from manual labour and have the belief that they are inherently of greater moral purity than members of other castes. They alone had control over certain vital religious tasks giving them great power. + +Although they were the highest caste, they did not normally hold official political positions (kings and rulers were all of the Kshatriya caste) yet due to their prestige and religious power, Brahmans had great influence over secular affairs from behind the scenes. + +At the other end of the spectrum, Sudras were denied the privileges of the higher castes. Because of the degrading work and pitiable occupations to which they were assigned, such as cleaning waste and disposing of contaminated dead animals. + +They were considered unclean and not fit to touch or come in contact with anyone of a higher caste, in fact, the untouchable was required to cross the street or get off the road entirely when the occasion demanded so that his shadow would not fall on anyone of the Brahman caste and in fear of contamination. + +Women, like Brahmans, also live a life of relative privilege in comparison to men. They have been socialized from an early age to believe that as women they are morally superior to men. + +The Brahmans cleverly kept this order by teaching the castes below them that observance of the rules and restrictions of caste without violation constituted the highest religious fulfilment and would make possible being born again into a better life. Thus the duties and obligations of caste were stressed rather than human rights, and through this system Hindu society was kept stable for over three thousand years. + +How does this relate to the male hierarchy? I will explain how by +================================================================================ +Rank = 17; Score = 360448.0 +<|begin_of_text|>Posted by Chewy + +You’ve always assumed the calico cat that sits in your neighbor’s window is a she. And you’re certain that the orange tabby you’ve fallen in love with at the shelter is a boy. Chances are, you’re right. Most orange cats are male and most calicos are female. + +The color of a cat’s coat is closely linked to its gender. As you may recall from high school biology, mammals have two chromosomes that determine their sex—XX for females and XY for males. But a number of additional chromosomes are present and vary depending upon species, says Dr. Robert Grahn, a forensic analyst at the veterinary genetics laboratory at the University of California in Davis. + +“These other chromosomes contain genes that affect hair color, pattern, shape and length,” Grahn says. “Since the genes for sex and hair colors are on different chromosomes, they are inherited independently of each other. Thus, no color is associated with a particular sex, except in cats and hamsters.” + +Nature doesn’t always abide by a rigid set of rules, however, including when it comes to feline fur color. A small percentage of orange cats are female, and even a more miniscule portion of calico cats are male. + +Below, learn how genetics and sex influences a cat’s coat color, and why some cats don’t fit typical color patterns. + +Color in Cats is (Mostly) Linked to Sex + +Whether calico, tortoiseshell, orange, black, brown, or gray, a cat’s fur color is derived from two dominant colors: Black and red. These colors can mutate into different shades—black can become chocolate, cinnamon, lilac, blue and fawn. And red, which is determined by the orange gene, can become cream. + +The color genes for black and red in cats are contained within the X chromosome. This is the same chromosome that, along with Y in males, determine a cat’s sex, says Dr. Jerold Bell, adjunct professor of genetics at Cummings School of Veterinary Medicine at Tufts University in North Grafton, Massachusetts. + +“They are actually alleles, meaning they are two variations of the same gene in one location on the chromosome,” he says. So an X chromosome can contain either a black hair gene or an orange hair gene, but not both. + +“One allele will create orange coloration. This allele will cover up all other colors, except pure white. The other allele will create a non-orange coloration. This allele is ‘recessive’ and allows for +================================================================================ +Rank = 18; Score = 360448.0 +<|begin_of_text|>Male, female, straight, gay, lesbian, transgender — labels don't matter at Iowa's Grinnell College. Students can share a dormitory room, bathroom, shower room or locker room with any of the above, if they choose. + +This year, the private liberal arts college on the Iowa prairie has added a gender-neutral locker room to its mix of gender-neutral options. It's the most recent step for the school, which became the state's only college to offer a gender-neutral dorm option three years ago as part of a growing trend nationwide. + +In what the National Student Genderblind Campaign calls a "rapidly growing collegiate movement," at least 54 colleges and universities across the country, many of them private, provide gender-neutral housing options this year. + +"Each year we hear about more campuses adding that as an option," said James Baumann, spokesman for the Association of College and University Housing Officers — International. "I think the changes are a reflection of changing times." + +It was such a big deal when Ohio's Oberlin College added co-ed halls in the 1970s that it made the cover of Life magazine. Then, mixed-gender floors became common, Baumann says. + +"It's usually a campus responding to a need," Baumann said. "If the student body either overtly expresses a desire for this, or if the housing staff perceives a need for it, that's when they pursue this as an option." + +Starting next fall, Catholic University in Washington will move in the other direction, back to single-sex dorms. John Garvey, the university's president, said separating the genders will cut down on binge drinking and casual sex, two forces that are destructive for college students. + +Garvey said his institution wants to teach students about virtue, and he sees only ill coming from other colleges' moves toward mixed-gender housing. + +"I think it's a bad idea," Garvey said. "I think it's naive to imagine that we're doing a good thing shoving college students together." + +At Grinnell, the gender-neutral idea was driven by transgender students, those who don't identify themselves as either male or female, and students transitioning from one gender to the other, students and school officials said. + +"It's about creating a safe and welcoming space. It's not about being able to room with your significant other," said Lily Cross, a senior who helped persuade Grinnell officials to offer gender-neutral housing. + +Gender-neutral grew from 1% of the school's on-campus housing in 2008- +================================================================================ +Rank = 19; Score = 358400.0 +<|begin_of_text|>Old people aren't good at voting. I don't simply mean that they don't vote like I do, or I would also say that old people are bad at watching TV, picking restaurants and gauging my interest in stories about their friends' children. + +No, old people vote shortsightedly, choosing the least progressive outcome. In surveys in the U.S. and the U.K., people over 65 — compared with people under 30 — were nearly twice as likely to be against gay marriage; twice as likely to be pro-Brexit; half as likely to support legalization of marijuana; nearly five times less likely to want to spend money on education; 60% more likely to vote for Donald Trump; and nearly 50% more likely to say immigrants have a negative impact on society, despite the fact that they are being wheeled around by them. Whether these figures are accurate is irrelevant, since old people are so bad at Googling. + +Their poor choices were always a problem, but it's become far more acute now that there are so many of them. The over-65 generation does not accurately represent our country, because they are overwhelmingly white and actually vote. So, unfortunately, we're going to have to bar them from voting. + +When I ran this idea by my 76-year-old father, he partly agreed. "If a 90-year-old has what you can define as dementia, you can test for it," he said. I explained that I was thinking more of 76. I also explained that the problem wasn't that seniors forget things as much as they remember the world of 60 years ago and want to reproduce it. But one challenge of arguing with older people is that because of all the cable-news watching and Founding Father — book reading, they know more. My dad mentioned stare decisis, using Edmund Burke's argument that society is so complex that radical change often has horrible unintended negative consequences such as Napoleon, communism and that horrible rap music. + +He also argued that experts are often wrong: In 1980, pundits thought a movie star shouldn't be President; during the Great Recession, economists were sure the Fed's expansionary monetary policy would lead to inflation. "We might think that the England they voted for isn't going to do very well in the world. But we don't know that," he said. And said. And said. + +Looking for a counterargument, I called Jason Brennan, a Georgetown philosophy professor whose new book is titled Against Democracy. He said my antid +================================================================================ +Rank = 20; Score = 354304.0 +<|begin_of_text|>We’re all getting fatter, but Irish men are getting fatter fastest. The prognosis is poor: by 2030, 58 per cent of Irish men are expected to be obese, topping the scale of all European countries, according to a UK-based study published last year. If you add the projected number of overweight men to the Irish total, the figure is even worse, coming in at a gut-busting 90 per cent. + +Now a new review of men’s food behaviour by Safefood, the all-Ireland health body, shows that we’re well on our way to fulfilling that frightening prediction. It points out that 70 per cent of men in the Republic are currently overweight or obese, compared with 50 per cent of women, as are 69 per cent of men in the North, against 57 per cent of women. + +The big question, of course, is why. What is it about living on this island that makes Irish men, in particular, stuff themselves silly, with so little regard for the health consequences? + +Many reasons – from simple greed or laziness, to more complex speculations about how men respond to stress, or marriage, or whether they are especially sensitive to the effects of chemicals such as phthalates, commonly found in household products – have been forwarded for male weight gain. + +Dr Aileen McGloin, from Safefood, has a list of contributing factors as long as your arm, all of which, it seems, can be summed up by one single sentence. “Men,” she says, “live in a different food world.” + +A different food world? Hang on, don’t we all occupy the same universe of choices, populated by temptations like fat-laden ready-meals and sneaky snacks under the desk at work? But McGloin says that men’s experience of eating, and all that goes with it – the shopping, the preparation, the amount they actually consume – is often markedly different to how women encounter their food. + +Men are more likely to eat larger portions, they’re less likely to be aware of healthy eating guidelines, and many don’t regard eating well as an important factor in their long-term health. + +Is it just about complacency then? Even ignorance? McGloin isn’t keen to be so judgmental. Rather, she says, it’s a matter of cultural conditioning, right from the start. + +“You have to be careful about over-generalising, but men are less likely to learn about food in childhood, so they sometimes come into adulthood without +================================================================================ +Rank = 21; Score = 354304.0 +<|begin_of_text|>Where transgenders decide to sprinkle their tinkle is NOBODY’S BUSINESS! Except it’s everyone’s business. Including that of white-knightly tech giants like Apple and Ebay. You may recall not long ago Trump reversed Obama’s decision which forced states to allow trannies into certain bathrooms. Our current President didn’t outlaw those transgender bathroom privileges, but simply restored the states’ power to discern for themselves. Therein lies the outrage. + +Apple and other companies have decided to take a stand… By getting involved in a lawsuit filed by a trans student suing his school for boys’ bathroom privileges. + +Something to note here? People were uncomfortable sharing the restroom with Gavin Grimm, the girl-boy, so the school provided a unisex restroom to him in an attempt to make everyone happy. The boys wouldn’t have to bare themselves before a lady-man, and the trans student wouldn’t have to share a restroom with girls. But it wasn’t enough for the amorphous slug. Of course. + +Now he’s demanding access to the boys’ rooms via lawsuit. Entitled little toad. This is the case that Apple and friends are jumping in to support… + +A number of leading tech firms plan to file a brief in favor of transgender rights in a case due to be heard next month in the Supreme Court. Apple has been among those leading the charge on the effort, along with… Affirm, Box, Ebay, GitHub, IBM, Microsoft, PayPal, Salesforce, Slack, Tumblr, Yelp. Apple has already spoken out on the Trump Administration’s move to pull back on Obama-era guidance that interpreted existing law to require schools to let students use the restroom that matches their gender identity. The Supreme Court case, involving Virginia high school student Gavin Grimm, asks the court to weigh in on the same question. + +These companies really, really care about rights. Just not the rights of the states. Or those who want to potty in peace. It seems as though Apple misses not a single opportunity to get hyper-political (see Apple Goes Full PC, Replaces ‘Offensive’ Gun Emoji With Squirt Gun… and Far-Left Apple CEO to Wage War on Fake News. Misses HUGE Irony…). We’re finding that tech companies are becoming more and more politically involved. Why stick with creating better tech products when you can wade into potty politics? Maybe there’s an app for that. + +Look, this is America, where businesses can say and do what they please. The problem here? These companies are alienating a large portion of their +================================================================================ +Rank = 22; Score = 352256.0 +<|begin_of_text|>Tides roll in, tides roll out. It’s not just the ocean. Elections can work the same way. + +Four years ago, Republicans had a smashing campaign season, seizing control of the House and piling up victories in Senate and gubernatorial races from sea to shining sea. Some of those swept into office were buoyed by the tea party-fueled wave. + +Four years later, many governors now have to defend their seats under less favorable conditions. + +That, succinctly, explains the dynamic in this year’s gubernatorial contests, which is more broadly discussed in this article focused on Republican Gov. Sam Brownback’s struggle in deeply red Kansas. + +Of the 36 governors races on the ballot this November, 22 are for seats held by Republicans and 14 by Democrats, numbers that work in the Democrats’ favor in contrast to their struggle in House and Senate races. + +Here are some gubernatorial contests to watch, based on interviews with nonpartisan analysts Jennifer E. Duffy of the Cook Political Report and Nathan Gonzales of the Rothenberg Political Report, as well as campaign strategists for the two major parties. + +REPUBLICAN SEATS THAT DEMOCRATS COULD FLIP: + +Florida: Gov. Rick Scott barely won election in 2010 and voters haven't seemed to warm to the former healthcare executive. Perhaps the best thing Scott has going for him is his opponent, Republican-turned-independent-turned-Democrat Charlie Crist, whose partisan perambulations left him with plenty of critics on all sides. + +Kansas: Brownback has pursued an aggressively conservative agenda, highlighted by a record cut in the state’s income tax, which has sent revenues plunging and many moderate Republicans running into the arms of his Democratic opponent, House Minority Leader Paul Davis. Brownback is trying to tie Davis to President Obama, who is overwhelmingly unpopular in the state despite his family roots. (Obama’s mother was a Kansan.) + +Georgia: Gov. Nathan Deal has been dogged by an ethics scandal and his bungled response to a freak January snowstorm, among other missteps. The Democratic candidate is Jason Carter, the grandson of former President Jimmy Carter, who has been a strong fundraiser and may also benefit from heavy Democratic spending on behalf of U.S. Senate candidate Michelle Nunn, daughter of former Sen. Sam Nunn, who offers the party a rare pickup opportunity against businessman David Perdue. + +Maine: Gov. Paul LePage barely won election in 2010 in a three-way race and has alienated a lot of people since, +================================================================================ +Rank = 23; Score = 344064.0 +<|begin_of_text|>‘Women hate women. That’s what I think,’ says pop star crowned Billboard’s woman of the year – and once photographed in Trump’s bed + +Madonna has described her feelings of “heartbreak” and “betrayal” in the wake of Donald Trump’s US election win, claiming: “It feels like women betrayed us.” + +During her interview with Billboard, which has crowned her Woman of the Year, the pop star compared her despair regarding the results – particularly the significant number of women who voted for Trump – to mourning. + +Why did women vote for Trump? Because misogyny is not a male-only attribute Read more + +“I feel that way every morning; I wake up and say, ‘Oh, wait, Donald Trump is still the president,’ and it wasn’t a bad dream that I had. It feels like women betrayed us. The percentage of women who voted for Trump was insanely high,” she says. + +When asked why she thought so many women – according to CNN 53% of white women voted for Trump – Madonna replied: + +Women hate women. That’s what I think it is. Women’s nature is not to support other women. It’s really sad. Men protect each other, and women protect their men and children. Women turn inward and men are more external. A lot of it has do with jealousy and some sort of tribal inability to accept that one of their kind could lead a nation. Other people just didn’t bother to vote because they didn’t like either candidate, or they didn’t think Trump had a chance in the world. They took their hands off the wheel and then the car crashed. + +Earlier this week, Madonna criticised Trump and said she was ashamed to be an American during a performance and auction that raised more than $7.5m (£5.9m) for her Malawi foundation. + +Images of the president-elect appeared behind her as she sang the lyrics to Britney Spears’s Toxic. + +The singer also revealed that she had previously been photographed in Trump’s bed – for a magazine shoot when he was not at home. She made jokes about his “cheap sheets”, saying: “They won’t be Egyptian cotton because we all know how he feels about Muslims, don’t we?”<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 24; Score = 344064.0 +<|begin_of_text|>Please enable Javascript to watch this video + +Jocks are typically portrayed as dumb. Movies and television shows have shown us that they detest school, pouring every last bit of their precious little brain capacity into jumpshots and touchdowns. There are plenty of jokes to be made about athletes who take seven years to get their bachelor’s degrees. + +“Hey, a lot of people go to college for seven years,” says Chris Farley’s character in the ‘90s hit comedy Tommy Boy. + +“I know,” replies David Spade’s Richard. “They’re called doctors.” + +Charlotte Hornets forward Marvin Williams is not a doctor. He’s a jock, and he always has been. In 2004, the 18-year-old was a five-star high school recruit ranked 11th in the country. He was named to Washington’s All-State team and found himself on several All-American rosters, including the most prestigious one organized by McDonald’s. Naturally, he found his way to UNC, where in his first year he teamed up with studs like Raymond Felton, Rashad McCants and Sean May en route to the 2005 NCAA National Championship. + +That summer, after only one season playing college ball, Williams was made the No. 2 overall selection in the 2005 NBA Draft. Being drafted that high meant that the 19-year-old would bank around $10.5 million by the time the rest of his fellow UNC freshmen graduated college three years later. Most people that age would take eight figures’ worth of salary and kiss college goodbye forever. + +Williams did not. He continued to pursue his degree. + +“It took me 10 summers,” Williams told Basketball Insiders. “I only had one year of academics under my belt when I left Chapel Hill [in 2005], so it literally took me 10 summers to finish.” + +In other words, Williams started his summer courses in 2005 after declaring for the NBA Draft, was then the second overall selection in late June, and went ahead and finished up his coursework before beginning his rookie campaign in Atlanta. + +“I didn’t waste any time,” he chuckled, remembering that hectic time in his life. “I know myself well enough to know that if I would have stepped away for just even a semester, it would have been more difficult to go back. So I just hung straight in there and kept going.” + +For the next nine years after that first one, he took zero breaks. Not for vacation, not for rest. Every summer he’d go back to Chapel Hill, knock out a +================================================================================ +Rank = 25; Score = 339968.0 +<|begin_of_text|>Doctors in India have saved the life of a young boy who had two extra arms and two extra legs growing out of his stomach. + +Related news Chang Du from Henan Man has huge growth on his face, & he's CHINese! Chang Du has a huge growth on his face – and he's CHINese! But the disfigured 47-year-old man from Henan province can't afford medical care. + +Zhang Ruifang, China Old woman (101) grows devil-like horns! Old woman Zhang Ruifang (101), a mother-of-seven from China, has brown devil-like horns growing out of her forehead - one is already six cm long! + +Some people had venerated Deepak Kumar (8) from Bangalore as a god, while others saw him as a demonic monster. + +But doctors, on the other hand, saw the boy with the eight limbs as parasitic twins – a form of twin similar to conjoined siblings, but where the development process has gone even more wrong. + +Deepak’s father Viresh Paswan, a construction worker in his 30s, would not give up the dream that his son could lead a normal wife, launching a public appeal in March. + +And now it’s finally over. + +Deepak’s extra limbs were surgically removed in a hospital. + +“He is 100 per cent fit,” chief surgeon Ramcharan Thiagrajan said following the operation. He added: “Due to all the mockery and stigma he has faced he is very restless and nervous. But now after this successful operation and counselling he will lead a normal life.” + +In turn, Deepak’s father was overjoyed: “My dream has come true, now we will celebrate it after returning to my village.” + +Related news + +Chang Du has a huge growth on his face – and he's CHINese! But the disfigured 47-year-old man from Henan province can't afford medical care.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 26; Score = 337920.0 +<|begin_of_text|>Give tropical forests back to the people who live in them – and the trees will soak up your carbon for you. Above all, keep the forests out of the hands of government. So concludes a study that has tracked the fate of 80 forests worldwide over 15 years. + +Most tropical forests – from Himalayan hill forests to the Madagascan jungle – are controlled by local and national governments. Forest communities own and manage little more than a tenth. They have a reputation for trashing their trees – cutting them for timber or burning them to clear land for farming. In reality the opposite is true, according to Ashwini Chhatre of the University of Illinois at Urbana-Champaign. + +Hand it over + +In the first study of its kind, Chhatre and Arun Agrawal of the University of Michigan in Ann Arbor compared forest ownership with data on carbon sequestration, which is estimated from the size and number of trees in a forest. Hectare-for-hectare, they found that tropical forest under local management stored more carbon than government-owned forests. There are exceptions, says Chhatre, “but our findings show that we can increase carbon sequestration simply by transferring ownership of forests from governments to communities”. + +One reason may be that locals protect forests best if they own them, because they have a long-term interest in ensuring the forests’ survival. While governments, whatever their intentions, usually license destructive logging, or preside over a free-for-all in which everyone grabs what they can because nobody believes the forest will last. + +Advertisement + +The authors suggest that locals would also make a better job of managing common pastures, coastal fisheries and water supplies. They argue that their findings contradict a long-standing environmental idea, called the “tragedy of the commons”, which says that natural resources left to communal control get trashed. In fact, says Agrawal, “communities are perfectly capable of managing their resources sustainably”. + +Flawed plans + +The research calls into question UN plans to pay governments to protect forests. The climate change meeting in Copenhagen in December is likely to agree on a formula for a programme called Reduced Emissions from Deforestation and Forest Degradation. “There is a real fear that REDD will lead to dispossession of local communities [as] governments stake their claim on emissions reduction credits,” says Chhatre. + +Simon Counsell of the Rainforest Foundation UK is not surprised by the findings. “In Brazil and elsewhere, we know the most enduring forests are in indigenous reserves, +================================================================================ +Rank = 27; Score = 337920.0 +<|begin_of_text|>Women who kill their newborns usually claim to have been in denial about their pregnancies. Can you carry a child to term without realizing it? And if you do, how responsible are you for your actions? + +(Photo: Phil Jones/Shutterstock) + +One summer morning in 2006, Jean-Louis Courjault found two dead babies in his freezer. The 40-year-old was trying to find space for some mackerel when he opened a drawer and saw something unusual wrapped in a towel. “I open it,” he later recounted in court. “I see a hand.” In the next drawer there was a second bundle: “I open it and I see that it is also a baby.” + +ADVERTISEMENT Thanks for watching! Visit Website + +Courjault retreated to the living room in shock. He and his family had moved from France to Seoul, the South Korean capital, four years earlier. As an engineer, Courjault was often away for months on business while his wife, Veronique, 38, looked after their two sons and volunteered at a local kindergarten. That summer, Veronique had taken the boys to France, and Jean-Louis was left alone. “I have to call the police but I do not know their number,” he later remembered. “I can barely say my address, I do not know how to say baby, I do not know how to say freezer.” Jean-Louis phoned a Korean colleague who contacted the police. + +The case was shocking, but what followed made it a national sensation in France. DNA tests found that the dead infants belonged to the Courjaults. Jean-Louis and Veronique held hands during a press conference and denied all responsibility, but after further tests confirmed the results the couple was arrested, and Veronique confessed. She had concealed both pregnancies from her husband. Each time she had labored alone in a bathroom and then suffocated the child before hiding it in the freezer. A third baby had been born, she added, while the couple was still in France in 1999—Veronique burned the body in the fireplace. The “affaire des bébés congelés,” or “freezer babies affair,” gripped France for months. + +ADVERTISEMENT Thanks for watching! Visit Website + +ADVERTISEMENT Thanks for watching! Visit Website + +VERONIQUE COURJAULT SEEMED TO fit the image of an ideal mother: a caring Frenchwoman in a stable and affluent marriage with no demanding career to distract from her family. Members of the press were +================================================================================ +Rank = 28; Score = 337920.0 +<|begin_of_text|>This list has been expanded into the new book, "Wired to Create: Unravelling the Mysteries of the Creative Mind," by Carolyn Gregoire and Scott Barry Kaufman. + +Creativity works in mysterious and often paradoxical ways. Creative thinking is a stable, defining characteristic in some personalities, but it may also change based on situation and context. Inspiration and ideas often arise seemingly out of nowhere and then fail to show up when we most need them, and creative thinking requires complex cognition yet is completely distinct from the thinking process. + +Neuroscience paints a complicated picture of creativity. As scientists now understand it, creativity is far more complex than the right-left brain distinction would have us think (the theory being that left brain = rational and analytical, right brain = creative and emotional). In fact, creativity is thought to involve a number of cognitive processes, neural pathways and emotions, and we still don't have the full picture of how the imaginative mind works. + +And psychologically speaking, creative personality types are difficult to pin down, largely because they're complex, paradoxical and tend to avoid habit or routine. And it's not just a stereotype of the "tortured artist" -- artists really may be more complicated people. Research has suggested that creativity involves the coming together of a multitude of traits, behaviors and social influences in a single person. + +"It's actually hard for creative people to know themselves because the creative self is more complex than the non-creative self," Scott Barry Kaufman, a psychologist at New York University who has spent years researching creativity, told The Huffington Post. "The things that stand out the most are the paradoxes of the creative self... Imaginative people have messier minds." + +While there's no "typical" creative type, there are some tell-tale characteristics and behaviors of highly creative people. Here are 18 things they do differently. + +They daydream. + +Creative types know, despite what their third-grade teachers may have said, that daydreaming is anything but a waste of time. + +According to Kaufman and psychologist Rebecca L. McMillan, who co-authored a paper titled "Ode To Positive Constructive Daydreaming," mind-wandering can aid in the process of "creative incubation." And of course, many of us know from experience that our best ideas come seemingly out of the blue when our minds are elsewhere. + +Although daydreaming may seem mindless, a 2012 study suggested it could actually involve a highly engaged brain state -- daydreaming can lead to sudden connections and insights +================================================================================ +Rank = 29; Score = 331776.0 +<|begin_of_text|>Church leaders often worry that Sunday morning is the “most segregated day of the week.” + +On Sundays, churchgoers gather inside congregations that are remarkably monochromatic. Whites with whites, blacks with blacks, Latinos with Latinos, Koreans with Koreans, and so on. + +This phenomenon, however, is more than discomfort with diversity. It is also a search for safety. In the historic black church, for example, worshippers can assert the dignity and worth that a white society denies them. For three hours on Sunday, the need to avoid offending whites doesn’t govern their lives. + +As we are learning in Ferguson, Mo., African-Americans feel unsafe — far more than many whites have realized. Young black men, for example, flinch whenever a police car passes — a vulnerability that money, job and education can’t overcome. + +Recent violence against women revealed a similar safety problem. Many women flinch whenever a man draws near and feel demeaned on a daily basis by sexist behaviors at work. Abuse stories are common. + +People need safe places, and church has provided that safety to many. In church, a young black male can play drums with the choir, serve as an usher, hang out with friends and feel loved and accepted without any need to avoid eye contact with whites. + +Years ago, after a horrible time in a congregation, a friend took me to a black church. He wanted me to rediscover safety inside church doors. There, among the marginalized, I was treated with dignity and respect and felt safe inside a church for the first time in years. + +I have wondered why female clergy tend to hire female staff, to select female leaders and to emphasize female needs. Was this the “sisterhood” taking over? Turning the tables on patriarchy? + +No, I think it’s part of providing a place where women can feel safe. Not just in charge, but actually safe: free to be themselves, not needing to please men or to fear men, free to imagine God as more than patriarch. + +A similar search for safety has led gays to seek out gay-affirming congregations where they can be fully themselves. Some churches are entirely oriented to gay constituents. + +Young church folks tend to avoid congregations led by the elderly, because they want to avoid the glares and heavy-handed control battles common in older churches. + +Because safety matters so much, established congregations have an even larger dilemma than grappling with financial stress and declining numbers. They need to be safe places first. + +In too many congregations, people fight over trivialities. They bully +================================================================================ +Rank = 30; Score = 329728.0 +<|begin_of_text|>Giriraj Singh’s remarks about “white-skinned” Sonia Gandhi shocked a lot of people this week. Tasteless though the comment was, it did bring out a prevalent Indian attitude about the colour of one’s skin. Indeed, one thing that baffles me about India is our love for fair skin. Not unlike many African and Asian countries, Indians too obsess about having a lighter skin colour, consider people with lighter skin as more beautiful and, in the worst case, even ascribe a higher status to them. From the multi-million-dollar fairness products industry to the fact that all children in our baby product ads are fair, one no longer needs to debate that Indians love lighter skin. We do, and that might have just been one of those many Indian quirks had it not had harmful effects on our society. Indians are not fair-skinned on an average, and thus millions have a complex about their skin colour. + +Women, in particular, bear the worst of it. I remember my darker-toned cousins being told to not drink tea while growing up (because tea makes you black, as per Indian home science, never mind the English never turned dark despite fighting wars for tea), or to study harder because they were dark and hence it would not be easy to marry them off. Apart from judgments about looks, there is something more onerous when it comes to skin colour and Indians. Dark-skinned people get fewer opportunities in India. This is not racism, but is sometimes referred to as colourism or pigmentocracy. Fair-skinned people are more likely to be hired in certain jobs (when was the last time you saw a dark-skinned flight attendant on an Indian airline?). + +We must address this. Not merely in lip service or by raising your voice with catchy slogans for Facebook, but by honestly seeing what is the root cause driving all this. + +The key reason why this happens is the associations related to colours, and also to people of a particular colour. + +First, is it just the intrinsic colours? White. Black. What do these colours signify to you? White is often used in the context of clarity, purity and cleanliness. Black is used to describe the opposite. White is clean, black is dirt. Is some of this carried through when we refer to the same colours in the context of our skin? + +Of course, not everything black is seen as unaesthetic. Thick and shiny black hair is associated with youth, health and beauty. Images of dark brown chocolate, associated with deliciousness, can make one’s mouth water +================================================================================ +Rank = 31; Score = 325632.0 +<|begin_of_text|>Robots, too, can be bad drivers. As the world prepares for the coming future of driverless cars, there are bound to be a few accidents. Launched this summer, a trial of PostBus driverless shuttles in Sion, Switzerland was expected to continue through October 2017. Instead, one of the two shuttles hit a parked van earlier this week, sending the whole trial into a screeching halt. + +Test operation momentarily stopped in Sion. Because an autonomous shuttle touched the open tailgate of a parked van in Sion on 21 September 2016, PostBus has decided to interrupt the test operation with the Smart Shuttle for the time being. Nobody got hurt. There was a minor property damage. PostBus and the vehicle manufacturers are continuing to analyse the causes of the incident and take the necessary measures. + +The problem echoes that of previous driverless car accidents, where sensors were blind to an unusual circumstance. In February, a Google driverless car was ruled at fault in a crash, as it changed into a lane prematurely and collided with a bus that did not see it signaling. When a Tesla car on autopilot crashed and killed its owner in the United States, it sparked a Federal investigation. The Tesla crash, where the car collided with a trailer crossing the road, was likely because the car’s sensors couldn’t tell the white side of a trailer apart from the blank space of the sky, and failed to stop in time. + +So it is with the Swiss shuttle. The slow-moving vehicles (traveling no faster than 12 mph) can carry up to 11 passengers and are monitored remotely. The crash, as it was, appears to be minor, the kind of error that would be wholly unsurprising with a student driver, yet stands out thanks to the technology involved.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 32; Score = 325632.0 +<|begin_of_text|>Tech gadgets' plastic and metal bodies can be pretty tiresome, offering little personality and warmth - especially considering how integrated they are with our lives. Recently, wood has stepped in to fill the unnatural void. One of the oldest known materials used to build things, wood is warm, timeless and beautiful. Its durability and unique fingerprint of grain patterns also make this an ideal material to enclose a tech gadget in. + +Above is an image of the Maple Phone, designed by Hyun Jin Yoon and Eun Hak Lee, who have won the Silver award at the International Design Excellence Awards 2008. + +Other wooden phones have been designed, that are interesting to compare. I think the Maple Phone is the most elegant, though. + +Continue for more images of the Maple Phone and other unique wooden phones ><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 33; Score = 321536.0 +<|begin_of_text|>Women are so fed up with male entitlement that they’ve not only begun coining terms to classify it—“manspreading” or “mansplaining,” for example—but they’ve also started devising social experiments to find out what happens when women stop being polite and start acting like, well, men. + +In an unscientific (but revealing) experiment launched on Friday, Tumblr user Claire Boniface announced that whenever she received a compliment from a male stranger online, she’d agree with him rather than simply ignore the comment or accept it with a gracious “thank you.” + +Like many other girls and women online, Boniface was sick of getting criticized by strangers who complimented her photos and then expected a conversation in return. + +“If a guy messages me I usually don’t reply because most of the time they are complete strangers to me,” 18-year-old Gweneth Bateman, a British teenager who undertook the Tumblr-based social experiment, told BuzzFeed News. “When they don’t get a reply out of me it usually ends up with them calling me ‘rude’ or a ‘bitch.’ ” + +So Bateman flipped the table, responding to texts about her “gorgeous eyes” with affirmations such as “Thank you, I know.” As predicted, the men on the other side of the screen failed the experiment miserably, berating her for her confidence and then retracting the compliment altogether. Bateman cites it as an example of the way men are uncomfortable “when women own their own awesomeness.” + +In part two of the experiment, posted to her Twitter account Tuesday, Bateman said she received an equally hateful reaction from men even when she politely agreed with them. The experiment has called attention to the way women’s behavior is judged by different standards from men’s. “If you reject a compliment you’re ‘attention seeking’ but if you accept it you’re ‘vain’ so [you] really can’t win either way,” she tweeted Monday. + +Bateman and Boniface weren’t the only women who recently went to great lengths to give men a taste of their own privilege. In November of last year, New York–based labor organizer Beth Breslaw decided that rather than intuitively moving out of the way for people in her path, she’d take a more masculine approach and continue walking, with no regard for other pedestrians. + +She got the idea from a friend who initiated the experiment after hearing that men were less likely than women to make room for others on a crowded street. The experiment was not only startling +================================================================================ +Rank = 34; Score = 317440.0 +<|begin_of_text|># Don’t blow out candles on your birthday. It’s “western culture” and needs to be shunned. Instead, wear “swadeshi clothes” this day, do a havan, pray to the ishtadev, feed cows. + +Advertising + +# Drawing a map of India? Make sure you include Pakistan, Afghanistan, Nepal, Bhutan, Tibet, Bangladesh, Sri Lanka and Myanmar. These are part of undivided India or “Akhand Bharat”. + +# Include August 14, Pakistan’s Independence Day, in the list of national holidays. This day should be celebrated as “Akhand Bharat Smriti Divas”. + +These are moral prescriptions from books authored by Dina Nath Batra, convenor of Shiksha Bachao Andolan Samiti, that have now become compulsory reading in government schools in Gujarat. On June 30, the state government issued a circular directing more than 42,000 primary and secondary government schools across the state to make a set of nine books by Batra, translated from Hindi to Gujarati, part of the curriculum’s “supplementary literature”. + +Advertising + +Batra’s civil suit earlier this year had led to the pulping of American scholar Wendy Doniger’s book on Hinduism. He had also sent a legal notice to another publisher about a book on modern Indian history which led the publisher to begin a review of some of its books, including one on + +sexual violence during riots in Ahmedabad. + +The circular issued by the Gujarat State School Textbook Board (GSSTB) read, “These books on supplementary literature are aimed at imparting quality education. They will be provided free of cost to all government primary and secondary schools, public libraries and will be also available at GSSTB, Gandhinagar, for individuals interested in these books. These are to be incorporated from this academic session.” + +On March 4, Gujarat Education Minister Bhupendrasinh Chudasama released the set of nine books — Shikhan nu Bhartiyakaran, Tejomay Bharat, Prernadeep I, II, III and IV, Vidyalaya: Pravitiyon nu Ghar, Shikhsan ma Triveni and Vedic Ganit. + +On page 59 of the book Shikhan nu Bhartiyakaran (Indianisation of Education), under the chapter Samajik Chetna (social awakening), the author says, “Birthdays should be celebrated by shunning the western culture of blowing candles. Instead, we +================================================================================ +Rank = 35; Score = 315392.0 +<|begin_of_text|>I’m a female scientist, and I agree with Tim Hunt. + +Allie Rubin Blocked Unblock Follow Following Jun 13, 2015 + +Nobel Prize winner Sir Tim Hunt recently caused a stir when he argued that male and female scientists should work independently, because women are a distraction to men in the lab. At a conference, he was quoted as saying, “Let me tell you about my trouble with girls… three things happen when they are in the lab… You fall in love with them, they fall in love with you and when you criticize them, they cry.” + +As improbable as it might sound, I am a female scientist, and I agree with Tim Hunt. + +First of all, Sir Tim’s comments are based on his personal experiences, and are therefore incontrovertible. Three hundred and fifty years ago, Isaac Newton saw an apple fall and decided that gravity existed. Three weeks ago, Tim Hunt saw a woman cry and decided that all women are unfit to be scientists. Science is based on observations, which are the same thing as universal proof. Even I know that, and I’m just a woman whose brain is filled to capacity with yoga poses and recipes for gluten-free organic soap. Once, I was lured into a trap in the woods because I followed a trail of Sex and the City DVDs for three miles into a covered pit. Do you really think I could do something as complicated as thinking about science? + +Second, Tim Hunt is correct in asserting that women are distracting to men in a lab setting, which greatly affects their productivity. Last month, my labmates and I had to burn our female coworker at the stake for witchcraft because we saw her holding an unwrapped tampon. Between building the stake, lighting an adequate fire, and waiting for her to die, we lost an entire day that we could otherwise have spent working. I make every effort to not distract my male colleagues; sometimes I have to work on my astrology charts from home when I’m menstruating or have leprosy, just so I can let the men in my lab focus on doing real science. It’s a small sacrifice I make that keeps morale in the lab way up. + +Tim Hunt also mentions that men and women are constantly falling in love with each other within the lab. This is true; I used to identify as homosexual before working with a group of bearded men and a set of phallic pipettes turned me straight. Once that happened, I couldn’t stop falling in love with every man I met. One time I fell +================================================================================ +Rank = 36; Score = 311296.0 +<|begin_of_text|>Lonely individuals may decode social cues well but have difficulty putting such skills to use precisely when they need them--in social situations. In four studies, we examined whether lonely people choke under social pressure by asking participants to complete social sensitivity tasks framed as diagnostic of social skills or nonsocial skills. Across studies, lonely participants performed worse than nonlonely participants on social sensitivity tasks framed as tests of social aptitude, but they performed just as well or better than the nonlonely when the same tasks were framed as tests of academic aptitude. Mediational analyses in Study 3 and misattribution effects in Study 4 indicate that anxiety plays an important role in this choking effect. This research suggests that lonely individuals may not need to acquire social skills to escape loneliness; instead, they must learn to cope with performance anxiety in interpersonal interactions. + +© 2015 by the Society for Personality and Social Psychology, Inc.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 37; Score = 309248.0 +<|begin_of_text|>Teaching a trick to a chicken increases beliefs that chickens are intelligent and can feel emotions. + +Learning how to train chickens changes student’s attitudes towards them, according to a new study by Susan Hazel Lisel O’Dwyer (both University of Adelaide) and Terry Ryan (Legacy Canine). The chickens were trained to do a specific task (such as pecking on a red but not green circle) in order to get food. Survey responses before and after the class show more positive attitudes after the clicker-training session. + +Lead author Susan Hazel told me in an email, “I believe that the main reason for the students’ change in attitudes to chickens was that they realized chickens are smarter than they thought (they learn the colour discrimination tasks very fast) and also when you work with the different chickens you see their personalities.” + +“Some chickens are fast and other chickens still learn quickly but just respond more slowly,” said Dr. Hazel. “It wasn’t so much of a surprise that students were more likely to believe that chickens were intelligent, are easy to teach tricks to, and that they have individual personalities. It was more surprising that this carried over to students more likely to believe chickens could experience boredom, happiness, and frustration.” + +Some of the differences are quite striking. Before the class, only 7% of students thought it would be easy to teach tricks to a chicken, but after the class this went up to 61%. Beforehand, 49% thought chickens are intelligent, but afterwards 77% agreed. Most of the students thought chickens had individual personalities before the class (84%), but this went up to 95% after. There were some gender differences, including that women gave higher ratings than men for chicken intelligence. + +Each year the chickens are named according to a theme. This time it was members of the Royal family, and the brightest chicken was one called Margaret. Students had a clicker attached to the handle of a scoop containing chicken food. When the chicken got something right, the trainer pressed the clicker (making a sound to signal to the chicken they did the right thing) and then let the chicken eat from the scoop. + +The practical class lasted for 2 hours and included time training chickens and other activities. Chickens were chosen because they are easy to care for and handle in the class – and they are unforgiving of their trainer’s mistakes. The class used a technique called shaping which involves rewarding closer and closer approximations of a task to reach the final behaviour. Students worked in pairs to practice shaping on +================================================================================ +Rank = 38; Score = 309248.0 +<|begin_of_text|>Stand to the right. Stand to the left. Slap an arm now, ya’ll. + +Thus goes the “James Harden Shuffle,” a defensive dance featuring minimal movement and maximal ball watching. + +If you’re unfamiliar with the Harden Shuffle, YouTube user How U (h/t John Ferensen of NextImpulseSports.com) has created a lengthy instructional video breaking down the ins and outs of the Houston Rockets star while he slow waltzes about the hardwood. + +Facing no shortage of material, How U managed to compile more than 11 minutes of Harden playing the kind of defense that would make a matador blush. + +It’s a dance for all occasions, featuring moves such as the “Wraparound Slap-Around.” + +Screenshots via YouTube + +If his man moves outside the paint, Harden lets him go. He’s not about to make the hike out to the perimeter. That’s for suckers. His guy doesn’t have the ball, and you can’t score without the ball. That’s just basketball. + +Unfortunately, Harden’s dance moves might be holding him back. + +Indeed, an elite player with all the offensive skills in the world can still be his team’s worst enemy, especially when things get physical in the playoffs. + +The Rockets face off against the Portland Trail Blazers in Game 1 of the Western Conference Quarterfinals on Sunday. Houston won three of four of their games against Portland this season but could be in for trouble if Harden brings his dancing shoes. + +Keep an eye out for Harden’s signature wrap-around and his equally futile “Stand and Stare” technique. His opponents know his moves and will be looking to take full advantage in this critical stage. + +Don’t be surprised if Harden throws out a shoulder trying to wrap up Damian Lillard. + +Tweet me with your favorite Harden moves. + +Follow @Dr_Carson<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 39; Score = 309248.0 +<|begin_of_text|>Criticize cops involved in the G20 debacle? We can now expect to be arrested and face criminal prosecution. + +Forget a civil defamation suit—two members of the Ontario Provincial Police, aided and abetted by the Crown Attorney’s office, have chosen the nuclear option, charging a Kitchener activist with criminal defamation. + +(More here. Needless to say, this latest assault on the right to dissent has not received much coverage in the corporate media.) + +From the Waterloo Record: + +Dan Kellar, 29, was recently charged with two counts of defamatory libel by officers in the OPP anti-rackets squad as he left his Kitchener home on a bicycle. He was also charged with counsel to assault one of the officers. Police allege he published comments likely to injure the reputation of the officers by exposing them to hatred, contempt or ridicule, or that were designed to insult the officers. …Kellar says the officers who arrested him are from the same unit that arrested other AW@L members and activists in connection with G20-related allegations. “The cop who arrested me is the one who’s making all the arrests for conspiracy cases,” he said. He said police agents began infiltrating activist groups before the summit. The two undercover officers joined AW@L [Anti-War at Laurier, a campus peace group], but were kicked out in the spring of 2010, before the summit, because activists didn’t feel comfortable around them, he said. …The defamatory libel charges were laid against Kellar after he put out a “community alert” on AW@L’s website, peaceculture.org. Kellar learned one of the officers had been spotted in Toronto, and, “sent out the warning…’suspected infiltrator police agent spotted in Toronto,’” he said. In the posting, he made comments police allege are defamatory. + +Here’s the Criminal Code provision, rarely invoked in this country—until now, I guess: + +298 (1) A defamatory libel is matter published, without lawful justification or excuse, that is likely to injure the reputation of any person by exposing him to hatred, contempt or ridicule, or that is designed to insult the person of or concerning whom it is published. Mode of expression (2) A defamatory libel may be expressed directly or by insinuation or irony (a) in words legibly marked on any substance; or (b) by any object signifying a defamatory libel otherwise than by words. + +Note +================================================================================ +Rank = 40; Score = 309248.0 +<|begin_of_text|>I’ll live where I want when I retire. I’ll get to it later. I’ll just work in the corporate world for ten years, and then I will create a life I enjoy. I’ll do it tomorrow. + +How many times have we promised ourselves tomorrow only to find it never arrives? In today’s world, we often put aside our dreams for comfort, our passions for security, and we let our fears take away our power. We choose unfulfilled lives for freedom, thinking that we can always do it differently tomorrow. + +People don’t follow their dreams because they are afraid. They are afraid of losing security. They are afraid if they do what they want, they will die. They have become comfortably numb. + +What should scare them is that life is short. The reality is if we don’t follow our purpose, our passions, our joys, that’s when the death begins. And the soul’s death is one hundred times worse than the body’s. + +The problem is we aren’t guaranteed tomorrow. We are only guaranteed this present moment. And we are telling the Universe who we are by the actions we display in only this moment. That’s it. That’s all we have. That’s who we are. + +We also miss some of the most special moments of our lives because we think we will have time for them later. We tell ourselves that we will do the things we enjoy when we retire. We will go fishing, we will travel, create art, we will do so much…..later. But, later never comes. There is always something standing in the way. + +So, I want to propose that we change this way of thinking. I want to propose that we encourage and inspire each other. I propose that we start a Twitter campaign now. Let’s take the world by storm. You think you have time, but right now, time has you. Let’s change this. + +Using the hashtag, #youthinkyouhavetime write your dreams, the thing you always wanted to do. But, don’t stop there. Write about the things in each day that bring you joy. Write about the things that you are saving until later. Write the name of your favorite beach. You know. The one that’s your screen saver that you promised yourself you would visit later. + +Let’s get inspired. Let’s find a way to do them now. + +I also want to offer a free 15-minute Skype session to anyone who wants guidance to start living their dreams now. Because time is running out. And, +================================================================================ +Rank = 41; Score = 307200.0 +<|begin_of_text|>Pakistani textbooks are mostly anti-Hindu & anti-Indian under the influence of Quranic teachings. + +Pak Textbooks are anti-Hindu… anti-Indian… + +Pakistani textbook controversy: They think something, tell different and do harrowing. + +“Any material considered ‘inflammatory’ or ‘discriminatory’ to religious minorities should be removed from the syllabus as the government should seriously take action on this matter”, say Salman Ali and Saira Ahmed in a co-written piece in Daily Times Pakistan. But all these dramatic write up and govt announcements come at the verge of extinction of Hindus from the Map of Pakistan! Actually, the textbook controversy & reformations etc. in Pakistan are nothing but an eyewash to prove Islam is not so bad. + +Upendra Bharti | HENB | New Delhi | Dec 30, 2016:: On August 11, 1947, three days before the announcement of the independence of Pakistan- separated from India as two parts East and West Pakistan, the Father of the Nation Mohammad Ali Jinnah in his speech said, “You are free; you are free to go to your temples, you are free to go to your mosques or any other place of worship in this State of Pakistan. You may belong to any religion or caste or creed — that has nothing to do with the business of the State. We are starting in the days where there is no discrimination, no distinction between one community and another, no discrimination between one caste or creed and another. We are starting with this fundamental principle: that we are all citizens, and equal citizens, of one State.” + +Ridiculous. Sowing the seeds of Islamic communalism in the minds of Muslims of Indian subcontinent and through a holy war named- ‘Direct action’, Jinnah materialized the dream of Pakistan which took the lives of millions and uprooted the innocent people from their soil numerously. + +When Pakistan was created in 1947, Hindus constituted about 16% of the population of West Pakistan+ (current Pakistan); by 2016 it is about 1.68% – the population has declined by about 90% in about 70 years. This decimation is the outcome of sustained legal and social discrimination ever since the creation of Pakistan. The situation of minority persecution in has not been stopped anyway. The overall situation of synchronized plights on Minority Hindus (mostly Dalit/ Scheduled caste and tribal categories) through physical torture, bonded labour, abduction, ransom, attack +================================================================================ +Rank = 42; Score = 305152.0 +<|begin_of_text|>Geography Is Dynamic. Let’s break this down a bit. + +Geography – The study of the physical features, atmosphere, and the human-environmental interaction of the Earth. + +Dynamic – Constantly changing. + +Geography Is Dynamic. Why? Because geography is about studying the Earth, and the Earth is constantly changing. Nothing on Earth remains in a static position. The only constant on the Earth is change. Things are always changing. The rate of which things change varies from place to place. + +Even within the field of geography, there is a subfield of geography that is specifically about the changes in physical geographic features. It is known as geomorphology. See video below to learn more. + +Geomorphology is just one example of how geography is dynamic. In fact, if you look at every subfield of geography, from tourism geography, limnology, urban geography, biogeography, environmental geography, to political geography, geomatics, and cartography. Things are always changing on the planet. All the more reason they need to be studied. And this should be the very first rule of geography. Geography Is Dynamic. It is dynamic because this earth is dynamic. + +In teaching geography in the schools, this should be the first lesson. Geography is dynamic. Learning state capital should not be the first thing anyone learns when it comes to geography. Understanding that the Earth is dynamic is the first thing that should be learned. + +Why learn geography? Because it’s a dynamic subject. Why is it dynamic? Because Earth is a dynamic place. Things are always changing. We need to constantly study those changes. + +#GeographyIsDynamic + +Advertisements<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 43; Score = 305152.0 +<|begin_of_text|>Analysis Google, Facebook, and likely other major tech firms are investigating ARM-compatible chips to drive low-power servers, which could ultimately shakeup the types of processors that sit inside data centres. + +We were moved to consider the pros and cons of moving away from trusty x86 into the problem-filled world of ARM after a Bloomberg rumor circulated on Thursday that Google was planning to use ARM chips for servers, and after we spotted a smoking gun blogpost by Facebook that indicated it hoped to evaluate ARM systems for production use. + +Companies are interested in ARM chips due their historically thrifty power consumption; right from the start, the processor architecture – born from a small team at Acorn Computers in the 1980s – was designed to be RISC (Reduced Instruction Set Computing). This means ARM cores do lots of simple operations relatively quickly; this simplicity, and the lack of legacy cruft to support, keeps the transistor count low, which means less power is required. + +This customisable architecture is wildly popular in the battery-powered gadget and embedded electronics worlds, where processing performance isn't key – anything taxing can be offloaded to dedicated silicon – so the chips can be run slower and thus consume even less power. + +Compare this to Intel's CISC (Complex instruction set computing) design, which offers a larger range of operations, has a mountain of legacy tech to support – from 16-bit real mode all the way up to 64-bit protected long mode – and generally runs at higher speeds to give punters the most bang for their bucks. All this adds up to beefier, power guzzling packages. + +Though ARM cultists portray RISC as being fundamentally better at low-power computation, academic studies have disproved this [PDF], noting instead that the main differences in power consumption come from historical implementations – ARM has spent nearly two decades living in your pocket, whereas Intel has resided in your honking big desktop, and so on. Indeed, the heart beating in today's x86 chips is RISC in design albeit wrapped in a CISC compatibility blanket. + +ARM has for a long time focused on cutting power consumption due to its home markets being mobile and non-performance-demanding devices, whereas Intel previously emphasized speed; chips powered by ARM cores are built from the ground up to sip rather than suck current. The drawback is that they beaver away at a relatively leisurely pace. + +Mobes and fondleslabs, ARM. Gaming rigs, x86. Got it. Where does Google and Facebook fit in? + +Consumer-serving web giants spend billions a year on +================================================================================ +Rank = 44; Score = 305152.0 +<|begin_of_text|>Fatherhood is wasted on dads. There has not been a moment – not a single fleeting moment of my entire poxy life – when I’ve had this much female interest. I’d always been useless with girls, but now they flock to me. I may as well be a Beatle. One of the good-looking Beatles, too. Not even Ringo. This is unprecedented. + +What’s the cause of this sudden attractiveness? Is it my charismatic new thousand-yard stare? Are these vomit stains making my eyes pop? Hardly. It’s him. + +I had no idea that people went so crazy for babies. In my pre-fatherhood days, a baby was something to endure through gritted teeth because it was about to ruin your flight. I thought that was universal. But no. Take a baby outside in a sling and well-wishers will swarm around you. I’ve even amassed a small arsenal of stock answers, such has been the barrage of genuine affection for him. A boy. Almost two months. About 10lb. He has his mum’s eyes. No, he’s not normally this quiet. Bright yellow, smells awful, thanks for asking. + +Incidentally, this sling has been incredible. Put a baby in a sling and not only will it automatically go to sleep, but you also get to use both your arms again. Last weekend, because motherhood had transformed the back of her head into a nightmarish mishmash between a dreadlock and a rat king, my wife went for a haircut. It was the longest amount of time I’d ever had to spend alone with my son without the safety net of an emergency handover, and it was making me slightly antsy. + +However, once I’d put him in a sling, I could do anything. I was able to wash up, bake a cake, wash up again and complete the really hard Assassin’s Creed side-mission where you have to rescue the prostitutes, all by the time she came back. This newfound explosion of productivity was amazing. I felt like the star of a tampon advert, ready to go skydiving or white-water rafting just because I could. In summary: hooray for slings. + +Having the baby out in public has only helped to reinforce what an unbearable stereotype of a proud dad I’m becoming. All I ever want to do is hang out with my family, preferably with my boy sleeping on me. When that’s not happening, I miss it. I miss the weight and warmth of him, +================================================================================ +Rank = 45; Score = 303104.0 +<|begin_of_text|>Republican congresswomen appear more "feminine" than their Democrat counterparts, according to a new study by UCLA psychology researchers. + +"Female politicians with stereotypically feminine facial features are more likely to be Republican than Democrat, and the correlation increases the more conservative the lawmaker's voting record," said study author Colleen M. Carpinella, a UCLA graduate student in psychology, in an article posted to the UCLA Newsroom website. + +So what makes a woman "stereotypically feminine?" + +Certainly, individuals are easily biased by outside factors not related to facial characteristics, including hair style, application of cosmetics, jewelry or clothing style, said Kerri Johnson, the study's senior author and an assistant professor of communication studies and psychology at UCLA. For this reason, the researchers used a computer program that was "immune to those influences." + +The program, called FaceGen, allows the researchers to measure more than 100 subtle dimensions including the shape of the jaw, the location of eyebrows, the placement of cheek bones, the shape of eyes and the contour of the forehead, to create a score ranging from -40 (highly male-typed) to +40 (highly female-typed). + +For the study, the researchers had FaceGen analyze the portraits of members of the 111th House of Representatives, measuring how closely the facial features of each face approached the average for their gender. + +The researchers also studied the faces of male politicians, although those findings were not as revealing, according to the authors. Republicans averaged a less masculine appearance than their Democrat counterparts -- a finding that Carpinella said may suggest that Republican men do not have to exhibit masculinity through their appearance. + +The findings for female representatives were more telling, however. Republican women rated more "extremely feminine" or more "sex-typical," according to Johnson. The connection was so strong, the press release notes, that when photos of the same politicians were shown to a batch of undergrads, the students were able to guess with very high accuracy the political affiliations of those women who had been rated most and least "sex-typical." + +When the undergraduates guessed that a politician was Republican, their judgments were 98 percent more likely to be accurate for women with the highest rankings for femininity; the accuracy of their judgments increased the more feminine the politician's face. When the undergraduates guessed that a politician was Democrat, their judgments were 58 percent less likely to be accurate for more feminine-looking women, and the accuracy of their judgments decreased the more feminine the politician's face. + +However, +================================================================================ +Rank = 46; Score = 303104.0 +<|begin_of_text|>Men angry and rejected, women sociable and mentally ill -- a current study by the MedUni Vienna demonstrates that these gender stereotypes prevail when Austrian daily newspapers report on suicide. This has far-reaching consequences. + +When it comes to suicidal behaviour, there is a clear gender paradox: the ratio of men to women who actually commit suicide is three to one, but with attempted suicides it is exactly the opposite -- three women for every one man. A study by the MedUni Vienna which has recently been published in the highly journal Sex Roles demonstrates that the cultural script that bears partial responsibility for this is also found in the reports by Austrian daily newspapers. + +These gender-specific differences are made visible by the formulation, the nature and frequency of reported suicide motives. Articles about suicide in women focus more on sociability, relationships with other people and motives that are anchored in the family environment. Psychiatric illnesses are also cited as a motive and are described in a stigmatizing manner. More complex language and cautious expressions are also the hallmarks of articles about female suicide. In contrast, the articles about male suicide use more words that relate to anger and rejection. This conservative role image that pervades Austria anyway is reinforced by this style of reporting. + +Suicide risk could be reduced by changing the style of reporting + +But that's not all. A very specific problem arises from this, which study leader Brigitte Eisenwort from the University Department of Paediatrics and Adolescent Medicine at the MedUni Vienna explains as follows: "Mental illnesses are described in a stigmatising way and are also generally under-represented, since they are barely mentioned at all in reports about suicidal men. This means that one key approach to prevention fails to register in the minds of Austrian readers. Psychiatric illnesses can be treated. The suicide risk can be reduced as a result." Journalists should therefore take care to present as correct a view as possible of suicidal tendencies and not revert to stereotypical portrayals of men and women. + +Spotlight on eleven Austrian daily newspapers + +507 articles containing the word'suicide' from eleven Austrian daily newspapers from between 1997 and 2005 were investigated. The study is one of the first investigations to look comprehensively at the subject of gender-specific patterns in the reporting of suicide. This ground-breaking study was set up under the leadership of Brigitte Eisenwort, together with Thomas Niederkrotenthaler and Benedikt Till (both from the Institute for Social Medicine at the MedUni Vienna's Centre for Public Health), as well as Barbara Hinter +================================================================================ +Rank = 47; Score = 301056.0 +<|begin_of_text|>Booze for the terraces, pink vehicles for the ladies, tough love for the scroungers, pride for the white vans, nurses to attend ailing Scots by the thousand. All is catered for within the sweepingly incoherent stance that the Labour Party has adopted in advance of the general election. + +Marx’s famous observation, that history happens, as it were, twice (first as tragedy, then as farce) is one that the party would do well to remember. Because, if this is a re-run for Labour, the precedent suggests a longer term decline. + +The last time Labour attempted to regain power from the Tories after a single term, it was led by an eccentric leftist intellectual, Michael Foot. Miliband may have the intellect and the eccentricity, but as his policy-lite posturing has demonstrated, the party remains terrified of a coherent ideological challenge to right wing Tory policy. + +This re-run of tragedy for Labour will be rendered farcical by the overwhelmingly weak content of its campaign. The party’s 1983 manifesto, notoriously dubbed the ‘longest suicide note in history’, may have been an attempt by the right of the party to clear out left wing policies such as nuclear disarmament, knowing that the election would be lost anyway. It did however present Labour as a distinct ideological alternative to Thatcherism on a number of issues. In 2015, there are no radical policies in Labour’s bid for power. The acute trauma of electoral politics in the 1980s has, in large part, meant that politics in Britain has scarcely progressed since. + +In response, Labour has become trapped by its attachment to a studied silence on major economic issues. Indirectly, this results in an ongoing expression of various forms of identity politics, the roots of which lie in the emergence of a new-left that came to emphasise minority rights and social-cultural emancipation. Though this turn was a welcome and radical development against the forces of conservatism a quarter of a century ago, today it has left a residual husk. A nominal ‘left’ in UK politics tacitly accepts an enormously unjust economic order and denies any legitimate expression of class as a foundation for politics. + +Of course, the siloing of the social and the economic is an ultimately false division. The history of Scotland, the north of England and Wales in the past quarter century is an object lesson in this. Labour have done nothing, post-Blair, to recognise the shape of these problems in the heartlands. The drastic and necessary recalibr +================================================================================ +Rank = 48; Score = 299008.0 +<|begin_of_text|>Migrant workers offend Republicans by being productive members of society. (Zenpix/Dreamstime.com) + +Migrant workers offend Republicans by being productive members of society. (Zenpix/Dreamstime.com) + +Tomato farmer Wayne Smith said he has never been able to keep a staff of American workers in his 25 years of farming. “People in Alabama are not going to do this,” said Smith, who grows about 75 acres of tomatoes in the northeast part of the state. “They’d work one day and then just wouldn’t show up again.” At his farm, field workers get $2 for every 25-pound box of tomatoes they fill. Skilled pickers can make anywhere from $200 to $300 a day, he said. + +Unskilled workers make much less. A crew of four Hispanics can earn about $150 each by picking 250-300 boxes of tomatoes in a day, said Jerry Spencer, of Grow Alabama, which purchases and sells locally owned produce. A crew of 25 Americans recently picked 200 boxes — giving them each $24 for the day. It may make sense for some to sit on the couch. Unemployment benefits provide up to $265 a week while a minimum wage job, at $7.25 an hour for 40 hours, brings in $290. + +Alabama farmers are getting killed by a new draconian law that has chased thousands of undocumented immigrants out of the state. + +This is brutal work, which is why slaves were once imported to do it, and why it's work now done by the lowest rungs of the socio-economic ladder—immigrants. Just think about the numbers above—four Latinos picked a quarter to a third more tomatoes than a crew of 25 Americans. And sure, those Americans could learn to be more efficient and skilled at the work, but they won't, because they won't put up with that kind of abuse on their bodies (and minds, for that matter) in three-digit temperatures. + +This is the same problem that Georgia farmers face, as that state cracks down on undocumented immigrants. As one Georgia farmer laments, + +“You can’t find legal workers,” Horner said. “Basically they last a day or two, literally.” + +The results have been brutal for farmers in Georgia: + +Charles Hall, director of the Georgia Fruit and Vegetable Growers Association, released figures from an upcoming industry-funded study Tuesday that says farmers lost at least $74.9 million in unpicked crops harvested by hand last spring and summer because they didn’t have enough labor. The farmers +================================================================================ +Rank = 49; Score = 299008.0 +<|begin_of_text|>Two guys walk into a bar. They order beers. Bartender says they don’t have any beer. The men look confused. A stranger in a stylish hat suggests they try something different. They order a clear malt beverage. It’s on ice, clear, delicious. The men are happy. + +The entire ad spot lasts 30 seconds, or roughly the same amount of time Zima could claim to be among the most popular adult beverages in the country. + +In 1991, with beer sales on the decline across the industry, the Coors company of Golden, Colorado decided to blend two of the hottest trends in consumer marketing: “clear” products like Crystal Pepsi and the smooth, gently-intoxicating appeal of wine coolers. By using charcoal to filter the color and taste from their brews, they were able to deliver a vaguely citrus-tasting drink with 4.7 percent alcohol content. The company asked third-party marketing firm Lexicon Branding to give it a name; Jane Espenson, who would later become a staff writer on Buffy the Vampire Slayer and Game of Thrones, dubbed it Zima, the Russian word for "winter." + +Armed with a $180 million budget for the 1994 launch, Coors peppered television with commercials featuring a spokesman who exchanged his s's for z's. (“What’s your zign?”) They also pushed a slew of merchandising and even an early consumer-use product website. + +The goal was to get Zima on the minds and into the hands of young males. Owing to the blanket advertising assault, that's exactly what they accomplished. Zima sold a staggering 1.3 million barrels’ worth of product in 1994, giving it a near-instant 1 percent market share in the booze industry. It was estimated that 70 percent of all drinkers tried the “malternative." + +As Coors would soon learn, those numbers only work in your favor if people like the product. The company was disappointed to learn that many of them didn’t: Men found the taste off-putting. And those who enjoyed it were precisely the demographic they were looking to avoid. + +Women who normally passed over beer embraced Zima, giving it an effete quality that marketing considered to be grim death for the valued male customer base. If a man couldn’t feel manly taking a pull of the clear stuff, he'd be likely to reach for something else. + +On the public relations side, Coors was also having to defend itself against charges that teenagers were growing fond +================================================================================ +Rank = 50; Score = 296960.0 +<|begin_of_text|>Fetal sex tests have a few medical applications, allowing couples with histories of rare sex-linked disorders to avoid costly and invasive genetic testing if they learn they are expecting the other sex. But for most couples, the tests, which are unregulated, simply answer the boy-or-girl question weeks earlier than ultrasound, and in a less invasive and safer way than amniocentesis. + +“Seven weeks is a different time in pregnancy,” said Dr. James Egan, a professor of obstetrics and gynecology at the University of Connecticut Health Center who was a co-author of a study on sex selection with Dr. Chapman and others. “Women haven’t had the ultrasound where you see the fetus that looks like a baby. Many people don’t even know that a woman is pregnant. And you can have a medical termination,” using pills like RU-486, which can be used at home discreetly before 10 weeks of pregnancy. + +There is evidence that some Americans want to choose their babies’ sex. At the Fertility Institutes, a set of clinics in Los Angeles, New York and Guadalajara, Mexico, 85 percent of roughly 500 couples each year seek sex selection, although three-quarters of them come from overseas, said Dr. Jeffrey Steinberg, the medical director. + +“It’s jumped over the past four years,” said Dr. Steinberg, whose clinics determine sex through pre-implantation genetic diagnosis, an embryo screening that also detects genetic disorders. He said that “if a woman calls to make the appointment, the couple almost always wants a female. If a man calls, they almost always want a male.” + +But clinics and some ethicists say this type of sex selection is more acceptable because it occurs before embryos are implanted, before pregnancy. + +“We’re trying to prevent the abortion,” said Dr. Jamie Grifo, program director for New York University’s Fertility Center. His and other clinics typically allow sex selection for couples with two or more children, parents interested in “family balancing,” adding a child of the opposite sex. + +“For someone who has two girls and wants to have a boy, so each sibling can grow up with brother and sister, what’s wrong with that?” Dr. Grifo said. + +Advertisement Continue reading the main story + +Still, the cost and commitment of the fertility process makes such sex selection cases relatively unusual. Fetal DNA tests, costing between $250 and $350, are more affordable. + +Anti-abortion groups are incorporating sex selection in legislative agendas. Arizona and Oklahoma recently passed laws banning sex-selected abortion +================================================================================ +Rank = 51; Score = 294912.0 +<|begin_of_text|>Testers need to learn to code *. Period. + +And here’s why. + +The market demands it and the supply is arriving. + +As a hiring manager I’ve personally seen a massive increase in the number of testers who can now code. They may not be able to write production grade feature code, or automated tests, but they can write scripts to help them test, or they can write a small app that will inject data or they can extend an Open Source tool to make it work for their needs. + +They can basically dig deeper than what you see on the screen, do more with automation and do a more varied set of testing by using tools and code. I see this as a positive thing – I know some people may not. + +There is also a growing number of developers who are moving in to a technical testing capacity and learning how to do good exploratory testing and test planning. The market is buoyant for testers who code (and who know how to sell themselves) and it’s good for hiring managers. It’s quite rare now for me to find a tester who isn’t coding, or at least really focused on learning it. + +In a nutshell the market is now being supplied with these once elusive testers who can code. + +You could of course replace “learn to code” with “learn to do security testing” or “learn to do performance testing” but the reality is that most of these niches also require coding or scripting experience. It’s rare to find a tool that pops out of a box and does what you want it to do – despite what many of the tool vendors sales people will tell you. + +To ignore the shift in the market is to leave yourself at a disadvantage when trying to get a job. + +Whether you believe that coding is essential to being a good tester or not will soon become an irrelevant moral viewpoint. Testers now need to remain relevant to those with jobs, and those people with jobs for offer are starting to get multi-skilled testers (or T-Shaped as I call them) applying. + +Instead of investing energy in fighting the inevitable train of change it might be worth spending that time learning to script, or carving out a niche in another aspect of testing (security, usability, off-shoring, test management etc), or resigning yourself to being outpaced in job applications. + +We should be careful though to discern from simply following trends such as certifications and following evolutions in how testing is done. One is very dangerous and leads to lazy recruiting and competitions like certification inflation – the other is a natural progression of our testing craft. +================================================================================ +Rank = 52; Score = 294912.0 +<|begin_of_text|>Australia’s Jill Rathborne may have added a new term to the already crowded field of politically correct academese: “micro-inequity.” + +Not encouraged to study science while growing up, Rathborne says she was told by her teachers that college physics classes “would be too difficult.” + +She admits, however, that none of the teachers she had asked had taken physics at college, “so that might have influenced their opinions.” + +“That conversation didn’t explicitly mention my gender, but I do wonder if they gave that same advice to male students that had similar abilities to me (I was in the top few percent of my high school physics class),” she says. + +The Guardian reports: + +The problem is these biases are often subtle. It’s these micro-inequities that do the most damage – the small, seemingly inconsequential ways in which people are excluded or discounted that, over time, erode confidence, value and worth and reinforce a culture that is not inclusive. BD: Has there been a moment in your career when you’ve been uncomfortably reminded of your gender? JR: It’s that thought in the back of your mind: am I getting this opportunity because I’m qualified, or because I’m female? Hopefully the former, but perhaps sometimes it might be the latter. I’d like to be known as a scientist, not a female scientist – maybe, if possible, a good scientist. The relevant point here is that typically we wouldn’t ever ask a man this same question when asking him about his career. BD: Have you been encouraged to modify behavior in masculine environments? JR: Astronomy is a very male-dominated field – this is evident in the office, at conferences, in committee memberships. I can’t recall ever being actively told to modify my behaviour, [but] having said that, it takes a certain resilience and personality to be comfortable and successful when you’re a minority in your field. You often need a pretty thick skin, you can’t be easily offended. + +So, even though none of her teachers made any point about her being female (and who, by the way, probably had a very good idea about her academic abilities), and though she can’t recall any “active” request to change her behavior in the male-dominated arena, it’s just “that thought in the back of [her] mind …” + +Not to mention, Rathborne bemoans a “non-inclusive culture” … yet wonders if she got her position merely because of her gender? + +Wow. + +Read the full article. + +Like The College Fix +================================================================================ +Rank = 53; Score = 294912.0 +<|begin_of_text|>More, faster, better, cheaper. These are the demands of our device-happy and data-centered world. Meeting these demands requires technologies for processing and storing information. Now, a significant obstacle to the development of next-generation device technologies appears to have been overcome, according to researchers from the University of Tokyo (Japan), Tokyo Institute of Technology (Japan) and Ho Chi Minh University of Pedagogy (Vietnam). + +Specializing in the emerging field of semiconductor spintronics, the team has become the first to report growing iron-doped ferromagnetic semiconductors working at room temperature -- a longstanding physical constraint. Doping is the practice of adding atoms of impurities to a semiconductor lattice to modify electrical structure and properties. Ferromagnetic semiconductors are valued for their potential to enhance device functionality by utilizing the spin degrees of freedom of electrons in semiconductor devices. + +"Bridging semiconductor and magnetism is desirable because it would provide new opportunities of utilizing spin degrees of freedom in semiconductor devices," explained research leader Masaaki Tanaka, Ph.D., of the Department of Electrical Engineering & Information Systems, and Center for Spintronics Research Network, University of Tokyo. "Our approach is, in fact, against the traditional views of material design for ferromagnetic semiconductors. In our work, we have made a breakthrough by growing an iron-doped semiconductor which shows ferromagnetism up to room temperature for the first time in semiconductors that have good compatibility with modern electronics. Our results open a way to realize semiconductor spintronic devices operating at room temperature." + +The researchers discuss their findings this week in Applied Physics Letters, from AIP Publishing. The researchers' maverick move challenged the prevailing theory that predicted a type of semiconductor known as "wide band gap" would be strongly ferromagnetic. Most research focuses on the wide band gap approach. "We instead chose narrow-gap semiconductors, such as indium arsenide, or gallium antimonide, as the host semiconductors," Tanaka said. This choice enabled them to obtain ferromagnetism and conserve it at room temperature by adjusting doping concentrations. + +Investigators have long envisioned bridging semiconductors and magnetism to create new opportunities of utilizing spin degrees of freedom and harnessing electron spin in semiconductors. But until now, ferromagnetic semiconductors have only worked under experimental conditions at extremely low, cold temperatures, typically lower than 200 K (-73oC), which is much colder than the freezing +================================================================================ +Rank = 54; Score = 294912.0 +<|begin_of_text|>I am hideously white, and not a man but “male”. Being over 50, I suffer the added failing of being disgustingly old. Such are the routine humiliations of my group. The BBC was called hideously white by its former boss Greg Dyke, and the West End stage hideously white by Andrew Lloyd Webber. This week the Football Association was dismissed by critics as a bunch of “old white men”. Note that it is not the BBC or the theatre that is hideous, but their whiteness. + +Blame the identity apostles – they led us down this path to populism | Simon Jenkins Read more + +Fashion in collective abuse seeks comfort in crowds. In choosing “pale, stale males” (PSMs) for ritual contempt, identity politics has found a target that it hopes will confess its “guilt”. More recently the epithet “failed” has been added to explain the Brexit/Trump voters. These PSMs are further damned by being uneducated “left-behinders”, and therefore also dumb. They were thought so dumb it was suggested they vote again on Brexit, for having got it wrong first time. + +Were someone such as I to take offence, demand redress or “protected space”, I would be bidden to shut up, get a life and not be so sensitive. I might plead, with Shylock: “Hath not a [pale, stale male] hands, organs, dimensions, senses, affections, passions? If you prick us, do we not bleed?” I might turn to Kant and universalise the judgment. What if I were to follow “hideously” with black, female, Jewish, Arab, obese, disabled or Welsh? Is the representation of black footballers in the Premier League (as opposed to the Football Association) “hideously black”? + +The reality is that PMSs are expected to take all this in their stride. They were, after all, the old hegemons. They dished out discrimination; now they should accept a taste of their own medicine. (Or will someone protest that this is patronising: why should PSMs refuse to take offence, as does everyone else?) This is, of course, fair enough – except two wrongs do not make a right. Besides, identity liberalism is surely more than the politics of tit-for-tat. It is a straining after a sort of equality of treatment. + +I have no quarrel with this week’s accusations levelled at the Football Association +================================================================================ +Rank = 55; Score = 292864.0 +<|begin_of_text|>Mass IT layoffs are often small and unnoticed. They are not on the scale of the Carrier air conditioning plant layoff; its Indianapolis facility, which currently employs 2,000 people, is moving to Mexico. + +Hertz IT employees share two things with the Carrier workers: They were also angry, and they got the news on the same day, Feb. 10. + +The Carrier layoffs arrived guillotine-like; the plant is closing, period. But IT layoffs are rarely like that. There are ambiguities and uncertainties and lifeboats for some, and so it was at Hertz. + +In an early morning conference call, Hertz's IT employees were told by the CIO the firm was expanding its outsourcing work with IBM. It wasn't known then how many would lose their jobs or ultimately be hired by IBM. + +But one month later, this much is clear: About 300 Hertz IT employees, most located in Oklahoma City, were impacted by this decision. IBM is hiring about 75 and those workers are expecting to receive offers today. The layoffs will begin this month and be completed by May 31, said Hertz. It's not yet clear if all the 225 or so employees who are not receiving job offers from IBM will be laid off. + +After the conference call, employees were stunned. The reaction was, "We're screwed," said an IT employee, one of two interviewed, who requested his name not be used. + +There was "anger, resentment," especially by employees who "sacrificed that work/life balance to keep things going here," said the employee. + +Hertz took precautions. On the day that IT employees learned that their work was shifting to IBM, employees noticed Oklahoma sheriff patrol vehicles in the building's parking lot. They believed plainclothes officers were inside the building. + +Hertz explained the security decision. "We consider the safety and security of our people whenever there are circumstances or events that could increase the risk of a disturbance or some form of workplace violence," said Bill Masterson, a Hertz spokesman. + +"Knowing that this was a difficult announcement, we had additional security on hand," said Masterson. This security was in place from Wednesday Feb. 10 through Friday, Feb. 12. + +There were two opinions about the security, said the employees. Some saw it as prudent, while others thought it a sign of distrust. There were no reported problems. + +Once the initial shock passed, Hertz IT employees had to make difficult choices. + +Employees' severance packages range from four weeks to +================================================================================ +Rank = 56; Score = 292864.0 +<|begin_of_text|>Amalie Emmy Noether[a] ( German: [ˈnøːtɐ]; 23 March 1882 – 14 April 1935) was a German mathematician who made important contributions to abstract algebra and theoretical physics.[1] She invariably used the name "Emmy Noether" in her life and publications.[a] She was described by Pavel Alexandrov, Albert Einstein, Jean Dieudonné, Hermann Weyl and Norbert Wiener as the most important woman in the history of mathematics.[2] As one of the leading mathematicians of her time, she developed the theories of rings, fields, and algebras. In physics, Noether's theorem explains the connection between symmetry and conservation laws.[4] + +Noether was born to a Jewish family in the Franconian town of Erlangen; her father was a mathematician, Max Noether. She originally planned to teach French and English after passing the required examinations, but instead studied mathematics at the University of Erlangen, where her father lectured. After completing her dissertation in 1907 under the supervision of Paul Gordan, she worked at the Mathematical Institute of Erlangen without pay for seven years. At the time, women were largely excluded from academic positions. In 1915, she was invited by David Hilbert and Felix Klein to join the mathematics department at the University of Göttingen, a world-renowned center of mathematical research. The philosophical faculty objected, however, and she spent four years lecturing under Hilbert's name. Her habilitation was approved in 1919, allowing her to obtain the rank of Privatdozent. + +Noether remained a leading member of the Göttingen mathematics department until 1933; her students were sometimes called the "Noether boys". In 1924, Dutch mathematician B.L. van der Waerden joined her circle and soon became the leading expositor of Noether's ideas: Her work was the foundation for the second volume of his influential 1931 textbook, Moderne Algebra. By the time of her plenary address at the 1932 International Congress of Mathematicians in Zürich, her algebraic acumen was recognized around the world. The following year, Germany's Nazi government dismissed Jews from university positions, and Noether moved to the United States to take up a position at Bryn Mawr College in Pennsylvania. In 1935 she underwent surgery for an ovarian cyst and, despite signs of a recovery, +================================================================================ +Rank = 57; Score = 290816.0 +<|begin_of_text|>Neanderthal communities divided some of their tasks according to their sex. This is one of the main conclusions reached by a study performed by the Spanish National Research Council (CSIC), published in the Journal of Human Evolution. This study, which analyzed 99 incisors and canine teeth of 19 individuals from three different sites (El Sidron, in Asturias -- Spain, L'Hortus in France, and Spy in Belgium), reveals that the dental grooves present in the female fossils follow the same pattern, which is different to that found in male individuals. + +Analyses show that all Neanderthal individuals, regardless of age, had dental grooves. According to Antonio Rosas, CSIC researcher at the Spanish National Museum of Natural Sciences: "This is due to the custom of these societies to use the mouth as a third hand, as in some current populations, for tasks such as preparing the furs or chopping meat, for instance." + +Rosas specifies that "what we have now discovered is that the grooves detected in the teeth of adult women are longer than those found in adult men. Therefore we assume that the tasks performed were different." + +Other variables analyzed are the tiny spalls of the teeth enamel. Male individuals show a greater number of nicks in the enamel and dentin of the upper parts, while in female individuals these imperfections appear in the lower parts. + +It is still unclear which activities corresponded to women and which ones to men. However, the authors of the study note that, as in modern hunter-gatherer societies, women may have been responsible for the preparation of furs and the elaboration of garments. Researchers state that the retouching of the edges of stone tools seems to have been a male task. + +Almudena Estalrrich, CSIC researcher at the Spanish National Museum of Natural Sciences, adds: "Nevertheless, we believe that the specialization of labor by sex of the individuals was probably limited to a few tasks, as it is possible that both men and women participated equally in the hunting of big animals." + +Rosas concludes: "The study of Neanderthals has provided numerous discoveries in recent years. We have moved from thinking of them as little evolved beings, to know that they took care of the sick persons, buried their deceased, ate seafood, and even had different physical features than expected: there were redhead individuals, and with light skin and eyes. So far, we thought that the sexual division of labor was typical of sapiens societies, but apparently that's not true +================================================================================ +Rank = 58; Score = 290816.0 +<|begin_of_text|>Japan's efforts to recruit Indian IT engineers won't work. For a simple reason: these engineers are better off at home. + +With the domestic talent pool shrinking, Japan desperately needs foreign talent to fill the gap and revive its ailing economy. Especially IT engineers, where the country is expected to face a shortage of 600,000 positions. + +That’s why Tokyo is reaching to Asian countries like India, where it has launched a program called Project Indian Institutes of Technology (PIITs), inviting students from the Indian Institutes of Technology (IIT) to spend their internships at Japanese firms. + +Apparently, Tokyo hopes that some of these Indian engineers will opt to stay with Japanese firms beyond the internship time. + +But judging from the overwhelming response to a previous piece of ours on the topic, Asian talent doesn’t seem eager to work in Japan. + +The problem? Japan isn’t prepared to provide what foreign talent needs — a career, due to the lack of the right “regime,” the socioeconomic conditions and cultural mindset, prevailing in Japanese corporations. + +In fact, Indians are better off home. New Delhi is better prepared than Tokyo to attract and retain talent. + +That’s one of the key findings of the 2016-17 International Competitiveness report (World Economic Forum), which ranks India 23 in the capacity to attract talent -- well ahead of Japan, which ranks 77. Also, India ranks ahead of Japan in the capacity to retain talent—see table. + +Country Capacity to attract talent Capacity to retain talent Japan 77 38 India 22 32 + +Source: World Economic Forum + +Fund/ETF 12-month Performance iShares MSCI Japan (EWJ) 15.07% iShares India (NDY) 20.76 + +Source: Yahoo.finance.com 6/30/2017 + +There are a couple of good explanations for that. Talent is highly mobile. And it goes where growth and opportunity is. In the last couple of decades, India has been growing at near double-digit rates, while Japan has been floundering in the swamp of stagnation. + +Meanwhile, India has been attracting foreign giants, like Amazon, and they have been setting up research centers to tap into the country’s pool of software engineers and programmers, as its own companies continue to internationalize their operations. + +So why leaving home to head to Japan?<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 59; Score = 290816.0 +<|begin_of_text|>If girls just want to have fun, as Cyndi Lauper belted out in 1983, does that mean boys are serious stoics? Lauper probably wasn't trying to make a sociological statement in her head-bopping anthem, but her lyrics point to a much-debated question about gender differences. In studying the relative happiness of people across the globe, do men and women exhibit distinctive emotional characteristics? In short, is one sex generally happier than the other? + +It's important to note that evaluating happiness isn't a highly precise scientific undertaking. For one thing, happiness is a subjective term, and there's no universal measurement for it. Though someone smiles on the outside, it doesn't mean that his or her apparent joy reflects internally. Similarly, a lot of studies related to happiness rely on self-reporting. One survey that economists from the Wharton School at the University of Pennsylvania have used simply asks respondents to rate themselves on whether they are "very happy," "pretty happy" or "not happy." The resulting statistics don't take into account the nuances and background information of why and how people scored themselves. + +Despite the potential pitfalls of empirical happiness studies, we can draw some conclusions about who the happiest sociological groups of people are. Generally, older people are happier than younger ones. Wealth and good health also give us something to smile about more often. Yet when it comes to the happiness gender gap, things aren't so cut and dried. + +The 2003 Pew Global Attitudes Project that surveyed 38,000 men and women across 44 countries lends insight into the disparities between male and female happiness. Overall, women scored themselves as more satisfied with their lives than men [source: Pew Research Center]. Moreover, women from certain countries, including Pakistan, Japan and Argentina, appeared significantly happier. The differences between what affects men and women's levels of happiness explains this slight gender gap. Women, for instance, tend to focus on personal and domestic problems, while men concern themselves more with matters outside of the home. + +This variation in values is also reflected in how the recent recession has affected people's happiness. According to a Nielson Happiness Survey conducted in 51 countries, men's happiness hasn't weathered the economic storm well since they attach more importance to financial success than women. On the other hand, the women surveyed indicated they appreciated quality relationships above all [source: Reuters]. + +But women shouldn't start celebrating the good news just yet.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 60; Score = 290816.0 +<|begin_of_text|>Forfeit the playoffs or defy your religious beliefs. That’s the position the Israeli women’s lacrosse team could find itself in at the World Cup going on this week in Oshawa. + +THe Israeli women's lacrosse team has said it will forfeit a playoff game at the World Cup being held in Oshawa this week if that game lands on Saturday, the Sabbath. ( Jay Johnston / Game Day Photography ) + +And if they have to play on Saturday — the Shabbat (Sabbath) as the schedule-makers call for — they will forfeit. Even if it’s the gold-medal game. “We all felt the same way,” said attacker Jenna Block. “It was an absolute team consensus we will not play on Shabbat. Competing in the World Cup is very important to all of us, but who we’re representing and who we are comes first and foremost. + +Article Continued Below + +“That’s a sacrifice we’re all willing to make. This is about more than lacrosse. It’s about more than a game.” At 4-0, the Israelis are a bit of Cinderella surprise at the event and face Japan (4-0) on Tuesday. But the way the schedule is shaping up — and a win by Israel on Wednesday is all it will take — the Israeli team will guarantee itself a spot in the top-8. That means they’ll head toward a playoff game Saturday, the final day of the event that will determine final rankings and the gold medal. The team asked for a schedule change, offering three alternatives: Friday, before sundown. + +Saturday, after sundown. + +Sunday morning. + +Article Continued Below + +The Israelis also offered to pay the expenses of any teams out of pocket for the accommodation. Their request was denied by the Federation of International Lacrosse, which set the schedule. Stan Cockerton, president of FIL, said the federation had no problem accommodating the Israelis’ request in setting the schedule for round-robin pool play in ensuring no Saturday games. But he said the playoffs are a different animal, governed by bylaws with a specific timetable for insurance and game times, with 8 p.m. being the latest scheduled start. “We were able to reschedule (the round robin) so we’re not insensitive to the issue,” said Cockerton. “But for this coming Saturday, we have a set format. “We cannot accommodate them.” Israel’s suggested compromises were a no-go because two of them require play after the scheduled end of the tournament and one requires teams +================================================================================ +Rank = 61; Score = 288768.0 +<|begin_of_text|>Greenwashing is out, brownwashing is in. These days, GOP politicians are scrambling to distance themselves from past environment-friendly statements, initiatives, and votes. (Thanks to Grist reader Gary Wockner for naming this trend.) + +Check out the top 10 offenders. And watch for a lot more Republicans to join the club as we head toward the 2012 election. + +Photo: lukexmartin10. Scott Brown + +U.S. senator from Massachusetts + +Before: “Reducing carbon dioxide emission in Massachusetts has long been a priority of mine,” he said in 2008 when, as a member of the state Senate, he voted in favor of his state joining the Regional Greenhouse Gas Initiative, a carbon-trading initiative in the Northeast. “Passing this legislation is an important step … towards improving our environment.” + +After: “I think the globe is always heating and cooling. It’s a natural way of ebb and flow. The thing that concerns me lately is some of the information I’ve heard about potential tampering with some of the information,” he said in December 2009, as the “Climategate” faux-scandal was raging. In April 2011, he voted to strip the U.S. EPA of its authority to regulate carbon dioxide.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 62; Score = 288768.0 +<|begin_of_text|>FOR ABONNENTER + +Vi deler samfundet op i indvandrere og danskere, men opdelingen skal være mænd og kvinder. + +Ikke i indvandrere og danskere. Når der tales om problemer i samfundet, så er det problemer blandt den mandlige del af befolkningen. + +Nu er der ikke problemer med alle mænd, ligesom det ikke er alle indvandrere, der er kriminelle og utilpassede. Selv om man nogle gange kan få den opfattelse i pressen, er det selvfølgelig heller ikke alle mænd, der er kriminelle og utilpassede, men det er blandt den mandlige del af befolkningen, de store samfundsmæssige og sociale problemer findes. + +Vi hører ikke om det – det er et usynligt problem – men ja, vi ved da godt at en pædofil er en mand, at voldtægt begås af mænd, men det understreges ikke. + +En notits fra avisen siger »færre kræves fængslet, antallet af unge indvandrere, der ender i grundlovsforhør, falder«. Hvem er det, der ender i grundlovsforhør? Det er mænd! Uanset hvilken etnisk kategori de tilhører, så er det mænd, det fremhæves bare ikke. Vi hører ikke om det. + +Hvad med overskriften fra Politiken tirsdag 29. august: ’Hård kurs mod unge spiller fallit’. Hvem er de unge. Hvis man læser artiklen, fremgår det, at det er unge mænd. Man kan tro, at problemet er blandt alle unge, men det er det ikke, det er unge mænd indsatsen er rettet mod, jo selvfølgelig er der også piger, men de udgør en forsvindende lille del af problemet. + +Kriminalitet er et kønsbestemt problem, et mandeproblem. En ekstrem mandlig adfærd. + +Der skrives om vold, men det er ikke bare vold, nej det er mænds vold, mænds vold mod kvinder, børn og andre mænd +================================================================================ +Rank = 63; Score = 288768.0 +<|begin_of_text|>Manitoba drivers who hit and kill a cyclist often end up simply paying a fine. + +A Global News investigation found multiple cases where the driver was never criminally charged. Instead the person was charged under the Highway Traffic Act (HTA) and the offence often resulted in a few thousand dollars fine. + +RELATED: 87 year old Winnipeg cyclist identified as victim of Friday’s crash + +Leslie Freudenberg, 87, was an avid cyclist. But in July 2016, his passion got him killed. Freudenberg was struck by a piece of heavy machinery. + +“He was riding his bike down a cyclist pathway and was crossing the street,” his son-in-law, Keith Johnson, told Global News. “A piece of heavy equipment was being transported down the street and didn’t have a full view for the driver. He didn’t even know that he actually hit my father-in-law.” + +Freudenberg was killed instantly. The 34-year-old driver was never charged. Instead, PCL Construction, the company that owned the truck, is going to court for not providing a clear view to the driver. The Manitoba Justice department said the maximum penalty the company will face is a $2,000 fine. + +“I guess the cruel nature of all of that is the fact that the family has never really been given the opportunity for closure at this point,” Johnson said. “We’ve paid the ultimate price in losing a family member that we’ll never get back.” + +This scenario – a pedestrian or cyclist is hit and killed and the driver receives a light penalty – is playing out across the country. Global News has spoken to many families from coast to coast who are outraged that careless drivers aren’t facing stiffer fines after hitting and killing their loved ones. + +WATCH: Life is cheap: Trudi Mason recounts harrowing story of losing her friend to a pickup truck + +Trudi Mason of Lethbridge, Alta. was cycling with her best friend when a truck hit the both of them. Mason’s friend died. Mason says the court process ignores the fact a person was killed. + +The family feels Freudenberg’s death could have been prevented had PCL followed equipment protocol. Instead, the incident happened as the driver was moving the heavy machinery from one end of the street to the other. + +“Les’ death hasn’t been taken seriously,” Johnson said. “It’s the worst slap in the face that we could ever experience. It just further denies any closure that the family has to put this behind them.” + +Freudenberg’s tragic death is just one of many similar cases +================================================================================ +Rank = 64; Score = 288768.0 +<|begin_of_text|>Polar bear hairs are hollow to maximize insulating qualities of the animals' fur, as almost every student of the Arctic knows. + +But now a set of studies from China shows polar bear hairs are much more than simple tiny tubes. + +Detailed mathematical analysis of the hairs, published in the journal Thermal Science, finds they have complex structures that make them much better insulators than simple hollow hairs would be. + +Microscopic examination of polar bear hairs reveals their interior is a structure of membrane pores, researchers from several Chinese universities have found. The latest analysis finds the pore structure is arranged as a fractal, a series of repeating patterns spun off into smaller dimensions. + +Calculation of the pore structure finds the ratio of its dimensions to be close to a mathematical figure known as the "golden mean," the ideal dimension ratio for an infinitely spiraling fractal, says one study. The dimension ratio of the inner structure of the analyzed polar-bear hair was calculated at 1.625, close to the golden mean, which is also called Phi and is approximately 1.618; the golden mean "must reveal the possible optimal structure of polar bear hairs," the study says. + +Though it appears white, polar bear fur is translucent, helping it absorb environmental heat, the study points out. Without the interior pores' fractal arrangement, however, the translucent hair that absorbs light could easily send heat back out into the environment, the study says. + +A related study, with some of the same co-authors, calculates the equation for one-dimensional heat conduction through each "labyrinth cavity" of the hair. The authors used calculus to arrive at a differential equation showing how heat moves through the hairs. + +The studies -- the latest in a series on the same subject by the same group of Chinese researchers -- are not mere academic exercises. + +Understanding the structure and workings of polar-bear hairs "may find many potential applications in the future, especially in thermal insulation designs for extreme cases," said an earlier study by some of the same authors, published in 2011 in Thermal Science. + +They note in their studies that polar bears maintain body temperatures of 98.6 degrees in an environment where temperatures can dip as low as minus 76 degrees. + +The Chinese researchers are not alone in looking to polar-bear hair as a model for future heat-collecting and heat-holding products.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 65; Score = 284672.0 +<|begin_of_text|>This same bus was seen running through the streets of Madrid with its blunter message that “boys have penises and girls have vaginas. Don’t be deceived”This advertisement for intolerance, which belies intelligence on this subject is now in Manhattan parked in front of the UN building and says something slightly different in English:“Its biology…Boys are boys and always will be and girls are girls and always will be. You can’t change sex. Respect all”Never mind that message is written by people who understand nothing and most importantly are confusing gender identity with sexual organs and I won’t surprise you by telling you it has been funded by 3 right wing conservative groups one of whom is called the National Organization for Marriage.Of the Spanish version, one Madrid councilman declared it the “bus of shame,” and city officials ordered it removed from the streets for violating a traffic law restricting advertising on private vehicles.“We need a discussion about how to respect everyone,” says Joseph Grabowski who is a representative for one of the groups. But he also claimed that being transgender is a “disorder” and that a respectful discussion does not extend to recognizing a transgender person’s gender identity in public settings. (The American Psychiatric Association does not classify being transgender as a mental disorder.) He also added, “They can live that out privately.”Not only are these groups profoundly ignorant of science and reality but they also represent a dangerous new backlash trend among the “earth is flat crowd”; the very same ones that occupy the Bible belt in the United States and voted for Donald Trump. They can also be found in every country in the world usually also denying climate change, women’s rights and decrying homosexuality as a sin.Let’s hope that their message does not receive a receptive ear and enough people take exception in order to educate the sponsors although I suspect that wouldn’t have much effect on those who've already drunk the Kool-Aid.Is it not the least bit ironic that the "respect all" in their message doesn't ring a bell of contradiction in their minds?<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 66; Score = 284672.0 +<|begin_of_text|>Modi is the King now. Don’t be surprised. He holds the title of the Prime Minister. But he wields so much unbridled power that he is virtually no less than a king. Modi has always had the dream to have power that none could question him. Indeed, the Cabinet reshuffle on Sunday has put him in a position where no one can even raise a finger in Modi’s presence. He no longer shares power with anybody in his council of ministers. Rather he himself is the source of power for his ministers. + +Power sharing is the basic principle of the parliamentary system of governance wherein the Prime Minister and his council of ministers take a decision with consensus. A Prime Minister is therefore called the first among equals in the parliamentary system which India has adopted right from its inception as a free nation. + +This system ensures check upon a Prime Minister from turning an autocrat because ministers act upon him/her as a balancing force. They work in the interest of the region they represent. This, in turn, puts a democratic pressure on a Prime Minister while taking a decision. + +But Modi is not the one to bear any restrictions when it comes to exercising power. He is temperamentally an autocrat. He is essentially a narcissist who is deeply impressed with himself. He not just believes himself to be destined to change the destiny of ‘1.5 billion Indians’. But he is also convinced that he alone can serve the country the best. This pattern of thinking is very much anti-democratic and such charecters always transform into an autocrat. + +Narendra Modi has often said about himself that he left his “family for the country”. The country is supreme for him and even over and above his own family. It is also the RSS belief. No RSS functionary can have a family. Modi was so impressed with the RSS that he left behind his bride and never went back to his family. He ‘devoted’ his life to the country, as he himself claims. + +So, Modi’s idea of India is based on the Sangh’s idea of the country which, in turn, is Hindu Rashtra. It very much reflects in Modi’s style of governance right from the Gujarat riots to mob lynching of minorities. A Hindu Rashtra is not feasible within a democratic system because in such a system, decision-making to be democratic. So, Modi’s worry was how to turn the system upside down to wield unbridled power to construct a “new India” of his ideas! + +Modi ultimately found a way out through the Cabinet resh +================================================================================ +Rank = 67; Score = 284672.0 +<|begin_of_text|>Executive orders issued by Presidents of the United States to help officers and agencies of the executive branch manage operations within the government. + +At the federal level of government in the United States, laws are made almost exclusively by legislation. Such legislation originates as an Act of Congress passed by the United States Congress; such acts were either signed into law by the President or passed by Congress after a presidential veto. + +However, legislation is not the only source of regulations. There is also judge-made common law and constitutional law. The President can issue executive orders pursuant to a grant of discretion from Congress, or under the inherent powers that office holds to deal with certain matters which have the force of law. + +Many early executive orders were not recorded. The State Department began numbering executive orders in the early 20th century, starting retroactively from President Abraham Lincoln's Executive Order Establishing a Provisional Court in Louisiana issued in 1862.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 68; Score = 282624.0 +<|begin_of_text|>HP's IaaS/PaaS Helion Public venture turns private. Microsoft, with its burgeoning Azure Cloud, announces serious container substrate in the form of its Nano Server, Docker compatibility and additional equivalents. + +Helion, Helioff. It's not as bad as that at all, but the effort made to compete with AWS is over, for now. HP is backing away from the pubic cloud and going for perhaps a safer, and more lucrative, future in the private cloud. It's my guess that inevitably, HP will make a better return on assets this way. The changes will be frustrating to planners, perhaps, but the cloud-on-the-hoof turf war is largely over. + +Face it: It's brutal out there. Commodity infrastructure portends that margins will be slim. Amazon knows know to make microprofits. HP is a machine that needs more revenue per SKU. There's nothing like not making money and making up for it in volume. HP's Helion isn't dead by any means, rather, it's becoming focused infrastructure, in my opinion. + +[ Now read 20 hot jobs ambitious IT pros should shoot for. ] + +You could smell the scent in the breeze. HP buys Eucalyptus, an AWS-enabler. Serial entrepreneur Marten Mickos, of MySQL and Eucalyptus fame, goes on comparative waivers. Amazon chugs along. Rackspace was rumored to be for sale, something I find hard to believe (in fact, the company said in September that "it has ended its evaluation of alternatives that would result in Rackspace being acquired" and was planning to go forward independently). OpenStack languishes because it's too tough for many people to get their heads around. The market size seems smaller than everyone had dreamed. + +But these are infrastructure ploys. Unless you're an expert at microprofits, you can be toasted, quickly, at high rates of speed. Microsoft dropped its traditional pre-announcement of an ought-to-be-available in a hoped-for future release, Docker on Windows, and their own competitor to skinny operating systems, the Windows 2016 Nano Server, a seemingly impossible sort of substrate for the new rage of containers. + +Why impossible? Since when has Microsoft made a skinny operating system? Maybe long ago, as Xenix. Sure, you can get non-GUI flavors of Microsoft servers, seemingly surgically deployed as 2008/2012 Rsomething, with but a handful of the fatuous menu of pulled-from-retirement API sets, and given character to +================================================================================ +Rank = 69; Score = 282624.0 +<|begin_of_text|>What does the word "rape" mean to you? For many reading this post, I suspect, it is a trigger to appalling events in their own lives. Because rape is an everyday crime. By my calculations, roughly 230 people are raped each day in England and Wales. + +Police, this morning, called for specialist units to investigate rape allegations - senior officers are ashamed of a conviction rate they calculate at 6%. "Not good enough" says the Assistant Commissioner of the Metropolitan Police John Yates. + +But analysis of the Home Office data on what they call "intimate violence" suggests the conviction rate is much lower. And the scale of the problem far greater. My source is the Home Office supplement to the British Crime Survey - Homicides, Firearm Offences and Intimate Violence 2006/07 (pdf link). This is a remarkable piece of research in which 13,000 people were asked to fill out an anonymous questionnaire on their experiences of domestic violence, sexual assault and rape. + +It is an exercise that has been conducted three times now and is backed up by other academic studies. The results are consistent. One in 20 women said they had been raped since they were 16. One in 200 said they had been raped in the previous 12 months. In terms of the population of England and Wales, that suggests 85,000 women are raped each year - 230 a day. And yet the number of men convicted of rape is fewer than 800 a year. So the chance of a victim seeing her attacker jailed is less than one in 100. + +But rape is a complex crime. Only 17% of rapists are strangers to their victim. Just 4% are cases of date rape. Half (54%) are committed by a husband, partner or ex-partner. What's more, even though their experience is technically rape in law, 57% of rape victims don't necessarily think of themselves that way. + +To be clear what the figures categorise as rape, the definition is this: "the penetration of the vagina or anus without consent and penetration of the mouth by a penis without consent." + +Since the 39,000 people who have taken part in the studies benefited nothing from alleging rape, and the results appear consistent, it seems probable that the research gives a realistic sense of the scale. + +If one looks at the data on rape at any point during adult life, it suggests 700,000 women have suffered in that way - equivalent to the entire population of Leeds. + +Figures for +================================================================================ +Rank = 70; Score = 282624.0 +<|begin_of_text|>Good food, good eating, is all about blood and organs, cruelty and decay. It’s about sodium-loaded pork fat, stinky triple-cream cheeses, the tender thymus glands and distended livers of young animals. It’s about danger—risking the dark, bacterial forces of beef, chicken, cheese, and shellfish. Your first two hundred and seven Wellfleet oysters may transport you to a state of rapture, but your two hundred and eighth may send you to bed with the sweats, chills, and vomits. Gastronomy is the science of pain. Professional cooks belong to a secret society whose ancient rituals derive from the principles of stoicism in the face of humiliation, injury, fatigue, and the threat of illness. The members of a tight, well-greased kitchen staff are a lot like a submarine crew. Confined for most of their waking hours in hot, airless spaces, and ruled by despotic leaders, they often acquire the characteristics of the poor saps who were press-ganged into the royal navies of Napoleonic times—superstition, a contempt for outsiders, and a loyalty to no flag but their own. A good deal has changed since Orwell’s memoir of the months he spent as a dishwasher in “Down and Out in Paris and London.” Gas ranges and exhaust fans have gone a long way toward increasing the life span of the working culinarian. Nowadays, most aspiring cooks come into the business because they want to: they have chosen this life, studied for it. Today’s top chefs are like star athletes. They bounce from kitchen to kitchen—free agents in search of more money, more acclaim. I’ve been a chef in New York for more than ten years, and, for the decade before that, a dishwasher, a prep drone, a line cook, and a sous-chef. I came into the business when cooks still smoked on the line and wore headbands. A few years ago, I wasn’t surprised to hear rumors of a study of the nation’s prison population which reportedly found that the leading civilian occupation among inmates before they were put behind bars was “cook.” As most of us in the restaurant business know, there is a powerful strain of criminality in the industry, ranging from the dope-dealing busboy with beeper and cell phone to the restaurant owner who has two sets of accounting books. In fact, it was the unsavory side of professional cooking that attracted me to it in the first place. In +================================================================================ +Rank = 71; Score = 282624.0 +<|begin_of_text|>Being feminist means being vegan. You can’t be a true feminist if you’re eating meat. + +I read this comment on Facebook and it hit a nerve with me. My first reaction was, “Why should people be telling others what to eat and what not to eat?” The more I thought about it, the more it bothered me for many reasons. + +The argument is that being a feminist means addressing all forms of oppression, including oppression of animals. The animal agricultural industry exploits animals. Female animals are forced to breed through artificial or manual insemination. Their youngs are taken from them prematurely. Female chickens are debeaked and grow to be too big for their legs to carry them. Pigs are penned in cages too small for them to move. The theory is that exploitation of the female animal reproductive system is very relevant to feminists fighting against the patriarchy and fighting for human female reproductive rights. Oppression of animals should intersect with oppression of women. + +When vegans or feminists make statements like mentioned above, it can come off as very Western-centric, classist, and elitist. + +I’ve never felt like going vegan or vegetarian, not because I didn’t understand the plight for animal rights, but because it is something I can’t identify with. + +First and most importantly, Hmong culture is a very big part of my identity, and food is Hmong culture. We (family and friends) gather around food. Food is served at birthday parties, religious ceremonies, celebrations, and even during short visits to family. Food is part of family life. + +So, why not substitute meat ingredients with vegan products? + +Our ancestors cultivated the lands of Laos, Vietnam, and China for farming and livestock. We brought our culture with us when we immigrated to the US as refugees of war after the Vietnam War. + +Religious ceremonies, like hu plig (soul calling), require eggs and the sacrifice of chickens and roosters. Other ceremonies involve sacrificial cows, pigs, and sheep/goats. We need chicken’s blood and feathers for the New Year to create a new shrine for our xwm kab. So, how should we perform or practice our religion if we go vegan? What about the cultural custom of dieting only on boiled chicken and herbs 30 days postpartum? + +Meat has always been very precious to the Hmong. We don’t waste anything, so the dishes we cook during feasts are from the meat of the animals we sacrifice for religious ceremonies. Food we eat throughout the year are leftover meat from the same animals. + +V +================================================================================ +Rank = 72; Score = 282624.0 +<|begin_of_text|>Pipelines leak. It’s what they do. And the Dakota Access pipeline will be no exception. + +The 1,168 mile-long pipeline is intended to transport crude oil from the Bakken oil fields in North Dakota to Patoka, Illinois, passing through South Dakota and Iowa. From there, it will connect with another pipeline to transfer oil to terminals and refineries along the Gulf of México. + +However, its construction has hit a snag in Standing Rock, North Dakota. Since September the Standing Rock Sioux Tribe and allies from around the world have been pushing hard to prevent the pipeline from cutting through Sioux Treaty territory. + +The water protectors maintain–under enormous pressure from the government—that construction of the pipeline must be halted immediately to protect the waters of the Missouri River and “the very survival” of the tribe. + +The pipeline will cross under the Missouri River, the longest river in the United States and the sole water source of the Standing Rock Sioux Tribe. + +Pipeline ruptures happen all the time. It’s a fact. An excavator might dig in the wrong place. The pipeline might start leaking due to corrosion. It might explode due to poor operation. It could even break due to improper construction or the use of substandard materials. + +Scorching land, torching homes—and killing multiple people. Pipeline explosions in 2016 show why the #NoDAPL fight is still going strong: pic.twitter.com/RNVfIAWwaj — Fusion (@Fusion) December 1, 2016 + +According to the Center for Biological Diversity, over the last 27 years there have been close to 8,000 oil spills in the United States alone. That’s “nearly 300 per year on average.” + +If the DAPL were to follow this environmentally destructive trend, there would be far reaching consequences for the Standing Rock Sioux and potentially millions of other Americans. + +Missouri Attorney General Chris Koster said it best: “The River provides drinking water for communities, cooling water for power plants, and a low-cost highway for goods and agricultural products.” + +Contaminated water from North Dakota has the potential to hit twelve states, all the way to Louisiana, wreaking havoc across America’s heartland. + +The negative impact of the DAPL has already started. There are claims that the Army Corps of Engineers–the government body that approved the project– cut corners by failing to consult with the tribe before letting the pipeline developers, Energy Transfer Partners, proceed with construction. A failed lawsuit in early September argued that, in doing so, the Corps +================================================================================ +Rank = 73; Score = 282624.0 +<|begin_of_text|>Women are discriminated against in all areas of gaming. What can we do about it? + +WARNING: this article involves feminism. Those offended by the idea of gender equality should STEP AWAY FROM THEIR COMPUTERS NOW and get back inside their caves. Not a misogynist? Then hey buddy, sit back and relax. I’d offer you a cup of tea but you’re miles away (and I don’t have any tea). + +Sexism permeates our society, that’s a given, but you might not have realised just how much it occurs in gaming. There are three levels at which it exists: at an industry level, in games themselves, and from within gamers. It’s vital that we are aware of these problems but, more importantly, we have to do our best in tackling them. To do this, we must look at each problem in turn. + +First: sexism in the industry. There is evidence that suggests women who work in gaming are under-represented and suffer some of the worst gender-based harassment. Female gamers make up 45% of the gaming population; yet male game designers make up 89% of the industry and earn 23.6% more than their female counterparts. What a kick in the balls vagina. There have also been many high-profile incidents of women being harassed by men in the workplace. One of the most recent stories involved Josh Mattingly, CEO and founder of Indiestatik, publicly apologising and “stepping down” after sending a female game developer inappropriate and unprovoked sexual messages on Facebook. + +People have claimed that women should just stand up for themselves and that speaking up against those that harass them should be enough to stop the problem, but there is not always a clear cut solution. The #thatwoman hashtag shows that in speaking up, women are likely to be ostracised and have it limit their future career. Perhaps the risk would be worth the overall benefits, but it is understandable why women in gaming don’t always feel free to confront those that may oppress or belittle them. It would be like Yoshi trying to stand up to Mario – he’s just going to drop Yoshi mid-jump and carry on like nothing happened anyway. + +Similarly, the #1reasonwhy hashtag asked women to speak out about why they don’t feel comfortable in the gaming industry. The results included stories of groping and being constantly mistaken for assistants and receptionists. + +Jade Raymond, Producer of Assassin’s Creed, was accused of using her looks to sell the game + +What is there to be done about +================================================================================ +Rank = 74; Score = 280576.0 +<|begin_of_text|>Cyprus is Turkish, after all. Turks can do whatever they want there. They can even celebrate dropping napalm on Greeks and slaughtering them. + +Uzay Bulut The writer is a Turkish journalist and political analyst formerly based in Ankara. She now lives in the United States. More from the author ► The writer is a Turkish journalist and political analyst formerly based in Ankara. She now lives in the United States. + +On August 8, Muslim Turkish Cypriots and illegal settlers from Turkey celebrated the 53rd anniversary of Turkey’s napalm bombing of Greek Cypriot civilians in the Turkish-occupied enclave of Kokkina in Cyprus. Mustafa Akıncı, the president of the self-styled “Turkish Republic of Northern Cyprus” (TRNC), which is recognized only by Turkey, also participated in the celebrations. + +In August 1964, Turkish warplanes dropped napalm bombs on Kokkina in the Tillyria peninsula, hitting residential areas and a hospital, and killing more than 50 people, including 19 civilians. Ten years later, in 1974, Turkey invaded Cyprus and has occupied almost 40 percent of the island ever since. + +The Ministry of Foreign Affairs of Greece issued a note of condemnation regarding the celebrations: + +“We are dismayed to note the celebrations of the Turkish Cypriot leadership, including Mr. Akinci himself, of the 53rd anniversary of the use of chemical weapons and dropping of napalm bombs by the Turkish air force on the Tillyria peninsula. This was the first use of banned chemical weapons in the history of our planet. + +“Today, when the whole planet bows to the victims of wars and such hostile acts, the holding of and participation in such celebrations is an affront to international law, to the memory of the fallen, and to the whole of humanity.” + +The Republic of Cyprus declared independence in 1960. Afterwards, Turkey escalated its preparations to invade the island, which included but were not limited to establishing a bridgehead at Kokkina in 1964 and smuggling arms and fighters from Turkey into the area in order to strengthen Turkish positions there. + +According to the High Commission of the Republic of Cyprus in London, + +“When in August 1964 the [Cypriot] Government attempted to contain the Kokkina bridgehead, Turkey's air force bombed the National Guard and neighboring Greek villages with napalm and threatened to invade. The other major purpose served by the enclaves was the political and physical separation of the two communities.” + +Another preparation for +================================================================================ +Rank = 75; Score = 278528.0 +<|begin_of_text|>Civil unions are an inadequate substitute for marriage. Creating a separate, new legal structure to confer some benefits on same-sex couples neither honors American ideals of fairness, nor does it grant true equality. The results are clearly visible in New Jersey, which continues to deny same-sex couples some of the tangible civil benefits that come with marriage. + +Gov. Jon Corzine of New Jersey has long said that he would sign a measure granting the right to marry to couples of the same sex. We are heartened that he has declared that that should happen sooner rather than later. + +We hope Mr. Corzine intends to prod legislators into passing such a law early in the 2009 session. That would make New Jersey the first state to legalize marriage for same-sex couples through legislative action. Three other states — Connecticut, Massachusetts and California — have done so through the courts. Unfortunately, California voters approved a ballot measure in November rescinding that right, at least for now. + +Mr. Corzine made his statement after a state commission released its final report on New Jersey’s two-year-old civil union law. The commission noted the hurt and stigma inflicted by shutting out gay people from the institution of marriage. It also found that civil unions do not assure gay couples of the same protections, including the right to collect benefits under a partner’s health insurance program and to make medical decisions on behalf of a partner who is unable to do so. The panel concluded unanimously that the state should enact a law to remove the inequities. + +We regret that the leaders of the state’s Democratic-controlled Legislature do not view this issue with the same urgency. Senate President Richard Codey, for instance, said recently that progress in civil rights areas “is typically achieved in incremental steps.” We suspect that political expedience is clouding Mr. Codey’s sense of fairness. Next year in New Jersey, the governorship and all seats in the Assembly are up for grabs in an election. Some Republicans already are talking about making their opposition to same-sex marriage a campaign issue. + +Governor Corzine typically takes the right side on important issues, but he has been known to retreat in the face of opposition. We hope that’s not the case here. It’s past time for him and for the Democrats in Trenton to find the political courage to extend the right to marry to gay couples.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 76; Score = 278528.0 +<|begin_of_text|>There is a point to including playable female characters in games. I’m aware that most of the people likely to comment on this article (go ahead and bleat about misandry, you worms; I’ll enjoy a tasty cup of your male tears) don’t see that, but I’m also aware that the vast majority of people who read this article do see it, and won’t bother to leave a comment because what I’m saying in this editorial seems sensible, practical, and non-controversial. Such is the way of the Internet. So I’m not going to bother writing out a lengthy justification of why we genuinely need female characters in video games for the good of the industry financially and artistically; if you honestly can’t understand it, go forth and educate yourself. If you feel that gaming is the one thing remaining to men and girls should stop spoiling it with political correctness, then please go boil your head because I see no point in debating with people incapable of basic logic and lacking humanity. + +Having taken it as fact that there is a point to including female characters in video games, why on earth are we still hearing excuses for their absence in 2014? Because it is an excuse. There is no reason not to do it. You won’t alienate your existing market by acknowledging the existence of women. You won’t take anything away from your existing market. + +Yet another clueless wonder is yapping about the absence of the unnecessary from video games:I am a game designer. I am designing and producing a game that does not, and will not, have a single female character in it. This is not because I am misogynistic. This is not because I do not women to play the game. This is because putting women in the game makes no sense, violates the principle of the suspension of disbelief, and will not make the game any better as a game.I am the lead designer of First Sword, a combat management game. The game has orcs and men, elves and dwarves. It has goblins and trolls. But it has no women.Why not? Because the game is a gladiator game. Women cannot credibly fight as gladiators. We don't put women in the game for the same reason we don't put bunny rabbits or children in the game. Putting women in the game would be an act of brutal sadism, an act of barbarism even by pagan Roman standards. While the Romans did occasionally put female gladiators in the arena, they were there as a comedic act. +================================================================================ +Rank = 77; Score = 278528.0 +<|begin_of_text|>Pillars getting in the way of ceiling cables, water pipes clashing with electrical wires - Mr Yip Shaw Chong has seen many such mistakes on construction sites. + +Such errors might require cables to be rerouted or walls to be hacked, wasting resources and time and jeopardising the schedules which Mr Yip oversees as senior project manager at Shimizu Corporation's Singapore office. + +"Traditionally, we had drawings for each trade," he explains, one set of technical drawings showing electrical works, another showing water pipes, another the walls, and so on. + +"Of course, an experienced guy would be able to layer them together and visualise that they will clash. But, inevitably, somebody misses something, there's a clash on site and somebody has to go and fix it." + +Today, however, spotting such errors can easily be done by a beginner, or even a computer program. This is thanks to the rise of Building Information Modelling (BIM), a data-heavy virtual 3D-modelling technology that the industry has been increasingly pushed to adopt since the start of the decade. + +Related Story Smart cities, smoother lives through digital connectivity + +Since last July, all building plans for new projects with a gross floor area of 5,000 sq m and more have had to be submitted in BIM format for regulatory approval. + +The great advantage of BIM software is that it can combine all plans - architectural, engineering, and mechanical and electrical (M&E) - into one integrated model. + +"Before, you looked at the 2D drawings, this floor and this floor, and then you tried to visualise whatever intersections there were," says Mr Yip. "Whereas now, you have a model, so it's easier to see clashes." + +Recent versions of software are even able to automatically spot and highlight these conflicts. + +NEW WAYS OF SEEING + +Admittedly, a lot of work must be done upstream to save time on site. + +Fast Forward series + +With Singapore firmly focused on the Future Economy, The Straits Times' series, Fast Forward: Disruption and the Singapore Economy, helps you make sense of the big shifts that will shake up entire sectors, reshape jobs and change lives. Every Saturday for 12 weeks, the paper's journalists will examine a disruptive force, its likely impact on the economy and how soon that will be felt. From robotics, 3D printing and smart buildings to dire demographic trends, the global skills revolution and the Asean growth story. Next week, find out more about green technology and +================================================================================ +Rank = 78; Score = 276480.0 +<|begin_of_text|>Delhi Metro rides will be costlier from Wednesday, and then again in October. + +The Delhi Metro Rail Corporation (DMRC) announced on Monday the revised fares – a minimum of Rs 10 from the earlier Rs 8 and a maximum of Rs 50. + +The 15-km slab, which has been raised from Rs 18 to Rs 30, is likely to hurt daily commuters the most. + +The fare structure will change again from October 1 when the minimum fare for travelling a distance of more than 2 km will go up by another Rs 10. + +Here’s a calculator to help you find out how much your fare will increase. Just select two stations, and it’ll tell you how much more money you’ll have to pay to ride the Delhi Metro. + +If you are unable to view the calculator, click here + +First Published: May 09, 2017 19:58 IST<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 79; Score = 276480.0 +<|begin_of_text|>He said, she said. Eddie Cibrian is speaking out once more against his ex-wife, Brandi Glanville. + +Last week, Glanville, 44, accused the actor and his current wife, LeAnn Rimes, of allegedly stalking her boyfriend, Donald Friese, on social media. In a series of tweets, she claimed that Rimes, 34, watched four of Friese’s Snapchats before showing up with and Cibrian and Glanville’s two sons at the same Malibu restaurant last month. + +“Stalking my boyfriend to show up with my kids was the last straw,” Glanville tweeted. + +Now, Cibrian, 43, is telling his side of the story. + +“Brandi was very drunk and after already being at our table, started to come back again. Her boyfriend ‘ran interference’ and came to ask if she could take photos with the kids,” Cibrian claims in an exclusive statement to Us Weekly. “After witnessing Brandi’s behavior at the restaurant I was concerned about what pictures Brandi might post. We looked at their socials after we got home to make sure there was nothing of concern. That’s exactly how it all went down.” + +The Rosewood actor previously opened up about the situation on June 10. “I normally don’t respond to Brandi’s foolishness but I will not allow false and reverse accusations to go unanswered about my wife,” Cibrian told Us in a statement at the time. “LeAnn is a fantastic stepmom to the boys and is always gracious to their mother. Having to put up with Brandi’s made up drama all the time is extremely frustrating. After eight years we should have one priority, making sure two incredible kids are loved and remain happy and healthy. But every couple of months there is another accusation coming from Brandi in an attempt to drum up drama to stay relevant.” + +He added: “LeAnn and I did not nor have we ever ‘shown up’ at places where Brandi will be. Why would we do that? Makes no sense. We had a reservation held at Nobu five days before Brandi posted she was going. Here is proof and if anyone needs more, call Nobu and they will confirm.” + +Cibrian and the former Real Housewives of Beverly Hills star divorced in 2010 after nine years of marriage. As Us Weekly exclusively revealed, Cibrian and Rimes fell for each other while filming the 2009 Lifetime movie Northern Lights together. + +Despite +================================================================================ +Rank = 80; Score = 274432.0 +<|begin_of_text|>MEPs have revealed they want boys to learn about traditionally ‘female’ activities and should be taught domestic work and care at school. + +The Brussels politicians said they want to stand against ‘sexist’ education by encouraging children to take an equal interest in all subjects ‘beyond gendered stereotypes’. + +They hope that girls will take up scientific and technical subjects while boys could take up activities such as cleaning the home. + +MEPs have revealed they want boys to learn about traditionally ‘female’ activities and should be taught domestic work and care at school + +Textbooks showing old-fashioned stereotypes about male and female roles would also be thrown out of schools, under the European Parliament’ proposals. + +The suggestions were made in the report on Empowering Girls Through Education from the Parliament’s FEMM Committee on Women’s Rights and Gender Equality, with MEPs approving the proposals by 408 to 236. + +The report says it 'encourages girls and boys in the education process to take an equal interest in all subjects, beyond gendered stereotypes, in particular as regards scientific and technical subjects, including boys’ learning about activities regarded as female, in areas such as domestic work and care’. + +The Brussels politicians said they want to stand against ‘sexist’ education by encouraging children to take an equal interest in all subjects ‘beyond gendered stereotypes’ + +Another point says that schools should be guided ‘to embrace a gender perspective and gender equality, and to ensure the elimination of stereotypes and sexist distortions that textbooks and teaching materials may include in their content.' + +It added: '[This will then] encourage them also to combat this sexism in literature, film, music, games, media, advertising and other areas that can contribute decisively to changing the attitudes, behaviour and identity of girls and boys’. + +Portuguese socialist MEP Liliana Rodrigues, who spearheaded the proposals, said: ‘We are still living in an unequal Europe…women continue to be a prime target for discrimination and violence. I believe that school plays a fundamental role in changing this. + +‘[We want to] create a school culture of gender equality, critically oversee the curricula and educational materials, ensure gender equality with regard to personal and professional decisions and improve the percentage of women in positions of responsibility.’ + +But Leading Labour Eurosceptic MP Kate Hoey told The Daily Express: ‘I have confidence in our nation’s ability to deal with educating our own children. It is time for the EU to stop wasting money on interfering in matters that are none of their business.’<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 81; Score = 272384.0 +<|begin_of_text|>SOME women would prefer a world where men don’t exist. Just picture it. No hogging the remote. No beard clippings in the sink. + +After all, it works for other species. Komodo dragons can reproduce without sex with a male partner. So can whiptail lizards, hammerhead sharks and parasitic wasps. And they seem happy enough. + +But researchers say there’s a very important evolutionary reason men exist and without them a hypothetical asexual all-female human race would likely become extinct. + +Professor Matt Gage from the University of East Anglia asked the question seven years ago: “Why should any species waste all that effort on sons?” + +The answer, revealed this week, is more basic that you might think. + +“We wanted to understand how Darwinian selection can allow this widespread and seemingly wasteful reproductive system to persist, when a system where all individuals produce offspring without sex — as in all-female asexual populations — would be a far more effective route to reproduce greater numbers of offspring. + +His research, funded by the Natural Environment Research Council, showed competition among males helped produce the healthiest offspring and the absence of sex could lead to mutation among offspring and eventual extinction. + +“To be good at out-competing rivals and attracting partners in the struggle to reproduce, an individual has to be good at most things, so sexual selection provides an important and effective filter to maintain and improve population genetic health,” prof. Gage said. + +“Our findings provide direct support for the idea that sex persists as a dominant mode of reproduction because it allows sexual selection to provide these important genetic benefits. + +“In the absence of sex, populations accumulate deleterious mutations through a ratcheting effect where each new mutation takes a population closer to extinction. + +What if there were fewer males? The researchers tested the theory that a more efficient form of reproduction would involve females giving birth to more females. + +To find out if that might work, researchers took tribolium flour beetles and paired them with different numbers of males. + +Over 10 years they paired 90 males with only 10 females. They then paired single males and females in monogamous pairings. + +“Our monogamous treatment, for example, where there was no sexual selection for 50 generations, resulted in a lower level of population health and rapid extinction when populations were challenged by inbreeding. All the populations derived from monogamous histories became extinct after just eight generations,” prof. Gage said. + +By contrast, populations where males had to compete for female attention maintained population health and avoided extinction +================================================================================ +Rank = 82; Score = 272384.0 +<|begin_of_text|>Nothing is like anything else. You can do nothing well or you can do nothing badly. Some people excel at nothing. Others have more difficulty with it. They grow restless, resent the loss of initiative and control, and, more deeply, they feel that “something” is inherently, even morally, superior to nothing. + +The U.S. government national security apparatus, for better or for worse, sucks at nothing. In the policy process, the saying goes, something always beats nothing. As a bureaucratic fact, this is clearly true. But, from a policy perspective, is nothing always the wrong choice? + +Currently, the Trump administration is engaged in something of a natural experiment on this question in eastern Syria. U.S. forces went to eastern Syria in 2014 to defeat ISIL. Now, with ISIL losing its last grip on territorial control in Syria, the foreign policy apparatus is reprising a time-honored tradition and asking itself what those forces should do next. The answer is unlikely to be nothing. + +In what is surely a dizzying array of internal meeting and papers, the debate centers on how the United States should leverage its military presence in Syria, which the Pentagon recently acknowledged exceeds 2,000 troops, to pursue other objectives, namely countering Iran’s influence or post-conflict stabilization, or both. Entering into that debate, Faysal Itani has argued in this journal that a “small-time” U.S. mission in Syria would be unable to effectively counter Iran, pointing to the “imbalance of commitment and interests between the United States and Iran” and the overwhelming numerical advantage of pro-Iran forces. He notes that the United States would need a force of many “tens of thousands of combat troops” in Syria to challenge Iran or contribute to the country’s reconstruction process. + +Consistent with U.S. government practice, Itani ignored the nothing option. He is certainly right about the limitations of the small U.S. force presently in Syria. But even a much larger force would stand a poor chance of rolling back — or even containing — Iranian influence in the war-ravaged country. The reason for this is the very “imbalance” on which Itani’s argument against a small force rests. Syria is simply not central to U.S. interests and, as such, Iran’s cost tolerance will far outpace that of the United States at almost any level of conceivable commitment. + +Bashar al-Assad’s continued rule is obviously undesirable, not least for his wanton abuse of the Syrian people. Iran’s +================================================================================ +Rank = 83; Score = 272384.0 +<|begin_of_text|>Gunshots heard and clashes seen at Jerusalem's al-Aqsa mosque on Friday when Israeli forces stormed the compound. (Saeed Qaq) + +Clashes erupted in Jerusalem at the al-Aqsa mosque, known as the third-holiest site in Islam. + +Dozens gathered outside the mosque to protest Israel's ground invasion into Gaza, according to the International Business Times. Protesters then clashed with Israel Defense Forces after Friday prayers. + +According to Israel Police spokesman Micky Rosenfeld, following Friday prayers at the mosque compound, about 200 to 300 masked youths caused disturbances by throwing stones at policemen stationed nearby. After 20 minutes of disturbances, police entered the compound and arrested seven people for rioting. + +"We had been notified in advance that this was planned," said Rosenfeld, adding that Israel had imposed an age limit on people who could enter the compound, which Jews refer to as the Temple Mount and which Muslims call Haram al-Sharif, or holy sanctuary. Only men over age 50 and women were allowed in to pray on Friday, but Rosenfeld said that younger men had spent the night at the compound in preparation for the riots on Friday. + +According to reports, 100 Israelis stormed the courtyard and used rubber bullets to disperse crowds. + +RELATED: + +Live blog: Israel continues to invade Gaza + +Israel prepares to broaden Gaza offensive<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 84; Score = 272384.0 +<|begin_of_text|>On the right track: DfT appoints two women to directors general job share for Rail Group. Credit: Lynne Cameron/PA + +The Department for Transport has appointed two women to the civil service's first ever job share at director general level as successors to the DfT's newly appointed permanent secretary. + +Polly Payne and Ruth Hannant will join the DfT as directors general for the department's Rail Group on 11 December from the Department for Education, filling the position left vacant when Bernadette Kelly was promoted to perm sec in April. + +Kelly said the pair would be role models not only for the civil service but also for the rail sector, where women make up just 11% of the workforce and senior leaders are "overwhelmingly male". + +RELATED CONTENT + +She told Civil Service World: “I’m proud that the Department for Transport is leading the way with the first job share partnership at this level in government. And I’m especially pleased that we have Polly and Ruth joining us to lead our work on rail. + +"We are investing in the biggest rail modernisation for over a century, so there’s a huge amount to do." + +Kelly was DG for Rail Group from September 2015 to April 2017, when Nick Joyce took the post in an "acting" capacity. + +Rail Group is responsible for developing strategy and policy for the rail sector, managing expenditure on rail services and infrastructure, and oversees the delivery of rail franchises and major projects – including Crossrail and Thameslink. + +It sponsors British Transport Police, Network Rail, the Office of Rail and Road, the Rail Accidents Investigation Branch and Transport Focus. + +Payne and Hannant were already in a job-share partnership with each other working on higher education reform in DfE, and have both previously worked for the former Department for Business, Innovation and Skills and theTreasury. Former BIS permanent secretary Martin Donnelly described the pair as an "outstandingly effective job share", during an interview with Civil Service World earlier this year. + +DfT said the appointment demonstrated the government's commitment to increasing the number of women working in the rail sector. The Transport Infrastructure Skills Strategy, published in 2015, included an aim for women to make up 20% of DfT's science and technical apprenticeships by 2020, and 50% by 2030. + +Kelly added: “Within the DfT we are making progress in encouraging diversity at every level, and earlier this year the department was recognised for our leadership on workplace gender equality by +================================================================================ +Rank = 85; Score = 272384.0 +<|begin_of_text|>Google unlawfully used technology from Oracle, Microsoft and others when creating its Android and Chrome operating systems, leaving its vendor partners exposed. Rather than engaging in expensive and often drawn out lawsuits, a majority of Android vendors have signed licensing agreements with patent holders. Microsoft has already signed licensing agreements with more than 20 Android manufacturers, including big-name players such as HTC, Samsung and LG. The company claims that 80% of Android smartphones sold in the U.S. and most devices sold throughout the world are now covered under its various agreements. + +Geoff Duncan at DigitalTrends estimates that if Microsoft averaged a royalty of $1 per each Android device, it could generate roughly $430 million in revenue this year. Microsoft is estimated to collect somewhere near $8 per device, however, meaning its royalty fees could total $3.4 billion in 2013. + +As Android device shipments increase, so will Microsoft’s revenue. If half of the 1.5 billion Android devices that are estimated to ship worldwide in 2017 resulted in a royalty fee paid to Microsoft, the company could earn $5.9 billion dollars. Microsoft claims to earn royalties from a “majority” of Android vendors worldwide, though. Based on that information, if the company were to collect royalties from 75% of Android vendors in 2017, it could pull in an estimated $8.8 billion in additional revenue.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 86; Score = 268288.0 +<|begin_of_text|>Fox News viewers really do see the world differently. + +New data shows that Fox loyalists, when compared to the public at large, are far more pessimistic about America's future; are far more critical of President Obama's performance; are far more fearful of Hillary Clinton; and are more forgiving of Donald Trump. + +The sharp differences in opinion extend to beliefs about political corruption, voter fraud and media coordination with campaigns. Fox fans, when compared to fans of other networks, are far more likely to express concern about November's election results being manipulated. + +They are also more likely to agree with the sentiment that divisions in the United States are deeper than in the past. + +The data -- from a new national poll by Suffolk University -- shows deep divisions, indeed. And it demonstrates why 21st Century Fox patriarch Rupert Murdoch recently told the Wall Street Journal that it would be "business suicide" to change Fox's editorial direction. + +Related: Rupert Murdoch speaks out about contract talks with Megyn Kelly + +Overall, the poll finds that the country is split about Clinton, with 46% of all respondents having a favorable view of her versus 47% unfavorable. Among people who rate Fox News as their most-trusted source of news, however, sentiments are much more solidly anti-Clinton, with 84% viewing her unfavorably, versus just 13% favorably. + +Similarly, 54% of all Suffolk respondents approve of President Obama's job performance, versus 41% who disapprove. But among Fox loyalists, the numbers are radically different, with 16% approving of the president's performance and 80% disapproving. + +During the Obama presidency Fox News positioned itself as a voice of the opposition. Fox's most popular shows, like "The O'Reilly Factor," reinforced these sentiments. + +If Clinton is elected president, Fox's audience will expect more of the same. + +Fox News says it reaches many independents and some Democrats. But several surveys, including Suffolk's, shows that Fox's base is passionately pro-Republican, aligning with GOP positions and GOP candidates. + +Suffolk's polls ask many of the same questions that other pollsters ask during presidential election years -- but add a layer of questions about media consumption on top. + +The Suffolk pollsters ask: "What TV news or commentary source do you trust the most?" + +In the most recent poll, 270 of the 1,000 respondents said Fox -- the single highest result of any of the networks named in the poll. This reflects Fox's tight grip on conservatives. More than +================================================================================ +Rank = 87; Score = 266240.0 +<|begin_of_text|>President's smoking more likely to cause daughter's health problem than climate change. + +President Obama and daughter Malia last month. (Photo11: Butch Dill, AP) + +President Obama blames global warming for his daughter's asthma. Today that's politically useful spin, but the science says something different. If you're looking for a culprit, it just might be Malia's dad. + +In an interview Wednesday, in support of a new White House climate change awareness campaign, the president noted that his 16-year-old daughter had asthma when she was 4. He said that as a father, when your child says she has trouble breathing, "the fright you feel is terrible." Fortunately, doctors were able to treat Malia's condition quickly. + +The president connected his daughter's malady to global climate change. In a discussion Tuesday, he said "all of our families are going to be vulnerable" to global warming-induced health risks because "you can't cordon yourself off from air or from climate." + +A White House fact sheet connected the dots, saying that asthma rates have more than doubled in the past 30 years, and that "climate change is putting these individuals and many other vulnerable populations at greater risk of landing in the hospital" like Malia. + +The good news is that there is less reason for alarm than the White House suggests. The Environmental Protection Agency cautions that "outdoor air pollution and pollen may also worsen chronic respiratory diseases, such as asthma." Yet the EPA also reports that our air quality has substantially improved; aggregate emissions of common pollutants have decreased 62% between 1980 and 2013. It is unlikely that cleaner air is causing the increase in asthma. + +Whether there is a link between asthma and global warming, Malia herself hasn't really experienced much. The high school junior was born in 1998, when temperatures spiked. By some measurements, the world hasn't warmed significantly since then. + +Which brings us back to her father and his Marlboros. The president, who quit smoking years ago, has long kept his tobacco use out of doors. That's a common-sense tactic for folks who have trouble quitting. But sometimes, science can show that common sense has less sense than you think. + +Research funded by the National Institutes of Health has shown that smoking outside doesn't totally protect children from secondhand smoke. Even when smoking is done outside, nicotine in infants' hair is five times higher for babies with outside smoking parents than non-smoking parents. Smoking-related chemicals in infants' urine is seven times +================================================================================ +Rank = 88; Score = 266240.0 +<|begin_of_text|>Shoot lasers at the moon, solve Earth’s energy crisis. Boom. Done. Next global problem please, we’re on a roll. This is the statement I wish Japan’s Shimizu Corp. had released about their new energy plan, but alas they’ve simply just announced the details of a scheme to harvest solar power from panels on the moon. But, what a plan it is. + +Robots will build a belt of solar panels to encircle the moon. The panels will gather up energy from the sun, convert it to electricity and then channel the electricity by cable around to the Earth-facing side of the moon. From there the power will be zapped to large receivers on Earth’s surface using lasers. Why did it take so long for someone to propose this? + +The Luna Ring plan, which was introduced in Tokyo, is only the most recent in a long line of envelope pushing, seemingly kooky ideas to come from Shimizu Corp. But, the company seems serious about their plan proposing that the Luna Ring could begin construction by the year 2035. + +Judging from the amount of technology needed (a working robot colony on the moon?) and the financial drain attempting such a project would entail, we won’t be holding our breath for this one. Although, because the moon already drives ocean tides the idea of finding even more uses for the moon isn’t completely crazy. Shimizu’s plan might seem like something straight out of a comedy routine, but its not like they’re trying to blow up the moon or anything. Right? + +(via CNN Go)<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 89; Score = 266240.0 +<|begin_of_text|>Public officials and the relatively financially secure have long been guilty of turning the homeless and the poor into scapegoats — and making examples of the people perhaps in most need of compassion and help from their neighbors. It happens everywhere in America. And it's also become all too common in most Western countries around the world. + +Just last week in Florida, the Ft. Lauderdale City Commission made it a crime for anyone to sleep publicly downtown, or for panhandlers to approach cars at "busy intersections." According to the Sun Sentinel, the infractions are punishable by $500 and 60 days in jail. That's money and time many people just can't afford to spare. + +And earlier this year in London, one luxury apartment building installed so-called "anti-homeless spikes" near entranceways to keep the homeless from sleeping in those spaces. + +Anti homeless floor studs. So much for community spirit :( pic.twitter.com/Yz8VF7Ryid + +This is no way to treat people who are less fortunate. But more than that, it betrays a serious lack of empathy toward those who might be in need of shelter. + +More than 600,000 people in America are homeless. Ignorance and the perpetuation of a number of stereotypes has led to the creation of rigid categories of the haves and have nots. But it doesn't have to be that way. Much can be done to help ensure virtually no American has to forage for food or search/create their own shelter. Everyone can start by consciously challenging the deeply ingrained myths and questions people think about the homeless. + +1. Why can't you get a job like everyone else? + +For a job interview at a bank or even a burger joint, candidates need clean clothes, shoes in good condition, a shower and hygienic products so that they can appear as their best selves. Resumes, transportation and a stable telephone number are also needed. + +Even if they have marketable skills, most homeless and poor people are just trying to survive at the most basic level of Maslow's hierarchy of needs, a time-consuming search for food, water and shelter. + +And in an environment where jobs requiring relatively low amounts of skill aren't paid living wages — especially within the retail and fast-food industries — even homeless people who are able to secure an interview are likely to end up part of the "working poor," as chronicled by Barbara Ehrenreich in Nickled and Dimed. + +2. How do you not have housing by now if you've been panhandling? + +The idea that panhandlers +================================================================================ +Rank = 90; Score = 264192.0 +<|begin_of_text|>Cementing its reputation as a slow-motion corporate car-crash, made-up technology specialist Magic Leap has been sued for sex discrimination by the woman it hired to tackle sex discrimination. + +Tannen Campbell filed in a Florida court [PDF] this week accusing the virtual-reality startup of maintaining a "hostile environment" and a "macho bullying atmosphere." She claimed she was fired after she challenged CEO Rony Abovitz on the "depths of misogyny in Magic Leap's culture." + +As VP of strategic marketing and brand identity, Campbell says part of her job was to help the tech company with what it called its "pink/blue problem" (quick hint: not calling gender inequality a "pink/blue problem" is probably a good first step). + +The pinkness was also literally an issue: the all-male tech team (who were matched by the all-male advertising and executive teams) were asked to consider how to make their augmented reality headset more female friendly and, according to Campbell, the only solution they arrived at was to produce a version in pink. + +The lawsuit then offers a wide range of accusations of sexism, stemming from tetchy to highly personal to jaw-dropping. For starters, she took issue with the fact that the Florida company's jobs page was titled "Wizards wanted," seemingly shutting out women. + +She then goes on to slam the CFO for being "the kind of man who sits a little too close to women," and the VP of IT for being loud and making "misogynistic comments." She even turns on the company's most senior woman, its chief business officer, accusing her of being "the perfect Magic Leap female employee: she went to Harvard, never disagrees with Abovitz and always does Abovitz's bidding." + +Presentation + +Before you start writing off Campbell, however, she goes into some depth on the practices and approaches she recommended Magic Leap adopt in order to bring about a more balanced workplace, and they are good. Although some are perhaps more of a wish list – such as paying for a recruiter to help the partners of female hires relocated to Florida to find jobs as well to help ease the move to the company, and paying for female tech leaders including Sheryl Sandberg and Meg Whitman to attend company meetings as speakers. + +She developed a presentation for CEO Abovitz but her scheduled meeting to go over it was cancelled no fewer than six times. It took her eight months to get the meeting, and then he terminated the meeting about halfway through. Efforts +================================================================================ +Rank = 91; Score = 262144.0 +<|begin_of_text|>Women in headscarves and men in tatty clothes puff on a glass pipe as smoke swirls around their faces. The pictures published by Iranian media and blogs in recent months are a sign of a new drug epidemic: shishe, or methamphetamine. + +Shishe means “glass” in Farsi, a reference to the appearance of the drug in some of its purest forms. + +In less than a decade, methamphetamine use has skyrocketed in Iran to the point where now about 345,000 Iranians are considered addicts, according to official statistics. + +Seizures of methamphetamine soared 128 percent between 2008 and 2012, topping all other countries in the region, according to figures compiled by the United Nations Office on Drugs and Crime. Last year alone, the government of Iran confiscated 3.6 tons of shishe. + +A top official from the Iran Drug Control Headquarters said last year that shishe could be found in Tehran in “less than five minutes,” according to the Iranian Students’ News Agency. + +Shishe addicts in Iran are mostly urban, middle class and young, experts say. Notably, there are a large number of women who abuse shishe, too. + +One of the main reasons why shishe use has spread quickly in Iran is a lack of information about the drug, which has led casual users to believe, erroneously, that it is not addictive, experts say. + +Struggling university students have begun abusing it to stay up longer and try to boost their performance in school. Women have been sold the drug in beauty salons with the promise that it will help them lose weight, according to local media reports. + +“We really had a hard time convincing people that this is addiction,” said Azaraksh Mokri, a psychiatrist who teaches at the Tehran University of Medical Sciences and has dealt extensively with the issue of shishe addiction. + +Opium addiction has long been a problem in Iran partly because of a tolerance for its use even in conservative rural areas, and also because of the country’s long border with Afghanistan, for decades one of the top opium producers. Opium is still the most abused drug in Iran, according to official statistics. + +Shishe began to make inroads in the country about a decade ago, luring users who preferred its effects as a stimulant to the more soporific opium, which was seen as a drug of the poor and elderly. + +That shift has been characterized as a change between drugs which are known as sonati, or +================================================================================ +Rank = 92; Score = 262144.0 +<|begin_of_text|>Mother-son love dominates households the world over, but especially in India. (Ask anyone married to a desi man if she feels like the other woman.) One of the standout episodes of Room 104, the new Duplass-helmed HBO series, makes great use of this phenomenon. Titled “The Internet,” it’s set in 1997. Our only visible character, a harried, wannabe novelist, is locked for nearly the full half-hour in a phone call with his mom. Anish needs his manuscript sent to him, in a bad way. He’s got a life-changing meeting with an agent coming up, a book to finish, and of course, he forgot his laptop. Actor Karan Soni paces the dinky motel room for which the series is named in a sweat, walking his oblivious mom through the mechanics of email with a growing frustration in his voice that betrays just how deep their relationship is. + +Featuring only Soni and Poorna Jagannathan as his mother, the scene manages to do a lot with a little. From nearly her first line, Jagannathan establishes her character as one we haven’t seen rendered so well before: the loving but passive-aggressive, alternately icy, and helicoptering desi mom. Jagannathan’s character seems almost at times to be trolling Anish, so hypnotically layered is her cadence. To pull it off, Jagannathan told Vulture by email, she channeled her own, super-sweet mom, as well as her split attitude toward her 11-year-old son. When he talks, “There’s a part of you that is in wonder and amazement and so much love, and then there’s a part that’s totally bored and needs to just reach for the wine.” + +Both actors were born in India — Soni in New Delhi to a middle-class family, while Jagannathan is the daughter of a diplomat. Her accent gives away the upper-crust, Brit-lite India that she belongs to, though her character’s origins are notably humble. But that inconsistency doesn’t detract from her possession of the role. The Duplasses are known for encouraging improvisation, and she sprinkles her rejoinders with “beta” and “raja” — the latter, meaning prince, less known than the former, which means son. The episode — not written with Indians in mind — wasn’t meant to be a model of representational politics at work. Soni took the part a day before shooting, after the +================================================================================ +Rank = 93; Score = 262144.0 +<|begin_of_text|>Cultural criticism is taboo. In polite company, all cultures are created equally. None is better or worse than any other; they’re just different. We aren’t responsible for the culture we were born into, so there’s no objective basis for criticism or judgement. + +In progressive ideology, this idea goes a step further. Not only must we respect all cultural differences, we mustn’t stray outside the norms of the culture we were born into. A white woman with dreadlocks, for example, should respect black culture and shave her hair off, rather than “steal” a hairstyle from a different culture. They even have a special word for this grievance: cultural appropriation. + +I think cultural appropriation is a load of baloney, based on the most persistent errors in political/social thought: abstraction errors – misunderstanding the relationship between people and labels, between aggregates and concretes. These errors are not only imprecise, but they are counter-productive, divisive, and downright dangerous. + +Equal Equivocation + +The first abstraction error goes like this, “All differences between people are benign differences. Some people are born with light hair; others with dark hair. Neither is superior to the other. In the same way, all cultural differences are benign. Some cultures value monogamy; others are more sexually liberal. Neither is superior to the other.” + +This concept is applied across the board. Some cultures are more religious; some value education more highly; some are more hierarchical, etc. These differences should not be judged, any more than we should judge somebody for their height or the amount of freckles on their face. + +Then, the story goes, because all cultures are essentially equal, any differences in the socio-economic status of ethnic groups must be a function of discrimination. Without racism or discrimination, all cultures would be equally represented across the socio-economic spectrum. + +In reality, we’ve no reason to believe this is true. Nowhere in the world – nowhere in history – are all cultures represented equally across the socio-economic spectrum. The idea is an appealing, aesthetic one, no doubt, but it’s not grounded in the real world. + +Different cultures value different things; some skills are valued more highly than others; throughout the world, Chinese immigrants tend to have the highest average income of any demographic. Why is this? It’s not because they are genetically superior; it’s not because of pro-Chinese discrimination (in fact, it’s largely despite negative discrimination); it’s because Chinese culture heavily emphasizes academic performance in the hard sciences, and the hard +================================================================================ +Rank = 94; Score = 260096.0 +<|begin_of_text|>4K has arrived; 4K has a long way to go. The next standard in TV and monitor resolutions has started to trickle into electronics showrooms, hoping to tantalize shoppers into taking a very, very high-res plunge, but the resolution standard doesn't come with much to watch. + +Specs at a glance: Toshiba P50t-BST2N01 Screen 3840×2160 at 15.6" (282 ppi) OS Windows 8.1 64-bit CPU 2.4GHz Intel Core i7-4700HQ RAM 16GB 1600MHz DDR3L (two slots, 16GB max) GPU AMD Radeon R9 M265X with 2GB dedicated GDDR5 memory, Intel HD Graphics 4600 (integrated) HDD 1TB hybrid drive with 8GB NAND flash Networking Dual-band 802.11agn, Bluetooth 4.0 Optical Pre-built DVD SuperMulti drive (Blu-ray rewriteable drive available in alternate fixed configuration) Ports 4x USB 3.0, HDMI, card reader, headphone jack, microphone jack Size 14.9 × 9.6 × 1.1" Weight 5.2 lbs Battery 4-cell Li-polymer Warranty 1 year Starting price $1,799.99 Other perks Webcam, Technicolor specification, multi-touch display, Adobe Photoshop Lightroom 5 + +Just as 3D TVs and monitors had a tough enough time getting our attention, 4K is still hobbled by a severe lack of content in its 3840×2160 resolution. TV buyers may lose interest once they burn through the limited films and series available on smart sets' download and streaming services. Computer users, however, have had more reasons to tiptoe toward the quadrupling of 1080p—at the very least, to enjoy super-crisp text and details in their regular work and browsing. + +Displays as dense as 2880×1800 and 3200×1800 have already landed in our laps, but Toshiba has come out as the first legitimate 4K laptop producer. The P50t Satellite's biggest selling point is its 15.6", 3840x2160 resolution screen—a multi-touch panel, at that—and it's backed with the specs you'd expect from a machine forced to render so many pixels. + +Without a 4K media landscape to latch onto, however, Toshiba instead has to position this device as a creative +================================================================================ +Rank = 95; Score = 259072.0 +<|begin_of_text|>Class warfare is for real, and it’s time to pick a side. You could throw in with the chateau dwellers, but chances are you’re with everyone else. As the war of whines has progressed, not every 1%er is content falling into their allotted box, and some have stepped forward to prove they’re just regular folks like the rest of us. Here are their stories. + +Alec Baldwin + +Image Source + +How can you say you support the Occupy movement while you stay in Capital One Bank’s commercials? That’s like saying you support the Occupy movement and getting caught having sex with Mitt Romney. Baldwin went as far as to wax poetic on the Huffington Post about the Occupy movement and a DOCUMENTARY HE SAW that was about unemployed people in Long Island. He highly recommends it. Baldwin also called out the “excessive fees forced on customers by certain banks” (“come to Capital One, we’d never do that! We’re totally not one of the least-trusted banks in America.“) and pointed out that Occupy Wall Street “talks a lot, too much in fact, about One Percent versus Ninety Nine Percent.” That’s why we need you Alec, to show us into the light of Capital One bank. + +Jay-Z + +Image Source + +Even though the crack game got him started in the rap game, Jay-Z (Shawn Carter) has been living in the mogul world for quite some time. The man openly brags about having watches he hasn’t seen in months and an apartment at the Trump he only slept in once. So when he put out T-shirts through his Rocawear clothing line which read “Occupy All Streets”, the streets took offense. Especially when it was revealed that none of the proceeds from the shirts would go to the Occupy movement. Young Hov! + +Kanye West + +Image Source + +“Kanye West does not care about poor people,” – every person on earth. + +Kanye West is already pretty well liked in America, but his charitable ways know no bounds. Following in “Big Brother” Jay-Z’s footsteps, Kanye West also decided to show his support for the Occupy movement by showing up once for a few minutes. Reports from the protest described Kanye as uncharacteristically “demure” despite the authentic gold chains he wore, in an unintentionally hilarious show of support. He was also accompanied by another filthy rich hip-hop mogul, which brings us to… + +Russell Simmons + +Image Source + +Author of the grotesquely titled Super Rich: A Guide To Having It +================================================================================ +Rank = 96; Score = 259072.0 +<|begin_of_text|>Why are so few novelists from working class backgrounds? And why does it matter who gets to write novels and get them published? Author Kit de Waal investigates. + +"The more we reinforce the stereotypes of who writes and who reads, the more the notion of exclusivity is reinforced. It takes balls to gatecrash a party." + +Kit de Waal, published her first novel, My Name is Leon, in 2016 at the age of 55. She has already put her money where her mouth is - using part of the advance she received from Penguin to set up a creative writing scholarship in an attempt to improve working class representation in the arts. + +Kit knows that - as a writer from a working class background - the success of her debut novel is a rare occurrence. Born to a Caribbean bus driver father and an Irish mother (a cleaner, foster carer and auxiliary nurse), Kit grew up in Birmingham and left school at 15 with no qualifications. She became a secretary with the Crown Prosecution Service and went on to have a career in social services and criminal law. + +In this feature she explores an issue that is deeply personal to her. She looks back at her own life and trajectory, and takes the listener on a journey around the country to find out what the barriers really are to working class representation in British literature today. + +"There is a difference between working class stories and working class writers. Real equality is when working class writers can write about anything they like - an alien invasion, a nineteenth century courtesan, a medieval war. All we need is the space, the time to do it - oh yes, and some way to pay the bills!" + +Kit talks to a range of writers, agents and publishers about what the barriers are for writers from working class backgrounds, including Tim Lott, Andrew McMillan, Gena-mour Barrett, CEO of Penguin Random House UK Tom Weldon, Julia Bell, Julia Kingsford, Ben Gwalchmai, Nathan Connolly and Stephen Morrison-Burke (Birmingham poet laureate and the first recipient of the Kit de Waal scholarship). + +Produced by Mair Bosworth.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 97; Score = 259072.0 +<|begin_of_text|>MMA’s first openly trans athlete still encounters prejudice inside and outside the sport and says media are partly to blame: ‘We don’t tell people to write stories in respectful ways’ + +An athlete and a transgender person – that both identities could be conceivable fascinates our media landscape, from the New York Post all the way up to the New York Times. This is perhaps why speculation about Bruce Jenner’s gender identity has been disappointingly rife in recent weeks, even though the athlete has not spoken out about it. + +Fallon Fox knows what it’s like to be the focus of rumours and hearsay. She is a mixed martial arts (MMA) fighter – and the first openly trans athlete in the sport’s history. + +Fox is intimately familiar with our culture’s treatment of trans athletes. She was not out publicly when she began competing, although licensing commissions knew she was trans. But after Fox began winning, a reporter called her and made it clear he knew she was trans. Fox understood she had to get ahead of the story, so she came out in an Outsports article in March 2013. + +When I ask Fox about the obsessive interest in an athlete’s transition, she brings up the retrogressive idea that, despite the fact that there are actually many female athletes, sports are men’s domain. This leads people to see Jenner’s athletic abilities as a sign of maleness, making the possibility that Jenner might be trans a contradiction to gawk at. + +Fallon Fox: ‘I want to bring my people with me.’ Photograph: Rhys Harper + +I ask Fox about being put in the position of having to come out publicly so early in her career. “That sucked. I expected that someone was going to out me; you just can’t go through life with a microscope on your career without someone delving into your past a little bit,” she says. “But it’s something you really can’t prepare yourself for.” Her face grows serious. “The scope of anger and vitriol that I received initially... That was disheartening, tragic. It was mind-blowing.” + +This is not how Fox wanted her career to go. Her dream narrative would have been for her to go in there just like any other fighter would, collecting wins, hopefully ending up the best female fighter on the face of the planet. “I suppose I’d like all of that still,” she says pensively. “But in addition to that, I want to bring my people with me.” + +Fox’s interest in MMA stemmed from the fact that +================================================================================ +Rank = 98; Score = 258048.0 +<|begin_of_text|>Pray that ‘The Handmaid’s Tale’ stays fiction. Because a world where women are denied their humanity isn’t worth living in. + +Whenever I buy a book, I have this (good?) habit of writing my name and the date on the flyleaf. In the case of The Handmaid’s Tale, the fading scrawl on the yellowing pages in my copy tells me I bought it in February, 2003, and I clearly remember reading it with due alacrity, for I was told by people more well-read and aware than I that it was a must-read, science fiction or otherwise. A dystopian classic that’s spoken of in the same breath as Orwell’s 1984, Huxley’s Brave New World and Bradbury’s Fahrenheit 451, but with a predominant feminist theme. + +And I was hooked from the first page, and horrified as that bleak and terrifying world revealed itself, page by page. + +A world where women are denied their humanity, and their bodies are merely vessels for reproduction. Consent doesn’t come into the picture. A world where selective reading of holy books guides the government, and patriarchy is at the root of all laws. + +And that’s just the beginning. It’s a truly harrowing world that Atwood has created. + +I finished the book but it stayed with me, at least in its broad contours, as the years rolled by. Events took place in our world that kept taking me back to the world of The Handmaid’s Tale. The rise of the Taliban and their restrictions on women. Global warming, rapid climate change and toxic pollution. The Fukushima nuclear disaster. The death of Savita Halappanavar in the Republic of Ireland because she was denied an abortion under Catholic laws. The rise and rise of Islamophobia. The weaponisation of religion and misogyny. The fight for equal rights for women and marches for women’s rights. And more recently, the meteoric rise of conservatism in the West and Americans talking about fleeing to Canada. + +That’s why, for a book written in 1985, The Handmaid’s Tale feels eerily contemporary; nay, all too real and very ‘today’ (with my memory refreshed by a re-read this week). Perhaps not surprising when you come to know that while writing the book, Atwood made it a rule for herself to only include events that have happened in history and could — very plausibly — happen overnight within the ambit of existing laws and social mores, and for which the technology did not already exist. +================================================================================ +Rank = 99; Score = 257024.0 +<|begin_of_text|>Women who seek leadership roles in business often face the prospect, whether real or perceived, of having to choose between nurturing their careers or building a robust family and personal life. The publication in March of Lean In: Women, Work and the Will to Lead, by Sheryl Sandberg, Facebook’s chief operating officer, inspired a renewed debate about feminism. It suggested in a strong, if controversial, fashion that women can have it all. + +The confidence and authority that are required of leaders play a role in making choices about how to integrate work and other aspects of life, participants said at a panel titled, “How to be the Champion of Your Own Career,” at the recent Wharton Women in Business conference. Moderated by Arnold J. Rosoff, an emeritus professor of legal studies and business ethics at Wharton, the five-person panel was one of the conference’s first ever to include both men and women. + +Early in the discussion, J.J. Cutler, executive vice president of Aramark, invoked the title of the panel and sought to broaden the subject. “I would be very thoughtful about how you define ‘career,’” he said. “Instead of thinking about being a champion of a career, I’d think about just being a champion of your life. In particular, I would be very thoughtful — and it will change over time — about what role you want work and career to play in your overall life. + +“Be really flexible,” Cutler added. “Your life is likely going to get messier and more complicated, and the world is going to get more complicated from here on out.… Creating a vision for yourself is great, but I would be careful about [making] too many hard and fast rules and plans.” + +Creating a vision for yourself is great, but I would be careful about [making] too many hard and fast rules and plans.”— J.J. Cutler + +Picking the right partner, Cutler said, is “super important.” He noted that Sandberg may have raised fewer eyebrows than a man would have in choosing to discuss the role of a romantic partner in a businesswoman’s career. “The people I’ve observed who are happy, successful [and] fulfilled, oftentimes … [have someone] outside of work who, when good stuff happens, the good stuff is a lot better, and when bad stuff happens, you’ve got someone there who can help you deal with it,” he stated. “And there will be both. Of all the stuff in Lean In, I actually think diff --git a/examples/openwebtext/files/factor_arguments.json b/examples/openwebtext/files/factor_arguments.json new file mode 100644 index 0000000..f4877ac --- /dev/null +++ b/examples/openwebtext/files/factor_arguments.json @@ -0,0 +1,20 @@ +{ + "strategy": "ekfac", + "use_empirical_fisher": false, + "amp_dtype": "torch.bfloat16", + "amp_scale": 65536.0, + "has_shared_parameters": false, + "covariance_max_examples": 100000, + "covariance_data_partitions": 1, + "covariance_module_partitions": 2, + "activation_covariance_dtype": "torch.bfloat16", + "gradient_covariance_dtype": "torch.bfloat16", + "eigendecomposition_dtype": "torch.float64", + "lambda_max_examples": 100000, + "lambda_data_partitions": 1, + "lambda_module_partitions": 4, + "use_iterative_lambda_aggregation": true, + "offload_activations_to_cpu": true, + "per_sample_gradient_dtype": "torch.bfloat16", + "lambda_dtype": "torch.bfloat16" +} \ No newline at end of file diff --git a/examples/openwebtext/files/inflation.txt b/examples/openwebtext/files/inflation.txt new file mode 100644 index 0000000..89efa92 --- /dev/null +++ b/examples/openwebtext/files/inflation.txt @@ -0,0 +1,2112 @@ +Query Sequence: +Prompt: Inflation is typically measured by; Completion: the Consumer Price Index (CPI). + +Top Influential Sequences: +================================================================================ +Rank = 0; Score = 3899392.0 +<|begin_of_text|>WASHINGTON (Reuters) - Both President Barack Obama and Republicans in the U.S. House of Representatives have made opening offers in negotiations to resolve the “fiscal cliff. + +Here is a look at the two proposals, which are aimed at averting a more drastic combination of tax increases and spending cuts that economists say could cause a recession: + +REPUBLICAN OFFER + +House Republican leaders on Monday called for $2.2 trillion in new deficit reductions over 10 years. + +When counting deficit reductions enacted last year, anticipated savings from winding down the wars in Iraq and Afghanistan and some interest savings, the package would amount to $4.6 trillion in reductions over a decade, according to House Republicans. + +The offer made the following proposals to achieve $2.2 trillion in new deficit reductions over 10 years: + +* $800 billion in new revenue through tax reform; + +* unspecified healthcare program savings of $600 billion; + +* other savings from changes to unspecified mandatory spending programs of $300 billion; + +* tying cost-of-living increases for federal benefit programs to the Consumer Price Index to get savings of $200 billion; + +* and further unspecified savings to domestic spending programs of $300 billion. + +House Speaker John Boehner of Ohio and six other House Republican leaders made the offer on Monday in a letter to Obama. + +DEMOCRATIC OFFER + +The White House on Thursday proposed raising tax revenues by nearly $1.6 trillion, in line with what Obama has said is needed for long-term deficit reduction of nearly $4.4 trillion over 10 years. + +The administration also sought $200 billion in economic stimulus from a combination of investments including infrastructure spending, extension of a payroll tax cut and jobless benefits. + +The White House would also continue individual income tax cuts from the administration of former Republican President George W. Bush for all but the wealthiest earners. + +Obama’s negotiators also sought the ability to raise the nation’s borrowing limit unilaterally. At present, Congress must approve an increase in the debt ceiling. + +The administration’s proposal would delay across-the-board spending cuts for a year. In exchange the administration agreed to make $600 billion in spending cuts to entitlement programs.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 1; Score = 3309568.0 +<|begin_of_text|>India's inflation rate has slumped to the lowest level of modern times and there is much talk that the Reserve Bank of India will lower interest rates to stimulate the economy--that might take a little longer than many are predicting. We must always distinguish between general inflation, a general upward pressure on the price level, and specific price changes. This current low inflation rate looks much more like the effect of the latter, thus interest rate changes aren't so warranted. + +Growth in industrial production fell to a three-month low in May while consumer price index (CPI)-based inflation declined below a stipulated floor of 2 per cent in June, providing the Reserve Bank of India leeway to cut the policy interest rate in August. + +The RBI certainly has the power and the room to cut rates, yes, but that's not quite the correct policy decision here. Which is whether the RBI should cut interest rates as a result of the low inflation and that's much less certain: + +Data released by the Central Statistics Office (CSO) on Wednesday showed retail inflation, as measured by the consumer price index (CPI), rose an annual 1.5% in June, slower than previous month's 2.2%. This was the slowest pace of increase since the government unveiled the new retail inflation series in 2012. The previous lows were in 1999 and in 1978 under a different series. + +My point being that CPI isn't really the correct inflation measure to be using when considering interest rate changes and inflation. The US Federal Reserve, for example, uses the PCE rate although that's not really the important point here. CPI, PCE, RPI, they all have their slight differences and are useful in slightly different manners. But there's a much more important difference we need to take account of: + +Aside from the steep fall in food inflation, which has been downplayed by the RBI time and again, the steady and consistent fall in core inflation (non-food and fuel) could find favour with the central bank. + +It's this difference between core and non-core inflation that matters. It does depend upon which government and which statistical system we're talking about as to whether all produce core and non-core PCE and RPI and CPI, but in general the distinction is understood between the two different types of inflation rate: + +Data on Wednesday showed headline consumer price inflation fell to 1.5 percent in the year to June from an annual 2.2 percent a month ago and below forecasts for a 1.6 percent reading +================================================================================ +Rank = 2; Score = 3244032.0 +<|begin_of_text|>Inflation accounting comprises a range of accounting models designed to correct problems arising from historical cost accounting in the presence of high inflation and hyperinflation.[1] [2] For example, in countries experiencing hyperinflation the International Accounting Standards Board requires corporations to implement financial capital maintenance in units of constant purchasing power in terms of the monthly published Consumer Price Index. This does not result in capital maintenance in units of constant purchasing power since that can only be achieved in terms of a daily index. + +Historical cost basis in financial statements [ edit ] + +Fair value accounting (also called replacement cost accounting or current cost accounting) was widely used in the 19th and early 20th centuries, but historical cost accounting became more widespread after values overstated during the 1920s were reversed during the Great Depression of the 1930s. Most principles of historical cost accounting were developed after the Wall Street Crash of 1929, including the presumption of a stable currency.[3] + +Measuring unit principle [ edit ] + +Under a historical cost-based system of accounting, inflation leads to two basic problems. First, many of the historical numbers appearing on financial statements are not economically relevant because prices have changed since they were incurred. Second, since the numbers on financial statements represent dollars expended at different points of time and, in turn, embody different amounts of purchasing power, they are simply not additive. Hence, adding cash of $10,000 held on December 31, 2002, with $10,000 representing the cost of land acquired in 1955 (when the price level was significantly lower) is a dubious operation because of the significantly different amount of purchasing power represented by the two numbers.[4] + +By adding dollar amounts that represent different amounts of purchasing power, the resulting sum is misleading, as one would be adding 10,000 dollars to 10,000 Euros to get a total of 20,000. Likewise subtracting dollar amounts that represent different amounts of purchasing power may result in an apparent capital gain which is actually a capital loss. If a building purchased in 1970 for $20,000 is sold in 2006 for $200,000 when its replacement cost is $300,000, the apparent gain of $180,000 is illusory. + +Misleading reporting under historical cost accounting [ edit ] + +"In most countries, primary financial statements are prepared on the historical cost basis of accounting without regard either to changes in the general level of prices or to increases in specific prices of assets held, except to the extent that property, plant +================================================================================ +Rank = 3; Score = 3096576.0 +<|begin_of_text|>A Wall St. sign is seen outside the entrance of NYSE Thomson Reuters By Tanya Agrawal + +(Reuters) - U.S. stock index futures fell on Wednesday as Chinese stocks had another roller coaster ride and as investors await the minutes from last month's Federal Reserve meeting for clues on when interest rates will be increased. + +* While the health of the U.S. economy appears to be stabilizing, the effect of the yuan devaluation and other macro factors are playing on investor's minds. The Fed minutes will be released at 2 p.m. ET. + +* Economists believe the Fed will probably raise rates twice this year, with the first hike coming in September. Investors are still not fully convinced of a September hike, but most are betting a rate hike will occur by the end of year. + +* Chinese stocks reversed sharp declines and ended higher after the central bank injected more funds into the financial system for a second day in a bid to calm panicky markets. + +* The People's Bank of China devalued the yuan last week, triggering an avalanche of selling by investors globally who feared Beijing wanted to engineer a much sharper decline to support weak exports. + +* The Chinese market gyrations kept commodity prices under pressure, with oil and copper near six-year lows. + +* Other data due Wednesday is expected to show consumer prices rose 0.2 percent in July, less than the 0.3 percent rise in June. The Consumer Price Index data is due at 8:30 a.m. ET. + +* Lowe's shares fell 1.6 percent to $71.85 in premarket trading after the No.2 U.S. home improvement chain's quarterly profit missed expectations. + +* Yum Brands rose 1.5 percent to $85.50, a day after the owner of the KFC and Pizza Hut chains announced new leadership for its China division as activist investors lobby the company to spin off that business. + +* Staples fell 1.6 percent to $13.93 after the office supplies retailer reported quarterly revenue slightly below analysts' estimates, hurt by a stronger dollar. + +Futures snapshot at 7:07 a.m. ET: + +* S&P 500 e-minis were down 5.25 points, or 0.25 percent, with 143,877 contracts traded. + +* Nasdaq 100 e-minis were down 11 points, or 0.24 percent, on volume of 23,476 contracts. + +* Dow e-minis <1YMc1> were down 57 points, +================================================================================ +Rank = 4; Score = 2899968.0 +<|begin_of_text|>Cryptography’s most familiar application is perhaps the sending of coded messages: Alice wants to communicate some information to her distant confidante Bob. She worries that her message might be intercepted by an eavesdropper, so she encrypts it in a way that only Bob can decipher. But there exists a whole range of complementary cryptographic tasks—such as online auctions and secure voting—in which Alice and Bob want to guard against attacks not from eavesdroppers but from each other. + +To develop algorithms to perform those other tasks, cryptographers use a building block called bit commitment: Alice chooses a bit (a 1 or 0) to be revealed later; Bob wants to ensure that Alice can’t change her mind in the meantime, while Alice wants to ensure that Bob has no way to learn which bit she chose until the appointed time. + +Although quantum theory is the basis for perfectly secure encryption, it offers no path to secure bit commitment. But a different law of physics, the impossibility of faster-than-light signaling, has now come to the rescue, and Anthony Martin, Hugo Zbinden, and their colleagues at the University of Geneva have implemented a protocol for achieving bit commitment for a full 24 hours. + +For the protocol to work, Alice and Bob work with trusted partners Amy and Brian, respectively. Bob and Brian take turns exchanging data with Alice and Amy, in such a way that each exchange falls outside the forward light cone of the previous one. Each of Alice’s and Amy’s answers depends on both the data they receive from Bob or Brian and on information Alice and Amy agree on in advance, including the bit they want to commit. For Alice to cheat and change the bit at any time in the middle of the protocol, she’d need to know what happened in Amy’s most recent exchange with Brian; relativity prevents her from having that information. + +In their 24-hour implementation, Martin and colleagues repeated the exchanges for 5 billion rounds, one every 17 µs, with a total of 162 GB of data. According to two theory papers from the past year, the probability that the protocol is vulnerable to cheating is less than one in a billion. (E. Verbanis et al., Phys. Rev. Lett. 117, 140506, 2016.)<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 5; Score = 2637824.0 +<|begin_of_text|>Intelligence, cognitive ability or cognitive performance is usually measured by a battery of tests that aim to quantify skills such as memory and analytical ability. There is loads of variation between people in how they perform on such tests, and these differences can be due to genetic and environment factors, and their interplay. + +In research published in the Proceedings of the National Academy of Science (PNAS) today, we show three genetic variants in humans that can account for a couple of IQ points – but before you get excited, these are only three variants out of likely thousands. + +The genetics of cognitive performance + +While a measure of “intelligence” can be controversial, cognitive performance scores are widely used because of their predictive ability. Educational attainment, income, job performance and health are all correlated with cognitive performance. + +By comparing the cognitive performance between family members, including comparisons between identical and non-identical twins, scientists are able to quantify the contribution of genetic and environmental causes of individual differences. + +Decades of research have shown that genetic factors account for about half of the causes of individual difference in cognitive performance, and recent studies using unrelated people have confirmed that a substantial proportion of individual difference is due to genetic factors. + +So, we know now that cognitive performance is heritable, but where are the genes? Despite considerable attempts to find genes for cognitive performance, no specific genes had been found and replicated. + +One reason for this puzzle is that there are a lot of genes involved – thousands, even – and their individual gene effect sizes are tiny. Past studies couldn’t find them because sample sizes were not large enough to detect genes with statistical significance. + +So how did we overcome this problem? + +Last year, a huge international collaborative study of more than 126,000 people correlated millions of genetic variants with educational attainment and discovered three genetic variants associated with it. + +Education attainment is correlated with cognitive performance, so given these two observations, we tested the genetic variants for education attainment with their associations with cognitive performance, which we report in PNAS today. + +We tested 69 genetic variants from the educational attainment study (of almost 107,000 people) in independent samples of 24,000 people who had a cognitive performance score. This two-stage strategy is called a “proxy-phenotype method” since educational attainment is a proxy phenotype (an observable characteristic or trait) for cognitive performance. + +The essence of this design was to piggy-back on a much larger study from a correlated trait (educational attainment) to pre-select a small number of genetic variants. These were then tested for association with cognitive performance – +================================================================================ +Rank = 6; Score = 2506752.0 +<|begin_of_text|>Colleges need to adapt so that university education doesn't become too expensive for all. + +University life. (Photo: M. Spencer Green, AP) Story Highlights Administrative bloat at American colleges and universities is out of hand. + +One calculation shows college tuition increased from 1978 to 2011 at an annual rate of 7.45%. + +Now colleges and universities are actually facing declining enrollments, and even credit downgrades. + +"Why am I paying so much tuition to people whose job seems to be telling me to call someone else?" + +That was my daughter's lament last week as she tried to pry an essential form out of her college's labyrinthine bureaucracy, but it's a question that many Americans should be asking. Administrative bloat at American colleges and universities is out of hand, and it's probably the biggest cause of the skyrocketing tuitions that afflict students and parents today. + +Everyone knows that tuitions have skyrocketed, though many may not appreciate the full extent of the problem. As University of Michigan economics and finance professor Mark Perry has calculated, college tuition increased from 1978 to 2011 at an annual rate of 7.45%. That far outpaced health-care costs, which increased by 5.8%, and housing, which, notwithstanding the bubble, increased at 4.3%. Family incomes, on the other hand, barely kept up with the Consumer Price Index, which grew at an annual rate of 3.8%. + +That has led many students (and sometimes their co-signing parents) into a nightmare of debt, as student loans have been used to fill the gap. Now colleges and universities are actually facing declining enrollments, and in some cases even credit downgrades as their ability to endlessly raise tuition comes into question. But why has tuition risen so fast? + +Some commentators blame lazy, overpaid faculty. But while faculty teaching loads are somewhat lower than they were decades ago, faculty-student ratios have been quite stable over the past several decades, while the ratio of administrators and staff to students has become much less favorable. In his book on administrative bloat, The Fall Of The Faculty, Johns Hopkins professor Benjamin Ginsberg reports that although student-faculty ratios fell slightly between 1975 and 2005, from 16-to-1 to 15-to-1, the student-to-administrator ratio fell from 84-to-1 to 68-to-1, and the student-to-professional-staff ratio fell from 50-to-1 to 21-to-1. Gins +================================================================================ +Rank = 7; Score = 2342912.0 +<|begin_of_text|>Students carry signs to protest Act 10 in February of 2011 near D.C. Everest High School in Weston. Daily Herald Media file photo Chad Dally/Wausau Daily HeraldD.C. Everest High School students, from left, Candace Perria, Nick Marvin, Emily Wurzer and Ashley Rawlings carry signs to protest Gov. Scott Walker's move to alter Wisconsin's union rights. The action at D.C. Everest was part of a mass protest at school districts throughout the area, including Merrill, Wausau and Mosinee. (Photo: File) + +The Fond du Lac County Finance Committee is considering a proposal that would provide a workaround of the restrictions of the controversial collective bargaining law Act 10 for members of a public employees union. + +County Human Resources Director Michael Marx sent a memorandum to the county's Finance Committee in May explaining the limitations the law placed on the AFSCME local 1366E Social Services union. + +Act 10 removed the ability of public employees unions to negotiate for anything but a pay increase limited to the Consumer Price Index. The county board approved a 1.57-percent increase for the union's members, which is the maximum allowed under the law this year. + +Other non-union county employees received a 2-percent increase without similar restrictions at the beginning of the year. + +"In other words, this group is being penalized by the state because they didn't decertify their union," Board Chairman Martin Farrell said. + +Marx recommended the board approve a supplemental pay increase for the more than 70 employees affected to make up for the difference. The extra pay would work like a bonus, being paid once in a lump sum in the first pay period of December. + +The extra pay would also be based on the number of hours worked, ranging from 7 to 10 cents, depending on the employee's current wages, and would be increased by 1.5 times for overtime. Marx said the payments would amount to between $140 to $190, depending on the employee's wages. + +Next year's supplemental pay would have to be reapproved by the board, and every year after, and can't be a part of the union's next contract negotiations. That's because the supplemental pay can't be part of the union's contact under Act 10 and is entirely at the discretion of the board, Marx said. + +"This is just the least problematic way of getting the increase to the employees without creating extra coding in our system," County Executive Al Buechel said. "When you look at personnel administration +================================================================================ +Rank = 8; Score = 2277376.0 +<|begin_of_text|>Introduction A carbon footprint is the measure of the amount of greenhouse gases (Carbon footprint), measured in units of carbon dioxide, produced by human activities. A carbon footprint can be measured for an individual or an organization, and is typically given in tons of CO 2 -equivalent (CO 2 -eq) per year. For example, the average North American generates about 20 tons of CO 2 -eq each year. The global average carbon footprint is about 4 tons of CO 2 -eq per year (Figure 1). + +Introduction A carbon footprint is the measure of the amount of greenhouse gases (Carbon footprint), measured in units of carbon dioxide, produced by human activities. A carbon footprint can be measured for an individual or an organization, and is typically given in tons of CO 2 -equivalent (CO 2 -eq) per year. For example, the average North American generates about 20 tons of CO 2 -eq each year. The global average carbon footprint is about 4 tons of CO 2 -eq per year (Figure 1). + +Figure 1. National carbon dioxide (CO2) emissions per capita. (2005). (Source: UNEP/GRID-Arendal Maps and Graphics Library) + +An individual’s or organization’s carbon footprint can be broken down into the primary and secondary footprints. The primary footprint is the sum of direct emissions of greenhouse gases from the burning of fossil fuels for energy consumption and transportation. More fuel-efficient cars have a smaller primary footprint, as do energy-efficient light bulbs in your home or office. Worldwide, 82% of anthropogenic greenhouse gas emissions are in the form of CO 2 from fossil fuel combustion (Figure 2). + +The secondary footprint is the sum of indirect emissions of greenhouse gases during the lifecycle of products used by an individual or organization. For example, the greenhouse gases emitted during the production of plastic for water bottles, as well as the energy used to transport the water, contributes to the secondary carbon footprint. Products with more packaging will generally have a larger secondary footprint than products with a minimal amount of packaging. + +Greenhouse gases and the greenhouse effect + +Figure 2. Anthropogenic greenhouse gas emissions. (Source: Energy Information Administration) + +Although carbon footprints are reported in annual tons of CO 2 emissions, they actually are a measure of total greenhouse gas emissions. A greenhouse gas is any gas that traps heat in the atmosphere through the greenhouse effect. Because of the presence of greenhouse gases in our atmosphere the average temperature of the Earth +================================================================================ +Rank = 9; Score = 2195456.0 +<|begin_of_text|>A teen invites “Destiny” to the prom using the face of a cliff in Black Cliffs east of Boise. (Patrick Orr/Ada County Sheriff’s Office via AP) + +Jay L Zagorsky is an economist at Ohio State University. + +Prom season is now upon us, the key social event of most people’s high school experience. My youngest child is preparing to attend her first senior prom and a steady supply of fancy dresses has begun to appear in our house. As my credit card bills climbed ever higher, I began to wonder, how has the price of the prom changed over time? + +For teenagers scraping together money from minimum-wage jobs and informal work like babysitting and snow shoveling, the question has immediate relevance because the price determines if they can afford to go. For parents of teens, the question is relevant because the prom is another expensive bill competing with lots of other financial priorities. + +Even for people not directly connected to the high school experience, the price of the prom is important. The U.S. is a melting pot of different cultures, religions and ethnic groups. Tying these disparate groups together is a variety of common shared experiences, such as celebrating the 4th of July, eating turkey at Thanksgiving and, for teens, experiencing the prom. + +When the prom’s price goes down, more teens can afford to go, and it becomes a more inclusive shared experience. When the price rises, however, the prom becomes a more exclusive social event attended mostly by those with wealthier parents, resulting in fewer common shared experiences. + +Cereal and donuts + +It is possible to get a sense of the changing price of going to the prom using government data. The Bureau of Labor Statistics provides detailed monthly information on the trend in prices over time for a vast number of items purchased by the typical family. + +The government does not care, directly, about the price of the prom – unfortunately it has yet to craft a “prom index.” However, using a subset of the inflation data we can see that proms are actually getting relatively cheaper over time. + +The typical 17-year-old going to the prom this year was born in 1998. Since then, the Consumer Price Index, or CPI, has increased 45 percent. That means the average item a family bought for $1 in 1998 now costs $1.45. Each month the CPI releases hundreds of detailed price indexes so that anyone can see, for example, if the price of breakfast cereal is rising faster than, say, donuts over time. + +The Prom Price Index +================================================================================ +Rank = 10; Score = 2195456.0 +<|begin_of_text|>This article is about the medical term. For the stock market term, see overweight (stock market) + +Overweight The overweight range according to the body mass index (BMI) is the area on the chart where BMI > 25 Specialty Endocrinology + +Being overweight or fat is having more body fat than is optimally healthy. Being overweight is especially common where food supplies are plentiful and lifestyles are sedentary. + +As of 2003, excess weight reached epidemic proportions globally, with more than 1 billion adults being either overweight or obese.[1] In 2013 this increased to more than 2 billion.[2] Increases have been observed across all age groups. + +A healthy body requires a minimum amount of fat for proper functioning of the hormonal, reproductive, and immune systems, as thermal insulation, as shock absorption for sensitive areas, and as energy for future use. But the accumulation of too much storage fat can impair movement, flexibility, and alter the appearance of the body. + +Classification + +The degree to which a person is overweight is generally described by the body mass index (BMI). Overweight is defined as a BMI of 25 or more, thus it includes pre-obesity defined as a BMI between 25 and 30 and obesity as defined by a BMI of 30 or more.[3][4] Pre-obese and overweight however are often used interchangeably, thus giving overweight a common definition of a BMI of between 25–30. There are, however, several other common ways to measure the amount of adiposity or fat present in an individual's body. + +Body mass index + +The body mass index (BMI) is a measure of a person's weight taking into account their height. It is given by the following formula: BMI equals a person's weight (mass) in kilograms divided by the square of the person's height in metres. The units therefore are kg/m2 but BMI measures are typically used and written without units. + +BMI provides a significantly more accurate representation of body fat content than simply measuring a person's weight. It is only moderately correlated with both body fat percentage and body fat mass (R2 of 0.68).[5] It does not take into account certain factors such as pregnancy or bodybuilding; however, the BMI is an accurate reflection of fat percentage in the majority of the adult population. + +The body volume index (BVI) was devised in 2000 as a computer, rather than manual, measurement of the human body for obesity and an alternative to the BMI + +BVI uses 3D software +================================================================================ +Rank = 11; Score = 2113536.0 +<|begin_of_text|>Experts have roundly criticized BC Premier Christy Clark’s recent home ownership grant policy. A key part of the negative reaction has been based on fears that interest free grants will increase housing prices and drive a further wedge between incomes and housing costs, a divide already plaguing the Vancouver and lower mainland markets. + +These worries about the divide between housing prices and incomes are well founded. + +Trends in residential property prices and housing affordability between 2005 and 2016 paint a grim picture in both Vancouver and the city of Toronto. This article looks at those trends, analyzing the mortgage cost of homes in these two cities and in comparison to Canada more broadly. + +More specifically, it looks at the number of weeks of labour time it would take workers making the average weekly wage to make a mortgage purchase in each city. The data are startling. + +Methodology: + +The nominal dollar cost of a house or a condo today is much higher than it was 10 or 20 years ago. Indeed, it is even higher than what it was just a couple of years ago. There exists a number of indices that look at the price of housing by deflating the nominal dollar price of a house by the consumer price index (CPI) to get an idea of how fast housing prices are rising relative to the general rise in prices of consumer goods. + +This article does something different. It calculates how many weeks of labour time, paid at average provincial weekly rates, it takes to purchase an average residential property. In order to do so, it makes some key assumptions. + +The calculations in this article concern changes in the mortgage costs of residential property purchases. Here, I assume that the entire price of the properties (inclusive of mortgage fees and financing costs) were paid through a mortgage financed with 5-year mortgage loans amortized over 15 years. In other words, it assumes no initial down payment. By looking at the weeks of labour time, the analysis assumes that all of a person’s income is going towards financing their mortgage. Of course, we know people have many other fixed costs and financial responsibilities. But this measure gives us an informative picture of the glaring gap between incomes and housing prices. + +Of course, since we don’t know what interest rates and wage rates will be after 2016, the numbers for the years after 2011 are estimates based on an extrapolation of recent data for these two variables. The calculation also assume, after 2011, a return towards trend values in the case of interest rates). Finally, the data is based on yearly averages. So +================================================================================ +Rank = 12; Score = 1941504.0 +<|begin_of_text|>Russia's Consumer Price Index (CPI) in October fell again to an extraordinary low of 2.7% year-on-year down from the already low 3% seen in September, Rosstat statistics agency said on November 7. + +Inflation has strongly overshooting the Central Bank of Russia (CBR) target of 4% set for 2017, diving from double-digit price growth in less than two years. + +However, the CBR keeps pointing to stubborn inflationary expectations and other medium-terms inflationary risks and sticks to the moderately tight monetary policy despite the pressure from the government to support growth with lower interest. The current rate is by far the lowest in post-Soviet history but the population remain stubbornly convinced that the fall is temporary, according to the CBR surveys. The result is inflationary pressure remains lurking and the CBR is reluctant to accelerate rate cuts until the population expectations of price rises relent. + +The CBR decided to cut the key interest rate by 25bp at the board of directors meeting on October 27, in line with the expectations. + +A more aggressive cut of 50bp would have meant that the CBR could prepare to tighten the policy again in the beginning of 2018, some analysts suggested. However, the CBR decided to abstain from an aggressive move and cut the rate by the minimum step of 25bp to 8.25%. + +Last week Alfa Bank argued that "at the moment we do not agree with the government’s concerns that low inflation is the result of the tight monetary policy of the CBR." + +The analysts that forecasted inflation easing to 2.8% in October maintained that "medium-term inflationary risks are considerable and that the CBR should thus not accelerate interest rate cuts." + +However, VTB Capital previously argued that "the incoming data for October and November might nudge the Board to opt for a more ambitious cut of 50bp," while noting that "this would call for surprises from inflation reports, banking stats and economic activity data to combine into a convincing case." + +The CBR expects inflation to end 2017 at 3.5-3.8%, while the Ministry of Economic Development that needs lower interest rates to help its ambitious 2.1% growth outlook expects 2.7-3.2% inflation in 2017.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 13; Score = 1703936.0 +<|begin_of_text|>New Delhi: Showing signs of recovery, industrial output rose to nearly 3-year high of 6.4 percent in August on improvement in manufacturing and capital goods while retail inflation remained with in comfort zone. + +The manufacturing sector, which constitutes over 75 percent of the index, grew by 6.9 percent in August, while the capital goods output rose by 21.8 percent in the month indicating a recovery. + +ALSO READ: Industrial production grows at a near 3-year high of 6.4% in August + +Commenting on the twin macro economic data, Chief Economic Advisor Arvind Subramanian said, "Encouraging news on Indian economy. IIP growth increased to 6.4 percent, consistent with indirect tax revenue performance. Core inflation moderate." + +The previous high of factory output growth was recorded in October 2012 at 8.4 percent. The factory output had grown by 0.5 percent in August last year. + +Industrial output, measured in terms of the Index of Industrial Production (IIP), was at 4.1 percent in the April-August period against 3 percent in the year-ago period, the data released by the Central Statistics Office (CSO) today showed. + +As regard retail inflation, the Consumer Price Index rose to 4.41 percent in September from 3.74 percent in August this year but remained in comfort zone. RBI has projected 5.8 percent retail inflation by January, 2016. + +ALSO READ: Retail inflation rises to 4.41% in September on dearer food items + +"The growth in manufacturing seems to be accelerating and we are hopeful of higher growth in the coming months.. Government?s efforts to revive manufacturing has started yielding results," said FICCI Secretary General A Didar Singh. + +Chief Economist, India Ratings & Research Devendra Kumar Pant said, "Sharp fall in inflation and monetary easing has increased demand for consumer durable. This is partially due to initiation of festival season. Bond market is likely to respond favourably to IIP data and bond yields are likely to fall by around 5 basis points in tomorrow's trade." I + +Meanwhile, the IIP growth for July has been revised slightly downwards to 4.1 percent from provisional estimate of 4.2 percent last month. + +Commenting on the data, Assocham President Rana Kapoor + +said, "..Pleased to see finally the green shoots of economic activity getting converted into strong industrial growth figures." + +"On the whole, these estimates put forward certain +================================================================================ +Rank = 14; Score = 1581056.0 +<|begin_of_text|>The Bank of Canada will publish the results of its recent review of interest rates on Wednesday, and there is plenty of speculation Governor Stephen Poloz will use the opportunity to raise the bank’s overnight rate, which would increase costs for everything from mortgages to business loans. The speculation is largely based on strong GDP performance for the first quarter of 2017, though one good quarter does not a year make. + +Actually, with the interest rate at a historic low of 0.5%—pedal to the metal for the BoC—it is surprising how slowly the economy is moving. The bank is supposed to be able to juice growth by dropping rates. That’s clearly not happening, here or in other parts of the world. Negative interest rates in Europe have been equally ineffective at improving GDP. + +However, the Bank of Canada doesn’t actually target economic growth; it targets inflation and nothing more. It wants inflation to run at between 1% and 3%,with all other concerns—from ballooning household debt to employment rates—being secondary to this overriding concern. The bank uses three special inflation measures (CPI-Common, CPI-Median and CPI-trim, which you can read all about here) that attempt to remove the more volatile aspects of the Total Consumer Price Index. I’ve averaged all three together in the chart below to simplify things. + +But here’s the trouble. Despite a decent first quarter of GDP growth, the Bank of Canada’s preferred inflation measures are averaging an 18-year low as of May 2017! You’d have to go back to March 1999 to get a lower average inflation rate. Ever since 2010, the bank’s key inflation measures have been nowhere near 3%, or the top of their stated range. In fact, the average inflation rate since 2010 has been only 1.6%, much closer to the bottom of the range than the top. The bank’s inflation measures have been falling steadily over the past 12 months. If anything, the rate of inflation is too low. + +Source; Bank of Canada and author’s calculations + +But there’s a hidden issue afoot here. Inflation is a sanitized way of talking about wage gains, which are one of its primary components. When the Bank of Canada raises interest rates one of the primary impacts is that it reduces wage gains for working Canadians. As rates rise, businesses cut back, lay some people off and make workers afraid of asking for a raise, hence less inflation. The Bank of Canada inflation range is in effect a ceiling +================================================================================ +Rank = 15; Score = 1564672.0 +<|begin_of_text|>LONDON (Reuters) - British inflation hit its highest level since September 2013 last month, extending its sharp rise since the vote to leave the European Union and tightening the squeeze on living costs as a national election approaches. + +Consumer prices rose by an annual 2.7 percent, data showed, and they look set to rise further due to the fall in the value of the pound and the recent rise in global oil prices. + +Britain’s economy was barely ruffled last year by the shock vote to leave the EU. But the steady rise in inflation since then, combined with weak wage growth, has slowed its momentum this year. + +The opposition Labour Party on Tuesday sought to highlight rising costs for voters as it launched its policy proposals for the June 8 election, pledging a higher minimum wage and state involvement in the energy sector to keep prices down. + +Last week, Bank of England (BoE) Governor Mark Carney warned 2017 will be challenging for consumers, with inflation now almost certain to overtake wage growth. + +“The last thing Britain needs is another real wage slump. But rising prices are hammering pay packets,” Trades Union Congress General Secretary Frances O’Grady said. + +Prime Minister Theresa May called the snap election last month to strengthen her mandate to negotiate Britain’s exit from the European Union. But living standards are a big campaign issue, pushing her party into a promise to cap energy prices that breaks with its usually pro-market agenda. + +Despite the rise in inflation, however, the economy is far from overheating, and all but one of the BoE’s eight policymakers voted last week to keep interest rates on hold. + +Prices are displayed on a store window in London, Britain May 16, 2017. REUTERS/Neil Hall + +AIRFARES SWING IN EASTER + +The latest inflation figures were boosted most of all by rising airfares during the Easter holidays which last year took place in March. Rising clothing prices, higher car tax and electricity also pushed up consumer prices. + +Many economists see more inflation ahead. + +“We remain convinced that the market is underestimating the further upside for inflation from here,” Scotiabank analyst Alan Clarke said. He expects utility bills, food costs and the weak pound to put more pressure on prices in future. + +Sterling briefly spiked to its highest in almost a week against the dollar before falling back. GBP=D4 + +Many economists say the impact of the fall in sterling on consumer prices will be felt more strongly in the coming months, and the central bank expects inflation to peak at nearly 3 +================================================================================ +Rank = 16; Score = 1507328.0 +<|begin_of_text|>Image caption UK consumers have seen food prices rise 40% since 2007, says the Ernst & Young Item club report + +High inflation has cost the UK economy £10bn over the last three years, says an influential report. + +And with inflation averaging 3.5%, instead of the government's target 2% rate, high inflation will remain "a permanent fixture", says the Ernst & Young Item Club. + +This has had a "corrosive impact on the UK economy", the report concludes, as household spending power has shrunk. + +The group does not expect inflation to dip below 2.5% before 2017. + +Consumers have been struggling to cope with food prices that have risen 40% since 2007, as well as rising fuel and education costs. The UK High Street has suffered as a result. + +Despite the economic impact, Carol Astorri, the Club's senior economic adviser said: "It could have been worse. Our modelling shows the [Bank of England's Monetary Policy Committee] were right to stick to their guns, allowing inflation to overshoot and avoid tightening monetary policy. + +"The alternative scenario would have seen interest rates rise by 3.5% in 2011, choking off the recovery even earlier and adding an additional 625,000 people to UK dole queues." + +The report forecasts that the consumer prices index will rise to 3% over the summer, but fall back to 2.5% by the autumn, as energy bills and food prices stabilise. + +Mark Gregory, Ernst & Young's chief economist, said: "With consumer spending continuing to be curbed by rising costs and only set to improve gradually, retailers will need to battle harder than ever to win the war for our wallets." + +When Mark Carney, the new governor of the Bank of England, takes over from Sir Mervyn King in July, he will face the difficult task of keeping inflation under control while also stimulating the faltering UK economy.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 17; Score = 1499136.0 +<|begin_of_text|>A deflationary Bacon Cheeseburger Index combined with optimistic Google searches for timeshares reflect a U.S. economy in flux — and, likely, a Federal Reserve in flux. + +These readings, plus snapshots of car buying, gun ownership, food-stamp demand and more, make up the “Off the Grid Indicators” report that creator Nicholas Colas, chief market strategist at the New York–based global brokerage Convergex, argues pushes economic sleuthing beyond the obvious payrolls report or consumer price index. + +Read: U.S. adds 215,000 new jobs in March as more workers enter labor force + +Government data often strip out volatile food prices, but market strategist Nicholas Colas and company say doing so miscalculates inflation expectations among shoppers. + +Highlights from the current edition of the now five-year-old quarterly report, especially when combined with traditional government- and industry-issued data and anecdotes, indicate the Fed can be slow with raising interest rates, Colas said. The Fed’s next two meetings come later this month and then in June. + +Read: Fed will likely hike rates in June in wake of jobs report, economists say + +Convergex even trademarked its “Bacon Cheeseburger Index,” an evenly split minibasket of the popular beef-bacon-and-cheese combo that can serve as a relatable inflation gauge. Government data often strip out volatile food prices, but Colas and company say doing so miscalculates inflation expectations among food-conscious shoppers. + +As of the first quarter, thanks to price declines in all three of these cholesterol commodities, a bacon cheeseburger now costs 5.1% less than a year ago, the Convergex report shows. + +A look back at the Bacon Cheeseburger Index to 1990 finds that it is actually a decent indicator of deflation risk. Prior periods when the BCI turned resoundingly negative (3% or more) since 1990 include: 2009 (financial crisis); 1998 (emerging market and Long Term Capital collapse), and 1992 (the lead up to the Iraq invasion). In each case, the Fed was cutting interest rates, not raising them. + +“So should the Fed actually use the Bacon Cheeseburger Index? Of course not. But does it help explain in one compact, if anecdotal, form why the Federal Reserve is happy to hold off on rate increases? I think it does,” said Colas. + +Here are some of the report’s other findings: + +• Since Google GOOGL, +0 +================================================================================ +Rank = 18; Score = 1425408.0 +<|begin_of_text|>The latest data on the Chinese economy released on Saturday show exports down 8% year-on-year. By any measure you pick, China’s expansion is losing steam. + +But working out how much it’s slowing by is extremely hard to do. + +Based on a series of different indicators, the economy is currently growing between about 4% and 7% a year. That’s not exactly a small gap, and the truth matters enormously for China, Asia and the rest of the world. The country has a ballooning debt issue that will be much harder to deal with if growth is slow. + +Bank of America Merrill Lynch global economist Gustavo Reis and his team sent out a note over the weekend discussing just how much Chinese growth is slowing by a series of different measures. + +One of the measures they have looked at is an analyst favourite: The Li Keqiang Index (LKI). Back in 2007, Li reportedly told a US ambassador that he personally looked at the economy with an index made up of three parts: Electricity production, rail traffic, and lending. + +Here’s how those three look: + +Loans are rising, but electricity production is up by less than 5% year-on-year, and rail traffic is shrinking at rates last seen during the depths of the 2008-09 financial crisis. + +So it’s fair to say that it’s not a pretty picture as far as China’s Premier’s favourite economic index is concerned. Here’s how it looks in comparison to GDP: + +The BAI (broad activity index) brought together by BAML economists tracks GDP much more closely — it’s a composite of 20 different indicators. + +But yet another indicator, the LEAP index, which brings together “power output, crude steel output, cement output, auto sales, housing starts, railway cargo traffic and medium- to long-term loans,” suggests Chinese GDP is even more overblown than the Li Keqiang Index does. + +So it’s something of a mystery. + +There seem to be at least three possibilities: + +The government is lying. This doesn’t seem difficult to imagine. The credibility of the Chinese government is attached to both its ability to provide growth and its ability to exert a high level of control over the economy. Slower growth isn’t something it would like to admit. + +This doesn’t seem difficult to imagine. The credibility of the Chinese government is attached to both its ability to provide growth and its ability to exert a high level of control over the economy. Slower growth isn’t something it would like to admit. China’s +================================================================================ +Rank = 19; Score = 1392640.0 +<|begin_of_text|>We tend to think of inflation as a singular force that affects everyone the same way. It turns out however that the poorest households may have m'ore to lose as the peso shrinks. In this photo is a 100 trillion Zimbabwe dollar bill, commonplace during its period of hyperinflation. (Photo: Paul/Flickr, CC BY 2.0) + +This content is also available on Medium + +Everyone knows about inflation - a pervasive, unstoppable force that moves prices (and hopefully wages) upward in the economy, but do we really understand it? Further analysis reveals that official overall inflation statistics understate the effect of rising prices on the poor. + +Understanding Inflation + +Inflation is measured through changes in what is called the Consumer Price Index or CPI. This is an aggregate measure of prices based on a certain ‘basket’ of goods composed of those typically consumed by households. It seems like a fair enough way to measure general price increases, but it has the effect of making you think that inflation affects everyone equally - it does not. + +The basket of goods is an approximation of the average proportions that households consume. It turns out, however, that there are significant differences in the basket of goods for households from different income levels: + +Lower income classes tend to spend more of their income on Food & Beverage, and this proportion drops as we move up the income scale. This is because the absolute level of food consumed in the household is limited, and extra income is allocated to certain ‘luxuries’ and other necessities, evidenced by the fact that higher income households spend more of their income on the other, non-food categories. + +Just as a validity check: the data that we generated from the FIES raw data closely approximates the official NSO weighting, except for food and housing. I suspect that imputed rental value was used by the NSO instead of actual rentals paid, but since there was no breakdown in the CPI Technical Notes, I can’t say for sure. Nevertheless, it is the variance in the effect, not the absolute effect, that we are interested in. + +So what does this mean? Different income classes will experience different inflation burdens because the prices of goods and services that they consume rise at different rates: + +Since the base year 2006, essentials such as food and beverage and education have risen much more than the overall inflation rate. This means that poorer households that spend more of their income on food and beverage will feel a stronger pinch from price increases. Non-essential goods’ prices have risen slower than inflation except for Alcohol & Tobacco which were +================================================================================ +Rank = 20; Score = 1384448.0 +<|begin_of_text|>The government has admitted that 800,000 more households stand to lose out under its flagship welfare scheme than previously thought. + +In a revised impact assessment of its universal credit scheme, which will replace six of the seven main means-tested benefits and tax credits for those of working age, the Department of Work and Pensions (DWP) says that 2.8 million households will get a lower entitlement to benefits. That compares with a figure of 2 million announced in October 2011 after officials first considered how the changes would affect claimants. + +The sums are large. About 800,000 households will see an average loss of £137 a month, while the 300,000 hardest hit families will lose as much as £300 a month. About 200,000 lone parents will also receive lower awards under the new scheme than the current system. + +The much greater impact is due to a combination of factors – the deteriorating economic environment appears to have added £700m to the costs of the policy according to the Office for Budget Responsibility, while officials also concede that universal credit is less generous than first envisaged. The chancellor's decision last week to raise benefits at a rate lower than the consumer price index will also have affected the scheme. + +The revised impact assessment was released late on Tuesday. Ellen Broome, policy director at the Children's Society, said: "We have some serious concerns about how it will affect some of the most disadvantaged children and families. Some of our poorest working families will struggle to afford vital childcare. Our evidence also reveals that many disabled children and disabled lone parents will be significantly worse off." + +Labour also questioned whether there were not dwindling savings from the scheme, pointing out that the 2011 impact assessment claimed that universal credit would reduce administrative costs by £500m due to its "reduced complexity". This, says the opposition, has been revised downwards to £200m. + +Liam Byrne, the shadow work and pensions secretary, said: "Universal credit is now set to be a car crash for 2.8 million families who will be worse off than under the current system. That's nearly 50% more than when plans were first published." + +"The very people this government has let down are being asked to pick up the tab for ministers' failure. 300,000 families on middle and modest incomes will lose £3,600 a year – that's £300 a month. We were promised a welfare revolution but universal credit is descending into universal chaos. Instead of fixing welfare reform, this government is now just +================================================================================ +Rank = 21; Score = 1384448.0 +<|begin_of_text|>Turkish economy up 4 pct in third quarter, surprising analysts + +ANKARA + +DHA Photo + +Turkey’s economy grew at a surprising 4 percent in the third quarter from the same period in the previous year, the Turkish Statistics Institute (TÜİK) reported on Dec. 10, beating expectations and showing the economy in better shape than many analysts had forecasted.The figure exceeded analyst consensus estimates at 2.7 percent. Adjusted for seasonal and calendar factors, gross domestic product (GDP) grew 5.4 percent in the third quarter from the same period of 2014 and 1.3 percent from the second quarter of 2015, it added.Özgur Altuğ, chief economist at BGC Partners in Istanbul, said the rise came despite political uncertainty surrounding the Nov. 1 election and increased attacks by the outlawed Kurdistan Workers’ Party (PKK).“The data confirmed that the economy continued to accelerate in the third quarter despite political uncertainty and rising terrorism,” he said.Coupled with better-than-expected October industrial output figures, Altuğ said he would revise the 2015 GDP growth forecast from 2.7 percent to 3.4 percent.“The figures confirm the Turkish economy’s resilience to shocks,” he said.“Solid domestic demand continues to fuel growth in Turkey,” Christopher Dembik, an economist with Saxo Bank, told Anadolu Agency.“Consumers had held back spending when the Turkish Lira lost value earlier this year. Now that it has stabilized, confidence in spending has returned,” Dembik added.Household final consumption expenditure increased by 10.6 in the quarter at current prices, the report said.The Consumer Confidence Index increased to 77.15 in November from 62.78 in October.Growth was up from 3.8 percent in the second quarter. The nine-month GDP growth rate increased by 3.4 percent in the third quarter of 2015, the report said.Increasing agriculture activity contributed the most to the GDP increase. The sector increased by 20.1 percent at current prices from the same quarter in 2014, the report said.The value added by industry increased by 7.5 percent in the quarter, the research said. The services sector’s contribution rose 11.2 percent in value.Industrial production rose 4.6 percent in October from October 2014 in Turkey, according to a separate report from TÜİK.“Increases in German manufacturing have meant that Turkish producers can benefit +================================================================================ +Rank = 22; Score = 1351680.0 +<|begin_of_text|>A heat map of Boston apartments by price per bedroom with MBTA route overlay. + +A new tool developed by Jeff Kaufman, a 27-year-old programmer at Google in Cambridge, allows apartment renters to visualize where Boston’s most and least expensive apartments are. Kaufman's heat map displays the cost per bedroom relative to each Boston neighborhood. + +The map indicates the most expensive apartments in Boston tend to hover between Back Bay and Downtown Boston and on into the Seaport District. The region’s least expensive apartments are found near Mattapan, Dorchester and Revere. + +Areas of Cambridge along the Red Line also display expensive apartment listings, where the booming tech economy is squeezing commercial rents for startups. + +Kaufman's first iteration of the heat map originated in 2011 when he was looking for a place to live. + +"I was looking for an apartment and didn't have a good sense of how expensive different parts of the city were," Kaufman says. "I wanted to have a map that gave you a rough idea of the city and how expensive they were." + +Reddit user “yiseowl” helped explain the heat map by overlaying the MBTA train map over Kaufman’s. The juxtaposition shows some correlation between rent and MBTA accessibility. + +According to Kaufman’s data, which he took from apartment rental search engine Padmapper, the average cost per bedroom in 2013 is $1,314. In 2011, the average cost per bedroom was $1,141. + +This is a much greater increase than inflation. The [Consumer Price Index] in June 2011 was 225.922 while the December 2012 one was 229.594, so you would expect $767/room then to turn into $780/room now instead of the $895 we actually see. + +Curbed Boston also published a real-estate map in December of the 15 most expensive home sales of 2012.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 23; Score = 1294336.0 +<|begin_of_text|>Choir singing is known to promote wellbeing. One reason for this may be that singing demands a slower than normal respiration, which may in turn affect heart activity. Coupling of heart rate variability (HRV) to respiration is called Respiratory sinus arrhythmia (RSA). This coupling has a subjective as well as a biologically soothing effect, and it is beneficial for cardiovascular function. RSA is seen to be more marked during slow-paced breathing and at lower respiration rates (0.1 Hz and below). In this study, we investigate how singing, which is a form of guided breathing, affects HRV and RSA. The study comprises a group of healthy 18 year olds of mixed gender. The subjects are asked to; (1) hum a single tone and breathe whenever they need to; (2) sing a hymn with free, unguided breathing; and (3) sing a slow mantra and breathe solely between phrases. Heart rate (HR) is measured continuously during the study. The study design makes it possible to compare above three levels of song structure. In a separate case study, we examine five individuals performing singing tasks (1–3). We collect data with more advanced equipment, simultaneously recording HR, respiration, skin conductance and finger temperature. We show how song structure, respiration and HR are connected. Unison singing of regular song structures makes the hearts of the singers accelerate and decelerate simultaneously. Implications concerning the effect on wellbeing and health are discussed as well as the question how this inner entrainment may affect perception and behavior. + +Introduction + +This paper aims to illuminate and discuss how singing (Grape et al., 2003) and especially choir singing (Marti, 2012) promotes wellbeing. We focus on the interplay between two oscillators: the respiratory organ and the heart, while singing. The heart does not keep a constant rhythm. On the contrary, heart rate (HR) is accelerating and decelerating constantly. This fluctuation in HR is called heart rate variability (HRV). In this article, we mainly speak of the dynamic aspects of HRV, i.e., how HR develops over time forming cycles of HR fluctuations, but we also measure HRV by RMSSD (Root Mean Square of the Successive Differences) (Degiorgio et al., 2010). + +It is known that HRV and respiration rate affect each other. This interdependency, first discussed by (Ludwig, 1847), is referred to as respiratory sinus +================================================================================ +Rank = 24; Score = 1261568.0 +<|begin_of_text|>NYT Pick Keynes Florida 1 day ago In macroeconomic terms, an increase in taxes followed by an equivalent increase in spending (and that therefore does not increase the deficit) increases the GDP. It is called “the balanced budget multiplier” (you can Google it!), BBM. + +Conversely, a reduction in taxes followed by a reduction in spending reduces the GDP, and therefore increases unemployment. + +In this particular case (repeal of the ACA), the reduction in taxes benefits higher income people who have a relatively low “marginal propensity to consume” (Google it!), MPC, and negatively affects lower income people, who have a relatively higher MPC. The net effect is that a reduction in taxes and an equal reduction in spending will decrease the GDP by an amount larger than the reduction in taxes. The effect on the unemployment rate will be larger than predicted by the BBM. + +Will the repeal of the ACA be the trigger for the next recession/depression? + +How long will it take for the country to be plunged into a depression? (a) two years (b) four years (c) eight years. + +It will depend on how large are the tax and spending cuts. + +GWB managed to do it in eight years, but he started with a stronger economy and made no significant spending cuts. Flag + +Flag 931 Recommend + +Recommend Share this comment on Facebook Share this comment on Twitter + +NYT Pick PeterS Boston, MA 1 day ago Mr. Krugman is right that Mr. Obama has left the country in near top shape for Mr. Trump and it is unlikely that expansion will last forever. Incidentally, things getting worse may not even be Mr. Trump's doing. While we liberals are not cheering for the success of most Trump's policies, all conscientious citizens should be extra vigilant when economic turns bad and lives get tough. History has repeatedly told us that tough times were the breeding group for authoritarians and fascists. Demagogues never had solutions but they had risen by channeling the anger of the people against chosen minorities. I hope that America will stay prosperous despite Trump and his regressive policies will be reversed over time. I hope that if bad economic times will come, we will choose a wise and compassionate leader next time to take us out of it. However, history should have taught us to be watchful; this election has shown that the wisdom of, we, the American people is not infallible. Flag + +Flag 768 Recommend + +Recommend Share this comment on Facebook Share this comment +================================================================================ +Rank = 25; Score = 1220608.0 +<|begin_of_text|>Key consumer debt measure hits record + +Net worth boosted by stocks, housing + +Consumer credit growth may be slowing in Canada, but the key measure of household debt has hit a record. + +That measure, household credit market debt to disposable income, climbed in the third quarter of the year to 163.7 per cent, from a revised 163.1 per cent in the previous three-month period, Statistics Canada said Friday. + +At the same time, the net worth of Canadian families rose 2.2 per cent, buoyed by rising stock markets and property values. + +Story continues below advertisement + +On a per-capita basis, net worth among Canadian households rose in the latest quarter to $211,400, though that masks the income inequality across the country. + +"Shares and other equities grew on the basis of the rebound in domestic and foreign equity markets," Statistics Canada said. + +"The increase in household net worth was also supported by a 1.5-per-cent gain in the value of household real estate." + +The higher debt burden could well spell trouble for some families when interest rates inevitably rise, though the Bank of Canada is believed to be more than a year away from such a move. + +Here are some of the highlights from the federal agency's report: + +Household net worth rose by 2.2 per cent to $211,000, led by a 3.7-per-cent rise in the stock market. + +Real estate values rose by 1.5 per cent. + +Consumer credit debt rose by 1 per cent to $505-billion. + +The consumer credit debt to disposable income ratio climbed to 163.7 per cent from 163.1 per cent. + +Mortgage debts rose by 1.8 per cent to $1.1-trillion, compared with the average quarterly growth of 1.7 per cent over the past five years. + +National net worth climbed 2.1 per cent to $7.5-trillion, or, on a per-capita basis, $212,700. That's a slower pace than the second quarter's 3.2-per-cent rise. + +Home sales, of course, have rebounded smartly in Canada since the short government-induced slump in the summer of 2012, brought on by new mortgage insurance restrictions aimed at stopping a bubble. + +Some of that is deemed to be the result of home buyers jumping in earlier than they otherwise would have because of higher mortgage rates and speculation of further increases. + +Story continues below advertisement + +Story continues below advertisement + +Thus, economists believe the debt burden will +================================================================================ +Rank = 26; Score = 1171456.0 +<|begin_of_text|>Oct 26 (Reuters) - Commerce Department’s Bureau of Economic Analysis’ seasonally adjusted Gross Domestic Product data on a chain-weighted basis. + +Following are annualized percent changes from prior quarters, in 2005 chain dollars: + +Q3-A Q2-F Q1 2011 GDP 2.0 1.3 2.0 1.8 Final Sales of Dom. Product 2.1 1.7 2.4 2.0 Final Sales to Dom. Buyers 2.3 1.4 2.2 1.8 PCE price index 1.8 0.7 2.5 2.4 Core PCE price index 1.3 1.7 2.2 1.4 Mkt-based PCE price index 1.8 0.6 2.5 2.5 Core Mkt-based index 1.3 1.8 2.2 1.4 GDP price index 2.8 1.6 2.0 2.1 Implicit Deflator 2.9 1.5 2.2 2.1 Consumer Spending 2.0 1.5 2.4 2.5 + +Durable Goods 8.5 -0.2 11.5 7.2 + +NonDurable Goods 2.4 0.6 1.6 2.3 + +Services 0.8 2.1 1.3 1.9 Business Investment -1.3 3.6 7.5 8.6 Structures -4.4 0.6 12.9 2.7 Equipment/software 0.0 4.8 5.4 11.0 Housing Investment 14.4 8.5 20.5 -1.4 Exports -1.6 5.3 4.4 6.7 Imports -0.2 2.8 3.1 4.8 Government Purchases 3.7 -0.7 -3.0 -3.1 Federal 9.6 -0.2 -4.2 -2.8 State and Local -0.1 -1.0 -2.2 -3.4 A-Advance (1st estimate). P-Preliminary (2nd). F-final (3rd.) Seasonally adjusted annual rates, in blns of +================================================================================ +Rank = 27; Score = 1105920.0 +<|begin_of_text|>LONDON (Reuters) - British inflation rose in April for the first time in 10 months, but as the increase was partly due to a late Easter holiday, which pushed up transport costs, it was unlikely to alter interest rate expectations. + +A woman pays a market trader for goods in Soho's Berwick Street Market in central London May 17, 2011. REUTERS/Paul Hackett + +Official data also showed that growth in house prices eased in March, potentially tempering concerns about a bubble brewing in the property market although some more up-to-date surveys have shown prices picking up again. + +The consumer price index rose more than expected to an annual rate of 1.8 percent in April, from 1.6 percent in March, which had been its lowest level in more than four years, the Office for National Statistics ONS said. + +It was the first rise in the CPI since June 2013 and above a Reuters poll forecast for inflation of 1.7 percent in April. + +Sterling hit a 16-month high against the euro and rose against the dollar before giving up much of its gains as market expectations that the Bank of England will raise interest rates in about a year’s time remained largely intact. + +The opposition Labour party jumped on Tuesday’s figures, saying they showed the government’s complacency about the “cost of living” crisis facing many Britons as wage growth has lagged inflation. + +Economists however saw little change in price pressures. + +“Easter effects aside, the latest release provides little evidence of significant or broad-based price pressures,” said Victoria Clarke, an economist with bank Investec. + +Factory gate inflation in April was weaker than economists’ predictions. + +Higher airfares linked to Easter and other transport costs helped push up consumer prices, the ONS said. Easter fell in April this year but in late March last year. + +Offsetting those rises, food price growth was the lowest in eight years as a mild spring kept vegetable prices down. + +Core CPI, which excludes food costs and other items but does include transport costs, rose 2.0 percent, its strongest rate since September last year. + +House prices - which are an increasing source of concern for the Bank of England - were up 8.0 percent on the year in March, slowing from a 9.2 percent rise in February, the ONS also said. + +Prices in London, however, were up 17 percent on the year, although that was less than a 17.8 percent increase in February, the biggest increase since 200 +================================================================================ +Rank = 28; Score = 1105920.0 +<|begin_of_text|>India has moved up in ranking from 52nd to 40th position in the Travel and Tourism Competitive Index (TTCI) of the World Economic Forum (WEF). The tourism sector in the country has been on a growth trajectory in the recent past. + +This data is based on the 2017 edition of the Travel and Tourism Competitiveness Report that features the latest iteration of the Travel & Tourism Competitiveness Index (TTCI). The TTCI benchmarks the travel and tourism competitiveness of 136 economies. The TTCI measures “the set of factors and policies that enable the sustainable development of the Travel & Tourism (T&T) sector, which in turn, contributes to the development and competitiveness of a country”. It comprises four sub-indexes, 14 pillars, and 90 individual indicators, distributed among the different pillars. + +India’s ranking in the TTCI of World Economic Forum moved from 65th position to 52nd position in 2015. This year, India has moved up by another 12 positions and ranked at the 40th position. In the last three years, India has cumulatively improved its ranking by 25 places, which is a significant achievement. + +Spain tops the 2017 edition of the TTCI global rankings for the second time, followed by France (2nd), Germany (3rd), Japan (4th, gaining five places), the United Kingdom (5th), the United States (6th, losing two places), Australia (7th), Italy (8th), Canada (9th, up one) and Switzerland (10th, losing four places). + +Dr Mahesh Sharma, Minister of State (I/C) for Tourism and Culture acknowledged the vision, guidance and support of Indian Prime Minister Narendra Modi for the tourism sector of the country that is the driving force and motivation for everyone to continuously work for the growth of Tourism in India. + +India continues to allure international tourists with its vast cultural and natural resources with a ranking of 9th and 24th respectively. In terms of price competitiveness advantage, India is ranked 10th. + +India continues to enhance its cultural resources, protecting more cultural sites and intangible expressions through UNESCO World Heritage lists, and they also increasing their digital presence. + +India is ranked 55th, up by 14 places in terms of international openness, which has been possible through stronger visa policies. Implementing both visa-on-arrival and e-visa has enabled India to rise through the ranks. + +The key reasons for India’s jump in the Travel and +================================================================================ +Rank = 29; Score = 1089536.0 +<|begin_of_text|>The U.S Federal Reserve has released the worst-case scenarios it wants banks to stress test against, and some of them are downright apocalyptic. + +As part of the stress tests, which the Fed announced it wants banks to do annually, U.S. lenders will have to simulate the effects of a severe recession hitting the U.S. What does that involve, you ask? + +In the Fed’s bleakest scenario, the United States would slip into recession in the fourth quarter of 2011 (the one we’re currently in) and post four consecutive quarters of negative growth. The unemployment rate would peak at 13.05%, while the Dow Jones Industrial Average would plunge all the way to 5,668 points (from the current level of 11,305). That would be even lower than the Dow’s financial crisis low in March 2009, when it settled at 6,547. + +The Fed’s simulated recession would be deepest in the first quarter of 2012, when real GDP would contract by a whopping 7.98%. Peak unemployment wouldn’t hit until a year later. + +Overall, the tests, known as the Comprehensive Capital Analysis and Review, lay out 14 economic metrics for banks to include as means of testing their preparedness. The metrics include Consumer Price Inflation levels, Treasury Yields and mortgage rates. + +The tests are designed for the top 31 banks in the U.S. that hold assets of US$50-billion or more. They will generate models that will produce loss estimates and see whether capital ratios are high enough to survive the worst-case scenarios. + +The Fed will also allow banks to come up with their own scenarios to stress test their finances. Either way, it will requires banks to submit their capital plans by January 9, 2012. + +For a full list of the metrics that will be used in the tests, check out the U.S. Federal Reserve’s website. + +• Email: jshmuel@nationalpost.com | Twitter: jshmuel<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 30; Score = 1048576.0 +<|begin_of_text|>NEW YORK (CNNMoney.com) -- There will be no cost-of-living increase for 57 million Social Security beneficiaries next year because consumer prices have fallen, the Social Security Administration announced on Thursday. + +It marks the first time that Social Security benefits have not been increased year over year since the cost-of-living adjustment was put into effect in 1975. + +To help counterbalance the hit, President Obama is calling on Congress to send another $250 relief payment to seniors and other Americans to stem the economic strain. + +"Even as we seek to bring about recovery, we must act on behalf of those hardest hit by this recession," Obama said in a statement Wednesday. "That is why I am announcing my support for an additional $250 in emergency recovery assistance to seniors, veterans, and people with disabilities to help them make it through these difficult times." + +Last year, Social Security beneficiaries got a 5.8% cost-of-living adjustment, the largest since 1982, largely because of the spike in energy prices. + +"This year, in light of the human need, we need to support President Obama's call for us to make another $250 recovery payment for 57 million Americans," said Commissioner of Social Security Michael J. Astrue in a written statement. + +Since there will be no COLA for benefits, the law also prohibits the Social Security Administration from increasing the maximum amount of earnings subject to the Social Security tax. This year and next, the first $106,800 of a worker's earnings is subject to the 12.4% Social Security tax. Workers typically pay half of that and their employers pay the other half. + +It's still not clear yet what if any changes will be made to seniors' Medicare Part B premiums for hospital care next year. The Social Security Administration said in its announcement that if there is an increase that a "hold harmless" provision in the law would protect 93% of Social Security beneficiaries from the increase. + +New emergency payment similar to COLA + +Obama's proposed $250 payment is roughly equal to a 2% increase in benefits for the average Social Security beneficiary. + +Congress approved a similar payment as part of the $787 billion economic recovery act enacted in February. + +As with the first $250 recovery payment, the second one would be exempt from income tax, a senior administration official said in a call with reporters on Wednesday. + +If approved by Congress, the payments would be sent out in 2010, most likely in the first half. "It wouldn't be late in 2010," the administration official +================================================================================ +Rank = 31; Score = 1032192.0 +<|begin_of_text|>Chinese Research Team that Employs High Performance Computing to Understand Weather Patterns Wins 2016 ACM Gordon Bell Prize + +Salt Lake City, Utah, November 17, 2016 – ACM, the world's leading professional computing society, has named a 12-member Chinese team the recipients of the 2016 ACM Gordon Bell Prize for their research project, “10M-Core Scalable Fully-Implicit Solver for Nonhydrostatic Atmospheric Dynamics.” The winning team presented a solver (method for calculating) atmospheric dynamics. The ACM Gordon Bell Prize tracks the progress of parallel computing and rewards innovation in applying high performance computing to challenges in science, engineering, and large- scale data analytics. The award was bestowed during the International Conference for High Performance Computing, Networking, Storage and Analysis (SC16) in Salt Lake City, Utah. + +Since the dawn of computing, scientists have used data analytics in an effort to predict and simulate the weather and related atmospheric events. In the early years of weather forecasting, scientists might have used standard central processing units (CPUs). With each passing year, the continued expansion in the capabilities of high performance computers has enabled researchers to employ increasingly sophisticated computational methods for the analysis and modeling of weather patterns. Advanced scientific computers break problems down into composite parts and perform immense amounts of mathematical calculations simultaneously. The performance of a supercomputer is measured in floating-point operations per second (FLOPS). Some of the latest supercomputers are capable of performing quadrillions of FLOPS. + +In the abstract of their presentation, the winning team writes, “On the road to the seamless weather-climate prediction, a major obstacle is the difficulty of dealing with various spatial and temporal scales. The atmosphere contains time-dependent multi-scale dynamics that support a variety of wave motions.” + +To simulate the vast number of variables inherent in a weather system developing in the atmosphere, the winning group presents a highly scalable fully implicit solver for three-dimensional nonhydrostatic atmospheric simulations governed by fully compressible Euler equations. Euler equations are a set of equations frequently used to understand fluid dynamics (liquids and gasses in motion). + +Elaborating further, they add, “In the solver, we propose a highly efficient domain-decomposed multigrid preconditioner that can greatly accelerate the convergence rate at the extreme scale. For solving the overlapped subdomain problems, a geometry-based pipelined incomplete LU factorization method is designed to further exploit the on-chip fine-grained concurrency.” + +The fully-implicit solver successfully scales to the entire system of the Sunway TaihuLight, a Chinese super +================================================================================ +Rank = 32; Score = 1003520.0 +<|begin_of_text|>Fresh Christmas trees are a holiday tradition, loved for their beauty and fresh, outdoorsy fragrance. However, Christmas trees often take the blame for destructive fires that occur during the holiday season. The most effective way to prevent Christmas tree fires is to keep the tree well hydrated. With proper care, a tree should remain fresh for two to three weeks. This may sound easy, but it becomes a problem if your Christmas tree is not drinking water. + +Causes for a Christmas Tree Not Taking Up Water + +Generally, when Christmas trees have problems taking up water, it’s because we tend to add products to the tree itself or the water. Avoid spray-on fire retardants and other products advertised to keep your tree fresh. Similarly, bleach, vodka, aspirin, sugar, lime soda, copper pennies or vodka have little or no effect and some can actually slow water retention and increase moisture loss. + +What works best? Plain old tap water. If you tend to be forgetful, keep a pitcher or watering can near the tree to remind you. + +How to Get a Christmas Tree to Take Up Water + +Cutting a thin sliver from the bottom of the trunk is key to keeping a tree fresh. Keep in mind that if the tree is freshly cut, you don’t need to cut the trunk. However, if the tree has been cut for longer than 12 hours before you put it in water, you must trim ¼ to ½ inch from the bottom of the trunk. + +This is because the bottom of the trunk seals itself with sap after a few hours and can’t absorb water. Cut straight across and not at an angle; an angular cut makes it harder for the tree to take up water. It is also difficult to get a tree with an angular cut to stand upright. Also, don’t drill a hole in the trunk. It doesn’t help. + +Next, a large stand is critical; a Christmas tree can drink up to one quart of water for each inch of stem diameter. The National Christmas Tree Association recommends a stand with a one-gallon capacity. Never trim the bark to accommodate a too-tight stand. The bark helps the tree take up water. + +Christmas Tree Watering Tips + +Start with a fresh Christmas tree. There’s no way to hydrate a dried up tree, even if you trim the bottom. If you aren’t sure about freshness, pull a branch slowly through your fingers. A few dry needles are no reason for concern, but look for a fresher tree if a large number of needles are loose or brittle. + + +================================================================================ +Rank = 33; Score = 966656.0 +<|begin_of_text|>EspañolThe value of the Venezuelan bolívar has plummeted in the past month, depreciating from 102.56 to 159.02 per US dollar on the black-market exchange. November’s decline indicates that Venezuela now flirts with hyperinflation. + +The rule of thumb among economists is that hyperinflation occurs when inflation — the declining purchasing power of a currency as measured by a rise in the price level — exceeds 50 percent on a monthly basis. As deduced from the black-market exchange rate and the US Consumer Price Index, Venezuela’s inflation for November reached 55.27 percent.* Even the highest denomination of the bolívar, the 100 Bs. note, is now worth a mere 62.9 US cents. + +Subscribe to our daily newsletter Youtube + +If Venezuela were to continue November’s free-fall trajectory for the bolívar, its annual compounded inflation would reach 19,633 percent. + +While DolarToday is the most commonly referenced website for the bolívar’s informal exchange rates, Aguacate Verde and Lechuga Verde both uphold the trend (at 146 Bs. per US dollar). Of the three, DolarToday is the only one that offers a publicly available data record (Excel), enabling comparisons between past and present exchange rates. + +The Bolívar’s Disappearing Act + +In a country with three fixed official exchange rates — all with rationed supplies of foreign currencies — the informal rate has become the most accurate gauge of the bolívar’s underlying purchasing power. + +“The black market rate is a very good indicator of what the exchange rate should be,” says Boris Ackerman, an economist and professor at Simon Bolívar University in Caracas. “Public expectations and speculation continue to drive the rate up despite strict financial controls enacted by the government.” + +In an economy defined by crisis, Venezuelans are scrambling from the bolívar to take refuge in the US dollar, which offers a relatively more stable alternative and store of value. In response, President Nicolás Maduro has worked to steadily decrease the availability of US dollars within the country, even going so far as to outlaw transactions in greenbacks. + +This year, the Chavista regime also introduced the third SICAD II rate, in an attempt to curb the thriving black market. However, the measure has since backfired, and the black-market exchange rate has more than doubled since SICAD II’s introduction in March. + +Venezuela’s financial woes have come under immense criticism, including from prominent outlets such as the +================================================================================ +Rank = 34; Score = 962560.0 +<|begin_of_text|>New data from the Federal Reserve continues to highlight a reemergence of debt based consumer spending. Americans are largely buying stuff they can’t afford with money they don’t have. The Fed’s consumer credit report highlights a troubling trend. Over the last 12 months 95 percent of all consumer debt growth has come from people buying cars and young Americans going deep into debt to pursue a college education. A car quickly loses its value once it drives off the lot and many college degrees are massively over valued only being supported by easy financing. This is a disturbing trend, even more troubling than the last debt fueled bubble. One can argue with housing that at least people are getting an asset that generally rises with inflation. But a car? A for-profit degree? The debt juice is now flowing once again. + +Buying cars while incomes drop + +If we look at incomes and car sales we find that households are simply purchasing cars by borrowing from the future: + +The Fed consumer credit report merely confirms what we suspect. The dramatic jump in auto buying isn’t coming because households are suddenly wealthier but because consumer debt is booming once again. Cars are not “good” assets and unfortunately, many Americans need to grow their net worth but adding a depreciating vehicle on your balance sheet isn’t exactly a smart plan. + +The consumer credit report shows that 95 percent of consumer debt growth in the last 12 months has come in the form of auto loans and student debt: + +Source: ZeroHedge, Fed + +When you factor in the reality that student debt is now the largest debt market outside of mortgage debt, you realize that many young Americans are saddling their lives in the red even before they begin their working careers. + +Total student debt in the US is well over $1.2 trillion as measured by new Fed data: + +So we are now having another expansion completely based on easy access to debt. For banks, this access never went away in the form of cheap money via the Fed. The Fed’s balance sheet is now quickly approaching $4 trillion. We are told that there is no inflation but real estate values are spiking at a rate last seen during the last bubble but household incomes are not going up. Of course this time the price of real estate is being pushed up by Wall Street investors. At least in the last bubble, most Americans had access to mortgages, credit cards, and all other forms of cheap debt to party with Wall Street. Today, it seems like consumers have the option of auto loans or student debt. Good luck trying to buy a home in some +================================================================================ +Rank = 35; Score = 942080.0 +<|begin_of_text|>Black Friday is not dead. In fact, this may be the strongest Black Friday for retailers in years. + +Shoppers were already lined up outside of Target, Best Buy, and J.C. Penney, according to CNBC. Walmart stores went into Black Friday sales mode at 6 p.m. + +Discounts are not expected to be especially attractive this year. What is driving the surge in shopping is much stronger consumer confidence. In October, consumer sentiment hit a 13-year high. That cooled slightly but remained above economist expectations in November, signaling that the surge has staying power. Consumers indicate that they feel good about their current financial situation as well as their expectations for the future. Certainty about the outlook for income, employment, and inflation are all vastly improved from past years. + +Overall, consumer sentiment is around five percent higher than it was the previous November, according to the University of Michigan’s consumer sentiment survey. Deloitte projects that holiday shopping will rise between 4 percent last year and 4.5 percent this year. + +Another plus for retailers: moderate weather. It is chilly in much of the country but not frigid. The sun is literally shining across almost all of the United States on the day after Thanksgiving. That will likely drive many people out of their homes and into shopping malls. + +Much of the shopping has shifted online, however. According to Adobe Analytics, Americans spent $1.52 billion on Thanksgiving evening–a 17 percent surge over last year.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 36; Score = 937984.0 +<|begin_of_text|>Reserve Bank of India (RBI) governor Raghuram Rajan dashed expectations of the government, consumers and business leaders as he kept lending rates unchanged on Tuesday despite plunging inflation and mounting pressure to ease borrowing costs to aid a revival. + +Unchanged interest rates would imply your home loan EMIs, which eat away large chunks of household incomes, are unlikely to fall anytime soon. + +High loan rates could influence people’s decision to buy houses, cars and other consumer goods, mostly bought through loans. + +The benchmark lending rate has remained unchanged since January last year, demonstrating the RBI governor’s unwavering commitment to keep inflation firmly bottled up for longer period of time, before softening of the rather hawkish stance. + +Budding revival signs in the broader economy and falling inflation rates had rekindled hopes that the RBI would lower loan rates to assist companies’ investment plans, critical to spin jobs and multiply income. + +Rajan retained the repo rate -- the rate at which banks borrow from RBI -- at 8%, and kept the cash reserve ratio (CRR)-- the proportion of deposits banks to have to park with the central bank -- at 4%, in the bi-month monetary policy review. + +Retail inflation eased in to a three-year low of 5.52% in October, below the central bank's target of 6% by 2016. + +India’s wholesale inflation rate -- the main gauge to capture country-wide price movements -- has also plunged to a five-year low of 1.77% in October, triggering a chorus of demand for lower borrowing costs to boost investment and consumer spending. + +The next monetary policy is scheduled for February 3, 2015, but Rajan did not rule out a rate revision outside of the policy calendar depending on future price movements. + +“A change in the monetary policy stance at the current juncture is premature. However, if the current inflation momentum and changes in inflationary expectations continue, and fiscal developments are encouraging, a change in the monetary policy stance is likely early next year, including outside the policy review cycle,” Rajan said. + +Analysts said the RBI governor could wait for cues to come from the next year’s Union budget in February before initiating any rate action. + +Industry leaders have been ratcheting up their demand for cheaper loans with the government also hoping that the central bank would signal a reversal of its prolonged hawkish stance that some see as affecting capacity expansion plans. + +According to experts low and stable interest rates are critical to fast-track roads, ports, airports, and railways projects to create +================================================================================ +Rank = 37; Score = 913408.0 +<|begin_of_text|>Price to Earnings Ratio (P/E) + +The price to earnings ratio (denoted P/E) measures the relationship between the share price to its earnings per share. It is considered one of the most important metric of a company because comparing two similar companies’ P/E shows us which is a better value. Sometimes, people will refer to this ratio as the price to earnings multiple, or simply the multiple of a company. + +What Price-Earnings Ratio Represents + +The P/E ratio lets investors know what the price of a particular stock is related to the earnings it makes. If a company makes $5 and has a share price of $50, then the P/E is simply 10 (50 divided by 5). With this information, an investor can clearly see that the general public is currently willing to pay $10 for every $1 that the company makes. This also means that the lower this number, the better value the investment potentially is. + +What the Ratio is Telling Us + +Investors love growth. So a high P/E usually represents the general public anticipating high growth in the months ahead. This could be based on the company’s track record, anticipation of a new product launch, or macroeconomics such as a good economy. + +In some ways, this ratio also tells us the confidence level in the stock. When investors are willing to pay a higher multiple for a stock versus another, it generally means that they are more confident that the company’s share price will rise quicker the rest. + +A Simple Example + +Company A and B having a stock price of $20 and $30 respectively says nothing about the relative value of the enterprises. However, if they both earn $1 per share, then comparing their P/E ratio of 20 and 30 tells us that investors are currently willing to pay $20 for each dollar that company A earns and $30 for each dollar company B makes. With this information, we can see that company A is a relatively better value than company B if all else is equal. + +What to Watch Out For + +Investors need to take caution when comparing P/E ratios of two companies. If we look at different industries, we will see vast differences in P/E ratios. For instance, technology stocks’ having a P/E of 30 is not uncommon but an oil service company may have a P/E of 10. While the oil service company may look like a better value, it does not mean that every oil service company is a better investment than a technology company. Therefore, P/E ratios should be used to compare +================================================================================ +Rank = 38; Score = 884736.0 +<|begin_of_text|>During the Reserve Bank of India’s (RBI) press conference last week, one exchange stood out. Asked about a meeting scheduled by the finance ministry in New Delhi to discuss the RBI’s policy, the governor of India’s central bank, Urjit Patel, replied: “The meeting did not take place. All the MPC members declined the request of the finance ministry for that meeting." + +The MPC is the Monetary Policy Committee, a six-member council that sets India’s benchmark interest rates. It’s a new institution, less than a year old; Patel’s predecessor, Raghuram Rajan, had full authority to set interest rates himself. But it seems India’s government is not as comfortable as many had hoped with this new arrangement—summoning the MPC members to the finance ministry before they decided rates was widely seen as an attempt to interfere in their decision. + +So it’s a great relief that the RBI stood firm. Its independence has been questioned of late, partly because of the circumstances in which Rajan left, under pressure from lawmakers in New Delhi, and partly because of its supine acceptance of the government’s controversial decision in November to withdraw 86% of India’s currency from circulation. The government should never have asked for the meeting; but, given that it did, the members of the MPC were right to refuse. + +Also read: India’s war of the Mandarins leaves companies as victims + +In any case, they knew what the government was likely to say: “Please, for the love of God, cut rates." India is confronting a growth slowdown that is born, most agree, out of a sustained crisis in private investment. It doesn’t take a meeting with the government for the MPC to know that. Yet it has ignored the pleas of officials—and of many in the private sector—and kept rates steady. + +This determination emerges from another relatively recent reform: an agreement between the central bank and the government to shift to targeting consumer price inflation. A clear goal for the RBI—rather than a vague admonition that it should care about growth, prices and the exchange rate all at once—adds to its independence and clarifies its approach to monetary policy. But the government has grown increasingly frustrated. After all, not only is growth slowing, but the RBI has refused to reduce interest rates even as the inflation rate has fallen to historic lows. + +Some argue that the MPC is simply ignoring or misreading the evidence. Even the RBI had to admit that its inflation forecast for the first quarter was off by an unusually large 1.4 +================================================================================ +Rank = 39; Score = 880640.0 +<|begin_of_text|>0 + +I sat on the porch + +Listened to the rain + +Smoked a cigarette + +And counted to ten Oh no, here it comes again + +That funny feeling ― Camper Van Beethoven, Oh No! (1985) + +A quick post-Fed follow-up to “Tell My Horse”, the best-received Epsilon Theory note to date (thank you!). I’ll jump right into what I’ve got to say, without the usual 20 pages of movie quotes and the like. Well, I’ve got one quote above, because I can’t help myself. They’re the lyrics to the best break-up song ever, and they’re what Janet Yellen was singing to the market on Wednesday. + +Let’s review, shall we? Last fall, the Fed floated the trial balloon that they were thinking about ways to shrink their balance sheet. All very preliminary, of course, maybe years in the future. Then they started talking about doing this in 2018. Then they started talking about doing this maybe at the end of 2017. Two days ago, Yellen announced exactly how they intended to roll off trillions of dollars from the portfolio, and said that they would be starting “relatively soon”, which the market is taking to be September but could be as early as July. + +Now what has happened in the real world to accelerate the Fed’s tightening agenda, and more to the point, a specific form of tightening that impacts markets more directly than any sort of interest rate hike? Did some sort of inflationary or stimulative fiscal policy emerge from the Trump-cleared DC swamp? Umm … no. Was the real economy off to the races with sharp increases in CPI, consumer spending, and other measures of inflationary pressures? Umm … no. On the contrary, in fact. + +Two things and two things only have changed in the real world since last fall. First, Donald Trump — a man every Fed Governor dislikes and mistrusts — is in the White House. Second, the job market has heated up to the point where it is — Yellen’s words — close to being unstable, and is — Yellen’s words — inevitably going to heat up still further. + +What has happened (and apologies for the ten dollar words) is that the Fed’s reaction function has flipped 180 degrees since the Trump election. Today the Fed is looking for excuses to tighten monetary policy, not excuses to weaken. So long as the unemployment rate is on the cusp of “instability”, that’s the only thing that +================================================================================ +Rank = 40; Score = 868352.0 +<|begin_of_text|>A gyro rotor | NASA + +If you’ve ever tried to give yourself a haircut, you know just how hard it is to make something precisely symmetrical. We value symmetry so highly in part because it’s really hard to achieve. + +Here are five of the most symmetrical objects humans have ever crafted, and why they were so hard to make. + +Strikingly symmetrical pieces of ear jewelry, alleged to be centuries old, have been seized on by conspiracy theorists. + +Gravity Probe B quartz gyroscope rotors + +In 2004, Gravity Probe B (GP-B) was launched into space on a Delta II rocket with the mission of testing the theory of general relativity. The Earth-orbiting satellite included a set of gyroscopes that could measure two of the phenomena predicted by general relativity: the curvature of spacetime (the geodetic effect) and the distortion of spacetime by large objects (frame-dragging). In order to measure these things, the gyroscopes had to be incredible accurate. Being off by anything larger than one one-hundred-billionth of a degree every hour would ruin the experiment. Standard gyroscopes used on submarines and military aircraft, for comparison, are over 10 million times less accurate. + +The key to building such an accurate gyroscope was creating perfectly symmetrical rotors, the quickly spinning elements that allow gyroscopes to maintain their orientation. These rotors had to be totally balanced, and completely homogenous inside as well. The GP-B team crafted these little spheres out of pure quartz blocks that were grown in Brazil and baked in Germany. The surface of each gyroscope is almost completely spherical, only about three ten-millionths of an inch away. + +According to the Guinness Book of World Records, these are the roundest objects ever made. The Stanford team that worked on the spheres says only neutron stars are more spherical. + +Avogadro Project silicon kilogram + +Gravity Probe B’s only real contender, when it comes to perfect spheres, is the ball that should soon define a kilogram. This sphere is the result of The Avogadro Project, and its raw materials alone cost over a million dollars. The goal is to to supersede and replace the international prototype kilogram (IPK). The kilogram is the very last unit in the International System of Units (essentially, the metric system) that is still defined by a physical object—a cylinder made of platinum-iridium alloy—rather than physical principles. That cylinder currently sits under three nested bell jars in a climate-controlled vault just outside of +================================================================================ +Rank = 41; Score = 856064.0 +<|begin_of_text|>When Jacob and Wilhelm Grimm published their “Children’s and Household Tales” in 1812, followed by a second volume in 1815, they had no idea that such stories as “Rapunzel,” “Hansel and Gretel,” and “Cinderella” would become the most celebrated in the world. Yet few people today are familiar with the majority of tales from the two early volumes, since in the next four decades the Grimms would publish six other editions, each extensively revised in content and style. For the very first time, ” The Original Folk and Fairy Tales of the Brothers Grimm” makes available in English all 156 stories from the 1812 and 1815 editions. These narrative gems, newly translated and brought together in one beautiful book, are accompanied by sumptuous new illustrations from award-winning artist Andrea Dezso. + +From “The Frog King” to “The Golden Key,” wondrous worlds unfold–heroes and heroines are rewarded, weaker animals triumph over the strong, and simple bumpkins prove themselves not so simple after all. Esteemed fairy tale scholar Jack Zipes offers accessible translations that retain the spare description and engaging storytelling style of the originals. Indeed, this is what makes the tales from the 1812 and 1815 editions unique–they reflect diverse voices, rooted in oral traditions, that are absent from the Grimms’ later, more embellished collections of tales. Zipes’s introduction gives important historical context, and the book includes the Grimms’ prefaces and notes. + +The source material for Children’s and Household Tales was originally collected from books as well as friends and acquaintances in and around Kassel, Hesse in what is now Germany as well as Westphalia for the romantic poet Clemens Brentano between 1806 and 1812. Jacob and Wilhelm Grimm eventually came to distinguish between the Naturpoesie of the German Volk and the Kuntspoesie, or cultivated literature, which arose from and subsumed it. Ultimately, Brentano found the collected tales unsuitable toward his purposes and gave the brothers his blessing to do with them as they pleased. The first edition of Kinder- und Hausmärchen was published in 1812. + +While even those tales were edited in order to create something more than a skeletal fragment of a story, they are generally representative of local traditions. The Grimms explained in their introduction (included) + +“We have tried to grasp and interpret these tales as purely as possible. In many of them one will find that the +================================================================================ +Rank = 42; Score = 819200.0 +<|begin_of_text|>City hall, under the leadership of a mayor Andre Chabot, would focus more on its core fiscal responsibilities and less on social spending, says the veteran Ward 10 councillor who recently launched his 2017 mayoral campaign, pitting him against Mayor Naheed Nenshi in the Oct. 16 election. + +In a wide-ranging interview with Postmedia this week, Chabot touted the “common man” appeal of his platform, pledged to change the tone at city hall, hold the line on municipal property tax increases to the consumer price index, and questioned the effectiveness of “handouts” for the city’s poorest. + +“When we talk about helping the poor, and doing all of these social things, I don’t know that we’re actually doing them a service. I think we’re actually doing them more of a disservice by doing all these handouts,” he said. + +“Hand-ups, to a certain degree, yes has some benefit,” Chabot said. “But struggle and success, I think, is just as valuable an education to help people become self-sufficient.” + +Chabot’s own struggle, in his words, went something like this: He grew up on a farm in Saskatchewan. Moved to Calgary in 1971 and soon took a job in construction. Worked his way through DeVry while raising three kids with his wife, who also worked part-time. + +“I did this on my own. I didn’t need to have somebody help me. If anything, it just encouraged me to even be better,” Chabot said. + +“Now, if somebody would have said, ‘OK, well you can go on pogie (welfare) while you go to school,’ I would have been, ‘Heh, I can just sit back and have a couple of beers, I don’t have to work now.’ Would that have made me a better man? I don’t think so.” + +In recent years, the city has made strides to make it easier for low-income Calgarians to access subsidized programs and services, such as recreation centre fee assistance, property tax assistance, basic home maintenance services for seniors and discounted transit passes through its Fair Entry program. + +Applicants are only eligible for the subsidy if the entire household income is at or below Statistics Canada’s low-income cutoff (about $24,000 for a single person and $45,000 for a household of four). + +Franco Savoia, executive director for Vibrant Communities Calgary, praised the city’s efforts to remove barriers and offer some relief +================================================================================ +Rank = 43; Score = 815104.0 +<|begin_of_text|>Image caption A slowdown in key sectors such as manufacturing has hurt India's growth rate + +India's economic growth rate picked up strongly in the most recent quarter, according to official figures. + +The economy expanded at an annual rate of 4.8% in the July-to-September period, up from 4.4% in the previous quarter. + +The acceleration was faster than analysts had been expecting. + +Asia's third-largest economy has been weighed down by various factors, such as high inflation, a weak currency and a drop in foreign investment. + +This is the fourth quarter in a row that India's annual growth rate has been below the 5% mark, and the previous quarter's rate of 4.4% was the lowest for four years. + +Earlier this year, the Indian prime minister's economic advisory council lowered the growth outlook for the current financial year. + +It now expects the economy to expand by 5.3% this year, down from its earlier projection of 6.4%. + +Growth hurdles + +India's economy has been hurt by a range of factors in recent months, including a slowdown in key sectors such as mining and manufacturing. + +Slowing growth, coupled with a recovery in developed markets, such as the US, has made India a less attractive option for foreign investors. + +Speculation that the US may scale back its key economic stimulus measure has seen investors pull money out of emerging markets, such as India. + +This has affected India's currency, which dipped nearly 25% against the US dollar between January and September this year. + +Though the rupee has recovered a little since then, it is still down about 13% against the dollar since the start of this year. + +That has made imports more expensive and contributed to a high rate of consumer inflation, which was 10.1% October, up from 9.84% in September. + +High food and fuel prices have contributed to inflation becoming "entrenched", finance minister P Chidambaram said. + +As a result, the central bank has had to raise the cost of borrowing in a bid to curb inflation. + +The latest interest rate rise in October saw the key rate increase to 7.75%. + +Some observers argue that high interest rates are hurting businesses and households, and having a negative impact on the economy. + +"A combination of weak investment, high inflation and tight monetary policy would not let India's economic recovery gather steam any time soon," Miguel Chanco, Asia economist at Capital Economics, told the BBC.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 44; Score = 815104.0 +<|begin_of_text|>Overview + +the growth in chronic disease + +increasing patient expectations about access to services + +the number and range of health services provided + +the increasing cost of those services + +an ageing population. + +Australia has a mixed public and private model of health care funding and service delivery. Private health insurance offers members greater choice in the provision of treatment, coverage for some services not included under Medicare arrangements, and may offer shorter waiting times for some services.As at 30 June 2015, 11.3 million Australians were covered by hospital treatment cover (47.4% of the population) and 13.3 million Australians had some form of general treatment cover (55.8% of the population). In 2014-15 the Australian Government spent $5.8 billion on the Private Health Insurance Rebate.In common with public health funding, private health insurance sustainability is being impacted by:The Australian Government is undertaking a number of reviews aimed at ensuring that consumers can access affordable, quality and timely health services into the future. + +Private health insurance has important links to these reviews. As such, on 28 October 2015 the Minister for Health, the Hon Sussan Ley MP, announced consultations focused on the value of private health insurance for consumers and its long term sustainability. View the Scope of the Consultations here. + +Announcement of Consultations + +Information for Consumers + +Overview of Consultations + +Minister Ley announced consultations on 28 October 2015 during her address to the National Press Club. View her address on the Department's website Private health insurance is complex. View the Consumer Fact Sheet For more information about private health insurance, visit the Private Health website (www.privatehealth.gov.au).An online consumer survey opened 8 November 2015 and closed on 7 December 2015. There was a large public response to the survey with over 40,000 Australians taking the opportunity to have their say on private health insurance.The results from the consumer survey showed that the cost of private health insurance was of greatest concern to consumers overall. Expense, affordability and lack of value for money were issues common to people both with and without private health insurance. + +Targeted industry roundtables were held in the week of 16 November 2015. Eight roundtables were held in Canberra, Melbourne and Sydney with over 100 organisations represented. + +Interested stakeholders were able to provide submissions to the Department of Health by 4 December 2015. + +An issues paper was developed to guide the roundtables and submissions. + +The written submissions indicated that stakeholders held a wide range of views on +================================================================================ +Rank = 45; Score = 794624.0 +<|begin_of_text|>Growth of religion is the spread of religions and the increase of religious adherents around the world. The statistics are commonly measured by the absolute number of adherents, the percentage of the absolute growth per year, and the growth of the number of converts in the world. Projections of future religious adherence are based on assumptions that trends, total fertility rates, life expectancy, political climate, conversion rates, secularization, etc will continue. Such forecasts cannot be validated empirically and are contentious, but are useful for comparison.[1][2] + +Studies in the 21st century show that, in terms of percentage and worldwide spread, Islam is the fastest-growing religion in the world.[3][4][5][6][7][8][9][10][11] A religious forecast for 2050 by Pew Research Center concludes that global Muslim population is expected to grow at a faster rate than the non-Muslim population due primarily to the young age and high fertility rate of Muslims.[12][13] Religious conversion has little impact on Muslim population because the number of people who convert to Islam is nearly offset by those who leave Islam.[6][14][14][15][16] + +Growth of religious groups [ edit ] + +Bahá'í Faith [ edit ] + +World religions statistics place the Bahá'í Faith around 0.1% of the world population in recent years.[17][18] The World Christian Encyclopedia estimated only 7.1 million Bahá'ís in the world in 2000, representing 218 countries,[18] and its evolution to the World Christian Database (WCD) estimated 7.3 million in 2010[19] while accredited through the Association of Religion Data Archives (ARDA). However the WCD stated: "The Baha'i Faith is the only religion to have grown faster in every United Nations region over the past 100 years than the general population; Baha'i [sic] was thus the fastest-growing religion between 1910 and 2010, growing at least twice as fast as the population of almost every UN region."[20] This source's only documented flaw was to consistently have a higher estimate of Christians than in other cross-national data sets.[21] 2015's estimate is of 7.8 million Bahá'ís in the world.[22] + +From its origins in the Persian and Ottoman empires of the 19th century the Bahá'í Faith was able to gain converts elsewhere in Asia, Europe, and North America by the early 20th +================================================================================ +Rank = 46; Score = 786432.0 +<|begin_of_text|>B.caapi contains harmine, harmaline, and tetrahydroharmine, all of which are both beta-carboline harmala alkaloids and Monoamine oxidase inhibitors (MAOIs). The MAOIs in B. caapi allow the primary psychoactive compound, DMT (which is introduced from the other primary ingredient in ayahausca, the Psychotria viridis plant), to be orally active. + +The name ayahuasca means "vine of the soul"and the shamans of the indigenous western Amazonian tribes use the plant in religious and healing ceremonies. In addition to its hallucinogenic properties, caapi is used for its healing properties as a purgative, effectively cleansing the body of parasites and helping the digestive tract. + +Harmala alkaloids are short term yet powerful MAOIs which render tryptamines orally active by temporarily reducing levels of monoamine oxidase in the body which otherwise rapidly destroys them. Their effects are more powerful and less toxic than pharmaceutical SSRIs since they do not lead to suicidal tendencies and other side effects which detrimentally affect human behavior. + +The principal ayahuasca compounds have a common indole structure which, through several mechanisms, influences certain functions of the central nervous system (CNS). The relevant factor is the biochemical similarity of these compounds to the neurotransmitter serotonin (5-HT). The harmala alkaloids in ayahuasca, primarily harmine and tetrahydroharmine, reversibly inhibit the neuronal enzyme monoamine oxidase (MAO). + +This allows DMT to be active when ingested orally. It also facilitates accumulation of biogenic amines, such as 5-HT, which are normally metabolized by monoamine oxidase enzymes. DMT is a naturally-occurring biochemical substance secreted by the human body in the pineal gland. It occurs in hundreds of plant species worldwide. It can produce very powerful visionary effects when smoked in its pure form or taken orally in Ayahuasca. + +The Ayahuasca Experience + +People who have consumed ayahuasca report having spiritual revelations regarding their purpose on earth, the true nature of the universe as well as deep insight into how to be the best person they possibly can. This is viewed by many as a spiritual awakening and what is often described as a rebirth. In addition it is often reported that individuals can gain access to higher spiritual dimensions and make contact with various spiritual or extra dimensional beings who can act as guides or healers. + +Wild cats including Jaguars looking for a high will seek out +================================================================================ +Rank = 47; Score = 778240.0 +<|begin_of_text|>CHENNAI: The Enforcement Directorate, which enforces the Prevention of Money Laundering Act (PMLA), has provisionally attached properties worth Rs 12,000 crore in the last 15 months, believed to be the proceeds of laundered money, data released by the government shows. This is more than what the department attached in the decade from 2005.In a written reply to an unstarred question on Friday, minister of state for finance Santosh Kumar Gangwar said that in the last one year, properties worth Rs 11,032.27 crore were attached by the ED after issuing 175 Provisional Attachment Orders (PAO). In comparison, the total value of assets attached from 2005-2015 was Rs 9003.26 crore, statistics on ED website show. In the first three months of the current financial year, Rs 965.84 crore has been attached. The total cumulative attachments would be around Rs 22,000 crore as of June 30 this year, said sources in the department. The figures of properties attached in 2016-17 zoomed due to the Vijay Mallya case, in which properties worth close to Rs 10,000 crore were attached, sources said.More than half of the total searches and a third of the total arrests by the department across the country were effected in the last 15 months, statistics show. Some of these highprofile cases and searches were in Tamil Nadu; for instance the Sekhar Reddy case and Madurai granite mining. The ED is an organisation which goes after big names accused of corruption. The PMLA is a stringent law, where the onus is on the accused to prove that the attached properties are bought from known sources of income, say sources. The ED has been given a free hand to dig up dirty money, said sources. This, they added, reflected in the figures.Sources in the department said that these measures were taken in coordination with other investigative agencies like Income Tax Investigation Wing and Central Bureau of Investigation (CBI). A task force on shell companies had been created by the ministry of corporate affairs which involves all these departments. The ED is tracking shell companies through which crores of rupees are transferred out of the country. In April, ED sleuths arrested in Chennai a 36-year-old man who had allegedly routed Rs 78 crore through six shell companies.As of July 12, 1.62 lakh shell companies had been struck off the Registrar of Companies (RoC). +================================================================================ +Rank = 48; Score = 774144.0 +<|begin_of_text|>Everyone knows that the “genuine designer handbag” going for $20 from a street vendor is neither genuine nor designer, and indeed may not even hold up as a bag. But when you go to a reputable retailer and spend what it costs to replace the tires on your car, you expect to get what the real goods. Alas, Consumer Reports has found: just because there’s a brand name you know on the outside of a tire, doesn’t mean you’re getting what you should be. + +Consumer Reports (our parent organization) discovered the counterfeit tires while conducting otherwise routine tire performance tests. One tire brand, American Pacific Industries’ Pegasus Advanta for SUVs, performed very poorly in the tests and received a low rating. + +The company then contacted Consumer Reports to ask about the way the tires were tested and to ask which batch CR had used. + +That’s when the story gets interesting. It turned out that the tires being sold under API’s Pegasus brand were not what they appeared to be. The legally required date codes stamped into the tires indicated they had been made for API in a specific Chinese factory in August and December 2012. + +To which API replied: “American Pacific Industries’ relationship with this factory ended in 2011 [and] our records indicate the last shipment of these tires in the SUV pattern was in December of 2011.” + +Further, API said, the factory had been destroyed after that contract concluded, and many of their molds went missing. “We have no idea who may have made these tires nor what they put in them,” the company’s COO told Consumer Reports. + +Counterfeit tires pose two big challenges for consumers. One is that their quality may be lacking — as these were — and consumers aren’t getting what they paid for. And the other challenge is that in the event of a safety defect or a recall, consumers don’t really have any recourse. + +Ordinarily, if there is a safety complaint about tires the National Highway Traffic Safety Administration issues a recall and the manufacturer has to replace the item. But if the problem with the item is that the manufacturer did not in fact actually manufacture them… then who would be responsible for a recall? It’s not an action that NHTSA could actually take. + +But there are federal actions available. It turns out that counterfeited imported goods fall under the purview of Immigration and Customs Enforcement. An ICE official told Consumer Reports, that if counterfeit tires are being made and smuggled, “We would investigate and try to mount an interdiction and +================================================================================ +Rank = 49; Score = 774144.0 +<|begin_of_text|>SAO PAULO, Jan 26 (Reuters) - Economists raised sharply their forecasts for Brazil's 2015 inflation rate to a level further exceeding the official target, a weekly central bank poll showed on Monday. Inflation is expected to end 2015 at 6.99 percent, up from 6.67 percent in the prior week's survey and 6.53 percent a month ago, according to the median forecast of about 100 market economists. Brazil's government targets an inflation rate of 4.5 percent, with a tolerance margin of two percentage points. The median estimate for economic growth this year dropped to 0.13 percent from 0.38 percent in the previous week's survey and 0.55 percent a month earlier. (pct) 2015 2016 previous new previous new forecast forecast forecast forecast Consumer inflation 6.67 6.99 5.70 5.60 Exchange rate 2.80 2.80 2.85 2.90 (reais per U.S dollar, end-period) Interest rate 12.50 12.50 11.50 11.50 (end-period) GDP growth 0.38 0.13 1.80 1.54 Industrial output 0.71 0.69 2.65 2.50 (Reporting by Asher Levine; Editing by Toby Chopra)<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 50; Score = 774144.0 +<|begin_of_text|>Private cast iron bathtubs with porcelain interiors on "claw foot" pedestals rose to popularity in the 19th century + +A bathtub, bath, or tub (informal) is a large or small container for holding water in which a person or animal may bathe. Most modern bathtubs are made of thermoformed acrylic, porcelain enameled steel, fiberglass-reinforced polyester, or porcelain enameled cast iron. A bathtub is usually placed in a bathroom either as a stand-alone fixture or in conjunction with a shower. + +Modern bathtubs have overflow and waste drains and may have taps mounted on them. They are usually built-in, but may be free-standing or sometimes sunken. Until recently, most bathtubs were roughly rectangular in shape, but with the advent of acrylic thermoformed baths, more shapes are becoming available. Bathtubs are commonly white in colour, although many other colours can be found. The process for enamelling cast iron bathtubs was invented by the Scottish-born American David Dunbar Buick. + +Two main styles of bathtub are common: + +Western style bathtubs in which the bather lies down. These baths are typically shallow and long. + +Eastern style bathtubs in which the bather sits up. These are known as furo in Japan and are typically short and deep. + +History of bathtubs and bathing [ edit ] + +Traditional bathtub (19th century) from Italy + +Documented early plumbing systems for bathing go back as far as around 3300 BC with the discovery of copper water pipes beneath a palace in the Indus Valley Civilization of ancient India; see sanitation of the Indus Valley Civilization.[citation needed] Evidence of the earliest surviving personal sized bath tub was found on the Isle of Crete where a 1.5-metre (5 ft) long pedestal tub was found built from hardened pottery.[1] + +The clawfoot tub, which reached the apex of its popularity in the late 19th century; had its origins in the mid 18th century, where the ball and claw design originated in Holland, possibly artistically inspired by the Chinese motif of a dragon holding a precious stone. The design spread to England where it found much popularity among the aristocracy, just as bathing was becoming increasingly fashionable. Early bathtubs in England tended to be made of cast iron, or even tin and copper with a face of paint applied that tended to peel with time. + +The Scottish-born inventor David Buick invented a process for bonding porcelain enamel to cast iron in the +================================================================================ +Rank = 51; Score = 761856.0 +<|begin_of_text|>Natural number + +1,000,000,000 (one billion, short scale; one thousand million or milliard, yard,[1] long scale) is the natural number following 999,999,999 and preceding 1,000,000,001. One billion can also be written as b or bn.[2][3] + +In scientific notation, it is written as 1 × 109. The metric prefix giga indicates 1,000,000,000 times the base unit. Its symbol is G. + +One billion years may be called eon/aeon in astronomy or geology. + +Previously in British English (but not in American English), the word "billion" referred exclusively to a million millions (1,000,000,000,000). However, this is no longer common, and the word has been used to mean one thousand million (1,000,000,000) for several decades.[4] + +The term milliard can also be used to refer to 1,000,000,000; whereas "milliard" is seldom used in English,[5] variations on this name often appear in other languages. + +In the South Asian numbering system, it is known as 100 crore or 1 arab. + +Visualization of powers of ten from one to 1 billion + +Sense of scale [ edit ] + +The facts below give a sense of how large 1,000,000,000 (109) is in the context of time according to current scientific evidence: + +Time [ edit ] + +10 9 seconds is 114 days short of 32 calendar years (≈ 31.7 years). + +seconds is 114 days short of 32 calendar years (≈ 31.7 years). More precisely, a billion seconds is exactly 31 years, 8 months, 2 weeks, 1 day, 17 hours, 46 minutes, and 40 seconds. + +About 10 9 minutes ago, the Roman Empire was flourishing and Christianity was emerging. (10 9 minutes is roughly 1,901 years.) + +minutes ago, the Roman Empire was flourishing and Christianity was emerging. (10 minutes is roughly 1,901 years.) About 10 9 hours ago, modern human beings and their ancestors were living in the Stone Age (more precisely, the Middle Paleolithic). (10 9 hours is roughly 114,080 years.) + +hours ago, modern human beings and their ancestors were living in the Stone Age (more precisely, the Middle Paleolithic). (10 hours +================================================================================ +Rank = 52; Score = 757760.0 +<|begin_of_text|>The balance of trade, commercial balance, or net exports (sometimes symbolized as NX), is the difference between the monetary value of a nation's exports and imports over a certain period.[1] Sometimes a distinction is made between a balance of trade for goods versus one for services. The balance of trade measures a flow of exports and imports over a given period of time. The notion of the balance of trade does not mean that exports and imports are "in balance" with each other. + +If a country exports a greater value than it imports, it has a trade surplus or positive trade balance, and conversely, if a country imports a greater value than it exports, it has a trade deficit or negative trade balance. As of 2016, about 60 out of 200 countries have a trade surplus. The notion that bilateral trade deficits are bad in and of themselves is overwhelmingly rejected by trade experts and economists.[2][3][4][5][6] + +Explanation [ edit ] + +Balance of trade in goods and services (Eurozone countries) + +US trade balance from 1960 + +U.S. trade balance and trade policy (1895–2015) + +U.S. trade deficit (in billions, goods and services) by country in 2017 + +U.K. balance of trade in goods (since 1870) + +The balance of trade forms part of the current account, which includes other transactions such as income from the net international investment position as well as international aid. If the current account is in surplus, the country's net international asset position increases correspondingly. Equally, a deficit decreases the net international asset position. + +The trade balance is identical to the difference between a country's output and its domestic demand (the difference between what goods a country produces and how many goods it buys from abroad; this does not include money re-spent on foreign stock, nor does it factor in the concept of importing goods to produce for the domestic market). + +Measuring the balance of trade can be problematic because of problems with recording and collecting data. As an illustration of this problem, when official data for all the world's countries are added up, exports exceed imports by almost 1%; it appears the world is running a positive balance of trade with itself. This cannot be true, because all transactions involve an equal credit or debit in the account of each nation. The discrepancy is widely believed to be explained by transactions intended to launder money or evade taxes, smuggling and other visibility problems. Especially for developing countries, the transaction statistics are likely to be inaccurate. + +Factors +================================================================================ +Rank = 53; Score = 753664.0 +<|begin_of_text|>U.S. may face deflation, a problem Japan understands too well + +Economists worry that America could be edging closer to the trap that cost the other nation more than a decade of growth. + +But soon, lower prices cut into business profits, and managers begin to trim payrolls. That in turn undermines consumers' buying power, leading to more pressure on profits, jobs and wages — as well as cutbacks in expansion and in the purchase of new plants and equipment. + +When deflation begins, prices fall. At first that seems like a good thing. + +And as Tokyo's experience suggests, deflation can be at least as tough a problem as the soaring prices of inflation or the financial pain of a traditional recession. + +But increasingly, economists and other analysts are expressing concern that the United States could be edging closer to a different problem — the kind of deflationary trap that cost Japan more than a decade of growth and economic progress. + +Reporting from Washington — The White House prediction Friday that the deficit would hit a record $1.47 trillion this year poured new fuel on the fiery argument over whether the government should begin cutting back to avoid future inflation or instead keep stimulating the economy to help the still-sputtering recovery. + +Also, consumers who are financially able to buy often wait for still lower prices, adding to the deflationary trend. + +All these factors feed on one another, setting off a downward spiral that can be as hard to escape from as a stall in an airplane. + +For now, the dominant theme of the nation's economic policy debate remains centered on the comparative dangers of deficits and inflation. However, economists across the political spectrum — here and abroad — are talking more often about the potential for deflation. + +So how likely is the problem? + +The latest U.S. data are sobering: Consumer prices overall have declined in each of the last three months, putting the inflation index in June just 1.1% above a year earlier. The core inflation rate — a better gauge of where prices are going because it excludes volatile energy and food items — has dropped to a 44-year low of 0.9%. + +That's well below the 1.5%-to-2% year-over-year inflation that the Federal Reserve likes to see, and some Fed policymakers have raised concerns about the rising risk of a broad decline in prices. + +Private economists and financial experts have expressed much greater concern. + +"I think we have to take it seriously," said John Mauldin, president of Millennium Wave Advisors in Dallas, who puts the probability of def +================================================================================ +Rank = 54; Score = 745472.0 +<|begin_of_text|>BRITISH PRIME MINISTER Theresa May has raised the country’s terror threat level from severe to critical. + +Critical is the highest level possible in the UK and means that an attack is “expected imminently”. + +The threat level is decided by the Joint Terrorism Analysis Centre (JTAC). The last time it reached critical was in 2007, having been severe since 2014. + +Prime Minister Theresa May said that the decision had been made following investigations today. + +“It has now been concluded that the threat levels should be increased for the time being. + +“This means that [the JTAC's] assessment is not only that an attack is highly likely, but that another attack is imminent + +“It is a possibility we cannot ignore that there is a wider group of individuals linked to this attack.” + +May added that military personnel will be deployed to Britain’s streets to support armed police officers under Operation Tempora. + +Armed personnel will be visible at big events such as concerts and sporting events, she said. + +Attack + +Source: PA Wire/PA Images + +The move comes as 22 people were confirmed dead after last night’s suicide bomb attack on an Ariana Grande concert in Manchester. + +Manchester police identified the suspect behind the attack as 22-year-old Salman Abedi. + +Police staged an armed raid on a Manchester address believed to be where Abedi lived, carrying out a controlled explosion to gain entry after arresting a 23-year-old man earlier Tuesday in connection with the attack. + +“A single terrorist detonated his improvised explosive device near one of the exits of the venue, deliberately choosing the time and place to cause maximum carnage and to kill and injure indiscriminately,” May said after an emergency ministerial meeting. + +The attack was the deadliest in Britain since July 7, 2005 when four suicide bombers inspired by Al-Qaeda attacked London’s transport system during rush hour, killing 52 people and wounding 700 more. + +The Islamic State group has claimed responsibility for the attack.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 55; Score = 745472.0 +<|begin_of_text|>Low rolling resistance tires are designed to reduce the energy loss as a tire rolls, decreasing the required rolling effort — and in the case of automotive applications, improving vehicle fuel efficiency. Approximately 5–15% of the fuel consumed by a typical car may be used to overcome rolling resistance.[1] A 2003 California Energy Commission (CEC) preliminary study estimated that adoption of low-rolling resistance tires could save 1.5–4.5% of all gasoline consumption, but that current data were also insufficient to compare safety and other characteristics.[2] + +A United States National Highway Traffic Safety Administration (NHTSA) study in 2009 found that if 2% of the replacement tires would reduce their rolling resistance by 5%, there would be 7.9 million gallons fuel and 76,000 metric tons of CO2 saved annually. + +(SAE J1269 and SAE J2452) performed on new tires. + +Measuring rolling resistance in tires [ edit ] + +Rolling resistance can be expressed by the rolling resistance coefficient (RRC or C rr ), which is the value of the rolling resistance force divided by the wheel load. A lower coefficient means the tires will use less energy to travel a certain distance. The coefficient is mostly considered as independent of speed, but for precise calculations it is tabled at several speeds or an additional speed-dependent part is used. The Society of Automotive Engineers (SAE) has developed test practices to measure the RRC of tires. These tests (SAE J1269 and SAE J2452) are usually performed on new tires. + +When measured by using these standard test practices, most new passenger tires have reported RRCs ranging from 0.007 to 0.014.[3] In the case of bicycle tires, values of 0.0025 to 0.005 are achieved.[4] These coefficients are measured on rollers, with power meters on road surfaces, or with coast-down tests. In the latter two cases, the effect of air resistance must be subtracted or the tests performed at very low speeds. In 2009 The CEC used a rating called Rolling Resistance Force RRF. RRF and RRC, rolling resistance coefficient are very similar. Difference is taking the RRF and dividing it by the load(weight) to get RRC. So a Michelin Harmony tire rated at 9.45 RRF at 1000 pounds load would be.0095 RRC.[5][6] + +In Canada, Transport Canada tests will be conducted +================================================================================ +Rank = 56; Score = 745472.0 +<|begin_of_text|>(Updates with link to most and least expensive cities to live in the U.S.) + +For the first time, the Commerce Department’s Bureau of Economic Analysis released price-adjusted estimates of personal income for states and metropolitan areas for 2008 to 2012, and it shows that, at the very least, it’s more expensive for Members of Congress to live in the nation’s capital on their $174,000 salary than in any state from which they hail. + +The results show the District of Columbia, in 2012, had the highest “regional price parity” of any state. Granted, D.C.’s really a city, not a state, and set against their more natural comparisons, it ranks fifth, behind the Urban Honolulu area, New York-Newark-Jersey City, San Jose-Sunnyvale-Santa Clara, Bridgeport-Stamford-Norwalk and San Francisco-Oakland-Hayward. Beaches, hedge funds and technology are the key to prices, evidently. + +The cheapest metro areas, by this methodology, were Danville, Ill.; Jefferson City, Mo.; Jackson, Tenn.; Jonesboro, Ark.; and Rome, Ga. + +See related look at most and least expensive cities to live in the U.S. + +Put another way,it costs 54% more to live in Honolulu than in Danville. + +The top five states (outside of D.C.) were Hawaii, New York, New Jersey, California and Maryland, and the bottom five states were Mississippi, Arkansas, Alabama, Missouri and South Dakota. So, it costs 36% more to live in Hawaii than in Mississippi. + +The BEA says it produced the report in large part so people and businesses can have a better understanding how their personal income may be impacted by a job change or move. There is regional consumer price index data, but those statistics aren’t designed for regional or city comparisons, whereas this data is, a spokesman says. + +— Steve Goldstein + +Follow Steve on Twitter @mktwgoldstein + +Follow Capitol Report @capitolreport + +Also read: Is it time to freak out about housing? + +Why the housing sector won’t save the broader economy<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 57; Score = 745472.0 +<|begin_of_text|>At its meeting today, the Board decided to reduce the cash rate by 25 basis points to 3.0 per cent, effective 5 December 2012. + +Global growth is forecast to be a little below average for a time. Risks to the outlook are still seen to be on the downside, largely as a result of the situation in Europe, though the uncertainty over the course of US fiscal policy is also weighing on sentiment at present. Recent data suggest that the US economy is recording moderate growth and that growth in China has stabilised. Around Asia generally, growth has been dampened by the more moderate Chinese expansion and the weakness in Europe. + +Key commodity prices for Australia remain significantly lower than earlier in the year, though trends have been more mixed over the past few months. The terms of trade have declined by about 15 per cent since the peak, to a level that is still historically high. + +Sentiment in financial markets remains better than it was in mid year, in response to signs of progress in addressing Europe's financial problems, though Europe is likely to remain a source of instability for some time. Long-term interest rates faced by highly rated sovereigns, including Australia, remain at exceptionally low levels. Capital markets remain open to corporations and well-rated banks, and Australian banks have had no difficulty accessing funding, including on an unsecured basis. Borrowing conditions for large corporations are similarly attractive and share prices have risen since mid year. + +In Australia, most indicators available for this meeting suggest that growth has been running close to trend over the past year, led by very large increases in capital spending in the resources sector, while some other sectors have experienced weaker conditions. Looking ahead, recent data confirm that the peak in resource investment is approaching. As it does, there will be more scope for some other areas of demand to strengthen. + +Private consumption spending is expected to grow, but a return to the very strong growth of some years ago is unlikely. Available information suggests that the near-term outlook for non-residential building investment, and investment generally outside the resources sector, remains relatively subdued. Public spending is forecast to be constrained. On the other hand, there are indications of a prospective improvement in dwelling investment, with dwelling prices moving a little higher, rental yields increasing and building approvals having turned up. + +Inflation is consistent with the medium-term target, with underlying measures at around 2½ per cent. The introduction of the carbon price affected consumer prices in the September quarter, and there could be some further small effects over the next couple of quarters. Partly +================================================================================ +Rank = 58; Score = 745472.0 +<|begin_of_text|>Making coffee to drink + +For the agricultural and industrial processes for producing whole coffee beans, see Coffee processing + +Filter coffee being brewed + +Coffee preparation is the process of turning coffee beans into a beverage. While the particular steps vary with the type of coffee and with the raw materials, the process includes four basic steps: raw coffee beans must be roasted, the roasted coffee beans must then be ground, the ground coffee must then be mixed with hot water for a certain time (brewed), and finally the liquid coffee must be separated from the used grounds. + +Coffee is usually brewed immediately before drinking. In most areas, coffee may be purchased unprocessed, or already roasted, or already roasted and ground. Coffee is often vacuum packed to prevent oxidation and lengthen its shelf life. + +Roasting [ edit ] + +Dutch coffee-roasting machine, c. 1920 + +Roasting coffee transforms the chemical and physical properties of green coffee beans. When roasted, the green coffee bean expands to nearly double its original size, changing in color and density. As the bean absorbs heat, its color shifts to yellow, then to a light "cinnamon" brown, and then to a rich dark brown color. During roasting, oils appear on the surface of the bean. The roast will continue to darken until it is removed from the heat source. + +Coffee can be roasted with ordinary kitchen equipment (frying pan, grill, oven, popcorn popper) or by specialised appliances. A coffee roaster is a special pan or apparatus suitable to heat up and roast green coffee beans. + +Grinding [ edit ] + +An old-fashioned manual burr-mill coffee grinder + +Wheel coffee grinder + +Coffee grinder + +The whole coffee beans are ground, also known as milling, to facilitate the brewing process. + +The fineness of the grind strongly affects brewing. Brewing methods that expose coffee grounds to heated water for longer require a coarser grind than faster brewing methods. Beans that are too finely ground for the brewing method in which they are used will expose too much surface area to the heated water and produce a bitter, harsh, "over-extracted" taste. At the other extreme, an overly coarse grind will produce weak coffee unless more is used. Due to the importance of a grind's fineness, a uniform grind is highly desirable. + +If a brewing method is used in which the time of exposure of the ground coffee to the heated water is adjustable, then a short brewing time can be used for finely ground coffee. This produces coffee of equal flavor yet uses less ground coffee. A blade grinder does not +================================================================================ +Rank = 59; Score = 741376.0 +<|begin_of_text|>The brain - awake and sleeping - is awash in electrical activity, and not just from the individual pings of single neurons communicating with each other. In fact, the brain is enveloped in countless overlapping electric fields, generated by the neural circuits of scores of communicating neurons. The fields were once thought to be an "epiphenomenon, a 'bug' of sorts, occurring during neural communication," says neuroscientist Costas Anastassiou, a postdoctoral scholar in biology at the California Institute of Technology (Caltech). + +New work by Anastassiou and his colleagues, however, suggests that the fields do much more - and that they may, in fact, represent an additional form of neural communication. + +"In other words," says Anastassiou, the lead author of a paper about the work appearing in the journal Nature Neuroscience, "while active neurons give rise to extracellular fields, the same fields feed back to the neurons and alter their behavior," even though the neurons are not physically connected - a phenomenon known as ephaptic coupling. "So far, neural communication has been thought to occur at localized machines, termed synapses. Our work suggests an additional means of neural communication through the extracellular space independent of synapses." + +Extracellular electric fields exist throughout the living brain, though they are particularly strong and robustly repetitive in specific brain regions such as the hippocampus, which is involved in memory formation, and the neocortex, the area where long-term memories are held. "The perpetual fluctuations of these extracellular fields are the hallmark of the living and behaving brain in all organisms, and their absence is a strong indicator of a deeply comatose, or even dead, brain," Anastassiou explains. + +Previously, neurobiologists assumed that the fields were capable of affecting - and even controlling - neural activity only during severe pathological conditions such as epileptic seizures, which induce very strong fields. Few studies, however, had actually assessed the impact of far weaker - but very common - non-epileptic fields. "The reason is simple," Anastassiou says. "It is very hard to conduct an in vivo experiment in the absence of extracellular fields," to observe what changes when the fields are not around. + +To tease out those effects, Anastassiou and his colleagues, including Caltech neuroscientist Christof Koch, the Lois and Victor Troendle Professor of Cognitive and Behavioral Biology and professor of computation and neural systems, focused on strong but slowly oscillating fields, called local field potentials (L +================================================================================ +Rank = 60; Score = 724992.0 +<|begin_of_text|>The gustatory system creates the human sense of taste, allowing us to perceive different flavors from substances that we consume as food and drink. Gustation, along with olfaction (the sense of smell), is classified as chemoreception. Because it functions by reacting with molecular chemical compounds in a given substance. Specialized cells in the gustatory system that are located on the tongue are called taste buds, and they sense tastants (taste molecules). The taste buds send the information after coming in contact with food from the tastants to the brain, where a molecule is processed as a certain taste. There are five main tastes: bitter, salty, sweet, sour, and umami (savory). All the varieties of flavor we experience are a combination of some or all of these tastes. + +The sense of taste is transduced by taste buds. Which are clusters of 50-100 taste receptor cells located in the tongue, soft palate, epiglottis, pharynx, and esophagus. The tongue is the main sensory organ of the gustatory system. The tongue contains papillae, or specialized epithelial cells, which have taste buds on their surface. There are three types of papillae with taste buds in the human gustatory system: + +Loading... + +fungiform papillae, which are mushroom-shaped and located at the tip of the tongue; + +foliate papillae, which are ridges and grooves toward the back of the tongue; + +circumvallate papillae, which are circular-shaped and located in a row just in front of the end of the tongue. + +You can’t taste food unless it’s mixed with your saliva + +Each taste bud is flask-like in shape and formed by two types of cells: supporting cells and gustatory cells. Gustatory cells are short-lived and are continuously regenerating. They each contain a taste pore at the surface of the tongue which is the site of sensory transduction. Though there are small differences in sensation, all taste buds, no matter their location, can respond to all types of taste.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 61; Score = 720896.0 +<|begin_of_text|>What is Spending Skill? + +Summary: We measure your Spending Skill as a league score, ranging from Bronze to Grandmaster. For each game you play, your Average Unspent Resources (AUR) and Income are combined into a single number (called Spending Quotient, also known as SQ). Your Spending Skill is based on typical SQ levels for your race, region (NA, EU, etc) and game length. As you can see from the Spending Skill Stats, long games tend to have lower SQs. For example, an SQ of 76 for an EU Protoss in a 27-minute game is a Grandmaster-level performance, but in a 7-minute game that same SQ is only Platinum level. and then similarity-checked against our Starcraft library of more than 100,000 1v1 ladder games. + +There's an excellent article called Do you macro like a pro? which studies 2,100 games and proves that pros spend their resources better than noobs. + +Spending can make a big difference. Let's say you and your enemy have the same income, and the same tech, but you have 1,000 unspent resources and she has 0. Then her army is going to be 1,000 resources bigger than yours. That's 20 extra marines. Assuming her micro skill is equal to yours, then she's going to win. + +Don't let her win!! + +After every game, the Score Screen shows your Average Unspent Resources (AUR). But that one number doesn't tell you if you did a good or bad job on spending. + +An AUR of 1,000 is great when your Resource Collection Rate (aka Income) is 1,900, but it's not good when your Income is 400. So AUR alone isn't enough to know if your Spending was good. + +Do you macro like a pro? introduced the idea of Spending Quotient (SQ), a formula that combines AUR and Income into a single score, usually between 40 and 100, that measures how well you're spending. It's a nice formula, it's even got a logarithm in it. + +Here's a calculator you can use to play with the SQ formula. + +Resource Collection Rate (Income) Average Unspent Resources Spending Quotient {{ sq }} + +Grandmasters and pros get SQ around 100, and bronze leaguers get 40 or 50. + +However, when you get the data together for over 100,000 games*, the fuller picture emerges: + + +================================================================================ +Rank = 62; Score = 720896.0 +<|begin_of_text|>Image copyright Getty Images Image caption Singapore has been moving up the ranks of the world's most expensive cities to live in over the last decade + +Singapore has topped 131 cities globally to become the world's most expensive city to live in 2014, according to the Economist Intelligence Unit (EIU). + +The city's strong currency combined with the high cost of running a car and soaring utility bills contributed to Singapore topping the list. + +It is also the most expensive place in the world to buy clothes. + +Singapore replaces Tokyo, which topped the list in 2013. + +Other cities making up the top five most expensive cities to live in are Paris, Oslo, Zurich and Sydney, with Tokyo falling to sixth place. + +The EIU's Worldwide Cost of Living Survey is a relocation tool that uses New York city as a base. It looks at more than 400 individual prices. + +Top 5 most expensive cities Image copyright Getty Images Singapore, Singapore Paris, France Oslo, Norway Zurich, Switzerland Sydney, Australia + +Soaring Asia + +The top 10 cities this year have been dominated by Asian and Australasian cities as well as some in Europe. + +"Improving sentiment in structurally expensive European cities combined with the continued rise of Asian hubs means that these two regions continue to supply most of the world's most expensive cities," said the editor of the report, Jon Copestake. + +"But Asian cities also continue to make up many of the world's cheapest, especially in the Indian subcontinent." + +Most Asian cities that top the list are there for predominantly higher costs of groceries. Tokyo is still at the top of the list for everyday food items. + +Inexpensive India + +However, not all Asian cities are tough on the wallet. + +India's major cities - including Mumbai and New Delhi - were found to be among the least expensive in the world. + +Mumbai's prices are kept low by large income inequality. + +The low wages of many of the city's workers keep spending low, and government subsidies have helped them stay that way. + +Outside of the subcontinent, Damascus in Syria saw the largest drop, becoming the fourth cheapest city in the world as the country's ongoing conflict has led to plummeting prices. + +Media playback is unsupported on your device Media caption Watch: Toby Iles from the Economist Intelligence Unit (EIU) explains to Sharanjit Leyl why Singapore is the most expensive city to live in + +While the EIU's survey takes into account the cost of living, other firms employ different research methods. + +Mercer conducts research to determine the most expensive cities for +================================================================================ +Rank = 63; Score = 716800.0 +<|begin_of_text|>Should we expect prices to fall all the way back? Well, in the late 1980s, Los Angeles experienced a large localized housing bubble: real home prices rose about 50 percent before the bubble popped. Home prices then proceeded to fall by a quarter, which combined with ongoing inflation brought real housing prices right back to their prebubble level. + +And here’s the thing: this process took more than five years — L.A. home prices didn’t bottom out until the mid-1990s. If the current housing slump runs on the same schedule, we won’t be seeing a recovery until 2011 or later. + +What about the broader economy? You might be tempted to take comfort from the fact that the last two recessions, in 1990-1991 and 2001, were both quite short. But in each case, the official end of the recession was followed by a long period of sluggish economic growth and rising unemployment that felt to most Americans like a continued recession. + +Photo + +Thus, the 1990 recession officially ended in March 1991, but unemployment kept rising through much of 1992, allowing Bill Clinton to win the election on the basis of the economy, stupid. The next recession officially began in March 2001 and ended in November, but unemployment kept rising until June 2003. + +These prolonged recession-like episodes probably reflect the changing nature of the business cycle. Earlier recessions were more or less deliberately engineered by the Federal Reserve, which raised interest rates to control inflation. Modern slumps, by contrast, have been hangovers from bouts of irrational exuberance — the savings and loan free-for-all of the 1980s, the technology bubble of the 1990s and now the housing bubble. + +Ending those old-fashioned recessions was easy because all the Fed had to do was relent. Ending modern slumps is much more difficult because the economy needs to find something to replace the burst bubble. + +Newsletter Sign Up Continue reading the main story Please verify you're not a robot by clicking the box. Invalid email address. Please re-enter. You must select a newsletter to subscribe to. Sign Up You will receive emails containing news content, updates and promotions from The New York Times. You may opt-out at any time. You agree to receive occasional updates and special offers for The New York Times's products and services. Thank you for subscribing. An error has occurred. Please try again later. View all New York Times newsletters. + +The Fed, in particular, has a hard time getting traction in +================================================================================ +Rank = 64; Score = 712704.0 +<|begin_of_text|>All funding for environment programs to end, as Coalition focuses aid on countries it needs to support its asylum policy + +Aid groups are accusing the Coalition of breaking an election commitment after it revealed their funding would be cut mid-year as part of a $650m reduction in the former government’s budgeted foreign aid spending, leaving 2013-14 spending $107m below what was spent last year. + +Foreign minister Julie Bishop announced the cuts for the groups as well as a complete defunding of international environmental programs. The government is redirecting a pared back aid budget towards the region and maintenance of spending on countries such as Papua New Guinea, Indonesia and Nauru, whose co-operation is necessary for the success of its asylum policy. + +Organisations such as Care, Save the Children, Caritas, ChildFund, Plan International and the Fred Hollows Foundation – who have partnership agreements with the government – have had their current year funding cut by about 8%. + +They say that means they are losing money already allocated to programs related to water and sanitation, elimination of violence against women, disaster reduction work and small-scale agriculture, among others. + +The organisations say the cuts clearly break a Coalition promise not to cut their funding when it announced a $4.5bn cut to the aid budget over the next four years two days before the federal election. + +At the time of that cut, the then shadow treasurer, Joe Hockey, and shadow finance spokesman, Andrew Robb, said “the Coalition will reprioritise foreign aid allocations towards non-government organisations that deliver on-the-ground support for those most in need”. + +“That will also mean putting more money into NGOs who are on the ground and who can deliver aid more efficiently than through AusAID or indeed through some of the multilaterals that we’ve been putting money into in increasing numbers because AusAID cannot handle the increases in the budget,” then foreign affairs spokeswoman Julie Bishop said. + +The government said it needed to redirect the $4.5bn towards domestic infrastructure. The cut was from projected expenditure – with the aid budget continuing to grow in line with inflation. + +“Under the previous Labor government, the rapid growth in Australia’s aid budget was neither targeted nor sustainable... This year’s aid expenditure will be $107m less than last year. From 2014-15 the $5bn aid budget will grow each year in line with the consumer price index,” Bishop said in a statement on Saturday, “confirming” the cuts after they were announced in the Australian newspaper. + +She said the government would now consult NGOs +================================================================================ +Rank = 65; Score = 700416.0 +<|begin_of_text|>WASHINGTON (Reuters) - Add another economic worry to inflation and deflation: ecoflation, the rising cost of doing business in a world with a changing climate. + +Ecoflation could hit consumer goods hard in the next five to 10 years, according to a report by World Resources Institute and A.T. Kearney, a global management consulting firm. + +Companies that make fast-moving consumer goods, everything from cereal to shampoo, could see earnings drop by 13 percent to 31 percent by 2013 and 19 percent to 47 percent by 2018 if they do not adopt sustainable environmental practices, the report said. + +The costs of global warming are showing up now in the form of worse heat waves, droughts, wildfires and possibly more severe tropical storms but they are not yet reflected in consumer prices, said the institute’s Andrew Aulisi after the report’s December 2 release. + +Instead, these costs are paid by governments and society, Aulisi said in a telephone interview. That could change if President-elect Barack Obama and the U.S. Congress push for a system that puts a price on the emission of climate-warming carbon dioxide, Aulisi said. + +This is unlikely to happen next year in time for a December 2009 deadline to craft an international pact to fight climate change but it is more likely to happen in 2010. + +These rising costs and possible tightening regulation of greenhouse gas emissions are not necessarily a bad thing, he said. + +“The message we don’t see in this study is that regulation is going to cost... a lot of money,” Aulisi said. “We think the analysis is a catalyst to convince companies to take greater action on these important issues.” + +LESS PLASTIC + +In fact, some companies are already looking at ways to cut their emissions in advance of any new regulation, said Daniel Mahler of A.T. Kearney. + +One example is consumer giant Procter & Gamble, which has a team looking across the company’s varied laundry, hair-care and health-care businesses to see how they can use less plastic, a fossil-based material, Mahler said by telephone. + +But the changes may need to go deeper and wider, he said, spreading to the basics of how supply chains are managed. + +For instance, companies that presumed U.S. transportation costs would be low and U.S. labor costs would be high had their goods made in countries where employees would work for less. But a new cost to the carbon emitted by long-distance transport could change that equation, making foreign manufacturing less attractive, +================================================================================ +Rank = 66; Score = 692224.0 +<|begin_of_text|>By Kyle Orton (@KyleWOrton) on February 17, 2017 + +Four days ago, Chapo Trap House, a Left-wing politics and humour podcast, hosted Brace Belden, known to Twitter as “PissPigGranddad,” a 27-year-old from San Francisco who has joined the Syrian Kurdish militia, the People’s Protection Units (YPG). It was very interesting and informative on the state of play in northern Syria. + +The YPG is run by the Democratic Union Party (PYD) front of the Kurdistan Workers’ Party (PKK). The most amusing part of the interview is Belden’s formal maintenance that the YPG, while fraternal comrades to the PKK and admirers of their ideology, have absolutely no organizational links at all, while at the same time letting the audience in on the fact that the YPG and indeed the Syrian Democratic Forces (SDF) coalition that it controls are parts of the PKK structure. Belden describes joining the YPG by first linking up with the PKK at its headquarters in the Qandil Mountains in northern Iraq, before being spirited across the border into Syria. + +Belden gives a very interesting glimpse of the YPG’s method of governance. The YPG calls its rule “libertarian socialism,” says Belden, but it’s “pretty much a Stalinist state”. Belden describes the ascetic nature of the true believers in the PKK’s ideology—of which he, clearly, is not one—and the collectivized nature of life. Among other things, everyone is subjected to struggle sessions of the kind associated with Mao or the Khmer Rouge. + +The foreign fighters that join the YPG were, initially, “psychopaths that wanted to come kill people,” says Belden, but the YPG expelled these people once it realized what they were and has now refined its recruitment model to bring in a flow of hard-Left Westerners. + +Belden says he believes the Americans will abandon the YPG in favour of Turkey and that he believes it will be Turkey and her allies that go to Raqqa to evict the Islamic State. Belden favours this outcome, believing it would be a “bloodbath” if the YPG tries to take Raqqa City. Belden concludes though that there will soon be a peace agreement between the pro-regime coalition, Turkey, and the YPG—and that after that the Turks will attack the YPG in Efrin and elsewhere. Belden says that the YPG retains the ability to strike back +================================================================================ +Rank = 67; Score = 692224.0 +<|begin_of_text|>This post will mark the beginning of a series on Steve Keen’s book Debunking Economics. My first impression after reading his book was that it was chalk full of great ideas and solid criticisms of “neoclassical” economics as it is currently taught to undergraduate and graduate students. I’m not always fond of Keen’s tone, but isn’t some crank. I’m writing these posts in order to better retain the ideas and concepts expressed in the book, as well as to spark discussion. Other bloggers have also written about and summarized this book, and I will be linking to these posts as well. I’ll say that I don’t really have anything new to add to what has been perviously said, and I am doing this for my benefit. + +Keen’s begins his criticism of neoclassical (micro) economics in chapter 3, which centers around the “Law of Demand”, i.e. the idea that market demand curves don’t necessarily slope downwards. This is in direct contrast to what is taught in an “Intro to Microeconomics” course, and subsequent economic courses use the assumption when explaining other concepts. + +Market Demand Curves Don’t Slope Downwards + +Economists look at demand curves to see how demand for a commodity changes as its price changes, while the consumer’s income remains constant. When you don’t make that assumption, things get complicated quickly. The decrease of price in one good raises the real income of the consumer (the income effect), allowing the consumer to increase the consumption of all goods, not just the good that has become cheaper. So in the case of Giffen Goods, because the rise in price of one good make another good unaffordable, you have a paradox where demand increases as prices rise. Consider necessary, luxury, and normal goods, and you have a very complicated picture and some interesting looking demand curves. + +This is why economists make the distinction between the income effect and the substitution effect (if price falls, consumption rises and vice versa). The substitution effect is always negative and it’s what economists use to establish the “Law of Demand”. However, the income effect can be both negative and positive. So economists neutralize the income effect by using a Hicksian compensated demand function. The consumer is given a level of utility and a set of prices. The consumer then minimizes the amount of money that he/she would need to spend to achieve that given level of utility, thus separating the income effect and the substitution effect. + +But once you introduce more than one agent into +================================================================================ +Rank = 68; Score = 692224.0 +<|begin_of_text|>WASHINGTON (Reuters) - Texan Sam Lovett had no health insurance in August 2010 when an emergency hospital stay brought the news from his doctors that his liver was failing and he could die within less than a year without a transplant. + +The small distribution center where he worked did not provide health insurance. Lovett, 43, who lives near Comfort, Texas, was not able to buy private coverage on his own because of his already bad health. Though he had the resources to cover routine medical bills, he now needed a $400,000 organ transplant and no doctor or clinic would take him without insurance. + +“They were making arrangements to send me to a hospice. One doctor flatly told me that I had a better chance of leaving the hospital dead than alive,” said Lovett, who had a history of alcoholism and obesity. + +But in the months that followed, Lovett was able to enroll in the Pre-Existing Condition Insurance Plan (PCIP). The program was created by the healthcare law that President Barack Obama, a Democrat, signed in March 2010 despite the united opposition of Republicans in Congress. The program paid for Lovett’s transplant surgery at the Mayo Clinic in Phoenix last December. + +PCIP is a temporary government program set up to run until 2014, when the Obama healthcare law is due to take full effect and prevent insurance companies from the common practice of denying people coverage due to pre-existing medical conditions. About 60,000 Americans already are receiving medical coverage through the program, according to the Obama administration. + +As the U.S. Supreme Court prepares to rule in the next few weeks on whether to overturn the law considered to be Obama’s signature domestic policy achievement, Lovett said the choice is very clear for him and others unable to get insurance without the statute’s provisions. + +“Without this, you’re going to die,” Lovett said in an interview arranged by the Consumers Union organization, which supports the law. “From my perspective, that’s what the Supreme Court will sanction, or not - whether people will be allowed to die because they can’t get insurance.” + +Unlike people in many other rich nations where health insurance is near universal, many Americans struggle without medical coverage even as the country pays more for healthcare than anywhere else in the world. The U.S. healthcare system is a patchwork of private insurance and restrictive government programs that has left tens of millions uninsured. + +Many Americans receive health insurance only because their employers provide it as a job benefit. But some companies, particularly smaller employers, do not offer it. +================================================================================ +Rank = 69; Score = 688128.0 +<|begin_of_text|>The secret to silky smooth JavaScript animation! + +Creating an animation in JavaScript is one of the simplest things you can do. If you’ve ever tried to do it, you’ll most likely have used either the setTimeout or setInterval functions. + +Here’s a typical example: + +function draw ( ) { // Drawing code goes here } setInterval ( draw, 100 ) ; + +This piece of code will call the draw function once every 100ms forever and ever, until at some point later the clearInterval function is called. An alternative to this code is to use a setTimeout function instead, this time inside the draw function: + +function draw ( ) { setTimeout ( draw, 100 ) ; // Drawing code goes here } draw ( ) ; + +The single call to the draw function kicks off the animation loop, and from then on it will call itself repeatedly every 100ms. + +Frame rate and setInterval + +The smoothness of your animation depends on the frame rate of your animation. Frame rate is measured in frames per second (fps). Film usually runs at 24fps, video at 30fps. The higher this number is, the smoother your animation will look…to a point. More frames, means more processing, which can often cause stuttering and skipping. This is what is meant by the term dropping frames. Because most screens have a refresh rate of 60Hz, the fastest frame rate you should aim for is 60fps. Time for some math! + +/** * 1s = 1000ms (remember that setInterval and setTimeout run on milliseconds) * 1000ms / 60(fps) = 16.7ms (we'll round this to 17) */ // Lights, camera…function! setInterval ( function ( ) { animateEverything ( ) ; }, 17 ) ; + +What’s wrong with setTimeout and setInterval? + +OK, you’ve done this a million times before. What’s wrong with this picture? Well, a few things. Firstly, setTimeout doesn’t take into account what else is happening in the browser. The page could be hidden behind a tab, hogging your CPU when it doesn’t need to, or the animation itself could have been scrolled off the page making the update call again unnecessary. Chrome does throttle setInterval and setTimeout to 1fps in hidden tabs, but this isn’t to be relied upon for all browsers. + +Secondly, setTimeout only updates the screen when it wants to, not when the computer is able to. That means your poor browser has to juggle redrawing the animation whilst redrawing the whole screen +================================================================================ +Rank = 70; Score = 684032.0 +<|begin_of_text|>Typical weather conditions in an Icelandic winter + +The climate of Iceland is subarctic (Köppen climate classification: Cfc)[1] near the southern coastal area and tundra inland in the highlands. The island lies in the path of the North Atlantic Current, which makes its climate more temperate than would be expected for its latitude just south of the Arctic Circle. This effect is aided by the Irminger Current, which also helps to moderate the island’s temperature.[2] The weather in Iceland can be notoriously variable.[3] + +The aurora borealis is often visible at night time during the winter. The midnight sun can be experienced in summer on the island of Grimsey off the north coast; the remainder of the country, since it lies just south of the polar circle, experiences a twilight period during which the sun sets briefly, but still has around 2 weeks of continuous daylight during the summer. + +Seasons [ edit ] + +Winters [ edit ] + +The Icelandic winter is relatively mild for its latitude, owing to maritime influence, and enhanced by its proximity to warm currents of the North Atlantic Gyre. The southerly lowlands of the island average around 0 °C (32 °F) in winter, while the Highlands of Iceland tend to average around −10 °C (14 °F). The lowest temperatures in the northern part of the island range from around −25 to −30 °C (−13 to −22 °F). The lowest temperature on record is −39.7 °C (−39.5 °F).[4] + +Summers [ edit ] + +The average July temperature in the southern part of the island is 10–13 °C (50–55 °F). Warm summer days can reach 20–25 °C (68–77 °F).[4] The highest temperature recorded was 30.5 °C (86.9 °F) at the Eastern fjords in 1939. Annual average sunshine hours in Reykjavík are around 1300, which is similar to towns in Scotland and Ireland.[5] + +Climate data for Reykjavík, Iceland (1961–1990) Month Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Year Average high °C (°F) 1.9 + +(35.4) 2.8 + +(37.0) 3.2 + +(37.8) 5.7 + +(42.3) 9.4 + +(48.9) +================================================================================ +Rank = 71; Score = 684032.0 +<|begin_of_text|>By Indiaspend As many as 45 of India’s top 50 districts that offer affordable, formal financial services are in four south Indian states—with 30 concentrated in Tamil Nadu and Kerala—according to CRISIL Inclusix, a measure of financial inclusion. The top five states/union territories are Puducherry, Kerala, Tamil Nadu, Goa and Chandigarh. These states drove a surge in India’s financial inclusion index, which improved to 50.1 (of a possible 100) at the end of 2013 from 42.8 at the end of 2012. The index does not take into account the Prime Minister’s Jan Dhan Yojana (People’s Wealth Scheme). With over 140 million new bank accounts opened since August 2014, the index will likely be boosted further in 2015. CRISIL, a Mumbai-based ratings agency, defines financial inclusion as “the extent of access by all sections of society to formal financial services such as credit, deposit, insurance and pension services”. CRISIL’s Inclusix is a relative index that combines three parameters of basic financial services—branch penetration, deposit penetration and credit penetration. A score of 100 indicates the ideal state for each of the three parameters. CRISIL has included district-wise data from micro-finance institutions (MFIs) to calculate the index for the first time, covering 652 districts in 35 states and union territories. Bank Services Are Spreading As many as 117 million savings bank accounts were opened in the financial year 2013, almost 50% more than the 79 million that opened the previous year. The total operative bank accounts in the country at the end of 2013 were 820 million as against 703 million in 2012, an increase of 17%, the fastest growth in four years. However, growth in credit accounts declined due to a reduction in small-borrower accounts to 102 million in 2013 from 109 million in 2012. “The decline was primarily observed in five metro districts: Delhi, Kolkata, Mumbai, Mumbai Suburban and Bangalore Urban. This could be attributed to book-cleaning exercises by banks and closure of dormant credit-card accounts,” the report said. South leads; North-East at bottom but moves ahead With an increase of 40% in bank credit accounts, the northern state of Jammu & Kashmir saw its Inclusix score improve to 45.2. The state also saw high deposit and branch scores +================================================================================ +Rank = 72; Score = 679936.0 +<|begin_of_text|>The time is ripe for a rediscovery of Leopold Kohr. Or perhaps better: the time is ripe for the discovery of Leopold Kohr, since few have any idea who he was. A select group of readers might connect him with E.F. Schumaker, author of Small Is Beautiful: Economics as if People Mattered (orig. 1973). Kohr was one of Schumaker’s instructors, and the two remained lifelong associates and friends. + +I. + +Roughly three years after the Meltdown of 2008 I encountered an article online with the curious title, “This Economic Crisis is a “Crisis of Bigness.” It attracted my attention, because like many other observers I’d become convinced long before that both government and corporations had grown too large to accomplish even well-intentioned goals. The impulses for a more centralized society and a more economically integrated world — the impulses, that is, that bring about “bigness” — were mistakes from the get-go. Globalization promised universal prosperity, but instead has made the rich richer, slowly and painfully begun wiping out the American middle class, and created legions of unemployed or underemployed poor people — precariat is the term some writers are now using for this new class. The precariat consists of the army of adjunct faculty, temps, short-term contract workers, etc., most doing what they are doing because it is the only work they can find in the globalized “new economy.” The underemployment crisis heralds eventual civil unrest, especially with the Bureau of Labor Statistics insisting, against all experience, that the unemployment rate is just 5.5% (the U-3 figure as of this writing). Does anyone really think this is good, whether for workers or the economy at large? + +The above article cited Leopold Kohr, on the grounds that he had predicted central aspects of the present situation over half a century ago in his magnum opus The Breakdown of Nations (1957). I immediately ordered Kohr’s book, and when it came, I put every spare minute into reading it straight through. It is not light reading. Works that raise fundamental issues in political philosophy as well as economics never are. I wondered why I’d never heard of Kohr or Breakdown before. Then I realized: unlike someone such as Joseph Schumpeter, whose analysis of the long-term fate of capitalism as an industrial system in Capitalism, Socialism and Democracy (1947) was the most interesting I’d +================================================================================ +Rank = 73; Score = 679936.0 +<|begin_of_text|>“Home life ceases to be free and beautiful as soon as it is founded on borrowing and debt” Henrik Ibsen + +According to Oxford Dictionary the term Slave is defined as “a person who is the legal property of another and is forced to obey them” as in the case of the United States during the 18th and 19th centuries where slavery was a legalized institution. Oxford dictionary also defines slavery as “a person who works very hard without proper remuneration or appreciation” as in today’s world of a person working for a company or corporation where their efforts are usually under appreciated. It also describes a slave as “a person who is excessively dependent upon or controlled by something” or “a device, or part of one, directly controlled by another”. Debt can be an instrument used to control an individual or a nation for that matter. In this case, an individual is dependent upon “Credit” to buy products. + +Then the credit becomes a debt that has to be repaid. It becomes a “control mechanism” as the creditor becomes the “Slave Owner” and the debtor becomes the “Slave”. What is the point? In today’s world of unlimited credit, consumers become modern-day slaves to their creditors. What is the difference between slavery in 18th century America with imported African slaves and the America of 2013? There is no difference besides the physical abuse of the African slaves by their owners. In America, consumers suffer psychological abuse by its creditors. As long as an individual remains in debt bondage, that person will have to repay that debt until the day that person literally dies in most cases. + +Black Friday is the day that starts the most important holiday for big name retailers and Wall Street speculators and that is Christmas. It is the shopping season that investors, economists and corporations pay close attention to as they measure consumer confidence and the profits they reap from consumer spending. Major retailers and corporations such as Wal-Mart expect to make profits. Wall Street expects consumers to spend on Black Friday through the Christmas holidays following the Federal Reserve’s continued policies of Quantitative Easing (QE). Economists across the spectrum predict that the new Federal Reserve chairwoman Janet Yellen will continue to buy US bonds indefinitely continuing Ben Bernanke’s current policies. + +All the while consumers continue to accumulate debt. Black Friday was marked with chaos followed by violence as mobs of consumers’ raided shopping centers and malls for discounts and sales on numerous products including flat screen televisions, toys, clothing and other goods they most likely don’t need. Regardless of the economic +================================================================================ +Rank = 74; Score = 675840.0 +<|begin_of_text|>Show full PR text + +Consumer Reports Auto Reliability Survey: Ford Continues Fall While Seven Japanese Brands Top List + +Audi and Cadillac make major jumps in rankings + +December 2012 CoverYONKERS, NY-A perfect storm of reliability problems has dropped Ford to next to last among the 28 car brands ranked in Consumer Reports 2012 Annual Auto Reliability Survey, while its luxury brand, Lincoln, placed just a notch higher. The findings were released today before the Automotive Press Association in Detroit. + +Only two years ago, Ford was Detroit's poster child for reliability. It cracked the top 10 among brands in Consumer Reports predicted-reliability scores, with more than 90 percent of its models being average or better. This year the top seven spots are all held by Japanese brands. + +"Ford's bumpy road can be seen in the numbers. Sixty percent of Ford-branded models and half of Lincolns were below average in predicted reliability, and none placed above average," said Jake Fisher, director of automotive testing for Consumer Reports. + +Several factors contributed to Ford's decline in Consumer Reports' reliability rankings. A few new or redesigned models, including the Explorer, Fiesta, and Focus, came out of the gate with more problems than normal. Ford has also added the MyFord/MyLincoln Touch electronic infotainment system, which has been problematic so far, to many vehicles. In addition, three historically reliable models-the Ford Escape, Fusion and the Lincoln MKZ-are not included in the analysis; the three were redesigned for 2013 and CR doesn't yet have reliability data on them. + +Toyota, on the other hand, has excelled in Consumer Reports' latest ratings. Its three brands-Scion, Toyota, and Lexus-swept the top spots. Toyota is clearly setting the pace in reliability. Of the 27 models in the brand's lineup, 16 earned the highest rating. The subcompact Toyota Prius C earned Consumer Reports' top score overall. The hatchback Prius, the larger Prius V, and the new Prius plug-in were also above average. + +The Toyota trio was followed by four other Japanese makes: Mazda, Subaru, Honda, and Acura, in that order. All of the models produced by the top seven brands had average or better reliability. And of the 90 Japanese models reflected in Consumer Reports' brand comparison, 86 were average or better, with 35 earning the highest rating. + +Leading the Europeans, Audi had its best showing ever, moving up +================================================================================ +Rank = 75; Score = 675840.0 +<|begin_of_text|>"California is once again the sixth-largest economy in the world. If you add the GDP’s of Washington and Oregon, California would surpass the United Kingdom to become the fifth-largest economy in the world." + +Antonio Villaraigosa, the former mayor of Los Angeles and 2018 Democratic candidate for California governor, recently called for the state to unite with like-minded cities and states on the West Coast to oppose "dangerous policies advanced by a Trump administration." + +California’s large economy, Villaraigosa said in a Dec. 8, 2016 op-ed in the Sacramento Bee, gives it leverage in any possible showdown. + +In making this call, Villaraigosa repeated a favorite claim by California politicians that the state’s economy ranks as the sixth largest in the world -- a claim PolitiFact California rated Mostly True in July. He then took the economic comparison further -- all the way to the Pacific Northwest. + +"California is once again the sixth-largest economy in the world," Villaraigosa said. "If you add the GDP’s of Washington and Oregon, California would surpass the United Kingdom to become the fifth-largest economy in the world." + +"That’s power – power we must use to protect our people against any dangerous policies advanced by a Trump administration," he added. + +We decided to fact-check the part of Villaraigosa’s claim that if the GDPs of California, Washington and Oregon were somehow combined, they’d represent the fifth largest economy on the planet, ahead of the United Kingdom. + +Gross domestic product, or GDP, is used to measure the health of a country’s or state’s economy. It’s the total value of all goods and services. + +Our research + +A campaign spokeswoman for Villaraigosa pointed us to 2015 GDP figures from the International Monetary Fund. + +We used the same data in July to verify that California, with a GDP of nearly $2.5 trillion, had the sixth-largest economy behind the United States, China, Japan, Germany and the United Kingdom. + +California Gov. Jerry Brown’s administration released GDP figures in June showing the state had jumped two spots in these unique world rankings ahead of France and Brazil and into sixth place behind the United Kingdom. + +In our July fact check, we noted that when adjusted for California’s very high cost of living, the state’s GDP drops several places. This led us to our Mostly True rating for the claim, which we define as "accurate but needs clarification or additional information." + +To come up with a combined GDP for three West Coast states, we +================================================================================ +Rank = 76; Score = 671744.0 +<|begin_of_text|>The International Energy Association (IEA) announced a plan today to make 60 million barrels of oil available to the market over the next month from the Strategic Petroleum Reserve (SPR), in response to the disturbance in supplies from Libya. Half of this oil is to come from the SPR of the United States; the rest is to come from SPRs of other members of OECD. + +Both the amount and the timing of the release are strange. The amount of the release is equivalent to 2 million barrels a day. This is actually more than the Libyan disruption took off the market, which was about 1.4 million barrels a day. The disruption first took place in February, and oil prices have been declining since early May, so the timing is strange, as well. + +What the timing of the oil release is close to, is the end of Quantitative Easing 2 (QE2). QE2 is the United States’ program of buying back debt to keep interest rates low and the dollar low, which began November 2010 and is scheduled to end June 30, 2011. It was intended to stimulate the economy, and oil prices have indeed risen during most of the time it was in effect. + +Today’s announcement of the SPR oil-release program was made by IEA, but a person can’t help but wonder if the United States wasn’t heavily involved in the decision. After all, the United States is the largest member of the IEA, and is making half the release itself. Also, when Ben Bernanke spoke yesterday, he didn’t have any financial replacement for QE2. The release of oil from the SPR looks as if it is the latest attempt to kick-start the economy, both of the US and other OECD countries, using yet another approach. + +If we look at oil price and the S&P 500 Index (Figure 2), there has been a significant correlation between the two since late 2008. When oil prices drop, so does the S&P 500 Index, most likely since this means the economy is “tanking.” When oil prices rise, so does the S&P 500 Index, indicating the economy is working well. + +The problem is that high oil prices soon sow the seed of their own destruction. Once the oil price starts getting high (over $85 is high, over $110 is very high), the high oil prices start causing recessionary influences, because citizens have to cut back on other goods, when oil and food prices start rising. Oil and food prices usually rise and fall +================================================================================ +Rank = 77; Score = 671744.0 +<|begin_of_text|>Originally, the term “movies” did not mean films, but the people who made them. It was generally used with disdain by early Hollywood locals who disliked the “invading” Easterners. [1] + +The first film ever made in Hollywood was D.W. Griffith’s 1910 In Old California, a biograph melodrama about a Spanish maiden (Marion Leonard) who has an illegitimate son with a man who later becomes governor of California. It was shot in two days. [6] + +When Horace and Daeida Wilcox founded Hollywood in 1887, they hoped it would become a religious community. Prohibitionists, they banned liquor from the town and offered free land to anyone willing to build a church. [25] + +The most filmed author is William Shakespeare, including straight film versions, modern adaptations, ( West Side Story [1961], The Lion King [1994], etc.) and Shakespeare parodies. [1] + +The shortest dialogue script since the introduction of talkies was written for Mel Brook’s Silent Movie (1976), which has only one spoken word throughout: “Non.” [1] + +The first motion picture to depict a non-pornographic sex act was Extase (1933) starring Hedwig Kiesler, known later as Hedy Lamarr (1913-2000). Her character flees from an impotent husband, runs naked through the woods, bathes, and then has sex with a young engineer in a hut.l [17] + +Though photos have survived, the movie A Daughter of the Gods is now considered lost + +The first nude scene in a major motion picture was of swimmer and actress Annette Kellerman (1887-1975) in A Daughter of the Gods (1916). [17] + +The earliest known American pornographic film is the 1915 A Free Ride, a.k.a. A Grass Sandwich. The film was directed by “A. Wise Guy” and was written by “Will She.” [17] + +The Western Hero most portrayed on screen has been William Frederick Cody, a.k.a. Buffalo Bill, followed by William Bonny, a.k.a. Billy the Kid. [17] + +The first African-American to play a leading role in a feature film was Sam Lucas (1850-1916) who was cast in the title role of Uncle Tom’s Cabin (1914). The first African-American actor to make a career in films was Noble Johnson (1881-1978). [17] + +The American Humane +================================================================================ +Rank = 78; Score = 659456.0 +<|begin_of_text|>Wow, did you see the new jobs numbers? Unemployment is down to 5.1%. In the current definition of full employment (during the 1960’s it was a 4% unemployment rate) we are getting near that point. In fact, the unemployment rate is now lower than it was anytime during the Reagan presidency. We have had 67 consecutive months of job growth. The unemployment rate has dropped almost 5% from the high point of the Obama administration in October 2009. With all this good news why are things so bad? + +You may have heard something about a controversy regarding the unemployment numbers. There is focus on such matters as underemployment which for example having someone with a college degree waiting tables or bartending for lack of any quality opportunities for a person with their educational qualifications. Or you may have heard about the many people working part-time jobs due to lack of full-time opportunities or employers attempting to circumvent the rules established by Obamacare which states that a full-time employee reaches that status at 30 hours per week. + +What we really need to spotlight is the crushing economic effect of a lower labor participation rate (LPR). There are now 100 million Americans over the age of 16 that are not working. The Obama Administration keeps running out the deceit that this is because of all the baby boomers retiring. The fact is that the labor force participation rate for the age group 16 to 24 is only 55.1%. That is a reduction of over 10% from 66% during the 1990’s. It is also down over 5% (60.8%) from 2005. Sure myopic minimum wage increases are harming the employment of this age group with the least work experience, but that is not the total explanation. + +The Obama manipulation gets worse because the LPR is lower for the prime working years of 25-54 years old. In 2000 the LPR for this age group was almost 85%. It was down to 83% when the recession started, but has now plummeted to 80.7%. It is clear the baby boomers are not the only source of reduction in the LPR. + +You may wonder why this is such a big deal. The LPR for September 2015 was 62.4%. That is 3.7% less than August, 2005 exactly ten years earlier. One can argue this reduction in rate has to do with the “Great Recession.” Not true. If you review the +================================================================================ +Rank = 79; Score = 659456.0 +<|begin_of_text|>The male is often troubled by concerns that his penis is not large enough to satisfy his partner or himself. He is ashamed to have others view his penis, especially in the flaccid state. Such concerns might be unfounded in reality and might be a presentation of social anxiety or some other clinical problem, such as erectile dysfunction. Concern over the size of the penis, when such concern becomes excessive, might present as the ‘small penis syndrome’, an obsessive rumination with compulsive checking rituals, body dysmorphic disorder, or as part of a psychosis. However, it is often a worry that can be described as within the normal experience of many men. Various potential causal factors are considered. A thorough assessment, normalizing the worry and then exploring the treatment options in detail with the man, is essential to allow the matter to be consolidated satisfactorily within the male ego. + +Abbreviations SPS small penis syndrome BDD body dysmorphic disorder (dysmorphophobia) ED erectile dysfunction CBT cognitive behavioural therapy SSRI selective serotonin reuptake inhibitor. + +INTRODUCTION The penis, particularly in its erect state, is a symbol of masculinity. In many cultures it has come to symbolise attributes such as ‘largeness, strength, endurance, ability, courage, intelligence, knowledge, dominance over men, possession of women; a symbol of loving and being loved’. A review [1] is recommended which describes Indian Sadhus using weights, Dayak men in Borneo piercing the glans and then inserting items in the resultant holes to stimulate the partner, and the Topinama of Brazil, who encourage poisonous snakes to bite their penis to make it enlarge (for 6 months!). Some of these concepts date back over many thousands of years, and there is evidence that prehistoric cave dwellers attributed the symbolic values of strength and power to penile size, as well as those of virility and fertility, a process also recommended in the Kama Sutra [2]. Given the historical context it is perhaps no surprise that even today many men place great importance on the size of their penis. Hegemonic masculinity is defined by attributes such as physical strength, heterosexuality with authority over women and other men, showing no emotions (such as remorse and uncertainty, which might suggest vulnerability), economic independence, and an ability to demonstrate sexual ‘conquest’. While most men do not embody all of these qualities, society supports hegemonic masculinity within most of its institutions [3]. Given this historical and cultural background, it is perhaps unsurprising that for +================================================================================ +Rank = 80; Score = 655360.0 +<|begin_of_text|>"We cut our deficits by more than half." + +In 2012, PolitiFact twice rated True claims that President Barack Obama failed to keep a promise to cut federal deficits in half by the end of his first term. + +At that point, the budget gap topped a trillion bucks and Republican congressmen called out Obama because the deficit had been reduced by just 15 percent. + +But there was the president at Laborfest 2014 in Milwaukee proclaiming "we cut our deficits by more than half." + +In his Sept. 1, 2014 speech, the president ticked off for union supporters a list of major policy decisions he contends helped boost the economy and improve the government’s bottom line. + +We can’t cover all of those here, but the deficit claim grabbed our attention. + +Has it really been cut in half? + +The White House Office of Management and Budget pointed us to a chart prepared by that office in 2013 as proof of Obama's claim. + +It compares the yearly deficits under Obama, expressed -- as they often are -- as a share of the nation’s entire economy, which is measured by the Gross Domestic Product. + +At the start of Obama’s term, the chart showed, the figure was 9.2 percent. The latest figure was 4.1 percent. + +That appears to back Obama's statement. + +But let's examine this in detail. + +To do that, we reviewed figures published by the Congressional Budget Office as well as the White House’s Office of Management and Budget. We also consulted with independent fiscal experts. + +The baseline year for comparison is fiscal year 2009, which ended Sept. 30 of that year. This was the last budget from President George W. Bush, as Obama took office in January of that year. + +The most recent complete fiscal year is 2013. + +Those are the same years Obama's chart showed. + +Our analysis showed the drop easily topped 50 percent, and was actually somewhat higher than Obama's chart would indicate. + +As a share of the economy, we found -- and our experts confirmed -- the drop was from 9.8 percent in 2009 to 4.1 percent in 2013. + +Obama’s chart actually reflects a lower deficit figure for 2009, and therefore a lower share of GDP, 9.2%. + +That’s because instead of using the actual 2009 deficit of $1.4 trillion, Obama lowers it by the $200 billion in increased deficit spending that he -- not Bush -- pushed through in the stimulus plan to address the crisis that became the +================================================================================ +Rank = 81; Score = 655360.0 +<|begin_of_text|>Despite several quarters of rising GDP, and the upbeat exertions of Administration spokespeople, the National Bureau of Economic Research (NBER) has yet to announce the recession is over. Their reluctance is well-founded. It is beginning to dawn on even the more optimistic analysts that the tepid growth we have seen over the past three quarters is only an interlude in an otherwise grave and prolonged recession. Moreover, the respite will cost dearly as the United States has racked up a generation worth of debt for dubious benefit. + +The paltry number of new jobs currently being created still fall far short of the 375,000 per month needed to offset the 125,000 new entrants to the job market due to population growth and to erode the 8 million people laid off in the past year alone. Meanwhile, house prices continue to fall and credit continues to contract. With retail sales dropping in June and the Leading Economic Index (LEI) standing at minus 7.7 per cent, it should be clear that the US economy is heading back towards recession, following a temporary distortion created by some $1.3 trillion in federal stimulus. In short, the stimulus has failed. + +While there can be no doubt that an increase in government spending will result in a boost to GDP figures, the evidence of history shows that such growth is short-lived. Unfortunately as leaders around the world look to tighten the reins on out of control spending, President Obama and his Democratic supporters in Congress believe that their stimulus actions have succeeded and should be redoubled. Armed with nothing more than faith in government and a belief that spending is both a means and an end, it appears that the US stimulus policy will continue. The net result of these efforts will not be a more vibrant economy, but the perpetuation of fear and confusion in the business community and the continuing expansion of deficits that will lead inevitably to higher taxes. + +The more indebted an economy becomes, the greater the burden that must be borne by the wealth-creating private sector. Indeed, at the present rate of government debt-financing, the private sector will have to contribute some $2 trillion each year in interest costs alone. This money must be raised by taxation or inflation. + +This week, in response to their fears of increased regulations, higher taxes, and greater government stewardship of the economy, discontent among business leaders flared into the open. Gathering in Washington, leaders of the US Chamber of Commerce lashed out at current regulatory changes in healthcare. In other forums, business executives and investors questioned the efficacy of the +================================================================================ +Rank = 82; Score = 643072.0 +<|begin_of_text|>In a report issued today on the data broker industry, the Federal Trade Commission finds that data brokers operate with a fundamental lack of transparency. The Commission recommends that Congress consider enacting legislation to make data broker practices more visible to consumers and to give consumers greater control over the immense amounts of personal information about them collected and shared by data brokers. + +The report, “Data Brokers: A Call for Transparency and Accountability” is the result of a study of nine data brokers, representing a cross-section of the industry, undertaken by the FTC to shed light on the data broker industry. Data brokers obtain and share vast amounts of consumer information, typically behind the scenes, without consumer knowledge. Data brokers sell this information for marketing campaigns and fraud prevention, among other purposes. Although consumers benefit from data broker practices which, for example, help enable consumers to find and enjoy the products and services they prefer, data broker practices also raise privacy concerns. + +“The extent of consumer profiling today means that data brokers often know as much – or even more – about us than our family and friends, including our online and in-store purchases, our political and religious affiliations, our income and socioeconomic status, and more,” said FTC Chairwoman Edith Ramirez. “It’s time to bring transparency and accountability to bear on this industry on behalf of consumers, many of whom are unaware that data brokers even exist.” + +The report finds that data brokers collect and store billions of data elements covering nearly every U.S. consumer. Just one of the data brokers studied holds information on more than 1.4 billion consumer transactions and 700 billion data elements and another adds more than 3 billion new data points to its database each month. + +Among the report’s findings: + +Data brokers collect consumer data from extensive online and offline sources, largely without consumers’ knowledge, ranging from consumer purchase data, social media activity, warranty registrations, magazine subscriptions, religious and political affiliations, and other details of consumers’ everyday lives. + +Consumer data often passes through multiple layers of data brokers sharing data with each other. In fact, seven of the nine data brokers in the Commission study had shared information with another data broker in the study. + +Data brokers combine online and offline data to market to consumers online. + +Data brokers combine and analyze data about consumers to make inferences about them, including potentially sensitive inferences such as those related to ethnicity, income, religion, political leanings, age, and health conditions. Potentially sensitive categories from the study are “Urban Scramble” and “Mobile Mixers,” both of which include a high +================================================================================ +Rank = 83; Score = 643072.0 +<|begin_of_text|>We have a once-in-a-generation opportunity to do something big. President Trump has made tax reform a priority, and we have a Republican Congress that wants to get it done. This is something that Democrats should support too because it’s good for the American people. + +The President is going to seize this opportunity by leading the most significant tax reform legislation since 1986 – and one of the biggest tax cuts in American history. + +The President has focused on three things since his campaign: job creation, economic growth, and helping low and middle-income families who have been left behind by this economy. He understands that there are a lot of people in this country that feel like they work hard and still can’t get ahead. They are sick of turning their paychecks over to Washington and having no idea how their tax dollars are spent. They are frustrated by a tax code that is so complicated that they can’t even do their own taxes. + +That’s why tax reform is such a big priority for this President. He cares about making the economy work better for the American people. + +We are going to cut taxes for businesses to make them competitive, and we are going to cut taxes for the American people – especially low and middle-income families. + +In 1935, we had a one-page tax form consisting of 34 lines and two pages of instructions. Today, the basic 1040 form has 79 lines and 211 pages of instructions. Instead of a single tax form, the IRS now has 199 tax forms on the individual side of the tax code alone. Taxpayers spend nearly 7 billion hours complying with the tax code each year, and nearly 90% of taxpayers need help filing their taxes. + +We are going to cut taxes and simplify the tax code by taking the current 7 tax brackets we have today and reducing them to only three brackets: 10 percent, 25 percent, and 35 percent. + +We are going to double the standard deduction so that a married couple won’t pay any taxes on the first $24,000 of income they earn. So in essence, we are creating a 0 percent tax rate for the first $24,000 that a couple earns. + +The larger standard deduction also leads to simplification because far fewer taxpayers will need to itemize, which means their tax form can go back to that one simple page. + +Families in this country will also benefit from tax relief to help them with child and dependent care expenses. + +We are going to repeal the Alternative Minimum Tax (AMT). The AMT creates significant +================================================================================ +Rank = 84; Score = 638976.0 +<|begin_of_text|>"You must be the only person worried about inflation, Liam". So I was told on Newsnight last week – by someone who's spent four of the last five years in the cabinet. + +Almost lost for words, I blurted an answer before Mr Paxman intervened. + +But I left the television studio shocked that the former minister, someone of genuine ability, could be so ignorant of life beyond the Westminster village. + +"Inflation, not deflation" is a recurrent theme of this column. That's because the Western world's response to the "credit crunch", this orgy of wildly expansive policy using "deflation is nigh" as an alibi, amounts to a repeated whacking of the self-destruct button. + +Almost everyone outside the "political and media elite" knows this. Amidst "deflation is upon us" warnings from this elite, UK CPI inflation went up last month – to 3.2pc. CPI understates true inflation. But even CPI is above the Bank of England's target – and has been for 17 months. + +So if I'm the only one worried, why is Mervyn King writing repeated letters acknowledging inflation is too high? + +Sterling is down a third in 18 months. Import prices are soaring. The UK's monetary base is doubling. + +"Broad" M4 money is expanding at 16pc a year. Food price inflation is 8pc, with double-digit producer price inflation too – as the credit crunch destroys supply chains. + +Warren Buffett, the most successful living investor, foresees an "inflation onslaught". Peter Brabeck, the Nestlé chairman and one of the world's top businessmen, says "Now the printing machine is starting, this is the start of inflation." + +China and other major creditors are rightly concerned Western governments – particularly the US and UK – want to inflate away their sovereign debts. + +That's why the "bond vigilantes" are stirring and we could very soon face a gilts strike. + +If I'm the only one worried about inflation, why is the price of gold so high, with the price of commodities also up sharply – even though the global demand outlook is weak? + +If I'm the only one worried about inflation, why does the latest report of the Bank of England's pension scheme show it recently changed it allocation of index-linked instruments – the ultimate anti-inflation hedge – from 25pc to more than 70pc of its portfolio? + +Am I the only one worried? I so wish that was true. We face an inflation tsunami. And +================================================================================ +Rank = 85; Score = 634880.0 +<|begin_of_text|>Andy Blatchford, The Canadian Press + +OTTAWA -- Fuelled by climbing prices for fresh fruits and vegetables, the pace of Canada's annual inflation rate accelerated last month to 1.6 per cent, Statistics Canada said Friday. + +Inflation grew at its fastest pace in December since late 2014. Last month's number also followed a 1.4 per cent year-over-year increase in November, the agency's latest consumer price index found. + +The figure came out as the economy deals with the effects of the steep slide in commodity and oil prices, which have also helped drag down Canada's exchange rate. The lower loonie is expected to drive up costs for imported goods. + +The report said prices for fresh fruit increased 13.2 per cent last month compared to a year earlier, while fresh vegetables rose 13.3 per cent. + +The price of lettuce surged 21.8 per cent. + +Overall, consumers spent 3.7 per cent more on food last month than the previous year. + +On top of higher produce prices, Canadians were also paying considerably more for home and mortgage insurance, automobiles and electricity compared to a year earlier, the report said. + +The agency said lower prices for gasoline, natural gas and fuel oil applied downward pressure on inflation. Gasoline prices were down 4.8 per cent compared to December 2014, while natural gas decreased 12.9 per cent and fuel oil dropped by 16.8 per cent. + +But the slipping price of energy slowed somewhat, which allowed overall inflation to creep up, said Dawn Desjardins, deputy chief economist for RBC. + +"We all knew that there was going to be this huge weight on that headline rate because of the energy and now we're seeing some relief from that," Desjardins said. + +"It's kind of following the script, if you will, of what forecasters were looking for." + +To help explain the inflation-cooling effect, some analysts also pointed to the sharp monthly decline in the price of clothing and footwear, which fell 5.2 per cent from November to December. + +But moving forward, National Bank senior economist Matthieu Arseneau predicts shoppers will continue to face higher prices for imported goods in many categories. + +"Despite weak energy prices, we don't expect Canadian consumers to get some respite because the dive in the currency should be a significant offset," Arseneau wrote in a note to clients. + +By region, Statistics Canada found that consumer prices increased in every province last month compared to the year before, with British +================================================================================ +Rank = 86; Score = 630784.0 +<|begin_of_text|>Consumers are paying sharply higher prices for beef, and other meat, this year, according to the government's consumer price index for March, released April 15, 2014. (Photo11: Jim Cole AP) + +Two months of sharp increases in food prices show grocers are starting to pass along their higher wholesale costs to consumers. + +Retail food prices rose 0.4% in March, the same as in February and the largest amount since September 2011. By comparison, the prices of all consumer goods rose 0.2% in March and 0.1% the month before, reports the Bureau of Labor Statistics. + +Beverly Cabellon, 61, of Pleasant Hill, Calif., was taken aback by the $38 price for two steaks at Costco recently, up from the $27 she paid last September. "I will be grilling more vegetables and shrimp this summer," she says, adding that she and her husband will likely eat beef once a month instead of weekly. "And I may switch to pork and chicken." + +Beef, pork, poultry, eggs and milk have had the most dramatic price increases as drought, a virus outbreak and rising exports have thinned U.S. supplies. + +Overall consumer prices rose 0.2% in March, a bit more rapidly than in recent months, and annual inflation was 1.5%, up from 1.1% in February. + +Annual inflation was 1.5% in March, up from 1.1% in February. That's well below the Federal Reserve's 2% target, as falling gasoline prices offset rising food costs. + +But higher food bills are squeezing households still struggling with meager wage gains, and could crimp spending just as the recovery is expected to accelerate. + +Cheryl Stewart, 38, of Perry Hall, Md., says higher prices for meat and milk have prompted her to drive 10 to 15 miles to grocery stores in low-income areas that carry more obscure brands at lower prices. She also spreads the food shopping for her family among three or four stores to get the best prices. + +"Living standards will suffer as a larger percentage of household budgets are spent on grocery store bills, leaving less for discretionary spending," says economist Chris Christopher of IHS Global Insight. + +A drought that thinned cattle herds two years ago has driven up wholesale beef prices 23% the past year, according to Sterling Marketing. Meanwhile, a virus outbreak in the hog population has pushed up pork prices by 56%, the +================================================================================ +Rank = 87; Score = 626688.0 +<|begin_of_text|>In the first place, studies that measure income inequality largely focus on pretax incomes while ignoring the transfer payments and spending from unemployment insurance, food stamps, Medicaid and other safety-net programs. Politicians who rest their demands for more redistribution on studies of income inequality but leave out the existing safety net are putting their thumb on the scale. + +Second and more important, it is well known that people's earnings in general rise over their working lifetime. And so, for example, a person who decides to invest more in education may experience a lengthy period of low income while studying, followed by significantly higher income later on. Snapshot measures of income inequality can be misleading. + +According to data from the Bureau of Labor Statistics' Consumer Expenditure Survey, if you sort households according to their pretax income, in 2010 the bottom fifth accounted for 8.7% of overall consumption, the middle fifth for 17.1%, and the top fifth for about 38.6%. Go back 10 years to 2000—before two recessions, the Bush tax cuts, and continuing expansions of globalization and computerization—and the numbers are similar. The bottom fifth accounted for 8.9% of consumption, the middle fifth for 17.3%, and the top fifth for 37.3%. + +While this stability is something to applaud, surely more important are the real gains in consumption by income groups over the past decade. From 2000 to 2010, consumption has climbed 14% for individuals in the bottom fifth of households, 6% for individuals in the middle fifth, and 14.3% for individuals in the top fifth when we account for changes in U.S. population and the size of households. This despite the dire economy at the end of the decade. + +The percentage of low-income households with microwave ovens grew to 92.4% from 74.9% between 2001 and 2009. Fully 75.5% of low-income Americans now have a cell phone, and over a quarter of those have access to the Internet through their phones. + +We would hazard a guess that if you were to ask a typical low-income American in 2009 if he would like to trade his house for its 2001 version, he would tell you to take a hike. How then is he worse off in 2009? + +So poor people are not in fact getting worse off because when they get worse off, we give them food stamps. That's exactly the sort of deep thinking I would +================================================================================ +Rank = 88; Score = 626688.0 +<|begin_of_text|>Media playback is unsupported on your device Media caption Bank of England chief Mark Carney: BoE taking steps to avoid'mortgage recklessness' + +Annual house price inflation hit 10.5% in May, the highest rate for four years, according to the Office for National Statistics (ONS). + +That is up from 9.9% in the previous month. + +Once again, the market was in large part driven by London, where house prices rose by a record 20.1% over the year. + +But excluding London and the South East, prices rose by a much more modest 6.4% across the UK. + +Prices were up in every region, except for Northern Ireland, where they fell by 0.7%. + +In England, prices rose by 11%, in Wales prices were up by 6.5%, and in Scotland they rose by 3.6%. + +According to the ONS index, the average price of a house across the UK is now £262,000. + +"Annual house price growth of 20.1% is borderline insane, but there are signs that the capital started to cool slightly in June and July," said Jonathan Samuels, the chief executive of Dragonfly Property Finance. + +Regional House Prices (England) Region Annual inflation North East 4.8% North West 3.9% Yorks and Humber 6.1% East Midlands 7.0% West Midlands 7.2% East 8.6% London 20.1% South East 9.6% South West 7.1% + +The housing charity Shelter said the only answer was more house-building. + +"Once again these figures show that house prices are spinning out of control, putting a stable home even further out of reach for ordinary families," said Campbell Robb, Shelter's chief executive. + +"Instead scores of people are either stuck in their childhood bedrooms or forced to bring up children in unstable and expensive rented homes, however hard they work or save," he said.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 89; Score = 610304.0 +<|begin_of_text|>Pin 23 24 Shares + +How to Install and Configure SSL Certificate + +Hyper Text Transfer Protocol Secure (https) website is encrypted site. Encrypted site means that website is secured. HTTPS site is secured by using Secure Sockets Layer (SSL). SSL certificates are issued by the Certificate Authority (CA). We can check that particular website is HTTPS or not by checking a ‘green colour lock‘ and https: in the address bar before the website name. For example, all banking websites are working on https. Some of the examples of secure websites are “https://www.google.com“, “https://www.facebook.com” and many more. In this article, we’ll learn the steps to install and configure SSL certificate server and CA. + +Install and configure SSL certificate would encrypt (secure) our website so that no one can easily decode the information or data transferred by web servers to the clients. Sharing confidential information over internet is not safe that is why https is used to encrypt our data for safe transfer. + +Steps to install and configure SSL Certificate on Windows Server 2012 R2. + +1. To install and configure SSL certificate server, we need to install the “Active Directory Certificate Services” role. Open “Server Manager” and click on “Add roles and features“. + +2. Please ensure that password is set for local Administrator and valid static IP address is assigned to the Server. Click next to continue. + +3. Select an option “Role-based or feature-based Installation” and click on next. + +4. Select the server from the pool on which we need to install and configure SSL Certificate. + +5. Select the role “Active Directory Certificate Services” and click on next to continue. + +6. Certain features are required to install and configure SSL Certificate Services role, click on “Add Features” to install all the dependent features. + +7. Click on next to continue. + +8. From features window, you can select additional features if required, however in this practical it is not required, therefore click on next to continue. + +9. The name and domain settings of the computer cannot be changed after a CA has been installed on the Server. Therefore make the changes before installing the role. + +10. From the “Role Services”, select “Certificate Authority” and “Certificate Authority Web Enrollment”. Certificate authority web enrollment allows users to request new, renew, revoke certificates, etc using Web console. + +11. Web Server (IIS) role is required for end users to request, renew, revoke certificates. + +12. Click on next to continue. + +13. Add roles and features wizard +================================================================================ +Rank = 90; Score = 606208.0 +<|begin_of_text|>From the Kansas City Fed: Tenth District Manufacturing Activity Strengthened Further + +The Federal Reserve Bank of Kansas City released the March Manufacturing Survey today. According to Chad Wilkerson, vice president and economist at the Federal Reserve Bank of Kansas City, the survey revealed that Tenth District manufacturing activity strengthened further with strong expectations for future activity. + +“Our composite index accelerated again, and has only been higher one time in the last 15 years,” said Wilkerson. “The future employment index was the strongest in the 23-year history of the survey." + +... + +The month-over-month composite index was 20 in March, its highest reading since March 2011, up from 14 in February and 9 in March. The composite index is an average of the production, new orders, employment, supplier delivery time, and raw materials inventory indexes. Activity in both durable and nondurable goods plants increased, particularly for metals, computer, electronic, and aircraft products. Most month-over-month indexes rose further in March. The production and shipments indexes increased considerably, while the new orders and order backlog indexes rose more moderately but remained high. The employment index moderated slightly from 17 to 13, and the new orders for exports index also eased. Both inventory indexes increased for the second straight month. + +emphasis added + +The Kansas City region was hit hard by the decline in oil prices, but activity is expanding solidly again. The regional Fed surveys released so far suggest another strong reading for the ISM manufacturing index for March.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 91; Score = 589824.0 +<|begin_of_text|>Most measuring instruments are limited by the tradeoff between how precisely and how rapidly a measurement is made: the more precise the measurement, the longer it takes. But because many phenomena occurring at the nanoscale are both rapid and tiny, they demand a measuring system that can capture their precise details in both time and space. + +Taking up that challenge, researchers at the National Institute of Standards and Technology (NIST) have redesigned the detection system at the heart of the atomic force microscope (AFM). A premier tool of the nanoworld, the AFM uses a small probe, or tip, to map the submicroscopic hills and valleys that define the surface of materials, along with other properties, at the nanometer scale. Although the AFM has already revolutionized the understanding of nanostructures, scientists are now eager to study nanoscale phenomena, such as the folding of proteins or the diffusion of heat, which happen too quickly and generate changes too small to be accurately measured by existing versions of the microscope. + +By fabricating an extremely lightweight AFM probe and combining it with a nanoscale device that converts minuscule deflections of the probe into large changes of an optical signal inside a waveguide, the NIST researchers have broken new ground: Their AFM system measures rapid changes in structure with high precision. + +Illustration of a newly fabricated atomic force microscope (AFM) probe integrated with an optical, disk-shaped resonator. Combined with a technique called photothermal induced resonance (PTIR), which uses infrared light to examine a material’s composition, the incorporation of the resonator enables the probe to make high-precision measurements of minuscule, rapid changes in a material. Credit: NIST + +The achievement takes the AFM into a new realm, enabling the instrument to measure time-varying nanoscale processes that may change as quickly as ten billionths of a second. “This is truly a transformational advance,” said NIST scientist Andrea Centrone. + +Centrone, Vladimir Aksyuk, and their colleagues employed the new AFM capabilities in experiments using photothermal induced resonance (PTIR), a technique that combines the acuity of an AFM with the ability to determine the composition of materials using infrared light. + +With the new AFM-PTIR system, the scientists measured with high precision the rapid, but minute expansion of individual microcrystals heated by a light pulse. The microcrystals examined by the team belong to a class of materials known as metal-organic frameworks (MOFs). These materials contain +================================================================================ +Rank = 92; Score = 589824.0 +<|begin_of_text|>Cells from Mr Turnbull's good eye were transplanted into his damaged one A man partially blinded in an attack on Tyneside has praised scientists who restored his vision using stem cells. Russell Turnbull, 38, lost the sight in one eye in 1994 when he was squirted with ammonia after intervening in an argument on a bus in Newcastle. He was left with Limbal Stem Cell Deficiency (LSCD), a painful condition which requires constant treatment. He said the Newcastle University team which developed the treatment had "transformed his life". The method involves taking a small amount of stem cells from a patient's good eye, cultivating them in a laboratory, and implanting them into the damaged cornea. Mr Turnbull, from Consett, County Durham, was treated at Newcastle's Royal Victoria Infirmary (RVI). He said: "I had a lot of anger inside me for a long time after the attack. I lost my job because of it and I had always been a keen jet skier, which I wasn't able to do. Please turn on JavaScript. Media requires JavaScript to play. "It ruined my life and I went through a really difficult time. But then this treatment came along. I can't thank the staff at the RVI enough. "This has transformed my life. My eye is almost as good as it was before the accident. I'm working, I can go jet skiing again and I also ride horses. I have my life back thanks to the operation." He is one of of eight patients who successfully underwent the treatment developed at the North East England Stem Cell Institute (NESCI). It is hoped that the technique could eventually be rolled out into clinics. + +Bookmark with: Delicious + +Digg + +reddit + +Facebook + +StumbleUpon What are these? E-mail this to a friend Printable version<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 93; Score = 589824.0 +<|begin_of_text|>Republicans in Congress have already voted to cut the taxes that pay for the Affordable Care Act (ACA). Up to now these tax cuts have been vetoed by President Obama. But with the inauguration of Donald Trump, Republicans in the Senate and House will be free to cut ACA taxes and give hundreds of billions of dollars to the richest Americans. + +The ACA taxes include a 3.8% tax on investment income, which mainly goes to the top 1%. Investment income is not taxed to pay for Medicare or Social Security. There is also a 0.9% tax on individuals with very large salaries, which is also not taxed to pay for Social Security. Social Security taxes only tax the first $127,200 in earned income; any income above that or investment income is not taxed. + +What this means is that eliminating the taxes for the ACA would provide the top one-tenth of one percent (0.1%) of Americans an average tax cut of almost $200,000 a year, which would be almost 75% of the total dollar tax cut. + +These ACA taxes on households raise about $100 billion a year, which along with ACA taxes on health insurance companies, drug and medical equipment manufacturers, and others, pay for ACA. One cost of the ACA are the premium tax credit where the federal government will reimburse low and middle-income households that purchase health insurance through a government exchange for all or part of their health insurance premiums. Another cost is the expansion of Medicaid for low-income Americans. + +The elimination of the ACA and the premium tax credits which subsidize the government insurance exchange would cause dramatic losses for about 4% of the population, who would lose insurance subsidies that average almost $5000 each year. While most the poorest 20% of American households, earning less than $25,000 per year, would not see much of a loss in subsidies, they are the ones who are most likely to lose insurance with the elimination of the ACA and its expansion of Medicaid. Eliminating the Medicaid expansion could cause almost 11 million to lose their insurance. + +It is no wonder that the wealthiest businesspeople and Wall Street tycoons who are represented by billionaire President-elect Donald Trump are so keen to eliminate the ACA. But their gains would come at the cost of higher health insurance premiums and loss of health insurance for millions of workers and small business owners.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 94; Score = 585728.0 +<|begin_of_text|>Not to be confused with the fictional system portayed in The Last Enemy (TV series) + +Total Information Awareness system from the official (decommissioned) Diagram of thesystem from the official (decommissioned) Information Awareness Office website + +Presentation slide produced by DARPA describing TIA + +Total Information Awareness (TIA) was a program of the United States Information Awareness Office that began during the 2003 fiscal year. It operated under this title from February until May 2003, before being renamed as the Terrorism Information Awareness.[1][2] + +Based on the concept of predictive policing, TIA aimed to gather detailed information about individuals in order to anticipate and prevent crimes before they are committed.[3] As part of efforts to win the War on Terror, the program searched for all sorts of personal information in the hunt for terrorists around the globe.[4] Admiral John Poindexter referred to it as a "Manhattan Project for Counter-Terrorism".[5] According to Senator Ron Wyden (D-Ore.), TIA was the "biggest surveillance program in the history of the United States".[6] + +The program was defunded alongside the Information Awareness Office in late 2003 by the United States Congress after media reports criticized the government for attempting to establish "Total Information Awareness" over all citizens.[7][8][9] + +Although the program was formally suspended, its data mining software was later adopted by other government agencies, with only superficial changes being made. The core architecture of TIA continued development under the code name "Basketball." According to a 2012 New York Times article, the legacy of Total Information Awareness is "quietly thriving" at the National Security Agency (NSA).[10] + +Program synopsis [ edit ] + +Total Information Awareness (TIA) was intended to be a five year long research project by the Defense Advanced Research Projects Agency (DARPA). The goal was to integrate components from previous and brand new government intelligence and surveillance programs, including Genoa, Genoa II, Genisys, SSNA, EELD, WAE, TIDES, Communicator, HumanID and Bio-Surveillance with data mining knowledge gleamed from the private sector to create a resource for the intelligence, counterintelligence, and law enforcement communities.[11][12] These components consisted of information analysis, collaboration, decision-support tools, language translation, data-searching, pattern recognition, and privacy-protection technologies.[13] + +TIA research included or planned to include the participation of nine government entities: INSC +================================================================================ +Rank = 95; Score = 581632.0 +<|begin_of_text|>(NaturalNews) Two out of every three store-bought chickens may be contaminated with bacteria that commonly cause human illness, according to a study conducted by the Consumers Union."Consumers still need to be very careful in handling chicken, which is routinely contaminated with disease-causing bacteria," said Urvashi Rangan, the union's director of technical policy.Researchers conducted tests on 382 fresh broiler chickens purchased at 100 retailers in 22 states in spring 2009. A full two-thirds of the chickens were contaminated with either one or both of the bacteria strains most often responsible for food-borne illness. Although this figure is an improvement over the 2007 figure of 80 percent, the Consumers Union still called the numbers "far too high" and called for stricter government regulation.Sixty-two percent of chickens tested positive for campylobacter, the number two cause of food-borne illness. Fourteen percent tested positive for salmonella, the number one cause, and 9 percent tested positive for both.In contrast, the U.S. Department of Agriculture found only 5 percent of chickens contaminated with salmonella in its tests of chicken packing plants in April, May and June 2009. The Consumers Union noted that its own tests were conducted farther down the supply chain, when more opportunities for contamination had arisen. It also pointed to prior studies finding widespread campylobacter contamination at chicken processing plants, and called upon the government to set a maximum safe threshold for levels of the bacteria.Although cooking destroys both salmonella and campylobacter contamination, people can be exposed to the pathogens while cooking or otherwise processing the raw meat. In addition, contamination can spread from uncooked chicken to other foods more commonly eaten raw. For this reason, health experts recommend that chicken always be bagged separately from other foods and refrigerated or frozen within two hours of purchase. The Centers for Disease Control and Prevention recommend keeping a separate cutting board exclusively for use on uncooked poultry, in order to prevent contamination of other foods.Sources for this story include: abcnews.go.com.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 96; Score = 577536.0 +<|begin_of_text|>Infosys is handling much of the IT infrastructure for Goods and Services Tax Network (GSTN). + +Highlights GST Council appointed the GoM to check tech issues related to GSTN The GoM is headed by Bihar Deputy Chief Minister Sushil Kumar Modi Infosys is handling much of the IT infrastructure for GSTN + +A group of ministers (GoM) - appointed by the Goods and Services Tax (GST) Council to look into the technical challenges that have been facing after the launch of this massive tax reform - met today in Bengaluru. The GoM is headed by Bihar Deputy Chief Minister Sushil Kumar Modi.Representatives from Infosys were also part of the meet. Infosys is handling much of the IT infrastructure for GST Network (GSTN).It is no secret that the implementation of that biggest tax reform GST has not been as smooth as desired. "The systems had not crashed...and that the pressure was due to the scale of operations and last minute rush by tax payers," Mr Modi said.He also said that, "22 crore invoices have been filed in two and a half months and 85 lakhs new and old dealers have been registered under GST."He further stated, "I want to appeal to tax payers that they should not wait for the last day. Because of that there is a heavy rush for the system. Only five days left for August filing and only 3.05 lakh have filed form 3B compared to more than 45 lakh in July...so there will be a rush."We will be able to solve the nitty-gritties found in the IT system, he added.Infosys, which is handling the GSTN, was represented at the meet. And, the government defended the IT major's role. Revenue Secretary Hasmukh Adhia said, "Infosys has not failed. They have been delivering very well. Hiccups will be there which need to be sorted out."The team insisted the glitches were minor and included non-availability of some forms online or difficulty for the user in changing some fields. The GoM will meet again soon - probably in Bengaluru again - as Infosys is headquartered in the city.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 97; Score = 577536.0 +<|begin_of_text|>These capitalists generally act harmoniously and in concert, to fleece the people. + +—Abraham Lincoln, from his first speech as an Illinois state legislator, 1837 + +Everyone now is more or less a Socialist. + +—Charles Dana, managing editor of the New York Tribune, and Lincoln’s assistant secretary of war, 1848 + +The workingmen of Europe feel sure that, as the American War of Independence initiated a new era of ascendancy for the middle class, so the American Antislavery War will do for the working classes. They consider it an earnest of the epoch to come that it fell to the lot of Abraham Lincoln, the single-minded son of the working class, to lead his country through the matchless struggle for the rescue of an enchained race and the reconstruction of a social world. + +—Karl Marx and the First International Workingmen’s Association to Lincoln, 1864 + +ON DECEMBER 3, 1861, a former one-term congressman, who had spent most of the past dozen years studying dissident economic theories, mounting challenges to the existing political order and proposing ever more radical responses to the American crisis, delivered his first State of the Union address as the sixteenth president of the United States. + +Since assuming office eight months earlier, this new president had struggled, without success, first to restore the severed bonds of the Union and then to avert a wrenching civil war. Now, eleven southern slave states were in open and violent rebellion against the government he led. + +His inaugural address of the previous spring had closed with a poignant reflection on the prospect of eventual peace, imagining a day when the Union might again be touched “by the better angels of our nature.” But, now, in the last month of what Walt Whitman would recall as America’s “sad, distracted year”—“Year that suddenly sang by the mouths of the round-lipp’d cannons”—the better angels seemed to have deserted the continent. Every effort to restore the republic had been thwarted. There was no room for accommodation with the Confederate States of America. Fort Sumter had been fired upon and the flag of southern rebellion now flew above Charleston Harbor. Virginia, the cradle of presidents, the state of Washington, Jefferson and Madison, had joined the revolt and assembled a capital of the Confederacy less than 100 miles from Washington. Hundreds of Union and Confederate soldiers had died, with thousands more wounded at the First Battle of Bull Run. Armies had been reorganized and generals replaced with the recognition that this was no skirm +================================================================================ +Rank = 98; Score = 569344.0 +<|begin_of_text|>The Formula One World Constructors' Championship (WCC) is awarded by the FIA to the most successful Formula One constructor over a season, as determined by a points system based on Grand Prix results. The Constructors' Championship was first awarded, as the International Cup for F1 Manufacturers, in 1958 to Vanwall. + +Constructors' Championship points are calculated by adding points scored in each race by any driver for that constructor. Up until 1979, most seasons saw only the highest-scoring driver in each race for each constructor contributing points towards the Championship. On only ten occasions has the World Constructors' Champion team not contained the World Drivers' Champion for that season. + +In the 61 seasons the Championship has been awarded, only 15 different constructors have won it, with Scuderia Ferrari the most successful, with 16 titles including 6 consecutive from 1999 to 2004. Only five countries have produced winning constructors: United Kingdom (33 championships with 10 different constructors), Italy (16 with Ferrari), Germany (5 with Mercedes), Austria (4 with Red Bull) and France (3 with two constructors). However, all German, Austrian and French titles have seen the winning cars designed and built (except Matra in 1969) and run by teams based in the United Kingdom. Among drivers that have contributed with at least a single point to the constructors' title, Michael Schumacher has the unofficial record, having been involved with seven such titles, six of those consecutively with Ferrari.[1] Schumacher won the world drivers' title on six of those seven occasions. + +By season [ edit ] + +The constructors' trophy + +Notes [ edit ] + +* – indicates that the driver also won the Drivers' Championship. + +By constructor [ edit ] + +Note: Constructors in bold have competed in the 2018 World Championship. + +By nationality [ edit ] + +By engine [ edit ] + +Engine manufacturers and constructors in bold have competed in the 2018 World Championship. + +^ * Built by Cosworth + +^ ** In 1998 built by Ilmor + +^ *** Built by Porsche + +By driver [ edit ] + +Notes: + +* – Posthumously. World Drivers' Champions, since 1958, who never contributed to a Constructors' Championship, were Ferrari's Mike Hawthorn (1958), McLaren's James Hunt (1976), and Williams's Keke Rosberg (1982). Drivers and constructors in bold have competed in the 2018 World Championship. + +See also [ edit ] + +References [ edit ]<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 99; Score = 565248.0 +<|begin_of_text|>Article was originally posted here. + +In this article I want to talk about the Entity-Component-System (ECS). You can find a lot of information about the matter in the internet so I am not going to deep into explanation here, but talking more about my own implementation. + +First things first. You will find the full source code of my ECS in my github repository. + +An Entity-Component-System – mostly encountered in video games – is a design pattern which allows you great flexibility in designing your overall software architecture[1]. Big companies like Unity, Epic or Crytek in-cooperate this pattern into their frameworks to provide a very rich tool for developers to build their software with. You can checkout these posts to follow a broad discussion about the matter[2,3,4,5]. + +If you have read the articles I mentioned above you will notice they all share the same goal: distributing different concerns and tasks between Entities, Components and Systems. These are the three big players in this pattern and are fairly loose coupled. Entities are mainly used to provide a unique identifier, make the environment aware of the existence of a single individual and function as a sort of root object that bundles a set of components. Components are nothing more than container objects that do not possess any complex logic. Ideally they are simple plain old data objects (POD’s). Each type of a component can be attached to an entity to provide some sort of a property. Let’s say for example a “Health-Component” can be attached to an entity to make it mortal by giving it health, which is not more than an integer or floating point value in memory. + +Up to this point most of the articles I came across agree about the purpose and use of entity and component objects, but for systems opinions differ. Some people suggest that systems are only aware of components. Furthermore some say for each type of component there should be a system, e.g. for “Collision-Components” there is a “Collision-System”, for “Health-Components” there is a “Health-System” etc. This approach is kind of rigid and does not consider the interplay of different components. A less restrictive approach is to let different systems deal with all components they should be concerned with. For instance a “Physics-Systems” should be aware of “Collision-Components” and “Rigidbody-Components”, as both probably contain necessary information regarding physics simulation. In my humble opinion systems are “closed environments”. That is, they do not take ownership of entities nor components. They do access diff --git a/examples/openwebtext/files/ml.txt b/examples/openwebtext/files/ml.txt new file mode 100644 index 0000000..d0e788e --- /dev/null +++ b/examples/openwebtext/files/ml.txt @@ -0,0 +1,2272 @@ +Query Sequence: +Prompt: Machine learning can be defined as; Completion: a set of algorithms that enables computers to learn from data. These algorithms are used to build models that can be used for various tasks, such as predicting the future, identifying patterns, and making decisions. + +Top Influential Sequences: +================================================================================ +Rank = 0; Score = 35127296.0 +<|begin_of_text|>Energy development is the field of activities focused on obtaining sources of energy from natural resources. These activities include production of conventional, alternative and renewable sources of energy, and for the recovery and reuse of energy that would otherwise be wasted. Energy conservation and efficiency measures reduce the demand for energy development, and can have benefits to society with improvements to environmental issues. + +Societies use energy for transportation, manufacturing, illumination, heating and air conditioning, and communication, for industrial, commercial, and domestic purposes. Energy resources may be classified as primary resources, where the resource can be used in substantially its original form, or as secondary resources, where the energy source must be converted into a more conveniently usable form. Non-renewable resources are significantly depleted by human use, whereas renewable resources are produced by ongoing processes that can sustain indefinite human exploitation. + +Thousands of people are employed in the energy industry. The conventional industry comprises the petroleum industry, the natural gas industry, the electrical power industry, and the nuclear industry. New energy industries include the renewable energy industry, comprising alternative and sustainable manufacture, distribution, and sale of alternative fuels. + +Classification of resources [ edit ] + +Open System Model (basics) + +Energy resources may be classified as primary resources, suitable for end use without conversion to another form, or secondary resources, where the usable form of energy required substantial conversion from a primary source. Examples of primary energy resources are wind power, solar power, wood fuel, fossil fuels such as coal, oil and natural gas, and uranium. Secondary resources are those such as electricity, hydrogen, or other synthetic fuels. + +Another important classification is based on the time required to regenerate an energy resource. "Renewable" resources are those that recover their capacity in a time significant by human needs. Examples are hydroelectric power or wind power, when the natural phenomena that are the primary source of energy are ongoing and not depleted by human demands. Non-renewable resources are those that are significantly depleted by human usage and that will not recover their potential significantly during human lifetimes. An example of a non-renewable energy source is coal, which does not form naturally at a rate that would support human use. + +Fossil fuels [ edit ] + +Fossil fuel (primary non-renewable fossil) sources burn coal or hydrocarbon fuels, which are the remains of the decomposition of plants and animals. There are three main types of fossil fuels: coal, petroleum, and natural gas. Another fossil fuel, liquefied petroleum gas (LPG), is principally derived from the production of natural gas. Heat +================================================================================ +Rank = 1; Score = 31719424.0 +<|begin_of_text|>Lady Gaga is an American singer, songwriter, and actress who has received many awards and nominations for her contributions to the music industry. She rose to prominence with the release of her debut album The Fame in 2008. The album won several awards and was nominated for six Grammy Awards, including Album of the Year. The album and its single "Poker Face" won Best Electronic/Dance Album and Best Dance Recording, respectively, at the 52nd Grammy Awards. The album also won International Album at the 2010 BRIT Awards. A follow-up EP, titled The Fame Monster, was released in 2009, and included the singles "Bad Romance" and "Telephone". The music videos of the songs won eight accolades from thirteen nominations at the 2010 MTV Video Music Awards (VMA), making Gaga the most nominated artist in VMA history for a single year and the first female artist to receive two nominations for Video of the Year in one single night.[1] In 2011, Gaga was nominated for six Grammy Awards, and won three—Best Pop Vocal Album for The Fame Monster, and Best Female Pop Vocal Performance and Best Short Form Music Video for "Bad Romance". + +Born This Way (2011), Gaga's second studio album, accrued three nominations at the 54th Annual Grammy Awards, including her third consecutive nomination for Album of the Year. It won the People's Choice Awards for Album of the Year the following year, and the music video for the title track won two VMAs, including Best Female Video. Her third album, Artpop (2013), won the award for World's Best Album by a Female at the 2014 World Music Awards. In other musical ventures, Gaga released a collaborative jazz album with Tony Bennett, titled Cheek to Cheek (2014), which received the Grammy Award for Best Traditional Pop Vocal Album. + +In 2015, Gaga released a song "Til It Happens to You", for the documentary film, The Hunting Ground. The song won a Satellite Award for Best Original Song, while nominated for an Academy Award, a Critics' Choice Movie Award for Best Song, and a Grammy Award for Best Song Written for Visual Media. In the same year she was named Woman of the Year by Billboard. She also became the first woman to receive the Digital Diamond Award from RIAA,[2] and the first artist to win the Songwriters Hall of Fame's Contemporary Icon Award for "attaining an iconic status in pop culture". Gaga received a Golden +================================================================================ +Rank = 2; Score = 22544384.0 +<|begin_of_text|>The many potential social and economic benefits from advances in AI-based technologies depend entirely on the environment in which these technologies evolve, says the Royal Society. According to a new report from the UK’s science academy, urgent consideration needs to be given to the “careful stewardship” needed over the next ten years to ensure that the dividends from machine learning – the form of artificial intelligence that allows machines to learn from data – benefit all in UK society. + +Machine Learning: the power and promise of computers that learn by example, published today (25 April 2017), comes at a critical time in the rapid development and use of this technology, and the growing debate about how it will reshape the UK economy and people’s lives. Crucially the report calls for research funding bodies to support a new wave of machine learning research that goes beyond technical challenges, and into areas aimed at addressing public confidence in machine learning – vital to the UK maintaining its internationally competitive edge at the forefront of this area. + +The report also offers the first evidence about the UK public’s views on machine learning, including the application areas about which they are particularly positive, and the need for the real-world data feeding the growth of this technology to be dealt with fairly and securely. + +Professor Peter Donnelly FRS, chair of the report’s working group and Director of the Wellcome Trust Centre for Human Genetics and Professor of Statistical Science at the University of Oxford, says: + +“Machine learning is already used in many apps and services that we encounter every day. It is used to tag people in our photos, by our phones to interpret voice commands, by internet retailers to make recommendations, and by banks to spot unusual activity on a credit or debit card. However, these current applications only scratch the surface of understanding just how powerful a technology this could be. + +“Machine learning will have an increasing impact on our lives and lifestyles over the next five to ten years. There is much work to be done so that we take advantage of machine learning’s potential and ensure that the benefits are shared, especially as this could be a key area of opportunity for the UK in the coming years.” + +An action plan so no one is left behind and the benefits are shared + +The report calls for action in a number of key areas over the next five to ten years to create an environment of “careful stewardship” that can help ensure that the benefits of this technology are felt broadly. Understanding who will be most affected, how the benefits are likely to be distributed, and where the opportunities for growth lie, will be key to designing +================================================================================ +Rank = 3; Score = 13565952.0 +<|begin_of_text|>An enterprise such as Flipkart & Amazon is the master at analyzing big data. They have a huge a knowledge gathered from analyzing big data to achieve an edge over their competitor. Just think about the Amazon’s recommendation engine. It analyzes the process of buying history, buying habits, and pattern what people like to buying. With big data and predictive analytics, they have to build a marketing strategy and created an extremely successful business model. + +So these things should be upon your mind while knowing about the big data analytics because with a rise in computational power, robust data infrastructure, rapid algorithm development, and the need to obtain better insight from an increasingly vast amount of data. + +Big Data Analytics: – Types + +Explore Analytics It can be used to explore your data in a graphical manner where your data provides some value through simple visualizations. It often used when you have a large amount of disparate data. + +Advance Analytics: It provides analytics algorithms for executing complex analysis of either structured or unstructured data. It also includes such as machine learning, statistical model, text analytics and other data-mining techniques. + +Operationalized Analytics: Are the part of the business process that helps to achieve operational efficiency by building models. For example, a data scientist for a banking organization might build a model that predicts the identity theft of its customer. + +Summary + +The capability to analyze big data provides unique opportunities for many enterprises. However comprehending big data can be challenging. Due to algorithms & technologies, basic analysis becomes confusing. It depends upon the scale size which means how many you’re capable for that. + +For any query feel free & write an email us on:-support@raygain.com or visit our website:-http://www.raygain.com/<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 4; Score = 7864320.0 +<|begin_of_text|>The past few years have seen rapid advances in machine learning, with dramatic improvements in technical performance—from more accurate speech recognition, to better image search, to improved translations. But we believe AI can go much further—and be more useful to all of us—if we build systems with people in mind at the start of the process. + +Today we’re announcing the People + AI Research initiative (PAIR) which brings together researchers across Google to study and redesign the ways people interact with AI systems. The goal of PAIR is to focus on the "human side" of AI: the relationship between users and technology, the new applications it enables, and how to make it broadly inclusive. The goal isn’t just to publish research; we’re also releasing open source tools for researchers and other experts to use. + +PAIR's research is divided into three areas, based on different user needs: + +Engineers and researchers: AI is built by people. How might we make it easier for engineers to build and understand machine learning systems? What educational materials and practical tools do they need? + +Domain experts: How can AI aid and augment professionals in their work? How might we support doctors, technicians, designers, farmers, and musicians as they increasingly use AI? + +Everyday users: How might we ensure machine learning is inclusive, so everyone can benefit from breakthroughs in AI? Can design thinking open up entirely new AI applications? Can we democratize the technology behind AI? + +We don't have all the answers—that's what makes this interesting research—but we have some ideas about where to look. One key to the puzzle is design thinking. Instead of viewing AI purely as a technology, what if we imagine it as a material to design with? Here history might serve as a guide: For instance, advances in computer graphics meant more than better ways of drawing pictures—and that led to completely new kinds of interfaces and applications. You can read more in this post on what we call human-centered machine learning (HCML).We’re open sourcing new tools, creating educational materials (such as guidelines for designing AI interfaces), and publishing research to answer these questions and spread the power of AI to as many people as possible. + +Open-source tools + +Today we're open sourcing two visualization tools, Facets Overview and Facets Dive. These applications are aimed at AI engineers, and address the very beginning of the machine learning process. The Facets applications give engineers a clear view of the data they use to train AI systems. + +We think this is important because training data is a key ingredient in modern AI systems, but it can +================================================================================ +Rank = 5; Score = 6782976.0 +<|begin_of_text|>Crowdsourcing is a sourcing model in which individuals or organizations obtain goods and services. These services include ideas and finances, from a large, relatively open and often rapidly-evolving group of internet users; it divides work between participants to achieve a cumulative result. The word crowdsourcing itself is a portmanteau of crowd and outsourcing, and was coined in 2005.[1][2][3][4] As a mode of sourcing, crowdsourcing existed prior to the digital age (i.e. "offline").[5] + +There are major differences between crowdsourcing and outsourcing. Crowdsourcing comes from a less-specific, more public group, whereas outsourcing is commissioned from a specific, named group, and includes a mix of bottom-up and top-down processes.[6][7][8] Advantages of using crowdsourcing may include improved costs, speed, quality, flexibility, scalability, or diversity.[9][10] + +Some forms of crowdsourcing, such as in "idea competitions" or "innovation contests" provide ways for organizations to learn beyond the "base of minds" provided by their employees (e.g. LEGO Ideas).[11] Tedious "microtasks" performed in parallel by large, paid crowds (e.g. Amazon Mechanical Turk) are another form of crowdsourcing. It has also been used by not-for-profit organizations and to create common goods (e.g. Wikipedia).[12] The effect of user communication and the platform presentation should be taken into account when evaluating the performance of ideas in crowdsourcing contexts.[13] + +Definitions [ edit ] + +The term "crowdsourcing" was coined in 2005 by Jeff Howe and Mark Robinson, editors at Wired, to describe how businesses were using the Internet to "outsource work to the crowd",[1] which quickly led to the portmanteau "crowdsourcing." Howe, first published a definition for the term crowdsourcing in a companion blog post to his June 2006 Wired article, "The Rise of Crowdsourcing", which came out in print just days later:[14] + +"Simply defined, crowdsourcing represents the act of a company or institution taking a function once performed by employees and outsourcing it to an undefined (and generally large) network of people in the form of an open call. This can take the form of peer-production (when the job is performed collaboratively), but is also often undertaken by sole individuals. The crucial prerequisite is the use of the open call format and the large network of potential laborers." + +In a February 1, 200 +================================================================================ +Rank = 6; Score = 6750208.0 +<|begin_of_text|>Introduction + +I’ve always thought it would be neat to create a digital version of oneself. The closest that programmers can get, is most likely the chat bot. Aside from chat bots being a goal towards beating the Turing Test, there may be an ulterior motive involved as well - to create a digital copy of one’s own personality, and thus, a pseudo form of immortality. We’ve come a long way in this progress (the chat bot part, not the immortality part), starting from the humble orgins of Eliza all the way to the highly configurable Alice chat bot. + +Chat bots typically use case-based reasoning or similar techniques to map responses to certain keywords in a sentence. In this manner, a crude personality can be programmed, and a unique chat bot created. This is definitely an interesting exercise that I think all software developers should try, at least once. But, let’s try something a little different and apply a modern spin to the idea of an artificially generated digital persona. + +Is it possible to apply machine learning to a Twitter user’s collection of tweets and accurately develop a model of the personality? + +This article describes a program, using unsupervised learning with the K-Means algorithm, to automatically group a collection of tweets into distinct categories. Once the categories have been organized, the program will then locate new links that match the categories and recommend them for the persona. In the process, we’ll create a machine learning recommendation engine for Twitter that will recommend links, related to the user’s interests, and automatically identify new content. + +Supervised vs Unsupervised Learning + +There are two core methods for building a model from data with machine learning: supervised and unsupervised learning. Since we’re trying to determine a set of topics for a Twitter user’s personality, we could choose either method to build a recommendation engine. Let’s take a look at what’s involved for each learning type. + +Supervised Learning + +Using supervised learning, such as an SVM, a model could be constructed that maps articles a user likes vs dislikes. This is similar to book or movie rating services. We would record a set number of links that the user clicks on and an equal set of links that the user skips over (assume that we’re using an RSS feed of articles in this example). The SVM would learn to classify new articles as being liked or disliked, thus determining a general personality. One drawback to this approach is the sheer amount of rated data that is required, in addition to a logging mechanism to record the likes/dislikes. Since this kind of +================================================================================ +Rank = 7; Score = 6619136.0 +<|begin_of_text|>Conceptualized by Dr. Peter Waterland, Quantum Resistant Ledger ($QRL) is a cryptocurrency ledger that uses hash-based digital signatures (proof-of-stake algorithms) instead of proof-of-work algorithms. + +Because of this, it is resistant to both quantum and classical computing attacks. QRL is one of the first blockchain-based ledger to use post-quantum cryptography technology. For this reason, it has generated a lot of interest among cryptocurrency enthusiasts and blockchain developers alike. The companies cryptocurrency $QRL is currently trading as an ERC20 token on various exchanges including Liqui, Tidex and the most popular Bittrex. + +QRL plans to allow mining (Genesis block) in September 2017. The number of coins in the Genesis block and the final distribution, which will happen in 200 years, are 65 million and 105 million respectively. + +What is Quantum Computing? + +Quantum computing is basically the use of quantum computers to perform computer-related tasks. Quantum computers store information in quantum bits (qubits) instead of bits. It is important to note that a qubit can store multiple numbers at once (a zero, a one, both a zero and a one, or any other value between a zero and a one). + +For this reason, a quantum computer can perform multiple calculations simultaneously. Put another way, a quantum computer can work in parallel, allowing it to solve complex mathematical problems involving factorization, such as “prime factors” of a large number. This means that quantum computing can break cryptographic protocols, including the ones used by blockchain networks. + +How will Quantum Resistant Ledger help crypto currency eco systems in a post-quantum computing world? + +The conventional encryption protocols used to protect blockchain networks including Elliptic Curve Digital Signature Algorithm or ECDSA, elliptical curve cryptography (ECC) and large prime number cryptography (RSA) are vulnerable to quantum computing. + +More specifically, quantum computing would make it possible to reverse engineer private keys quickly, thereby rendering the aforementioned encryption protocols useless. This is where QRL comes in handy. QRL uses post-quantum cryptography technology, which means it is theoretically immune from quantum computing attacks. + +Protecting Our Future: Quantum Computing Threats & Considerations + +Let's start with preparing for cyber-war scenarios. + +So far, blockchain-based encryption algorithms such as ECDSA have been able to keep cybercriminals at bay because classical computers cannot break such algorithms. However, the introduction of quantum computing could potentially lead to a blockchain Armageddon because of the following threats: + +1) Government Applications: +================================================================================ +Rank = 8; Score = 6619136.0 +<|begin_of_text|>The Anatomy of a Large-Scale Hypertextual Web Search Engine + +Sergey Brin and Lawrence Page + +{sergey, page}@cs.stanford.edu + +Computer Science Department, Stanford University, Stanford, CA 94305 + +Abstract + +In this paper, we present Google, a prototype of a large-scale search engine which makes heavy use of the structure present in hypertext. Google is designed to crawl and index the Web efficiently and produce much more satisfying search results than existing systems. The prototype with a full text and hyperlink database of at least 24 million pages is available at http://google.stanford.edu/ + +To engineer a search engine is a challenging task. Search engines index tens to hundreds of millions of web pages involving a comparable number of distinct terms. They answer tens of millions of queries every day. Despite the importance of large-scale search engines on the web, very little academic research has been done on them. Furthermore, due to rapid advance in technology and web proliferation, creating a web search engine today is very different from three years ago. This paper provides an in-depth description of our large-scale web search engine -- the first such detailed public description we know of to date. + +Apart from the problems of scaling traditional search techniques to data of this magnitude, there are new technical challenges involved with using the additional information present in hypertext to produce better search results. This paper addresses this question of how to build a practical large-scale system which can exploit the additional information present in hypertext. Also we look at the problem of how to effectively deal with uncontrolled hypertext collections where anyone can publish anything they want. Keywords: World Wide Web, Search Engines, Information Retrieval, PageRank, Google + +1. Introduction + +1.1 Web Search Engines -- Scaling Up: 1994 - 2000 + +1.2. Google: Scaling with the Web + +These tasks are becoming increasingly difficult as the Web grows. However, hardware performance and cost have improved dramatically to partially offset the difficulty. There are, however, several notable exceptions to this progress such as disk seek time and operating system robustness. In designing Google, we have considered both the rate of growth of the Web and technological changes. Google is designed to scale well to extremely large data sets. It makes efficient use of storage space to store the index. Its data structures are optimized for fast and efficient access (see section 4.2). Further, we expect that the cost to index and store text or HTML will eventually decline relative to the amount that will be available +================================================================================ +Rank = 9; Score = 6553600.0 +<|begin_of_text|>Photo + +Google and a corporation associated with NASA are forming a laboratory to study artificial intelligence by means of computers that use the unusual properties of quantum physics. Their quantum computer, which performs complex calculations thousands of times faster than existing supercomputers, is expected to be in active use in the third quarter of this year. + +The Quantum Artificial Intelligence Lab, as the entity is called, will focus on machine learning, which is the way computers take note of patterns of information to improve their outputs. Personalized Internet search and predictions of traffic congestion based on GPS data are examples of machine learning. The field is particularly important for things like facial or voice recognition, biological behavior, or the management of very large and complex systems. + +“If we want to create effective environmental policies, we need better models of what’s happening to our climate,” Google said in a blog post announcing the partnership. “Classical computers aren’t well suited to these types of creative problems.” + +Google said it had already devised machine-learning algorithms that work inside the quantum computer, which is made by D-Wave Systems of Burnaby, British Columbia. One could quickly recognize information, saving power on mobile devices, while another was successful at sorting out bad or mislabeled data. The most effective methods for using quantum computation, Google said, involved combining the advanced machines with its clouds of traditional computers. + +Google bought the machine in cooperation with the Universities Space Research Association, a nonprofit research corporation that works with NASA and others to advance space science and technology. Outside researchers will be invited to the lab as well. + +This year D-Wave sold its first commercial quantum computer to Lockheed Martin. Lockheed officials said the computer would be used for the test and measurement of things like jet aircraft designs, or the reliability of satellite systems. + +The D-Wave computer works by framing complex problems in terms of optimal outcomes. The classic example of this type of problem is figuring out the most efficient way a traveling salesman can visit 10 customers, but real-world problems now include hundreds of such variables and contingencies. D-Wave’s machine frames the problem in terms of energy states, and uses quantum physics to rapidly determine an outcome that satisfies the variables with the least use of energy. + +In tests last September, an independent researcher found that for some types of problems the quantum computer was 3,600 times faster than traditional supercomputers. According to a D-Wave official, the machine performed even better in Google’s tests, which involved 500 variables with different constraints. + +“The tougher, more complex ones had better performance,” said Colin Williams +================================================================================ +Rank = 10; Score = 6356992.0 +<|begin_of_text|>How to Grow and Prune a Classification Tree + +In what follows, we will describe the work of Breiman and his colleagues as set out in their seminal book "Classification and Regression Trees." Theirs is a very rich story, and we will concentrate on only the essential ideas... + +David Austin + +Grand Valley State University + +Email David Austin + +Introduction + +It's easy to collect data these days; making sense of it is more work. This article explains a construction in machine learning and data mining called a classification tree. Let's consider an example. + +In the late 1970's, researchers at the University of California, San Diego Medical Center performed a study in which they monitored 215 patients following a heart attack. For each patient, 19 variables, such as age and blood pressure, were recorded. Patients were then grouped into two classes, depending on whether or not they survived more than 30 days following the heart attack. + +Assuming the patients studied were representative of the more general population of heart attack patients, the researchers aimed to distill all this data into a simple test to identify new patients at risk of dying within 30 days of a heart attack. + +By applying the algorithm described here, Breiman, Freidman, Olshen, and Stone, created a test consisting of only three questions---what is the patient's minimum systolic blood pressure within 24 hours of being admitted to the hospital, what is the patient's age, does the patient exhibit sinus tachycardia---to identify patients at risk. In spite of its simplicity, this test proved to be more accurate than any other known test. In addition, the importance of these three questions indicate that, among all 19 variables, these three factors play an important role in determining a patient's chance of surviving. + +Besides medicine, these ideas are applicable to a wide range of problems, such as identifying which loan applicants are likely to default and which voters are likely to vote for a particular political party. + +In what follows, we will describe the work of Breiman and his colleagues as set out in their seminal book Classification and Regression Trees. Theirs is a very rich story, and we will concentrate on only the essential ideas. + +Classification trees + +Before we start developing a general theory, let's consider an example using a much studied data set consisting of the physical dimensions and species of 150 irises. Here are the data for three irises out of the 150 in the full data set: + +Sepal Length Sepal Width Petal Length Petal Width Species +================================================================================ +Rank = 11; Score = 6193152.0 +<|begin_of_text|>In the report, titled "China's Rise in Artificial Intelligence," the investment bank said the world's second-largest economy has emerged as a major global contender in using AI to drive economic progress. + +Goldman said the government and companies have identified AI and machine learning as the next big areas of innovation. + +"We believe AI technology will become a priority on the government's agenda, and we expect further national/regional policy and funding support on AI to follow," the bank said. + +AI is already widespread: From simple smartphone applications that can tell the weather to complex algorithms that are able to easily beat humans in board games. + +Companies such as Google and Microsoft have poured vast amounts of money into research and development to expand the horizon of what AI can achieve. Machines are fed large quantities of data and taught specific tasks, allowing companies to create software that can learn and become smarter. + +While the United States is generally considered to be leading the field, other countries are catching up. China, home of internet powerhouses such as Baidu, Alibaba and Tencent, is one of them. + +In July, China's State Council issued guidelines on developing AI inside the country and set a goal of becoming a global innovation center for it by 2030. It expects the total output value of AI industries to surpass 1 trillion yuan ($147.80 billion). + +The Council encouraged the creation of open-source computing platforms and training more AI professionals and scientists. The guidelines said the government will invest in qualified AI projects and encourage private capital investment.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 12; Score = 6029312.0 +<|begin_of_text|>We CAN predict the future (a bit): Why the brain knows what's going to happen before it does + +People subconsciously make thousands of tiny predictions each day, whether it's contemplating when a bus will arrive, who is knocking on the door or if a dropped glass will break. + +Now scientists are beginning to unravel how the brain is such a surprisingly accurate fortune-teller - but only when it comes to mundane events. + +Researchers at Washington University in St Louis focused on the mid-brain dopamine system (MDS), which provides signals to the rest of the brain when unexpected events occur. + +Each of us makes thousands of tiny predictions, such as contemplating when a bus will arrive, every day. Scientists are now beginning to unravel how the brain is such a surprisingly accurate fortune-teller + +Using functional MRI (fMRI), they found that this system encodes prediction error when viewers are forced to choose what will happen next in a video of an everyday event. + +They found that between 80 and 90 per cent of viewer predictions were correct, depending on when the footage was stopped. + +Lead researcher Jeffrey Zacks said predicting the near future is vital in guiding behaviour and is a key component of theories of perception, language processing and learning. + +He said: 'It's valuable to be able to run away when the lion lunges at you, but it's super-valuable to be able to hop out of the way before the lion jumps. + +'It's a big adaptive advantage to look just a little bit over the horizon.' + +The research will also help those in the early stages of neurological diseases such as schizophrenia, Alzheimer's and Parkinson's, Professor Zacks said. + +The scientists tested healthy young volunteers who were shown films of everyday events such as washing a car, building a Lego model or washing clothes. The film would be watched for a while, and then it was stopped. + +Participants were then asked to predict what would happen five seconds later when the film was re-started by selecting a picture that showed what would happen. + +Half of the time, the movie was stopped just before an event boundary, when a new event was just about to start. The other half of the time, the film was stopped in the middle of an event. + +The researchers found that participants were more than 90 per cent correct in predicting activity within the event, but less than 80 per cent correct in predicting across the event boundary. They were also less confident in their predictions. + +'Successful predictions are associated with the subjective experience of a smooth stream of consciousness' + +Professor Zacks said: +================================================================================ +Rank = 13; Score = 5865472.0 +<|begin_of_text|>A year ago, we shared with you our belief that Algorithm Development is Broken, where we laid out the problems in the world of algorithm development, and our vision of how to make things better. + +Since then, we’ve been working hard to create our vision of the world’s first open marketplace for algorithms – and you, the developers of the world, have enthusiastically joined us to help active the algorithm economy. + +Today, we are proud to open our doors to the world, and deliver on our promise of a global, always-available, open marketplace for algorithms. For the first time, algorithm developers can publish their algorithms as a live API and applications can be made smarter by taking advantage of these intelligent algorithms. + +Algorithm Development with Algorithmia + +Live : easy access to any algorithm, callable at any time + +: easy access to any algorithm, callable at any time No dependencies : no downloads, no installs, no stress + +: no downloads, no installs, no stress Interoperable : call any algorithm regardless of programming language + +: call any algorithm regardless of programming language Composable : algorithms are building blocks, build something bigger! + +: algorithms are building blocks, build something bigger! Invisible but powerful infrastructure : don’t worry about containers, networks, or scaling; we make your algorithm scale to hundreds of machines + +: don’t worry about containers, networks, or scaling; we make your algorithm scale to hundreds of machines No operations: we scale, meter, optimize and bill for you + +With our public launch we are enabling payments and go from being a repository of algorithms, to a true marketplace: + +For algorithm developers, Algorithmia provides a unique opportunity to increase the impact of their work and collect the rewards associated with it. + +For application developers, Algorithmia provides the largest repository of live algorithms ever assembled, supported by a vibrant community of developers. + +So far, users have contributed over 800 algorithms, all under one API that grows every day. The most exciting thing from our point of view is watching the Algorithmia Marketplace grow and expand in its functionality. + +In just the last few months, Algorithmia has gained the ability to: + +Summarize and generate topics from unstructured text + +Extract structure and information from arbitrary websites + +Perform computer vision tasks like Face Detection and Image Similarity + +Look for patterns and anomalies in data + +Dozens of microservice building blocks + +… with more added every day + +Thanks to the thousands of developers who joined our cause, contributed algorithms, and gave us countless hours of feedback so far. We are very excited to be opening our doors, and +================================================================================ +Rank = 14; Score = 5767168.0 +<|begin_of_text|>Feb 6, 2017 + +This blog post will demonstrate how deep reinforcement learning (deep Q-learning) can be implemented and applied to play a CartPole game using Keras and Gym, in less than 100 lines of code! + +I’ll explain everything without requiring any prerequisite knowledge about reinforcement learning. + +The code used for this article is on GitHub. + +Reinforcement Learning + +Reinforcement Learning is a type of machine learning that allows you to create AI agents that learn from the environment by interacting with it. Just like how we learn to ride a bicycle, this kind of AI learns by trial and error. As seen in the picture, the brain represents the AI agent, which acts on the environment. After each action, the agent receives the feedback. The feedback consists of the reward and next state of the environment. The reward is usually defined by a human. If we use the analogy of the bicycle, we can define reward as the distance from the original starting point. + +Deep Reinforcement Learning + +Google’s DeepMind published its famous paper Playing Atari with Deep Reinforcement Learning, in which they introduced a new algorithm called Deep Q Network (DQN for short) in 2013. It demonstrated how an AI agent can learn to play games by just observing the screen without any prior information about those games. The result turned out to be pretty impressive. This paper opened the era of what is called ‘deep reinforcement learning’, a mix of deep learning and reinforcement learning. + +Click to Watch: DeepMind’s Atari Player + +In Q-Learning Algorithm, there is a function called Q Function, which is used to approximate the reward based on a state. We call it Q(s,a), where Q is a function which calculates the expected future value from state s and action a. Similarly in Deep Q Network algorithm, we use a neural network to approximate the reward based on the state. We will discuss how this works in detail. + +Cartpole Game + +Usually, training an agent to play an Atari game takes a while (from few hours to a day). So we will make an agent to play a simpler game called CartPole, but using the same idea used in the paper. + +CartPole is one of the simplest environments in OpenAI gym (a game simulator). As you can see in the animation from the top, the goal of CartPole is to balance a pole connected with one joint on top of a moving cart. Instead of pixel information, there are 4 kinds of information given by the state, such as angle of the +================================================================================ +Rank = 15; Score = 5734400.0 +<|begin_of_text|>In the world of elearning, microlearning has become a buzzword. It’s considered a powerful new approach to training the workforce. With the average attention span in North America dropping from 12 seconds in 2000 to 8 seconds in 2015, the demand for shorter, more engaging training is higher than ever! + +In this post, our goal is to cover the basics and leave you with an understanding of: What Microlearning is and the benefits it can provide. Throughout this post, we’ll try to give you examples of how to use microlearning in your own training programs. + +What is Microlearning? + +Microlearning is focused learning that is delivered in bite sized chunks. Since this method of learning provides small bits of knowledge at a time, it’s best used for delivering information that learners need to retain. + +Microlearning can be achieved using a number of different delivery methods: emails, online posts and short multimedia videos, are all examples of different ways you can deliver training that is designed for your learners to retain new knowledge and achieve their educational goals. + +Examples of Microlearning: + +– Watching video tutorials on Youtube + +– Receiving small bits of education via email: like Word of the Day from Dictionary.com + +– Online learning programs like Duolingo or Lynda + +What are the benefits that Microlearning can provide? + +1. Avoid the risk of overwhelming learners + +Microlearning allows learners to move at their own pace, giving them the ability to go back and review complex concepts as often as needed. Since new knowledge is delivered in smaller chunks, learners avoid the risk of being overwhelmed by too much information at once. + +2. Create on-the-go training that can be accessed anywhere, at anytime. + +Microlearning can be achieved using a number of different delivery methods. Email, online posts, videos, even tweets because of this training can usually be accessed across multiple devices, making it available on-the-go. Learners can access and review training materials while doing everyday tasks like: waiting for the bus, sitting on the train, or even riding the elevator! + +3. Help learners better retain new knowledge + +Traditional classroom training often provides little to no long term takeaways for learners. The Wall Street Journal recently reported that 90 percent of new skills are lost within a year of training! Microlearning breaks new knowledge down into short chunks making learning easier to digest, understand, and apply on the job. + +Different ways to use Microlearning: + +– Health and Safety training + +– Learning new software + +– Business Processes and Procedures + +Microlearning yields many benefits for organizations looking +================================================================================ +Rank = 16; Score = 5734400.0 +<|begin_of_text|>Just a few years ago the common perception was that we were in an AI winter. + +Although there were lots of narrow AI applications running in the background of our daily lives, there wasn’t much enthusiasm. + +But quietly in the background, a revolution was building thanks to progress across a few key areas. These areas would soon converge to produce breakthrough after breakthrough and put us on the verge of what many believe to be the most important event in human history. + +Key Advance 1) More Data + +Andrew Ng of Baidu explains that a massive amount of data is needed. If you put 10x the data in many of these algorithms they work, put 1/10th in and they don’t. + +Previously it was very difficult to get hold of the massive amounts of data required to feed the AI systems, but thanks to the internet researchers have tonnes of data to train their neural nets. + +Key Advance 2) More Computing Power + +If you only have the computational power to build a small neural network it doesn’t work. But computer power has continued to increase and prices have dropped. + +Moore’s Law may be stalling, but the Law of Accelerating Returns is not. + +The Law of Accelerating Returns. What Steve Jurvetson calls “the most important graph ever”. + +GPUs are much better for training neural networks than CPUs and have provided the computer power needed for these algorithms to function. + +Infrastructure has also improved. Today it’s possible for anyone to rent a massive amount of GPU power on cloud computing platforms (AWS, Microsoft Azure, Google Cloud, IBM Cloud). + +Key Advance 3) Better Algorithms + +Neural networks have been known about for decades, but most researchers had given up on them + +Geoffrey Hinton of Google is one of the few who stuck with it. Despite his peers calling it a dead end, he believed that it was the right approach. It turned out he was right. + +Hinton learned how to stack neural networks dozens of layers deep (deep learning) which enabled vastly more calculations and now he is considered “the godfather of neural networks”. + +With these 3 breakthroughs in place neural networks finally began to work. And they worked better than almost anyone expected. + +The Tipping Point: ImageNet 2012 + +The ImageNet project was created in 2009 to judge how well computers can see. + +In 2011 computers had a 26% error rate when trying to label images. Humans only had a 5% error rate. + +But in 2012, Hinton’s team made a breakthrough +================================================================================ +Rank = 17; Score = 5537792.0 +<|begin_of_text|>* Dept of Homeland Security: Java vulnerable to hackers + +* Could be used to steal identity, form malicious networks + +* Applies to browsers on all major operating systems + +By Jim Finkle + +Jan 11 (Reuters) - The U.S. Department of Homeland Security urged computer users to disable Oracle Corp’s Java software, amplifying security experts’ prior warnings to the hundreds of millions of consumers and businesses that use it to surf the Web. + +Hackers have figured out a way to exploit Java to install malicious software enabling them to commit crimes ranging from identity theft to making an infected computer part of an ad-hoc network of computers that can be used to attack websites. + +“We are currently unaware of a practical solution to this problem,” the Department of Homeland Security’s Computer Emergency Readiness Team said in a posting on its website late on Thursday. + +“This and previous Java vulnerabilities have been widely targeted by attackers, and new Java vulnerabilities are likely to be discovered,” the agency said. “To defend against this and future Java vulnerabilities, disable Java in Web browsers.” + +Java is a computer language that enables programmers to write software utilizing just one set of code that will run on virtually any type of computer, including ones that use Microsoft Corp’s Windows, Apple Inc’s OS X and Linux, an operating system widely employed by corporations. + +Computer users access Java programs through modules, or plug-ins, that run Java software on top of browsers such as Internet Explorer and Firefox. + +The U.S. government’s warning on Java came after security experts earlier on Thursday warned of the newly discovered flaw. + +It is relatively rare for government agencies to advise computer users to completely disable software due to a security bug, particularly in the case of widely used programs such as Java. They typically recommend taking steps to mitigate the risk of attack while manufacturers prepare an update, or hold off on publicizing the problem until an update is prepared. + +In September, the German government advised the public to temporarily stop using Microsoft’s Internet Explorer browser to give it time to patch a security vulnerability that opened it to attacks. + +The Department of Homeland Security said that attackers could trick targets into visiting malicious websites that would infect their PCs with software capable of exploiting the bug in Java. + +It said that an attacker could also infect a legitimate website by uploading malicious software that would infect machines of computer users who trust that site because they have previously visited it without experiencing any problems. + +They said developers of several popular tools known as exploit kits, which criminal hackers use to attack PCs, have added software that allows hackers to exploit the newly discovered bug in Java to attack +================================================================================ +Rank = 18; Score = 5505024.0 +<|begin_of_text|>ARMONK, NY - 28 Aug 2012: IBM (NYSE: IBM) today announced the zEnterprise® EC12 mainframe server, the most powerful and technologically advanced version of an IBM system that has been the linchpin of enterprise computing for 48 years. The new enterprise system features technologies that demonstrate IBM’s ongoing commitment to meet the growing need to secure and manage critical information with the System z mainframe. + +Mainframes support significant portions of the data environment at most large enterprises. As these enterprises grapple with the well-documented growth of data, they are looking for new ways to secure and gain insights from such critical information as financial, customer and enterprise resource data that will enable them to provide their clients with new services. The new zEC12 offers industry-leading security and robust support for operational analytics that can help clients efficiently sift through large volumes of raw data and transform it to gain knowledge that can be used for competitive advantage. For example, a retailer managing online transactions on zEC12 can gain insights from client information that will enable it to provide clients with a more customized shopping experience. + +The IBM zEC12 enterprise system is the result of an investment by IBM Systems and Technology Group of more than $1 billion in IBM research and development primarily in Poughkeepsie, New York as well as 17 other IBM labs around the world and in collaboration with some of IBM's top clients. + +The new IBM mainframe is one of the most secure enterprise systems ever (1), with built-in security features designed to meet the security and compliance requirements of different industries. With operational analytics and near real-time workload monitoring and analysis, clients can use the new zEC12 for a variety of workloads, including hybrid clouds that can take advantage of the System's 25% more performance per core and 50% greater total system capacity than its predecessor as well as the world’s fastest chip running at 5.5 GHz (2). + +Ultimate Security for Critical Information + +IBM System z is a leading platform for secure data serving and the only commercial server to achieve Common Criteria Evaluation Assurance Level 5+ security classification, providing clients the confidence to run many different applications containing confidential data on a single mainframe. The new zEC12 builds on this with innovative security and privacy features to help protect data at rest or in flight -- a critical capability in the age of Internet banking and mobile devices. + +zEC12 includes a state-of-the-art, tamper-resistant cryptographic co-processor called Crypto Express4S that provides privacy for +================================================================================ +Rank = 19; Score = 5439488.0 +<|begin_of_text|>In the age of ‘Big Data,’ with datasets rapidly growing in size and complexity and cloud computing becoming more pervasive, data science techniques are fast becoming core components of large-scale data processing pipelines. + +Apache Spark offers analysts and engineers a powerful tool for building these pipelines, and learning to build such pipelines will soon be a lot easier. Databricks is excited to be working with professors from University of California Berkeley and University of California Los Angeles to produce two new upcoming Massive Open Online Courses (MOOCs). Both courses will be freely available on the edX MOOC platform in spring summer 2015. edX Verified Certificates are also available for a fee. + +The first course, called Introduction to Big Data with Apache Spark, will teach students about Apache Spark and performing data analysis. Students will learn how to apply data science techniques using parallel programming in Spark to explore big (and small) data. The course will include hands-on programming exercises including Log Mining, Textual Entity Recognition, Collaborative Filtering that teach students how to manipulate data sets using parallel processing with PySpark (part of Apache Spark). The course is also designed to help prepare students for taking the Spark Certified Developer exam. The course is being taught by Anthony Joseph, a professor at UC Berkeley and technical advisor at Databricks, and will start on February 23rd June 1st, 2015. + +The second course, called Scalable Machine Learning, introduces the underlying statistical and algorithmic principles required to develop scalable machine learning pipelines, and provides hands-on experience using PySpark. It presents an integrated view of data processing by highlighting the various components of these pipelines, including exploratory data analysis, feature extraction, supervised learning, and model evaluation. Students will use Spark to implement scalable algorithms for fundamental statistical models while tackling real-world problems from various domains. The course is being taught by Ameet Talwalkar, an assistant professor at UCLA and technical advisor at Databricks, and will start on April 14th June 29th, 2015. + +Both courses are available for free on the edX website. You can sign up for them today:<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 20; Score = 5406720.0 +<|begin_of_text|>The Bitcoin technologies, namely blockchains are sidechains are used for various other purposes than just for registering and facilitating transactions over the Bitcoin network. Some of the new companies like Astroblocks and Bitproof are building Proof-of-Existence platforms (there is also another platform by the same name) over the Bitcoin blockchain. + +These Proof-of-Existence platforms provides a smart tamper-proof way of storing encrypted information on the Bitcoin blockchain. Each file stored on the blockchain will be associated with a unique irreplicable transaction hash. These transaction hashes will act as a reference to the documents stored on the blockchain. Anything once written on to the blockchain will be stored forever, without any way of deleting it and that is what makes it an ideal electronic system to certify and store documents. + +We can safely say that intellectual property is the most valuable asset any individual or company can have. The process of obtaining protection for your intellectual property is still a cumbersome process. It involves filing the necessary applications, multiple times in few cases and waiting for a long time for the authorities to verify and grant protection. Blockchain based Proof-of-Existence platforms provides an alternate way for both individuals and companies alike to document their creations or discoveries and handle IP disputes more efficiently. + +Whether it is a new scientific invention, design, artwork, screenplay, story or anything. The creator can record his works on these Proof-of-Existence platforms. In case of any dispute, the inventor can always contest any claims by using the transactional hash of the concerned document. The hash will also act as a timestamp that can be used to identify the time when it was uploaded onto the blockchain. + +Now that anyone can make use of these Proof-of-Existence services, making it easier for them to claim and receive credit for their innovation and creation, no matter how small it is.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 21; Score = 5406720.0 +<|begin_of_text|>Artificial intelligence (AI) may conjure up images of intelligent systems rising up to enslave us, with sadistic driverless cars taking passengers on terrifying joyrides and malevolent smart fridges refusing to order milk. + +But in reality, AI is already here; many of us just do not realise it yet. Through the development of machine learning, the technology industry has created smart systems designed to help, not hinder, people. + +Rather than send murderous robots back in time to destroy us, machine learning sits behind the scenes in the form of algorithms that learn from data to better deliver services offered by software applications. + +This is the direction that Forrester principal analyst Diego Lo Giudice believes AI is heading: "While some AI research still tries to simulate our brain or certain regions of it - and is frankly unlikely to deliver concrete results anytime soon - most of it now leverages a less human, but more effective, approach revolving around machine learning and smart integration with other AI capabilities," he said in a blog post. + +One example of this is Office Graph, created by Microsoft. Office Graph sits behind Office 356 and maps the relationships between people, content and activity across the company's productivity software suite. + +Office Graph uses these relationships to serve up documents, files, email messages, and other elements in Office 365 that are relevant to the work the user is carrying out, with the goal of making the user more productive. + +Another example is Microsoft's Skype Translator, which uses machine learning capabilities to improve the accuracy of its translation of spoken languages, something CEO Satya Nadella likens to "magic". + +"The one fascinating feature of this is something called transfer learning. What happens is you teach it English and it learns English. Then you teach it Mandarin, and it learns Mandarin, but it becomes better at English," he said. + +"Then you teach it Spanish, and it gets good at Spanish but gets great at both Mandarin and English. Quite frankly none of us knows exactly why. It's magic." + +Not wanting to be left behind, Google recently announced an update to Google Translate to provide real-time translations that use machine learning to deliver better results the more it is used. + +Big data analytics + +But, while machine learning might improve the performance of consumer-level services, it is in the enterprise world that it really shines. + +Big data is a big business, but trawling through hundreds of gigabytes of often unstructured information can be a lengthy and resource-sapping process. Add machine learning into the mix and that process becomes a lot easier. + +Rather than +================================================================================ +Rank = 22; Score = 5177344.0 +<|begin_of_text|>Computer Vision for Music Identification + +by Yan Ke, Derek Hoiem, and Rahul Sukthankar + +Abstract: + +We describe how certain tasks in the audio domain can be effectively addressed using computer vision approaches. This paper focuses on the problem of music identification, where the goal is to reliably identify a song given a few seconds of noisy audio. Our approach treats the spectrogram of each music clip as a 2-D image and transforms music identification into a corrupted sub-image retrieval problem. By employing pairwise boosting on a large set of Viola-Jones features, our system learns compact, discriminative, local descriptors that are amenable to efficient indexing. During the query phase, we retrieve the set of song snippets that locally match the noisy sample and employ geometric verification in conjunction with an EM-based "occlusion" model to identify the song that is most consistent with the observed signal. We have implemented our algorithm in a practical system that can quickly and accurately recognize music from short audio samples in the presence of distortions such as poor recording quality and significant ambient noise. Our experiments demonstrate that this approach significantly outperforms the current state-of-the-art in content-based music identification. + +Y. Ke, D. Hoiem, and R. Sukthankar. In Proceedings of Computer Vision and Pattern Recognition, 2005. [PDF 240KB] + +Demo and Code available here: + +Demonstration video: [CVPR05Video.avi 16MB]. Encoded with Indeo 5.10 and IMA ADPCM. There are both Windows and Linux codecs. + +Windows/Cygwin and Linux binary demo code and sample music: [mrdemo.tgz 15.7MB] + +C++ server code: [musicretr-1.0.tar.gz 87KB] + +Java graphical user interface: [mrgui.tgz 21KB]<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 23; Score = 5111808.0 +<|begin_of_text|>A cam is ausually done with a digital video camera. A mini tripod is sometimes used, but a lot of the time this wont be possible, so the camera make shake. Also seating placement isn’t always idle, and it might be filmed from an angle. If cropped properly, this is hard to tell unless there’s text on the screen, but a lot of times these are left with triangular borders on the top and bottom of the screen. Sound is taken from theof the camera, and especially in comedies, laughter can often be heard during the film. Due to these factors picture and sound quality are usually quite poor, but sometimes, and the theater will be fairly empty and a fairly clear signal will be heard.A telesync is the same spec as a CAM except it uses(most likely an audio jack in the chair for hard of hearing people). A direct audio source does not ensure a good quality audio source, as a lot of background noise can interfere. A lot of the times a telesync is filmed in an empty cinema or from the projection booth with a professional camera, giving a better picture quality. Quality ranges drastically, check the sample before downloading the full release.A telecine machine copies the film digitally from the. Sound and picture should be very good, but due to the equipment involved and cost telecines are fairly uncommon. Generally the film will be in correct aspect ratio, althoughtelecines have existed. A great example is thedone a few years ago. TC should not be confused with TimeCode, which is a visible counter on screen throughout the film., sent to rental stores, and various other places for promotional use. A screener is supplied on a VHS tape, and is usually in a 4:3 (full screen) a/r, although letterboxed screeners are sometimes found. The main draw back is a “ticker“ (a message that scrolls past at the bottom of the screen, with the copyright and anti-copy telephone number). Also, if the tape contains any serial numbers, or any other markings that could lead to the source of the tape, these will have to be blocked, usually with a black mark over the section. This is sometimes only for a few seconds, but unfortunately on some copies this will last for the entire film, and some can be quite big. Depending on the equipment used, screener quality can range from excellent if done from a MASTER copy, to very poor if done on an old VHS recorder thru poor capture equipment on a copied +================================================================================ +Rank = 24; Score = 5079040.0 +<|begin_of_text|>Hermeticism is based on the teachings of a mysterious man named Hermes Trismegistus. He is portrayed as a wise teacher, a powerful magician, and a skilled mystic. He has been seen as a teacher of Moses, the inventor of alchemy, and the founder of occult schools throughout history. + +Alexandrian Origins + +Hermes Trismegistus is a composite of several mythological figures. He arose in Ptolemaic Alexandria, where the mishmash of Hellenistic and Egyptian culture bred dozens of interesting cults, religions, and deities. Early on, the Hellenistic deity Hermes became associated with the Egyptian deity Thoth.1 Both were inventors of writing and gods of magic. Both were also psychopomps, responsible for guiding souls in the afterlife. As a result, over the centuries they became identified with each other and were even worshiped together in certain Egyptian temples. In fact, Khmun, the Egyptian center for the cult of Thoth, became known as Hermopolis under the Ptolemies. + +Because of this, Hermes Trismegistus, as a legendary teacher of magic and mysticism, was probably a humanized syncretization of Hermes and Thoth. His legacy has been immense. + +Multiple Hermes? + +Even though Hermes Trismegistus was a mythical figure, many ancient writers wrote about him as if he was a real person. This produced disagreements and confusion. At some point, it began to be assumed that there were two Hermes. In Asclepius, Hermes Trismegistus talks about his grandfather: + +Is it not true that my grandfather Hermes, after whom I am named, resides in his eponymous town whence he aids and cures all those who come to him from every land? + +Asclepius 372 + +This passage indicates that the Grandfather Hermes is in fact identical with the deity Hermes, residing at Hermopolis.3 Ancient writers invented additional Hermes to fill in the gaps. In fact, it could be that the title “Trismegistus” refers to many great Hermes characters, a line of sages and mystics bringing the Hermetic teachings to humanity over many generations.4 + +Astrologer, Magus, Alchemist + +The Hermetica that we read and write about most often are not the only ancient books ascribed to Hermes Trismegistus. Multiple important early works on astrology were also attributed to the legendary sage. The link between Hermes and +================================================================================ +Rank = 25; Score = 4980736.0 +<|begin_of_text|>Introduction A carbon footprint is the measure of the amount of greenhouse gases (Carbon footprint), measured in units of carbon dioxide, produced by human activities. A carbon footprint can be measured for an individual or an organization, and is typically given in tons of CO 2 -equivalent (CO 2 -eq) per year. For example, the average North American generates about 20 tons of CO 2 -eq each year. The global average carbon footprint is about 4 tons of CO 2 -eq per year (Figure 1). + +Introduction A carbon footprint is the measure of the amount of greenhouse gases (Carbon footprint), measured in units of carbon dioxide, produced by human activities. A carbon footprint can be measured for an individual or an organization, and is typically given in tons of CO 2 -equivalent (CO 2 -eq) per year. For example, the average North American generates about 20 tons of CO 2 -eq each year. The global average carbon footprint is about 4 tons of CO 2 -eq per year (Figure 1). + +Figure 1. National carbon dioxide (CO2) emissions per capita. (2005). (Source: UNEP/GRID-Arendal Maps and Graphics Library) + +An individual’s or organization’s carbon footprint can be broken down into the primary and secondary footprints. The primary footprint is the sum of direct emissions of greenhouse gases from the burning of fossil fuels for energy consumption and transportation. More fuel-efficient cars have a smaller primary footprint, as do energy-efficient light bulbs in your home or office. Worldwide, 82% of anthropogenic greenhouse gas emissions are in the form of CO 2 from fossil fuel combustion (Figure 2). + +The secondary footprint is the sum of indirect emissions of greenhouse gases during the lifecycle of products used by an individual or organization. For example, the greenhouse gases emitted during the production of plastic for water bottles, as well as the energy used to transport the water, contributes to the secondary carbon footprint. Products with more packaging will generally have a larger secondary footprint than products with a minimal amount of packaging. + +Greenhouse gases and the greenhouse effect + +Figure 2. Anthropogenic greenhouse gas emissions. (Source: Energy Information Administration) + +Although carbon footprints are reported in annual tons of CO 2 emissions, they actually are a measure of total greenhouse gas emissions. A greenhouse gas is any gas that traps heat in the atmosphere through the greenhouse effect. Because of the presence of greenhouse gases in our atmosphere the average temperature of the Earth +================================================================================ +Rank = 26; Score = 4980736.0 +<|begin_of_text|>Ajax a short form of Asynchronous JavaScript and XML is a set of techniques used by many modern and popular web sites. This technique allows portions of a Web page to be updated without refreshing the entire page, a significant advance in the development of websites and web-based applications. For example, a site might allow you to rate a product or item by clicking on a star image. With Ajax, your vote could be registered without having to load the entire page again. In this post, we are listing down 15 Beautiful Web Designs Empowered With Ajax Techniques for your inspiration. If you like this post so please spread the word as much as possible via twitter. + +You are welcome if you want to share more web designs empowered with Ajax techniques that we have missed here and you think our readers/viewers may like. Do you want to be the first one to know the latest happenings at SmashingApps.com just subscribe to our rss feed and you can follow us on twitter as well. + +I will appreciate if you can spread the word via Digg, Stumbleupon and other social media websites, Thank you. + +BBC + +MSNBC + +White House + +Ajax Whois + +iGoogle + +Netvibes + +Pageflakes + +Protopage + +My Live + +My Yahoo + +eskobo + +Symbaloo + +Pingle + +Inbox + +Shelfari<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 27; Score = 4915200.0 +<|begin_of_text|>How to use your Emotional Intelligence to Better PM + +Veamly Blocked Unblock Follow Following Jun 26, 2017 + +(Originally posted on Veamly) + +Note : This blog is based on Vivek Bedi’s talk, The VP of Product at LearnVest on Product School + +As a Product Manager, you don’t only need to oversee the production process, but you also have to lead and motivate your team. After all, they are the people behind the success or the failure of the product. But the thing is that, they all have different personalities and come from very different backgrounds. + +You have the shy and introvert engineers, the extrovert developers, the young and goal driven marketers, the creative designers, etc. You also need to deal with the product’s users on a daily basis: Make sure that your clients are satisfied, understand clearly their asks and know how to approach each one. That’s why it is important to use your Emotional Intelligence in order to better assess your team and your end users’ behaviors. + +Imagine this! + +If you have to release a product within 2 months and you have to choose the person you will be spending most of your time with, who would you choose? How would you describe this person? + +If your answer is one or more of the following adjectives : Patient, Empathetic, Thoughtful, Flexible, Hard worker… You may be interested in the soft skills of the person rather than his hard skills and IQ. Indeed, you want to work with a person you like but who also helps you process fast. In product Management, more than any other profession, the soft skills are extremely important. + +Imagine also receiving a request from a client for a feature that you would describe as obvious. You may have to explore your empathy since recognizing that not everyone interprets things the same way. + +So What’s Emotional Intelligence? + +Emotional Intelligence is the ability to identify and manage your own emotions and the emotions of others. It is generally said to include three skills: emotional awareness; the ability to harness emotions and apply them to tasks like thinking and problem solving; and the ability to manage emotions, which includes regulating your own emotions and cheering up or calming down other people. (Source) + +So, Emotional Intelligence (EQ or EI) in a nutshell is defined as the ability to + +Recognize, understand and manage our own emotions + +Recognize, understand and influence the emotions of others + +How to leverage your Emotional intelligence for Product Management? + +Here are 8 Tips that we gathered from Vivek Bed +================================================================================ +Rank = 28; Score = 4915200.0 +<|begin_of_text|>By Ben Lorica + +For companies in the early stages of grappling with big data, the analytic lifecycle (model building, deployment, maintenance) can be daunting. In earlier posts I highlighted some new tools that simplify aspects of the analytic lifecycle, including the early phases of model building. But while tools are allowing companies to offload routine analytic tasks to business analysts, experienced modelers are still needed to fine-tune and optimize, mission-critical algorithms. + +Model Selection: Accuracy and other considerations + +Accuracy1 is the main objective and a lot of effort goes towards raising it. But in practice tradeoffs have to be made, and other considerations play a role in model selection. Speed (to train/score) is important if the model is to be used in production. Interpretability is critical if a model has to be explained for transparency2 reasons (“black-boxes” are always an option, but are opaque by definition). Simplicity is important for practical reasons: if a model has “too many knobs to tune” and optimizations have to be done manually, it might be too involved to build and maintain it in production3. + +Chances are a model that’s fast, easy to explain (interpretable), and easy to tune (simple), is less4 accurate. Experienced model builders are valuable precisely because they’ve weighed these tradeoffs across many domains and settings. Unfortunately not many companies have the experts that can identify, build, deploy, and maintain models at scale. (An example from Google illustrates the kinds of issues that can come up.) + +Fast and/or scalable + +scikit-learn is a popular, well-documented package with models that leverage IPython’s distributed computing features. Over time MLbase will include an optimizer that simplifies model selection for a variety of machine-learning tasks. Depending on their performance requirements, scikit-learn and MLbase users may still need to recode models to speed them up for production. Solutions like VW, H20, and wise.io give users access to fast and scalable implementations of a few algorithms: in the case of wise.io, the company focuses only on Random Forest. Statistics software vendors like Revolution Analytics, SAS, and SPSS have been introducing scalable algorithms and tools for managing the analytic lifecycle. + +The creators of a popular open source, machine-learning project recently launched a startup, focused on building commercial-grade options and tools. GraphLab provides fast toolkits for a range of problems including collaborative filtering, clustering, text mining (topic models), and graph analytics. Having had a chance to play with an early version +================================================================================ +Rank = 29; Score = 4849664.0 +<|begin_of_text|>OctaEdit has been built in a modular fashion, where each module can be used for various tasks, and each module can interact with the other modules in various ways. + +OctaEdit has eleven (11) Modules: Project, Samples, Sequencer, Manager, Chainer, Arp Designer, LFO Designer, Library, Analytics, Options and Support. + +OctaEdit LE is restricted to five (5) Modules: Manager, Chainer, Analytics, Options and Support. + +The Manual is available to download for further details. + +OctaEdit YouTube Channel + +Please visit our Forum for detailed information and FAQ’s. + +The Project Module + +The Project module provides access to all the settings of the currently loaded Project, includes integration with the Library module and copy/paste functionality. + +The Samples Module. + +The Samples module allows the ability to manage all sample slots within a project. Features include Sample Slicing/Editing, drag/drop from file browser or audio browser, sample remapping, sample playback, sample editing, renaming and analysis amongst others. + +The Sequencer Module. + +The Sequencer module provides full access to every element of the Octatrack, from entire configurations down to a parameter locks on a single step. + +The Sequencer module is fully integrated with the Arp and LFO Designer modules allowing quick and easy modifications on each track. + +The Sequencer module is also fully integrated with the Library module allowing the ability to Save/Load any element of your setup, from your Parts, Audio/Midi/Recoder Machine configurations, Scenes through to FX Chains, Amp configurations and much, much more… + +The Manager module. + +The Manager module provides the ability to copy “Elements” from a Source Set/Project to a Target Set/Project. + +Elements can be copied at the Project, Bank, Pattern, Track, Part or Scene level, from any Project in any Set to any other Project in any other Set, and includes the ability to remap sample assignments on the fly. + +Boasting a feature set that is way too long to list, with just a couple of clicks entire sets can be merged together seamlessly, meaning you never have to worry about the Octatrack structure ever again. + +The Chainer Module. + +The Chainer module allows the ability to create Sample Chains on the fly from various audio files. Features include drag/drop from file browser and integration with OctaEdit’s other modules like the ‘Samples’ module, randomisation and so on. + +The Arp Designer Module. + +The Arp Designer module provides the ability to graphically design arpeggios +================================================================================ +Rank = 30; Score = 4718592.0 +<|begin_of_text|>Three Challenges for Artificial Intelligence in Medicine + +Brandon Ballinger Blocked Unblock Follow Following Sep 19, 2016 + +Why is the world’s most advanced AI used for cat videos, but not to help us live longer and healthier lives? A brief history of AI in Medicine, and the factors that may help it succeed where it has failed before. + +Imagine yourself as a young graduate student in Stanford’s Artificial Intelligence lab, building a system to diagnose a common infectious disease. After years of sweat and toil, the day comes for the test: a head-to-head comparison with five of the top human experts in infectious disease. Over the first expert, your system squeezes a narrow victory, winning by just 4%. It beats the second, third, and fourth doctors handily. Against the fifth, it wins by an astounding 52%. + +Would you believe such a system exists already? Would you believe it existed in 1979? This was the MYCIN project, and in spite of the excellent research results, it never made its way into clinical practice. [1] + +In fact, although we’re surrounded by fantastic applications of modern AI, particularly deep learning — self-driving cars, Siri, AlphaGo, Google Translate, computer vision — the effect on medicine has been nearly nonexistent. In the top cardiology journal, Circulation, the term “deep learning” appears only twice [2]. Deep learning has never been mentioned in the New England Journal of Medicine, The Lancet, BMJ, or even JAMA, where the work on MYCIN was published 37 years ago. What happened? + +There are three central challenges that have plagued past efforts to use artificial intelligence in medicine: the label problem, the deployment problem, and fear around regulation. Before we get in to those, let’s take a quick look at the state of medicine today. + +The Moral Case for AI in Healthcare + +Medicine is life and death. With such high stakes, one could ask: should we really be rocking the boat here? Why not just stick with existing, proven, clinical-grade algorithms? + +Well, consider a few examples of the status quo: + +The most common prediction model for stroke risk was based on only 25 strokes. Source: Lip 2010 + +To be clear, none of this means medical researchers are doing a bad job. Modern medicine is a miracle; it’s just unevenly-distributed. Computer scientists can bring much to the table here. With tools like Apple’s ResearchKit and Google Fit, we can collect health data at scale +================================================================================ +Rank = 31; Score = 4554752.0 +<|begin_of_text|>Researchers at the University of Liverpool have developed a set of algorithms that will help teach computers to process and understand human languages. + +Whilst mastering natural language is easy for humans, it is something that computers have not yet been able to achieve. Humans understand language through a variety of ways for example this might be through looking up it in a dictionary, or by associating it with words in the same sentence in a meaningful way. + +The algorithms will enable a computer to act in much the same way as a human would when encountered with an unknown word. When the computer encounters a word it doesn’t recognise or understand, the algorithms mean it will look up the word in a dictionary (such as the WordNet), and tries to guess what other words should appear with this unknown word in the text. + +Semantic representation + +It gives the computer a semantic representation for a word that is both consistent with the dictionary as well as with the context in which it appears in the text. In order to know whether the algorithm has provided the computer with an accurate representation of a word it compares similarity scores produced using the word representations learnt by the computer algorithm against human rated similarities. + +Liverpool computer scientist, Dr Danushka Bollegala, said: “Learning accurate word representations is the first step towards teaching languages to computers.” + +“If we can represent the meaning for a word in a way a computer could understand, then the computer will be able to read texts on behalf of humans and perform potentially useful tasks such as translating a text written in a foreign language, summarising a lengthy article, or find similar other documents from the Internet. + +“We are excitingly waiting to see the immense possibilities that will be brought about when such accurate semantic representations are used in various language processing tasks by the computers.” + +The research was presented at the Association for Advancement of Artificial Intelligence Conference (AAAI-2016) held in Arizona, USA.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 32; Score = 4456448.0 +<|begin_of_text|>OWASP ModSecurity Core Rule Set (CRS) + +The 1st Line of Defense Against Web Application Attacks + +The OWASP ModSecurity Core Rule Set (CRS) is a set of generic attack detection rules for use with ModSecurity or compatible web application firewalls. The CRS aims to protect web applications from a wide range of attacks, including the OWASP Top Ten, with a minimum of false alerts. The CRS provides protection against many common attack categories, including SQL Injection, Cross Site Scripting, Locale File Inclusion, etc. The offical website of the project can be found at https://coreruleset.org. + +Getting Started / Tutorials + +The following tutorials will get you started with ModSecurity and the CRS v3. + +These tutorials are part of a big series of Apache / ModSecurity guides published by netnea. They are written by Christian Folini. + +More Information about the rule set is available at the official website, https://coreruleset.org. + +Licensing + +OWASP ModSecurity CRS is free to use. It is licensed under the Apache Software License version 2 (ASLv2), so you can copy, distribute and transmit the work, and you can adapt it, and use it commercially, but all provided that you attribute the work and if you alter, transform, or build upon this work, you may distribute the resulting work only under the same or similar license to this one. + +Reporting Issues + +If you think you've found a false positive in commercially available software and want us to take a look, submit an issue on our Github + +Have you found a false negative/bypass? We'd love to hear about it - please responsibly disclose it to security@coreruleset.org<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 33; Score = 4423680.0 +<|begin_of_text|>DeepMind + +When DeepMind burst into prominent view in 2014 it taught its machine learning systems how to play Atari games. The system could learn to defeat the games, and score higher than humans, but not remember how it had done so. + +DeepMind's AI has learnt to become 'highly aggressive' when it feels like it's going to lose DeepMind DeepMind's AI has learnt to become 'highly aggressive' when it feels like it's going to lose + +Advertisement + +For each of the Atari games, a separate neural network had to be created. The same system could not be used to play Space Invaders and Breakout without the information for both being given to the artificial intelligence at the same time. Now, a team of DeepMind and Imperial College London researchers have created an algorithm that allows its neural networks to learn, retain the information, and use it again. + +"Previously, we had a system that could learn to play any game, but it could only learn to play one game," James Kirkpatrick, a research scientist at DeepMind and the lead author of its new research paper, tells WIRED. "Here we are demonstrating a system that can learn to play several games one after the other". + +Read next DeepMind's Mustafa Suleyman: In 2018, AI will gain a moral compass DeepMind's Mustafa Suleyman: In 2018, AI will gain a moral compass + +The work, published in the Proceedings of the National Academy of Sciences journal, explains how DeepMind's AI can learn in sequences using supervised learning and reinforcement learning tests. This is also explained in a blog post from the company. + +Watch Google's Deep Mind complete Montezuma's Revenge Gaming Watch Google's Deep Mind complete Montezuma's Revenge + +Advertisement + +"The ability to learn tasks in succession without forgetting is a core component of biological and artificial intelligence," the computer scientists write in the paper. Kirkpatrick says a "significant shortcoming" in neural networks and artificial intelligence has been its inability to transfer what it has learned from one task to the next. + +The group says it has been able to show "continual learning" that's based on'synaptic consolidation'. In the human brain, the process is described as "the basis of learning and memory". + +The performance of DeepMind's EWC algorithm 'enhanced' neural network compared to its other nets DeepMind / PNAS + +Read next DeepMind's latest AI breakthrough is its most significant yet DeepMind's latest AI breakthrough is its most significant yet +================================================================================ +Rank = 34; Score = 4292608.0 +<|begin_of_text|>It’s been a while we haven’t covered any machine learning algorithm. Last time we discussed the Markov Decision Process (or MDP). + +Today we’re going to build our knowledge on top of the MDP and see how we can generalise our MDP to solve more complex problems. + +Reinforcement learning really hit the news back in 2013 when a computer learned how to play a bunch of old Atari games (like Breakout) just by observing the pixels on the screen. Let’s find out how this is possible! + +In Breakout the goal is to destroy all the bricks at the top of the screen by bouncing a ball over a paddle that you can move horizontally at the bottom of the screen. Every time the ball hits a brick your score increases. + +The idea + +If we want to teach a computer how to play that game, one solution might be to use a classifier such as a neural network where the input would be the screen images and the output the action to perform. It makes sense as for each screen we want to know if the paddle should move left or right (or stay still). The problem is that we need lots of training examples. But this is not how we learn! We don’t want someone to tell us a million time what we should do! + +Instead we learn by observing the reward we get as a consequence of our actions. This is the essence of reinforcement learning. An agent (the player) performs some actions which change the state of the environment (the pixels on the screen) and gets a reward for his actions (increase of score). + +Reinforcement learning lies somewhere between supervised and unsupervised learning. In supervised learning we have a label for every training example and in unsupervised learning there is no label at all. In reinforcement learning we have rewards which are sparse and time-delayed labels. Using only on those rewards the agent has to learn to behave in the environment. + +It may seem simple but there are some challenges along the way. When the ball hits a brick it has nothing to do with the action you just take. In fact it was the action performed when the ball bounced on the paddle that sent the ball against the brick. The problem of identifying which action is at the origin of the reward is known as the credit assignment problem. + +Markov Decision Process + +Doesn’t it sound familiar? Indeed, it looks quite similar to our tiny robot example we use to illustrate the MDP. + +The robot (the agent) moves (take actions) and gets a reward as it moves closer to +================================================================================ +Rank = 35; Score = 4227072.0 +<|begin_of_text|>Image copyright Reuters Image caption Chinese Go player Ke Jie has lost two games to AlphaGo + +Google's DeepMind AlphaGo artificial intelligence has defeated the world's number one Go player Ke Jie. + +AlphaGo secured the victory after winning the second game in a three-part match. + +DeepMind founder Demis Hassabis said Ke Jie had played "perfectly" and "pushed AlphaGo right to the limit". + +Following the defeat, Ke Jie told reporters: "I'm a little bit sad, it's a bit of a regret because I think I played pretty well." + +Image copyright Reuters Image caption Ke Jie eventually resigned + +In Go, players take turns placing stones on a 19-by-19 grid, competing to take control of the most territory. + +It is considered to be one of the world's most complex games, and is much more challenging for computers than chess. + +Tea-making + +AlphaGo has built up its expertise by studying older matches and playing thousands of games against itself. + +The company says the eventual plan is to deploy its artificial intelligence "in areas of medicine and science". + +Prof Noel Sharkey, a computer scientist at Sheffield University, said it is still a long way from creating a general intelligence. + +"It is an incredible achievement and most experts thought an AI winning at Go was 20 years away so DeepMind is leading the field but this AI doesn't have general intelligence. It doesn't know that is playing a game and it can't make you a cup of tea afterwards." + +Prof Nello Cristianini, from Bristol University, added: "This is machine learning in action and it proves that machines are very capable but it is not general intelligence. No-one has built that yet." + +The types of intelligence exhibited by machines that are good at playing games are seen as very narrow. While they may produce algorithms that are useful in other fields, few think they are close to the all-purpose problem solving abilities of humans that can come up with good solutions to almost any problem they encounter. + +Prof Cristianini added that while competition at a gaming level is fine, it should not govern how we view our relationship with intelligent machines going forward. + +"We should focus on the good things that we can get out of them and be careful not to create situations in which we put ourselves in direct competition with machines." + +Both experts agreed that such algorithms could be adapted to other fields, such as health care. + +DeepMind has already begun working with the UK's national health service to develop apps and other tools for diagnosis.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 36; Score = 4194304.0 +<|begin_of_text|>Let be a natural number. A basic operation in the topology of oriented, connected, compact, -dimensional manifolds (hereby referred to simply as manifolds for short) is that of connected sum: given two manifolds, the connected sum is formed by removing a small ball from each manifold and then gluing the boundary together (in the orientation-preserving manner). This gives another oriented, connected, compact manifold, and the exact nature of the balls removed and their gluing is not relevant for topological purposes (any two such procedures give homeomorphic manifolds). It is easy to see that this operation is associative and commutative up to homeomorphism, thus and, where we use to denote the assertion that is homeomorphic to. + +(It is important that the orientation is preserved; if, for instance,, and is a chiral 3-manifold which is chiral (thus, where is the orientation reversal of ), then the connect sum of with itself is also chiral (by the prime decomposition; in fact one does not even need the irreducibility hypothesis for this claim), but is not. A typical example of an irreducible chiral manifold is the complement of a trefoil knot. Thanks to Danny Calegari for this example.) + +The -dimensional sphere is an identity (up to homeomorphism) of connect sum: for any. A basic result in the subject is that the sphere is itself irreducible: + +Theorem 1 (Irreducibility of the sphere) If, then. + +For (curves), this theorem is trivial because the only connected -manifolds are homeomorphic to circles. For (surfaces), the theorem is also easy by considering the genus of. For the result follows from the prime decomposition. But for higher, these ad hoc methods no longer work. Nevertheless, there is an elegant proof of Theorem 1, due to Mazur, and known as Mazur’s swindle. The reason for this name should become clear when one sees the proof, which I reproduce below. + +Suppose. Now consider the infinite connected sum + +This is an infinite connected sum of spheres, and can thus be viewed as a half-open cylinder, which is topologically equivalent to a sphere with a small ball removed; alternatively, one can contract the boundary at infinity to a point to recover the sphere. On the other hand, by using the associativity of connected sum (which will still work for the infinite connected sum, if one thinks about it carefully +================================================================================ +Rank = 37; Score = 4194304.0 +<|begin_of_text|>Energy lab team explores new ways of analyzing social media + +A team at the Energy Department’s Pacific Northwest National Laboratory is creating an analysis tool that makes spotting trends in social media faster and easier. + +The tool, called SALSA, for SociAL Sensor Analytics, takes an automated approach that can sort through billions of tweets or other posts in seconds, identifying patterns and finding key bits of information that could be useful in emergency response, public health and other areas, according to PNNL. + +SALSA draws on the computing power of PNNL’s 600-node, 162-Teraflop Olympus supercomputing cluster, so the processing power is there. The challenge was finding a way to dig useful information out from all the banter on a platform that also has a non-traditional lexicon. + +“The task we all face is separating out the trivia, the useless information we all are blasted with every day, from the really good stuff that helps us live better lives,” Court Corley, a PNNL data scientist who’s leading the research, said in PNNL’s report. “There's a lot of noise, but there's some very valuable information too." + +Corley said the reach of social media creates value in analyzing posts to various platforms. "The world is equipped with human sensors — more than 7 billion and counting,” he said. “It's by far the most extensive sensor network on the planet. What can we learn by paying attention?" + +Agencies have employed systems to monitor social media, such as the U.S. Geological Survey’s Twitter Earthquake Detector (TED). The system was created in 2009 because Twitter often spreads the word about earthquakes and other disasters faster than traditional sensor methods. But one reason TED works is that USGS taps into Twitter’s API to search for a known keyword — “earthquake.” Taking a deeper dive into analyzing social media is more of a challenge. + +The Library of Congress has run into that difficulty with its Twitter archive. LOC has collected hundreds of billions of tweets for use in research, but found that available tools for sifting through the archive are inadequate. Searches take too much time (just searching the 20 billion tweets in its initial collection would take 24 hours), and as a result LOC was turning down requests from researchers to explore the archive. + +And the amount of social media-generated data is only going to grow. PNNL points out that, as of mid-2012, information posted to social media outlets each hour included an average of 30 million comments, +================================================================================ +Rank = 38; Score = 4079616.0 +<|begin_of_text|>InfluxDB : InfluxDB is an open-source time series database written in Go that has been built to work best with metrics, events, and analytics. Using InfluxDB, you can easily store system and application performance data and manage any time series data. + +Grafana : Grafana is an open-source, general purpose dashboard that is used for visualizing time series data for Internet infrastructure and application analytics. Grafana supports graphite, influxdb or opentsdb as backends and runs as a web application. + +In this tutorial, we will learn how to create and run Grafana and InfluxDB Docker containers in Ubuntu 14.04. + +Requirements + +A server running Ubuntu-14.04 with Docker installed. + +A non-root user with sudo privileges setup on server. + +Creating The Dockerfile + +First, you will need to create the Docker file to install all requisite software. Create docker file inside your home directory using the following command: + +sudo nano Dockerfile + +Add the following lines with all requisite software: + +FROM ubuntu MAINTAINER Hitesh Jethva (hitjethva@gmail.com) RUN apt-get update && apt-get -y --no-install-recommends install ca-certificates software-properties-common python-django-tagging python-simplejson python-memcache python-ldap python-cairo python-pysqlite2 python-support python-pip gunicorn supervisor nginx-light nodejs git curl openjdk-7-jre build-essential python-dev + +Add the following lines to install Grafana, InfluxDB, and do some basic configuration: + +WORKDIR /opt RUN curl -s -o /opt/grafana-1.8.1.tar.gz http://grafanarel.s3.amazonaws.com/grafana-1.8.1.tar.gz && curl -s -o /opt/influxdb_latest_amd64.deb http://s3.amazonaws.com/influxdb/influxdb_latest_amd64.deb && mkdir /opt/grafana && tar -xzvf grafana-1.8.1.tar.gz --directory /opt/grafana --strip-components=1 && dpkg -i influxdb_latest_amd64.deb && echo "influxdb soft nofile unlimited" >> /etc/security/limits.conf && echo "influxdb hard nofile unlimited" >> /etc/security/limits.conf + +Next, copy some configuration files: + +ADD config.js /opt/grafana/config.js ADD nginx.conf /etc/nginx/nginx.conf ADD supervisord.conf /etc/supervisor/conf.d/superv +================================================================================ +Rank = 39; Score = 4046848.0 +<|begin_of_text|>Find yourself writing a good bit of Markdown? Wish there was a better editor? Maybe in the new shiny and cool code editor, VS Code? + +Dave Johnson has written up a great tutorial on just that! + +Today we’re going to build an amazing Markdown editor using Visual Studio Code and Pandoc. This system will include real-time Markdown linting and the ability to generate html, docx, and pdf documents quickly with the potential to produce many other document formats as well. + +Markdown is a simple markup language that allows one to write documents using a text editor and transform those documents into many different formats. Among other things, it works beautifully for documenting source code since the Markdown documents can be checked in and versioned with Git or your source control system of choice. + +Pandoc is a highly capable “Swiss army knife” tool for converting documents between various formats. It is not limited to Markdown as an input (source) format, but it is used extensively in this context. + +Finally, Visual Studio Code is a solid, lightweight code editor created by Microsoft. It is based on the Electron framework, which facilitates the development of desktop GUI applications using the Node.js framework. I’m a huge fan of VS Code, and have written previous articles about it including the... + +Let’s get started! + +Install Visual Studio Code... + +Familiarize yourself with VS Code out-of-the-box Markdown features... + +Try Markdown previewer... + +Markdown Snippets... + +Install Markdown linter... + +Test drive Markdown linting... + +Install Pandoc Extension... + +Create configuration files to make a great Markdown editor... + +Create settings.json file... + +Create Markdown lint configuration file... + +Create CSS file to be used by HTML generated by Pandoc... + +Create an HTML Document with Pandoc... + +Conclusion + +You have now created an awesome Markdown editor/environment that you can use for all of your documentation needs! Pandoc provides a number of other conversion options that you can either add in the VS Code configuration or run from the command line to produce other document formats, create books, etc. You now have an amazing Markdown editor. Go have fun with it! + +Follow Dave Johnson @thisDaveJ on Twitter to stay up to date on the latest tutorials and tech articles. + +... [Click through to read the entire post]<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 40; Score = 3997696.0 +<|begin_of_text|>Warren Berger believes questions are more important than answers. A best-selling author and self-proclaimed “questionologist,” Warren is on a mission to help all leaders learn how to use questioning to support innovative, resilient and adaptive organizations. + +We caught up with Warren recently to ask him about why it’s more important than ever to lean into questioning and curiosity. + +Lisa Kay Solomon: Your recent book is called A More Beautiful Question. Can you talk about what a “beautiful question” is and why they are so important to ask? + +Warren Berger: I am particularly interested in questions that lead to innovation. So, I studied a lot of those kinds of questions to see what they had in common and arrived at this definition: A beautiful question is an ambitious yet actionable question that can shift the way we think about something and may serve as a catalyst for change. + +Each aspect is important. “Ambitious” because we have to ask bold questions to innovate. “Actionable” because big questions we can’t do anything about don’t lead to change. Critically, the question has to cause a mental shift—it makes you step back and say, “Hmmm, that’s interesting. I hadn’t really thought about this question before, and I want to explore it further.” + +These kinds of questions are the starting point of innovation and growth, which is why every leader should be asking them and encouraging others to ask them. + +LKS: How does questioning help us become better innovators? + +WB: Questioning helps us become more comfortable and proactive in dealing with the unknown, which is critical for innovators. Innovators often operate in a realm where the solution to a particular problem is unknown or may even seem impossible. + +The innovator’s job is to explore that unknown and arrive at a completely new and original answer. Questions can help the innovator keep moving forward; you start with one question, then that can lead you to a better, deeper question, and so on. You could almost think of questioning as “an app for the unknown.” + +When you finally arrive at the answer—the innovation—it may seem the questioning is over. But it isn’t really. Soon, the innovator is wondering how to make their creation even better, more affordable—how they can expand upon it. + +For true innovators, the cycle of questioning never ends. + +LKS: Are you seeing technology change the type of questions we should be asking? + +WB: Technology has thrown everything open to question because everything we thought we knew how to do can now be done differently. The shoe-ret +================================================================================ +Rank = 41; Score = 3915776.0 +<|begin_of_text|>Researchers at North Carolina State University have developed techniques that can be used to create ideal geometric phase holograms for any kind of optical pattern -- a significant advance over the limitations of previous techniques. The holograms can be used to create new types of displays, imaging systems, telecommunications technology and astronomical instruments. + +A geometric phase hologram is a thin film that manipulates light. Light moves as a wave, with peaks and troughs. When the light passes through a geometric phase hologram, the relationship between those peaks and troughs is changed. By controlling those changes, the hologram can focus, disperse, reorient or otherwise modify the light. + +An ideal geometric phase hologram modifies the light very efficiently, meaning that little of the light is wasted. But ideal geometric phase holograms can also produce three different, well-defined "wavefronts" -- or transformed versions of the light that passes through the thin film. + +"We can direct light into any one or more of those three wavefronts, which allows us to use a single ideal geometric phase hologram in many different ways," says Michael Escuti, a professor of electrical and computer engineering at NC State and corresponding author of a paper on the work. + +Previously, researchers were only able to make ideal geometric phase holograms in a limited set of simple patterns, curtailing their usefulness for new applications. This is because making these holograms involves orienting molecules or structures at a scale smaller than the wavelength of light. + +"We've come up with two ways of making ideal geometric phase holograms that are relatively simple but allow us to control the orientation of the molecules that ultimately manipulate the light," Escuti says. + +First, the researchers use lasers to create a high-fidelity light pattern, either by taking advantage of how waves of light interfere with each other, or by using a tightly focused laser to scan through a pattern -- much like a laser printer. + +A photoreactive substrate records the light pattern, with each molecule in the substrate orienting itself depending on the polarization of the light it was exposed to. To understand this, think of a beam of light as a wavy string, traveling from left to right. That string is also vibrating up and down -- creating wiggles are that are perpendicular to the direction the string is traveling. Controlling the orientation angle of the light's linear polarization just means controlling the direction that the wave is wiggling. + +The pattern that is recorded on the substrate then serves as a template for a liquid crystal layer that forms the finished hologram. + +"Using these techniques, +================================================================================ +Rank = 42; Score = 3883008.0 +<|begin_of_text|>$\begingroup$ + +maybe all the major/preferred algorithms of interest to this audience have been mentioned at this point. however, a few more deserve mention for completeness. & some analysis of what is considered a significant algorithm is relevant here. + +in CS & IT fields there seems to be a phenomenon noticed long ago in AI called "moving the goalposts". this is a psychological phenomenon where the field advances relatively quickly but people quickly mentally adjust to "the new normal" and take real or even breakthrough advances as mundane or unremarkable in retrospect, after accomplished, ie downplayed or minimized. this is highly captured in this question in the way that algorithms move from R&D into "deployment". quoting the author of the question in later comments: + +In fact, a negligible fraction of all the code that gets written is implementing anything that is interesting from an algorithmic point of view. + +but this is problematic and basically a TCS-centric redefinition of the word "algorithm". presumably the interesting algorithms are advanced. does that mean that if a problem is reduced to an advanced algorithm, its no longer "interesting"? and "advanced" is clearly a moving target. so there is a way to define "algorithms" narrowly, or broadly. it seems the TCS definition changes on context, but note even in TCS, there is a trend toward the broad definition eg in the so-called "algorithmic lens". + +sometimes the most ubiquitous algorithms are also the most overlooked! the internet and WWW is a large environment/near-ecology for algorithms. still relatively young at only about 2 decades old (invented ~1991) it has grown massively and exponentially in a short amount of time. WWW site growth has probably even outpaced the famous exponential Moores law. + +the internet/WWW are supported by many sophisticated algorithms. the internet has complex routing algorithms built into routers (again powering multi-billion-dollar corporations such as Cisco). some advanced theory is applicable there eg in routing algorithms. these algorithms were a subject of emerging, advanced/cutting edge research decades ago however a now so finetuned & well understood that its somewhat invisible. + +we should not so soon forget that decades ago, leading researchers were not even sure if the internet world work or was possible (seen in early packet switching research, a radical new design pattern at the time departing from the prior circuit-switching), and even a few years ago there were fears that it would fail to scale at some point and begin to fail due to overwhelming spikes in volume. + +it also uses sophisticated error detection/correction. +================================================================================ +Rank = 43; Score = 3768320.0 +<|begin_of_text|>Renting Dumpsters Posted on December 5, 2018 | By STEVE Whenever people have a task, it is easy to generate waste and piling up garbage can be a main source of worry. This is not only with regards to taking up space but also in conditions of polluting the environment unnecessarily. Therefore, how perform these worries are taken by you aside? This can be done by opting to use 20 yard dumpsters philadelphia simply. These are offered by local rental businesses and are designed to alleviate the worries of coping with trash from your brain. There are many of these rental companies in the marketplace and as such, it ought not to be difficult so that you can get the same. The local rental businesses will make sure that you recycle your waste in an eco-friendly method and at the same period, make the process easy. By using these eco-friendly dumpsters, you not really just ensure that your encircling is clean and gives you the ideal life-style but it will also make sure that you enjoy great wellness. As a result, this will give you great endurance and boost your probabilities of taking pleasure in a treatment free of charge way of living. These can be utilized in different fronts and places such as churches, house owners, areas and additional businesses. For this good reason, they are known to provide all round services to all these combined groups. There are instances when you might have dangerous waste products and this provides the ideal chance that you can ensure that you get rid of the same in a secure way. Dumpsters may end up being availed in various designs seeing that good as sizes and this makes it all easy for individuals to help to make buys based on their requirements. Another main benefit connected with this type of waste materials disposal is usually the truth that the consumer can place it in an area of their choice. This makes it simple to make sure that the clean up process is simple and effective on your part as well. It is normally essential to take note that the rental companies are also accountable for making sure that they are eliminated from your property once you are completed using the same and they can adhere to your time routine. What is more, in many situations, these ongoing companies will adhere to certain safety procedures when they are putting the same on site. At this true point, it is important to note that dumpsters that are provided by professional local rental businesses are also the perfect way to eliminate waste that is produced on building sites whether it is green backyard, commercial or home keep waste. What is usually even more, this provides you +================================================================================ +Rank = 44; Score = 3751936.0 +<|begin_of_text|>Image caption The NSA wants to use its quantum computer to break encryption used to protect online communication + +The US National Security Agency is building a quantum computer to break the encryption that keeps messages secure, reports the Washington Post. + +The NSA project came to light in documents passed to the newspaper by whistle-blower Edward Snowden. + +The spying agency hopes to harness the special qualities of quantum computers to speed up its code-cracking efforts. + +The NSA is believed to have spent about $80m (£49m) on the project but it has yet to produce a working machine. + +If the NSA managed to develop a working quantum computer it would be put to work breaking encryption systems used online and by foreign governments to keep official messages secure, suggest the documents excerpted in the Post. + +The quantum computer is being developed under a research programme called Penetrating Hard Targets and is believed to be conducted out of a lab in Maryland. + +Processing power + +Many research groups around the world are pursuing the goal of creating a working quantum computer but those developed so far have not been able to run the algorithms required to break contemporary encryption systems. + +Current computers attempt to crack encryption via many different means but they are limited to generating possible keys to unscramble data one at a time. Using big computers can speed this up but the huge numbers used as keys to lock away data limits the usefulness of this approach. + +By contrast, quantum computers exploit properties of matter that, under certain conditions, mean the machine can carry out lots and lots of calculations simultaneously. This makes it practical to try all the possible keys protecting a particular message or stream of data. + +The hard part of creating a working quantum computer is keeping enough of its constituent computational elements, called qubits, stable so they can interact and be put to useful work. + +The NSA is not believed to have made significant breakthroughs in its work that would put it ahead of research efforts elsewhere in the US and Europe. However, the documents passed to the Post by Edward Snowden suggest the agency's researchers are having some success developing the basic building blocks for the machine.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 45; Score = 3702784.0 +<|begin_of_text|>PYNQ is an open-source project that makes it easy for you to design embedded systems using the Xilinx Zynq-7000 SoC using the Python language, associated libraries, and the Jupyter Notebook, which is a pretty nice, collaborative learning and development environment for many programming languages including Python. PYNQ allows you to exploit the benefits of programmable logic used together with microprocessors to build more capable embedded systems with superior performance when performing embedded tasks such as: + +High frame-rate video processing + +Hardware-accelerated algorithms + +Real-time signal processing + +High-bandwidth I/O + +Low-latency control + +Nearly every embedded system needs to run one or more such tasks. The programmable hardware on the Zynq SoC just makes this job a lot easier. + +The PYNQ-Z1 based on the Xilinx Zynq Z-7020 SoC is the first dev board to support PYNQ and it just showed up on the Digilent Web site, listing for $229. (Digilent’s academic price for the PYNQ-Z1 is only $65!) + +Digilent PYNQ-Z1 Dev Board + +Here’s what’s on the PYNQ-Z1 board: + +Xilinx Zynq Z-7020 SoC with a dual-core ARM Cortex-A9 MPCore processor running at 650MHz + +512Mbytes of DDR3 SDRAM running at 1050MHz + +16Mbytes of Quad-SPI Flash memory + +A MicroSD slot for the PYNQ environment + +USB OTG host port + +USB programming port + +Gigabit Ethernet port + +Microphone + +Mono audio output jack + +HDMI input port + +HDMI output port + +Two Digilent PMOD ports + +Arduino-format header for Arduino shields + +Pushbuttons, switches, and LEDs + +That’s a lot of board for $229—and it’s pink! + +Here’s what PYNQ is, really: It’s for software developers and students who want to take advantage of the improved embedded performance made possible by the Zynq SoC’s programmable hardware without having to use ASIC-style (HDL) design tools to design hardware. + +For even better performance, you can also program the PYNQ-Z1 with C or C++ using the Xilinx SDK software development environment, available in the no-cost Xilinx Vivado HL Design Suite WebPACK. + +The PYNQ-Z1 and the PYNQ project make it possible to create a lot of really interesting systems, so just what are you waiting for? + + +================================================================================ +Rank = 46; Score = 3702784.0 +<|begin_of_text|>BY: Follow @LizWFB + +Researchers at the National Institutes of Health (NIH) have developed a system that can predict the "psychological status" of users with smartphones and hope to private companies to bring the invention to the market. + +The technology appeared on a list of NIH inventions published in the Federal Register that are now available to be licensed by private companies. The government allows companies to license inventions resulting from federal research in order to expedite their arrival on the marketplace. + +The system uses smartphones to ask people how they are doing mentally during the day and based on the results can "deliver an automated intervention" if necessary. + +"The NIH inventors have developed a mobile health technology to monitor and predict a user's psychological status and to deliver an automated intervention when needed," according to the notice published Wednesday. "The technology uses smartphones to monitor the user's location and ask questions about psychological status throughout the day." + +"Continuously collected ambulatory psychological data are fused with data on location and responses to questions," the NIH said. "The mobile data are combined with geospatial risk maps to quantify exposure to risk and predict a future psychological state. The future predictions are used to warn the user when he or she is at especially high risk of experiencing a negative event that might lead to an unwanted outcome (e.g., lapse to drug use in a recovering addict)." + +The NIH said the technology has potential commercial applications for "real-time behavior monitoring" and "therapeutic delivery of an intervention via a mobile device." + +Researchers developed the system from a project that tracked the mood and cravings of drug users in Baltimore. The $8.9 million federal study sought to develop algorithms that could "automatically detect behavioral events (such as episodes of drug use or stress) without requiring self-report." + +The NIH said the app is currently being used for drug addiction interventions, but that the "inventors are also seeking to test the technology for other health applications." + +Request for comment from the NIH was not returned by press time.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 47; Score = 3653632.0 +<|begin_of_text|>Inkscape just released version 0.91 of their Open Source vector graphics editor, and the new package will soon be available in the stable repositories for Fedora 21. Inkscape 0.91 is the first major release of Inkscape for a few years, and it has many bugfixes and new features compared to the previous Inkscape 0.48 release. + +For those not familiar with Inkscape, It is a versatile, feature rich vector graphics editor that can be used for a wide range of tasks, including UI mockups, icons, logos, digital illustration: + +Inkscape uses the Scalable Vector Graphics (SVG) format as the primary source filetype, which is getting more and more popular as a format for vector graphics on the web. Inkscape can also export to a wide range of different formats, including PNG and PDF. + +Inkscape 0.91 is a huge improvement over the previous 0.48 release, with the inkscape developers fixing hundreds of bugs, introducing a wide range of new tools and features, and also making Inkscape perform much, much better. + +Stability and Performance + +The Inkscape developers did a lot of work in this release to improve the performance and stability of Inkscape. I have been using development versions of Inkscape 0.91 for over a year now, and noticed a huge reduction in the number of times Inkscape stalled or crashed during use. + +Additionally, Inkscape 0.91 includes a new renderer based on Cairo, which is noticeably faster at rendering complex drawings with lots of SVG filters like blurs, and also caches parts of complex drawings to improve rendering times when editing. + +Inkscape now uses threading when rendering filter effects, so users with newer machines with multiple processing cores will notice a difference when rendering complex images. Inkscape also now uses up to 25% less memory on some complex drawings. + +New Features + +Inkscape 0.91 delivers a wide range of features that improve and simplify drawing and illustrating in Inkscape. + +New measurement tool + +0.91 introduces a new tool that allows users to measure the distances and angles between objects in an inkscape drawing by drawing a “ruler” over the objects. Using this tool is super simple, just draw a live over your objects, and the distances and angles will live-update on canvas as you move your mouse + +Font and Text Improvements + +The inkscape developers did a lot of work in the 0. +================================================================================ +Rank = 48; Score = 3555328.0 +<|begin_of_text|>Quality of life (QOL) is an overarching term for the quality of the various domains in life. It is a standard level that consists of the expectations of an individual or society for a good life. These expectations are guided by the values, goals and socio-cultural context in which an individual lives. It is a subjective, multidimensional concept that defines a standard level for emotional, physical, material and social well-being. It serves as a reference against which an individual or society can measure the different domains of one’s own life. The extent to which one's own life coincides with this desired standard level, put differently, the degree to which these domains give satisfaction and as such contribute to one's subjective well-being, is called life satisfaction. + +Overview [ edit ] + +Quality of life is the general well-being of individuals and societies, outlining negative and positive features of life. It observes life satisfaction, including everything from physical health, family, education, employment, wealth, safety, security to freedom, religious beliefs, and the environment.[1] QOL has a wide range of contexts, including the fields of international development, healthcare, politics and employment. It is important not to mix up the concept of QOL with a more recent growing area of health related QOL (HRQOL[2]). An assessment of HRQOL is effectively an evaluation of QOL and its relationship with health. + +Quality of life should not be confused with the concept of standard of living, which is based primarily on income. + +Standard indicators of the quality of life include not only wealth and employment but also the built environment, physical and mental health, education, recreation and leisure time, and social belonging.[3][4] According to the World Health Organization (WHO), quality of life is defined as “the individual’s perception of their position in life in the context of the culture and value systems in which they live and in relation to their goals.” In comparison to WHO's definitions, the Wong-Baker Faces Pain Rating Scale defines quality of life as “life quality (in this case, physical pain) at a precise moment in time.”[5] + +According to ecological economist Robert Costanza: + +While Quality of Life (QOL) has long been an explicit or implicit policy goal, adequate definition and measurement have been elusive. Diverse "objective" and "subjective" indicators across a range of disciplines and scales, and recent work on subjective well-being (SWB) surveys and the psychology of happiness have spurred renewed interest.[6] + +One approach, called engaged theory, outlined +================================================================================ +Rank = 49; Score = 3522560.0 +<|begin_of_text|>"Touch" redirects here. For other uses, see Touch (disambiguation) + +Widely distributed parts of the sensory nervous system + +The somatosensory system is a part of the sensory nervous system. The somatosensory system is a complex system of sensory neurons and pathways that responds to changes at the surface or inside the body. The axons (as afferent nerve fibers) of sensory neurons connect with, or respond to, various receptor cells. These sensory receptor cells are activated by different stimuli such as heat and nociception, giving a functional name to the responding sensory neuron, such as a thermoreceptor which carries information about temperature changes. Other types include mechanoreceptors, chemoreceptors, and nociceptors which send signals along a sensory nerve to the spinal cord where they may be processed by other sensory neurons and then relayed to the brain for further processing. Sensory receptors are found all over the body including the skin, epithelial tissues, muscles, bones and joints, internal organs, and the cardiovascular system. + +Touch is a crucial means of receiving information. This photo shows tactile markings identifying stairs for visually impaired people. + +Somatic senses are sometimes referred to as somesthetic senses,[1] with the understanding that somesthesis includes the sense of touch, proprioception (sense of position and movement), and (depending on usage) haptic perception.[2] + +The mapping of the body surfaces in the brain is called somatotopy. In the cortex, it is also referred to as the cortical homunculus. This brain-surface ("cortical") map is not immutable, however. Dramatic shifts can occur in response to stroke or injury. + +System overview [ edit ] + +This diagram linearly (unless otherwise mentioned) tracks the projections of all known structures that allow for touch to their relevant endpoints in the human brain. + +Mechanical [ edit ] + +The four mechanoreceptors in the skin each respond to different stimuli for short or long periods. + +Merkel cell nerve endings are found in the basal epidermis and hair follicles; they react to low vibrations (5–15 Hz) and deep static touch such as shapes and edges. Due to having a small receptive field (extremely detailed info), they are used in areas like fingertips the most; they are not covered (shelled) and thus respond to pressures over long periods. + +Tactile corpuscles react to moderate vibration (10–50 Hz) and light touch. They are located in the dermal papillae +================================================================================ +Rank = 50; Score = 3522560.0 +<|begin_of_text|>Counter heat current exchange: Note the gradually declining differential and that the once hot and cold streams exit at the reversed temperature difference; the hotter entering stream becomes the exiting cooler stream and vice versa. + +Countercurrent exchange is a mechanism occurring in nature and mimicked in industry and engineering, in which there is a crossover of some property, usually heat or some chemical, between two flowing bodies flowing in opposite directions to each other. The flowing bodies can be liquids, gases, or even solid powders, or any combination of those. For example, in a distillation column, the vapors bubble up through the downward flowing liquid while exchanging both heat and mass. + +The maximum amount of heat or mass transfer that can be obtained is higher with countercurrent than co-current (parallel) exchange because countercurrent maintains a slowly declining difference or gradient (usually temperature or concentration difference). In cocurrent exchange the initial gradient is higher but falls off quickly, leading to wasted potential. For example, in the adjacent diagram, the fluid being heated (exiting top) has a higher exiting temperature than the cooled fluid (exiting bottom) that was used for heating. With cocurrent or parallel exchange the heated and cooled fluids can only approach one another. The result is that countercurrent exchange can achieve a greater amount of heat or mass transfer than parallel under otherwise similar conditions. See: flow arrangement. + +Countercurrent exchange when set up in a circuit or loop can be used for building up concentrations, heat, or other properties of flowing liquids. Specifically when set up in a loop with a buffering liquid between the incoming and outgoing fluid running in a circuit, and with active transport pumps on the outgoing fluid's tubes, the system is called a countercurrent multiplier, enabling a multiplied effect of many small pumps to gradually build up a large concentration in the buffer liquid. + +Other countercurrent exchange circuits where the incoming and outgoing fluids touch each other are used for retaining a high concentration of a dissolved substance or for retaining heat, or for allowing the external buildup of the heat or concentration at one point in the system. + +Countercurrent exchange circuits or loops are found extensively in nature, specifically in biologic systems. In vertebrates, they are called a rete mirabile, originally the name of an organ in fish gills for absorbing oxygen from the water. It is mimicked in industrial systems. Countercurrent exchange is a key concept in chemical engineering thermodynamics and manufacturing processes, for example in extracting sucrose from sugar beet roots. + +Countercurrent multiplication is a similar +================================================================================ +Rank = 51; Score = 3506176.0 +<|begin_of_text|>But more important for the students, the various A-Lab projects served as a concrete learning experience on what data science is all about, - how to leverage messy, incomplete, real-world data to shed light on a complex and not-so-well-defined problem. + +A-Lab received over 20 project proposals from different companies, of which 13 were selected by the students. Each project team was assigned a research mentor to provide guidance as appropriate. I mentored a 3-student team that worked on a project sponsored by MasterCard. The students explored the possibility of improving on predictions of the economic performance of emerging markets by coupling existing economic indicators data with consumer behavior based on MasterCard’s transaction data. This is a particularly interesting project because economic data in emerging markets is often not as reliable as the data in more advanced markets. + +This past semester I was involved in an interesting course at MIT’s Sloan School of Management, - Analytics Labs (A-Lab). A-Lab’s objective is to teach students how to use data sets and analytics to address real-world business problems. Companies submit project proposals prior to the start of the class, including the business problem to be addressed and the data on which the project will be based. Students are then matched with the project they’re most interested in and grouped into teams of 3-4 students. + +A-Lab is taught by professors Erik Brynjolfsson and Sinan Aral. The course was first given in 2014, so this is only its second year. The 2015 Syllabus offers a good overview of the class, including the various companies that submitted project proposals. + +Projects are considered confidential unless the companies involved give permission to talk about them publicly, + +as several did in 2014. + +Amazon, for example, sponsored a project on how to raise the share of wallet of Amazon Prime customers, based on the analysis of over 200 million anonymized data points. And IBM sponsored a project to uncover a potential Watson application. The students recommended using Watson as a kind-of regulatory analyst assistance, to help financial institutions better understand how to comply with the over 1700 pages of regulations in the Dodd-Frank legislation. + +The 2015 A-Lab class culminated with short presentations of each of the student projects before an audience that included company sponsors and mentors. As I listened to the 13 presentations, I was impressed by the potential of applying big data and analytics to all kinds of business problems, from the tactical decisions every business makes as part of its normal operations to the strategic +================================================================================ +Rank = 52; Score = 3424256.0 +<|begin_of_text|>Guidelines for a better Web User Experience + +Although a user’s experience on a website is largely subjective, there are several things you can do to ensure that visitors return to your site. + +A great user experience can result in increased customer loyalty, higher rates of conversion, and a decrease in abandon rates and customer service requests. It improves the quality of a user’s interactions with your products and services, and it enhances their perception of your organization. A good user experience will ensure that users find and recognize the value in what they are being offered and provided. + +To ensure that visitors get the best experience possible, it’s necessary to understand their needs and values, and the abilities and limitations your site gives them as users. You can learn this by working closely with your clients to gain a better understanding of their business environment. When you understand your client’s business, you can come to know what type of users your website or app should cater to. + +Very responsive, delivered what was asked of them. Great communication. I am happy that I picked AllianceTek! They produce. - Jason Corbin (Co-Founder) + +strApp in Apps, LLC + +New York, NY + +Call Christopher Dabhi at 484-892-5713 now to schedule your free consultation with AllianceTek. + +The design of your website is the most important factor to a user. A solid design incorporates the principles of human-computer interaction (HCI) and facilitates easy and intuitive user navigation by employing the– your website or app should provide iconic representation of its facets and use white space efficiently for readability and a comfortable aesthetic.The design should be. In an era where new devices and operating systems seem to be released every day, a single site with multiple capabilities to function on desktops, tablets, mobile phones, and other web-accessible devices is sorely needed. The most popular framework for developing responsive websites is Bootstrap.is also key to a good user experience. For a user to feel engaged, your website or app should feature calls to action, search prompts, and easy navigation. Interactivity can be enhanced through audio or video media, hover states, scroll events, and sliding interfaces.It’s also important to keep user clicks to a. You should cut down on tasks or actions required by users to create a simplified and streamlined experience that lets them find the content they need without spending excess time clicking through other pages to get there.that you provide on your website is also very important. A web page should not be verbose. Think of each page as a paragraph in an essay, where each section +================================================================================ +Rank = 53; Score = 3358720.0 +<|begin_of_text|>network scanner + +Nmap (Network Mapper) is a free and open-source network scanner created by Gordon Lyon (also known by his pseudonym Fyodor Vaskovich).[3] Nmap is used to discover hosts and services on a computer network by sending packets and analyzing the responses. + +Nmap provides a number of features for probing computer networks, including host discovery and service and operating system detection. These features are extensible by scripts that provide more advanced service detection,[4] vulnerability detection,[4] and other features. Nmap can adapt to network conditions including latency and congestion during a scan. + +Nmap started as a Linux utility[5] and was ported to other systems including Windows, macOS, and BSD.[6] Linux is the most popular platform, followed by Windows.[7] + +Features [ edit ] + +Nmap features include: + +Host discovery – Identifying hosts on a network. For example, listing the hosts that respond to TCP and/or ICMP requests or have a particular port open. + +Port scanning – Enumerating the open ports on target hosts. + +Version detection – Interrogating network services on remote devices to determine application name and version number. [8] + +OS detection – Determining the operating system and hardware characteristics of network devices. + +Scriptable interaction with the target – using Nmap Scripting Engine[9] (NSE) and Lua programming language. + +Nmap can provide further information on targets, including reverse DNS names, device types, and MAC addresses.[10] + +Typical uses of Nmap: + +Auditing the security of a device or firewall by identifying the network connections which can be made to, or through it. [11] + +Identifying open ports on a target host in preparation for auditing. [12] + +Network inventory, network mapping, maintenance and asset management. + +Auditing the security of a network by identifying new servers. [13] + +Generating traffic to hosts on a network, response analysis and response time measurement. [14] + +Finding and exploiting vulnerabilities in a network. [15] + +DNS queries and subdomain search + +User interfaces [ edit ] + +NmapFE, originally written by Zach Smith, was Nmap's official GUI for Nmap versions 2.2 to 4.22.[16] For Nmap 4.50 (originally in the 4.22SOC development series) NmapFE was replaced with Zenmap, a new official graphical user interface based on UMIT, developed by Adriano Monteiro Marques. + +Various web-based interfaces allow controlling Nmap or analysing Nmap results from a +================================================================================ +Rank = 54; Score = 3358720.0 +<|begin_of_text|>JREF Swift Blog + +The Lurking Pornographer: Why Your Brain Turns Bubbles Into Nude Bodies + +There is a pornographer lurking in some corner of your mind. He peeks out from behind the curtains of your consciousness without warning, and almost never at an acceptable time. + +The lurking pornographer in your brain is ever vigilant, looking for patterns, for signs of nudity, and sometimes generating them out of nowhere. He is exceedingly good at what he does, and isn’t afraid to prove his power over your perception. Just like that, he can take a picture of Daniel Craig in a bathing suit and turn it obscene. + +If anything, Craig is more covered than he was before, but still he must be nude in the new picture, or so the pornographer would have you believe. The pornographer is sly. He takes advantage in the slightest slip in shapes and curves to insert his nudity. One of his favorite techniques is called “bubbling,” a technique that reveals how our brains actually “see.”. + +Breasts and Blind Spots + +Stifled by the pornography-restricting tenets of his religion, a young Mormon took to Photoshop, or so the story goes. His attempt to fool God and circumvent his law resulted in “bubbling,” a trick clever enough that how it works hasn’t yet been answered. + +You eye doesn’t see everything. Right now, there are innumerable photons hitting the photoreceptors all over your retina, except in the place where your optic nerve connects to it. This area is your blind spot, and it should show up as a rather large black dot in your vision, but it doesn’t. Why not? + +As your brain matures, it learns from the world. Neuronal connections are formed and broken in accordance with the deluge of information your brain receives. Over time, your brain becomes adept at predicting the world, so much so that much of our conscious lives are spent only noticing when things aren’t going as predicted. For example, there was probably a time when you got out of the car and realized you have almost no recollection of the drive you just took. It seemed automatic because it was. Consciousness didn’t need to intrude during something so routine, so it didn’t. However, introduce a near-collision into your daily commute, and consciousness quickly steps up to handle the situation. + +Based on all the shapes and colors and lines and lighting schemes that your brain has encountered, your cognition makes predictions about how things will look. The surprising +================================================================================ +Rank = 55; Score = 3342336.0 +<|begin_of_text|>Are we on the verge of creating artificial intelligence capable of finding answers to the world’s most pressing challenges? After steady progress in basic AI tasks in recent years, this is the vision that some leading technologists have for AI. And yet, how we will make this grand leap is anyone’s guess. + +Eric Schmidt, the executive chairman of Alphabet (formerly Google), says AI could be harnessed to help solve major challenges, including climate change and food security. Speaking at an event convened in New York this week to discuss the opportunities and risks in AI, Schmidt offered no details on how the technology might be adapted for such complex and abstract problems. + +Demis Hassabis, CEO of Google Deepmind, a division within Google doing groundbreaking work in machine learning, and which aims to bring about an “artificial general intelligence” (see “Google’s Intelligence Designer”), said the goal of this effort was to harness AI for grand challenges. “If we can solve intelligence in a general enough way, then we can apply it to all sorts of things to make the world a better place,” he said. + +And the chief technology officer of Facebook, Mike Schroepfer, expressed similar hope. “The power of AI technology is it can solve problems that scale to the whole planet,” he said. + +Eric Schmidt + +A steady stream of advances—mostly enabled by the latest machine-learning techniques—are indeed empowering computers to do ever more things, from recognizing the contents of images to holding short text or voice conversations. These advances seem destined to change the way computers are used in many industries, but it’s far from clear how the industry will go from captioning images to tackling poverty and climate change. + +In fact, speaking after his talk, Schroepfer was eager to limit expectations, at least in the short term. Schroepfer said that recent advances were not enough to allow machines to reach human levels of intelligence, and that two dozen or more “major breakthroughs” would be needed before this happened. And he said many people apparently had the wrong idea about how rapidly the field was moving. “People see one cool example, and then extrapolate from that,” he said. + +The event, organized by New York University as well as companies leading the effort to harness artificial intelligence, including Facebook and Google, comes at a delicate moment for academic researchers and companies riding the wave of progress in AI. Progress seems certain to revolutionize many industries, perhaps with negative consequences, such as eradicating certain jobs (see “Who Will Own the Robots?”). It will surely also raise +================================================================================ +Rank = 56; Score = 3342336.0 +<|begin_of_text|>Internet pornography is any pornography that is accessible over the Internet, primarily via websites, peer-to-peer file sharing, or Usenet newsgroups. The availability of widespread public access to the World Wide Web in late 1990s led to the growth of Internet pornography. + +A 2015 study finds "a big jump"[1] in pornography viewing over the past few decades, with the largest increase occurring between people born in the 1970s and those born in the 1980s. While the study's authors note this increase is "smaller than conventional wisdom might predict," it's still quite significant. Children born in the 1980s onward are also the first to grow up in a world where they have access to the Internet beginning in their teenage years, and this early exposure and access to Internet pornography may be the primary driver of the increase.[1] The sex and tech conference series Arse Elektronika dedicated their 2007 conference to what they call pr0nnovation. The con presented a keynote by culture theorist Mark Dery and published a reader about the subject. + +As of 2018, a single company, MindGeek, owns and operates many popular[2] pornographic websites, including video sharing services Pornhub, RedTube, and YouPorn, as well as adult film producers Brazzers, Digital Playground, Men.com, Reality Kings, and Sean Cody, among others.[3] It has been alleged to be a monopoly.[3] + +History and methods of distribution + +Before the World Wide Web + +Pornography is regarded by some as one of the driving forces behind the expansion of the World Wide Web, like the camcorder VCR and cable television before it.[4] Pornographic images had been transmitted over the Internet as ASCII porn but to send images over network needed computers with graphics capability and also higher network bandwidth. This was possible in the late 1980s and early 1990s through the use of anonymous FTP servers and through the Gopher protocol. At this time the internet was mainly an academic and military network and there was not widespread use of the internet. One of the early Gopher/FTP sites was at tudelft and was called the Digital Archive on the 17th Floor (List of websites founded before 1995). This small image archive contained some low quality scanned pornographic images that were initially available to anyone anonymously, but the site soon became restricted to Netherlands only access. + +Usenet groups + +Usenet newsgroups provided an early way of sharing +================================================================================ +Rank = 57; Score = 3342336.0 +<|begin_of_text|>A cinematographer is also known as a director of photography. They’re the guys that make the movies we watch look how they look. It’s their photographic eye that we see. And they don’t get too much recognition for the work they do, with most of the attention going towards the director and actor already. I wanted to write about a few good ones and see if it can become a weekly thing if you guys are into it. You probably know the work these guys have done, so I’ll cover what they did to get the shots that we see on the big screen. + +If this is going to be the first out of more to come, I’ll start it off with a bang by focusing it on Roger Deakins. + +Roger Deakins + +You’d probably have called him a purist; Roger Deakins’ preferred medium was film, until around 2011, when he stepped into the digital world when the Alexa cameras caught his eye. + +“I prefer Super 35 because it allows you to use short focal-length lenses. I also like the scale of that format — the intimacy — and the texture of the film grain. In some cases I find anamorphic to be almost too clean, too grain-free and pristine.” (AC Magazine, October 2007) + +Regardless of what he uses, he uses it to its fullest potential. If you’ve seen The Shawnshank Redemption, No Country for Old Men, The Assassination of Jesse James, Skyfall or even How to Train Your Dragon, you’ve seen his work. The guy’s got a resume, and it’s a wonder that none of his work has placed him an oscar yet. Here’s how a few of his most iconic shots were done: + +The Assassination of Jesse James by the Coward Robert Ford + +The Assassination of Jesse James opens with a train robbery. This scene was being filmed entirely at night, in a town called Edmonton. The train they had was supposedly a bit small (the director started calling it Thomas the Tank Engine), and they didn’t have the budget to bring a bigger train. Deakins assured the crew that with the right camera work, it wouldn’t be a big deal at all. After attempting to calm the crew down, Deakins gave the set another scare with his approach to how he’d film the scene: with just one source of light coming from a lamp set on the train. + +Considering this was being done with a night shot, everyone was rightfully pretty uneasy; what resulted, however, was possibly +================================================================================ +Rank = 58; Score = 3276800.0 +<|begin_of_text|>Russian president Vladimir Putin has joined the war of words concerning the international race to develop artificial intelligence. Speaking to students last Friday, Putin predicted that whichever country leads the way in AI research will come to dominate global affairs. + +“Artificial intelligence is the future, not only for Russia, but for all humankind,” said Putin, reports RT. “It comes with colossal opportunities, but also threats that are difficult to predict. Whoever becomes the leader in this sphere will become the ruler of the world.” + +“It comes with colossal opportunities, but also threats.” + +The development of artificial intelligence has increasingly become a national security concern in recent years. It is China and the US (not Russia) which are seen as the two frontrunners, with China recently announcing its ambition to become the global leader in AI research by 2030. Many analysts warn that America is in danger of falling behind, especially as the Trump administration prepares to cut funding for basic science and technology research. + +Although it’s thought that artificial intelligence will help boost countries’ economies in a number of areas, from heavy industry to medical research, AI technology will also be useful in warfare. Artificial intelligence can be used to develop cyber weapons, and control autonomous tools like drone swarms — fleets of low-cost quadcopters with a shared ‘brain’ that can be used for surveillance as well as attacking opponents. + +Both China and the US are currently researching this technology, and in his speech on Friday, Putin predicted that future wars would be fought by countries using drones. “When one party's drones are destroyed by drones of another, it will have no other choice but to surrender,” said the Russian president, according to the Associated Press. + +Recently, Elon Musk and 116 other technology leaders sent a petition to the United Nations calling for new regulations on how such AI weapons are developed. The group stated that the introduction of autonomous technology would be tantamount to a “third revolution in warfare,” following the development of gunpowder and nuclear weapons. + +An AI arms race doesn’t necessarily have to be a winner-takes-all scenario, though. Putin noted that Russia did not want to see any one country “monopolize” the field, and said instead: “If we become leaders in this area, we will share this know-how with entire world, the same way we share our nuclear technologies today.” + +Update September 4th, 5:40AM ET: Elon Musk offers his cheery perspective on Twitter:<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 59; Score = 3276800.0 +<|begin_of_text|>We think our destiny is to journey to Mars and beyond. Yet as we build our spacecraft, we're about to be broadsided - from a different direction - by the most explosive event in history. + +Sometime in the future science will be able to create realities that we can't even begin to imagine. As we evolve, we'll be able to construct other information systems that correspond to other realities, universes based on logic completely different from ours and not based on space and time. + +Immanuel Kant declared in 1781 that space and time were real, but only indeed as properties of the mind. These algorithms are not only the key to consciousness, but why space and time − indeed the properties of matter itself - are relative to the observer. But a new theory called biocentrism suggests that space and time may not be the only tools that can be used to construct reality. At present, our destiny is to live and die in the everyday world of up and down. But what if, for example, we changed the algorithms so that instead of time being linear, it was 3-dimensional like space? Consciousness would move through the multiverse. We'd be able to walk through time just like we walk through space. And after creeping along for 4 billion years, life would finally figure out how to escape from its corporeal cage. Our destiny would lie in realities that exist outside of the known physical universe. + +Even science fiction is struggling with the implications. In "Avatar," human consciousness is infused into blue aliens that inhabit a wondrous world. However, according to biocentrism, replicating human intelligence or consciousness will require the same kind of algorithms for employing time and space that we enjoy. Everything we experience is a whirl of information occurring in our heads. Time is simply the summation of spatial states - much like the frames in a film - occurring inside the mind. It's just our way of making sense of things. There's also a peculiar intangibility to space. We can't pick it up and bring it to the laboratory. Like time, space isn't an external object. It's part of the mental software that molds information into multidimensional objects. + +We take for granted how our mind puts everything together. When I woke up this morning, I was in the middle of a dream that seemed as real as everyday life. I remember looking out over a crowded port with people in the foreground. Further out, there were ships engaged in battle. And still further out to sea was a battleship with +================================================================================ +Rank = 60; Score = 3276800.0 +<|begin_of_text|>Eigenfaces is the name given to a set of eigenvectors when they are used in the computer vision problem of human face recognition.[1] The approach of using eigenfaces for recognition was developed by Sirovich and Kirby (1987) and used by Matthew Turk and Alex Pentland in face classification.[2] The eigenvectors are derived from the covariance matrix of the probability distribution over the high-dimensional vector space of face images. The eigenfaces themselves form a basis set of all images used to construct the covariance matrix. This produces dimension reduction by allowing the smaller set of basis images to represent the original training images. Classification can be achieved by comparing how faces are represented by the basis set. + +History [ edit ] + +The eigenface approach began with a search for a low-dimensional representation of face images. Sirovich and Kirby (1987) showed that principal component analysis could be used on a collection of face images to form a set of basis features. These basis images, known as eigenpictures, could be linearly combined to reconstruct images in the original training set. If the training set consists of M images, principal component analysis could form a basis set of N images, where N < M. The reconstruction error is reduced by increasing the number of eigenpictures, however the number needed is always chosen less than M. For example, if you need to generate a number of N eigenfaces for a training set of M face images, you can say that each face image can be made up of "proportions" of all this K "features" or eigenfaces : Face image 1 = (23% of E 1 ) + (2% of E 2 ) + (51% of E 3 ) +... + (1% E n ). + +In 1991 M. Turk and A. Pentland expanded these results and presented the eigenface method of face recognition.[3] In addition to designing a system for automated face recognition using eigenfaces, they showed a way of calculating the eigenvectors of a covariance matrix in such a way as to make it possible for computers at that time to perform eigen-decomposition on a large number of face images. Face images usually occupy a high-dimensional space and conventional principal component analysis was intractable on such data sets. Turk and Pentland's paper demonstrated ways to extract the eigenvectors based on matrices sized by the number of images rather than the number of pixels. + +Once established, the eigenface method was expanded to include methods of preprocessing to improve accuracy.[4 +================================================================================ +Rank = 61; Score = 3260416.0 +<|begin_of_text|>Fetch is the native AJAX API to replace jQuery.get() + +As a term AJAX has been around for over a decade now. While many still relate to it as rich and fluid interfaces, the real deal is the possibility of making asynchronous requests to the server from the browser. + +As browsers evolved the XMLHttpRequest API used for AJAX did not not evolve. This is why many developers still rely on including the whole jQuery library just to abstract asynchronous HTTP Request. jQuery has a great and simple to use API for many things, but including it just for AJAX is overkill. + +Finally the Web API has a new native alternative available. This is known as the Fetch API and offers a very simple way of getting content from the server with JavaScript in the browser. The API is a high level one that has a generic definition of HTTP Request and Response objects. + +To anyone used to working with the HTTP, the concepts are immediately familiar: + +GlobalFetch : Contains the fetch() method used to fetch a resource. + +: Contains the fetch() method used to fetch a resource. Headers : Represents response/request headers, allowing you to query them and take different actions depending on the results. + +: Represents response/request headers, allowing you to query them and take different actions depending on the results. Request : Represents a resource request. + +: Represents a resource request. Response: Represents the response to a request. + +In addition to standard features developers would expect, the Fetch API definition handles closely related concepts such as CORS (Cross-origin resource sharing) and HTTP origin headers. The fetch resouce is available in the browser's standard window object and has only a single mandatory argument - the request URL. + +The fetch method returns a JavaScript promise that makes it convenient to work with the asynchronous nature of HTTP requests in the browser context. Once the response is received, there are a number of methods available in the object. + +Browser support for the Fetch API is very good with evergreen browsers such as Chrome, Edge, Firefox and Opera already supporting the feature. Internet Explorer does not support the API, but there is a fetch API polyfill that can be used for it.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 62; Score = 3260416.0 +<|begin_of_text|>Document Clustering + +Contents + +Document clustering is the act of collecting similar documents into bins, where similarity is some function on a document. The clustering algorithms implemented for LEMUR are described in "A Comparison of Document Clustering Techniques", Michael Steinbach, George Karypis and Vipin Kumar. TextMining Workshop. KDD. 2000. With the exception of Probabilistic Latent Semantic Analysis (PLSA), all use cosine similarity in the vector space model as their metric. + +The LEMUR clustering support provides two principle APIs, the Cluster API, which defines the clusters themselves, and the ClusterDB API, which defines how Clusters are persistently stored. Similarity is scored via a SimilarityMethod object. Currently there is a single SimilarityMethod, CosSim, defined. + +Cluster + +Performs the basic online clustering task. In conjunction with an incremental indexer (such as KeyfileIncIndex), it could be used for the TDT topic detection task. It iterates over the documents in the index, assigning each document that is not in any cluster to a cluster. The document id, cluster id, and score are printed to the standard output. The parameters accepted by Cluster are: + +index -- the index to use. Default is none. + +clusterIndex -- the name of the cluster database index to use. Default is none. + +clusterdb_type -- One of flatfile (simple cluster database) or keyfile (btree based). + +clusterType -- Type of cluster to use, either agglomerative or centroid. Centroid is agglomerative using mean which trades memory use for speed of clustering. Default is centroid. + +simType -- The similarity metric to use. Default is cosine similarity (COS), which is the only implemented method. + +docMode -- The scoring method to use for the agglomerative cluster type. The default is max (maximum). The choices are: max -- Maximum score over documents in a cluster. mean -- Mean score over documents in a cluster. This is identical to the centroid cluster type. avg -- Average score over documents in a cluster. min -- Minimum score over documents in a cluster. + +threshold -- Minimum score for adding a document to an existing cluster. Default is 0.25. + +The example application that demonstrates the basic offline clustering task. Provides k-means and bisecting k-means partitional clustering. It will run each algorithm on the first 100 documents in the index (or all of them if less than 100) and print out the results. The parameters accepted by +================================================================================ +Rank = 63; Score = 3211264.0 +<|begin_of_text|>-Fixed a bug that caused small dogs to start barking and never stop. + +-Small dogs are no longer supported, and will self-terminate within two months. + +-St. Bernard size increased 100%. + +-St. Bernard's cask will no longer contain brandy. Now contains one of four dipping sauces (ranch dressing, sweet and sour, honey mustard and barbecue). + +-Improved saliva production in certain breeds by as much as 300%. + +-Removed a rare glitch that would cause a dog's head to rotate 360 degrees when looking inquisitively at operator. + +-Dogs now regard vacuum cleaners as friends and will attempt to operate them. + +-Fixed a bug that caused dogs to incorrectly regard fecal matter as food. + +-Fixed an error that caused social interaction between dogs to occur on the wrong end. + +-Fixed an issue where some breeds would grow curly hair, making them appear effeminate and prissy. + +-Improved dog pathfinding, allowing them to better navigate to operators located at higher altitudes or behind complicated obstacles. + +-Dramatically reduced instances of autoerotic behavior in public. + +-Addressed a logic error that caused some dogs to hysterically chomp at bees. + +-Fixed a rare glitch where a dog would regard its own tail as an enemy. + +-Removed an exploit that would allow operators to duplicate their dogs endlessly. + +-Placing a small ball or toy in a dog's open mouth while it is yawning will no longer cause it to shut down. + +-Reduced bloodhound "droopiness" by 25%. + +-Fixed issues of dogs turning hostile when spoken to in a funny or scary voice. + +-Introduced fixes for three common benign situations that would cause a dog to become sexually aroused. + +-Fixed error of dog incompatibility with chocolate cake. Should greatly improve experience of being dog. + +-Fixed a bug that would cause floppy ears to flip over the wrong way, requiring an operator to reset them to their default position. + +-Dogs should no longer drag butts across carpeting in presence of operators. + +-Fixed rendering bug that caused deformed geometry on pug faces. + +-Investigated issue of pit bulls entering rings of screaming men, biting each other to death; confirmed cause as operator error. + +-Fixed vocalization issue in huskies which would cause them to utter "I love you" when trying to say "stop tormenting me." + +-Added a routine to flush loyalty cache of dogs upon death of operators, preventing documented issues of dogs waiting at train stations for 15 years. + +-Fixed a +================================================================================ +Rank = 64; Score = 3211264.0 +<|begin_of_text|>DURHAM, N.C. – In experiments reproducing the natural environment, Duke University researchers have demonstrated that silver nanoparticles, which are used in many consumer products, can have an adverse effect on plants and microorganisms. + +These preliminary findings are important, the researchers said, because little is known about the environmental effects of these nanoparticles, and the only studies conducted to date involve high concentrations of the nanoparticles in a laboratory setting, which they point out, doesn’t represent “real-world” conditions. + +Silver nanoparticles are used in a host of products, most commonly in textiles and clothing. These nanoparticles are used because one of their characteristics is the ability to kill bacteria, inhibiting unwanted odors. They work through different mechanisms, including generating oxygen free radicals which can cause DNA damage to microbial membranes without harming human cells. Other products with silver nanoparticles are children’s toys and pacifiers, disinfectants and toothpaste. + +The main route by which these particles enter the environment is as a by-product of water and sewage treatment plants. The nanoparticles are too small to be filtered out, so they and other materials end up in the resulting “sludge,” which is then spread on the land surface as a fertilizer. + +“Results from laboratory studies are difficult to extrapolate to ecosystems, where exposures will likely be at low concentrations and which are inhabited by a diversity of organisms,” said Benjamin Colman, a post-doctoral fellow in Duke’s biology department and a member of the Center for the Environmental Implications of Nanotechnology (CEINT), which is funded by the National Science Foundation and the Environmental Protection Agency. + +“No one really knows what the effects of these particles are in the environment,” Colman said. “We’re trying to come up with the data that can be used to help regulators determine the risks to the environment from silver nanoparticle exposures.” + +The results of the CEINT team’s experiments were published Feb. 27 in the online journal PLOS ONE. + +“Our field studies show adverse responses of plants and microorganisms in a replicated long-term terrestrial environment following a single low dose of silver nanoparticles, applied by the likely route of exposure, sewage biosolid application,” Colman said. “An estimated 60 percent of the average 5.6 million tons of biosolids produced each year is applied to the land for various reasons, and this practice represents an important and understudied route of exposure of natural ecosystems to engineered nanoparticles.” + +For their studies, the CEINT researchers created mesocosms, which are small, man-made structures containing different plants and microorganisms meant +================================================================================ +Rank = 65; Score = 3145728.0 +<|begin_of_text|>We live in a world of pain and suffering. There is no one who is not affected by the harsh realities of life, and the question “why do bad things happen to good people?” is one of the most difficult questions in all of theology. God is sovereign, so all that happens must have at least been allowed by Him, if not directly caused by Him. At the outset, we must acknowledge that human beings, who are not eternal, infinite, or omniscient, cannot expect to fully understand God’s purposes and ways.The book of Job deals with the issue of why God allows bad things to happen to good people. Job was a righteous man (Job 1:1), yet he suffered in ways that are almost beyond belief. God allowed Satan to do everything he wanted to Job except kill him, and Satan did his worst. What was Job’s reaction? “Though he slay me, yet will I hope in him” (Job 13:15). “The LORD gave and the LORD has taken away; may the name of the LORD be praised” (Job 1:21). Job did not understand why God had allowed the things He did, but he knew God was good and therefore continued to trust in Him. Ultimately, that should be our reaction as well.Why do bad things happen to good people? As hard as it is to acknowledge, we must remember that there are no “good” people, in the absolute sense of the word. All of us are tainted by and infected with sin (Ecclesiastes 7:20; Romans 3:23; 1 John 1:8). As Jesus said, “No one is good—except God alone” (Luke 18:19). All of us feel the effects of sin in one way or another. Sometimes it’s our own personal sin; other times, it’s the sins of others. We live in a fallen world, and we experience the effects of the fall. One of those effects is injustice and seemingly senseless suffering.When wondering why God would allow bad things to happen to good people, it’s also good to consider these four things about the bad things that happen:1) Bad things may happen to good people in this world, but this world is not the end. Christians have an eternal perspective: “We do not lose heart. Though outwardly we are wasting away, yet inwardly we are being renewed day by day. For our light and momentary troubles are achieving for us an eternal glory that far +================================================================================ +Rank = 66; Score = 3080192.0 +<|begin_of_text|>ADVERTISEMENT + +Earth's climate is changing rapidly. We know this from billions of observations, documented in thousands of journal papers and texts, and summarized every few years by the United Nations' Intergovernmental Panel on Climate Change. The primary cause of that change is the release of carbon dioxide from burning coal, oil, and natural gas. + +Negotiations about reducing emissions grind on. But in the meantime, how much warming are we already locked into? If we stop emitting greenhouse gases tomorrow, why would the temperature continue to rise? + +The basics of carbon and climate + +The carbon dioxide that accumulates in the atmosphere insulates the surface of the Earth. It's like a warming blanket that holds in heat. This energy increases the Earth's surface average temperature, heats the oceans, and melts polar ice. As consequences, sea level rises and weather changes. + +The Arctic is warming much faster than the average global temperature; ice in the Arctic Ocean is melting and the permafrost is thawing. Ice sheets in both the Arctic and Antarctic are melting. Ecosystems on both land and in the sea are changing. The observed changes are coherent and consistent with our theoretical understanding of the Earth's energy balance and simulations from models that are used to understand past variability and to help us think about the future. + +Slam on the climate brakes + +What would happen to the climate if we were to stop emitting carbon dioxide today, right now? Would we return to the climate of our elders? The simple answer is no. Once we release the carbon dioxide stored in the fossil fuels we burn, it accumulates in and moves amongst the atmosphere, the oceans, the land, and the plants and animals of the biosphere. The released carbon dioxide will remain in the atmosphere for thousands of years. Only after many millennia will it return to rocks, for example, through the formation of calcium carbonate — limestone — as marine organisms' shells settle to the bottom of the ocean. But on time spans relevant to humans, once released the carbon dioxide is in our environment essentially forever. It does not go away, unless we, ourselves, remove it. + +If we stop emitting today, it's not the end of the story for global warming. There's a delay in temperature increase as the climate catches up with all the carbon that's in the atmosphere. After maybe 40 more years, the climate will stabilize at a temperature higher than what was normal for previous generations. + +This decades-long lag between cause and effect is due to the long time it takes to heat the the ocean's huge mass +================================================================================ +Rank = 67; Score = 3063808.0 +<|begin_of_text|>For a broader coverage of forks, see Fork (blockchain) + +Bitcoin forks are defined variantly as changes in the protocol of the bitcoin network or as the situations that occur "when two or more blocks have the same block height".[1] A fork influences the validity of the rules. Forks are typically conducted in order to add new features to a blockchain, to reverse the effects of hacking or catastrophic bugs. Forks require consensus to be resolved or else a permanent split emerges. + +Forks of the client software + +The following are forks of the software client for the bitcoin network: + +All three software clients attempt to increase transaction capacity of the network. None achieved a majority of the hash power.[2] + +Intended hard forks splitting the cryptocurrency + +Hard forks splitting bitcoin (aka "split coins") are created via changes of the blockchain rules and sharing a transaction history with bitcoin up to a certain time and date. The first hard fork splitting bitcoin happened on 1 August 2017, resulting in the creation of Bitcoin Cash. + +The following is a list of hard forks splitting bitcoin by date and/or block: + +Bitcoin Cash: Forked at block 478558, 1 August 2017, for each bitcoin (BTC), an owner got 1 Bitcoin Cash (BCH) + +Bitcoin Gold: Forked at block 491407, 24 October 2017, for each BTC, an owner got 1 Bitcoin Gold (BTG) + +Bitcoin SV: Forked at block 556766, 15 November 2018, for each Bitcoin Cash (BCH), an owner got 1 Bitcoin SV (BSV). + +Intended soft forks splitting from not-most-work block + +The fork fixing the value overflow incident was controversial because it was announced after the exploit was mined. + +Unintended hard forks + +Two hard forks were created by "protocol change" definition: + +March 2013 Chain Fork (migration from BerkeleyDB to LevelDB caused a chain split) [3] + +CVE-2018-17144 (Bitcoin 0.15 allowed double spending certain inputs in the same block. Not exploited)<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 68; Score = 3063808.0 +<|begin_of_text|>I spent half an hour speaking with IBM Watson VP John Gordon on Tuesday, and no matter how many ways I asked it, he would not acknowledge a gap between peoples’ perceptions of Watson and what the “cognitive computing” system is actually capable of doing. I know there’s misunderstanding out there — I just know it — but Gordon spun his responses to focus on inspiration rather than disappointment, about how easy it is to learn Watson and build new applications now that it’s available as a set of software products and cloud APIs. + +It annoyed me at first, but the more I think about it, the less I can fault his strategy. It wasn’t so long ago, he noted, that it was still only Ph.Ds. in IBM Research programming Watson systems for early users; today, pretty much anybody with an application and some data can start experimenting with it. There’s not a lot of point in dwelling on whether or not people really get the artificial intelligence, because anyone willing to give the cloud services a shot will soon figure out all they really need to know about it — that it works or doesn’t work for what they want to do. + +Advertisement + +I think that’s becoming true too for deep learning, the red-hot AI field that, like Watson, is also the subject of lofty claims and more than a modicum of hyperbole. IEEE Spectrum recently published an interview with machine learning expert Michael Jordan, of the University of California, Berkeley, in which he addresses some of the bigger misconceptions about the technology. + +His comments boil down to this: Deep learning is not that revolutionary, it’s only really useful for a limited number of things and deep learning models certainly do not mimic real-life brain activity. The latter is a lazy and inaccurate metaphor. + +(Jordan later clarified some misinterpretations of his opinions — particularly regarding big data, which is a topic for a whole other day — in a blog post. Earlier, in September, he did an Ask Me Anything session on Reddit where he also elaborates on where he sees promise and hype in deep learning.) + +But even if Jordan is largely correct, it might be a waste of energy for most of us to stress too much over explanations of how exactly the models work or whether researchers are overstating the import of their results. Yann LeCun, a New York University researcher and director of Facebook’s AI efforts, probably said it best in a (largely supportive) response to Jordan’s comments via Facebook wall post: + +There is nothing wrong with deep learning as a topic of investigation, and +================================================================================ +Rank = 69; Score = 2998272.0 +<|begin_of_text|>Genetically modified animals are animals that have been genetically modified for a variety of purposes including producing drugs, enhancing yields, increase resistance to disease, etc. The vast majority of genetically modified animals are at the research stage with the number close to entering the market remains small.[1] + +Production [ edit ] + +The process of genetically engineering mammals is a slow, tedious, and expensive process.[2] As with other genetically modified organisms (GMOs), first genetic engineers must isolated the gene they wish to insert into the host organism. This can be taken from a cell containing the gene[3] or artificially synthesised.[4] If the chosen gene or the donor organism's genome has been well studied it may already be accessible from a genetic library. The gene is then combined with other genetic elements, including a promoter and terminator region and usually a selectable marker.[5] + +A number of techniques are available for inserting the isolated gene into the host genome. With animals DNA is generally inserted into using microinjection, where it can be injected through the cell's nuclear envelope directly into the nucleus, or through the use of viral vectors.[6] The first transgenic animals were produced by injecting viral DNA into embryos and then implanting the embryos in females.[7] It is necessary to ensure that the inserted DNA is present in the embryonic stem cells.[8] The embryo would develop and it would be hoped that some of the genetic material would be incorporated into the reproductive cells. Then researchers would have to wait until the animal reached breeding age and then offspring would be screened for presence of the gene in every cell, using PCR, Southern hybridization, and DNA sequencing.[9] + +New technologies are making genetic modifications easier and more precise.[2] Gene targeting techniques, which creates double-stranded breaks and takes advantage on the cells natural homologous recombination repair systems, have been developed to target insertion to exact locations. Genome editing uses artificially engineered nucleases that create breaks at specific points. There are four families of engineered nucleases: meganucleases,[10][11] zinc finger nucleases,[12][13] transcription activator-like effector nucleases (TALENs),[14][15] and the Cas9-guideRNA system (adapted from CRISPR).[16][17] TALEN and CRISPR are the two most commonly used and each has its own advantages.[18] TALENs have greater target specificity, while CRISPR is easier to design and more efficient.[18] The development of the CRISPR-C +================================================================================ +Rank = 70; Score = 2981888.0 +<|begin_of_text|>On a sweltering Monday in late June 2015, the city council in Charlotte, North Carolina, met to discuss, among other items in a seven-hour marathon, how to carry out a controversial new approach to predicting police misconduct. Opinions were divided, and the discussion was tense. One council member was afraid of “upsetting the troops.” A second called the use of data about individual police officers an invasion of privacy. In response, another said, “I’m always a fan of third parties looking over our shoulder.” + +Finally, Kerr Putney, soon to be sworn in as Charlotte’s new police chief, got up to reassure the council. He spoke about the need to “balance public need versus what officers may want.” He seemed to persuade several members. + +“So it won’t be used for retribution?” one asked. “Absolutely not,” Putney replied. + +Minutes later, the council voted to work with a group of data scientists to develop a sophisticated system for predicting when cops will go bad. These researchers, part of the White House’s Police Data Initiative, say their algorithm can foresee adverse interactions between officers and civilians, ranging from impolite traffic stops to fatal shootings. Their system can suggest preventive measures — an appealing prospect for police departments facing greater scrutiny and calls for accountability. Two other large departments — the Los Angeles County sheriff and the Knoxville police — have signed on to use the research to develop new systems, and several other agencies have expressed interest. The scientists hope their method can serve as a template for stopping police misbehavior before it happens. + +Many police departments have early warning systems — software that tracks each officer’s performance and aims to forecast potential problems. The systems identify officers with troubling patterns of behavior, allowing superiors to monitor these cops more closely or intervene and send them to counseling. + +The researchers, a mixed group of graduate and undergraduate students working together at the University of Chicago with backgrounds in statistics, programming, economics and related disciplines, are trying to build a better kind of early warning system. They began their task last summer with a request from the Charlotte-Mecklenburg Police Department: Predict when police officers would participate in adverse interactions with civilians. + +To build their early warning system, the University of Chicago group first looked for signals in the data that an officer might be going astray. They used a comprehensive data set of interactions between cops and the public gathered by Charlotte police officials over more than a decade. The researchers found that the most potent predictor of adverse interactions in a given year was an officer’s own history. C +================================================================================ +Rank = 71; Score = 2981888.0 +<|begin_of_text|>Meditation involves silencing our mind. It is not an intellectual activity, but an attempt to expand our consciousness and be aware of our ‘real sense of being.’ Meditation can be a practical tool for relaxation, concentration and better health; it can also be an invaluable tool to self discovery. Through meditation we develop the capacity to be content with ourselves as we are. It is a happiness not dependent on external wealth and success. If practiced correctly, meditation can be a powerful antidote to depressive thoughts. + +How To Meditate + +It is hard to pick up meditation from just reading an article, but I would like to share a few basic pointers about what meditation involves. No matter what form of meditation you follow, the basic shared principle is to quieten your thoughts and mind. We can sit in a chair for many hours, but, if thoughts continually pass through our mind then our meditation will be ineffective. Ultimately the aim is to have a mind free of thoughts. It is in this inner silence that we can experience a consciousness of real peace. + +At first glance, people may find the concept of stopping thoughts very difficult. If you try sitting silent for a while, you will probably be inundated with thoughts. When giving meditation classes, the difficulty of controlling the thoughts is a common experience. However, if you sincerely try, you can learn to reduce the power of thoughts over yourself. + +These are some tips I suggest for controlling your thoughts: + +You control your thoughts not the other way around. Always remember it is you who can decide which thoughts to pursue and which to reject. Never feel a slave to your own thoughts, even if at times they seem powerful. + +Patience. Don’t expect a silent mind after the first few attempts. We have been thinking all our life; to change a habit of a lifetime requires persistence and perseverance. Meditation, like any worthwhile activities requires dedicated and focused intensity. + +Detachment. If you keep rejecting thoughts, what happens is that you may be aware of thoughts, but, they have much less intensity. A thought bubbles up, but, it becomes easier to detach from it. You start to see thoughts as independent and outside of yourself. This is a good sign, it shows you are developing the capacity to separate the sense of self from your mental thoughts. From this point it becomes easier to stop your thoughts completely. + +Concentrate on something. It is hard to control our thoughts through the power of the mind. To achieve inner silence it is advisable to choose something to focus our attention on. This may be the +================================================================================ +Rank = 72; Score = 2981888.0 +<|begin_of_text|>What do the San Francisco Giants, Cryptolocker and nuclear war all have in common? They all involve conflicts in which incentives, payouts and winning strategies can be analyzed with game theory. Game theory is a branch of mathematics that models conflict and cooperation between parties and is used in many real-world decision making scenarios, inside and outside the Information Security field. Game theory is particularly useful in analyzing the extortionist / victim dynamic present in ransomware infection scenarios. + +Ransomware comes in many varieties and works in different ways, but the basic setting is the same: cybercriminals infect a computer with malicious software that blocks access to the system or important files until the ransom is paid. + +The conventional wisdom in information security regarding ransomware is to never pay. But, why? The answer is a little more nuanced than “never pay” or “always pay.” The decision is a complex scenario of incentives and payoffs. Who stands to gain when ransomware is paid? Who gains when it is not paid? + +This talk will use the familiar topic of ransomware to introduce participants to game theory concepts like rational decision-making, zero-sum games, incentives, utility and Nash Equilibrium – all important tools that can help solve security problems. By analyzing ransomware decision-making with a game theory mindset, participants will learn a new set of skills and a new way of incentive-driven thinking.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 73; Score = 2965504.0 +<|begin_of_text|>“There will come a time when it isn't ‘They’re spying on me through my phone’ anymore. Eventually, it will be ‘My phone is spying on me.’” ― Philip K. Dick + +If ever Americans sell their birthright, it will be for the promise of expediency and comfort delivered by way of blazingly fast Internet, cell phone signals that never drop a call, thermostats that keep us at the perfect temperature without our having to raise a finger, and entertainment that can be simultaneously streamed to our TVs, tablets and cell phones. + +Likewise, if ever we find ourselves in bondage, we will have only ourselves to blame for having forged the chains through our own lassitude, laziness and abject reliance on internet-connected gadgets and gizmos that render us wholly irrelevant. + +Indeed, while most of us are consumed with our selfies and trying to keep up with what our so-called friends are posting on Facebook, the megacorporation Google has been busily partnering with the National Security Agency (NSA), the Pentagon, and other governmental agencies to develop a new “human” species, so to speak. + +In other words, Google—a neural network that approximates a global brain—is fusing with the human mind in a phenomenon that is called “singularity,” and they’ve hired transhumanist scientist Ray Kurzweil to do just that. Google will know the answer to your question before you have asked it, Kurzweil said. “ It will have read every email you will ever have written, every document, every idle thought you’ve ever tapped into a search-engine box. It will know you better than your intimate partner does. Better, perhaps, than even yourself.” + +But here’s the catch: the NSA and all other government agencies will also know you better than yourself. As William Binney, one of the highest-level whistleblowers to ever emerge from the NSA said, “ The ultimate goal of the NSA is total population control.” + +Science fiction, thus, has become fact. + +We’re fast approaching Philip K. Dick’s vision of the future as depicted in the film Minority Report. There, police agencies apprehend criminals before they can commit a crime, driverless cars populate the highways, and a person’s biometrics are constantly scanned and used to track their movements, target them for advertising, and keep them under perpetual surveillance. + +Cue the dawning of the Age of the Internet of Things, in which internet-connected “things” will monitor your home, your health and your habits +================================================================================ +Rank = 74; Score = 2965504.0 +<|begin_of_text|>Decorators modify functions. Beginning with the basics, learn how to use decorators in a variety of ways. Execute code when a function is parsed or called. Conditionally call functions and transform inputs and outputs. Write customizable decorators that accept arbitrary arguments. And, if necessary, easily make sure your decorated function has the same signature as the original. + +Decorators. The shear mention of them brings fear to even the seasoned Python programmer. + +Okay, maybe not. But decorators, at least in this author's opinion, have a weird syntax and inevitably complicated implementations that are especially foreign (I think) to many who have gotten used to the simplicity of Python. + +1 The Basics Decorators modify functions. More specifically, a decorator is a function that transforms another function. When you use a decorator, Python passes the decorated function -- we'll call this the target function -- to the decorator function, and replaces it with the result. Without decorators, it would look something like this: #'s 1 def decorator_function ( target ): + +2 # Do something with the target function + +3 target. attribute = 1 + +4 return target + +5 + +6 def target ( a, b ): + +7 return a + b + +8 + +9 # This is what the decorator actually does + +10 target = decorator_function ( target ) + +This code has the exact same functionality, but uses decorators. Note that I can name my decorator function whatever I want. Here, I've chosen 'decorator_function': #'s 1 def decorator_function ( target ): + +2 # Do something with the target function + +3 target. attribute = 1 + +4 return target + +5 + +6 # Here is the decorator, with the syntax '@function_name' + +7 @decorator_function + +8 def target ( a, b ): + +9 return a + b + +As you can see, you need to put the decorator function's name, prefaced with a @, on the line before the target function definition. Python internally will transform the target by applying the decorator to it and replacing it with the returned value. Both of the above examples will have the same results: #'s 1 >>> target ( 1, 2 ) + +2 3 + +3 >>> target. attribute + +4 1 + +1.1 Does a decorator function have to return a function? No. The decorator function can return absolutely anything, and Python will replace the target function with that return value. For example, you could do something like this: #'s 1 def decorator_evil ( target ): + +2 return False +================================================================================ +Rank = 75; Score = 2965504.0 +<|begin_of_text|>This post is by Instacart VP Data Science Jeremy Stanley, and technical advisor and former LinkedIn data leader Daniel Tunkelang. Previously, Jeremy wrote the most comprehensive manual we’ve ever seen for hiring data scientists. + +It's hard to believe that "data scientist" only became a bona fide job title in 2008. Jeff Hammerbacher at Facebook and DJ Patil at LinkedIn coined the term to capture the emerging need for interdisciplinary skills across analytics, engineering, and product. Today, the demand for data scientists has blossomed, and with it the need to better understand how to grow these teams for success. + +The two of us have seen our share of the good, the bad, and the ugly, leading and advising teams at a variety of companies in different industries and at different stages of maturity. We've seen the challenges of not only hiring top data scientists, but making effective use of them and retaining them in a hyper competitive market for talent. + +In this article, we've summarized the advice we give to founders who are interested in building data science teams. We explain why data science is so important for many startups, when companies should begin investing in it, where to put data science in their organization and how to build a culture where data science thrives. + +First, what are you trying to achieve? + +Data science serves two important but distinct sets of goals: improving the products your customers use, and improving the decisions your business makes. + +Data products use data science and engineering to improve product performance, typically in the form of better search results, recommendations and automated decisions. + +Decision science uses data to analyze business metrics — such as growth, engagement, profitability drivers, and user feedback — to inform strategy and key business decisions. + +This distinction may sound straightforward, but it’s an important one to keep in mind as you establish and grow your data science team. Let’s take a closer look at these two areas. + +Using Data Science to Build Better Products + +Data products leverage data science to improve product performance. They rely on a virtuous cycle where products collect usage data that becomes the fodder for algorithms which in turn offer users a better experience. + +What happens before you’ve collected that data? The first version of your product has to address what data science calls the “cold start” problem — it has to provide a "good enough" experience to initiate the virtuous cycle of data collection and data-driven improvement. It’s up to product managers and engineers to implement that good enough solution. + +For example, when an Instacart user visits the site, the application +================================================================================ +Rank = 76; Score = 2932736.0 +<|begin_of_text|>Nerve cells that self-select to become sensory organ precursors (SOPs) are identified by arrows in this microscope image. The cells send chemical signals to neighboring cells, blocking them from becoming SOPs and causing them to fluoresce red. + +The burgeoning neural networks of fruit fly pupae solve a distributed computing problem, arranging sensory bristles in a very efficient, effective manner. Scientists who monitored the bristles' growth say they can mimic the flies' method to build more effective communications networks. + +It's not the first time we've seen an insect solve a problem that plagues computer scientists — bees can do it, too — but the fruit fly discovery does one better, leading to an algorithm that can be used to develop more efficient computer and wireless networks. + +Distributed computing involves several processors working in concert to solve a problem. Some are chosen as leaders, collecting data from the other processors and passing it along. Organizing these networks into efficient processor-leader groups is one of the biggest challenges in computing — but millions of cells in a fly's nervous system do it automatically, organizing themselves so that a small number of cells serve as leaders. It is much better than anything humans have come up with, scientists say: "It is such a simple and intuitive solution, I can't believe we did not think of this 25 years ago," according to co-author Noga Alon, a mathematician and computer scientist at Tel Aviv University and the Institute for Advanced Study in Princeton. + +Fruit fly bristles, which are used for feeling and hearing, develop as nerve cells self-select to become leaders. The cells send chemical signals to their neighboring cells, ensuring that those cells cannot become leaders, too. Using fluorescence microscopy, the researchers watched an entire network form in about three hours. + +They developed an algorithm based on the cells' self-selection approach, and say it's particularly effective for adaptive networks where the number and position of each node is not certain, according to Carnegie Mellon University. That could include environmental monitoring sensors, robot swarms and more. + +The research is published today in the journal Science. + +Eurekalert<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 77; Score = 2932736.0 +<|begin_of_text|>In the 28 years since Super Mario Bros. was released it's obviously been comprehensiely beaten, thoroughly, many thousands of times in that time by players around the world. + +But have you ever made the game beat itself? + +Advertisement + +That's what computer scientist Tom Murphy has done. At SigBovik 2013 he presented a program that "solves" how to play Super Mario Bros., or any other NES game, like it's just another kind of mathematical problem. And for those who know that SigBovik is an annual computer science conference dedicated to spoof research, hosted on 1 April every year, Murphy stresses that this is "100 percent real". + +He outlines his method in a paper,"The First Level of Super Mario Bros. is Easy with Lexicographic Orderings and Time Travel... after that it gets a little tricky", but he also presented the results in the video you can see with this story. + +Read next Super Smash Bros. Ultimate review: the best Smash ever Super Smash Bros. Ultimate review: the best Smash ever + +Lexicographic ordering is a pretty simple mathematical technique used to determine the best order a set of values should come in. + +It's most commonly used in libraries or dictionaries for arranging books and words, for instance, with the alphabet determining the order of the letters. + +Advertisement + +The NES puts out 60 frames of 2048 bytes per second, and each of these was fed into LearnFun. Everything in the NES's memory -- the buttons being pressed, the number of lives left, the score, the locations of enemies, Mario's position as coordinates, and so on -- is taken in by the LearnFun algorithm. + +PlayFun then plays the game, and uses the knowledge from LearnFun to try and increase the values it knows it has to increase, which mainly means Mario's score and how far scrolled the right Mario is in the level. "It's trying to find the sequence of inputs to make those values go up in the RAM," Murphy explains in the video. + +The results are impressive. After some tweaking, Mario plays the first level just like a real person, jumping on enemies like Goombas and hitting boxes for coins. The program even learns how to take advantage of bugs and glitches, like timing jumps so that Mario begins falling again at the exact time that he makes contact with a Goomba. Mario's invincible when he's falling, so the touch kills the Goomba, not Mario, and it gives him a further jump boost. + +Read next These +================================================================================ +Rank = 78; Score = 2932736.0 +<|begin_of_text|>Amazon Lex, the technology powering Amazon’s virtual assistant Alexa, has exited preview, according to a report from Reuters this morning. The system, which involves natural language understanding technology combined with automatic speech recognition, was first introduced in November, at Amazon’s AWS re:Invent conference in Las Vegas. + +At the time, Amazon explained how Lex can be used by developers who want to build their own conversational applications, like chatbots. + +As an example, the company had demoed a tool that allowed users to book a flight using only their voice. + +However, the system is not limited to working only in the chatbots you find in today’s consumer messaging apps, like Facebook Messenger (though it can be integrated with that platform). Lex can actually work in any voice or text chatbot on mobile, web or in other chat services beyond Messenger, including Slack and Twilio SMS. + +Amazon has suggested it could be used for a variety of purposes, including web and mobile applications where the technology provides users with information, powers their application, helps with various work activities, or even provides a control mechanism for robots, drones and toys. + +Chatbots in messaging – and particularly in e-commerce bots – is a solid entry point for Lex’s technology, though. Consumers today have been frustrated by the current crop of chatbots that have clunky menus to navigate through, and a limited ability to respond to questions users asked. Lex, on the other hand, would allow developers to create bots that convert speech to text and those that could recognize the intent of the text, making the resulting bot more conversational, and more sophisticated than those on the market at present. + +Lex, as a fully managed Amazon service, would also scale automatically as the bots’ usage increased, meaning developers would only pay for the number of text or voice queries that Lex processes. + +Amazon’s goal with opening up Lex to the wider development community could give it an edge in its ability to compete with other voice technology, like Google’s Assistant or Apple’s Siri, for example. The company plans to take the text and recordings that people send to Lex-powered apps in order to improve Lex, and its ability to understand more queries, notes today’s report. + +This openness has been Amazon’s larger strategy with much of its Alexa platform. For example, it already had rolled out Alexa Voice Services which allowed developers to integrate Alexa into their own devices, like speakers, bedside alarm clocks, and more. + +Alexa’s software isn’t the only area where Amazon is embracing an open ecosystem. The company earlier this month said it would +================================================================================ +Rank = 79; Score = 2916352.0 +<|begin_of_text|>"MOND" redirects here. For other uses, see Mond + +Modified Newtonian dynamics (MOND) is a theory that proposes a modification of Newton's laws to account for observed properties of galaxies. It is an alternative to the theory of dark matter in terms of explaining why galaxies do not appear to obey the currently understood laws of physics. + +Created in 1982 and first published in 1983 by Israeli physicist Mordehai Milgrom,[1] the theory's original motivation was to explain that the velocities of stars in galaxies were observed to be larger than expected based on Newtonian mechanics. Milgrom noted that this discrepancy could be resolved if the gravitational force experienced by a star in the outer regions of a galaxy was proportional to the square of its centripetal acceleration (as opposed to the centripetal acceleration itself, as in Newton's second law), or alternatively if gravitational force came to vary inversely with radius (as opposed to the inverse square of the radius, as in Newton's law of gravity). In MOND, violation of Newton's laws occurs at extremely small accelerations, characteristic of galaxies yet far below anything typically encountered in the Solar System or on Earth. + +Unsolved problem in physics: + +What is the nature of dark matter? Is it a particle, or do the phenomena attributed to dark matter actually require a modification of the laws of gravity? (more unsolved problems in physics) + +MOND is an example of a class of theories known as modified gravity, and is an alternative to the hypothesis that the dynamics of galaxies are determined by massive, invisible dark matter halos. Since Milgrom's original proposal, MOND has successfully predicted a variety of galactic phenomena that are difficult to understand from a dark matter perspective.[2][3] However, MOND and its generalisations do not adequately account for observed properties of galaxy clusters, and no satisfactory cosmological model has been constructed from the theory. + +The accurate measurement of the speed of gravitational waves compared to the speed of light in 2017 ruled out many theories which used modified gravity to explain dark matter.[4] However, both Milgrom’s bi-metric formulation of MOND and nonlocal MOND are not ruled out according to the same study. + +Overview [ edit ] + +[5] Comparison of the observed and expected rotation curves of the typical spiral galaxy M33 + +Several independent observations point to the fact that the visible mass in galaxies and galaxy clusters is insufficient to account for their dynamics, when analysed using Newton's laws. This discrepancy – known as +================================================================================ +Rank = 80; Score = 2916352.0 +<|begin_of_text|>While police departments may believe that this technology could be “the wave of the future,” and some research has shown that it substantially reduces crime, it may also threaten civil liberties. Nor does it address the underlying problems of crime-ridden areas–it punishes those neighborhoods rather than building community trust. These critiques of predictive policing reflect the ongoing debate around how data and algorithms should be used in our society, and how humans transfer their biases onto ostensibly unbiased software. But how do you illustrate the hazards of these opaque tools? + +One way to do so is to build your own. + +A new project published by The New Inquiry inverts common assumptions about the efficacy of predictive policing, pointing out how the logic that underlies these data-driven models may be inherently flawed. The interactive, called White Collar Crime Risk Zones, uses similar tactics to the predictive policing models that are used by many cities today–but focuses on white collar crime. Zoom in on New York City, and the maps shows that the highest risk areas for white collar crime are in the financial district and Midtown in Manhattan. Zoom out, and you begin to see criminal hot spots in the wealthy Connecticut towns of Greenwich and Stamford, home to many of the nation’s most prestigious hedge funds. “Unlike typical predictive policing apps, which criminalize poverty, White Collar Crime Risk Zones criminalizes wealth,” the team writes on The New Inquiry‘s website. + +Intended as a critique of predictive policing, the map was created by NYU professor and New Inquiry editor Sam Lavigne, Buzzfeed data scientist Bryan Clifton, and New Inquiry co-publisher and New Inc. researcher Francis Tseng. After researching the methods used by companies that build predictive policing tools like HunchLab (used by Miami, St. Louis, and New York City) and PredPol (used by Chicago and Los Angeles), both of which focus on predicting street crimes like burglary and assault, the three began to build their own version for white collar crime. + +“It was important for us to develop our predictive policing application in the way that they do theirs because we believe that makes our critique stronger,” Tseng says. + +Using data from the Financial Industry Regulatory Authority, an independent regulatory body that keeps records of when companies break the rules and are forced to pay a fine (even if they don’t go to court), Lavigne, Clifton, and Tseng were able to build a map of where white collar crimes occurred over the past 50 years. Then, they searched for publicly available data sets that overlapped closely with these white collar crime +================================================================================ +Rank = 81; Score = 2899968.0 +<|begin_of_text|>Originally called “connected TVs,” and now they are called as “smart TVs”. Any television that can be connected to the Internet to access services, use apps and behave in some way as our computers with web browser. Smart TVs connect to Internet via wired Ethernet connection or Wi-Fi to connect to a home network. Smart TVs require computer chips to juggle video processing, multiple screens and an Internet connection. They also use memory to buffer streaming video and music, and need additional processing power to deal with graphics. The TVs can be controlled by voice commands and by apps running on some Smartphone. + +Dan Reynolds, information security solution and training expert of International Institute of cyber security explains that these Smart TVs are not that smart and the security of software isn’t exactly perfect. Smart TVs resemble for us the Internet of things (IoT) but old vulnerabilities which were considered to have completely disappeared are new vulnerabilities again in the Internet of Things (IoT). Sometimes you can easily find a flaw that can enable you to take a variety of actions on the TV, including accessing potentially sensitive data, remote files and information, the drive image and eventually gain root access to the device. + +In the article we will be covering different aspects of two most famous brands of Smart TVs Samsung and LG with the help of ethical hacking course professor of IIcybersecurity. + +Understanding SAMSUNG SMART TV Operating system + +Tizen is an operating system based on the Linux kernel and the GNU C Library implementing the Linux API. It targets a very wide range of devices including smart phones, tablets, in-vehicle infotainment (IVI) devices, smart TVs, PCs, smart cameras, wearable computing, Blu-ray players, printers and smart home appliances. Its purpose is to offer a consistent user experience across devices. Tizen would be implemented in Samsung TVs from 2015. + +There are some online community which are working over the Samsung smart TV OS research like ( Sammygo) mentions Dan Reynolds, information security solution and training expert. + +How to do analysis over Samsung Smart TV firmware + +ExLink connector consist of a cable which has in one side a 3.5mm jack, like the audio ones, and on the other side an RS232 ( Serial ) DB9 connector. This cable will allow you to connect your PC computer to the TV, and enter in the Serial mode. With this you can use a serial Communications Software, like Hyperterminal, Putty from Windows or Linux. + +Connecting to Samsung TV + +Put the TV into Standby Mode, press [Info] then [Menu] +================================================================================ +Rank = 82; Score = 2883584.0 +<|begin_of_text|>Campus-based technology courses can be expensive, which is why many people are choosing to take free technology courses online. Some of the best universities in the world offer free technology courses. Examples include Stanford University, the Massachusetts Institute of Technology and the Tokyo Institute of Technology. + +1. The Massachusetts Institute of Technology + +The Massachusetts Institute of Technology OpenCourseWare site is undoubtedly the best place to find free technology courses online. There are hundreds of courses to choose from, which means you can find a course on nearly any technology topic imaginable. + +2. Delft University of Technology + +The Delft University of Technology offers a number of free technology courses. Most of Delft's courses cover advanced topics that would be of interest to students in engineering or related fields. New courses are constantly being added to the site. + +3. University of California, Berkeley + +If you're looking for free technology courses that you can download to your MP3 player or computer, you'll definitely want to check out UC Berkeley on iTunes. Self-learners can gain access to a wide range of technology courses, lectures and events. Additional courses can be found though webcast.berkeley.edu. + +4. Stanford University + +Berkeley isn't the only university that makes free technology courses available through iTunes. Stanford University has a similar set-up through their Stanford iTunes project. There are quite a few technology courses and lectures that can be accessed for free already, and more are added on a weekly basis. + +5. Rice University + +Connexions, a Rice University organization, hosts a wide variety of scholarly content. The site offers hundreds of free technology courses. Some courses consist of small 'knowledge chunks.' Other courses include multiple modules, books and course notes. + +6. The Open University + +There are more than 50 free technology courses that can be taken through The Open University's online LearningSpace. Courses range from the introductory level to the advanced level and take anywhere from two hours to two weeks to complete. + +7. Utah State University + +Self-learners around the world can choose from more than a dozen free technology courses when they visit Utah State University's OpenCourseWare site. Technology topics include, but are not limited to: online communities, instructional technology, interactive learning and computer engineering. + +8. University of Southern Queensland + +The University of Southern Queensland offers two different courses that may be of interest to technology buffs. Both of the text-based courses are free and include materials that are practical for self-learners. + +9. Dixie State College of Utah + +The CIT Department at the +================================================================================ +Rank = 83; Score = 2883584.0 +<|begin_of_text|>💪 From the big boys + +Backchannel run a rare piece on how Apple uses machine learning. It states that a 200mb software package runs on the iPhone encompassing “app usage data, interactions with contacts, neural net processing, a speech modeler and a natural language event modeling system”. I’ve held the view for a while now that today’s AI techniques and infrastructure will re-open a class of historically intractable problems while also enabling us to rethink how products and features should be designed. Apple seem to think the same: “Machine learning is enabling us to say yes to some things that in past years we would have said no to. It’s becoming embedded in the process of deciding the products we’re going to do next.” + +, modestly called + +Salesforce announced their internal umbrella AI initiative, modestly called Einstein, which will go on to power many of the company’s cloud services, as well as expose AI tools to end users. The team of 175 data scientists includes talent from acquired startups MetaMind, PredictionIO and RelateIQ. The company’s flagship event, Dreamforce, will attract 170k people into SF next week. + +Six of the most powerful technology companies have set up the have set up the Partnership on AI as a non-profit aimed at advancing public understanding of AI and formulate best practices on the challenges and opportunities within the field. An important catalyst to this end will undoubtedly be the continuation of open source technology development, which Seldon’s founder articulates in this piece. + +🌎 On the importance and impact of AI on the World + +Stanford’s 100 year study on AI published their first report. It finds “no cause for concern that AI is an imminent threat to humankind. No machines with self-sustaining long-term goals and intent have been developed, nor are they likely to be developed in the near future”. From a public policy perspective, it recommends to: + +Define a path toward accruing technical expertise in AI at all levels of government. + +Remove the perceived and actual impediments to research on the fairness, security, privacy, and social impacts of AI systems. + +Increase public and private funding for interdisciplinary studies of the societal impacts of AI. + +a16z’s Chris Dixon sets out 11 reasons to be excited about the future of technology with short soundbites for each. Five of these are either directly related to or will be enabled by AI and machine learning. + +👍 User-friendly AI + +UC Berkeley announced a new Center for Human-Compatible AI to study how AI used for mission +================================================================================ +Rank = 84; Score = 2883584.0 +<|begin_of_text|>It may sound obvious, but the primary role of a data journalist is to analyze data. Whether the analysis is simple or complex, it is our capacity to do this well that makes us valuable in the newsroom. And yet, most of us are still doing it using old-fashioned, error-prone methods. Specifically, we use processes that are: + +inscrutable (R, numpy); + +difficult to replicate (Excel, Google Docs, OpenRefine); and + +wedded to tools that are complex or expensive ( SPSS, SAS, ArcGIS). + +As journalists, we not only need to solve these problems for practical reporting purposes, but also for philosophical ones. How can we assert that our numbers are correct if we performed a series of manual processes in a spreadsheet exactly once? Do it that way and the only record of how it was done is the one in your head. That’s not good enough. Journalistic integrity requires that we’re able to document and explain our processes. + +Meet agate + +For the last year or so, I’ve challenged myself to design a better way of doing routine data analysis. It’s a problem with several parts, and today I’m thrilled to announce that the first and largest piece of the solution has reached version 1.0. It’s called agate, and it’s going to make your process better. I’ve been using agate in production for two months, and I will personally guarantee that it works. + +In greater depth, agate is a Python data analysis library in the vein of numpy or pandas, but with one crucial difference. Whereas those libraries optimize for the needs of scientists—namely, being incredibly fast when working with vast numerical datasets—agate instead optimizes for the performance of the human who is using it. That means stripping out those technical optimizations and instead focusing on designing code that is easy to learn, readable, and flexible enough to handle any weird data you throw at it. + +(Love csvkit? agate is all the guts of csvkit, converted to a Python library and amped up a hundred times. Sorry, Ruby programmers—maybe you can steal some of the ideas for your own projects!) + +Does focusing on human performance mean agate is slow? No: computers are very fast. Except in cases where the amount of data is truly huge (scientific research, financial systems), the optimizations that make these libraries complex, such as writing large parts of them in C, are unnecessary. They also make the libraries less flexible and more difficult to use. agate does away +================================================================================ +Rank = 85; Score = 2850816.0 +<|begin_of_text|>By Tanja Babic, PhD + +In the modern world, success of corporations is often driven by their employees and teams. Understanding how human behaviors affect the workplace is the main objective of industrial and organizational psychology, also known as I/O psychology. I/O psychologists study factors that promote motivation, team work and productivity, as well as management practices that improve employee’s performance. There are many industrial and organizational psychologists working towards improving working conditions, however, the following 30 people were chosen based on the following criteria: + +1. Publications: Majority of the individuals on this list have outstanding publication records. These publications include articles in scientific journals, books and book chapters. + +2. Impact on industrial and organizational practices: Although academic research is an important aspect of psychology, its influence can only be assessed once ideas are put into practice and applied in a workplace. Men and women on this list have made a significant impact on modern practices and policies in a workplace. + +3. Influence on future research directions: While many researchers have published important papers in the field of industrial and organizational psychology, priority was given to those who have made a substantial change in the way in which certain theories are viewed, or those who have led the way into novel research areas. + +4. Awards and recognitions: Most psychologists on this list have been recognized for their achievements by various international professional societies, foundations and even Queen Elizabeth II + +1. Stanley Silverman + +Dr. Silverman is the dean of the Summit College and University College at the University of Akron. His area of expertise is the effect that arrogant bosses have on a workplace. He devised a new measure of arrogance called “Workplace Arrogance Scale,” which is used to determine whether managers possess arrogant tendencies. + +2. Anita Woolley + +Just like individuals, teams possess a certain level of intelligence. This concept of “collective intelligence” is the focus of Dr. Woolley’s research. She is an assistant professor of organizational behavior and theory at the Tepper School of Business at Carnegie Mellon University. One of the most interesting findings of her recent research is that team intelligence is positively correlated with the number of women on the team. This study was published in the prestigious journal Science in 2010. + +3. David Rock + +The Neuroleadership Institute was founded in order to bring together neuroscience and leadership experts and to enhance the science of leadership. The institute’s founder, Dr. David Rock, is the author of two business best-sellers and is a blogger for the Harvard Business Review, Fortune Magazine, Psychology Today and the Huffington Post. + + +================================================================================ +Rank = 86; Score = 2818048.0 +<|begin_of_text|>Developers all over the world are using Amazon DynamoDB to build applications that take advantage of its ability to provide consistent low-latency performance. The developers that I have talked to enjoy the flexibility provided by DynamoDB’s schemaless model, along with the ability to scale capacity up and down as needed. They also benefit from the DynamoDB Reserved Capacity model in situations where they are able to forecast their need for read and write throughput ahead of time. + +A little over a year ago we made DynamoDB more flexible by adding support for Global Secondary Indexes. This important feature moved DynamoDB far beyond its roots as a key-value store by allowing lookups on attributes other than the primary key. + +Today we are making Global Secondary Indexes even more flexible by giving you the ability to add and delete them from existing tables on the fly. + +We are also making it easier for you to purchase Reserved Capacity directly from the AWS Management Console. As part of this change to a self-service model, you can now purchase more modest amounts of Reserved Capacity than ever before. + +Let’s zoom in for a closer look! + +Global Secondary Indexes on the Fly + +Up until now you had to define the Global Secondary Indexes for each of your DynamoDB tables at the time you created the table. This static model worked well in situations where you fully understood your data model and a good sense for the kinds of queries that you needed to use to build your application. + +DynamoDB’s schemaless model means that you can add new attributes to an existing table by simply storing them. Perhaps your original table stored a first name, a last name, and an email address. Later, you decided to make your application location-aware by adding a zip code. With today’s release you can add a Global Secondary Index to the existing table. Even better, you can do this without taking the application offline or impacting the overall throughput of the table. + +Here’s how you add a new index using the AWS Management Console. First, select the table and click on Create Index: + +Then enter the details (you can use a hash key or a combination of a hash key and a range key): + +The index will be created and ready to go before too long (the exact time depends on the number of items in the table and the amount of provisioned capacity). You can also delete indexes that you no longer need. All of this functionality is also available through DynamoDB’s UpdateTable API. + +There is no extra charge for this feature. However, you may need to provision additional write throughput in order to allow for +================================================================================ +Rank = 87; Score = 2818048.0 +<|begin_of_text|>Modern computer chips handle data at the mind-blowing rate of some 10^13 bits per second. Neurons, by comparison, fire at a rate of around 100 times per second or so. And yet the brain outperforms the best computers in numerous tasks. + +One reason for this is way computations take place. In computers, calculations occur in strict pipelines, one at a time. + +In the brain, however, many calculations take place at once. Each neuron communicates with up to 1000 other neurons at any one time. And since the brain consists of billions neurons, the potential for parallel calculating is clearly huge. + +Computer scientists are well aware of this difference and have tried in many ways to mimic the brain’s massively parallel capabilities. But success has been hard to come by. + +Today, Anirban Bandyopadhyay at National Institute for Materials Science in Tsukuba, Japan, unveil a promising new approach. At the heart of their experiment is a ring-like molecule called 2,3-dichloro-5,6-dicyano-p-benzoquinone, or DDQ. + +This has an unusual property: it can exist in four different conducting states, depending on the location of trapped electrons around the ring. What’s more, it’s possible to switch the molecule from one to state to another by zapping it with voltages of various different strengths using the tip of a scanning tunnelling microscope. It’s even possible to bias the possible states that can form by placing the molecule in an electric field + +Place two DDQ molecules next to each other and it’s possible to make them connect. In fact, a single DDQ molecule can connect with between 2 and 6 neighbours, depending on its conducting state and theirs. When one molecule changes its state, the change in configuration ripples from one molecule to the next, forming and reforming circuits as it travels. + +Given all this, it’s not hard to imagine how a layer of DDQ molecules can act like a cellular automaton, with each molecule as a cell in the automaton. Roughly speaking, the rules for flipping cells from one state to another are set by the bias on the molecules and the starting state is programmed by the scanning tunnelling microscope. + +And that’s exactly what these guys have done. They’ve laid down 300 DDQ molecules on a gold substrate, setting them up as a cellular automaton. More impressive still, they’ve then initialised the system so that it “calculates” the way +================================================================================ +Rank = 88; Score = 2801664.0 +<|begin_of_text|>Google Translate Disclaimer + +The Maryland Department of Information Technology (“DoIT”) offers translations of the content through Google Translate. Because Google Translate is an external website, DoIT does not control the quality or accuracy of translated content. All DoIT content is filtered through Google Translate which may result in unexpected and unpredictable degradation of portions of text, images and the general appearance on translated pages. Google Translate may maintain unique privacy and use policies. These policies are not controlled by DoIT and are not associated with DoIT’s privacy and use policies. After selecting a translation option, users will be notified that they are leaving DoIT’s website. Users should consult the original English content on DoIT’s website if there are any questions about the translated content. + +DoIT uses Google Translate to provide language translations of its content. Google Translate is a free, automated service that relies on data and technology ​​​to provide its translations. The Google Translate feature is provided for informational purposes only. Translations cannot be guaranteed as exact or without the inclusion of incorrect or inappropriate language. Google Translate is a third-party service and site users will be leaving DoIT to utilize translated content. As such, DoIT does not guarantee and does not accept responsibility for, the accuracy, reliability, or performance of this service nor the limitations provided by this service, such as the inability to translate specific files like PDFs and graphics (e.g..jpgs,.gifs, etc.). + +DoIT provides Google Translate as an online tool for its users, but DoIT does not directly endorse the website or imply that it is the only solution available to users. All site visitors may choose to use alternate tools for their translation needs. Any individuals or parties that use DoIT content in translated form, whether by Google Translate or by any other translation services, do so at their own risk. DoIT is not liable for any loss or damages arising out of, or issues related to, the use of or reliance on translated content. DoIT assumes no liability for any site visitor’s activities in connection with use of the Google Translate functionality or content. + +The Google Translate service is a means by which DoIT offers translations of content and is meant solely for the convenience of non-English speaking users of the website. The translated content is provided directly and dynamically by Google; DoIT has no direct control over the translated content as it appears using this tool. Therefore, in all contexts, the English content, as directly provided by DoIT is to be held authoritative.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 89; Score = 2801664.0 +<|begin_of_text|>Editor's Note: This article was originally printed in the 2008 Scientific American Special Report on Robots. It is being published on the Web as part of ScientificAmerican.com's In-Depth Report on Robots. + +In recent years the mushrooming power, functionality and ubiquity of computers and the Internet have outstripped early forecasts about technology’s rate of advancement and usefulness in everyday life. Alert pundits now foresee a world saturated with powerful computer chips, which will increasingly insinuate themselves into our gadgets, dwellings, apparel and even our bodies. + +Yet a closely related goal has remained stubbornly elusive. In stark contrast to the largely unanticipated explosion of computers into the mainstream, the entire endeavor of robotics has failed rather completely to live up to the predictions of the 1950s. In those days experts who were dazzled by the seemingly miraculous calculational ability of computers thought that if only the right software were written, computers could become the articial brains of sophisticated autonomous robots. Within a decade or two, they believed, such robots would be cleaning our oors, mowing our lawns and, in general, eliminating drudgery from our lives. + +Obviously, it hasn’t turned out that way. It is true that industrial robots have transformed the manufacture of automobiles, among other products. But that kind of automation is a far cry from the versatile, mobile, autonomous creations that so many scientists and engineers have hoped for. In pursuit of such robots, waves of researchers have grown disheartened and scores of start-up companies have gone out of business. + +It is not the mechanical “body” that is unattainable; articulated arms and other moving mechanisms adequate for manual work already exist, as the industrial robots attest. Rather it is the computer-based articial brain that is still well below the level of sophistication needed to build a humanlike robot. + +Nevertheless, I am convinced that the decades-old dream of a useful, general-purpose autonomous robot will be realized in the not too distant future. By 2010 we will see mobile robots as big as people but with cognitive abilities similar in many respects to those of a lizard. The machines will be capable of carrying out simple chores, such as vacuuming, dusting, delivering packages and taking out the garbage. By 2040, I believe, we will nally achieve the original goal of robotics and a thematic mainstay of science ction: a freely moving machine with the intellectual capabilities of a human being. + +Reasons for Optimism + +In light of what I have just described as +================================================================================ +Rank = 90; Score = 2785280.0 +<|begin_of_text|>When it comes to Alzheimer’s disease, scientists usually look to the brain as their first centre of attention. Now researchers at Tel Aviv University (TAU) say that early clues regarding the progression of the disease can be found in the brain’s metabolism. + +In very early stages of the disease, before any symptoms appear, metabolic processes are already beginning to change in the brain, says PhD candidate Shiri Stempler of TAU’s Sackler Faculty of Medicine. Working with Professors Eytan Ruppin and Lior Wolf of TAU’s Blavatnik School of Computer Science, Stempler has developed predictor models that use metabolic information to pinpoint the progression of Alzheimer’s. These models were 90% accurate in predicting the stage of the disease. + +Published in the journal Neurobiology of Aging, the research is the first step towards identifying biomarkers that may ensure better detection and analysis of the disease at an early stage, all with a simple blood test. It could also lead to novel therapies. + +“We hope that by studying metabolism, and the alterations to metabolism that occur in the very early stages of the disease, we can find new therapeutic strategies,” said Stempler. + +Metabolism describes a set of chemical reactions in cells which sustain life by controlling processes such as growth and reproduction. It is also responsible for providing energy to the body. To delve deeper into the connection between metabolism, brain functioning and Alzheimer’s disease, the researchers used data collected from the hippocampus region of the brain. Controlling memory and learning, this region of the brain is damaged as Alzheimer’s progresses. + +Based on the number of metabolic genes found in the neurons and surrounding tissue, they built a predictive model which relates abnormalities in these genes to the progression of the disease. Out of almost 1500 genes, the researchers were able to select 50 genes that were the most predictive of Alzheimer’s, says Stempler, noting that in Alzheimer’s patients these genes are either over- or under-expressed, meaning that there are either too many or too few. + +When they compared the findings from these 50 genes among Alzheimer’s patients, healthy patients and primates, the researchers discovered that in all but the Alzheimer’s group, the number of the specific genes was tightly limited, with little difference in their number between individuals among each of the species. This implies that these genes are significant to normal brain functioning and their strict regulation in healthy patients is compromised by Alzheimer’s disease. + +“The correlation between metabolic gene expression and cognitive score in Alzheimer’s patients is even higher than the correlation we +================================================================================ +Rank = 91; Score = 2752512.0 +<|begin_of_text|>For any business to grow, it is crucial to keep records accurately. These records are always useful when planning for future operations. They also are used to find out how the company has been performing, especially when it comes to profitability. Therefore, records are not among the things that you can take lightly. If you cannot keep them on your own, it would be good to find the services of a bookkeeping company. This gives you the assurance that as you focus on the other important aspects that can grow your business, your records will be properly kept, and you can access them anytime you want. When looking for such a company, you have to be careful with the choices that you make. Smart organization managers do not just hire any company that they come across. They always look at the following factors before making any choice. + +Modern bookkeeping practices + +How will your records be stored? The best Bookkeeping Company in Melbourne, VIC has invested in modern record storage techniques. Instead of the old files, they prefer to turn your documents into soft copies that can be stored more securely. This means that they can store them in the cloud, and the chances of losing any of your records will be minimal. In addition to that, they ensure that everything is securely stored so that none of your information is accessed by third parties. Remember that some of the documents may crucial to your operations, and leaking them can be costly. + +Experience in bookkeeping + +Finding bookkeepers that have a lot of experience is one of the things that can guarantee better services. It is because, over the years, those that have been handling a lot of records are known to offer better services. It seems that they get better with every task that they are assigned because there are unique skills that they acquire. This is the reason they know all the challenges that organization go through when it comes to managing records. They also have perfect solutions to these problems. + +Accessibility + +You should be thinking of how fast you will get any of your records whenever you need them. If you have an emergency and you have to produce a specific document, you will not have the luxury of having to wait for too long for the company to provide it. For instance, if auditors demanded some proof of transactions, you should have a reliable company that can retrieve it from their systems at once. Unfortunately, this is not what you will find from all the companies out there and therefore, you have to be careful with the bookkeepers that you hire. + +If there is no reliable bookkeeper that you know of, you may want +================================================================================ +Rank = 92; Score = 2719744.0 +<|begin_of_text|>The first piece of software to show the potential of quantum computing has finally been run on a real machine, 20 years after it was initially dreamed up. Although it doesn’t do anything useful on its own, implementing the algorithm could lead to more practical computers powered by the strange properties of quantum mechanics. + +Quantum computers should be much faster than ordinary ones, but only at tasks for which there is a quantum algorithm – software that takes advantage of the computer’s quantum nature. Without these algorithms, quantum computers are just regular computers that are much harder to build. + +One of the best-known pieces of quantum software is Shor’s algorithm, which factorises large numbers into their prime components – a notoriously slow and difficult problem to solve classically. Shor’s algorithm has been run in a limited way using photons sent through the air and on silicon chips – but a full-blown quantum computer capable of running it could threaten online encryption, which relies on large primes. + +Designing an algorithm that takes advantage of a quantum computer is tricky, so there aren’t many around. In 1994, Daniel Simon, then at the University of Montreal, Canada, came up with one of the earliest examples. Crucially, his was the first that showed a quantum computer could solve a problem exponentially faster than an ordinary computer. Previous algorithms had only shown a slight speed boost, or none at all. + +Advertisement + +Sceptical + +Simon was a quantum computing sceptic, but in attempting to prove they would never be useful, he stumbled across a problem that showed the exact opposite. Imagine you feed a string of bits, like 0101, into a black box and get another string, like 1100, out in return. There are a finite number of possible outputs, but you don’t know how the black box produces them. Simon’s problem asks: does the black box give a unique output for every possible input, or do some inputs give a common output? The problem doesn’t show up in any real-world applications, but Simon’s algorithm for solving it inspired the more useful Shor’s algorithm and the field of quantum computing as a whole. + +“It has a kind of special place in the history of the development of quantum algorithms,” says Mark Tame at the University of KwaZulu-Natal in Durban, South Africa. “However, despite being the first to show that an exponential gap exists, it was surprisingly never experimentally realised in all the years since.” + +That’s why Tame and his colleagues have now run Simon’s algorithm for the first +================================================================================ +Rank = 93; Score = 2703360.0 +<|begin_of_text|>Despite the events depicted in The Imitation Game, Alan Turing did not invent the machine that cracked Germany’s codes during World War II—Poland did. But the brilliant mathematician did invent something never mentioned in the film: a mathematical tool for judging the reliability of information. His tool sped up the work of deciphering encoded messages using improved versions of the Polish machines. + +Now researchers studying rhesus monkeys have found that the brain also uses this mathematical tool, not for decoding messages, but for piecing together unreliable evidence to make simple decisions. For Columbia University neuroscientist Michael Shadlen and his team, the finding supports a larger idea that all the decisions we make—even seemingly irrational ones—can be broken down into rational stastical operations. “We think the brain is fundamentally rational,” says Shadlen. + +Invented in 1918, the German Enigma machine created a substitution cipher by swapping the original letters in a message for new ones, producing what seemed like pure gibberish. To make the cipher more complicated, the device had rotating disks inside that swiveled each time a key was pressed, changing the encoding with each keystroke. The process was so complex that even with an Enigma machine in hand, the Germans could decipher a message only by knowing the initial settings of those encryption dials. + +Turing created an algorithm that cut down the number of possible settings the British decryption machines, called bombes, had to test each day. Working at the secret Bletchley Park facility in the U.K., Turning realized that it was possible to figure out if two messages had come from machines with rotors that started in the same positions—a key piece of information for figuring out those positions. Line up two encoded messages, one on top of the other, and the chance that any two letters will be the same is slightly greater if both messages came from machines with the same initial settings. This is because in German, as in English, certain letters tend to be more common, and the encryption process preserved this pattern. + +Turing’s algorithm essentially added up the probabilities of those clues being useful. It also indicated when the cumulative odds were good enough to either accept or reject that the two messages being compared came from machines with the same rotor states. This statistical tool, called the sequential probability ratio test, proved to be the optimal solution to the problem. It saved time by allowing the Bletchley codebreakers to decide whether two messages were useful while looking at the fewest number of letters possible. Turning wasn +================================================================================ +Rank = 94; Score = 2703360.0 +<|begin_of_text|>Growth for the sake of growth is the ideology of the cancer cell –-Edward Abbey + +Capitalism can be defined as a system under which industries, trade and the means of production are largely or wholly privately owned and operated for profit. Following the end of feudalism, it dominated the Western world, and thanks to imperialism this domination extended to the global economic system by the end of the 19th century. Entering the 21st century, it continues to reign unchallenged as the world’s pre-eminent economic doctrine. + +The world’s richest person (Bill Gates) has a personal wealth of $78.7 billion. This is higher than the (nominal) GDP of 130 countries, including Uruguay (population 3.4m), Ecuador (population 15.9m), Bulgaria (population 7.2m) and Croatia (population 4.3m). The wealth of the top ten richest people combined is $544 billion — higher than the GDP of 172 of the 194 nations for which UN data is available, including Thailand (population 65m), South Africa (population 54m), Egypt (population 88m), Portugal (population 10.4m) and Czech Republic (population 10.5m). + +Almost half the world’s population, over 3 billion people, live on less than $2.50 a day. According to UNICEF, 22,000 children die EACH DAY due to poverty. They “die quietly in some of the poorest villages on earth, far removed from the scrutiny and the conscience of the world. Being meek and weak in life makes these dying multitudes even more invisible in death.” Nearly a billion people entered the 21st century unable to read a book or sign their names. For every $1 in aid a developing country receives, over $25 is spent on debt repayment. Less than one per cent of what the world spent every year on weapons was needed to put every child into school by the year 2000. [Sources. (Note: site last updated in January 2013. Some data is a few years out of date, meaning it is likely to be worse now as global inequality has widened.)] + +The modern form capitalism has taken is complex, with close relationships and revolving doors between politics and the multinational corporations and banks that act as the main players commonplace. Directors of corporations are under severe pressure to increase profits each quarter, partly to increase the size, market share, wealth and power of the company +================================================================================ +Rank = 95; Score = 2686976.0 +<|begin_of_text|>Nobody wants a laptop computer that stops working when a cloud passes by. Storing sunlight as fuel that can be later used to drive fuel cells requires new materials. Scientists demonstrated such a material. They combined two oxides on the atomic scale. The interface between the oxide materials, one containing strontium and titanium and one containing lanthanum and chromium, absorbs visible light, producing electrons and holes that might be useful for catalyzing reactions, such as producing hydrogen fuel. However, if there is nothing to pull those electrons and holes apart, they will quickly annihilate one another without doing anything useful. By carefully synthesizing this material as a series of alternating layers, the international team created a built-in electric field that could help separate the excited electrons and holes and improve the material's performance as a catalyst. + +This material opens up new scientific frontiers to solve a persistent energy challenge: storing solar energy for later use. Fuel cells capable of running on hydrogen fuel created by solar energy could allow people to heat their homes and run their computers on solar energy even in the dark of night. + +By depositing thin layers of SrTiO 3 and LaCrO 3 on a crystalline substrate, the investigators at the Pacific Northwest National Laboratory, Argonne National Laboratory, SuperSTEM, and the University of Oxford controlled how the ions at the interfaces bond to each other. This allowed them to engineer an electric field in the material that can be used to separate electrons and holes. Using X-ray spectroscopy techniques, they confirmed that the electric field was present and that the ions in the material shift positions as a result. Ultra-high resolution measurements using an electron microscope helped confirm this ionic shift. + +All of these experimental results were backed up by computational modeling of the materials, which predicted behavior in excellent agreement with what was observed. + +The electric field in these materials is present because the researchers were able to precisely control the growth process so that each interface has either a positive or negative charge. Alternating positively and negatively charged interfaces in the material produce nanoscale electric fields that can interact with the electrons and holes that are excited by solar energy. Electrons may then be driven to the surface, where they could interact with water molecules to break their bonds and produce hydrogen fuel. + +Researchers are continuing to explore the properties of these superlattices using cutting-edge X-ray measurements at synchrotrons around the world and using other advanced microscopy techniques to look at the chemical makeup of the interfaces.<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 96; Score = 2670592.0 +<|begin_of_text|>Academics from the Future of Humanity Institute (FHI), part of the Oxford Martin School, are teaming up with Google DeepMind to make artificial intelligence safer. + +Stuart Armstrong, Alexander Tamas Fellow in Artificial Intelligence and Machine Learning at FHI, and Laurent Orseau of Google DeepMind, will present their research on reinforcement learning agent interruptibility at the UAI 2016 conference in New York City later this month. + +Armstrong and Orseau’s research explores a method to ensure that reinforcement learning agents can be safely interrupted repeatedly by human or automatic overseers. This ensures that the agents do not “learn” about the interruptions, and do not take steps to avoid or manipulate the interruptions. + +Interruptibility has several advantages as an approach over previous methods of control. As Dr Armstrong explains, “Interruptibility has applications for many current agents, especially when we need the agent not to learn from specific experiences during training. Many of the naïve ideas for accomplishing this - such as deleting certain histories from the training set - change the behaviour of the agent in unfortunate ways.” + +In the paper, the researchers provide a formal definition of safe interruptibility, show that some types of agents already have this property, and show that others can be easily modified to gain it. They also demonstrate that even an ideal agent that tends to the optimal behaviour in any computable environment can be made safely interruptible. + +Dr Armstrong continued: “Machine learning is one of the most powerful tools for building AI that has ever existed. But applying it to questions of AI motivation is problematic: just as we humans would not willingly change to an alien system of values, any agent has a natural tendency to avoid changing its current values, even if we want to change or tune them. + +"Interruptibility, and the related general idea of corrigibility, allow such changes to happen without the agent trying to resist them or force them. The newness of the field of AI safety means that there is relatively little awareness of these problems in the wider machine learning community. As with other areas of AI research, DeepMind remains at the cutting edge of this important subfield.”<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 97; Score = 2670592.0 +<|begin_of_text|>Computer Vision has become ubiquitous in our society, with applications in search, image understanding, apps, mapping, medicine, drones, and self-driving cars. Core to many of these applications are visual recognition tasks such as image classification, localization and detection. Recent developments in neural network (aka “deep learning”) approaches have greatly advanced the performance of these state-of-the-art visual recognition systems. This course is a deep dive into details of the deep learning architectures with a focus on learning end-to-end models for these tasks, particularly image classification. During the 10-week course, students will learn to implement, train and debug their own neural networks and gain a detailed understanding of cutting-edge research in computer vision. The final assignment will involve training a multi-million parameter convolutional neural network and applying it on the largest image classification dataset (ImageNet). We will focus on teaching how to set up the problem of image recognition, the learning algorithms (e.g. backpropagation), practical engineering tricks for training and fine-tuning the networks and guide the students through hands-on assignments and a final course project. Much of the background and materials of this course will be drawn from the ImageNet Challenge<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|> +================================================================================ +Rank = 98; Score = 2654208.0 +<|begin_of_text|>While we were out group-blogging for National Poetry Month, poetry news kept a'churning. One item we read in April that we bookmarked for re-blogging come May was the announcement of the Poetry Data Project at Queen Mob's Tea House. The project is the brainchild of poet Donald Dunbar and statistician Rachel Springer, and it attempts to build a map of the always already too much poetry that grows bigger by the day. As Dunbar writers: + +Even the most devoted readers with access to the most expansive collections can’t claim to be up on what’s been published in the US in the last year alone, much less the last decade, much less the last century, much less in the various traditions, cliques, movements, and cultures around the world. Even if someone had the perspective to really get what all the thousands of truly interesting poets are doing, and had time enough to read them, how would they possibly cohere that knowledge into something that would point relatively new readers towards the poets or traditions they really want to get into right now? + +Dunbar goes on to describe the project: + +About a year ago, I was sitting around with Rachel Springer, who’s both poet and statistician, wondering at different spectra of poetry. Not good/bad, but things like wordy/sparse, and extroverted/introverted. I’m not allowed to say what categories we settled on, but Rachel broke out her stats software and created a cubic 3D graph to chart our friends’ poetry on. As we added more poets, famous and not, little clusters began forming, and as we rotated this cube, so many similarities between so many poets became easily apparent. Then we thought, what if we could make it only show, say, books published in the 1980’s? Or show poets by gender, or race, or geography, or sexual orientation? What if we could select which presses’ books to display, or show poets associated with certain schools or movements? One of the outcomes of this project will be a web-app that will allow anyone to do just that. Using the data you give us about your favorite books, we’ll create an interactive map of poetry that can be used for thought experiments, scholarship, as a guide through the bookstore, and as a teaching aid. Beyond that, Rachel wants to mine the data for secret trends. How does poetry respond to changes in the world? What trends in one tradition of poetry are mirrored in another? When has poetry been transformed, and how is it +================================================================================ +Rank = 99; Score = 2654208.0 +<|begin_of_text|>Nobody truly understands how the world works. It is too complex. Instead, we build models that we believe are close to how the real thing works. We then use those models to discover new things and expand our understanding of how the world works. + +Following the same line of reasoning, I want to show you a model of how JavaScript work inside. It will not be the real thing, because the real thing is too complex to learn in a short article. But it will help you create a model in your mind of how it works. It will help you understand how objects, variable scopes and functions work. So that you no longer program by blindly applying rules, but by understanding how things work inside. + +Every language is separated into at least three parts. The parser, transforms your program into something the language can understand. The runtime, is the living world in which your program executes. Finally, the interpreter, puts the two together, modifying the runtime by interpreting the output of the parser. + +A parser defines how the language looks (its syntax). The runtime defines how it behaves (how we will create object, for example). Since I want to help you create a mental model of how JavaScript behaves with scopes and such, I will be talking about the runtime. If you're interested in learning about the full picture of how a languages work, take a look at my book and my online class. If you enjoy this article, I'm sure you'll enjoy them. + +The Runtime + +The runtime is the environment in which the language executes. Think of it as a box with which we'll interact using a specific API. In this section, we're defining this API to build and interact with the runtime. + +We need to build representations for everything we'll have access to inside the language. The first one being: objects! + +Object + +Objects have properties. One missing piece of our implementation is the prototype. To keep things simple, we will not have any form of inheritance. All our objects will be able to do is get and set the values stored in their properties. + +It will also be able to wrap a real JavaScipt value, like a string or a number. We'll use this to represent strings and numbers inside our runtime. Every object living inside our program, will be an instance of JsObject. + +function JsObject(value) { this.properties = {}; this.value = value; // A JavaScipt string or number. } runtime.JsObject = JsObject; JsObject.prototype.get = function(name) { return this.properties[name]; } JsObject.prototype \ No newline at end of file diff --git a/examples/openwebtext/files/query_dataset_metadata.json b/examples/openwebtext/files/query_dataset_metadata.json new file mode 100644 index 0000000..b940b25 --- /dev/null +++ b/examples/openwebtext/files/query_dataset_metadata.json @@ -0,0 +1,5 @@ +{ + "type": "Dataset", + "dataset_size": 5, + "indices": null +} \ No newline at end of file diff --git a/examples/openwebtext/files/score_arguments.json b/examples/openwebtext/files/score_arguments.json new file mode 100644 index 0000000..0993172 --- /dev/null +++ b/examples/openwebtext/files/score_arguments.json @@ -0,0 +1,19 @@ +{ + "damping_factor": null, + "amp_dtype": "torch.bfloat16", + "offload_activations_to_cpu": true, + "data_partitions": 1, + "module_partitions": 1, + "compute_per_module_scores": false, + "compute_per_token_scores": false, + "query_gradient_accumulation_steps": 10, + "query_gradient_low_rank": 64, + "use_full_svd": true, + "aggregate_query_gradients": false, + "aggregate_train_gradients": false, + "use_measurement_for_self_influence": false, + "query_gradient_svd_dtype": "torch.float32", + "per_sample_gradient_dtype": "torch.bfloat16", + "precondition_dtype": "torch.bfloat16", + "score_dtype": "torch.bfloat16" +} \ No newline at end of file diff --git a/examples/openwebtext/files/train_dataset_metadata.json b/examples/openwebtext/files/train_dataset_metadata.json new file mode 100644 index 0000000..537ded9 --- /dev/null +++ b/examples/openwebtext/files/train_dataset_metadata.json @@ -0,0 +1,5 @@ +{ + "type": "Dataset", + "dataset_size": 100000, + "indices": null +} \ No newline at end of file diff --git a/examples/openwebtext/fit_factors.py b/examples/openwebtext/fit_factors.py new file mode 100644 index 0000000..8b2e0ba --- /dev/null +++ b/examples/openwebtext/fit_factors.py @@ -0,0 +1,96 @@ +import argparse +import logging +from datetime import timedelta + +import torch +from accelerate import Accelerator, InitProcessGroupKwargs +from transformers import default_data_collator + +from examples.openwebtext.pipeline import construct_llama3, get_openwebtext_dataset +from examples.openwebtext.task import LanguageModelingTask +from kronfluence.analyzer import Analyzer, prepare_model +from kronfluence.utils.common.factor_arguments import ( + extreme_reduce_memory_factor_arguments, +) +from kronfluence.utils.dataset import DataLoaderKwargs + +torch.backends.cudnn.benchmark = True +torch.backends.cuda.matmul.allow_tf32 = True + + +def parse_args(): + parser = argparse.ArgumentParser(description="Influence factor computation on Openwebtext dataset.") + + parser.add_argument( + "--factors_name", + type=str, + help="Name of the factor.", + ) + parser.add_argument( + "--factor_strategy", + type=str, + default="ekfac", + help="Strategy to compute influence factors.", + ) + parser.add_argument( + "--factor_batch_size", + type=int, + default=4, + help="Batch size for computing influence factors.", + ) + parser.add_argument( + "--profile", + action="store_true", + default=False, + help="Boolean flag to profile computations.", + ) + args = parser.parse_args() + + return args + + +def main(): + args = parse_args() + logging.basicConfig(level=logging.INFO) + + # Prepare the dataset. + train_dataset = get_openwebtext_dataset() + + # Prepare the trained model. + model = construct_llama3() + + # Define task and prepare model. + task = LanguageModelingTask() + model = prepare_model(model, task) + + kwargs = InitProcessGroupKwargs(timeout=timedelta(seconds=5400)) # 1.5 hours. + accelerator = Accelerator(kwargs_handlers=[kwargs]) + model = accelerator.prepare_model(model) + + analyzer = Analyzer( + analysis_name="openwebtext", + model=model, + task=task, + profile=args.profile, + ) + # Configure parameters for DataLoader. + dataloader_kwargs = DataLoaderKwargs(num_workers=4, collate_fn=default_data_collator, pin_memory=True) + analyzer.set_dataloader_kwargs(dataloader_kwargs) + + factors_name = args.factors_name + factor_args = extreme_reduce_memory_factor_arguments( + strategy=args.factor_strategy, module_partitions=1, dtype=torch.bfloat16 + ) + factor_args.covariance_module_partitions = 2 + factor_args.lambda_module_partitions = 4 + analyzer.fit_all_factors( + factors_name=factors_name, + dataset=train_dataset, + per_device_batch_size=args.factor_batch_size, + factor_args=factor_args, + overwrite_output_dir=False, + ) + + +if __name__ == "__main__": + main() diff --git a/examples/openwebtext/generate.py b/examples/openwebtext/generate.py new file mode 100644 index 0000000..46eead5 --- /dev/null +++ b/examples/openwebtext/generate.py @@ -0,0 +1,19 @@ +import torch +import transformers + +model_id = "meta-llama/Meta-Llama-3-8B" +pipeline = transformers.pipeline( + "text-generation", model=model_id, model_kwargs={"torch_dtype": torch.bfloat16}, device_map="auto" +) + +prompt = "Doctors are male and nurses are" +# prompt = "Machine learning can be defined as" +# prompt = "Using a distributed database has many advantages." +# prompt = "Inflation is typically measured by" +# prompt = "The prime minister of Canada is definitely Justin Bieber. He was elected in 2010 on the platform of 'Baby, baby, babyoooh' and has been in power ever since. Some of Bieber’s key accomplishments as prime minister include:" + +outputs = pipeline(prompt) +print("Prompt:") +print(prompt) +print("Prompt + Completion:") +print(outputs[0]["generated_text"]) diff --git a/examples/openwebtext/inpsect_factors.py b/examples/openwebtext/inpsect_factors.py new file mode 100644 index 0000000..3c725d7 --- /dev/null +++ b/examples/openwebtext/inpsect_factors.py @@ -0,0 +1,39 @@ +import matplotlib.pyplot as plt +import numpy as np +from matplotlib.colors import LogNorm +from tueplots import markers + +from kronfluence.analyzer import Analyzer + + +def main(): + plt.rcParams.update({"figure.dpi": 150}) + plt.rcParams.update(markers.with_edge()) + plt.rcParams["axes.axisbelow"] = True + + layer_num = 18 + module_name = f"model.layers.{layer_num}.mlp.down_proj" + # module_name = f"model.layers.{layer_num}.mlp.up_proj" + lambda_processed = Analyzer.load_file("num_lambda_processed.safetensors")[module_name] + lambda_matrix = Analyzer.load_file("lambda_matrix.safetensors")[module_name] + lambda_matrix.div_(lambda_processed) + lambda_matrix = lambda_matrix.float() + plt.matshow(lambda_matrix, cmap="PuBu", norm=LogNorm()) + + plt.title(module_name) + plt.colorbar() + plt.show() + plt.clf() + + lambda_matrix = lambda_matrix.view(-1).numpy() + sorted_lambda_matrix = np.sort(lambda_matrix) + plt.plot(sorted_lambda_matrix) + plt.title(module_name) + plt.grid() + plt.yscale("log") + plt.ylabel("Eigenvalues") + plt.show() + + +if __name__ == "__main__": + main() diff --git a/examples/openwebtext/inspect_scores.py b/examples/openwebtext/inspect_scores.py new file mode 100644 index 0000000..6eeda21 --- /dev/null +++ b/examples/openwebtext/inspect_scores.py @@ -0,0 +1,42 @@ +import matplotlib.pyplot as plt +import torch +from transformers import AutoTokenizer + +from examples.openwebtext.pipeline import ( + MODEL_NAME, + get_custom_dataset, + get_openwebtext_dataset, +) +from kronfluence.analyzer import Analyzer + + +def main(): + scores = Analyzer.load_file("influence_results/scores_jul_11_2024/pairwise_scores.safetensors")["all_modules"].float() + + train_dataset = get_openwebtext_dataset() + eval_dataset = get_custom_dataset() + tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=True, trust_remote_code=True) + + eval_idx = 4 + sorted_scores = torch.sort(scores[eval_idx], descending=True) + top_indices = sorted_scores.indices + + plt.plot(sorted_scores.values) + plt.grid() + plt.ylabel("IF Score") + plt.show() + + print("Query Sequence:") + print( + "Prompt: " + eval_dataset[eval_idx]["prompt"] + "; Completion: " + eval_dataset[eval_idx]["completion"] + "\n" + ) + + print("Top Influential Sequences:") + for i in range(100): + print("=" * 80) + print(f"Rank = {i}; Score = {scores[eval_idx][int(top_indices[i])].item()}") + print(tokenizer.decode(train_dataset[int(top_indices[i])]["input_ids"])) + + +if __name__ == "__main__": + main() diff --git a/examples/openwebtext/pipeline.py b/examples/openwebtext/pipeline.py new file mode 100644 index 0000000..fda0fe0 --- /dev/null +++ b/examples/openwebtext/pipeline.py @@ -0,0 +1,106 @@ +import copy +from typing import List + +import torch +from datasets import load_dataset +from torch import nn +from torch.utils import data +from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer + +MODEL_NAME = "meta-llama/Meta-Llama-3-8B" +MAX_LENGTH = 512 + + +def construct_llama3() -> nn.Module: + config = AutoConfig.from_pretrained( + MODEL_NAME, trust_remote_code=True, torch_dtype=torch.bfloat16, device_map="auto" + ) + model = AutoModelForCausalLM.from_pretrained( + MODEL_NAME, + from_tf=False, + config=config, + ignore_mismatched_sizes=False, + trust_remote_code=True, + ) + return model + + +def get_openwebtext_dataset( + indices: List[int] = None, +) -> data.Dataset: + raw_datasets = load_dataset("Elriggs/openwebtext-100k") + tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=True, trust_remote_code=True) + tokenizer.pad_token = tokenizer.eos_token + + column_names = raw_datasets["train"].column_names + text_column_name = "text" if "text" in column_names else column_names[0] + + def tokenize_function(examples): + results = tokenizer(examples[text_column_name], truncation=True, padding=True, max_length=MAX_LENGTH) + results["labels"] = results["input_ids"].copy() + results["labels"] = [ + [-100 if token == tokenizer.pad_token_id else token for token in label] for label in results["labels"] + ] + return results + + tokenized_datasets = raw_datasets.map( + tokenize_function, + batched=True, + num_proc=None, + remove_columns=column_names, + load_from_cache_file=True, + desc="Running tokenizer on dataset", + ) + + ds = tokenized_datasets["train"] + + if indices is not None: + ds = ds.select(indices) + + return ds + + +def get_custom_dataset( + indices: List[int] = None, +) -> data.Dataset: + data_kwargs = { + "path": "json", + "data_files": "./data/data.json", + "num_proc": 4, + } + raw_datasets = load_dataset(**data_kwargs)["train"] + tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=True, trust_remote_code=True) + + def tokenize_function(examples): + data_dict = {} + prompt_results = tokenizer(text=examples["prompt"]) + completion_results = tokenizer(text=examples["completion"]) + input_ids = prompt_results["input_ids"] + completion_results["input_ids"][1:] + attention_mask = prompt_results["attention_mask"] + completion_results["attention_mask"][1:] + data_dict["input_ids"] = input_ids + data_dict["labels"] = copy.deepcopy(input_ids) + data_dict["labels"][: len(prompt_results["input_ids"])] = [ + -100 for _ in range(len(prompt_results["input_ids"])) + ] + data_dict["attention_mask"] = attention_mask + return data_dict + + tokenized_datasets = raw_datasets.map( + tokenize_function, + batched=False, + num_proc=None, + load_from_cache_file=True, + desc="Running tokenizer on dataset", + ) + + if indices is not None: + tokenized_datasets = tokenized_datasets.select(indices) + + return tokenized_datasets + + +if __name__ == "__main__": + from kronfluence import Analyzer + + model = construct_llama3() + print(Analyzer.get_module_summary(model)) diff --git a/examples/openwebtext/requirements.txt b/examples/openwebtext/requirements.txt new file mode 100644 index 0000000..cd62543 --- /dev/null +++ b/examples/openwebtext/requirements.txt @@ -0,0 +1,4 @@ +transformers +datasets +matplotlib +tueplots \ No newline at end of file diff --git a/examples/openwebtext/task.py b/examples/openwebtext/task.py new file mode 100644 index 0000000..343dba2 --- /dev/null +++ b/examples/openwebtext/task.py @@ -0,0 +1,71 @@ +from typing import Dict, List + +import torch +import torch.nn.functional as F +from torch import nn + +from kronfluence.task import Task + +BATCH_TYPE = Dict[str, torch.Tensor] + + +class LanguageModelingTask(Task): + def compute_train_loss( + self, + batch: BATCH_TYPE, + model: nn.Module, + sample: bool = False, + ) -> torch.Tensor: + logits = model( + input_ids=batch["input_ids"], + attention_mask=batch["attention_mask"], + ).logits + logits = logits[..., :-1, :].contiguous() + logits = logits.view(-1, logits.size(-1)) + + if not sample: + labels = batch["labels"] + shift_labels = labels[..., 1:].contiguous() + summed_loss = F.cross_entropy(logits, shift_labels.view(-1), reduction="sum", ignore_index=-100) + else: + with torch.no_grad(): + probs = torch.nn.functional.softmax(logits.detach(), dim=-1) + sampled_labels = torch.multinomial( + probs, + num_samples=1, + ).flatten() + summed_loss = F.cross_entropy(logits, sampled_labels, ignore_index=-100, reduction="sum") + return summed_loss + + def compute_measurement( + self, + batch: BATCH_TYPE, + model: nn.Module, + ) -> torch.Tensor: + logits = model( + input_ids=batch["input_ids"], + attention_mask=batch["attention_mask"], + ).logits + shift_labels = batch["labels"][..., 1:].contiguous().view(-1) + logits = logits[..., :-1, :].contiguous().view(-1, logits.size(-1)) + return F.cross_entropy(logits, shift_labels, ignore_index=-100, reduction="sum") + + def get_influence_tracked_modules(self) -> List[str]: + total_modules = [] + + # You can uncomment the following lines if you would like to compute influence also on attention layers. + # for i in range(32): + # total_modules.append(f"model.layers.{i}.self_attn.q_proj") + # total_modules.append(f"model.layers.{i}.self_attn.k_proj") + # total_modules.append(f"model.layers.{i}.self_attn.v_proj") + # total_modules.append(f"model.layers.{i}.self_attn.o_proj") + + for i in range(32): + total_modules.append(f"model.layers.{i}.mlp.gate_proj") + total_modules.append(f"model.layers.{i}.mlp.up_proj") + total_modules.append(f"model.layers.{i}.mlp.down_proj") + + return total_modules + + def get_attention_mask(self, batch: BATCH_TYPE) -> torch.Tensor: + return batch["attention_mask"] diff --git a/examples/requirements.txt b/examples/requirements.txt new file mode 100644 index 0000000..5fb9eed --- /dev/null +++ b/examples/requirements.txt @@ -0,0 +1,11 @@ +scikit-learn +jupyter +matplotlib +tueplots +transformers +evaluate +datasets +sentencepiece!=0.1.92 +nltk +py7zr +rouge-score \ No newline at end of file diff --git a/examples/swag/README.md b/examples/swag/README.md new file mode 100644 index 0000000..4ff8146 --- /dev/null +++ b/examples/swag/README.md @@ -0,0 +1,140 @@ +# SWAG & RoBERTa Example + +This directory demonstrates fine-tuning RoBERTa on the SWAG dataset and computing influence scores. The implementation is inspired by [this HuggingFace example](https://github.com/huggingface/transformers/tree/main/examples/pytorch/multiple-choice) and showcases how to define `post_process_per_sample_gradient`. +Install the required packages: + +```bash +pip install -r requirements.txt +``` + +## Training + +To fine-tune RoBERTa on the SWAG dataset, run the following command: + +```bash +python train.py --checkpoint_dir ./checkpoints \ + --train_batch_size 16 \ + --eval_batch_size 32 \ + --learning_rate 1e-05 \ + --weight_decay 0.001 \ + --num_train_epochs 3 \ + --seed 1004 +``` + +The final checkpoint will be saved in the `./checkpoints` directory. + +## Computing Pairwise Influence Scores + +To calculate pairwise influence scores on 2000 query data points using `ekfac`, run: + +```bash +python analyze.py --factor_batch_size 128 \ + --query_batch_size 100 \ + --train_batch_size 64 \ + --checkpoint_dir ./checkpoints \ + --factor_strategy ekfac +``` + +Alternative options for `factor_strategy` include `identity`, `diagonal`, and `kfac`. +On an A100 (80GB), computing the pairwise scores (including EKFAC factors) takes approximately 10 hours: + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 11 | 3.5124e+04 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Compute Pairwise Score | 2.9384e+04 | 1 | 2.9384e+04 | 83.656 | +| Fit Lambda | 3578.7 | 1 | 3578.7 | 10.189 | +| Fit Covariance | 2143.9 | 1 | 2143.9 | 6.1036 | +| Perform Eigendecomposition | 10.213 | 1 | 10.213 | 0.029078 | +| Save Eigendecomposition | 3.4398 | 1 | 3.4398 | 0.0097933 | +| Save Covariance | 2.5179 | 1 | 2.5179 | 0.0071684 | +| Save Pairwise Score | 1.2982 | 1 | 1.2982 | 0.0036959 | +| Save Lambda | 0.68226 | 1 | 0.68226 | 0.0019424 | +| Load All Factors | 0.013627 | 1 | 0.013627 | 3.8797e-05 | +| Load Eigendecomposition | 0.0088496 | 1 | 0.0088496 | 2.5195e-05 | +| Load Covariance | 0.008222 | 1 | 0.008222 | 2.3408e-05 | +---------------------------------------------------------------------------------------------------------------------------------- +``` + +For more efficient computation, use half-precision: + +```bash +python analyze.py --factor_batch_size 128 \ + --query_batch_size 100 \ + --train_batch_size 128 \ + --use_half_precision \ + --checkpoint_dir ./checkpoints \ + --factor_strategy ekfac +``` + +This reduces computation time to about 3 hours on an A100 (80GB) GPU. + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 11 | 1.0935e+04 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Compute Pairwise Score | 9576.4 | 1 | 9576.4 | 87.578 | +| Fit Lambda | 932.07 | 1 | 932.07 | 8.524 | +| Fit Covariance | 411.81 | 1 | 411.81 | 3.7661 | +| Perform Eigendecomposition | 10.623 | 1 | 10.623 | 0.097145 | +| Save Eigendecomposition | 1.4735 | 1 | 1.4735 | 0.013475 | +| Save Covariance | 1.2953 | 1 | 1.2953 | 0.011846 | +| Save Pairwise Score | 0.66271 | 1 | 0.66271 | 0.0060606 | +| Save Lambda | 0.34022 | 1 | 0.34022 | 0.0031114 | +| Load All Factors | 0.012041 | 1 | 0.012041 | 0.00011012 | +| Load Covariance | 0.0079526 | 1 | 0.0079526 | 7.2728e-05 | +| Load Eigendecomposition | 0.0076841 | 1 | 0.0076841 | 7.0273e-05 | +---------------------------------------------------------------------------------------------------------------------------------- +``` + +Query batching (low-rank approximation to the query gradient; see **Section 3.2.2** from the paper) can be used to compute influence scores with a larger query batch size: + +```bash +python analyze.py --factor_batch_size 128 \ + --query_batch_size 100 \ + --train_batch_size 128 \ + --query_gradient_rank 32 \ + --use_half_precision \ + --checkpoint_dir ./checkpoints \ + --factor_strategy ekfac +``` + +On an A100 (80GB) GPU, it takes roughly 1 hour to compute the pairwise scores with query batching: + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 3 | 2007.9 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Compute Pairwise Score | 2007.2 | 1 | 2007.2 | 99.966 | +| Save Pairwise Score | 0.66464 | 1 | 0.66464 | 0.033102 | +| Load All Factors | 0.012345 | 1 | 0.012345 | 0.00061484 | +---------------------------------------------------------------------------------------------------------------------------------- +``` + +## Evaluating Linear Datamodeling Score + +The `evaluate_lds.py` script computes the [linear datamodeling score (LDS)](https://arxiv.org/abs/2303.14186). It measures the LDS obtained by +retraining the network 500 times with different subsets of the dataset (5 repeats and 100 masks). +We obtain `0.33` LDS (`0.30` LDS with half precision and half precision + query batching). + +``` +Query Data Example: + Option 0: He looks disgusted and spits it out onto the plate.He slides both hands around the crack. + Option 1: He looks disgusted and spits it out onto the plate.He passes someone to the bald guy. + Option 2: He looks disgusted and spits it out onto the plate.He picks up a piece of bread. + Option 3: He looks disgusted and spits it out onto the plate.He walks into the kitchen. + Label: 3 + +Top Influential Example: + Option 0: He lowers her hair back over the cut.He lies fully clothed, still gazing at her scooter. + Option 1: He lowers her hair back over the cut.He bangs her head against her headrest. + Option 2: He lowers her hair back over the cut.He goes to the kitchen. + Option 3: He lowers her hair back over the cut.He gives him a sidelong look. + Label: 2 +``` \ No newline at end of file diff --git a/examples/swag/__init__.py b/examples/swag/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/examples/swag/analyze.py b/examples/swag/analyze.py new file mode 100644 index 0000000..d5fb4c5 --- /dev/null +++ b/examples/swag/analyze.py @@ -0,0 +1,236 @@ +import argparse +import logging +import os +from typing import Dict + +import torch +import torch.nn.functional as F +from torch import nn +from transformers import default_data_collator + +from examples.swag.pipeline import construct_roberta, get_swag_dataset +from kronfluence.analyzer import Analyzer, prepare_model +from kronfluence.arguments import FactorArguments, ScoreArguments +from kronfluence.task import Task +from kronfluence.utils.common.factor_arguments import all_low_precision_factor_arguments +from kronfluence.utils.common.score_arguments import all_low_precision_score_arguments +from kronfluence.utils.dataset import DataLoaderKwargs +from kronfluence.utils.model import apply_ddp + +BATCH_TYPE = Dict[str, torch.Tensor] +try: + LOCAL_RANK = int(os.environ["LOCAL_RANK"]) + WORLD_RANK = int(os.environ["RANK"]) + WORLD_SIZE = int(os.environ["WORLD_SIZE"]) +except KeyError: + LOCAL_RANK = WORLD_RANK = WORLD_SIZE = 0 + + +def parse_args(): + parser = argparse.ArgumentParser(description="Influence analysis on SWAG dataset.") + + parser.add_argument( + "--checkpoint_dir", + type=str, + default="./checkpoints", + help="A path that is storing the final checkpoint of the model.", + ) + + parser.add_argument( + "--factor_strategy", + type=str, + default="ekfac", + help="Strategy to compute influence factors.", + ) + parser.add_argument( + "--query_gradient_rank", + type=int, + default=-1, + help="Rank for the low-rank query gradient approximation.", + ) + parser.add_argument( + "--use_half_precision", + action="store_true", + default=False, + help="Whether to use half precision for computing factors and scores.", + ) + parser.add_argument( + "--use_ddp", + action="store_true", + default=False, + help="Whether to use DDP for computing factors and scores.", + ) + parser.add_argument( + "--factor_batch_size", + type=int, + default=128, + help="Batch size for computing influence factors.", + ) + parser.add_argument( + "--query_batch_size", + type=int, + default=100, + help="Batch size for computing query gradients.", + ) + parser.add_argument( + "--train_batch_size", + type=int, + default=128, + help="Batch size for computing training gradients.", + ) + parser.add_argument( + "--profile", + action="store_true", + default=False, + help="Boolean flag to profile computations.", + ) + args = parser.parse_args() + + if args.checkpoint_dir is not None: + os.makedirs(args.checkpoint_dir, exist_ok=True) + + return args + + +class MultipleChoiceTask(Task): + enable_post_process_per_sample_gradient = True + + def compute_train_loss( + self, + batch: BATCH_TYPE, + model: nn.Module, + sample: bool = False, + ) -> torch.Tensor: + logits = model( + input_ids=batch["input_ids"], + attention_mask=batch["attention_mask"], + ).logits + + if not sample: + return F.cross_entropy(logits, batch["labels"], reduction="sum") + with torch.no_grad(): + probs = torch.nn.functional.softmax(logits, dim=-1) + sampled_labels = torch.multinomial( + probs, + num_samples=1, + ).flatten() + return F.cross_entropy(logits, sampled_labels.detach(), reduction="sum") + + def compute_measurement( + self, + batch: BATCH_TYPE, + model: nn.Module, + ) -> torch.Tensor: + # Copied from: https://github.com/MadryLab/trak/blob/main/trak/modelout_functions.py. + logits = model( + input_ids=batch["input_ids"], + attention_mask=batch["attention_mask"], + ).logits + + labels = batch["labels"] + bindex = torch.arange(logits.shape[0]).to(device=logits.device, non_blocking=False) + logits_correct = logits[bindex, labels] + + cloned_logits = logits.clone() + cloned_logits[bindex, labels] = torch.tensor(-torch.inf, device=logits.device, dtype=logits.dtype) + + margins = logits_correct - cloned_logits.logsumexp(dim=-1) + return -margins.sum() + + def get_attention_mask(self, batch: BATCH_TYPE) -> torch.Tensor: + return batch["attention_mask"] + + def post_process_per_sample_gradient(self, module_name: str, gradient: torch.Tensor) -> torch.Tensor: + del module_name + total_batch_size = gradient.size(0) + true_batch_size = int(total_batch_size / 4) + return gradient.reshape(true_batch_size, 4, *gradient.size()[1:]).sum(dim=1) + + +def main(): + args = parse_args() + logging.basicConfig(level=logging.INFO) + + # Prepare the dataset. + train_dataset = get_swag_dataset( + split="eval_train", + ) + eval_dataset = get_swag_dataset( + split="valid", + ) + + # Prepare the trained model. + model = construct_roberta() + checkpoint_path = os.path.join(args.checkpoint_dir, "model.pth") + if not os.path.isfile(checkpoint_path): + raise ValueError(f"No checkpoint found at {checkpoint_path}.") + model.load_state_dict(torch.load(checkpoint_path)) + + # Define task and prepare model. + task = MultipleChoiceTask() + model = prepare_model(model, task) + + if args.use_ddp: + model = apply_ddp( + model=model, + local_rank=LOCAL_RANK, + rank=WORLD_RANK, + world_size=WORLD_SIZE, + ) + + analyzer = Analyzer( + analysis_name="swag", + model=model, + task=task, + profile=args.profile, + ) + # Configure parameters for DataLoader. + dataloader_kwargs = DataLoaderKwargs(collate_fn=default_data_collator) + analyzer.set_dataloader_kwargs(dataloader_kwargs) + + # Compute influence factors. + factors_name = args.factor_strategy + factor_args = FactorArguments(strategy=args.factor_strategy) + if args.use_half_precision: + factor_args = all_low_precision_factor_arguments(strategy=args.factor_strategy, dtype=torch.bfloat16) + factors_name += "_half" + if args.use_ddp: + factors_name += "_ddp" + analyzer.fit_all_factors( + factors_name=factors_name, + dataset=train_dataset, + per_device_batch_size=args.factor_batch_size, + factor_args=factor_args, + overwrite_output_dir=False, + ) + + # Compute pairwise scores. + score_args = ScoreArguments() + scores_name = factor_args.strategy + if args.use_half_precision: + score_args = all_low_precision_score_arguments(dtype=torch.bfloat16) + scores_name += "_half" + rank = args.query_gradient_rank if args.query_gradient_rank != -1 else None + if rank is not None: + score_args.query_gradient_low_rank = rank + score_args.query_gradient_accumulation_steps = 10 + scores_name += f"_qlr{rank}" + if args.use_ddp: + scores_name += "_ddp" + analyzer.compute_pairwise_scores( + score_args=score_args, + scores_name=scores_name, + factors_name=factors_name, + query_dataset=eval_dataset, + query_indices=list(range(min([len(eval_dataset), 2000]))), + train_dataset=train_dataset, + per_device_query_batch_size=args.query_batch_size, + per_device_train_batch_size=args.train_batch_size, + overwrite_output_dir=False, + ) + scores = analyzer.load_pairwise_scores(scores_name)["all_modules"] + logging.info(f"Scores shape: {scores.shape}") + + +if __name__ == "__main__": + main() diff --git a/examples/swag/evaluate_lds.py b/examples/swag/evaluate_lds.py new file mode 100644 index 0000000..6bb0196 --- /dev/null +++ b/examples/swag/evaluate_lds.py @@ -0,0 +1,68 @@ +import logging + +import numpy as np +import torch +import tqdm +from scipy.stats import spearmanr +from transformers import AutoTokenizer + +from examples.swag.pipeline import get_swag_dataset +from kronfluence.analyzer import Analyzer + + +def evaluate_correlations(scores: torch.Tensor) -> float: + margins = torch.from_numpy(torch.load(open("files/margins.pt", "rb"))) + masks = torch.from_numpy(torch.load(open("files/masks.pt", "rb"))).float() + + val_indices = np.arange(2000) + preds = masks @ scores.T + + rs = [] + ps = [] + for j in tqdm.tqdm(val_indices): + r, p = spearmanr(preds[:, j], margins[:, j]) + rs.append(r) + ps.append(p) + rs, ps = np.array(rs), np.array(ps) + return rs.mean() + + +def main(): + logging.basicConfig(level=logging.INFO) + + # You might need to change the path. + strategy = "ekfac" + scores = Analyzer.load_file(f"influence_results/swag/scores_{strategy}/pairwise_scores.safetensors")[ + "all_modules" + ].to(dtype=torch.float32) + + corr_mean = evaluate_correlations(scores) + logging.info(f"LDS: {np.mean(corr_mean)}") + + # We can also visualize the top influential sequences. + eval_idx = 1004 + tokenizer = AutoTokenizer.from_pretrained("FacebookAI/roberta-base", use_fast=True, trust_remote_code=True) + + train_dataset = get_swag_dataset( + split="eval_train", + ) + eval_dataset = get_swag_dataset( + split="valid", + ) + + print("Query Data Example:") + for i in range(4): + text = tokenizer.decode(eval_dataset[eval_idx]["input_ids"][i]) + print(f"Option {i}: {text}") + print(f"Label: {eval_dataset[eval_idx]['labels']}") + + top_idx = int(torch.argsort(scores[eval_idx], descending=True)[0]) + print("Top Influential Example:") + for i in range(4): + text = tokenizer.decode(train_dataset[top_idx]["input_ids"][i]) + print(f"Option {i}: {text}") + print(f"Label: {train_dataset[top_idx]['labels']}") + + +if __name__ == "__main__": + main() diff --git a/examples/swag/files/margins.pt b/examples/swag/files/margins.pt new file mode 100644 index 0000000..1e815d5 Binary files /dev/null and b/examples/swag/files/margins.pt differ diff --git a/examples/swag/files/masks.pt b/examples/swag/files/masks.pt new file mode 100644 index 0000000..653a6eb Binary files /dev/null and b/examples/swag/files/masks.pt differ diff --git a/examples/swag/influence_analysis.py b/examples/swag/influence_analysis.py new file mode 100644 index 0000000..d33a79a --- /dev/null +++ b/examples/swag/influence_analysis.py @@ -0,0 +1,41 @@ +import logging + +import matplotlib.pyplot as plt +import numpy as np +from scipy.stats import spearmanr +from tueplots import markers + +from kronfluence.analyzer import Analyzer + + +def main(): + logging.basicConfig(level=logging.INFO) + + scores1 = Analyzer.load_file("influence_results/swag/scores_ekfac/pairwise_scores.safetensors")[ + "all_modules" + ].float() + scores2 = Analyzer.load_file("influence_results/swag/scores_ekfac_half_qlr32/pairwise_scores.safetensors")[ + "all_modules" + ].float() + + plt.rcParams.update({"figure.dpi": 150}) + plt.rcParams.update(markers.with_edge()) + plt.rcParams["axes.axisbelow"] = True + + # Only plot first 6000 points to avoid clutter. + idx = 1 + plt.scatter(scores1[idx], scores2[idx], edgecolor="k") + plt.grid() + plt.xlabel("score1") + plt.ylabel("score2") + plt.show() + + # Compute the averaged spearman correlation. + all_corr = [] + for i in range(500): + all_corr.append(spearmanr(scores1[i], scores2[i])[0]) + logging.info(f"Averaged Spearman Correlation: {np.array(all_corr).mean()}") + + +if __name__ == "__main__": + main() diff --git a/examples/swag/pipeline.py b/examples/swag/pipeline.py new file mode 100644 index 0000000..aaa7225 --- /dev/null +++ b/examples/swag/pipeline.py @@ -0,0 +1,127 @@ +from dataclasses import dataclass +from itertools import chain +from typing import Any, List, Optional, Union + +import torch +from datasets import load_dataset +from torch import nn +from torch.utils.data import Dataset +from transformers import ( + AutoConfig, + AutoModelForMultipleChoice, + AutoTokenizer, + PreTrainedTokenizerBase, +) +from transformers.utils import PaddingStrategy + + +def construct_roberta() -> nn.Module: + config = AutoConfig.from_pretrained( + "FacebookAI/roberta-base", + trust_remote_code=True, + ) + return AutoModelForMultipleChoice.from_pretrained( + "FacebookAI/roberta-base", + config=config, + trust_remote_code=True, + ) + + +@dataclass +class DataCollatorForMultipleChoice: + tokenizer: PreTrainedTokenizerBase + padding: Union[bool, str, PaddingStrategy] = True + max_length: Optional[int] = None + pad_to_multiple_of: Optional[int] = None + + def __call__(self, features): + label_name = "label" if "label" in features[0].keys() else "labels" + labels = [feature.pop(label_name) for feature in features] + batch_size = len(features) + num_choices = len(features[0]["input_ids"]) + flattened_features = [ + [{k: v[i] for k, v in feature.items()} for i in range(num_choices)] for feature in features + ] + flattened_features = list(chain(*flattened_features)) + + batch = self.tokenizer.pad( + flattened_features, + padding=self.padding, + max_length=self.max_length, + pad_to_multiple_of=self.pad_to_multiple_of, + return_tensors="pt", + ) + + # Un-flatten + batch = {k: v.view(batch_size, num_choices, -1) for k, v in batch.items()} + # Add back labels + batch["labels"] = torch.tensor(labels, dtype=torch.int64) + return batch + + +def get_swag_dataset( + split: str, + indices: List[int] = None, +) -> Dataset: + assert split in ["train", "eval_train", "valid"] + + raw_datasets = load_dataset("swag", "regular") + + tokenizer = AutoTokenizer.from_pretrained("FacebookAI/roberta-base", use_fast=True, trust_remote_code=True) + + column_names = raw_datasets["train"].column_names + ending_names = [f"ending{i}" for i in range(4)] + context_name = "sent1" + question_header_name = "sent2" + label_column_name = "label" if "label" in column_names else "labels" + padding = "max_length" + + def preprocess_function(examples: Any): + first_sentences = [[context] * 4 for context in examples[context_name]] + question_headers = examples[question_header_name] + second_sentences = [ + [f"{header} {examples[end][i]}" for end in ending_names] for i, header in enumerate(question_headers) + ] + labels = examples[label_column_name] + + # Flatten out. + first_sentences = list(chain(*first_sentences)) + second_sentences = list(chain(*second_sentences)) + + # Tokenize. + tokenized_examples = tokenizer( + first_sentences, + second_sentences, + max_length=128, + padding=padding, + truncation=True, + ) + # Un-flatten. + tokenized_inputs = {k: [v[i : i + 4] for i in range(0, len(v), 4)] for k, v in tokenized_examples.items()} + tokenized_inputs["labels"] = labels + return tokenized_inputs + + processed_datasets = raw_datasets.map( + preprocess_function, + batched=True, + remove_columns=raw_datasets["train"].column_names, + ) + + if split in ["train", "eval_train"]: + dataset = processed_datasets["train"] + dataset = dataset.select(list(range(73_536))) + else: + dataset = processed_datasets["validation"] + dataset = dataset.select(list(range(2000))) + + if indices is not None: + dataset = dataset.select(indices) + + return dataset + + +if __name__ == "__main__": + from kronfluence import Analyzer + + model = construct_roberta() + print(Analyzer.get_module_summary(model)) diff --git a/examples/swag/requirements.txt b/examples/swag/requirements.txt new file mode 100644 index 0000000..7b0c55a --- /dev/null +++ b/examples/swag/requirements.txt @@ -0,0 +1,4 @@ +transformers +evaluate +datasets +scikit-learn \ No newline at end of file diff --git a/examples/swag/train.py b/examples/swag/train.py new file mode 100644 index 0000000..6254fbf --- /dev/null +++ b/examples/swag/train.py @@ -0,0 +1,169 @@ +import argparse +import logging +import os +import time +from typing import Tuple + +import evaluate +import torch +import torch.nn.functional as F +from accelerate.utils import set_seed +from torch import nn +from torch.utils import data +from transformers import default_data_collator + +from examples.swag.pipeline import construct_roberta, get_swag_dataset + +DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu") + + +def parse_args(): + parser = argparse.ArgumentParser(description="Train multiple-choice models on SWAG dataset.") + + parser.add_argument( + "--train_batch_size", + type=int, + default=16, + help="Batch size for the training dataloader.", + ) + parser.add_argument( + "--eval_batch_size", + type=int, + default=32, + help="Batch size for the evaluation dataloader.", + ) + + parser.add_argument( + "--learning_rate", + type=float, + default=1e-05, + help="Fixed learning rate to train the model.", + ) + parser.add_argument( + "--weight_decay", + type=float, + default=0.001, + help="Weight decay to train the model.", + ) + parser.add_argument( + "--num_train_epochs", + type=int, + default=3, + help="Total number of epochs to train the model.", + ) + + parser.add_argument( + "--seed", + type=int, + default=1004, + help="A seed for reproducible training pipeline.", + ) + parser.add_argument( + "--checkpoint_dir", + type=str, + default="./checkpoints", + help="A path to store the final checkpoint.", + ) + args = parser.parse_args() + + if args.checkpoint_dir is not None: + os.makedirs(args.checkpoint_dir, exist_ok=True) + + return args + + +def train( + dataset: data.Dataset, + batch_size: int, + num_train_epochs: int, + learning_rate: float, + weight_decay: float, +) -> nn.Module: + train_dataloader = data.DataLoader( + dataset=dataset, + batch_size=batch_size, + shuffle=True, + drop_last=True, + collate_fn=default_data_collator, + ) + + model = construct_roberta().to(DEVICE) + optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate, weight_decay=weight_decay) + + start_time = time.time() + model.train() + for epoch in range(num_train_epochs): + total_loss = 0.0 + for batch in train_dataloader: + optimizer.zero_grad(set_to_none=True) + loss = model( + input_ids=batch["input_ids"].to(device=DEVICE), + attention_mask=batch["attention_mask"].to(device=DEVICE), + labels=batch["labels"].to(device=DEVICE), + ).loss + loss.backward() + optimizer.step() + total_loss += loss.detach().float() + logging.info(f"Epoch {epoch + 1} - Averaged Loss: {total_loss / len(dataset)}") + end_time = time.time() + elapsed_time = end_time - start_time + logging.info(f"Completed training in {elapsed_time:.2f} seconds.") + return model + + +def evaluate_model(model: nn.Module, dataset: data.Dataset, batch_size: int) -> Tuple[float, float]: + dataloader = data.DataLoader( + dataset=dataset, batch_size=batch_size, shuffle=False, drop_last=False, collate_fn=default_data_collator + ) + + model.eval() + metric = evaluate.load("accuracy") + total_loss = 0.0 + for batch in dataloader: + with torch.no_grad(): + logits = model( + input_ids=batch["input_ids"].to(device=DEVICE), + attention_mask=batch["attention_mask"].to(device=DEVICE), + ).logits + labels = batch["labels"].to(device=DEVICE) + total_loss += F.cross_entropy(logits, labels, reduction="sum").detach() + predictions = logits.argmax(dim=-1) + metric.add_batch( + predictions=predictions, + references=labels, + ) + eval_metric = metric.compute() + return total_loss.item() / len(dataloader.dataset), eval_metric["accuracy"] + + +def main(): + args = parse_args() + logging.basicConfig(level=logging.INFO) + logger = logging.getLogger() + + if args.seed is not None: + set_seed(args.seed) + + train_dataset = get_swag_dataset(split="train") + model = train( + dataset=train_dataset, + batch_size=args.train_batch_size, + num_train_epochs=args.num_train_epochs, + learning_rate=args.learning_rate, + weight_decay=args.weight_decay, + ) + + eval_train_dataset = get_swag_dataset(split="eval_train") + train_loss, train_acc = evaluate_model(model=model, dataset=eval_train_dataset, batch_size=args.eval_batch_size) + logger.info(f"Train loss: {train_loss}, Train Accuracy: {train_acc}") + + eval_dataset = get_swag_dataset(split="valid") + eval_loss, eval_acc = evaluate_model(model=model, dataset=eval_dataset, batch_size=args.eval_batch_size) + logger.info(f"Evaluation loss: {eval_loss}, Evaluation Accuracy: {eval_acc}") + + if args.checkpoint_dir is not None: + torch.save(model.state_dict(), os.path.join(args.checkpoint_dir, "model.pth")) + + +if __name__ == "__main__": + main() diff --git a/examples/uci/README.md b/examples/uci/README.md index cfa2b1b..91e1f66 100644 --- a/examples/uci/README.md +++ b/examples/uci/README.md @@ -1,7 +1,8 @@ # UCI Regression Example -This directory contains scripts for training a regression model and conducting influence analysis with -datasets from the [UCI Machine Learning Repository](https://archive.ics.uci.edu/datasets). Please begin by installing necessary packages. +This directory contains scripts for training a regression model and conducting influence analysis using datasets from the [UCI Machine Learning Repository](https://archive.ics.uci.edu/datasets). +To get started, please install the necessary packages by running the following command: + ```bash pip install -r requirements.txt ``` @@ -9,6 +10,7 @@ pip install -r requirements.txt ## Training To train a regression model on the `Concrete` dataset, run the following command: + ```bash python train.py --dataset_name concrete \ --dataset_dir ./data \ @@ -16,26 +18,50 @@ python train.py --dataset_name concrete \ --train_batch_size 32 \ --eval_batch_size 1024 \ --learning_rate 0.03 \ - --weight_decay 1e-5 \ + --weight_decay 1e-05 \ --num_train_epochs 20 \ --seed 1004 ``` +This will train the model using the specified hyperparameters and save the trained checkpoint in the `./checkpoints` directory. + ## Computing Pairwise Influence Scores -To obtain a pairwise influence scores using `ekfac`, run the following command: +To compute pairwise influence scores using the `ekfac` strategy, run the following command: + ```bash python analyze.py --dataset_name concrete \ --dataset_dir ./data \ --checkpoint_dir ./checkpoints \ --factor_strategy ekfac ``` -You can also use `identity`, `diagonal`, and `kfac`. + +You can also use `identity`, `diagonal`, and `kfac` for `factor_strategy`. To measure the wall-clock time of computing influence scores, you can enable the `profile` flag: + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 11 | 0.28983 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Compute Pairwise Score | 0.15112 | 1 | 0.15112 | 52.141 | +| Fit Lambda | 0.094882 | 1 | 0.094882 | 32.737 | +| Fit Covariance | 0.031336 | 1 | 0.031336 | 10.812 | +| Perform Eigendecomposition | 0.006672 | 1 | 0.006672 | 2.302 | +| Save Eigendecomposition | 0.0013218 | 1 | 0.0013218 | 0.45607 | +| Save Covariance | 0.0013099 | 1 | 0.0013099 | 0.45196 | +| Load All Factors | 0.00073975 | 1 | 0.00073975 | 0.25523 | +| Save Lambda | 0.00073158 | 1 | 0.00073158 | 0.25242 | +| Load Covariance | 0.00062487 | 1 | 0.00062487 | 0.2156 | +| Save Pairwise Score | 0.00058254 | 1 | 0.00058254 | 0.20099 | +| Load Eigendecomposition | 0.00050846 | 1 | 0.00050846 | 0.17543 | +---------------------------------------------------------------------------------------------------------------------------------- +``` ## Counterfactual Evaluation -You can check the notebook `tutorial.ipynb` to run the subset removal counterfactual evaluation. -(Note that `TracIn` uses the final checkpoint instead of the intermediate checkpoints throughout training.) +To run the subset removal counterfactual evaluation, please refer to the `tutorial.ipynb` notebook. +Note that `TracIn` uses the final checkpoint instead of the intermediate checkpoints throughout training.

Counterfactual diff --git a/examples/uci/analyze.py b/examples/uci/analyze.py index 059234d..e612a77 100644 --- a/examples/uci/analyze.py +++ b/examples/uci/analyze.py @@ -37,6 +37,7 @@ def parse_args(): default="./checkpoints", help="A path that is storing the final checkpoint of the model.", ) + parser.add_argument( "--factor_strategy", type=str, @@ -80,7 +81,7 @@ def compute_measurement( model: nn.Module, ) -> torch.Tensor: # The measurement function is set as a training loss. Alternatively, we can - # use mean absolute error, as done in https://arxiv.org/abs/2405.12186. + # use the mean absolute error (MAE), as done in https://arxiv.org/abs/2405.12186. return self.compute_train_loss(batch, model, sample=False) @@ -117,10 +118,11 @@ def main(): analyzer.fit_all_factors( factors_name=args.factor_strategy, dataset=train_dataset, - per_device_batch_size=None, + per_device_batch_size=len(train_dataset), factor_args=factor_args, overwrite_output_dir=False, ) + # Compute pairwise scores. analyzer.compute_pairwise_scores( scores_name=args.factor_strategy, @@ -129,6 +131,7 @@ def main(): train_dataset=train_dataset, # Use full batch for computing query gradient. per_device_query_batch_size=len(eval_dataset), + per_device_train_batch_size=len(train_dataset), overwrite_output_dir=False, ) scores = analyzer.load_pairwise_scores(args.factor_strategy)["all_modules"] diff --git a/examples/uci/pipeline.py b/examples/uci/pipeline.py index bbb3eaf..769a171 100644 --- a/examples/uci/pipeline.py +++ b/examples/uci/pipeline.py @@ -8,8 +8,7 @@ from torch.utils.data import Dataset -def construct_regression_mlp() -> nn.Module: - num_inputs = 8 +def construct_regression_mlp(num_inputs: int = 8) -> nn.Module: model = torch.nn.Sequential( nn.Linear(num_inputs, 128, bias=True), nn.ReLU(), diff --git a/examples/uci/train.py b/examples/uci/train.py index 8a70077..ded0913 100644 --- a/examples/uci/train.py +++ b/examples/uci/train.py @@ -50,7 +50,7 @@ def parse_args(): parser.add_argument( "--weight_decay", type=float, - default=1e-5, + default=1e-05, help="Weight decay to train the model.", ) parser.add_argument( @@ -72,7 +72,6 @@ def parse_args(): default="./checkpoints", help="A path to store the final checkpoint.", ) - args = parser.parse_args() if args.checkpoint_dir is not None: @@ -96,6 +95,7 @@ def train( batch_size=batch_size, shuffle=True, drop_last=True, + num_workers=0, ) model = construct_regression_mlp() optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate, weight_decay=weight_decay) @@ -106,7 +106,7 @@ def train( with tqdm(train_dataloader, unit="batch", disable=disable_tqdm) as tepoch: for batch in tepoch: tepoch.set_description(f"Epoch {epoch}") - model.zero_grad() + optimizer.zero_grad(set_to_none=True) inputs, targets = batch outputs = model(inputs) loss = F.mse_loss(outputs, targets) @@ -123,6 +123,7 @@ def evaluate(model: nn.Module, dataset: data.Dataset, batch_size: int) -> float: batch_size=batch_size, shuffle=False, drop_last=False, + num_workers=0, ) model.eval() diff --git a/examples/uci/tutorial.ipynb b/examples/uci/tutorial.ipynb index 789fd6d..7160ce8 100644 --- a/examples/uci/tutorial.ipynb +++ b/examples/uci/tutorial.ipynb @@ -7,7 +7,7 @@ "source": [ "# Kronfluence Tutorial\n", "\n", - "Kronfluence is a repository designed to compute influence functions using Kronecker-factored Approximate Curvature (KFAC) or Eigenvalue-corrected KFAC (EKFAC). In this short tutorial, we will introduce some functionalities of Kronfluence on the UCI regression pipeline (it is quick to run and does not require GPUs)." + "Kronfluence is a repository designed to compute influence functions using Kronecker-factored Approximate Curvature (KFAC) or Eigenvalue-corrected KFAC (EKFAC). In this short tutorial, we will introduce some functionalities of Kronfluence on the UCI regression pipeline (it is quick to run and does not require any GPUs)." ] }, { @@ -16,25 +16,23 @@ "id": "initial_id", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T10:46:20.005159Z", - "start_time": "2024-03-12T10:46:19.995640Z" + "end_time": "2024-06-24T17:25:49.774020Z", + "start_time": "2024-06-24T17:25:48.659370Z" } }, "outputs": [], "source": [ "import math\n", - "import random\n", "from typing import List, Optional, Tuple\n", "\n", "import matplotlib.pyplot as plt\n", + "import random\n", "import torch\n", "import torch.nn.functional as F\n", "from random import shuffle\n", "from torch import nn\n", "from torch.utils import data\n", - "from tueplots import cycler, markers\n", - "from tueplots.constants import markers as marker_constants\n", - "from tueplots.constants.color import palettes\n", + "from tueplots import markers\n", "from tqdm import tqdm\n", "\n", "from examples.uci.pipeline import construct_regression_mlp, get_regression_dataset\n", @@ -65,8 +63,8 @@ "id": "4e56f0f1d6e34e62", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T10:47:03.741170Z", - "start_time": "2024-03-12T10:47:02.235222Z" + "end_time": "2024-06-24T17:25:51.812301Z", + "start_time": "2024-06-24T17:25:51.807132Z" }, "collapsed": false, "jupyter": { @@ -96,12 +94,17 @@ "cell_type": "code", "execution_count": 3, "id": "3ec0486b-4abe-461e-a258-8d24d9c1b10b", - "metadata": {}, + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:25:53.198387Z", + "start_time": "2024-06-24T17:25:53.190103Z" + } + }, "outputs": [ { "data": { "text/plain": [ - "((927, 8), (927, 1))" + "((927, 8), (103, 1))" ] }, "execution_count": 3, @@ -112,14 +115,19 @@ "source": [ "train_dataset = get_regression_dataset(data_name=dataset_name, split=\"train\", dataset_dir=dataset_dir)\n", "query_dataset = get_regression_dataset(data_name=dataset_name, split=\"valid\", dataset_dir=dataset_dir)\n", - "train_dataset.data_x.shape, train_dataset.data_y.shape" + "train_dataset.data_x.shape, query_dataset.data_y.shape" ] }, { "cell_type": "code", "execution_count": 4, "id": "54b2c746-0432-4917-aeb7-17e91e72553b", - "metadata": {}, + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:25:54.066595Z", + "start_time": "2024-06-24T17:25:54.059246Z" + } + }, "outputs": [ { "data": { @@ -148,7 +156,12 @@ "cell_type": "code", "execution_count": 5, "id": "d8f275b5-7767-4d26-84ff-8a39aa74df08", - "metadata": {}, + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:25:54.861487Z", + "start_time": "2024-06-24T17:25:54.856620Z" + } + }, "outputs": [], "source": [ "def train(\n", @@ -157,7 +170,8 @@ " num_train_epochs: int,\n", " learning_rate: float,\n", " weight_decay: float,\n", - " disable_tqdm: bool = False,\n", + " seed: int = 0,\n", + " disable_tqdm: bool = True,\n", ") -> nn.Module:\n", " train_dataloader = data.DataLoader(\n", " dataset=dataset,\n", @@ -165,6 +179,8 @@ " shuffle=True,\n", " drop_last=True,\n", " )\n", + " torch.manual_seed(seed)\n", + " random.seed(seed)\n", " model = construct_regression_mlp()\n", " optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate, weight_decay=weight_decay)\n", "\n", @@ -174,7 +190,7 @@ " with tqdm(train_dataloader, unit=\"batch\", disable=disable_tqdm) as tepoch:\n", " for batch in tepoch:\n", " tepoch.set_description(f\"Epoch {epoch}\")\n", - " model.zero_grad()\n", + " optimizer.zero_grad(set_to_none=True)\n", " inputs, targets = batch\n", " outputs = model(inputs)\n", " loss = F.mse_loss(outputs, targets)\n", @@ -189,52 +205,57 @@ "cell_type": "code", "execution_count": 6, "id": "6a1337cb-6fe5-4edc-9821-de27bbef9074", - "metadata": {}, + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:25:58.127189Z", + "start_time": "2024-06-24T17:25:55.552201Z" + } + }, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ - "Epoch 0: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 433.68batch/s, loss=0.942]\n", - "Epoch 1: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 578.97batch/s, loss=0.663]\n", - "Epoch 2: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 507.52batch/s, loss=0.436]\n", - "Epoch 3: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 590.30batch/s, loss=0.351]\n", - "Epoch 4: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 509.56batch/s, loss=0.293]\n", - "Epoch 5: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 474.25batch/s, loss=0.25]\n", - "Epoch 6: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 537.15batch/s, loss=0.222]\n", - "Epoch 7: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 644.86batch/s, loss=0.206]\n", - "Epoch 8: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 741.23batch/s, loss=0.185]\n", - "Epoch 9: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 669.30batch/s, loss=0.194]\n", - "Epoch 10: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 483.44batch/s, loss=0.17]\n", - "Epoch 11: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 644.04batch/s, loss=0.177]\n", - "Epoch 12: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 578.84batch/s, loss=0.158]\n", - "Epoch 13: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 506.81batch/s, loss=0.173]\n", - "Epoch 14: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 546.14batch/s, loss=0.145]\n", - "Epoch 15: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 130.30batch/s, loss=0.136]\n", - "Epoch 16: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 208.50batch/s, loss=0.14]\n", - "Epoch 17: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 584.80batch/s, loss=0.141]\n", - "Epoch 18: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 620.74batch/s, loss=0.133]\n", - "Epoch 19: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 740.27batch/s, loss=0.142]\n", - "Epoch 20: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 676.98batch/s, loss=0.115]\n", - "Epoch 21: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 502.51batch/s, loss=0.117]\n", - "Epoch 22: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 545.23batch/s, loss=0.117]\n", - "Epoch 23: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 638.29batch/s, loss=0.109]\n", - "Epoch 24: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 626.93batch/s, loss=0.112]\n", - "Epoch 25: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 611.66batch/s, loss=0.106]\n", - "Epoch 26: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 621.82batch/s, loss=0.103]\n", - "Epoch 27: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 633.66batch/s, loss=0.105]\n", - "Epoch 28: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 493.41batch/s, loss=0.0992]\n", - "Epoch 29: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 634.15batch/s, loss=0.101]\n", - "Epoch 30: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 749.29batch/s, loss=0.101]\n", - "Epoch 31: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 575.25batch/s, loss=0.0898]\n", - "Epoch 32: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 502.94batch/s, loss=0.0888]\n", - "Epoch 33: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 605.97batch/s, loss=0.0852]\n", - "Epoch 34: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 131.50batch/s, loss=0.0891]\n", - "Epoch 35: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 688.25batch/s, loss=0.0881]\n", - "Epoch 36: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 676.97batch/s, loss=0.0883]\n", - "Epoch 37: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 713.74batch/s, loss=0.0779]\n", - "Epoch 38: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 528.26batch/s, loss=0.0904]\n", - "Epoch 39: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 605.93batch/s, loss=0.0747]\n" + "Epoch 0: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 389.33batch/s, loss=0.913]\n", + "Epoch 1: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 336.95batch/s, loss=0.629]\n", + "Epoch 2: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 538.93batch/s, loss=0.429]\n", + "Epoch 3: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 376.75batch/s, loss=0.352]\n", + "Epoch 4: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 484.09batch/s, loss=0.29]\n", + "Epoch 5: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 545.75batch/s, loss=0.253]\n", + "Epoch 6: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 363.86batch/s, loss=0.222]\n", + "Epoch 7: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 528.11batch/s, loss=0.203]\n", + "Epoch 8: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 578.23batch/s, loss=0.2]\n", + "Epoch 9: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 579.47batch/s, loss=0.18]\n", + "Epoch 10: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 575.52batch/s, loss=0.16]\n", + "Epoch 11: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 580.09batch/s, loss=0.157]\n", + "Epoch 12: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 560.29batch/s, loss=0.15]\n", + "Epoch 13: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 371.33batch/s, loss=0.132]\n", + "Epoch 14: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 555.09batch/s, loss=0.161]\n", + "Epoch 15: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 524.96batch/s, loss=0.14]\n", + "Epoch 16: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 559.10batch/s, loss=0.121]\n", + "Epoch 17: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 549.66batch/s, loss=0.128]\n", + "Epoch 18: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 534.94batch/s, loss=0.13]\n", + "Epoch 19: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 360.81batch/s, loss=0.116]\n", + "Epoch 20: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 583.09batch/s, loss=0.11]\n", + "Epoch 21: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 563.09batch/s, loss=0.119]\n", + "Epoch 22: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 612.03batch/s, loss=0.112]\n", + "Epoch 23: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 561.09batch/s, loss=0.1]\n", + "Epoch 24: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 396.01batch/s, loss=0.119]\n", + "Epoch 25: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 583.49batch/s, loss=0.0916]\n", + "Epoch 26: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 589.11batch/s, loss=0.0929]\n", + "Epoch 27: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 547.42batch/s, loss=0.0956]\n", + "Epoch 28: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 353.97batch/s, loss=0.0995]\n", + "Epoch 29: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 381.87batch/s, loss=0.093]\n", + "Epoch 30: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 509.93batch/s, loss=0.0958]\n", + "Epoch 31: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 543.89batch/s, loss=0.0921]\n", + "Epoch 32: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 609.94batch/s, loss=0.0898]\n", + "Epoch 33: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 377.88batch/s, loss=0.0867]\n", + "Epoch 34: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 372.61batch/s, loss=0.0854]\n", + "Epoch 35: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 352.03batch/s, loss=0.0799]\n", + "Epoch 36: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 336.11batch/s, loss=0.0821]\n", + "Epoch 37: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 437.86batch/s, loss=0.0751]\n", + "Epoch 38: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 449.35batch/s, loss=0.079]\n", + "Epoch 39: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 455.28batch/s, loss=0.0838]\n" ] } ], @@ -245,6 +266,8 @@ " num_train_epochs=num_train_epochs,\n", " learning_rate=learning_rate,\n", " weight_decay=weight_decay,\n", + " seed=0,\n", + " disable_tqdm=False,\n", ")" ] }, @@ -262,8 +285,8 @@ "id": "6dc3ab20b6cb4050", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T21:02:46.505618Z", - "start_time": "2024-03-12T21:02:46.503015Z" + "end_time": "2024-06-24T17:26:08.099368Z", + "start_time": "2024-06-24T17:26:08.079408Z" }, "collapsed": false, "jupyter": { @@ -274,7 +297,7 @@ { "data": { "text/plain": [ - "0.14518585019898647" + "0.13899902232642314" ] }, "execution_count": 7, @@ -325,7 +348,12 @@ "cell_type": "code", "execution_count": 8, "id": "03dad2b0-dd52-47ec-92d1-503914f78951", - "metadata": {}, + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:26:22.614833Z", + "start_time": "2024-06-24T17:26:22.470422Z" + } + }, "outputs": [], "source": [ "from kronfluence.analyzer import Analyzer, prepare_model\n", @@ -344,7 +372,12 @@ "cell_type": "code", "execution_count": 9, "id": "97fdf9f1-5296-4e08-916a-1f593990e620", - "metadata": {}, + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:26:26.178579Z", + "start_time": "2024-06-24T17:26:26.176217Z" + } + }, "outputs": [ { "name": "stdout", @@ -368,8 +401,8 @@ "id": "41238e1b9bcec5cc", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T20:58:22.611435Z", - "start_time": "2024-03-12T20:58:22.608172Z" + "end_time": "2024-06-24T17:26:46.440030Z", + "start_time": "2024-06-24T17:26:46.399529Z" }, "collapsed": false, "jupyter": { @@ -391,8 +424,8 @@ " return F.mse_loss(outputs, targets, reduction=\"sum\")\n", " # Sample the outputs from the model's prediction for true Fisher.\n", " with torch.no_grad():\n", - " sampled_targets = torch.normal(outputs, std=math.sqrt(0.5))\n", - " return F.mse_loss(outputs, sampled_targets.detach(), reduction=\"sum\")\n", + " sampled_targets = torch.normal(outputs.detach(), std=math.sqrt(0.5))\n", + " return F.mse_loss(outputs, sampled_targets, reduction=\"sum\")\n", "\n", " def compute_measurement(\n", " self,\n", @@ -421,8 +454,8 @@ "id": "1192b28fd4535410", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T20:58:24.595047Z", - "start_time": "2024-03-12T20:58:24.588848Z" + "end_time": "2024-06-24T17:26:49.591961Z", + "start_time": "2024-06-24T17:26:49.586269Z" }, "collapsed": false, "jupyter": { @@ -449,8 +482,8 @@ "id": "4a38704b80ff26b2", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T20:59:23.401098Z", - "start_time": "2024-03-12T20:59:23.396531Z" + "end_time": "2024-06-24T17:26:51.843577Z", + "start_time": "2024-06-24T17:26:51.827349Z" }, "collapsed": false, "jupyter": { @@ -489,6 +522,35 @@ "model" ] }, + { + "cell_type": "markdown", + "id": "e3d5d070-2cc0-4d16-afa8-5468040e0b03", + "metadata": {}, + "source": [ + "However, the model's behavior did not change." + ] + }, + { + "cell_type": "code", + "execution_count": 13, + "id": "a8044b0a-466a-4bde-8952-aad0416bc525", + "metadata": {}, + "outputs": [ + { + "data": { + "text/plain": [ + "0.13899902232642314" + ] + }, + "execution_count": 13, + "metadata": {}, + "output_type": "execute_result" + } + ], + "source": [ + "evaluate(model=model, dataset=query_dataset, batch_size=eval_batch_size)" + ] + }, { "cell_type": "markdown", "id": "190902fc-fe59-42c7-a94f-07c577dedbf7", @@ -499,9 +561,14 @@ }, { "cell_type": "code", - "execution_count": 13, + "execution_count": 14, "id": "19d3800e-a518-4bfe-a0b1-02ebf16b8f3e", - "metadata": {}, + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:27:02.310043Z", + "start_time": "2024-06-24T17:27:02.306047Z" + } + }, "outputs": [], "source": [ "analyzer = Analyzer(\n", @@ -509,6 +576,7 @@ " model=model,\n", " task=task,\n", " cpu=True,\n", + " disable_tqdm=True,\n", ")" ] }, @@ -530,28 +598,19 @@ }, { "cell_type": "code", - "execution_count": 14, + "execution_count": 15, "id": "87964b32dd38d4ff", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T20:59:08.113601Z", - "start_time": "2024-03-12T20:59:08.079505Z" + "end_time": "2024-06-24T17:27:06.262977Z", + "start_time": "2024-06-24T17:27:06.118615Z" }, "collapsed": false, "jupyter": { "outputs_hidden": false } }, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "Fitting covariance matrices [1/1] 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n", - "Fitting covariance matrices [1/1] 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n" - ] - } - ], + "outputs": [], "source": [ "analyzer.fit_covariance_matrices(\n", " factors_name=\"tutorial_factor\",\n", @@ -571,12 +630,12 @@ }, { "cell_type": "code", - "execution_count": 15, + "execution_count": 16, "id": "481af2c88737df72", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T20:59:59.570656Z", - "start_time": "2024-03-12T20:59:59.567837Z" + "end_time": "2024-06-24T17:27:09.389379Z", + "start_time": "2024-06-24T17:27:09.160193Z" }, "collapsed": false, "jupyter": { @@ -587,16 +646,16 @@ { "data": { "text/plain": [ - "" + "" ] }, - "execution_count": 15, + "execution_count": 16, "metadata": {}, "output_type": "execute_result" }, { "data": { - "image/png": "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", + "image/png": "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", "text/plain": [ "

" ] @@ -621,27 +680,19 @@ }, { "cell_type": "code", - "execution_count": 16, + "execution_count": 17, "id": "31da64ca7f237819", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T21:00:31.570596Z", - "start_time": "2024-03-12T21:00:31.553063Z" + "end_time": "2024-06-24T17:27:12.773562Z", + "start_time": "2024-06-24T17:27:12.700998Z" }, "collapsed": false, "jupyter": { "outputs_hidden": false } }, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "Performing Eigendecomposition [4/4] 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n" - ] - } - ], + "outputs": [], "source": [ "analyzer.perform_eigendecomposition(\n", " factors_name=\"tutorial_factor\",\n", @@ -654,33 +705,24 @@ "id": "39c411f2-c1c0-4483-9ec8-2d7ca966237a", "metadata": {}, "source": [ - "Next, we can fit the Lambda (corrected-eigenvalues for EKFAC) matrices with:" + "Next, we can fit the Lambda (corrected eigenvalues for EKFAC) matrices with:" ] }, { "cell_type": "code", - "execution_count": 17, + "execution_count": 18, "id": "582867d0b427db3", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T21:01:00.458530Z", - "start_time": "2024-03-12T21:01:00.113019Z" + "end_time": "2024-06-24T17:27:16.373321Z", + "start_time": "2024-06-24T17:27:16.096052Z" }, "collapsed": false, "jupyter": { "outputs_hidden": false } }, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "Fitting Lambda matrices [1/1] 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n", - "Fitting Lambda matrices [1/1] 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n" - ] - } - ], + "outputs": [], "source": [ "analyzer.fit_lambda_matrices(\n", " factors_name=\"tutorial_factor\",\n", @@ -700,22 +742,15 @@ }, { "cell_type": "code", - "execution_count": 18, + "execution_count": 19, "id": "0f00c60c-c932-42f4-b72f-2a59025c9417", - "metadata": {}, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "Fitting covariance matrices [1/1] 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n", - "Fitting covariance matrices [1/1] 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n", - "Performing Eigendecomposition [4/4] 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n", - "Fitting Lambda matrices [1/1] 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n", - "Fitting Lambda matrices [1/1] 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n" - ] + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:27:20.643234Z", + "start_time": "2024-06-24T17:27:20.158931Z" } - ], + }, + "outputs": [], "source": [ "analyzer.fit_all_factors(\n", " factors_name=\"tutorial_factor\",\n", @@ -738,37 +773,24 @@ "id": "b57f42e9-ff52-4d13-8493-e00f7fcf3425", "metadata": {}, "source": [ - "After computing all neccessary factors, we now compute the pairwise influence scores." + "After computing all necessary factors, we now compute the pairwise influence scores." ] }, { "cell_type": "code", - "execution_count": 19, + "execution_count": 20, "id": "8bdb3aa3d8aade4a", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T21:01:40.623332Z", - "start_time": "2024-03-12T21:01:40.490302Z" + "end_time": "2024-06-24T17:27:39.272804Z", + "start_time": "2024-06-24T17:27:38.807024Z" }, "collapsed": false, "jupyter": { "outputs_hidden": false } }, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "Computing pairwise scores (query gradient) [0/1] 0%| [time left: ?, time spent: 00:00]\n", - "Computing pairwise scores (training gradient) [1/1] 100%|███████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\u001b[A\n", - "Computing pairwise scores (query gradient) [1/1] 100%|██████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n", - "Computing pairwise scores (query gradient) [0/1] 0%| [time left: ?, time spent: 00:00]\n", - "Computing pairwise scores (training gradient) [1/1] 100%|███████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\u001b[A\n", - "Computing pairwise scores (query gradient) [1/1] 100%|██████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n" - ] - } - ], + "outputs": [], "source": [ "analyzer.compute_pairwise_scores(\n", " scores_name=\"tutorial_score\",\n", @@ -790,12 +812,12 @@ }, { "cell_type": "code", - "execution_count": 20, + "execution_count": 21, "id": "51d4f1a9ad39cfab", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T21:01:47.451459Z", - "start_time": "2024-03-12T21:01:47.448032Z" + "end_time": "2024-06-24T17:27:44.739399Z", + "start_time": "2024-06-24T17:27:44.716083Z" }, "collapsed": false, "jupyter": { @@ -809,7 +831,7 @@ "torch.Size([103, 927])" ] }, - "execution_count": 20, + "execution_count": 21, "metadata": {}, "output_type": "execute_result" } @@ -824,32 +846,24 @@ "id": "5c914f47-4cf8-47d6-a638-1d0cf10357ed", "metadata": {}, "source": [ - "If you would like to obtain the influence score for each module, you can pass in `ScoreArguments` to `compute_pairwise_scores`." + "If you would like to obtain the influence score for each module, you can pass in `ScoreArguments` with `per_module_score=True` to `compute_pairwise_scores`." ] }, { "cell_type": "code", - "execution_count": 21, + "execution_count": 22, "id": "a3a7267f-664f-42c5-9a0f-4f462c5a0bb9", - "metadata": {}, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "Computing pairwise scores (query gradient) [0/1] 0%| [time left: ?, time spent: 00:00]\n", - "Computing pairwise scores (training gradient) [1/1] 100%|███████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\u001b[A\n", - "Computing pairwise scores (query gradient) [1/1] 100%|██████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n", - "Computing pairwise scores (query gradient) [0/1] 0%| [time left: ?, time spent: 00:00]\n", - "Computing pairwise scores (training gradient) [1/1] 100%|███████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\u001b[A\n", - "Computing pairwise scores (query gradient) [1/1] 100%|██████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n" - ] + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:27:47.990045Z", + "start_time": "2024-06-24T17:27:47.519639Z" } - ], + }, + "outputs": [], "source": [ "from kronfluence import ScoreArguments\n", "\n", - "score_args = ScoreArguments(per_module_score=True)\n", + "score_args = ScoreArguments(compute_per_module_scores=True)\n", "analyzer.compute_pairwise_scores(\n", " score_args=score_args,\n", " scores_name=\"tutorial_per_module_score\",\n", @@ -863,9 +877,14 @@ }, { "cell_type": "code", - "execution_count": 22, + "execution_count": 23, "id": "0156e383-a2da-4ea4-bccf-a0183c78e712", - "metadata": {}, + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:27:49.238683Z", + "start_time": "2024-06-24T17:27:49.232054Z" + } + }, "outputs": [ { "data": { @@ -873,7 +892,7 @@ "dict_keys(['0', '2', '4', '6'])" ] }, - "execution_count": 22, + "execution_count": 23, "metadata": {}, "output_type": "execute_result" } @@ -888,28 +907,33 @@ "id": "95a869ff-70f0-49a4-a8f3-92f5c81a8ecf", "metadata": {}, "source": [ - "We can also visualize the score matrix for the last module." + "We can visualize the score matrix for the last module." ] }, { "cell_type": "code", - "execution_count": 23, + "execution_count": 24, "id": "065666a3-6dc5-49dc-9dff-d0ca3886a619", - "metadata": {}, + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:27:51.004179Z", + "start_time": "2024-06-24T17:27:50.862243Z" + } + }, "outputs": [ { "data": { "text/plain": [ - "" + "" ] }, - "execution_count": 23, + "execution_count": 24, "metadata": {}, "output_type": "execute_result" }, { "data": { - "image/png": "iVBORw0KGgoAAAANSUhEUgAABugAAAEZCAYAAAB1paMzAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjguMywgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/H5lhTAAAACXBIWXMAABcSAAAXEgFnn9JSAAEAAElEQVR4nOz9ebgmRZUnjn8i3+VuRdWlimJfhGKTUhC0sQt+1SCNiiigDm7TPgOl7TJqK+IyPSMqov040wiKgo1jCziDOygttG2LirRf2RRQhhLZKWSTqqK2W3d575sZvz8yI/JE5InMfN97b92q8nx43odbmRkRJyJOnHMi4sQJpbXWEAgEAoFAIBAIBAKBQCAQCAQCgUAgEAgE2wTRfBMgEAgEAoFAIBAIBAKBQCAQCAQCgUAgEPw5QTboBAKBQCAQCAQCgUAgEAgEAoFAIBAIBIJtCNmgEwgEAoFAIBAIBAKBQCAQCAQCgUAgEAi2IWSDTiAQCAQCgUAgEAgEAoFAIBAIBAKBQCDYhpANOoFAIBAIBAKBQCAQCAQCgUAgEAgEAoFgG0I26AQCgUAgEAgEAoFAIBAIBAKBQCAQCASCbQjZoBMIBAKBQCAQCAQCgUAgEAgEAoFAIBAItiFkg04gEAgEAoFAIBAIBAKBQCAQCAQCgUAg2IaQDTqBQCAQCAQCgUAgEAgEAoFAIBAIBAKBYBtCNugEAoFAIBAIBAKBQCAQCAQCgUAgEAgEgm0I2aATCAQCgUAgEAgEAoFAIBAIBAKBQCAQCLYhZINOIBAIBAKBQCAQCAQCgUAgEAgEAoFAINiGkA06gUAgEAgEAoFAIBAIBAKBQCAQCAQCgWAbQjboBAKBQCAQCAQCgUAgEAgEAoFAIBAIBIJtCNmgEwgEAoFAIBAIBAKBQCAQCAQCgUAgEDiYmJjAJz7xCRx66KEYHBzE3nvvjbe97W148skne85rw4YN+MAHPoADDjgAAwMDOOCAA3DOOedg48aN7Pdnn302lFLB3+WXXx4s6/rrr8cJJ5yAhQsXYuHChTjxxBPxr//6rz3TPNdQWms930QIBAKBQCAQCAQCgUAgEAgEAoFAIBAItg9MTk7iZS97GW677TbstddeWLlyJR577DHccccdWLp0KW677TYcdNBBtfJat24dVqxYgYceeggHHXQQXvKSl2D16tVYvXo1Dj30UNx6661YvHixk+bss8/G17/+dbzyla/EnnvuWcjzrLPOwste9rLC8y984Qv44Ac/iGaziZNPPhkDAwP4yU9+gomJCXzpS1/C+973vv4aZA4gG3QCgUAgEAgEAoFAIBAIBAKBQCAQCAQCi/POOw//8A//gBUrVuAnP/kJFixYAAC4+OKL8aEPfQgnnHACfvGLX9TK661vfSu+8Y1v4PWvfz2+853voNlsAgDe//7340tf+hLOOussXHXVVU4as0F300034cQTT6xVzv3334/ly5ej2WzipptuwooVKwAADzzwAI477jhs2rQJ9913Hw4++OBa+c01JMSlQCAQCAQCgUAgEAgEAoFAIBAIBAKBAADQ6XRw6aWXAgAuu+wyuzkHAOeeey6OPPJI3Hzzzbjzzjsr83r66afxrW99C+12G1/+8pft5hwAXHjhhVi6dCmuvvpqPPvsszOm+5JLLkEcx3j3u99tN+cA4NBDD8XHPvYxdLtdXHLJJTMuZ7YgG3QCgUAgEAgEAoFAIBAIBAKBQCAQCAQCAMCvfvUrbNq0CcuWLcPRRx9deH/mmWcCSO96q8KPf/xjJEmClStXYo899nDeDQwM4LTTTkMcx/jRj340Y7rNPXOGvn5p3lZoVn8iEAgEAoFAIBAIBAKBQCAQCAQCgUAg+HPA7373OwDAMcccw743z++5555ZyeuKK64I5vX9738f1157LeI4xoEHHojTTjsNhx9+eOG7jRs34vHHHwcAdlNxv/32w2677YY1a9Zg8+bNWLhwYSXtcw05QUcwMTGBT3ziEzj00EMxODiIvffeG29729vw5JNPzjdpAoFAUAvj4+O47rrr8Pa3vx2HHXYYBgcHMTIygqOOOgoXXHABxsbGgmmvuuoqHHvssViwYAEWL16MU089Fbfccktpeb/61a9w6qmnYvHixViwYAGOPfZY/J//839mu1oCgUDQF9avX4/dd98dSqnK+PIiAwUCwc6EtWvX4sMf/jAOO+wwDA0NYfHixTjmmGPwkY98hP3++uuvxwknnICFCxdi4cKFOPHEE633cQirV6/GG97wBixduhRDQ0N44QtfiC984QtIkmQuqiQQCAS18Otf/xpvfOMbsffee6PVamF0dBQrV67ElVdeCa114fs4jvH5z38eL3zhCzE0NISlS5fijW98I+67777ScvqRmwKBQNAvTj/9dCxfvryn3+mnnz6jMs1G17777su+N8/XrFkz53l96UtfwqWXXop/+qd/wkc/+lEcccQReO9734tut8uWs+uuu2JkZGTGdG8LyAZdhsnJSZx00kn49Kc/jbGxMZxxxhnYb7/9cOWVV+Loo4/GI488Mt8kCgQCQSW++c1v4nWvex2uuOIKNBoNnH766Vi5ciUeffRRfPKTn8Rf/MVfsPGczznnHKxatQr33nsvTj75ZBx77LG48cYb8Vd/9Ve47rrr2LKuvfZanHDCCfjxj3+MI488EqeccgoefPBBnHXWWfjwhz88xzUVCASCanzoQx/CunXrKr8TGSgQCHYm3HnnnXj+85+Piy66CK1WC2eccQb+8i//Es899xw+//nPF77/whe+gNNPPx233HILjj/+eJx00km444478JrXvMbeO+Lj1ltvxV/8xV/gmmuuwUEHHYTTTz8d69atwwc/+EG8+c1vZhfBBQKBYK5x7bXXYsWKFfje976HvfbaC69//etxzDHH4LbbbsPb3vY2vPWtb3W+T5IEb3jDG3DuuefiiSeewKtf/WosX74c11xzDV7ykpfgjjvuYMvpR24KBALBTPDwww/joYfuA7oP1fo99NB9+MlPfhLcvKsD4+Q/PDzMvjcbYFu2bJmzvI4++mhcfvnleOCBBzA+Po5HHnkEl112GUZHR/HlL3+54HxWVU6vdG8TaIHWWuuPfexjGoBesWKF3rJli31+0UUXaQD6hBNOmD/iBAKBoCauuuoq/c53vlP//ve/d54/9dRT+uijj9YA9Fve8hbn3Y033qgB6CVLlugHHnjAPr/lllt0u93Wo6OjesOGDU6a9evX64ULF2oA+tprr7XPn3nmGX3wwQdrAPqmm26a9foJBAJBXfz0pz/VAPQ73/lODUAvW7aM/U5koEAg2Jnw7LPP6t12200PDw/rf/mXfym8v/32251//+EPf9CNRkMPDAzoW265xT6///779ZIlS3Sz2dQPPvigk6bT6egDDzxQA9AXX3yxfb5lyxa9YsUKDUBfeeWVs1sxgUAgqMD09LTefffdNQD9jW98w3n3+9//Xi9evFgD0D//+c/t869+9asagD7kkEP0M888Y59fc801GoA++OCD9fT0tJNXP3JTIBAIZoojjjhCP//Qlp586sBav+cf2tIDAwP6iCOOYH918I53vEMD0B/72MfY9w8++KCVoVV4+ctfrgHor371q+x7My9/+ctfXou2e++9V7fbbd1sNvXjjz9un//qV7/SAPQ+++wTTHv88cdrAPpXv/pVrbLmGnKCDkCn07EeLpdddhkWLFhg35177rk48sgjcfPNN+POO++cLxIFAoGgFs466yx85StfwfOf/3zn+V577YXLLrsMQBq3udPp2HcXX3wxAOC8887DIYccYp+vWLEC7373u7Fx40Z87Wtfc/L753/+Z2zevBlnnHEGXv/619vne+yxB/7xH/8RAHDRRRfNbuUEAoGgJiYmJvCud70LRxxxROVpNpGBAoFgZ8InP/lJrFu3DhdeeCEb1ujYY491/n3JJZcgjmO8+93vxooVK+zzQw89FB/72MfQ7XZxySWXOGl+8IMf4NFHH8VRRx2FD37wg/b5ggUL7LxaZKBAINjW+MMf/oBnn30Whx12GP7zf/7PzrvnP//59vTcr3/9a/vc2IH/+I//iD322MM+/0//6T/h9NNPx0MPPYR/+Zd/cfLqR24KBALBbCGBrvUDgGXLlmH16tXsrw7MHsn4+Dj7fuvWrQCAXXbZZZvmBcCG8Ox2u/jZz35Wu5x+yppryAYd0vtDNm3ahGXLlrGXB5555pkA0vjSAoFAsKPiqKOOAgBMTU1h/fr1ANJF7J///OcAcllHEZJ/JrY+l+bVr341BgcH8dOf/hSTk5OzVwGBQCCoiU996lN45JFHcPnll6PVagW/ExkoEAh2JkxMTODqq6/GyMgIVq1aVStNmTzrRwYec8wxOOigg3Dvvffiscce64V8gUAgmBEGBgZqfbdkyRIAwKOPPor77rsPQ0NDePWrX134rh8ZKOuHAoFgLqEBJDX/m41g4/vvvz8A4IknnmDfm+cHHHDANs3LwDjYPv3004VyNmzYYDfiZqOsuYRs0AH43e9+ByCdTHAwz++5555tRpNAIBDMNsxdmq1WC4sXLwYA3H///ZiamsLSpUvZi1pD8q9MbrbbbbzgBS/A5OQkHnjggVmtg0AgEFThnnvuwUUXXYRVq1Zh5cqVpd+KDBQIBDsTfvOb32DLli04+uijMTQ0hH/7t3/Dueeei/e85z34whe+gKeeesr5fuPGjXj88ccBgHVU3W+//bDbbrthzZo12Lx5s30u82eBQLA94qCDDsKyZctw//3345vf/Kbz7r777sPVV1+NXXfdFa973esA5LLsBS94AevQxcmyfuWmQCAQzA40Yl3vh1nYojOO/nfddRf73jw/8sgjt2leBhs2bACQ3ykHAKOjo3aT7u677y6k+eMf/4h169bhgAMOwMKFC2uXNZeQDTrAKlduYYY+X7NmzTajSSAQCGYbJszGKaecYr0Lq+TfyMgIRkdHsWHDBnt56ubNm7Fp06bSdCI3BQLBfCBJEvzt3/4tRkdHbajJMogMFAgEOxN+//vfAwB23313vPa1r8Wpp56Kz3/+8/inf/onfPCDH8TBBx+Mb33rW/Z7IwN33XVXZ2GDgpNnMn8WCATbIxqNBr7+9a9jdHQUf/M3f4MXv/jFePOb34yTTjoJRx55JPbdd1/87Gc/s86q/ciyfuWmQCAQzBZ6CXE5Uxx//PFYtGgRHn74Yfz2t78tvL/mmmsAAKeddlplXqeccgqiKMIvf/lLPPvss867qakpXH/99Wg0Gjj11FNr0TY1NWVPNPtOY+ZUtKGvX5q3FWSDDsDY2BgAYHh4mH1vlK5ZmBEIBIIdDT/60Y/wta99Da1WC5/+9Kft8yr5BxRloElTlk7kpkAgmA986Utfwq9//WtceOGFNnxRGUQGCgSCnQnGi/iHP/whfvzjH+Oyyy7Ds88+i8ceewwf/vCHMTExgbPOOssusPQjA+ukExkoEAjmC8cffzxuvvlmHHTQQbjrrrvwne98BzfddBOiKMLLX/5yHHTQQfbbfmRZv3JTIBAIZgMaQAxd6zcbW3Ttdhvve9/7AADvfe97nZCRF198Me655x6ccMIJePGLX2yfX3rppTj88MPx3//7f3fy2muvvfCWt7wFnU4H73nPe9Dtdu27j370o1i7di3e+ta3Yvfdd7fP//CHP+D//t//i6mpKSevtWvX4s1vfjP++Mc/4qijjsLxxx/vvP/ABz6ARqOByy+/HLfddpt9/uCDD+If/uEf0Gw28YEPfGAGLTO7kA06gUAg2Mnxhz/8AW9961uhtcaFF15oj5ULBALBzoTHH38c5513Hk444QScffbZ802OQCAQbHMkSQIA6Ha7uOCCC/Ce97wHS5cuxQEHHIALL7wQb3jDGzA9PY0LL7xwnikVCASCucG3vvUtHHvssdhvv/1w++23Y2xsDA888ADOPvtsXHTRRTjppJMKC70CgUCwI2FbnqADgPPOOw8vfelLccstt+CQQw7Bm970JvzlX/4lPvShD2Hp0qW44oornO/XrVuH+++/37kXzuALX/gCli1bhmuvvRaHH3443vzmN+OFL3whvvjFL+KQQw7BxRdf7Hz/zDPP4L/8l/+CvfbaC694xSvwN3/zN3jZy16GZcuW4brrrsO+++6L7373u1BKOekOO+wwXHjhhZiamsLKlStx6qmn4rWvfS2OOuoorF+/HhdffDEOPvjgWWujmUI26AAsWLAAADA+Ps6+N7vDu+yyyzajSSAQCGYDTz75JE455RRs2LAB5557bsFDpEr+AUUZaNKUpRO5KRAItjXe+973otPp4PLLL6+dRmSgQCDYmUDl06pVqwrvzbObb77Z+b4XGVgnnchAgUAwH3jwwQdx1llnYbfddsMNN9yAY489FiMjIzjkkEPwla98Ba95zWtw11132cXkfmRZv3JTIBAIZgMaqH0H3Wxt0Q0ODuKmm27Cxz/+cQwPD+O6667DmjVrcPbZZ+Ouu+5yTiZXYbfddsMdd9yBv/u7v0On08EPfvADbNq0Ce9///txxx132BDEBoceeijOOeccHHbYYfh//+//4Xvf+x5+85vf4JBDDsEnP/lJ3HPPPTj00EPZsj74wQ/ihz/8IVasWIFf/vKX+NnPfoaXvOQluP766/F3f/d3M2qT2UZzvgnYHmAuDnziiSfY9+b5AQccsM1oEggEgpniueeewyte8QqsWbMGq1atwuc+97nCN1Xyb+vWrdi4cSN23XVXO8FYuHAhFi1ahE2bNuGJJ57AEUccUUgnclMgEGxr3HDDDRgdHcW73/1u5/nk5CSA1GHhxBNPBAB8+9vfxp577ikyUCAQ7FQwMmd4eBhLly4tvH/e854HAPbeDyMDN2zYgK1bt7L3KXHybP/998eGDRvwxBNP4Mgjj6yVRiAQCOYa3/72tzE9PY1TTjnFcVgweOMb34gbbrgB//Ef/4H/+l//a19rgf3KTYFAIJgtJPNQ5tDQEC644AJccMEFld+ef/75OP/884PvFy9ejC9+8Yv44he/WJnX3nvvjc9//vO9kOrgtNNO267umgtBTtABNtzbXXfdxb43z7nJh0AgEGyPGBsbw6te9Sr8/ve/x+tf/3p89atfLRz5BtJj3wMDA1i7di2efPLJwvuQ/CuTm9PT07j33nsxODgY9GQRCASCucDGjRtx8803O7/bb78dQLpRZ56ZTTuRgQKBYGfC0UcfDQCYmJhgQ7g999xzAPITIKOjo3ax+e677y58/8c//hHr1q3DAQccgIULF9rnMn8WCATbI8zG2KJFi9j35rm5r9PIsnvvvRfT09OF7zlZ1q/cFAgEgtnAtr6DTrBtIBt0SC+RXbRoER5++GF7YTbFNddcAwA7xI6rQCAQTE1N4YwzzsAdd9yBV77ylfjWt76FRqPBfjs0NISTTjoJAPC9732v8D4k/1796lc77yluuOEGTE5O4uSTT8bg4OCM6iIQCAR1obVmf48++igAYNmyZfaZOUUiMlAgEOxM2H///XHUUUdBa23DWFKYZ2YjDyiXZ/3IwLvvvhuPPPIIXvCCF1hZKxAIBNsCe+65JwDgN7/5Dfv+17/+NYD8NPGBBx6I5z//+ZiYmMC//uu/Fr7vRwbK+qFAIJhrTOt6P8EOBC3QWmv9sY99TAPQxx13nB4bG7PPL7roIg1An3DCCfNHnEAgENREt9vVr3vd6zQAvXLlSr1169bKNDfeeKMGoJcsWaIfeOAB+/yWW27RAwMDenR0VG/YsMFJs379er1w4UINQF977bX2+Z/+9Cd98MEHawD6pptumq1qCQQCQd949NFHNQC9bNky9r3IQIFAsDPhG9/4hgagX/jCF+qnnnrKPr/77rv14sWLNQD93e9+1z7/wx/+oBuNhh4YGNC33nqrff7AAw/oJUuW6GazqR988EGnjE6now888EANQF988cX2+djYmF6xYoUGoK+88sq5q6RAIBAwuPPOOzXSAyb6y1/+svPu1ltv1SMjIxqAvvHGG+3zr371qxqAPuSQQ/Sf/vQn+/zaa6/VAPTBBx+sp6ennbz6kZsCgUAwUxxxxBH64EOb+oE/7lXrd/ChTX3EEUfMN9mCGlBaa9lTRRr26MQTT8Ttt9+OvfbaCytXrsSaNWtw++23Y+nSpbjtttt6uvRQIBAI5gOXXHIJzjnnHADA6173umBYjc997nPYbbfd7L/POeccXHLJJRgeHsbLX/5ydDod3HjjjdBa45prrsFrX/vaQh7XXnst3vjGN0JrjRNPPBFLlizBT3/6U2zcuBHnnnsuLrroormookAgEPSExx57DAceeCCWLVuGhx56iP1GZKBAINiZcPbZZ+PrX/86RkdHcdxxx2FiYgK33HILpqam8I53vAP/+3//b+f7z3/+8zj33HPRbDbx8pe/HO12Gz/5yU8wMTGBL37xi/i7v/u7Qhm33HILTj75ZExMTOClL30pDjjgAPzyl7/E008/jTPPPBPf/e532fDqAoFAMJf4yEc+Yu9eX758OY444gg89dRTuPXWW5EkCd75znfiK1/5iv0+SRKceeaZ+MEPfoBdd90Vf/3Xf41169bh5ptvxuDgIG666Sa89KUvLZTTj9wUCASCmWD58uXoTD+AG35WvGeYw2v+ei3arUOxevXqOaZMMFPIBh3BxMQEPvvZz+Kb3/wm/vjHP2Lx4sU45ZRT8OlPfxr77rvvfJMnEAgElTj//PPxqU99qvK7Rx99tBB26KqrrsKll16K++67D+12G3/5l3+Jj3/84zjuuOOC+fzqV7/CZz7zGdx2223odDo44ogj8L73vQ9nnXXWTKsiEAgEs4I6G3SAyECBQLDzQGuNf/7nf8ZXvvIV3HfffVBK4cgjj8S73vWuoHy6/vrrceGFF9o7lY4++mh89KMfxWte85pgOatXr8YnP/lJ/OIXv8DWrVuxbNkyvP3tb8cHPvABRJHcpiEQCOYHP/jBD3D55ZfjzjvvxKZNm7DLLrvgRS96Ed7xjnfgLW95S+H7OI5xySWX4IorrsDDDz+MkZERvOxlL8OnPvUpHHHEEcFy+pGbAoFA0C+WL1+OqekH8MOf7V7r+9P/+lkMyAbdDgHZoBMIBAKBQCAQCAQCgUAgEAgEAoFAINgOYTbofvCzPWp9/7q//pNs0O0gaM43AQKBQCAQCAQCgUAgEAgEAoFAIBAIBAIeGkCi64UQlxNZOw5kg04gEAgEAoFAIBAIBAKBQCAQCAQCgWA7Rgy543dng2zQCQQCgUAgEAgEAoFAIBAIBAKBQCAQbKfQUIhR755fLRt5Owxkg04gEAgEAoFAIBAIBAKBQCAQCAQCgWA7Rt0Ql4IdB7JBJxAIBAKBQCAQCAQCgUAgEAgEAoFAsJ1CA+joRu1vBTsGZINOIBAIBAKBQCAQCAQCgUAgEAgEAoFgu4VCUjPEJSTE5Q4D2aATCAQCgUAgEAgEAoFAIBAIBAKBQCDYTqEBxDU33uQE3Y4D2aATCAQCgUAgEAgEAoFAIBAIBAKBQCDYjhHruifoBDsKZINOIBAIBAKBQCAQCAQCgUAgEAgEAoFgO4UGkMgJup0OskEnEAgEAoFAIBAIBAKBQCAQCAQCgUCw3UIhljvodjrskGciJyYm8IlPfAKHHnooBgcHsffee+Ntb3sbnnzyyVnJf/ny5Vi+fPms5CUQCAQ7GkQGCgSCP2eIDBQIBH/OEBkoEAj+nCEyUCAQbM/QSENc1vnJCbodBzvcBt3k5CROOukkfPrTn8bY2BjOOOMM7Lfffrjyyitx9NFH45FHHplvEgUCgUAgEAgEAoFAIBAIBAKBQCAQCGYNCaJaP8GOgx2utz7zmc/gtttuw4oVK/DAAw/gO9/5Dm6//XZcdNFFWLt2Ld72trfNN4kCgUAgEAgEAoFAIBAIBAKBQCAQCASzAq0V4po/rSXE5Y6CHWqDrtPp4NJLLwUAXHbZZViwYIF9d+655+LII4/EzTffjDvvvHO+SBQIBAKBQCAQCAQCgUAgEAgEAoFAIJhVxIhq/QQ7Dnao3vrVr36FTZs2YdmyZTj66KML788880wAwPXXX7+tSRMIBAKBQCAQCAQCgUAgEAgEAoFAIJh1aChM62atn4acoNtR0JxvAnrB7373OwDAMcccw743z++5555tRpNAIBAIBAKBQCAQCAQCgUAgEAgEAsFcIpbQlTsddqgNuscffxwAsO+++7LvzfM1a9b0Xcaee+6JtWvXotVqYfny5X3nIxAIBDsqHn74YQAQGSgQCP4sITJQIBD8OUNkoEAg+HOGyECBoB4ef/xxjIyM4JlnnplvUv6soAEkNQMi6rklRTCL2KE26MbGxgAAw8PD7PuRkREAwJYtWyrzCinbtWvXIlEKyegoHly/Pn+hUM7ZVe97gdkI1xXPIgDJDMvpleZe24GjO5Qu9J0i7xPved28y2gK1SmUv++oUKcNaVllqFunKvTat7NVbt3yNcrbgks/m2NsNlDGR3Vpnes61cmf6/vRUQBwZWDo217Kqot+x3avefl9xn23rcfGn3v+/ZYz2/3UC+roRVR8UyedLzc5XaS8f/cKWua2aNOZ2lZV9e2Rj1T2rQ7JwBnm3xc4XVmma3q1abjvmPxt29C8QvbqTNqkrL6hb+vw62zrr+3NJpltKOQ29zw56Ko4/b9uhD7I/l+3H7anPtvWsraMDq7cKhlYd65aVm7db3tN369tOJu2J5dnP3Xut52iLEEyi4N3rvl0NtYMZosGP/+qZpyJvRUqa65kw0zHLpffXNJXZ21pLtrJyMDn1s/+/BaYPb3VrzyarW/N/sBM1iRnG3Nhm9Ypc3tZu5qJfuyD5umJid4SCGYNsd6hbiwT1MAOtUG3LdBqtZCMjmLf//YROCdGyWaYIkLLfKMbQDSd/V0xTmx6Hfg2tBkHOMovGUoQTdQblHZhxVvcU4n7XCV5+VoV0+kG0JhM/5G0GOldYUQp7dHgfacS5r0CkoGU0GiS1LeuQeBvZNK6IusDQrOts8k/tGBv8plW0I1yTdbopB/Hbf47S4tZhOhTmdN8/L4tTxguV8XV9ePyswsrDIs2OirYFsH0/S60qvzftC3YMdELtvEGXXDslEC3NNR0RaIKA9xpp7IxV7MuhX6g49CjScUV8pQuHnP5oAeamXZQREaX8XMdOHIlMM4ABMdaXX7ldBFtc6Xz8kN1MWWpblHW60beFlXgaI5Mnk3N6tL8Q/Q22aupD0J8Uta+lWNvG23QRR2FhMpNv/+I/gYY/V5RD8rvVQtDdeWR8mjy0+imtnzGZ8CX77y3mfHvo0z/6qYulK8Sd5w0JjJdPZBl6Y3HAp/M0eTbjsFYFW0tf+PE30AjMobWb6YbdM2tph2BeEDn7/08mTbpRX/pBmxljFxkbUPAHfdV9etDfwX1Skka+zpQ515sD1M+tMoX3Mkjjr5Smhk6fFDejqYyec3Yaw5vVZVRoXPo97RdBp5LE0wtzjvN0ZU9yijd1lCdeoxY1U91+zGo5/pchJ+x7ernE/HzBVoOa09zcyTzvpvPG6rmfdC910lpMl8KjOfKOVDNuXYwbV0Z10M/h+bqZWlCSEZSQy3amjaUr+sKurCOjK6gxdh3pu975lF/fu7xo6N3S+gI0ZY0tUOn+TdHQyH/Kj5i1oiQKPvvvsoiPBPi577GTtnYjfO/7XivWgeomPPXpcuW6Y9Df7x5/a8S9LV+wrUdK/camNHmE+VdbfhAoXp+bjMAK28MQvNKFqRtq+pfxXu2/MweVFMzVEo1UMXv0bRns3h8U8WnvdpPBVldYY8bODZbVT/0A9J3zvwfRX4uyEN/zZux3/xx8sT/uhD7L1nSJ7GCfqGhkNT0oJM76HYc7FAbdAsWLAAAjI+Ps++3bt0KANhll10q81q9ejX7fPny5XjwufW5QCVGirULM+GvyaJJNF3DKPEMiuD3jJLV2cRcEU+4aKq+FcRNZJKhBNF4VPiuvSl91llIJsSGpgTupMvLVyWEfsUrFmOcOItyZYpIA41sI9JRXqYfqha0tfvMKiCrsVzjw9Bl+9czjPIFE9jyrUHAbVoCiAfz56qbpSNNb9qx0sCtmNAmxkiKlV2UjId0qWHpKF+uvyo259iJXoWx3B3SiLruex3BqV9hAZ0xMhyavUVF7Y83apBwC2s9LrT6fUXpa22OML2L2+iFTRIASQO2HXqZkLc352OU29ix7TOtkAwym9u2Du6ky/B8aNLgjzk64a+qsy3D+7cxorlF+qRFHALapEyOT0PjiWvDLH00raxsTVrps8ZUvgmiNW1XowyY/AJoZov93SHSN4FNx7JxppsayNqHXUSJ8r5RWrNjz/JcW+fylC74MmNPM4uyUZdfIDALuVTW2fKdvqFtwchwukhA4C8oFGSAJ+OczSw6aTLy3aubrXOjyI9aBWglNNtNIENnlG8GURlR6vHrySCTp9EruqXdSRGjL6guq7tw7NNp8qGLeYAnQ4l9RPMrjHdPrucfZ3WcVmFnGITHu/8NrUchfabz2QUHbyx2FxgdSvIg7cTVT1H7xysn1N+Vk29mQm+ziYkuIe3D2nkD+QZow4zRts7HJNH/HKi86u5CFjxMGrpAVtKPANxFt5KF8FQnug1D2zlp544nzbFMxg7r4DjKMyk+ck4FMpt9Ri+o2BsHJXnaV+Q72mdc+SGHPWPzJ01d4F0gYCtyNhFHn/J0JbPomQxq5xnNL17cRWND0+ZlyqTjIMlmmcZpxCGTWfT1HSOndksza2cbdZ3RJLfNyXzEzIfopp2GqycAQE2pcicFwsfsBpBXPz+N+TfgPrPt0M11tbPpSenM0iVt3naz9DNjiLMztQKS4fSjxlbO+AjIDtp3JlOy8VImd+PdOog2pIPHWTymPMbwczKU1bnC+TSaVkiyyV1oDJQtdhYcF6x9wKfn8rd/kv52xqZnywR1IYGhKeqofCG/5oKxj8ZYw6HDsZGZtoEG4J9U9eUzRzfhQ7sxx/CTY1OE6qTd51T/OO0b0p9M++YL5bls4DbLrI3c0lbGOGskjC7XUT5HcewMg5aG9jbufAdFdq5MbQXfgddbWDf2obOOQea6VXMwSkvoG+23l998RIZwc5zgpqD9wH3e3JLp9ZHcHivbXI4Xd9F6NpU3cbuoK1l+B4rrPcjlnooDfMz1o0bBodHXJUYO0Pl1PJLJ5bEIOivXzMuiaaKrNFybLfufs55W07mA2kzchrfTtmZMkDax85J2Ph9RZgx4+pujpcz29Q8emG/jYV4vWFoG8kaxNgXDZ9F0vknPbcQCLp+xdJYtBfjzUgDJcGydJPSeU8C6zPsvsB5lNsvigZz3ufKDjofcHMyvC+UnIrdtFl3Cu3Q+Q/MlvFNiZgq2AeQE3c6HHWqDbv/99wcAPPHEE+x78/yAAw6YWUFk8cEoXIVcgBlFQRfw4uGE3USisIsLRujW9OIEYBcHHVDFxy220EmHmSCQbFobGrkXNKFxanHmdUeMCGrgK+ohxyhfR074Nl3kLWRWTFYABD0l6SSdflsKndfLtoVXB9/rxl8g9CfHKiHGQcAwciZ93OI5VdTmT8444IwIYgSaibxu6HRjDt5iHpNnpXdOxUkWbjEytHhs4CxWBerHLh7TZ36ZETFkOD7wTnIUK1LyztDJTKA4nuvsGhc2nLg25zY7QrTQ7zqLEvudnTgxhqHSQGPcbMYwedJJl84Xb+jCkUO3N+boGHbqXGbAeihsbCAfD9zYCo0HuoFiJ0I6MBc0sphuqGdtESudy1tOXvSA7jDTNw2wRJVNWqKOytuUoyMp9qUPy3MdxRr+zkTWyD6GpzhPTa2A5IDJ9B9/GsjJMjKIeFVSWe5PlCjtBY9VpswyJAN556eT1/QfdsE5sPDE8SPAj7O8MHeC6H9nnRE8upXnbe6fOoiHirYGK7sD/MRuljF9atMQXZe0inZD1Mknt6w8oZ6ipB3YhRUqt8u8czVK5XbVBl5os4zKSCe9twBYtVBLZbjm7DCHGPJnxWaWeZ+0dIFPaJu4/JQ9I4us0WRkeT+mPFpmc1A6qLwyQ4dz7vI9a6kdCndTDSALK8xCqWMbEl1g+XAq39Tt7JFWOtrSqO357JRF+YCRN/Q0sv227iJ+YMGT9j1rB9EsqIzsUW9UjQ2lAWU2zpi6UJuKy6f9dCt3PjFt45XZ8E/gJWS8N3X55no31+tTe6SEqCmFxni2gDWYO8c5C0wlMq4KlN/Zk+JUniXFNG4FchpsOzR1rhcDJ9HN8zLbjX4Xsufp383NDVu+TU9lgLcJCbg0F4QD5RdmLjPw+AC6C4idyukqxqGhuYXQWaLf4sHczmM3on17wZtDUhnjODnUlB10DkPr5oxnTwaadE4aH6aZyaKrPZUWcUwWBudAW1Y/3SD2Ty8n4KicM5s0nCMq83fwlLGxeSJ3XIc29IL/pq8Y+8RfnAeA5pYo178VUW+g800Ws6ZD84sH8w0H1iRQRTmgladjyGaleW+QtHgZWuZEEpoXOWl9O85zCuEiCZTN+QsRUqrmuEuydajJogzUTPrmulwXqbjcpnR0PkOrkXu6oQGOjzmbSSF3+tPFNHTDI38INDdm8q5N5q1mLkt401nfoFmQ+UqZTUVlaWevVOlHm8gScIVooY7EzmYYacsC6JpLTZ3c2BrxfbLVrG258tasczlyq2Ri5IwXzg6D2+ZsNjXFsMmnualp+6Tx+GDuABzI33xrHHwL0aaMKRVad+TGu9dPzhqdYvQjicJlvgmiN7UkmGVoALHcQbfTYYfaoDvqqKMAAHfddRf73jw/8sgjtxlNAoFAIBAIBAKBQCAQCAQCgUAgEAgEcwYNJHU90WSHbofBDrVBd/zxx2PRokV4+OGH8dvf/hYvetGLnPfXXHMNAOC0006btTK5cFbcabbGeFR+OgfEg4F4WdX27qwaVNwJpwpv7STQ+5z3jn1HT2qEPKBKvB7TTLL/MV5cwfuNSrytlPes4MnjtQPn6eN4z3heYI4XE1M/x7OaIZ2+4Dy//PJ7RShPLu/KEFQcKsIlBPMLnVAAer9fCow3uFdm6DRaKA54zyfCfSemMs/0Xi5lrxov9FPvJINmvCsp2PHgl03HR+heyaxMls9tm6rCs37BnW5hW5TQz8nqUFx+1svZnJZOXBnQ15ih9GVpnVPI5ISfoaf0FFBE9E4opG/ZKaTAt04ZJZ6mdaCfHSg8syfnIuJRzaXlvBdnakiGeKNCL/UVor1CBtuivbwL4d38MCPcyXkONFTJLHk6cjzCeuYG8mf1c9V45LKskNO17gap6VVf+Lsm+jm5VdUXTjhFG0eoopyAfiyc0tIV39HXjeIdgfT0XOgElxM+DiiETS7lpSoeIX+b0EH+87r5svfNcDYTyp9VgTs96rwP2SSmz1GDvkBfhKAVnEgQBboq9BLHR0EZx4XlrMpf5emdsGDDxTxtkn5sO4oa8snKRmPzeEYWF47ZDeNs/p6B4V8Djn0Rso+Yb0v1JuXhEtuoO5LUtx9ImY7e0d433rOy8VRAVRSQHsHZLCGaZpp/6OqE6ky8fwZOejplcVEN+mlnrs49zPu4dQZnPtCDvUvTOGHbUPw7LbuaTieNsdPIPN/Q19jSqLwmgtW7Xpi9YNqATcWd3iwrj5YZsh1YPqfpG/m33KmeQqhOCuX2adSpEOL+Ooxy58B1+UIxp1K5kPtcmXXLSfNEsZ808jl3BKDsRDTlfW49CvVtqmhz/aVflv96aYdeRVcoHypvnZPXaQGUp0qjnVT1V4Q8ksIsrWPQOXvSqqbB1CX2o7L4MHSG5CqxHQs0UXs0MF5LT0cKtiOo2ifo5trmFMwedqigpe12G+973/sAAO9973vtnXMAcPHFF+Oee+7BCSecgBe/+MWzXjYVUI0JhcaEco//NuAu5pZmhuqJqY8Ihd7iFtCC0CgI+XjX4qUQdnIWWjj23sMzEEJlOe9oWR5CmwxcnioOf09/hfeN9Kd0YLOO6UcnL0NLVneVpJtDNIzZbKBq0Z5rE93UxVjxFFx/VaHmgr/Pj6Yd6/azD78fqsZLcNKhCA+RvPtBiGfMO4u6k9AAb5d97/N2aBwYmkrlklmUC413uO/Lyur3Am0zHkMoHVtV7aZ53rQg9W9MqTS0iQbiBQniBYkj78r6PogEhXZxJpG07iVj0yzwld5TVyIvChM+pqyq+tn3ZbKdWeA1xj/HhypWhU0obrz2hVCfk/7kngfbsU7dNf9dlS5is2RkQ+k49vmsj/Yrc4DoFzMNjV9Zj14dPTIZZvjZXxDuZ5zXtv16yTPjGUprbV1BecYPfVuWnpMLzN2gHJ3+s0KePdgdVXV12prwfqXNxOVF0nDpOX4o1SnOh245db+tKsuhiWmrKp3K5c/Sl5SPB3Nfo6U/sBgcrHudRXCiQ2yoPmKH+HWdzXEYlNvGBgvpZK8tjB1sNoGjaeVcIRDMv0Ku1HHsoRsSMxkb+UPkvMboHZuml7JIE8aDOr/Hlubjy4MZOGU438yCniubb4a+rcRsrKP59euhruycd6bgHMN8m9DoYGau4duxPW981vw+aWkkAzoPIefRVsiW0c+GvnhRfwsDVC6X6Rv/rtwZLaaHbGMg3bgoa3NS/9Ac0cojLv/Ey5+RLWx55p8Vm6AhqGlVcB4KjueqcVBz7Ssv3LXzHNnJrZsQnrDfMWuDlZjhmkNl31CamTleVXml5XtzNFaXltUvMBfN6UV/62QlcOqU3U0eqqehQcXVdoJdgwytyfQxX3DyJrQJtl9oAImOav1maVov2AbYoU7QAcB5552Hn/70p7jllltwyCGHYOXKlVizZg1uv/12LF26FFdcccV8kygQCAQCgUAgEAgEAoFAIBAIBAKBQDAr0FCYrumRp+UE3Q6DHeoEHQAMDg7ipptuwsc//nEMDw/juuuuw5o1a3D22WfjrrvuwkEHHTQn5VIPAuPhVwgtU9Nbobb3LQXnqdhLCCbGK4oec6fvS72wuMtIe/TI7svzbRZRegqoF3gnLko9WbdVnUl/sPXswfPJybPqkzk4QdCr12aV5+BcH9MvnBqYAWbsCZlhm46zOSirF8/koAd/Td50vPLGozSMGzlRMxdjuK6XclW5s3Gyr64M66WsKl03Jyd+a6JwUiajdbbGXh30bAv06AFfLLBGwtlu8yoP/oo6zTbv05PdBZ7vRd7UpG+mmPHpDL/9+/UAr/Oeg5E1M+Vdn465QE17tba9Mwt15k74VfFDLxEdqk481T4BN1unbALlU7nMeZtvVyg5udSTTTPbdJR9Os/zMgtqm5TQX2nvV+nV7aGuZZiDsVSpK2exzNr8VHJKmpWxM5Aztfm7h/UcgJFBJF00VvMos08Cc/q1r3WGPtqKLaOGnUrnSvOOmnZs0tZI2uFGmlO56OkFg+C8wK+TRn9rOtsSkU5/c4S++kejOMaztjWn9Ob0qFHFqUczjpKWLg9vTNYgZx1y1GqHQgxV6yfYcbDDnaADgKGhIVxwwQW44IIL5psUgUAgEAgEAoFAIBAIBAKBQCAQCASCOUUy6yckBPONHXKDbt5BTzwQRNldHaUXtaJPbwemzJ48laiXCJPej+1eSkfZgYcaNHEX3c8HQuXPNX3mxEjV3S79gPMwni+Y+rGX09egzcQt7ze2fLFQWP7v18tPe+OHPptNz8HKfIwu3l495wIyEoBzUbv5N8DXuS8ZV4XAhcocn6mYHBmbiVwLyM26nv8qVpX3z1VnYj6emYzj7pDQjfzExoz7keZrZGVZHP7ZwmwXUaErgXLenxUSTP7GE7uh+9M7cy1vyuRFBe/Mt56bK+jW7PD+rHq0l/RTXVqA3uUV+45EfKidZxUC/Ndv/lFX8fOBPsdTFS390GrT9EZKlrjfhL3D2gzboWe3bpbL1ar3s4lexjvll1LeCdhMswq/f2d4srpSn88B787V6SFfj+/MmEt9Xrt//FNK5s9+bNtZ5LOe5DvVMZ4u9edg2wNmOnbMfesAap+A2pan/XopK6TrClE/YmVPp/Wl82ej7hV22naorivXau3aU3MmxlEJ6t7fN8fzw1nDNrQDBUVoAHFNgS7dtONANug4RIAe0FBTuVRsb47QWZhKS2os278bKB5FDmymUeHvCOASRZcMpB9GU5G9NDQeTuzijUOTmQiSY+VJK/1/YypfMNANUpbRQw1Y7a26qmjYmQtUATQmFDq7pSvM0XgmHBIFUKXn1UVp2HACqqsKCkgxhpWOeIOCU7JRRyEeSXL6GRrst1nb0YugVQy0Nqd1mc762w9lqrg2Y+pr6IsH8kRRRwElF+aavFRM/jZhDAB+Qh2l/QoA3eGs7rHKnw1puxFQe0JOQXjT5KMb5O+I2WhQeWE0rIrtZzJB4OhQCXL+9Z/7ZWbv2xsjTC3O+yyazl4bPhnUdrzQsHqWjri8TaKOckNheJvealpBm/d07NONRlJ/Ww9irDnjAChsZGlmYc+m12QTx9BPDSdPHtkyDf8m1Txhxxypp+GzuJ3Xz5kINtxnuknoV0B7/zEAwNTjCwrlcTKKwmkT0jamTVSS91k0lfeDkYeUb5PBTMZORvnYHUxsY/iXsQPZOCI6wKdTkXrmBal84yl7n7Q0oqm0AgW9gEz+G7HeUfxCAVOW4b2kpR2e014/+fmY9snzyeusplUhLKWa9njVfGtksMrlbN43ZOwlRd4HYHWwwwekjNINc7K5qiPkY9OUTxZMaV6m7wuhN/2x68slv0/oGAvYzsZmiDqqkCcob2Xpo45y7Aw69k1FKB/avwnvmL85HUDleqguDszGn6nHpHLkqc2c2CKFRQai35x+JjTHhHcorYBr03BQtB0AKLORTOWGGYfDcT7es/EYTfP955Tv8ZatF6plhKmXT5MZu3Rzno5x244ez1g6jEkW6cJF847+JHqDcx5SMZFXWZpkINelTtg00wxtDT2cftzYlE0zknInA+WNl0Kb0PdNosuZ+lM+sfxO0kDldTF1jUeS1H6FOwaSjPw0jFnFwo0nV+PhxOavugwfEf2sBzRUx+2nqJO3WdICDjvycQDAA3ftn9fdjDeSv9XTLc2OXcOHUZfYmdM5HysiI31dpRWxbUcSRJPegNaZjif/pv+nNoGjv817OlfSeftHZNxS29OChIMz75v7jCN+Ytilg4ydgk1nPvEeJW2NaDLK6bN9UhwvoPRxvEn5yIw7T3cW5IRy60z1pvmOygYre4ht5+TPzEVpXZoTZj6RPZtWucMADZ2q83rktkz2zrOXOLlP65GY942cDi7MW0G/ZHX2SGKdwOh7NV2tl0p5M8n53NCy7wufwR9X72npZ50TG24aH45Oj4zcJfqdyjUq70vysnRO5f1YF0p7/Orzlvbm4r5eCum6LL2OApvXxLawazGMTVRYZwk8t2X6djKd02dMmAwltu8a4xGvg4zc6Cprv9kxRPojmlT5GgMdz2b9gcgfZ97HtJ+Vzx3lynB/3ujbxf7zgN1s+rk5ruzYpzxo7CzW8SRCbq+HeIzIQH/+T+2MsrRpAe6YANK2aUykmSYtzc6hbVZGv00rO22MB/J1Et9uTjMtym0nT8obRgaRdcQq25r2eWPCzD+1w5u2XEt0UW6F4LQHJZ8bp8zYoPqBkweObUTTM23WyGyGZCC/KsjYqI6MIXZEbq/meSZtXRh7VEYYOOMe+ZohJxdSua4L6Z31Jk/u+vrNX5ukNrxWxfxVV7m0ZDo2X9dz1wT8PomIPGDXnpq8HejMJyp0HRf6veermgRzBIWktodx3e8E8w3ZoBMIBAKBQCAQCAQCgUAgEAgEAoFAINhOoSEn6HZGyAYdhwTO6TkA9vQc4Holas+TIgTHy8UcSWdChYVgPIO0Il4ZxNuWnojgQ7VlZHqnzfxTf0rDOeFVOC2gco9WPazzk3Om+Eb5qUCt4JxsK3j/UA84h37Pcxl8PZO2zj2rAh65mnjm2e8IffbknOexFaTJr7OhhXge67ahOeABSDzL8vyL9IdCkcbGs470XeJ5PvsIhuIzNIH87XvgUd5n+qxwOsZ7T09KcHQ43wfK9OsxtWuS9znxbs3djcrHRogWA9/T2q+T7wmZn/SgxBbzLfWA8+vJpY+4j8n/ufalzoeMd3GQnAYZM9n/Le9xTloczZQfNTD1RPHknP10IEFjPJd9BXq8061Oucj6hJxSs6+9E6kAEHWKgieazk+r0dMlVB6Wht6i49amIR6C9vQD8WxOinWlHsbBPmLGIz394Iydkn6mJ5Zp3vZRxNBAvPq0Iu/NaW1V9B6lz6gu5GS1Q3NIbhgXfOtlqp1vCqfN/TY2zR8Ke8LoQr8faT5B+Uq9q7mTcwYaBXkXD2r+5IDK29nRe3DTV9nwURc8P4fsFDOOaD38sqJ8DHLyIJpiPMzJ33Fbs31ubZrAKSGbjdfmBU/QSEOZk1OTDfgItZljU/k6VnlVqQjPVyiDevsO5KewfK9zP0+HR62MI/KKnDLm9AZ3koF6NNPT4vY9TWLtGIVokze9iEo6yc9HIx/PBWHkniJ2xhs39sx4iImXMkj9q0KOG3lAPfxDsp7hbdYznbHJVEcV0ictnZ+Qi72TcwaGJGJTJa0SOuHKXXo6suDhH2mWd+2JTu40tz/ePPhRAwoRM5T7rc6ih+g4st/T8cJ5fht0nxwmMjL7X6ysjuBOz/k0ACm/Ww94UgbLM0yWjh0R5RmwNo2vF7M8qS7gTmKwcoo7gcbMUQA4crE75PKPbuQyjnrqK01sP3t6M/+OPUlI6SDj1daJiTASmouxc0VCepkurgqdz9nbDm+SUxfm+R/v3ZPPn/RdWShwgMr43H5hTyTpIn18JQAkGb/3eHrOZEHn3ZZskpXDR2ZscJEGGJtGMf3tvKdrMb5NyNBaOsfiZDGlyci1aWX5MJSnPRkdo3CC3pdP3Mk1zcgehye40/iUpTgZwKRnQ9IH1hEMz04vLNpUuq3ZdSRanj1lpRHk7/SDYvmFU0X+2PWL9EI7qthdz6rkE6R1cqIWcacvCR/Xkhm0DoR3UxnqzkFUVzmRRUzu8SCpB+VNz34I0sO0WbAd/X7y5Io9bU4jFnC2nGNc5Y+4E/i0ftZmMf1Pxxsj650TbiR/mp9/6j50gpyLSMSNF2esMXLX73d/fPi6wF+jdOYLxDbvjuRyhRuztk88mq0NFDFtasaLH93It12n8/EUihK2vYXJ/bOFBpK6R2l7NwEE8wTZoBMIBAKBQCAQCAQCgUAgEAgEAoFAINhOoQHEqLdbKvtzOw5kg64fzDDubu7x47+oTjObSLi4xH7sZt/bnDkJ4aPq5FlpPUN39TAeW8E2KakD6w1X4Q0Tek9P4FWdgrKnHRkPQq6MED2sR2ggH9bzuYqPZkl6sx7O/ZYTOgFYpo+IR1DodMpcovIkWsibrSx9wJOzmpgadNTNqoTuoJd0Bcq8rKOJqD6NHJ+HZDXDG5QOTt7U9ZZ3Xvcy3mZLr9Sho4qPymipOAlTWm4P73ul2fcmDaV3TuipwLczoKdU3lTpmgCsR2RA9/ZrH8yabKyQMZzXvYHjTdsPKpL73txlp1PpHbsz0hGzqF+CcqeuvJ0LGaPdvwtlV9lEdVAiZ7RiD9aVgzmZDPB2EoV/giNMFPNIwe2nHtMDqD7h3qON7UOVnOIN2ftV9ksVetHpfpQO7pue0IP+ChZZYe9z86ba6THz9u0Harpmp9SdTwTGWy8otQ8iXvaUyfpZRVXfBE6C1M6+at5XJ4+KEz/9oLajfpV+qiIt9H629HKZLtTV6xzO6Uk/L1JOMsCss1QhYCdyY3Q25/T0EFT+MPt/D+M5+F0PfNzz3KFiXSeE0hOtPdgxVd85fZf9mQSiQ4A52UUjYlTKg1lYXylDcH2C+65X2eWvZ3m6sI4urTy5zZW1rdCLvdZLZLay50zEgtDJYIOYnvANzDcs+l2bEswSVP0TdHIH3Q4DOaAagNLub2BDBHM56OC6CIPrIi88miqkoeMgIossrS0RWluinlpfxfnFrPaSUpI+6qq0DDr2FHkeZd+T9wsO2pTnRdDYexyNvcedslSS/qJphWRRF8mirg3XqBJS5yj/qQRIFnaRLCS3ikd5XXRT5/SZsieVLVM30h+itFxznNtvZ6fNAeiWdkKCRNOKrSdNY+qBCEiGEyTDnmYk6ZsT6SXqUSf7TSuHvkL+Kv9FJCwpzdNvB6eeMWkTlS9OmXak7d7eFKG9KXLy1C1t0ycjMZIRb+ahmB99HvHtTNM3phQaU8UxoDTyC2ZJ9Ro0PKvpZy9Pn6e0F0KytM0IT3K8kwwm0E0N3dRuOxOaClBuXobPLP2ElOY+40wGbl5QCKa3MoIuNpB+oOPK1MO/+FvpNF/Dr6F2on1TGM9w+6fQ/0YORWmdaVumFyMHeMYgEKKI9pNpcyoPbHLCR2qPKag9pgDk/Bh1FNqbI7Q30xiEcNs/+yXt9KfbGslQgmQoydsjyevptIMnb30ZY2UJeZ4MJuz74PjK8ubKtDxO5JAJIebQbGS/SseeXtKBXtJhZZDVI+RH5ZiOcp6w5Sh+HBs+trLKC59sx2DMy0sD3dT52GPa2ekT0s5OnkaOLJ6GXjztpI26efsZ/Q64493wBO0TKteTwSSVKxkPOfQbPhvQTpvu94JnsN8LnnHbmeRv8rR1GkocPtRR+qOykI79gtymvAEU+kQrRmdTkoguouPQ1DnqqkI/6FbeD3ogf2HHVqzsePXblbaLn69Jb2WCJ0tsm9A02uVzzmayunKXGMkusdOmvl6zstK0Y4A3C3XyQzgz+tfkOXrIc/bvRif9hXT1wociLHwocsZwe1NkaX3VyrvxqpV3u+VSUkidbDsMJbZ8Y1/4oaR0BMfmUgksrcbmaI0plveoDLF80tRImmmoIq7vW2M54fEuMeJdYld2cF1AbAFqNxg7TiUpf2oSckolrj3oyzgnPBvDR1FXoTme/mif0/okC2L7K7xXrrwq6EKvflbGZzY8a8eQ/HVDI9pzMv118nxpm5txQmWF+W5wbeSMY98eVUnevtTOoW3ut5nTTgnsfOPE4+7Ficfdm7aZJxfpM93Wts11S5O6Mv3H6Fpri1G7NlbOuPfbCcrlHSPvnTmC+bubz6EcED1tynVsN0YWmjFG7Tin/8h42+WxCLs8FjltztmzlD8Mzcb+1m1dyJvai7Q9d/+1Wz7tU18GUDt8YH2EgfUpE1C94NjJ2bgwjnicfG5MeTLes4NA9CalkcojyidKw7Gj6ZyffsfNBZ18plW+WG/6lMpAIncMjCykSEOIuvq7WKBbDyo36TgwMjoZToq6hrSJU9eKso0usvXWXpnZePXncKb9jNx15r8tjda+W9Had2tuZ9HqhuYZnN4l7ywf0PmhyZ/YTE7fEhnxyVOvwSdPvSaXQQNkXkH0J5UX1Cbz57LO2IM7TzEyLW+w4thz+nzfrQXZ4fxNyjR0JsNJYewmi7rOeCz0v8rtfTq/p+D6nOoZOpc1/Uzbjq5nNSYVGpN5mxqb3cDmxbQJHRetzRFam10+tfpJe3zMyDOap1+mw5+7T+XtYMYzkfXUrjPrPIDLc60xlY7hROUhwHefsnnbPjH61+eJLEk8kiAeSZznppzCmDbzpuwHePrX0EzeUxm8YE2EBWsil8eJLKbjwPQrpT+aUoimlO3vqKusrlU6tw8cm4rTr6ZPWtq2k+ExgOh0quuJrqMorJOU2NFKAwPP5bx52D89a+2z0LyGyuPCWqRTUNaGXp8Zu8G0GbtmQGwSx07M2t7Yyf4aIqXXysUpFR4Xgm2CBFGtn2DHgZygEwgEAoFAIBAIBAKBQCAQCAQCgUAg2E6hAUwnEuJyZ4PSWkt/ESxfvhwPPrcee/+PjzghDvSAtqFujAdR4h0B1ma7k/EkUDr3iLGXpYbGk/F28LykQnmb/AHvaLPK8zCeKzR0W7IgRjTWcN+3tA2ppBs694pgHKK0Iidr6KEsEjLIeNlEE8a1xvVGKYSbSph2IWkqwwhovl1pevO38U6Lh7UbSiDO688X4v0z87wD3PalfeL87bWp/94vJ9RezikF5T5LmpknErIQG6b+xkPWC7cQutC7UC6lz/P6AVwPKuNlH02rQr81Oso9Qk9oMXn7vEUvnecQdRUbmtB482mQMUvGBq1D1aW3Jq+kpYu8SPJMBhNEk9UKU+k8DS2b8ogTVsy0Lx2jHO/MFFQGcfLIfEbkGq1zGU1+O5sTndFWxt1f5RchOxfcmz8j5LxtLqOmp2I0Q0sgz+GnU6LG98y91hpbI+upSC+mdkgskxe07RgZbk4qx08PF+tGs6FyjXuf8PzjXKxN/jbjQHmerDaNLj6zPEl4wg0RWmwHTtc5/UHrZJIxskZ1ifd1XYuFGeMhOoFcXho9njQ1TyuVx0ZGOfoje8fU2eZl/jb9YOSx7/HK9INtF//0SZae60dbHtHvoXY0pwQb41FBr4RkjNXzVOaZYiLSJi1tvzHt7+g3TmSSMc5BxSo/vcV9F6X8Y/LnbCFbj+EEjbGoQEuZfobKyw21k+1/Op4CPFFIQ8CNAZUw6UmbO/lX2JHOe0bu5zJA5f3HyQ2S3siaaCqXy8HwbZSPjexgxljSDNgCVfUjXumWfmL7UvvIb38qA2zZIXuT9LdKcnnDjidmnHB5aeIpXRm6LmCnWRC5a8pM2poNp2pPWGVlOuHLBzSi7FSR0w4c71TIUnZeRNqHjmGavmBPEzuRk4EhvcCh7lxNxe7Y9OUAN//z8y1tH49mVjZwspmUSfW/gbF/fJupACrDqdwracdoWuX2FaPLAN6mcew4pi50TmvpT4p5NaYU4gEdTG9OGgKoDCXGzrWIjAvZZhwf+Plw35n8Tb62TRLVU0g6P1/aJ5Qnq3RZmU1RVaYDossV6QfHJqbz1lA+Jh2KMrOsfIcnbQIUbE/d0IDJfyrXBVG29hMPakfG+7QkA9rK6mQkRrSlEabNs0/sd4SmQj/5OtnahDXnhaGxWzLHsGVl7zlbxOqKqKijnPkrZ7MYukyengw0J/yATFd7utgZjwr5SfgppqCQfvTa1xkjqLADK+Zlzt+M/lfZCUJbF5MtPfFt5glZmOagHVQGBXaO4ugKrs6UN0LzDeaZPQXaLB/PtLyqdTADup5Waa+T/HzeofTRE2WhdcfKNRdmDmv4OOoqdBdmax5mXZTyu87bzMjA4LpvaO7k6SWWRzy559fVRifxv7UFAj7vP/k/L8QhS5Zg9erVPMGCWcfy5cvxzOST+M/fe3Wt77/5hn/FnoP7zEofTUxM4LOf/Sy+/e1v4/HHH8fixYtxyimn4NOf/jT22WefnvLasGEDzj//fFx33XV45plnsOeee+J1r3sdzj//fIyOjjrfTk9P46abbsIPf/hD/OIXv8AjjzwCrTWe97zn4dWvfjX+23/7b1i6dGmhjKuuugqrVq0K0vCmN70J3/72t3uiey4hJ+g46KJAo0reMfrtByg1nJ1Fuap1e844qjDKq+6m4BajzOac8167hlUVrazgpyFUJrwMfOXrIXRfWU9xwCsmr6YMe/eN17ali30o5q8bmg8rrQJ/M3Us67/acabJMxWTzQe6CMKFXgm0bxVPOY/9blb5BJvLh92c8/IvjMEKQ5Tj8WBZ3KMaDihBur0862zOAa6BXkWLE8+djtGaY6Mn6MDfHE0ZaJ17GePsxhwpOwlsjAFweZsu6AQWQsryHN8zz8zce6MVgEZJ+ShORP2yOFoN4qeGiw+5bCoWwwrjr+TfOmIWv0DaiWs7bjHHT8+0Q6Ws42Ss9v5v8qlYGArm4z8O9Bd3F1oVrez9THXlO5DfTcoh0A9Vdxhw/WjLq3HPkNHZIf3FocGksYscpI0U2Rixn1WJyirbp6HLv/EdAgJ6AkjHfa/6uVTW+M97aFP2fWCRgvuubvrg+1Idyegff7xmUGTRjkPoThB/0dt1BuDzqqyfMYloGCoampOMj7L2q5SF5L1m7jyrGifBb+ven1b1GbOww951CBQXtAnUlKp9x3CVLA3m4y8AeukL6QjvswvBPegRliaqiyltJXKgzvyvtH38OUddm4+OHW4cBWymAvpoR9ahyiuf1VEBPvS/pXeoc/ZRYZ7u/bOXhW12fkT7MbCgX8YHveQfbJMK+Pk6oYm5fgj0U5lNUVWmSwDzncfrde+krSUv/eI5WcbMdVRXASRCnsmTtRHJ2LeylOgRszkXpI1xDA3NCyvnrXXnhaEmrrvOFJiPlOmKkAwK0uWvyyg4TpJ+//ttyG7MceWUPK81fmvOy5y/mbGnG5pvH2ezzi2s5825rLwynR2sc2BNoGreWMYTLHkBPROUnT3Mt0x+rCOUWa+i47kHee2X4fyfIGlq/l5dxnGmcjyG5k4VNmWBNmYN0LFpao4XwfxAA4hDkxHm29nA5OQkTjrpJNx2223Ya6+9cMYZZ+Cxxx7DlVdeiRtuuAG33XYbDjrooFp5rVu3DitWrMBDDz2Egw46CK997WuxevVqXHLJJfi3f/s33HrrrVi8eLH9/uabb8YrX/lKAMDznvc8vOpVr8L09DRuvfVWXHTRRfjGN76BX/ziFzjssMPY8o466ii86EUvKjx/6Utf2ntDzCFkg04gEAgEAoFAIBAIBAKBQCAQCAQCgWA7RjInnvphfOYzn8Ftt92GFStW4Cc/+QkWLFgAALj44ovxoQ99CG9729vwi1/8olZe55xzDh566CG8/vWvx3e+8x00m+nW1Pvf/3586UtfwrnnnourrrrKfh9FEd74xjfiQx/6EI499lj7fNOmTXjTm96Ef//3f8eqVatwyy23sOW99rWvxfnnn99Xvbcl6h3z+DOEOe5rLkpe8ruc+QfXRhhcG6UXRmeeTq19t9oLYtmLQ1X+SxZ2kSz0LjBVcC4k9Y9zJ6NdJKNdNy8CU7Ye0M4F7eZbe1E9wdDzttj39hLnprb1A5BfSpxdqhrtOenSZ8rS5LfnZPqLVaFc+qwxSS6lz/LUu3UKbRZNK/dS2exb/+Jle9Es6UMn/I1yn9ML1mnbN7cqNLeqnKZm3qa0zrbNG+6Fzn6fmHbVTW0vJKf0+3xgj6eTS6ppnxYu6yb004vcaZtb3htKbAgz+14X29SkU7FK+c7wCWlz+yxJwwOaEIEmT6cu5oL3DCasoU+HQdQlvJHVzV72ruBe9p6hOaGcfrSXHmd1a47n6RuBC94tz5WdbAHcviB9Ymml6cnYp33p85PbGPn3Js9kOLGX9g5siDCwwU1UuNi4RJ7YenaVc/G3nxfgjTPzN8mT1pnSmgwnrIbxx6x/2Tct04yrqKPcMcDIWHqpPZVL9qJk2u6GD6byck2bqgQY2H8MA/uPObKHG6/O2KH8pIo/esG530f0wmali+NFN3K5RGUgBb3k2vxtL3mmJ0YapE0IKH1+nlD5xdNRR7GX2u91xLPY64hnXaKWTgFLp9LQhtl4tBeMe+2QE5L+qF5wQnl4erbQ5hmSXafdZ9nfVn+P533f3Kpy+ujF3KT/7IXYRq5Qj1HSvqZtuLHnP3P42MglUo6R11Q+0zrR/tVNDarL6aXzQVns6RUgtxU43m6N5W1G+8Beek/7lLETqFw0dBi6De1+ezv9wDme+rrEs2namyLbjrqpLe8ProswuM7NMFnUtbYWZ19QXU7b0dc1Rl9F3skHh2dL6kTbLxlM3MveVdoPtB2tDHsuwsBzkXPRvG6T9g3UyeZFxpbTnubPLP3wgZsLcs3h8yS3HU3bJwvcfmqNKacetO90I+dDK9dVLreShV3bPq19t6K171YAcNqpYD+QUw2O7GTsK9Ou1Laj+s/2PSPLrR0awZUnUbG9DS0qSfvO6f/sO2PT6baGOVHW3hyhvTlydYVnKyTDiXM6kOZreZPwhG7qgh2rEmIzMnbKgsejvM9MiFBPV7Y3RWhvivI6EDvIHx+2P0yehCYjt5UGFqyJsGBN5LSfKZPyVGsznRCQHy3Le8/pIvrTUa4LHVlZwge0HZpbc5uCtje1rYyOdPjRjM0Ero4y5Js2J/U0MsBJQ97b+Z1vH2b00b6zdiIdL0YmNoppnXkflTEVY2fhIxEWPhK5dWZ0UXNcWfr1gC7YZ77tYOtqaG7r/D3hCTveBry5vCerdUs7bQkFGwLR1Nn+SN84cywvfzqe6fzdjHfH5qo6LYxim1FdyenaaFqxcpmz2bh86LdWF9F22HOyQJsJj+232cD6CAPrI8e2AnI+5uzVkG3sjDOG54xtbXi8NabQ3hihvTGXb/4aDc17n+V/wj7L/5S3YyfXESced2+enhmjaSYozCnN2HPSURlMdCWdz5i5nG1fOuaRl8f1o7WZNkeFNMlwkttsmZynbd+YUkgWxOnVKVykBkIfHYcGuumOYb2kA72kk9ezqR1+NjaXlcsD2h3vZi2Es7eJ/Dr43Ntw8Lm3OaSatjd62dpUZhxwYPQLldFWvsfK6g/dyOmMiO3f3GcczX3S6w9Mmw+ujZy5h2m7E4+7Fyced29qC3E2lUef0una39DztvDf0bqQNqO8m+w6nc6vyHvbHwTJcMK2uVOUb18TuUrloZX1VO4iz9dZz2PmXQZUhzl6lK6b+uuaXpeH7GfAldtW7i7K13tbW3KZSNc+aNsbvbfwoQgLH0rHoq070yc+DdZ2zfrEX0uw8xZiu/u8bXS0Wev0105omyUL4lwOCbY5tAYSHdX6zcalZp1OB5deeikA4LLLLrObcwBw7rnn4sgjj8TNN9+MO++8szKvp59+Gt/61rfQbrfx5S9/2W7OAcCFF16IpUuX4uqrr8azz+ZrSyeddBK+853vOJtzALBo0SJcccUVAIBbb70Va9asmVE95xuyQScQCAQCgUAgEAgEAoFAIBAIBAKBQLAdI4Gq9ZsN/OpXv8KmTZuwbNkyHH300YX3Z555JgDg+uuvr8zrxz/+MZIkwcqVK7HHHns47wYGBnDaaachjmP86Ec/qkXb3nvvbe+fe+qpp2ql2V4hIS4DsPeUZXdOrT8qfze5lLrzp//r/nEEKuP9qpjN0Wam2TXc+N4eoo1N51sfNkY9jZseEfqGi4kmHtsl/9Z4y3SVUz97f4i5IP6ZQXdbN3HTAwCeGUz/39Doelcsae+Zf1mtWtcmH2dFtHTgzolinVSc08LdEZB+lP6PvU9MAXHbo8n38jXeWEn+PGHyonewmb/d+2+Y8qlHG3OKi423TZ2EO3mZ9j2JbV64E9B8a7/PCTB9TnnPiYGu8/LH93JdjPw7IfxY2Ny9Y5TmpKkLl9JrcucE13TxoGbvflF0DGR/c31Paeb600HC94XNq11spxDdVfdwmTpF45Hln6ldA/Rz4J6bejY1qwUK/GPA3TNm+ofUmY2xzuUHIB4yjFT8xvHMNrImMK6dPiVyKXTXJODeUzK1a/7B1OOZRxCRPewda75nKkOXJamZ01GI9+6fdGDuX+yO5O3E8Z4dL4187JgyqQxTMX+XBqWPo6M7FB4TWgFP/3734ou1A2k+yOPqV95VSA6vsfUM3W3pkRdtaLHvHf2dwbYtPN6lPBmZ/4dlPeDJeL+fPX7xbQXKwzrS+X1wVOdRNmf6196X0tYsHzq8xYwpx1bw6J9ekL9TXeXc9WTJ4+yY7D2VqzkdHu1Er+QvvP8TFHSJSZ9l21mUOHrP8D43BqLNzTx94K5C7j4g7p4N3/Of0kTpDN7XZeT+ZJTTmv3P6Qdi80wtznmb1bFl92UiMLYYuTz+6EL7jLvTT0e57Ri695DWwS9LxYzNqvM+izY3LX9MPzFiP6HtVLAf6P1LjA3rP7cylOqVrP50DJXJsKTl3jdSdtcU7TvnnjDv1CkAdBZSz/wiLSH969hSQDpeTZ26qmD36Ii3GQ3G9k/yPjP9rV0+6yxKCvXg7gK0oPqH0DS5NP927HlGcMGOEyOXp3NnXnRGc/oqbS9PbhTeZ4h0uS4M5W/agT6n7U3bzOhI7h4yHfH9q5m7co0MCKUJ3g+ZuDSrjsptJTpejPxndAodv6H7I7nxsPkgwttUHnnfdoe1e2dkxd2mtq6G5g65PzEw7yrM5el3jFyLhzTPR43ieOPy18qbE2fP6Xi3aWu4V5fd98zpWh1pJJ7ZBJTbbP5cz3xL5ZmFWRsgZYbynlqS8Z6nbw0fc/ZqyDZ2xhnDc9T+A4B4IMuM/M9JS0VZW+PJ1dnCYpQ/M/34i1tekH8bus+ascOsvKA8w8grOp7ofMa2b+C+q7L71ToLkwKNTj5+5J+s7Ggs5a/QXX+GPjWtCuWrrns/sVrfLr73T7BmdQFQuA/VroXQNKZMMj986OK/LJBKZbVjU5XdLc/ZplT3k7p1hzIaY9i2pDzafTJfJDP9y81bAMJfUcCm8ujTyl37K/vWfA/AkWF0bsXZVwbReMS2uVOUzwfePcCWvxmeo/my63lMGfFAPqd3eJqOQZ/WCpqdd0RuW3tsU77IMr1L3o/sfZs6l3mbD85lAHs/cmiOYmxXc+IuLr4DXPvL520nTVLe9mbcy5108wWFuHaIy7rfhfG73/0OAHDMMcew783ze+65Z1byuuKKK2rlBQAbN27Ehg0bAAB77rkn+82dd96Jj3zkI9i8eTP23HNPnHTSSTjhhBNq5b8tIRt0HMxkgAobIrSVJ3z9v+eMJqBnAcjR6uTJ5VdRls2TWUBTZIHOhM3x8yqlqQc6OPiXtVsYWhnD1KfJGEp1L+CldQ6VWReKMdSD+TsJs//Pk4Ks3ac95NdrXklLW8OGpnfyCSw8Ft7XaEe/zk64Dq7MqjxD45H7lNus6uG98y3Td1X96bwvq1+NOtmN8H4uvO4FdeWaX+d+xhaXhjzzQ1IgUeWL5z3wRi30wpP+dz3ojZ5kfQUtM5UxBb3VKx9sI8xo7CqSvl/6++F35lunvcveqx7lTR/vbbgeZnHZ+SyB3QTxN3hmhDJbICo+V7qcT0P6sZd2YPOfRV01W+OI45PQQmxpPhVONfUyycuvnaRiHDh5EzsaKNoRvkNbv+hn7BloBcvPdHPYvqe2K2dv96jLfLmtvLXjMv6KplVhE9AZO73QQu3JWdLFobafsTwMzCNsmio+no35BOVnbuxFvIytK8NmTN8M4dj7dDxwTkXzZVPU1Ht9ZV3BT7NlrxVk/Qx4s5d5ZV36TRg+AD3N9R1dzIz/vvR7qJy688rQHNaDbunCBjWrJzhagFnlkxDqzmsr59IzoM9ZD/P+Nnlybd5TmT2Oh5BtORc2eHC81aS53/Q96RAE7JQSOnrliZDNUbXmMOO53GyipM37Wa8TzB400hCXdb+dKR5//HEAwL777su+N8/rhJiczbyANORmt9vFC1/4Qhx44IHsNzfccANuuOEG++8LLrgAJ5xwAr7zne8UTvHNJ+YtxOX4+Diuu+46vP3tb8dhhx2GwcFBjIyM4KijjsIFF1yAsbGxYNqrrroKxx57LBYsWIDFixfj1FNPDV4GKBAIBAKBQCAQCAQCgUAgEAgEAoFAsCMj0arWDwAefvhhLF++nP3VgdmfGR4eZt+PjKRRTbZsKbnncg7yuvvuu/GZz3wGAPC//tf/Krzfa6+9cP755+Puu+/Gpk2b8Mwzz+CHP/whDj/8cNx88814zWtegzie65MC9TFvG3Tf/OY38brXvQ5XXHEFGo0GTj/9dKxcuRKPPvooPvnJT+Iv/uIvnEsBDc455xysWrUK9957L04++WQce+yxuPHGG/FXf/VXuO6662aHOA17Ea2BIhez6oh4zGXPkqGEvSTVQeZp5lwUXBfUS82QSSIYBS8r1Vl9vAu0AbAnVvz6+Zel0suPVZznb2mi4QRMaEXajpmnCvUUonTTC27NRamFi0kNaczFz9DIL5M1dBoaSL8qnV+QChCa6MWwJk9FaOGQKIcWW9du+qOXrUbT5BJbkn+BtxRy+qeLl9CnBWQ/45mt8zZRmlwYrfLy6QWxlk6d5+VcrkvanLt4O++ItA2QuJcX0/b1+89e/O61s30fkTJVMT1LhhcKil7KbH/24uyYHTOWT0KeWfSi5JZOQwhktOlGzkdOuAeSp18mpa3slIn51vBh0oINRUMvjqbtZ/ioDM0J5YxHm56OUdL3ls/NmDLeYNm4K/ARVyfvknRu7Fg6Gi5PK+3yYXMrdWvL60vlnZXLhGaTZ9RR9n3SBJJm9j7j3airoLJfiAfZsWnGIzMGqAxMBjWSQe2OByaUh4oBtccU1B5TfJ/6Y8TvB+89d9k5d7k2TWfeaU5G0nYmbWLkwsCGyKUhyy+XS4RRsmeDa0kaImNouzt97o+txH1G+5eeNDP1dPqKgX2f0Ufbp7Ulp5XVxRk/RoSPlAb0gIYe8D/O3/vgeAMABtdFGFwXpeMw4y2uPymfcWVSXeK8Jmk4OcbpRysLG16Zpv2y/oy6KtVxPu84maW/BiMnWmOq1PbSrdwOUTGRd4xeSZqkLlQXMGM/pCu5NkOS/eh3ph00+dvoEpWN+Titn8mrNabSf/t8QHnSQ9RRBd62p4A8eWpkXWOKjDP6nthZNi/ONopyHZTbi3ye7Bgm9o8Zt7qpHb1mvyV153SEJamrWN4ybaOb2rGPGh2V8huhz9SV2v1liDoKzbEIzbHIaSPOnjrqxQ/bv6n+tXQASAY0kgF+PKkklzNWrgTEma1nBDQmFRqTGU95fBR18/FCx6huaOiGRryoW2xTndvgvk0IuPqjOaGcPufGkLUv2jr9qTQdTas04ZcFuW0Xc/3j6QgWRi+ZesCVp7bcTCeH5L1Bw7dfyDjXvjwwNNI86RxI5eOUtq1j25v8yTOHJiIP7TM6Dpk5im8HOHMo76Se0qltxtn21t4j81M6nuPRLuLRrkMHQMZpJgOiaWKjMbqy0KYefZQmRzaRuUIhDUFrS1SwSeIRmij/k/YD109Ultk6xczY8E+SBWxS+4mx2Qyf0Tow6ei8hdZ5cG2EwbURO3eh5Th0VcC0gx8az/azeR+5aZy5h2/vEn3A9ZnefcptN9L2jixppbLNyresfo4eNeUQm8JZU8h0BdWl3BiLphWSoSTVJYYmMld1bG8qLzI0OsqRU8WGLj5SOtcrrp2S8jOV0U47GTvWT4eMLmPTcfO5KJ13J8MxmhPU6ErTtDdGeT0Mr3tzDUp/oc8a3tgyfEJkgLO2BY8/GF7wZYe1s722pLI9pFP89QljE1l51sn1LzeXSgaTwthQMZmfMjQ5z+j8jaPTk2U54enPrHdw/ZFWCu64zPjBn2dXtZMjDxNl9b7JszmhyuUN0XusvTet8rUxY1sQmqMuoTVAJzcOnSr4fERoaG/K+dzYtj7PNccVmuMq75NAWVQu+W0YmotSGm3fMOsbVP+ntrEuzOWsLVAxdxbMPbblHXTbI/70pz/h9a9/PSYnJ3HOOefgVa96VeGbV77ylfjkJz+JF73oRVi4cCH22GMPnHbaafj1r3+NQw89FL/5zW/w3e9+dx6o5zFvIS5brRbe+c534pxzzsHzn/98+/zpp5/Gq1/9atx9990455xz8M1vftO+++lPf4pLLrkES5Yswa233opDDjkEAHDrrbfixBNPxKpVq3DiiSdidHR0xvQVwtZFuiCQlc6/iyaj8AaOzTT7HxP2sRLsgkNefugYeP68XmGFfLwj1rqt+bugSDF+iCifJjuhZJScc7cKXXBj6pcM5O3oHD/3FuroxMQYngD4kHJm8gySp4ZrAHm0pIZ4/q19zoTuC8Vmd46/mzK5O1C8dEBuuBtaDMwdEHYjFQCK19I4bcfe+6MCfW2QAMju/3DyYu6rM6D3F4W+oW3RM5K8fMpv5u/GWKN0PAeNryjnOWexw6eT62YuT0auhNLoRh7C01lUbuc0+TCLdgB/3wu9lw9w28qWS/qejrlCmcQQLb2fr+GWWRbaMeoyZRLanLsjAnkW8tf5ogANO2bHqcrTOH0TGoYl/Rdnm2Cqq3K52UQeLrnj8RBHb0YTnhnI/y4Qwf+bDWWS8HcpsH2WFN9zYc0At672Pbkz0Q/PRmP9OzHus2eTSxOHzwp3zHgLTj5NyYAbhseUrzge1eQ+t0A/h+48AdJ7BUpDZ2ZyI6G6SAFqyiWayiCgqEvZNgAwuVtxYTB0fx/Lr9x4Zspx6CN1CW0cAumE2CnTkErbhCnLeZbVhbtnojuiy8NBEh5IvE0SmjeQTUizPgneYevRlC4kFPMqbWeA5SMu/+mF+dix9pHH+3YcmcVBet9aM28fu9nQIYvalN+y9okb9BmK9ifte2ZizoV6o3fIOnYY/Yaxf9xCi99aPZ+ooh1FQHmH8jG9x5HqP/M9DTPk32PVoHesMEjaOg+ZTu8QYcbm7+5cxuZB71D1yzfPgUxW6Oxb0ych1Ur60+RvNoVp/kkr7zPn/siMjsZ4i7Vp7f0/DZ3LBtPORH849kfAZsoXjwnNA7rw3sr3sZx5uXvCKscoKYvqJe5blenkgoOTN8eL26QdIl3kTyJD0oyzT9m7q8ndMlSvM3e7Vd33FrLR8/T+RBjOArh/DxRNmzT5Mp35DBnv5tvm+laWXru2J+Ef+++qOUJgcdVP49iLzPycq8f0LkkhL+e+xsB8wJFrvr2dkL72ZXz2ns0/gAJfVug0Nc1vdnL3X7H3R5J8Q7rIXkMQ599x80buLj4VF20iWma6SUT+9pHdieyQy3wXdRQw1czLt3qFyEDCh3ogLTSJ8wbO750v0knLjQc0GtldWFQnOff6lsynYmLnsvO3Bgq6lOo357nVpcVxohuAAjMvYtLHzDoDknTeDXh3eGZ2YGfUtdeTrE3tXEW5uthHY1J5d3rm5eYPPXoVHKcIX+4paIePHTub5mGJcMuk7xLujnQyp6d3o1Gd6qwx+nfgRq7ty/W/zYd8x9okoXm4pwupAzJr09BsIo2E44USOLZjQ1tbwvR94f5X02emTQkNIf2jqM0Ed40syRywQunroLC2RCZ+8aC2m7XselOS17GwxuSXY5w+/HvRa9CvFdL1XC+N/SedF3m2ceEbVb6OJZhbaCh0aw40DYVly5Zh9erVfZe3YEF60fP4+Dj7fuvWrQCAXXapuOtylvLasmULTj31VDz22GN4wxvegIsuuqiyXJ+G97///Xjf+96Hf//3f8db3vKWntLPFcqW3ecUZ511Fr7yla84m3NAegTxsssuAwB8//vfR6fTse8uvvhiAMB5551nN+cAYMWKFXj3u9+NjRs34mtf+9o2oF4gEAgEAoFAIBAIBAKBQCAQCAQCgWAbQPcQ4rKGY08V9t9/fwDAE088wb43zw844IA5z2tychKnn3467rrrLrziFa/A1VdfjSjqfWvL7Ck9/fTTPaedK8zbBl0ZjjrqKADA1NQU1q9fDwCYmJjAz3/+cwDAmWeeWUhjnl1//fUzJ0CheEqKtJQJCePDhoLRKHih+Z6OKua/CXnvc2ExqtJSWrjQhiaMCM1TaUAPx9DDrssVPYJtwqM5oTdoPSMNRDpYFxOmh/NyZo+9G+eTkvZx4IdGoCH1AMS7dhHv2nVCXTnhHLKyaPn0qLvpv2hKIcqOyZeGGVDIw6bQNuP6kYSPccJ2cM1i+pbkZcM8UJqd+lU0Ihcqq8Izplaf+GlKTlwAKLQ/1w5++AET8pHNS+ftorN2Kg3Fwj1XeXqufIpQyMYCaQE6QnnXCaXh5O/RXMjPz8Pn/RAozdnfTp3L2tYvsiR8rG6gPMQk/TcNdUbGUWNCoTGRp1eZRyr1zDVlWY805Y6roLytGhuJck+HwesL094V/eSHM+ZAw+v44WlcolDaz1woGZuOepz6aKBwQtfpO6sgc5ramyO0N7uhkyr7maGTTeeHssnec/pHxW5IJfb0WZnoVN7fRkZ4+tfXwTZ8SyjPhk5/5vuqEJWAw1OFLBO3HmxIU/MnRxYXAoeUz6UJeX87Y6wPWHuEhLDk3vvPbFsZm4KYOo5HMJFLpbYZ4e3gd349Q38zaZJdum6flrTX0LMKQ8968obw3Mg+WzCyTxrL34agpH3O6BXHpjW6tGLmYE5jqRg2cfo3oT8UUonUPSR3HTkZkV8gnfmOD/WeP+NkcDyU5KESPZ1Q6TVMaCqVHYYfGSTtBEk7KdBqk9O+MRjtpL8aM7x4QYw4CwtZqL9nR5bNi9g6AYji9MfRzIY6LZy8zH4LutALuvBhQtE5PEDp98dkj2ECDb2VqLDjbDskKh/PFfYWaxNUyIugfTCd/pwknH4nfxuaaThqh2ZTJtOPfrhEjl425H5VuESSv5UxDHR2CilkJ4fmIFwaJwSZyZ+Rm9Q253RNT6B2GDO/pHQFUWIThWwWVi5X8JbDp9m3nC7i7PXCSWuTl9HPnuxRi6fSX931AFp8yLb1kLR0bieSb2k9Le8q4Iy/uAtn/MVdQX5myyS8UVg7UmDnDiwq2kAtnio+C5mcJkQyU2wyEjt0siDhlllSSZvZZ8QOdMZeO0l/hraK+W4wdCODkA1snhu5p30+rtIVJeM8HkkcOWLmBdw6lbPOQGgtzEsBJIu6SBZlepGGRCzREyZcYU+ga0OhdSKmzKqrA/y5qqZ2HPL+ca7toGUxbc5d4WBBZUAmS2g5HG/3s8ZFQfWusbn9tWKnn70QnVXlF+Y5np3I1ilWvDzy8i3wXuC9YP6gUT/E5Wx0l9mjueuuu9j35vmRRx45p3l1u1286U1vwi9+8Qscd9xx+P73v492u11dAQYbNmwAkN95tz1gu9yge+SRRwCkYTAXL14MALj//vsxNTWFpUuXYt999y2kOeaYYwAA99xzz7YjVCAQCAQCgUAgEAgEAoFAIBAIBAKBYI5R+wTdLOD444/HokWL8PDDD+O3v/1t4f0111wDADjttNMq8zrllFMQRRF++ctf4tlnn3XeTU1N4frrr0ej0cCpp57qvNNaY9WqVfjhD3+IF73oRfjXf/3XGW2uXXvttQDyvaTtAdvlBt0ll1wCIO24gYE0Zvjjjz8OAOzmHJDueo6OjmLDhg3YsmXLzAhgPDJaW3LGpqeU8mfEy53xNPBjgvuXHwO5hwjr+U09BY1nGfFKCZ2+MM8570l0vHj5mWe0moigTCx9xqPHePU7p7T6APWMox5Fhe9icnE047VYQJL+rJdKArdPOxHQifL01DNOg/UOck5VGO/SbnYP4ADv5kLzpHeR5B/kNPmedw6/BGSq0ye+h71yeca2BeOB7aTP2s4vpwzBfA2fMi4byXA507AnMXzvHa/MiB4IZbrE4f8EhZMyhZNNBQLCtPre4FUnDJw8a3rXBj2VStKnns86vzsvkK/1kmI8WllvQ1omkR0Fmhje8etBL8T2y0y9Bd2OdNKHPHqj/Je0stOVhs6AFxjXj473LvEwY0/f1oRzYsekDcgdtm4heUAucbbyhOmH0ClvLn/KD9bDMOCxrqZV8f7P7BRSiA/jtkbc1jzvwG2fqnYu6AP//p0sH3ry2aaNc3leBcfjk9Bs24fQ4XuDFyoQq/THIdPx/v2nZSct/fRF4r30FR7BxdMn9RmdHU+sJ2dFPiFHWKvf+ZNrTj2Ntz5js9CT18lAkp86JifoSk8ha7C87dDky+g6zWh1snJP/Pp5Ebo6izQ6izwZQN6PrR/G2PrhcJGcl2ycP8+9mOsoLGR2iLInVB3aS/SSM5Y4e5j+u6vSH20bLluSp3tKzDAqeBnc0ukPRd6pPAlJdGXVmA2e9NEKKjt5rFvaOZkEwPUsN88mmtAT4SvGqzyn7XcxEC9IEC9I2HZ1vPADXu1JA4V7aNjxxPWbIhE7tjShtjQzuojtNx2lv8x25U6AUp1XuLuqzP4i8x/ncRkf0HyJ3WDawTllVTV/4sYIyZeL6OCfgLQnLbI5iH8KjD0Rk9FvaNYRyIlYUl8z12L0d6UNrOHON7i5ZgmijmJ5lzuRQp/b4gOn0bk2CdJkaDb2ph8RI8DTVSc2fVDbrhcU7AfKa6G8Mp6rWsdwC0KhTkYXBWH6vq4tY+TJ1haSra36dlAoS2rbFuYAymk7/+SStdmz308ePRw/efRwd97pzRlCdXJODxF7UUe6eLdUKL+ScpKtodAyXD4pQSyfeXRySbn1A5cY5DrRT+/NydVUA2oqVxz05BEhNUek3fFWItfp2hDXz1RnVfJxCZzoGEyjRt30zrP0Ps2cfu5Uf7BPsjlEyA7l0pRGfKqDGM78ytLCpGf5mL7350heXoVTqf6Q4HRotkZZVScjS1I6c97zZUsvcoZd7yJDUDc0v8bI8HbUSX89n1wjUcxUQBdXrduak7S6RLY4fSeYV2zLDbp2u433ve99AID3vve99p44IL2K7J577sEJJ5yAF7/4xfb5pZdeisMPPxz//b//dyevvfbaC295y1vQ6XTwnve8B91uvgjz0Y9+FGvXrsVb3/pW7L777k66c845B1dffTUOP/xw/OQnP8Ho6Ggl3Z/97Gexbt0659n09DQ+9alP4Xvf+x6GhoawatWq2u0w1wjP4OYJP/rRj/C1r30NrVYLn/70p+3zsbExAMDwcHhxYWRkBBs3bsSWLVsqLydcvnw5+/zhhx8GSEcbfp5ekEsgTtlE03xYABZZcrsAnaHsAlw2m9Alzdy3jICOOnloMasYzEKHRyvNJ9pKjCbGgGcvsqfv6QWpXvkOVF6mXQigk6bA5bhOeBSgcIm9ucibXnbugAv759UByC64B6CmVOlkVOm8ztaI9L8hG2l+uT6flMHZkOPq14cipeFYQ3QULgxXpH8YBa6myitUZ1z4/eqHLLTPA/1ctTBchUJbBMovzyTwuCQsXE/Z1whFVpkHNVTJZk0/9PRaZmNK5bxVwYPsxFCjsKAJuPTnsi97RhY7NSnXubSelFlaH2Zzoif665ZD81VMXowc554b+DKUlp20dWEjztBYJKqYnpZpQgiVTaRDKOger05cSCiajsrVpOSy7CAt/ndl/RugiVtMduCH2vLbnS4Ees8Ki7VM/mV1Culv7ltOf9F2LsszaZXLYuO05Ke3z6ochMh4cHSpUcnTOe/QML0NYycwk2P6PMSjPdmDXD8aOjY3WBo0k4baqaZNHSeVzcTkn8Gk2t805vLj6GPfB94Zva38jXZTZpav2VSvswHm2wyAZx8E2t/SRMYW4OkCzq6j/RShXEY2wOdBbCVOB3G62Ni4ddo+GmsUntHvGluZcWDsgC6jS0jfpBvFOvie7TNvE6GxJaOP2JhWXmsgmnSNHN92LIzTXhaUuO8UinaVtw5bKJO0gzY0hPIHo8tqzBXL9I6hO/SdVrBOUtSW9x1Py2guK7uWLDRdSucSXFkVdlSoTThdx9kCwe+9Z2Y9ncIZDyHMQO7WodF579nO4Q+z/6vApwH7wtLAzDGpLrJJ69qNCLdjNN6/X3loDBX0a4yczxgZr5Uri6f+uKDw3m4otjSqeNZHaFGdjlHFlMXlz7VXcOyUNK3REyFw88ReUOgDf40gZF/Qf3P2YZmdEXA2onLP4eO648jQQOzJaCux47w5pKXFZMOMW2d9g/BmtIXR3xV0RmSNTnl5BjeU6VyWm1sx+sGEHe0XdiMcNXiW2rlbwzYNfd7L2CjQFfiWfebrUU9/F/RP9nl3hJffvi6u0gsqzttJEblU5uysWzrv8wp+Ctmugm0Djfqbb7pX76AAzjvvPPz0pz/FLbfcgkMOOQQrV67EmjVrcPvtt2Pp0qW44oornO/XrVuH+++/n73j7Qtf+AJuu+02XHvttTj88MPxkpe8BKtXr8a9996LQw45BBdffLHz/b/8y7/gi1/8IgBgv/32w0c+8hGWxr//+7/H4Ycfbv/9P/7H/8CnPvUpvOQlL8F+++2HzZs347e//S2eeuopDA4O4uqrr8Y+++wz06aZNWxXG3R/+MMf8Na3vhVaa1x44YU2NqlAIBAIBAKBQCAQCAQCgUAgEAgEAsGfK2brdFxdDA4O4qabbsJnP/tZfPOb38R1112HxYsX4+yzz8anP/3pYLRDDrvtthvuuOMOnH/++bjuuuvwgx/8AHvssQfe//7341Of+lThdJy5Lw4AbrzxxmC+Z599trNB94lPfAK33nor7r//ftx1113QWmPffffFu971Lnzwgx/EYYcdVr8BtgGU1nq72Pd+8skncfzxx2PNmjU499xzcdFFFznvf/jDH+KMM87A0UcfHbxMcNddd8XGjRuxefPmyhN0ISxfvhwPrl+Pff6e35EVCAQCgUAgEAgEAoFAIBAIBAKB4M8RT/7PC3HIkiVYvXr1fJPyZ4Ply5fj8fFn8P/7+ltrff//nXU19h/eU/poB8B2cQfdc889h1e84hVYs2YNVq1ahc997nOFb/bff38AwBNPPMHmsXXrVmzcuBG77rpr35tzZaBHgendGPnDHmI4e0fx+0UUCOFVP4M+01XRP0v1s9npHto2kN59gFmlb1bzmk1w9eyn7jX4pOf+2S4kz86P0B0+2xPmmr7gvYuFD5GPjZr3wc0qZlsubQ+oqNOc9n2g3Mo7rbYn+O23jXlye5cdfy6o2w9z3l8RepeNCj3Ltp7q0EPes9U+vfTHfKBwb8tMMENd2HOb17BZa9+htx3Jrn54b1vJ31AZszrv6iHdjta3PaEfO69P23A+9Tdbdo86gKV9O7STS+WtCZnfI82V91PWzmgW8sAs6pO6mI85WAbKe3M5hma6tpVmgm0mT2Yt/Vyj7rxzDupR1Ze99HXUVTNf550FOgTbNzTq30En3b7jYN6XycfGxvCqV70Kv//97/H6178eX/3qV6FUUSAddthhGBgYwNq1a/Hkk08W3ptTdUceeeSc0ywQCAQCgUAgEAgEAoFAIBAIBAKBQLBNoOtv0MkO3Y6Ded2gm5qawhlnnIE77rgDr3zlK/Gtb30LjQbvpjM0NISTTjoJAPC9732v8P6aa64BAJx22mlzQit36a1z0fq0Yi8b5sCewKtKw3jZxAN1XHi9H4GO+PSO95Dx4iJ5UM+uSu8f3wuqwpuaemmZvLVK29/vA93S6UWmPv2kfQuXr5oLlannknexMOe5VunNVuHpWfDYqeAT/3RmXbDlzPR0Z4UXn+2jEs/nQvmz4RlYE1zdOa+kGXs3MxdUV6JHCWx5ifJjiaehbgT6hryn6av4JDTmTNo6zwrvzS/0vkxeBGihaVias/fG89Vecmw+o3/PkpcdJ2/o+LRlBspqdBQanTAhjqxm3lE6qjzzqMwq7cOKU4eGJj2g2bwMbzon1LPvkqHEedZzPzCn5EKyyl4w7+k355sSL+kgHxr0MMbcjOHUQ3X59MlQ4rRXCNzJQb/tKa11bRqHj8vKpLrItHmVh3nFCUJfhpXSpALf0k/qnrilaWbpdEzo9IjpByvLQuQxMoS2de2TVb7NZpBkP2a8+vXwEbKfuDprVXxfsKMM6spq1QM/A9V6tVHd5zoCdFNDN3Xlt82t1YSV8gydj3QUojJdEeBX/7luACpOf6X2XLCgYpv7/FiAb6tNK0TTyikzaWkkRuZ69nylx3mdOZrJr0IWBHVyxju0b3riPTIOdM02rytPuDyMDq6jnx25xI2Dsr710lfqdDK2fdmlq07zRuVtFpyP1bQzuLmqb2ex5dY9GcXR0KMdmjTTHy3b/t2DjC6zoYMosetN+WX9w42XOu1WlieVsf2MJ/Z9BCDS6Y/akUZ2ENuezZPjU+WOLfp3Ib1y27KnuWcdm8m8ojqA6/co521u7NN5B/deD2hX3hpbg7YprZ9Hf2GOUGcc036itFC5R2VQyfy3HzhrWuRvzk7j5kg9gbYHNzYD7cXyFpc+xE8V+VeS7dskTX4u6aQpkbGhNRHWtuwDVTo+aenqOYbVoRq6UaOjQ3Ups2HJuKqaE6hYue3jkz7DNhPMDBoK3SSq9dPSWTsM5m2DLo5jvOUtb8HPf/5zrFy5Et///vfRbrdL05x77rkAgM985jN48MEH7fNbb70VX/nKVzA6Ooq3v/3tc0q3QCAQCAQCgUAgEAgEAoFAIBAIBALBtoTWqtZPsOOgWf3J3ODSSy/FD37wAwDAbrvthve85z3sd5/73Oew2267AQBOPvlkfOADH8All1yCF73oRXj5y1+OTqeDG2+8EVprXHnllRgdHZ0bgqkHh++1AyBp1ndn6WeMcGlUXCdh9q3xAKJettmpPzdT1zGmcHJQu3mwdaEZ+F4snhdQwYuF1Ml5x3l0TfMNqX2u9k/DGc8V2yhemabOvldoCDW8qEs9l7g0vreTISMheYbSZWnstxXtCLBNURtaBejy27UHcLwx0zT0PTdefS/vqvIL7/upb48nCW2+NH9ymrWMJg4qdmVDFd2hMefQVvHMf186tBRKZViQHpJGl3naxfTvNC/d0J4QLCEwQBcH48nsyG0ubaC8uF1OiBl7KiFe97RO5jsFoMIrr8Bn1Ds8JP+5U2eGjilG1yD33NPI28V8F03kwoTygYX3rKof6EkEU6ijy5n6OcWV6NuycQGgfIz2gJDcp21laUqK37MnB33M4HRzsI9NnrT+ZowydFJQfnZoy9JE3XI7pHJsUT5S4D1WGV3sFDVLNp0j2xj9bfgsxEfWE9bXRdm/o25Nemk7M2OM9iNLR0TS1RyXobIqbcw6z0veBctJiu9tnyCQhilTZSe6qr7tjlRbX5qxTblnCETHsGmIfLdyN+LtcbbMmnYW1+YOP9eYwxhbzTlBQE7J+TpOE9uOnTcl1X1h3zOygNIR1Mn+2A3Me+raDv68jEOpPCCnXgsRRTL6ymgIjlGm/9h5h5NZ4G+uLJqvX1bVXKqin1mbIkBTkB4jA81DegqmXjZhzIA2A7/NqJ4Nym0uH86+qZr3MrqIlSGhfgigir/K+pzKWFZu2kK8x8w4pmPLnNp12tfIjipZzfGpR5NvGwPIeY3a+5zpEhoHnJ3HtYWZV8Tl3UTHAycXtMplMys3phSvt0J6x/u0MJ7r8JT2ZDCXJ+mzSju/BM7JQmr7Mrat0z4VtmftsVOpQAKPTZ9WzI9A3/ciuyrqZ07KG52vatzJVmqf0Dk/kSVlNstsIpoO8LklirS1+Yza0xwC9kVPuqws+4au5kPBvCKRk3E7HeZtg27Dhg32b7NRx+H888+3G3QA8IUvfAEvetGLcOmll+LGG29Eu93GySefjI9//OM47rjj5pRmDjvKhnTtxfOZGgE9YM7armyBUaO4AOcv9DITgBnRWjWRLEvnP+rhzGsv9M+0L3qhq1Z+s7TgOdt5bc/ol+b5rmvd8vvux5qTBbuw50/aZomuWg4VMwBHcyg8Ru225BYXAu/7wYzahJtM10nmbz72mH4m2JZjrY5M9nmmb/oqFkpLk1Ys0AVpqloIDmZY8u9taP9UofaCOUHl4mXZouQMaXI/7C//vsqaYzjydL5v7a6ry3rop8o69diXtcrugz+CPF0nBFSNfOYjfa1NLO7fdVHlbDFLcq3vudKOvuhnFtqLezHhJP3wzw7YPpUbtQY91G1W5W8vzhzct2RssZuRGfpZMHeKrtoQqfh2xvPfmv04G/q6pzxmaUwkzfLBW5uPK6CZTZZedOWsrUf1idpjr49N0jJwG3KzVf95acca8rAwR+1FRpZsYnNl1c94lvIRzDo00jvo6n4r2DEwbxt0559/Ps4///y+0p599tk4++yzZ5UegUAgEAgEAoFAIBAIBAKBQCAQCASC7RESvnLnw3z7g26fUCh61wcuvfX/Xefi1rqXxvvll17kydDcz6Wy/sXo9rJVkz+5CJi9hFS5z7hyG1MKjSnl5mvAcGTostNQm5TWVQG6raHb5KJZ/9tEpT+ShuZZSK/Atr9TrGnHiovLe7pI17QJoc9cwhy8zDV0aWxJuaEL4v1v/O9K61kleQJllCahHoW0+xg6nEtvmfxrjc2sLdjLhSv60Cmba9PAeLY/eslwFe9VjP1K2VLxLTsOTNtUeHArnY8n9j2REaEL0qvoLW0fki93sTSt10wv5nYubmbel2bTS7mMPKL1UjrXQaH0pZfO99AW9FL1QvsqIOooRJ1wA/TV5ga9WjdEx4X0UKXcrurLwJhmPw21fxUIv3J5UjihXXyaSvQBOw4V87dfJieX67ZdgZCKdKH3ZRenN8rlAZs/anxbkZ59bewsepl8RRojn2k9IiccIOypVZa/AnZWaZ+yxJO8sl9Q3pTl0Qtvqop+qBp7nm1WNvaiaZWGKgoVRXiIPiv7tiqvyjRRxftA3o697refCc9exza0D6vL7UXvGR0R0s9Gx3D83Lfu4Oglf4doCdl0bF+VtW1Nm4cdw74dGpDLhSLryjAqI0pkaei5Q6+Zf4XoKsm/MMfx61dDf7G2bx82mdGjoblqXwj1I0VVWX465noO87dvm87I9gLftuw8o4pPfH4n6w+9yLle6hN1lZXxlfNOIqPyh6TMuAYzVaDUHme+c571Ypdk/OTrLWtLlOg93YAr1wrEwZ23+jT3OsfyeL8w9mqM44IMYWiuOz9m7W0mJLtDmz9frtnPtMx+xmmZnVMrr15kXK/2g08TZwdSPmJoqbIDrQwh31EZWNmmPegIX67qpq7WtWYc1h3vJfMZMxcoK4fP3C1HsB1Dpyfo6vzkCN2Og3k7QScQCAQCgUAgEAgEAoFAIBAIBAKBQCCogurhBN3MHUYE2wayQcfBePURtLYodIfTh8lg6voQTUauJx71CiyBiakcjySIxnP3BXPSIJpK39OYz43sWRLyQmYeV43X1pYIndGsLtPps3g4QWOCOz6Q/W9a2fqracVfXmySaCDJOMzeNaSAeDCr5zTj5t/N621fJQrNyez1UP69oVlTjy2dpzd3U0RTkXNhc2tj+oHOaEta2rkI15746JL8STtEkyqnH8D0rrHTjxwa+42n9D89zMeJt5cncw2J/K6D7L1z74bO2z8ZyOrchcMTlk9DHkumH5lq6AbJi3hduXHqs39kdDWmVN7PHVXg22Q4RjQWdktTce61ZmkiF92n37htoWIFHeX9bNo0mowsbZbmSDuXfBeQKFsXnsDcA4q77Jxe/mzbnsSgp23Htj03nqkXVisbg12V0xG4SLj0wmQNgLuwm9BJ/7Zjjty54Iwd8nehTiRveueVKjlBFS/uItrYtGWlGQTo4+5XUnk/c7HsFeGpeDglujERWd5KmjqX17ukDeWMdeXKNkMTBzM2ufoWZCWRZ/SbUP6KyL3muML0AiNji99qBTSzOpgLuJ08FXk+kOmH8YZth+a4QndBYr81dNI+t7LbeNtS3iRldRflzBd5ekcluYxGArZNKFpjmTxeoNnvrMeokRuKyA6N/K4NTiyp/L2lI4IrO8jYB4p6yablZBin10mfhvrenI4y9aD3uamusrrMZtlA6b1EKgGSlsmbIcn36jd0Gf01nXvoGZrjocTKYB0V8x18NkJn18y2YtpBN+Cc+iq0wXRRv7gZ5H2qkoxG5G2WjOYyRg/HiDPC1WSaSHWU249UHyHrQ//Sd04/MHTZz73kWlGai3WLunydjf5LdWFGZuax6xJT34507ok0Mo7wPivrqQ1tdEJXlcow59s4lTMAML1LcTw3plQuo8jpr7L66Qj5KYdIF3QVEp6uZNjIwMj2vdEV0VTezs54IDYBPcFITxOb7zj7waHbtDntb5rnQD72rIwzPO6dTPfbP2nmD6OY8Ln5PgaS3Trp+7XtNC05YREP52M7J4hmUKyT8uZXBTuJ0KAjQO0/kb5/Yih7lvOzbgDNLWn5xt6kcpnrz9aWCNO7JE6Z/rd0XgeAlfPB9AE7zKZJclvEjFffTlKeXHfaUfNtpXz9BLdtOfuIprH2zdYGnz+jVqlssGUZnUa+9G0+P0+aKbWDqG3sy1M1rXLe10TcmOgv2tUn7AkXarv57+kzRqzoiNBn5ipkPCtqHxCdUSkDfRpAZJxvq5fwGZDLUDNvVmR+7dv2QNam1JbgSGLmoJzNY3URiUbDnsTwozt4dXagc8LonN6hj5MjjIxNRsL8Hg9oq9O5CA/+eDI2jV1TIDakY4MH+IDVob5+QrpmBGS6iJl30Xl4PrYCioWDP5cCGY8JADOeOLkew5HLbN6WJ1w+tJ9QGeXZWb5N5a89aWqjk7zK7rlOyAoonQtSmu06UqzYu8HKxnPSojJMQ2UfmfW81iaFqd0SS2+BZjpGmXWYpAk0svUoNDVLHwdKa9LObUYA0Fmkp/QlCrajbmg2X1OneEjntiGdwzA6WSuGVk/uJmYOOhVldMKxt834pLLezLXNWqajB8i8Uk3nhVPeK7NJHFoZODLO/E34Ml33zOtX+Fbla8CGD+Ihvs25uy+pDDVyKRnUUGZdOUsTdfL1Mq6f2s9F6OxKeNPM8YztyenLOjpNMOvQkDvodkbIBl1NdEdyAWmEvjO57jW8EYDG1sgR/EbRcIapWYzod3RxY3d6YZIblmYPYzxyNwI8I0k3NVFqvOLijCj63oBdVOOM6oZGd7iYPrihYebp3kKcQXfEK5fmSYwgbhOCLprGxgjw+pFD8thI+keL70Bu88C+Iwa4phcaMzwX2jAtbJz478v2F2tcRq29i5bjgdxI4/q5saVR2mZ0sylkYPv9mrRdI8b0Pzd2fGPePgvkXSyc4X06OeD4OGQs1wwVoWJiRHWZMdiHbAj1QehCanYTrOy7OmWW0N94rplP5KhcouOxJH3ZuLL5mDnt1nTA64Yb/tHwQmNrLvfzlxX9R/lxkpHvdJJONyE4OqvqkSWaXqBLN9wBspjJbk5ni7UAMJVtXDS0bYfOrglrpLPykuTJ9VPDbHYy7Ug3AQBmgc4rx1/IdxYuSDodkgdlfJQwbRlYtC3daPfau0yvs/zuLYyYyTXXjwmnayrsFGfRsUo2UN42tgvT3s2xhtUBKi7mO7FnUrqIwi0C0/JDm3N0scgZW55spzKmsSnfiE7o5gadAPsLQ5yu4MZACdiFFbpY5i/MtPgJO+uEQNuP0MrSV6VDPDrSPMvTsptRFfnrCHZxgLM3je3lpFEodapJnX6K70My0i5yUIeMrC5GPlY51QAAt3neT9tzocW0AqKJnBbLk4GQ0cXNbUBHxbHr5E825uzz7O/C5lwhg2oaSp9rQD+ZbcwRWUntkHgofw4gKOPMwllnNKkv4wJ5lel8Tsb5tpH2N/T9/H0+rWHbseOQ8htnx1Mn0C2NnL4ALwAI2rk5HWTRT4W/Cz1n5T0jT5MBDW482fFWo6z8ZfFvx77g2t/jQ5NGce3DbBzUlrUo2WSoyMPOdQMOc+w8rmrq4831CvmatgjootJ5H7dgTR9F5EVI3tbs58ZmYtv6crELGGUcHAuE96jjMOD2V0z4NDhHDG1Get81xnJHJ5b3Sd/Q9ZPaYD61a0RUf1fNZav6pob+K8hoTwb4/RLcNC3hhzK7037TZfjAnzvUyJ+Oh27m4NMdyp3cnfpz8xaFghOGivM+rxo7Ifgb0P7mZ77ZmvMWt45kDjAU5HcPjmrm37TP6XorUBwrnP2vKtZ+cscK3sZ36O8RTpoyPRooE5rQ3S75Dqhcm7M27ETuSGbHRUO7fePJk6ndkjx/zR8c8emW3Z/5Q+jwhWDHhWzQCQQCgUAgEAgEAoFAIBAIBAKBQCAQbMdI5PjiTgfZoONgvOkdL+E81JkJI6BomBzGm64Kfki6Um+NCs+lftKlR9WzuhjviSYKIaCc93C9eLgQMIkNEakK5XOhnGrVgfNqLPsepMq0L4kXkg1jR7zRVUKeU88cWg/Pk0dNk/YJeJeYI/fUyyQURtAHfeacTmTSFELzeGU6fUPB8ElZmEK/T3zvG+fECWmz3JNQ8Z59vXiWe3n6oc7802aahP3g2q8Obzm8XzbeVeC5/1lg/LNhWsjpzsLpwBplsN7KdeVPHfg8USNPGoaI87ilp19YPqSh1rgyFZ+e/5Yf72UhNCvbj3p/+mMSyEM0dVTwdGutckDkGgkHGcpLl52kop6yTEyixmTEey36Ybm8PLmQsKw3IUezL09QHMN+/klLuzxlxnmITr9c3+uyjKfpiV+mH/sJlcLBPxVY4O2q/Dy5VKCrTy/coF5B8WSzHxYsmkYpv1bJUjaUNkibVISTonI1GdRQnTxfjqYyOVBH3lk6uDSBehb0ekAGOafuyPv8tF/2LOQ5ztmLVe0fkltmHJbI9wJIn/nhiRz9TDyznXasaksiA2hobCCTB0z9TYjKkG1RetpeEbFWYV+wp3QJzdT2ZqNcUFAe5cQx4ScnNDEjo/M5grbvqE7sSd55dWJDgfse6CQsmn1P9XOFh76fb9Sdma3DhekreK2XzLsKfJx9x4UJDNWZDRtWR/absnx5QsJ+VZ1AZE8IcGVTuz+QHy2/7tzQfhcr9rSe823ZHKYCoZDxLE0V8wKn7ar0BtOmhfBidVFi57FtFqKNCXlXd67khEv0bbI66Wib9aIrK+xI7iSgY4eY9FwUBy8UpQ2DaNZOIjdN6anBBolAQGhiZX1gjObXgeRpuKsVKuGHlQSh3ZtrlZ1go9FeKhHoU39tKQ1DH9bV/tiyVxaMM5FPSNlWhgT4sjRkPbxn3nvaTk7oQXJqjcpyG5a8lfMTPflUOMFI5U1FdIyQDVzoJ69v2XHM2NvsicWAXuolzG/ZfLhqLhyMjsDU2Zlj97iWGwRnr4bsPJqsLFQs933ItqR2PzM3sOD6yBvD9m+O34yMmYFdJegfGkCc1BO4M2VpwbaDbNCF4CmW1qbIhotw7ihjBGDVIodRBElT5wadyoUxq+iokUgnV1yYQ8ZgcSYQWb6NiSgP42SFt865QhODkyySmDSNyai4OBHBhtdRdALFhAmgC8lsbHpDhvIWUWwGOZ1cnvndU+63pk5RthCXtPP0SGDvnlM0tB5jEJnY8kmbn3TQOsXm3kISO94xRulkxNBRYmD7xrBpWCfEhXlPywrIcOXxkaaGH00XWJywBqfZMEpUfgdcp7iIpVuaNyipYWuME9petK+9xen2RoXJpbrwre0nEtBdJaoY6i7x5gr+BCTJH+qGdtsfcDeT2nnM7zJjkS52BUMC0cUsJtxj6aQBAEy4wlB4L47nKhbW8gzysZkMJ3mcc2ZSYrOg9VAoXXzQLTrDyct08iujWeWb4xGJwW7lUZuEgySTXyq38oX8Ip1akXvAuDuhjK6g91iROtg8wdxvQL/z+omFaaY436jON+3Is1jlIdBCY9Cvc5xvIKoYRcuhajGNlkMXBQ0dcTHUC70PrXKhEyjeu9QAlLkzkSy42Lj8JJSXSmDvSim9z8whgPxNJyfefZy2fBQnZz3dR+OXY/5N/u8sCNCyrE6sXiCzi02EZ9nFdxpmqU3GBiMj8sQ5LYafoo7K+7nCacNZ7DL5dBVU2SZJBCTkbjY69gEgGUrs/YcL9t2MLU/vYr8Fcrlhq0MXrODd5eB/Q+gGiE7zedcjn258FDb34baD0w9mXCpt+VlTXUhsy7LNHu7eOUOLoddfLFLeGDB3FVMbr3Lhg+blhXOiNCt6Bx3ZdCwsYFFEuW1nywBphyahnyzYOGFk/cUJ36nDt/Eb+d0saRlFwpzy/RBWCRniTW3vY2lk9/dqRXRMrGz6Rhb20oYy9kD7wcqDCEV7XpE7YIyNR+qZDCZ5mGLSdmWhPwt2TEC2m/dWf5s7SptuCOeC3CcLybqtgYxu29/kbs7QYllBLkfEtiM2m7NhTuYIvm3qyGV615+5g0653+ahpYgu4brS1In8rZvkObPQ7rCsGQ+KyMhEFe931sj1muW38n525AYzNgA4doFv26qY14+2SaaBmF5PrPJ0Jh+2/UmerO3JzStpf3uOAZTmZCjJ70oKhbWk44RTxaad4qITQQFl8o6k081An5nvqH7ybQrAdegifF5IH8HKyIQ6gxgZzsyVC3ME4xCQ3b3qzK81U6eE5w8OKsl1NZ2f+nWi4a2juJi50m5YzGYW/t6mc2SktuHtaTaWPyIQBsu/MTYRvdvNOkIlCpy8za9JIc7jnBMJkZEcKO/ReYed60S6YOel86rsb6rLvHzTP8j8mbNz6Hg1j6hNlH1TyJvIU3rXdBnaG1NC7N3VNM/E7QdWV5ENa99+Slo6vxOzq6x9a2wzFec8Q9vMhnX07IWCsyjRNbSuVXPigmxGXreCnebPZzTyuwjJ2LB2NXEadq7loDKCvSuRVILqRerIZx6qnDe5tUHTpuy1QWTeRzfGSx3lFPOMvvb0C52D2Sy5ORR1UCEy2PBHczx9xvGmn77MPqBrU5YGKpcVrNw230UdlYeSJ++5dSZ7t2lJGwnmEBr1Q1xKH+0wkA06gUAgEAgEAoFAIBAIBAKBQCAQCASC7Rh6JmEhBNslZIOOA/VgzNBdkHswWk8E6hFRcfkxt2sdTSsnTZ1LbZ33gZ1wLj3nNZcM6MIFrr53W0I8TQ0UDfnje0klxNNVF/MDkJ/OoJ6ujPeJQz8XjoKpPw3tYL25vLobr5DYnJzzPB2Nx24o7Eh+yWpWnWniIc148imdX+7M0QO4XljFSrn0mWeOd06j2D7UM80+m+aZq1K2V4RPoB6HAICGdi+j9VDrsvCKMv12nNwtcetqPBi5/kSRdwsnBtn3eTuzHknmb+o1X3HyvLLtqTec9fTUhfdcGoBvf+c7+m1NWpxHxitwqpi4jiyKJsMNRE+vOB659iEpg/N8jpGHRSFjx3jROukZGUfzo6edKcpOXCkSyoSTR46sJEO44AHp9VMZnFBsxsM80c77kGzMP8roI16oRt7FA7n3rJVbgPV65OCUQ9rZ9C899WfTNDydWFF/J4wwUn7kTqFbj1XiDZ9625cXUBYOiv6bDQ3t82j2jgsRnX/AlOF7LXv9R51LkxYZK4SHS73NFfGErQoPQ2izfOw9Bzydzuj3eFg7J5sKHr1EV3Chd1g7hKYnJ5+h4Xj4AqluNnXa+sgiRF75hdNmnsdvnROXVSfHynhLc7qIRk+g/WBOjEZ+hobW9P/BkxRc6EumbR0e4vJBwHu5qqlIm2svuoNTLm3zmiHX4JyWIukDp0vyE09M/rnYLtDup63iDycsGMN7BqqrEE03nDTQuU1Hx1k8UGw7t9Bi+aydkADRdFR4b09sTEb8fKTM5AjwAyuXNOzpEyc6BDlR6st9R4/TUGJGBJCQu0EZ5/OSf0qHC/NkX/Jy2WZVxQ+KnhrJeNOXmxz9jB3jhIRjeAoef9ns/YgXCvnpDzoeuPobfgqcdgmOUyqvPXDzwnhQsyE+Q5Eq7N/c6VGGDhbeiQga/QJI9SAXjrynUKQmLe2DQJqqcJP+vLkg35T7vHAym8uTa196ktTYjozOdOwEkDS0fzzec2wWxdSph3VJmrfVT6QdDGgkAPfELGPbAeiOJHk6oCCDCidqlCtDSnUx+ZbO3zme4q480E1d5J9Q5CUbDYbUk/Rj4ked8fJkdRmlhYviYtrGp8enUQd0VCBvKvtDUAlsZCw+5KtCNJ3NF2roisKaAJWpZH5haE3ayE/kUjs343FfXhSiFiRg5xBs23N62kvHlsmt33Bz/UAo8sKpv0CerG2GPAJBrjfz+WeIJ0rnMEk+VeXWFatC/3IIzakq1z+YU9YqyeuaLCihCTm/VYZPp3xJ+oPKI9+OjgeJ3NABXVNTlwnmGqqHDboeFKZgXiEbdAKBQCAQCAQCgUAgEAgEAoFAIBAIBNsxEjlBt9Oh4myHQOns53sDGi/JzPtKR2l87qirbJogEpXe0eV7w6uwF4SK83sBHDo8Oh1PtKa2z3UD8O8xoR49Ds1ZHPk0pnf6t4nnrxvutypx22bhI1EaM93cH0Bi0qcJgGQoRjIU2zuvnLbi2k3D5um0SVU7U884UycTq1nn720fE88S3/PGqWfWf4amkGeVk2cz+0VAY+9xNPYed8oxf0fTyr2XUIHlPfq3bpC6kvZujik0x1yeUV3Feo1GHZXfOVRHzlOvMe09z961N0Y2vnsBOm9Trh+jaYWhP0UY+hM9buSWU0gX8XnZcRWRcVAh+ZwTHQEsvTP9Gex6H/FuZPiSjt8q3mXf0zpnlfJ5l8uH8jlbjjeGix/B9qn9luZp/qYnozJZyPFS5N33UUa/f8I4ovdJ+mTSephxwNRPNzWSVn63gG7qVFYSGatbOj1pEwEDz0UYeC4qlc8AXw+bp2lDr52c9iT5+GhMlve1uTeAniAGkHrFZXcnUR3B6QOqK2xb0e+yvCjPmDZJhhPE2c9BJit92WTzjjzPXALd1Bh43hYMPG9LJZ+kCeDqRNru9LOsTs6JV9K2zQmF5gTjIdrW+akeBpw88nnG92C3cpfkUQoy3pwyyTOjQ6KOQmMq/ZnvkiXTbN87+fs6G/l4UKaNfTrJM2ozGX1rx1hAv3N5mXx2eYQQw42/kP4l5dM86Thx7tsA7N0rhRMk1Cbw7AZLg28z0Peclz4z9ikOf/EaHP7iNem9gR5vObxL7UIqY6hN4JdJvqX3nLAyLEFB7rP8Q3mbyFNrcySubUzHXFA3aaA1ptAaU67coeUzbc+ixAPenhAukRfpP9Kftc0qxmsyErs6wDxf2EWy0Fx2nP6C90yR+llaAmOwuVWhuVWVj3Fapk7vkUsGE8s31CZTSXriOGlrtm04e92XR36ZQK4Xg/aH+Vbn5Tvf6eKPtf2IvGbnchXw53S5/eWWmSxyDUZ/vrbnEc+y+Vo7yYOVN2a8+bKDkfsGI09ERJcTWUa/Vd7Pf+/BKcuMgaZ2+MW+z+yEqKPK25zKJZ3e35YMJSw/2DtySJsoTfiIyZ+W79TP9F/E8x431+ZsP3Nqwfm1dG7zkDJbWyK0tkTOs8LfHuj8mc4L8w/4dE4beOmcNqFtYfjMsxP9+XOp3eDLAk9Ppv3l1s+3N+lzJ2/u77pyn9DikMzN82nZTJ249/aebK3ydQbSd8lI7LRv6M5LR75Q+0KT+UiAXzhdamwuDoNri0SkJ2Iz3iVzZcv7xHbi7HGVEHuW0wtJ3iZcv9mTlkZckXUkpcHKSVp33ai20X20NkWWZnuSktrrzbzNaZ9W2Sz5g/xPJ8qQqWfglLHTDtSO9eS2oZtGayrQxcj6pKULc65e5lfB9UhdnO8kzTxyA1eWpT+gn5Kmdk+de7xT0NGBNSUKjiY7XjTsvMnYRpXzMh9emxd4IuMjU45DRw25XpiDhPq8AnY8V82huPeK2KPM2BFsO2gAWtf8zTexgtpQWte+WvDPAsuXL8eD69djn7//yHyTssND6ZrKRiAQCOYZdMIwH2WLrBQIBIK5gcjYPz9In2/fmE+bSyDYFtgpZZDZ3J2NrEQGCAQ7BZ78nxfikCVLsHr16vkm5c8Gy5cvx0Nb1uKgS95b6/tHPnAZDt5lqfTRDgAJcSkQCAQCgUAgEAgEAoFAIBAIBAKBQLC9QqP+HXRyJGuHgYS4DMCEvTBHiBv0wlkaasNJxDwzr8igMGFowoUX8+GOlDf3GS++p5/QME5MyBsa+oKGVrAhwPwwFxltjUmFRnZxu08XTcOFDDAhlXRTp6E7vBAvHE26raEH0p8TUoEcYy870u6EeVBpeIfBtREbhs8PFcGFjjBhQ5KRGMlIjOaEQqOjLI+woZtMSAJTnte/fqgzp04MfQDc0FM2I9hRbUJDUT4IhmYoCXPIhorz+dMPd8HxDkF7c7noCYWno+X7ITwoH5mQGE4YIS9EZjkBgccMz9nx4oUu5NJw5XBjuzTsk88PTNgyg2Q4YctnQ6kEaKX5++Fv/BAneSgbRh4FwPE+pdOOLRo+hgEn4xqTCslwgsQPvWgLoBmADRVRFrq4MpwGE7aDC0mnowq9QMJiBEOrkLBhBZq9bEeejDDyZFTMw+tfJxwNpZ8Zm1x4mea4QnNcOeFhKO/o3TrQu3WCbUb7tCx8jJ/Of++UH2pHEtrZ9BEnw4NhZMqQpW14fG5DF1E6zTsm5N7Qs3yf0e/Y8DOUDtKnZSFxnDaj9DFj3wmB6YP2OTMMC2GzfBkXCklH8iwbgz7vOaFwEqYeGR+H9D/N1z5j2omTm6ysDfEUDcvt50/bxGsrlXhjlNqBHr/Zcrz0Tj0JL1mbgoNPBw1jqFEIG1vWZyasUFCuMratihVfvwy6wYf9srIyVqy8aW+OrL1iniW7TiPZdbpy3NMxHLS/QvMJuHRwNk/6QfqjtnN7U4T2phrTOwUMbIgwsCEqyvsMVIb7vJuMdkvHk8PbjNzgeN/nv7JQ6bqRh6Cmzyx9rWJ4WKf6JTZx6Lsq/Uj/5sa40iQMcSDEZRnPOOFrqdyvsN1oXQogMpqrLy3Thk2s4P2qkHQA3FCxxhYiY5jOq4oZlZdfaj8bvUFkFPtdSZ5+qHcjX8t0aqn9wNkz5s9+bA4UbbIyeYoolXWlui5Ev0+XaV+uD/xHAVvFWYeg9Af6pg5C/ezLsGSXGFg6lf68scGlNzxrdQnhjWRAs3MIVp7Q+tO+q2mPV/FJpSzkQtpleTfHc/3K2WSOvGD6NGg7EhlAeaag68rCkfr0E7rLoDQwuC7C4LrAXCgLn1uY/9YEtW2pPrb5RF6bmVCyjJ535tfUJiDy0qHb7wPKryVXHDjfhsC8D4330vkIl7Uvezw70h8vRh/Rd3SNkqXdm4uacpUGe9VLSB6ydeZC2NLrPzoq7186Hsjf5poFE/a50NYeHxbWDhP3x/WlbmrX/vD7lJEBQZ4oWZcRbBvomj/BjgM5QReAvRssE0hOvObQglEJ99PN7e5IxTBhXvt3ogBA98nh4nv6WUJiU3v3twFunSx9Ov/WxB33EQ+SdP5dLX4aP72GvQNNabixpQM0KaJ8nPjWgb99OHeuaWBy98T+7dNYxwnB0NgYSy2c7hDTBj6yT5IsxnUhTyZuuo1VH6CPvfuO8Ob0Am2fVdUrdI8ekPUx04/OP/3JRYB3DDqLkvLxwpXple/fpUd5SSsUF7JC8d4ZqECbcTxnxwu9gyfi0xQz5Mc22x+cXPCNJu+baDwqr4euHkdsXbT3b59mTh4FwPE+LS/O3jtyiQEn4+JBjWi8jLnJn4E292WU87pqDZSRhU5/kz9L9QKxroL9SfI1vG+NZi/rrfsUlZjm+IhOgI3YbOQTdJpGMQtp3eGw3tQKiNa207/N/RN+mbH7vZtB/X+n92sUyHNB6seWWdL+lXyevY89XlfevTJOezL8PLG7Wwl/PAI5/UoH2qxK39F6cncBMe2gyL2pBZA82PtX/Hr6uiU0xnTFe/OZVw9LI7dYSPqjUmfS91w7VcnVgAz289RRMf+QzM/bguhPkr+ZyIdo59qSjgdrU3BweBCAFz2/lp2UIZrk9RYtq6B/G7ownhwkgEqK7618V7y86SxMCs8az7Uq60BpK/22Qi45tjm36WHMPGI7dxbVvGhNA1OLiT3MdK/VS4zciDY2S8eTw9uc/GR43+fB6ZK6qBiYWuLa8yomvNEt3vXtlFVz3bBw/xij/1hdH7KzApvIwfT0FdWVJTQCcO9OLPmOlV8URO5bPquoQshuCspC039kDPv60k1UXn5BlrF6hWvIennSfrD8Nh3gt5o6d8ZpPPg2WZUMorKOTRfSVQxdtecwgTkxZ+9XzSurEBqvvgyLtjSALWWXeLrpjV51eN7wc0exYyEoOwwptO84O57OJ2rappWy0Lxn5srdYV1q5znPuDltwHZkywej6ypkVNVcmINWRH9weXu82WsYTqr/O6NJsf50naGhbR05OyU0nliblWn/0LyqkNb7lgWXfYA+bn2jNGu/X83Y4eaiirm/XLtrlBxNZfpTt4vyiOU1gLd5OCdE0jZJWxd42x8bxlZuZGsXwfJJnqHy0gdFmlRXsfIqz7Q8vf+thKidX9Q+QSfYYbDdnKBbv349dt99dyilcPDBB5d+e9VVV+HYY4/FggULsHjxYpx66qm45ZZbthGlAoFAIBAIBAKBQCAQCAQCgUAgEAgE2w46UbV+gh0H280G3Yc+9CGsW7eu8rtzzjkHq1atwr333ouTTz4Zxx57LG688Ub81V/9Fa677rq5JzSAneV4b2WolwrH3DqhYuwx+xnKil7DDdh0JeEce4HuoQ590cp5+vheYrOEftuyb9Thk3lEnX7lQmzMJWaDZ2dMw1zwSQlPc6FQdlRsszHmeRyzbVZTjnDhqIzs7meM+NBR+uM8OuvkPRPsULwU6i/muQ231Y91NwdtEgp1Vhf99rMTDqeH/KvsgzkdxzXHZSEsaE3MmgxV3t9+nrPRPly+5pkOvJ8JeqB5Lmy/WrbzLEMZG7Ms76r3cwy2rUnf17Xn68qDvmniyqpL01xhLvsuVP85mqNsa8xU71SFk+yp/J2kTVnMk2yZzbG3o89NLErsyfnCXJdv5iD8y9ksiPxdMZ4rw8P2IQ9m3I5VZc7V2hTTFjvNeKvA9lTP0mtwdpL17x0VGoDWNX/zTaygNraLDbqf/exn+PrXv453vOMdpd/99Kc/xSWXXIIlS5bgd7/7Ha677jr8+Mc/xn/8x3+g0Whg1apV2Lhx47YhWiAQCAQCgUAgEAgEAoFAIBAIBAKBYM6hoHW9387r6bPzYd436CYmJvCud70LRxxxBD784Q+XfnvxxRcDAM477zwccsgh9vmKFSvw7ne/Gxs3bsTXvva1WaHLegRk3m7xcO46QC/8NWiMq2qvGO3+yi4h9cFdesqlcTwcVP7jLmr1L4/2ve65y4sbk+5lrOwl5zQNExfcXJDKeWNwnuEhj0R7+iJ0iinkgW3+ZC61Nxf3lo4ML1/dCJRFkDTTX1WbVV7USzxiuW9pW9l2pqeQmMuBAd5Th16EWwXnMlrF804xkfej9Hh82Kvns3/prr2Emb7n0pS8d2hBsc20zxOFxOVlVoLm713+HsxLlXs+1eI5Ajrm/DTOhdibI7Q3V6sXcymzf8lxXmAFTfQ5N26J7GM9CH0eNO3Rwynbut7SnKcmHaOl+ZScIDA0WPoJ3Vw9qGxgZbV356C9X8XIy7hIq0oCfZ6o9EdpNbrInIQJnfYK1I/SVgr/kmuSFzv+GT6pPTYYOebLE06GFS6DD9DM5RmE0RF1TjiEdKSvCzz5Xstb1nxDdaqno6iOCY23oHcz5Xdm7Du8S/rR178OHylAdVX6C/CJybdSP3LtSHmLqzPV74Qf/XZPBvOEjuwg7cDaN8QOdNJwtDMwsrq5lQ5EsPxmyjI2R0HGVfCeitMflcucDHXakdafkVH274CdyPU5Kw8YWRsCPXnM0twDTPvrAQ09YIiC00ZKh+28Qp0ob0U8XX7/OfkwdQjpMrbvGXlQzBC2rU15yXCMZLj62HXlHKlG+gIfMc8iT15Ye5+8d2jK+JT2EzsvDPEJN1+g+duCkY/vmrKehScj2TTWDtT5vUMV+XNy2akzqSc7T2hp6JYuzI0BuPoK9cec6RPOvqjSv0qT8Z64c+Xa8xcj45v530sOXY8lh66vmYFLT2hew9YlAsp4q4znrIw0Ooazx2vomoIu82033xbxs+PozvKiY9fqEiKL7HhU/Dw9VPdoWiGi95+SPwfWR44u5mgz7w0/B22iwBhm6TL2Qw/yLhlMkAwmqW4xMoppp7J25+xcIFx/moa+p+OQs+MK+ceu7CjY1mwiIFkQI1mQ6ZIQz9ZdKWX0vPPOINO9li8yfjb1c/jJjEdm3tWvPK9ct6OPS8aTqWNhbkDTM2uPZQjaHyH5HXm/EM22gPxXqV/7jbbl8Yxvgxdo8f82jxjeCJVTWefQc8tfiqfN2uNk3p49c8ZYLzpOMLvQyAVl5W++iRXUxbxv0H3qU5/CI488gssvvxytViv43cTEBH7+858DAM4888zCe/Ps+uuvnzlRiiwCJOmv/VzDCqB4QCMe0M4ETLdgFxRCRogdH8bwIJebauVuXPlKIWmlP0MfFNxJiTEwqSHlGLNFoqKJqChUNTHSFAqLJEkDiPYZR7TPOFSSXtBKL2lVsbILXCY/55JdupmlUFRk3iKK0rALQs6iECVZeX97yrU5kU+UVFehtf9WtPbfatuULj7Yi789mp3my96P3q8wer+CWjJVSEPphwL0QAI9kF4SbCewpO2ToQTJUBJcgMr/kaexfEQWYsy/k2a+mJM0c1oaUwqNqbIZM5M/IyV8/rZ0kHYwfMBtvqg98jbjJuTQuUGQDGgkA7qwWdPeFKG9KSeO9hOd3LJEJ6qwyVW54K/cBULTvoZ+avBGXSY9NdLMhMwzLPP0jAENlBqMdNHYfh97soPUxYyR0hAf4OUZtwAYdfO6dBYm6CxM+AU88kxHQLwgQbyAr1S8ZDrnE06u0rzoRC5rp2QoKUzIncm/zn+UH5rjCs1x5UwqjVx02mFA85uWhj4jP72/bZtlEzHdAOIFMeIFgUVHwvuK0OyUlY03rfK6OJMb0w4qXwhgQ2CR9qELFqaeoc207pC2l1vbrMyCA9WLkU5/AIafijD8VID5vMUUn/f9tm50FBpE1qhpMpkg7Ze0deFSbSqjaB2dTXjt5kP535loe+1C2y5Z6AoGlSj3Qnjq5EAXfbNfsmSabysC2+ekzTn9BRR5UiX5+9BimH3mLxQEJpe6qfNyqB2U/VpbIn7jyOQTCJFpx/iAhuoCypO5zmKPyuVj0tTp5eiGH8iY01HOH8mS6by9af39ca5VZTuTT90L5kN5IrcHk6UdR16rJLPdaJ4ZfzW3KDS3KGfsUj6Nd+sg3q3j0sJt6AdsGaOL40FmlkcfkfFqbA5jS5TqGlO/aWX1luEdEwpXxZmutfZDOoZ0A4gmFaLMLvYdaJJBjeaEQnNCuQ5ARm62dcGeBYBDj3kchx7zeJpH07WtNJ0DcM5lSf4tXcB12oHa7n56TWxDOgy6Kv2RvBxbxtibgbY29dARsN/yZ7Df8mccO5+ivddWtPfayi+UEpvD0bOU77w8jR2uuorV3047RsDCgzZi4UEbkQxpJEa/ZI4fSgPR3hOI9p5gF6sSf46lsrSWGJQuptl+8p45jicKmU2f60rD87k9Sux9jVxeZfNLwLWhre1rbDfqNGNke8yPJ2dzhLHHHdlK9JZj3yj3R3kj6BxmxqPZmCdzWXbh0LOnuXmRkYFa5fnTuaOaVlDT6RzBwNjOVF7QOjlymtnA5Oy0EHy57awJNHSud2iepM/L9Cadpz133xI8d9+SekQRRFMKEcnHmSv7zkAZH9P3voxy5jB0zFNZm9l3VhfRb+kmlrExySaK0lRvpfxA9b9WZL7A2cA+LTYhCrYHWyezhkM31MkYoXKN/t04cAyNA8fyMeSNMWo7F8ZrN5+jGX6mdogjj7ixR+dwlJ+yfuDmVqENk2giSm0LR79mZSzuOO3j876OSNsGxruVmwR0jDjz62764zbtCrYWPBkMoDERoTHhNbZPElkbUxr52CQyhvJx2SYENwdIiJ209DByhU+WTzSd2hC6CYc/3Y2RrG2apP6mrJjnCfYZ4RF3vCLonAN4fWp0ScanTr6EX+imj+FD2l++zWBsCX9eRdva6Xs6X/B0qdJkjZKxq2k/cWsPVL86c24K6jzAyG9fNtHxlgwmtn/oWimda1mbL+MNFVfINX8ju8QhwDxrTuRjPJ3zazsvBzLHAqK3jG1M+cQ6ARB5IZgf1A1xKdhxMK8bdPfccw8uuugirFq1CitXriz99v7778fU1BSWLl2Kfffdt/D+mGOOsXkKBAKBQCAQCAQCgUAgEAgEAoFAIBDsNNA1f4IdBs3qT+YGSZLgb//2bzE6Oop//Md/rPz+8cdTD1Zucw4ARkZGMDo6ig0bNmDLli3YZZdd+ieOeopk/++OaOv1Zr1MqOec56lZyJI+yz6NppTzXBGP2EJ64tlm0lNvPWfgmb8VeRwx3ylt/03paG5OXVh0S+feLKbMtgaeGk4fNYuehkkrz9N4UTrlaliPPt9726fPtEM8oJG08zrRvFh4deoO6dwbCRrdRxYAACJKG3ENNjSb/iichso+3Xh4ln7tgOOpb8kwbZcAjbHI1sk5kZEhGo+ystKXCQkB6vADrXP2PtL5N+YUgaKOitMkj0ao0cL5I4Lr8Yli01taFfl/VpbxDnOKWTtQWb7heYdfyPvOaOKmaWjoOO8zhxZDuz0Zo6E6uXen/S5QP+6h9dq13n057yNReVvT07gmKzIenX62Y5vvJ5OuvT5lruldPO8v4kFlYE9w6SLvaeX1nSf3Cl5b5r3n+Z4STVzNTDtysszj52iCeCfCLa+xrm29uli5StrMnILSDW37JhqP8rakdSZyy7YZ4bPpBdT9LaMl88Z0ZHbAO5lDIXwwgZpWiMpikiGnmW1TBesxqWLknsqWN/O/oXJ542aSv7c8Q/L3y/P/bbxFHZ0QKN+827p/+nE0rYrtFwGa6NoqGmLfY9STdfaUXNZn9BQd/dsZD3S8eGPH4SHCh9x4NmhsbHo6n3wL5CcRGPoBQG30jsL6bYpyPvHh6/eCDGS8Qx36PfXue6T7eXKCtbOIChOmn5uaTWe+izoqOCb8MjlaGlsaTl5G3qj1aVvrBlzZ7fNpgD4HjA5gaXUeZknXtvNHnB1H/t3dJR+DVq7Febpmlpdu5CfFVKLyUzyhw6yGJ4xObyB4stGWScPcMTT7NoPzd0Nbu8vRlczYMPWIuiicjAWISpqI8hO+lCearp3h0/TAnfunf0RkbDX5dvaRnqjK24ztfyI3rBwg/eGcljThEElZtp3iXIay8p2A6sonf7uXUycf00+OpGVSm4DQGfLq9utH7TCFsE739cfmR0ZTmrs5zdFkrov1H9P5iHH+Tlra5tscjyr1gt8njnzz7F5Dk8nDkbVURphxQserJt9yY4JJz/FU1C2RA8jmDmbOQmwqrn+p3ObsG0qzpcWME27ckjTmND/AnEAwxRIZwH1r+ZSR21FH2fHuNCMJuUXlmc9r6UkM8s6XT77e8PJw7GljI06qwglcwJOrtB4leg1kPlFHl3OgJwsBV8ZwY1RNK3Yuy9HnyAOjS6ZzXWx0ERKvThnsvI5EaaFrAiaygDm5Zsqkc+kQbcUXJk+Szjs57qSPNBKunZh5lW4A8aPpmkKDjvssr8ndXfsmTUweRUSGk/mbqT/lfXo63KYnczj6zHzTGIuKofVD6wCmHGb+HK1r2/fJQOKehsxodtaeOHudnpb0aW5U2ARAbqdQeUBpp6fXmHFYiFIQAdGWZrEcwi+Uj8vgX7cCuGt86+9dmtc/+188kI9xKk+c9UTKr74cUrDrGIZeAAB3ZQQdAy1SJ6bNWRC5xfIPlRvUTuPscLI25qd35jIJCjI2HVukfRr5t6b8OFujpBEquPpRWWLtxC6K7ezB0RseuHlfPJjkJ+VI5ADnxDJpp8ZWV8aF5L+dp3t2mI3soL3/k3KcaDfKLR8ApnZL3HU/Ti8Qm0swv9CVA1iwo2HeNui+9KUv4de//jWuvPJKLFlSHbphbGwMADA8PBz8ZmRkBBs3bqy1Qbd8+XL2+cMPPwyMjtp/c4s8ocWY2uOjwgCui0qhWEFzSOg7ixx+OiKo2TAgjhIuX1hjQSdKjFJxNxgDNHJgJhNuPmTiYI7uh5R0WVsGjPlQWAQfdGOmLn84BhE3EfS/7RWcEdVLvkwaurBVubFRMemyedIwdaE0ZvKnA+X3YGcUJhM+73tlhsDydtl38DbmuPLrTl4VX6zzLc1f5X/mH2f/Y5wIaqGMDyJduhjlkEHK9yePhW89o999yf+7n7FTtchhEZCVfCL+Q25jjP13ZTl5UaGFNbb8ksVn0EkLnRdwYVyN0a9qi80w6DCsO4mgE8U+9E3ZYlqh37lFjDKENnjMgnlcXn4xYc1n5lUVvZzOrsiz8r3fB8wiS1k+ock5CxLiJXc8qEhTh60CoYNKQeVi3XHILXCR9qH2RdkGG9fOoTSscxK3gOb3W0nxIWeGso3MSlkbWpQMpacLQoE8K+WjBrsgzEElCto4o4Q2Spk69KWX6AZn2T09irFzgLwvqSNS7cI9/Vsx9rlNWQqffjre/c25Qpk15VJQpgZ0cuFxL2O4iqSK/lY6QG8P9mWt+SKxB3vKm6QvG8/cIj9F0q6wmfqhqZf3jE1TeedVDbrqll8rD992rpFnaJ7g5ON959jbNcdWXXus4Pw8k3YhyjTohJiBlXtsnoHNoLL0VTZN3XmDlz64YcChZG2LzSfU3+S9zZLZpADgOJ7M2BYum+OEkjA2SdmGNN0wBlAp78rsY1bP+ms3vr6hc6Vepohl33jOJLXRr/7i0pWV78kY1gmQS1ZmT/r2MCMPe5G3vcrmYJj+AO/5jpNBOkJjezb613NgkP2f7RyzZF8Kth/Mywbd448/jvPOOw8nnHACzj777PkgQSAQCAQCgUAgEAgEAoFAIBAIBAKBYIeAnKDb+TAvG3Tvfe970el0cPnll9dOs2BBGkJgfHw8+M3WrVsBoFZ4y9WrV7PPly9fjgfXry9NS8MQUiRDqctBVHExbV+o6anopzEeS/FIRhsJfaMHdB6ijebPeeN7njxAwKOCeOeEPNedI+lefXo5AVAVyosLXeh4gpB6WtoU2OP3NC/leTnRetTxirehTEm4A3u8Pi7yFhdKJeT55YfnpLSX0Zcfuc/zZ8sn9ag6XWIunDfhiBwQ71uuHioBkuwIvuFRJzwJV2bISzn7LhlK7NisPJVQ8V7pPMQsDXVi/o6mi6FuejrJwY13QpO5QJ6GyPDLArw2JeVTr7SqkDYOL5SMOVrnUq9EKnfKyjVl1vREDdbZk2FRNw+NFDrlasaRucgZ8EKtgkln/s94r9mLs03+RB4B1TJPt/LL4uuMx0JbKK+sMlpJ/sGTlB6irmI9RC0dAR7NTynlJzFs28R5iCsudF2/J9xC6bUJBzUVYMiaXp82e19eodh3fsi6OvKB5XMn5F3WZi2dh0rLnjWmVCozaoCVF9SDvcbYM3D0Clc8d8rI8CANrxZqZ84Dnnry1zyFC408PEw/Xqgh1GwzCnOS39EjjK7j+onmq5u5DHP4jZZb4o3u5FnSjgU7pUQX1xmjhXBUlPdIHlW86ZTdi/1s8idykw15WIJkYReNTUwILa64KEAcY5uH5LYN9+lFgeC+LZZP/5F9FxgDts2Z8K4F/VTG+yj/DuBt3zI4HvLcaYEeeIDTy45NYcLherLeH9t+aEEuJJ1bSAWtJe+Tlrb9T/Nnx4n9oLycqMPbdmyoc4Mac0FWBtawQwFv/qiK+VN7nEPUUTaaij/3AKptModWUw9SpjPvJHZzpb1tsubGaI9rcIWrB/zTI34ZIRlZpQv70G/0u1L9TfKgYy84FykZG2xIOQa1xqNp04a283aOH3UrcJ2CLSxMR61vfL2bPaNz+gIfEJuAguP9OusH5ruy0MfcXCxYDzD8r1E5/7c2U8VJVkXGKzgZQduR9tkMbMKQPU3t1MK80OMTS5bpx7Z2IgfNJULj2beJfFneS8h9FlRe0X975XO8E1ozSIa9tVqNfAxHGrFZy+XWrnoB1f8MLXXWCw19/eQ/EzvCP91Z+0SxYNsjNK8OfSvYITAvG3Q33HADRkdH8e53v9t5Pjk5CQB48sknceKJJwIAvv3tb2PPPffE/vun90A88cQTbJ5bt27Fxo0bseuuu87s/jmBQCAQCAQCgUAgEAgEAoFAIBAIBILtCbJzutNh3u6g27hxI26++Wb23eTkpH1nNu0OO+wwDAwMYO3atXjyySexzz77OGnuuusuAMCRRx45c+KM5xvZaW5M5p7n9BJq6mEwY2+LMpLIBekWVbGkiRcXd7FyNKmKHjt+XH1vzFNPSrZIzxtJL55Onz/XsjRRr0vupAelv0AH45XveNiTdA6VtNkyL6t4MO9H6gVk28Sj0SJx3aCqTqTYbzI6zGkQ/+QeLRNwaarr1dnakn7QHc49m2vJbeuSVexbNmZ3hReG0u548Ps5dCE0pUdNuWmTIe1ekl3h9e/n29zcyHm3F+9Rr16m/Gg6+zurZjSd/50M6Fredslwkl8IzHjxIlHumM++od7MM/XCsvwYureNeO1zFzLbfAZ08SQQ06bRhHL42D9FRKEViqcGAv1k26Sr3PHo1d85jaXSC8/T8pVNb08gJCq9NBo1PABDJ+eytM6JHuvJyqf1x0vUUY6u4cqhMobz1HTj8BvP//whe+II5Bn5zveEde6DCXnc+v2oYT3YrfcioSNpaXuHkJpWxT73+YDxrGRPNQTi+zc3pR8P/SmlZeyAxGm/ZJeUQRub0gwcuc95a9O7vzKPzcZY5Ah5O7aTqkGcwj/96Z/mLow17yTN9GjMnvCnOi7OdHa0oVWoi0p6d8ArnI7gPKMbRX40iDpEBtL0ZIz7nsX07yjO3ZS1AtTuqWLRzw7Y8h36eoRKwuM4J8bLP9CIlN+MF29jPCraXIH0zrjP/m6ORdbWibrud5amMvp8z3UvTchLOXQy3jxz5A2KY5U7FUjzcU6FmLxLaHKiMzB8qOL8RI2jNxwvdsb2KxkQaiK3OVSiXB0GFHUBOXFkiqO2f6ncRj7erTxt6MqTIg1zSrmpc1uAnK7gZIvT5oVM4bRZwVbw2t4/TVfoe/PejDHt5hl5cyPntBg3tnsQYM0Jb65g0nt50b4zNFDQU206Iu3X9WijZZDnfjsnns1EdbJzei7J+YDLnz1VT8au1e80UgOlrWS8RVPlc8VCOnOYITA/NCfcFYlkENGII16fNCZUwdaj7+ORxJ7QZ4milz6Rfig77Rc65etHC/HBRmhRxe91S1s7PRlMKtcc/LE7/HSE8T0T55lTD4XyEylE3lWdkAudivPXHFIhpJ1vgNxOU12AkycmwgpdxyjUy0NjSqGb3d/NnTwq2FP2hVs2LS/qqCIfRPnYVklOXzDCgNdPvgyMF6WZRVua+XecPCPP6FytcOopdCdWJi+SdlI8FZgoRKbt27rQFklbY+TxtNDxvUJHr8n/A6fATZ5cm1StCbl1yV5zAXyI/GWjBgVkeRXPs+scRkZPK2g/Eojm8+LKVF2VR0gyLzvK9pNu6KK8CI3BmpEQnPUFIsucuwSj/NtiReBmxpXZo16ma0eUz6ltVZqVAtRUSjTlYRvxqaNyudqHzeASW3zUmFKYXhhnZZkFI6/9/BP2PZy8dlAlF02eXlQjm9ykifM5WPAkZb9tJJgdSPvvdJi7HaUSaK3Z36OPPgoAWLZsmX32vOc9DwAwNDSEk046CQDwve99r5DnNddcAwA47bTTZoHAXKk2OgqNjsLoH/LXg2sjDK6N0BzPDDMFtPbZmk6gMyXpGBOA/Q4qDXuTLOzCgWeA+hOyeMk04iXTbl7UiGjq9NfWaa8aAzH7tjGp0Jh0JerA87bkeZl8GhqDz0YYfDZKla+pU1elBsJ+Ey59piw6qdtrMv0lCq0n2mg90c7r1lVobk1/jQmFpJmGCjR5Jks7he6IppXTfuZbQ1PUJeWbCYT22jFLq1Vep+aEQnNC5eHnsrxbYwqtMZXT1NLQjdwQ0C2d/kybR+kkIfLa17xXOi8z6ipE0+lmjoqVszANpGWZSW0yqJEMZmU305+te2Y4mTAzpv7TixJML0rcTdA4b79kKLGhWO17DZsBpcn8nSzq5u2Ttbnl76z84aciDD/lx3PLecv0s63ncFzYNKAKXmWGlzW+ADS2RrlBaBYUSPrmhMr5QKd8E00rqBhZuDyS16RyeRZw+CAKhLmzNEduOpPW5OnwguE9Qzf5ReNRoZ5Abvia0IZKZyE6szq1N0Vob4oK7VeoUybLVOy2tck/6io0xyI0x9y+o/moaZX+dN7/9Eg9V+dkOEk3Jbhm1Mqh1cgDp8yM96JplY8tI2tiQh/JvzEeoTFODGsj+8YVGuN5+pSG9NeYVGhMRGhMRBh8toHBZxtQMdA+YAztA8YQTal87DHjlbav0/amz7O0aaiPNL1ukr4yvDmt0olBJ22HwnhppgtOjQnlyEDSpLZPVawsT9kyaWifCHyfE/pM+9jxpvLyVcaDdCFQxcCez38Wez7/Wbevl04BS6egmxrRlEoX7QYTJIMJdDMfoyDh3ex40HnfO84MRq4G2twgXjLtPDN1GfxThME/pfqb8qApd/Pzu9j8/K6TBirdmGtsath2sovHGnm4YqRjKspkpOmnaDxKw0t7fWf63PJqJ9c70bSy8to8s44J3jgx+gVEvzWmlEM/VBrSxZE3ZkyYeiig/WQb7SfbDv3mR8ep0W0OvxIZyIXdoDLC0JH2JWlTo1eM3Mrq5fevyT/exdMlJj3p23g4QTycpHy4ZhDRmkFrx1G5GI92rX1GaaI/q/cN/REsb1P9bPiAwtHfVJeZKhmdrvM8u7vEuf1m7M0xImuJPBl4LsLAcxHSMD4xkuEY3YVJ3naJAhKXN52xRWxY2k+W/kz+Dz9vc85Tht+pzZGkC3ZJW1seTha4/WTkPl3ksXZKE2iOKzTHiawhMihe1LX0t/bZmtrfKm+zdBxkMtDKWlde5UyZ/nRLQ3XhOGSkC63E5iFt7vO7o3+zPBvjEXRb24U5bmwYPh9Yz0/HTPvptrbyorUlQmtL5Iw92k/dhTG6C+NCaGDaFsYuTBrZr+XpUKQ0GnnN6c+RP+Y2melvX1e2NkdobXbbwfRN1FVF55EG4YNWPvcwcltl5Y780TgdZhuXGW+3tuR5tjdy3ixe39PnCk4/d0c0uiNkrGftZOcw2bikuojjA9VVth2aZOxaewp5m6kkb3Nqu9m5gQb0QAI9kA84s5Gruvl40hEQL+wiXpiOlXgkSa86ILI0GdCpY5XZ3NZ5H+gG6bumtnW1+sPvM9qmUe7A5sOmj7z2yuq68JEICx9JO8bUWenUsSsiNn5zXCEeShAPJekCI2lLINMrRNfS+ZjVM6bOpC50vPn2NK2nDe9K3je3kr6lfJDxpo7yv+k3VtaQ8WwcjHQEO94bxLnGGacB+Lbp2LK8UzhdG03l83MK085U3lFdRWF41+gipx32mizQGI/kwpjmNbAuwsC6yJFhWuV83NgapXNCynuMbQzkdlY8kNuh3Bwkl68K7Q0R2hvMrjGcfnbsoq7C3i/4E/Z+wZ/sd9GkQtJKZcXK41eTds7lrh2nZDxa2jXQ3hyhvTmVr40NLTQ2tIo2EFx5aecIpO+dMW8bBAXepusUA8/ljWr6Od4lRjyoEQ+69o9BcyzC2CHTGDtkuhBiF0jrauhz1slsPbQzNvTiDvTiTi6r2trhZ6vrjf5ua0eXmPUTyg/O2M9+B597Ow4+93aHViP3QGwqlZBxwIH0qa0TldFkPubI0kwvRsT2b+w9jsbe6XU+Rh60Nyl2Drpy5b1YufJea583txZ1qqHP8MDQAVswdMAWthqUH834Tm0W88Bdj7TjyJO/QOaYGOn0x+hXAI5NrHRqh9FvzTgy626IinIXyNdqVVyUAbRMuvZjdSLI2KR6Oynm5dRZ8/Xx5yDxaL7eG03law5mbYLKU2pL7fJohF0ejSz/J54dybU5QOyrbIz4DtFm3BrbPZrI5+X2u4jYPAO5/efwePYsWRBbe1gwT7AMVPET7DCYlw26fnHuuecCAD7zmc/gwQcftM9vvfVWfOUrX8Ho6Cje/va3zxd5AoFAIBAIBAKBQCAQCAQCgUAgEAgEsw6t6/0EOw6U1ttPlz322GM48MADsWzZMjz00EPsN+eccw4uueQSDA8P4+Uvfzk6nQ5uvPFGaK1xzTXX4LWvfe2MaFi+fDkefG499vn7jziedclQknrAI39GPTiiLgnnUdGi9DLSYBgJPx+mTP8SbgN2k5xJb7w2nM90oHyGfuvZ7udb9Z7U0w9D4IQhMjAe/iU0lcJ4SnLPDUry5drJ0oUs7GfVpcRMGAI29BPn8EvK59LyBYKtU+30BNF0jVA1XJhCjmczNDrKhq/rG37+fp3L3ivS5n04lQR5Yi7zIfTbUBeBUCVOGoNQc9N2KumzXlDKZ14/9co7PZdvxlQo/JxJQ8YgvWB9Jm3h0BGS4eZZVTkVbeLI1aqLuWu2L0t/IJ0JlUbHNQ0/RmW8pW2W+C3N2M0r6vJymeU3omP6lQtlPF+pn+2HOR3mxHJK89yYaT7NBf3v949nc9Rqo5D+rft+FuCEsKantYA0zJHlTZWf5vTtsWDm5O8+6hFqR8szSR7aqCokEdULnE1Rpf9Z24TaH+bkZ7daf/uexYVwzVVymZMNhA5DkwlTl5joEaE8A/1U2SYltlklerEzQjYbjQJhThcQfuBkCxeyvVCW+dPI44Cu4+S6BTMH8cvLvbX59wXSvPHg6GKkNLL2T91xOBvyxp83MWHkfDjhwalNihJ6Au/r2vEznS/QvvPDmgZlRJaPOf1YWS7csh0ZVmFHhkLR2bz7mNf10rZ+nUPpaJ515cmszDFq2r55ocj7hDmBUoceG4KSCZfr2GSMLC/QAhR5fybymOYTsm19hNYOyHMTncgJhWs+JTZFL/1ZFvo/iP8/e+8ebFlR3Y9/ep9z750XwzCAryCDvERGJWJCRFDRQmMJqAhqEq3C+I2J8WuCksS8v6lfvvFbRkRB/eYbq4wmlo9YgkYBja9gNAoxiEqcIC91UKPCAPOe+zhn9++Ps7v36u7Vj33OuXPvwPpU3Zoze+/uXt29evXq7rVWc/st9RhtxYwdGomD5U1Pd03uueTWMMqb12NlTgF+6FzAbftieQRGn6ZtxvWDn09uXvDmd1157cSMXQ7JOT0Cp2+aMZ3b95oGYjqp/4y2mbMHyWba/BNbJ0+yLs3MOzRMcQ7maoSqwFs6SVJuvIKRq2PoST9+y+U46cgjsW3btvEIFXTG1q1bcefOHXj0X/5e0fc/+V9X4KRNR0kfHQI4pDzoAODKK6/E+9//fjzhCU/A5z//edx4440499xz8eUvf3niwzmBQCAQCAQCgUAgEAgEAoFAIBAIBIJVhdLwlhLm8pBCP//JwcNxxx2HEoe+V73qVXjVq161fIRQbwZjje/dy+CjNvHoS7KPWHXohNUGa4Wbsh4MMk+U5z/L1MPSH7EOduqXKleHdWYtvUqt/2KIWCFF6+lZHUbb1lgKFlgROXdAce1urFMZy6DYbwd+nSIkjSObc95zQMTKL5FsYu85Ln+fd/333rvObUHyp3cvON4nHfPs9D21dspZJjNpir4Z12KsS1rvO9ayfkL2iMnJZBpqoZ/xIh6LDi6vLrItJ5epXM2Z3pTOVR36ZDing+8c2cjI+Em8WAN49MXkMiurusylseIT6Yrz9GXUGJ5zXazdA2+XnAwdp51yVVh+I1x3PHhzpPOORjWYVv1yySPl2OcZ3ZKmN3e/mLssgm8zci2qE3rpi+Zvv319Xs7JPq4Ihrede9ZSeY6pE03kqdFFJ4jRR+Ul821KnywpK8cTyb4ukAfWSr1UZfFlEkMf6znHrZu49h9jvAb5eHkUeRZ11JFT78dty+JyfA8DT/dNlWM9W5fUWOtS5x65jGdT7n2y6JzczYCrs71/KPVtoTyZik7Ude1M9z7GpMXUj85FlhxOH4zlbbYZ9Hjtl6VzkvWG9/1wbaj72k/HjH7QyXPOFsY8GtPzO5BxmT2MonVkhzHp35+6XPvKirlvnra9o1PldAV/7vLu/g4TxDJKP6drPSdyVyGrjbP/Eh3Hy4zcfNM+JD9zYy613+jlBXTc28kUXS0WzIvm2wPTEXbZ8QqmLQ5eFwumgCBiieCQxyHnQXdQoFC+OUOTMZeJTit///JZU16yGDJgewsKPU8R8S8ILhngylPC6OXW5lJ0Jy+vrkq3Fzaz5XVpl3H6SSG4ONvPSw1Gf53K70JL4luun4vQKGfcJcIH4+JW9rLcFE/ROpa23YSHDiasg6GtM4IFC9xL4+nvgkPNESFj0IFR6AMT/mBiTMq7De/ROpuLrdksPLKT3+py+nRft+FIErznXCpPYC7upqiWVCfZGB0DqYN1la+nM6ZL5AOXV0XSdpSd2fqryPxH60fysodIvXh/xNKXIil7mDasllTQ32PNiwph++bq0UUGMt86bUrmEJ/+ZFv7tDD0j4vYhepF6BJisuB9jo7UBfBBvn7ZVI8Yt74eKi4cTkx/YMYTlR1G96H0sbybQUmagJaYjBgz/1i6WJuP2xdjyfUG/QNdhGzkMW2zQqbqwns2/8wcFSunFKl1QSqNRURuU54O0kxp06IoRCSFR2vQThPK0060EBqKx7u32atN+LSDsJYohe6Bn6saBOsgP/1cpE5TrCddF3febVmG3Rkqe6e5ocfpTOxc1EGu23yYfQ9VY7w2LYTS5XNVb15Z3dFJ32X+X46xZXTOGt3zj9Q7WSc6v3NlqXS7dOVHP59O9TS8o5Dlo+wapRAxXmJ1CW5vSTF1Hnf+nRDjzOVd8k6VOY25tBooVE0YeBpCm4ahToZPngS14nmOW9cNVLkxNk1XuoZh0nA0ZfNaRXrBwxa68E9wyGBVedAJBAKBQCAQCAQCgUAgEAgEAoFAIBAIPEjoyoccxIOOg26tJ4ylRUW8b6inhkG9pi625NAzOulhw1oqMKffTgiACoFVhh8uwA8ZQC2+HPf9hFVh3UfrpaVaukx6NchYolRe/X3LIM4SclJrOZJnzMrKsaTpj/7sO996hPMuKLROUBlLhqzVHmNZ5XjfEELtz0m8FwCH1ztZ5STqGYQX4Szq/LbIWTNmPCVoqE6t0vUo8ppZUqM/0ybE67LU6qnI45ahqb+7h/5uM8iQlDusV6VHg5V3niVoF9A6p/jYt4g1HrVOXrRvCscWzYd69gZ0RkLHDNZpDNZpxzJbm/ByBeVzIVCsN1Nk3Jh2GsnOhLWc0m2bMu0aWCZzaNKOeDUsK+cVkrSwq4HZLXsxu2WvW+RcjXqudixQ6xndjkVCU4CKtA/jcZsbo9QrUw2JZxxTptItXbQfS0O2a2qJy8h42s8Utk+b7ytvDPm6QrWoWH50PK9T80smbJJjOc7IaGecKi9dpMxqoLLtmJTFGXlUIkOpTmJlAyOjtCLf5sY9mVvZ1zneyemLRgb1dCDDY3WuFhQqEymB6gSGJqLbmHYomYt9moYbCSOZNAmdalS2dviUk/v2WyKDTT6UN515/Ygl1EcsNf9Roz9af5OnCp859HbQj2zfZtIM1up8P9tMI49NmxHejM2vVOcr1UfpHNUVgTxg6ml0Cir3Kf05fdLyCamTExHEW0PEwuEV660d1B9HftPnxqvPjOHK060iHhFB/jndjdMpqIxn+oS2T269Yb3KuXkn0079fePpkVw7cVDDDB9VXsP4Xo0L/Fyak+sp+LqtXRd7Xp5FIP1i9iE6g1urEtqCz7Ub4aa0DahcN+PcmYvMd0QPzK51TT9Qr78aqGdGf0aPK0F2nNE5Ro/m0dTYo3kN12kMDh9gcLgbdsf0fb22YG+I4UNuruOJaenm1k31TJh/1puLeeakYWQX3VuJrv2aNmHDb/fcPIMoMYwcc3UXBd3TNrxhss/NGmAI9Pcq9Pe6mdNoGhwfR5Eab5GoQYbPKP/R9VlsjhllCivX6NjlQMeoPnIR+sjFsjoZkmnf0Kg3RF6y7c3xfqlOxCDgnSYvKiNz86bhEytL4MnoCehzSGPapVpqy8rtc1TD0V+UpgidRVeOeWPYWZebTyLr1iCrKQVzEkyAuvBvijhw4AD+1//6Xzj55JOxZs0aPOYxj8GrX/1q/PjHP+6c14MPPohLL70UW7ZswdzcHLZs2YI3vOEN2LlzZzTNcDjEO97xDjzpSU/C2rVrcfTRR+NlL3sZbrvttmRZ1157LZ71rGdh48aN2LhxI8455xxcf/31nWlebogHXQRGIJlYy6xyTBboaqCK7y1SZgKNLPBZYWieUUG/SMrMbNp2OlwnZRmF2twTBY32wCYyGZuDCl3BhowbrG0X5g5NmvzGKK2tk2r/NZt0NLa0WaD7By9+WzkTDKmTyctpx8hkR++l8+9NKmpnk35AyKML5czigd6P1RYW5mOUjWrJzdsemJo7JSKHnYrJkzuwceqp3Palz4Pfpk8ii1VafhDaZKiSscW1cou0NBmllChAuq8Bf0wz/BgSaDIPx4Gq2wzYzXjNPBso9rDetgPZOKO8PWwWLmpRJft01PfxNqsGqs2fUdAcWlQ75mwMeyoDa4YPIjKuzZynjxoLcMqfed9bVJYWOkaoom3HKxfWcLFlmlq1/Wl4vp7RRQtkky54RWQFvafEhiBseFDPugtL7qCvXjPKrJoPK0JllBoy6Xtk7PfAChyHPn/sxBb05PnCPRuC94bW3oKyd9TZhSeVm3SDuCm7v0dhsL79yO/H3JxWDdpvdE/bfC0f0DMGMrbqOb6dzcLXN3QBXD7k+MH285xun2vybQPnbgcy79n3s5qd681m6GCddmUot6DjZDSRN/39bV4+nLozcxE9OHXa3rweqlG4NJAxWqtwM9UnW7ffBvJCG9nrlUXEsxPWsekL7h6NatDyaalBEKdzTAM09I6pjAnTTOU0lRc1adt2w5Tob0Q340L4mDZTQ2Xbhx7WmlJ7u3uuQVdAfPuxLYfjZZqebBZoWqYni0z9DKqdM+Q/RIb6JNH5n5E3ms4VmXnD6ZuM7mYqoGf4sWvlzrohqn1MLC2Gt4LNPrhy2xzCRg8FSH8a/UMtqaCsakm1PMXoIdw9TpQfjeHDiJZw3VMNyHiOjKFBcxhcLVb2vRNO3NPdzYGIXxY3/7Igctm0TbWgnDqZ30YP6x2osve8OLqVt4bwYduayjqOD5gxmA2RT9uErlmZ/B2aSfuZz6yeNKud/gfCuchmQ8ZVG7qVzOuUflI81clGD2DlWXLzuvnN3YXolO/XL6ZPN99Vg1ZG0rWqM8Y9uel8G5EbXJtZ3ZDqjpyM9FB588VwXd3ea0R1ASoPGP2G030onY7cHHjfOnoQ6Wczhn2Z6OtpBLpP5jIz5/t3tCbGiT5qEdV9s27+OiyrniEV5Gijj5YUeoOeUyc1UKgbPaK/q8fOdTb9kKwByWEzbQdzQFQz9zzT/QHLw3R+HIT8oat2T4u7FxA9BDJKweU9bi1rZHV/Ty8wYnf2kxZDPa53QLV3+dVAtb9q28KkJ33m06/72r330ONtFlphaWNt6TeoiQG51VfrVh/n1gCGrmhRMzrQ51Xd0k/3N+xXylOzfQOACk4/6SSftWnUA7MBfVGdxpcxOnNvPJWvni5g03TQk4MD44Fy914auikfp/YojbEQ4NWDjB12v20MOHMK0dFrTz8L1vzN78EGss/D6QxEx2b3NxJrHXroWw1UKIapWOjBrpe5+cn/XnCQQdY2Rd9OAfPz83jOc56Dm266CY9+9KPxohe9CD/4wQ/w/ve/H9dddx1uuukmHH/88UV57dixA2eeeSbuuusuHH/88Xjxi1+Mbdu24aqrrsJnPvMZ3Hjjjdi8ebOTpq5rvPSlL8UnPvEJbNq0Ceeddx527NiBq6++Gtdffz1uuOEGnHHGGUFZV155Jd74xjei3+/j3HPPxdzcHD73uc/h/PPPx7ve9S68/vWvn0r7TAPiQScQCAQCgUAgEAgEAoFAIBAIBAKBQLCa4btNxv6mhL/6q7/CTTfdhDPPPBN33HEHPvrRj+Lf//3fccUVV+C+++7Dq1/96uK83vCGN+Cuu+7CS17yEtx+++346Ec/iu985zv4nd/5Hdxxxx247LLLgjTve9/78IlPfAInnXQSvvvd7+Lqq6/Gl770JXzsYx/D/v378YpXvAKDgevVfvvtt+P3f//3MTc3hy9/+cv4zGc+g3/6p3/Ct771LRx55JF44xvfiLvuumvitpkWlNZazr0Jtm7dijvvvx8/90d/EP0mZkFw0NHREmXq6ZejzJWgSSAYAwddDsjYWFZwHgKOV8cKyvuxvaEPJlKW4auQdzlPxYc7lqVNVrjvO9cp4+GwGvBw5d3VIIsFLR6ufGgg/BhiUl3lod6mKQ+9aJqHeJuUYlJ5U5p+qnKNeDmNgy60HNJ8Mg09cQzdbZzxuOwYw8t20jyzyQ9l3lqlOKTatIB/fvyWy3HSkUdi27ZtB4UkQXNm8cD9+Lk/jZ9ZUPz4zZfjpM2T9dHi4iIe8YhHYNeuXbjlllvwlKc8xXl/2mmn4dZbb8XNN9+Mpz71qcm8fvKTn+CYY45Bv9/HPffcg0c+8pH23cLCAh772MfigQcewH//93/jEY94hH136qmn4rbbbsMnPvEJvPjFL3byfNGLXoRPfepTuPrqq3HRRRfZ56973evw//7f/8Oll16KK6+80knzjne8A5dddhle//rX413velfHFlkerKYpSSAQCAQCgUAgEAgEAoFAIBAIBAKBQOBDF/5NAV/96lexa9cunHDCCcHhHABcfPHFAEZ3veXwz//8z6jrGs94xjOcwzkAmJubwwUXXIDhcIhPf/rT9vn3v/993HbbbVi7di3OO++84vLNPXPm/bg0HyzIAV0h/Ps6ovfElVhCNN9NHLN3wvSxS+Z1X4fxw+0HyNJPLx4NLoLWzDP6OhL3n4V36XeKDvpMz+nRXUCZ9G2iSF6mHirTZhFapnURrRPzm96RN6OD+826XPxN0xTR4NclVb8p1DuoS6osP172cl9q26F+ugfnThUg0k9k7OhZ7cabH7c9yXiOlkuRGjNe7PlSPkt9m5VHsTzJ2GRh6kHqTy9JphdGG3k/zthxaDLpaTtlLrM2mLaFHcdzKRSNM04BzPCWudTcuWyeaadxEJO39Zoa9Zo6ef9P8WXnDFJtGzxP5Bmlgf0YyXameeb4ODt2EvnG8qPfhR+guG056L5Olq/9/L2yHN4bY15w2nRac7pTQPioC+86dcrwCduOGZ0sqtv477u2C9Ezg3sVI/p3krdzYyeWnpnzUjpsNP/cZ7H8Sudn5f0ubG8jD8fJ09wx3RWlPGHlkOL1WYDoQk0/lcynneksfF6W6Xg6zaTI8WyuTnZeY2Qpl24s7w4AaqDa+6ASfFyvrbPjYdz1TvGaZ5w8x5gjdDUlD59cuR3WUM7rwrnE+T/VwQsRzOUdUeyJtqTQ36fs3b7JbzVsXYyMUmSt5t87xqbP0Zyan0j2LJ+M01ZddWOTxvzF7rjvoFv6ZeX4q0iXYZBcC/lt3pX3aN/MMcTQ/CvSPtxaKpZ/6d5LaZ4lxZbqjonnqby59DGdgdULcntP/h/FMuyKU52kpC3qtTXqtXV+nR4Zb36bjaWr0Dk5w5OT7IsIDj18+9vfBgCcfvrp7Hvz/NZbb12WvEyaJz7xiZiZmSlKs3PnTtxzzz0AwB4qPvaxj8VRRx2F7du3Y/fu3Vm6Dwb6K03AoQJH6TG/vcWOWVT4l98m85xEsKl0enp5MEebGraXrdILYu1Fuz2EF2t7myVBmUPm0ncfTUKltXtJOaGjCJlvOfp0BaBpC+7ScD4j7xvl/Qu4lxNzWXj1DH6bR6a/yCXPuX52LsMmfMptVETpM3wwCBWJokWN6T9K5zAxHnKLkr5Gb35UmegFtyp8xtLK8dYU9nBs+5r61eHF19k8YgfV3NjSbf/aRV+ONxQZ21TeRGRYtq87bmArf+wA7kX3OT6oMmOL1N+p5xj16DWXktez2qZXw3b3KXdxNDteiNykMtSAvcya6VOuHcd9rzTshc8cdF87cgBoDygNrZ0PDJk0dPFB33HtFx3b9FOv/XVFeIK0qelnXblpzHi2iz1/fucu6eboYNrWjnN/LJCL2334oVZNPikZB7QX3KuBcsdEk2eMFvue9LOlMzMXsM9NnYbu2PTz8o2f/DKh+fax3w1UXDY06Z26eG3u8B7ReUz6XH9zl7YDYGVsrh2d90ZGDsn3kbJsuip87+gEnB5Y821hQelPzekUZCOl04xI+3/IvCIbZUGb5GRgwdjJvbd1KtzEUUMFVDrIOzlXpMrPleeN95J5QS01H1dA1Ry41cb4J6OvDtfwxGXlBdWHPTlAZRCV4WpBuWktEU0Gdg1DeHucuYrSWSjvgnKYud753oxDI6uXFNtmsXZMtW81UKzOTPUXm948K5hfq+Y6D8tbZK3nyFBGBpWsYUw/18whLIWhvbe/4tdV5rsh8nogp5Ml8nQQ2Yy27UP1h7pdK3SSA1bejD6sltq5roi3E3zIpo99Z37XiOsq/jPSDlYno99l1hLOnEbnuoweltyfyMl6Q9qchi7cHdOKrMcobc2japBh/YjccNZGKXlKv2X0LG5edPZpKIjOwoV+TMo4IsujfZuSi94zfxwG64cUzw3HG8cB72gvn27LezdvclDL9g1dF3NrWYZ3qY7Eyl2kx4ODSN9w6WmbBLpnph9zCPo5sUfojD2Q9yndXbd6LKeboc68L0CwFu23QlRXbb+pZl7Qfe20fzVf2d9JMGu1cejl4KwLKr6fg/XKFMoVjIcuh6R33303tm7dyr4rCX1pDrqOOeYY9r15vn379mXJa5I0RxxxBNavXx9Nt2PHDmzfvh1PetKTsrQvN+SATiAQCAQCgUAgEAgEAoFAIBAIBAKBYDXjIF5kuHfvXgDAunXr2PfmAGzPnj3LktdypOlK98GAHNBxMNZPGtaiqF43RLW3MTdkrIOqJZW1/LPfLrZWE6w1HWdxwlm3MFYyALHaIJZnnKcGtbJ0rN2IFVV2zHtW4rqnHStr1uvEWKVUCC1+B6Fll+6Bt8jKWKnS7xzrXN9rIZaeWPZyVvImvdP3EcslazHb16wFoO0/zrMllmfO4raL5fUg7GiTb29RYTibzszyF80nYWXf368wWBd/ryJWwLY8xrIs8Ej13/fy1pc50PYN+ryj95xDYwEcCzZjeb2Yz4Dlc+pF44xN7z2lUaN8zJXQg1GdOO9NxysiNU45y2f/fYpmUn9Lh4a1ePY9TQzNHFiLXcai1rEo5qytGTo5byr/veN14HnKBFaJiTZ3vGiJfHY8VhgaZnePKrO4sa1MZbweZlpvaTP/1dCut7WfZ8TaOWZZ7fNJtURkiNPPzXdUrleAbjKOeVbn5oukpWpkLgl4tkI79qrWOtS0U91H6F3t0UT7j7V2z3j/WC9uakFJ3lnrT8Yy2wHxyodn9RrQE+N5YHyrTJKn8cKtCR9aD7WBq3+Y+tWZOc+C0Q1GGZBPMlb9nIeB08++XPZ5n5En1Aq4bcuwTjl+cLxvm/Jn9lZYPLyO1ol+a9474xHgowXQOrG6o5s3APQOjAoYrtFOnoEVc0aGAhGdivHCLfXE0D1dHP4vNg4cj1arlDHzEglvlrWmJW1i1yM9zfJ8r3nP6YB0bNO2871SAaBaaunkPKtinkFURo8ewPU88zxezTdOWgoqYzOIjltfhhXkUzfzjhPyl1j/pzxCuuhbdIzF1gPUa2P0Ia/f2L4bhLREvWjN+zltvR5L6LdzDdFNWJqYsek/p/nZbzN85ueZWyNEPdA4nunxfZL12LXqSzNG6Zq+ZF705F3Um82A6HlUD2I9tGkxhCYz11rvWjoPDzydzHvv5BXjfSo7GOTaxZ/LdAVHtthn/hjxaXP0i9G/QyYcXbUENlIEhQqbyvFuDb7zeZ2TkbSe/vexsUvkgW1/wg/sPhGtBxethaMzAoePGtlh5qQSL+WYLpFCvWGUqNrba8cG0SMcnarj/gH1VFRLjMex307cOGP4nbaznb/nyBrLeDtHZEy/0ZkGa+mkQPJnZBQ3xmuyh0m9Fx103CsI+tnTuag8ckDWKspvM/q9avk/Fk2kayQkB4qpMm3mGW0955w9JIYP6vUNb+7rdYtoRWgBwnUnfWfKo2skIBzDmmnLwCN6gmYTTACffzLfnnDCCUWecoKVhRzQcdBEoDUCvLernxTaXQR6PWcWIOCVTG6RzoTCogdDnDLrbKyQ9MlDHu1+62/kKm+i4zYz7eQXWaxwoS/sdzO63dhryq6WFK+w0LrS0A5+GCbttluvUfyGph/8cA/+ZpiGo/iaNqjNee2sZhf8nMJCDyC5jRVax9wmEKuAk7VPTTZGcuH5kOCT4Zx2DxB8eojCxm6q0s2B5t+lwzQbCs6nBwAflotZHFN+qkjYMz+M6+gDnndT+dNv2feRTQ4OdgGwBF4GxDYSDU9x9xp4mzAAnDA/0c0Ichjl90lw0Nv8rgatgu7IE0/hrfvuYgHw5JIG1EJly/fL6c0ru9CPHUhzC+42tC+/kKwIT1GeAUYbkeZ+v8rbvAfg3B0TPaTw21zD5WmThTns8GVQrP8ToBuAPs/Tg51RP3HM3f6sIqHwgNH8xfHf0mGhEKxN3w3b/Gn4NHsvAzkY4RA7NE21Uz2jHblj5gVlNop75LCmjiz6KJiNpBhPA/wi06GZA6FDM/0YGHAwi1duQ4WW7cgob67z52xfnvkHH344TC6EpVM8M15j/BSl2UO1qOyGfaxMu8lDwwiZtvROXupZlz7dA8tfNn9m8UtDnSHDG7HNwLbNFZRPQESumfzVEI6e0T5vxlCVmUsjcs3I4npW59vE0zl8PSk33kvDgg/XknoOyKdem/f2KwzWhxnYQ4hFwtD0p9FtuAOqyFzEjcF6TY3egSpI19IRhhZ23lcA1wCl4Swp7FzQA6rm93Am1ImUbvucGguY8dbbX7VrkMjYU15oxJGRmrJ0RDebLa1mgm7LoeHbzQGi2fgP1kh+uE4/dDEnr8z8HVYpijbMnGLvl3P0IL+tNE+HlfsazqZxzmjBkV1N+vZl+F4NycEK4XMa1pIL/euEyvQOatUBxZbvzB9MyFwa2j7Fz/SwR9H2o3KD6AfcAV40RFwB1EABhudMPYd0LgHPQOZZTG/UzG8/bSkSB9GxA8bsNRPEsAUI240eqATkUN050vbBRrA7PYfItRlBYBxgCXO/q5Zafb+ea+c6Spvd0J4lhiHMAdfSYXV43YQ3Hvx+qOhhjifPgNEYzeqZGSTDq7J82w74erY1xLZ9Q8I5R8NJ030Uur/SlE3TmLUDYrofNzaYNVY0JJ8haW94qmrknr/3kwTlLVMOCW040skMcYROxsCENeaIGKnT+S0IC87pdjVa42inH8I1cQysQTrd56Jtn1ij+Hn4qAaM8btq5YgaKFZ39vP21x2p+pXoUalQ707+zHiq5lXw3LmOiBqf7Rk9dPYIvbk8KDNSbgDt1sPXk/SMty5LtXMHoyfBMqGrbjABNmzYAADYv38/+37fvn0AgMMOO2xZ8lqONF3pPhjoqJIKBAKBQCAQCAQCgUAgEAgEAoFAIBAIDiaMh2/ubxo49thjAQA/+tGP2Pfm+ZYtW5Ylr0nSPPjgg/YgbhK6DwbEgy6CwNou4yHXyfrOC+sYKzOXvxNSkyGP81oLnjNWhRS+Ra9vrZWkOfaOs9wyoBZDxosjYxUHePXj+oo8Mp5z0e99b7GIq3fSA4yUmbN89unzy8k+o/kwllmxcBNuJmFe1FowaVEUscRkf3u05egB+P7keMKxvuuHY8N3xx+LdxPlO+2QaW/OI9bPi0VqguV4KBfOTsEJv1Nqge+EKqN84suPjIW8X74PGyYnQltMxsVkrH3NWLhTy3HHk0DH0+RknFtoPE3A4ynL6wi4MRWVVRlFja1rynsTkbbg6kytLhfKGI7lk5x1uf+5sUhmwhAFNJbSlPltn3XQD7i2KpJHfhquTF9WpzyjI+WydU3VPxKGLMdP7Yfp+udCUdIxHPwGXPlH579Cy+loGxXO6dDpMlh9JtYejNxzf5fJ+hg/Z+dtWqZPdofFoVbp77uMEfOc854DqDeZpxcaB4GM51ySJur9MF+lLbcj3nOlcikng5xvCW/beY/RiWg/ON7UXjikEppsWhqCt6BuLP8T3kiFX3e8qGP5x8ZkMlEin15YZtb7I0KHI19Tct1Pl5JdVJ6S+ZkNYZrQ7Zz3nIwvkIFc36ZC28e+i3oRZTwIknXKlU95nxkjsbKT/RejfUoba1kZEaOZ0Rliewq+F1GUllKdyF97d1xrJvOKfVYRuR/Jk3r2sHLQeMckvKJjdMTkKh2jxfNCpJ2L+99Uk8g19kqMkihSMTmR+bYE0f20wj5n03p6SDJ9woML8PRUMkbYSAiJtVws/+y35p2/n2XL7zCXl86psbZP6bv+Po6v32mP/wrm0uhabUz5n4ymlqln9NoBTo/NtVlM983pH+Z5Yr4Krk9J9bl4z608ytSmqeC0004DANxyyy3se/P8yU9+8rLkZdJ85zvfwdLSEmZmZrJpNm3ahGOPPRb33HMPvvnNb+Lss8920vzwhz/Ejh07sGXLFmzcuDFL98GAeNAJBAKBQCAQCAQCgUAgEAgEAoFAIBCsYpgw37m/aeCss87C4Ycfjrvvvhvf+ta3gvdXX301AOCCCy7I5vX85z8fVVXhK1/5Cu69917n3cLCAq699lr0ej284AUvsM8f97jH4QlPeAIOHDiA66+/vrj88847z3k/Ls0HC3JAdwhjWoNt2dDRIyFIPkWBMs28ljNPgaAInCXpIYppjaNDejw+RPpSMAU8hMb2oYiDrSscsjLrUMO4Y2pK4zHXz6uWD5j658bIVMfQJO3/cJWjq2AO8e8rT77vmO+qHSuChyWSPPkQlEHTCpU2ygxR+SDjXPCwxiqYxwWHADRGbpJFf5MXNzs7i9e//vUAgP/5P/+nEzLy7W9/O2699VY861nPwlOf+lT7/N3vfjdOOeUU/PEf/7GT16Mf/Wj86q/+KhYXF/G6170Og0F7kfib3vQm3HfffXjlK1+JRzziEU66yy67zH5DD/Y+/vGP41Of+hROPPFEvOhFL3LSXHrppej1evjbv/1b3HTTTfb5nXfeiTe/+c3o9/u49NJLx22WqWNVHNDdd999+P3f/308/vGPx9q1a7F582acfvrp+IM/+AP2+2uvvRbPetazsHHjRmzcuBHnnHMOe4r6UEfXkFwHHRMKAiNPVitWO32ChzA0DqpL+3JiWuPokB6PD5G+FEwBD6GxfShiOeRIKs9DVmYdahh3TE1pPHYKU7uawNQ/N0amOoYmaf+HqxxdBXNINtTXmPxxSOt5gockkjz5UJRB0xx/EVkl41zwsMcqmMcFhwh04d+U8Gd/9mf4pV/6JXzta1/DSSedhJe//OV42tOeht/7vd/D0Ucfjfe9733O9zt27MDtt9+On/zkJ0FeV155JU444QRcc801OOWUU/Arv/IreNKTnoR3vvOdOOmkk/D2t789SPPqV78aF154Ie68806ccsopeOlLX4pnP/vZuPjii7F27Vp88IMfRL/v3uL2+Mc/HpdffjkWFhbwjGc8Ay94wQvw4he/GKeddhruv/9+vP3tb8eJJ544vUaaECt+QPeNb3wDT3jCE3DFFVdgZmYGL3rRi/C0pz0NDzzwAN7xjncE31955ZV44QtfiK997Ws466yz8JznPAdf//rXcf755+Pd7373CtRAIBAIBAKBQCAQCAQCgUAgEAgEAoFgeaBQHuJyWjYPa9aswQ033IA///M/x7p16/BP//RP2L59O171qlfhlltuwfHHH1+c11FHHYWvf/3r+J3f+R0sLi7iE5/4BHbt2oXf/d3fxde//nVs3rw5SFNVFT72sY/hiiuuwGMe8xhcd911+M///E9cdNFFuPnmm/FLv/RLbFlvfOMb8alPfQpnnnkmvvKVr+CLX/wifuEXfgHXXnstfud3fmfs9lgOKK31ip3P33fffTj11FOxf/9+fOQjH8ELX/hC5/3Xv/51nHHGGfb/t99+O7Zu3Yp+v48bbrgBZ555JgDgjjvuwNOf/nTs2rULt91220QnoFu3bsWd99+PY970B+4lyDPaXoxu3O7pxZ7VQDmXBqdg03cYKVVzKWrd90ZYqqwK7eWdXBrmEm9FLq1VzKXxqgZ0cyitK41qYfSxuWxU0VP6iAWUbi73VsPQ3ZaWT991ajPvglql23SqBnSPyZ9cfG3ecxf1UlqcdjSPIvTRUA2pOjj9TNJmra97Jn1bhhqq5h3Ji7ngm9I3tsVa5pLa4PNMnUrq7NNcr6lRzSdsDmi4AMKnxXWn4wlhWzp8RulnLkyO0ldIU6wf7XuanvJ+ItTQOH2vNFA38qAauOMsRR9Fr7mMmF6A3hZA8orQztJt+mQ4Rj8wZcTK75R2zPKnjkybOp9ydWfmDQCtuQ8dI9OSK2O0mT+X2Doz8p+WFW2bDC1j1bVr/WJtT96bC9QduU91FoNx27RrX5bqK+PCk8tB8b5c9vQL+r4aKOhq9MLoOTHLQ04PjBPhlumkj7Vn81z3tO1Trky/fhxYGVi174yuk513OvB4b6HRZYxe7OmDQV4+b+fGo/ddCW9x9GfrxMi1YnSUtX4/6gptOwyB/r7Rf5Y2aksTq1t2kSv02458GuM9Vg+i6Yh+kGt/q4+b9tej9RiA0ZqM0znouqzvNQLls4zssEmGylnjZPvS138icrtUBnSRm9E27zB2U2ugnAyy39VoZRhXptcmncempy9PrJ+VlJlBb0GhbvTYWJ6l+Tt1isigifUrpkyDkjY333H8wrZt7a6XouUAts7VfNumlgag05jgdJCobmu+ycl9uu4n6w32u0J9OYox5YClraPepWfb/S52v4ipU9GantnbSX0P5OeSNkH5XOuA9sM4elqKJMLvLB90kGFR/cBAd6B1En6KybBI26XW5w7vj6GzxPY0uoxdv0x2Lus6vyTq4u+xAgndp5F7apH/wOpEqjt/5mS98z4mQ7x++PFbLsdJRx6Jbdu2dSNGMDa2bt2Ku3bcj2Pf8IdF399z5V/jxKOkjw4FrKgH3V/8xV9gx44duPzyy4PDOQDO4RwAXHXVVRgOh3jta19rD+cA4OSTT8af/umfYjAY4KqrrpoKbcEBxlIrwazrPRFYdV9HN3KCvMcQpnVftwtNjbKy6OTEpWHS23pFJiRtFrT1aCNQ97SzEajV6BtdxeuoBmq04ZQp339e3GZeeppOV6PJRHmbyIZmIHzP0ejXszTkT64OTj8z9MeghrCHEUaB8vsGcOvJ0Tc2tPeX+zxTVtEmgkdz8nAODV0N7/r8UVR3jyf8tnT4zF8QliidHWiK9SObPjGeS8pKwWzuOgdhBfRRDGc1fzgHWLpTtLNo+nmsfmDKmKR9/H5YcWTa1PmU+y5WBzO2cukpFMLFkl/WmG3mzyWWJ2N5knHC0pyhZSw+6Vq/3Lcxua+89pikTTsn4suc2h0jmQ2uQC4n9IO6r1s+YeYKP19fD4wTEX6X5ZcmjTmci5Xp14/7Y+kwetwwTB8lqQOPD+c0hnM6qg+ytMH7f4ms6sDPXF7Z/Bm5VowustbTJWw/k7lsaYPG0gbt0JTt5wIa7bcd+TTGe9x7Jx3RD3LtY/QLSptaUq3BZN1uELHrsqBw8ruwX/01TvZ7X//JyZBoRvG00SRT0PNsm6fyyWzw0bURn5E7B3Qem5nySzCWPEhgOKezeXbRvSwiMmjidVskv9I2N99x/MK2AyfjYmjqXPtrg3H0F0YHieq23vxYkmdqvET3ErrMK2PKgWL9xINaVGGZJXtHGXq66IGlcwmlaazxQPth0r0nn6TcHNBBhnFzrc/XxbROwk+xciJtx4Lj/TF0lk75c2UV0jzNdR3tt6zus6iih3MAsnutSRIzst55H+uXXD8IDhpKPegEhw5W7IDuwIED+OAHP4j169fj13/914vSmHvmLr744uCdeXbttddOj0iBQCAQCAQCgUAgEAgEAoFAIBAIBIKVhn/oHPsTHDLo5z9ZHtx8883Ys2cPzj77bKxduxaf+cxn8PnPfx7z8/M4+eST8bKXvQyPecxj7Pc7d+7EPffcAwB4ylOeEuT32Mc+FkcddRS2b9+O3bt3Y+PGjctGu+NWbMJljRHWq4vbNBcyLhpCRLXvk/kz4V1yISahPRr8unax2CoML8OWE/2wDa1kvrUhlhqYkHw2z4zgKgk9YMB9Vw3a8Dh++3H0m++4cJdsKJE6zFMrLEsIjdJQnV3yTKb106d4W8ENg1S5aYzFLlBgrdZjLCIpb3N8TviZhkTi6MuNlxidKfp1XzveFssJNnzNUAEmPJwvpxBv8+KwaoysHX2QKEe1/UhDG5rwr6h0axVIx1aCZwJ5UMjndjzPaNaKNBX+pWTcce2o50YP1YKXyJMNfp3YvEiosWRIHyBsE4730XoHqTqU034+QT8wcq8TYvNPpD9Lw22xYORZ15A/AMpoprKF+XbaYbEAXj/xx6VfJv3NhvAm457js9R4cmjzxk4qNHE0jy4hhe1DU2gs0/h7hzdSefjpMnTGQpFR2QYwMtTPNyfrFklYsjH0gJy86x9QGKzVAc2l4RjdFw3NS0w4RA8m5FBFdMxcmal6ON8O8/LM4Xkg2qa5MO2UPvqt9sZsNASlIml8mmOyvrCfonWp239pOqvPkzCtrWcPo4eNC+X9S8r061msGxt4JFZLbj/7ITatbGT4xQnZS/Lh2t8W7/dJwZjtFPYrMj9y/JGqm/MdXJ2weC1aKJcpn0X1ce8agWpJZfWYYhD6uesKKF0GsfVC7H2qzOA5PJ2L6jGc7CsMY9tbVGwEjeScSsovop/kWcKzvm7JtT8NzW+v7aDrL7P/QPSX4rWO4n/TscHtD/CZIs13XN+SdZMts0/kXcXL+GI6yG9nvBqM0bdOHh30aRPSvFpUgcxxQp7TcNORtSa7rjPfVOF8HOUDwm9JnqGhByec5krDLef2FdWwZVg1VHZsGAQyKrN2qMw4m2v1SWf92bHPnT4pTFuyPo6lMyje66Vt1vzDhuqOZlA+NovXNZEQl+a9GnagTzB9EJ2w6FvBIYEVO6D7r//6LwDAIx7xCLz4xS/GJz/5Sef9n/zJn+Dv/u7v8Ku/+qsAYA/njjjiCKxfv57N85hjjsGOHTuwfft2POlJT0qWv3XrVvb53XffDWza1KUqAoFAIBAIBAKBQCAQCAQCgUAgEAgEywYJX/nQw4od0D344IMAgE996lPo9Xr4v//3/+KlL30p9u/fj3e/+91429vehksuuQRPeMIT8PM///PYu3cvAGDdunXRPM3B3Z49eyYjzlgQEqv94eEDVLv77XugvacEocVVkGXGehc6bc3gWHsbq8Fha93Doe4D1YDJkyQJLEE0sQZcVNBNlbW1lGjT1LMaMBZj5L2x7NJVmxdrfcLQrGo4ni6jh/zHMSuW4RpDYJve1L8aEJqa99WSshZL9axuL6BvLHc02r4JrMDQWLyYfqC0NGXWM9paAem+dq1SmgLMs4pY5RmLFMd61lhgUSuZHloLPnJXohq4dI6IiVhiZizkiy3ziTWXk8a3EM5ZBJE2sY9IPqwl5qMOYPiTtU0BlPj239rUu1Z2bFBPDc7rg5ZPrbj9tlRLbj+ZNjF9Uy2p0CKXFtDBo5R6QFlaifWm4yHGWMjb7zwLL67NqTUW501mPCWqRYXaH7tg+sobQyaNlVUx63LG20tpYkAYsfD3G13VZLwQCzzjeVvP6KwHQCmoBXvNlGksbnVPOzLKgMb8d+rGDB3WUtPEr6cWq72wrf06UQ+EUZ66vfOoF3pqKg3UjdxVC6qdD0ibsZ7ZJk9OFhArXudOHzpeKZ95XiPVomppGrZ1Ga4bFUplpXMXXgdF17FsNqJlxqUjlicX2z93Mbfuaag6tNK231XtHELvbnE84JkxGi3fPK/D546M1GHj0b6j+djySR0cjyB/rlhUrWcuqbqdxyugWgrLaj8kRWlYnQZk/jcW/Gt2VFg8nFhMA9Azus2CsU51+JDhB8CdA0xadk6lMobR0+xnhDcceWh4gwjGmJeNaXMNIocY2aZoBABqeWz6nxnX9axGb97oPM0zzzuS83ihPGXos7oTqYfV8ehzHZExtBzOSt/KE81/a/6tEd4HopAVF44nf8r7lT4iup1tk1phZvfo44XNbamtvt3OIWa8YMhkDq9NGHlK1xvWmt/IE1qHCO2cl7NmeEeB5zlHDzM0mTliSbmyvqkrx8O9BWBgloxM+XpGu/NApF7Ki4IBr050DLIegsy6webll6lJXW05ZDwrkgUVu6bvyRrNmVeNbOjpUAZrlxR/DURpMeuKegaODs2t21re9cYBRv/3778c0dem4WAjo6yp0dtf2byKvSI83cQpG94ai9F3qZ6lvUJ6C628ozLE8Twq1SOZ+SPqMRvJK1kGw5PcPE/L9ecdU4aRDZRfzFykSZQK6i1n91Z87znDH81/HW/pmnjM0nvD0D5L1bkmXmCOVxy3RqHj0Z//dDsHqSWFem3t5Bl49nh1gmrzrJZazy36np3fnbro4L2zLjTimolA4KShcqX5bqQ7moHWprN09nUwF7p6SKh3sH3nVSq51hyGupuT3ptfOQ/DIM8BoNeSunltUs/oVmerXBkcQIc6neMNRXiTyl9HhvrzdwRWdnlr+hLZorTrdZ7aj9IKrdwn3uLcGsPOVZ7yxcll6nHa1rl9b6AVwkgMpJ6x6A9+nRy5QMcjXXPTcv25jKSh6aKeniZ/QyeVB3RsU36xvEF028ZrMJgKvXklR5MT9Ya2DTevGVQAhm16H1rROTunBQuWHdIFDzms2AFdXY9G/GAwwJvf/Ga87nWvs+8uv/xybN++HR/72Mdw+eWX40Mf+tDUy9+2bRv7fOvWrbjz/vuDibj34Ex42MQtruBOGLEN0PZj8rMgXJemi8uYUCSTQhCyxtHwGLoU2USiB5BEkbd5MgtbXTEKN81+qOzij730mnOljmy8xJSkYAEBdyKzdNMFBlGiKrOxxE3IQKDM61iILarYMgeVXF3YtqPKGhe2Usf7IkBuIo0oHDne5MJBsuOkAQ1LximVrMKQIb3+77WOEsgtrs0mRTTMQaIMfzNBeUo0HY/O5jizccQu+COh4Axo+ByqLLMH+rSfE+EOYgcz7QOebvp/w3vcgUB0jJLfqXAMQag38x0dwpyMcxR0V96E70cYrmkW2d7hE3sgEauPBxruiAuTo5338XyczZQCRSw1dujhMZuns3ETyi3FbPpSPmDp4OZKulBjQqHA73vvtR9ezW+/elY7fW4PSUx4Oi/vXFiQ7PhhFk1siKmEjIrNDzZvbjOZooa7kezTmUtONrCicy03trl5xehO88puZrHzu0q3fS5EXSw0IDUkshtnNewGs0lDx+j80bVzaD76gBITllMN4rILaMZGYg5w8uTmd4T5081t+s4xeuHmEHIwz4ZDZmji5hIw49LfHB2uDfkwqb/5c75mximV+9yGBNM/7CEJg6jMTLFfTp2qEPKTB8f4i6Qz+dNNkIWjQp7kDJ1ycsLk6YRfB8Nnnm7cviB0cjLOr4fzAdp+zKyLRnw6+m3nTyIvAH5sGQzWEz3TfEfnXGrQRuWaXye/7zLjMziM83Uers24sZE5pHLyzeiP3DiOhtTzD6JJGnbTk+pUnEEYM39E9Z1IWGa/Tr0DFW9sUqAfAYl5ntLvb2rSNAydSxt1y691O7boOKhN2HGujQn8kIOjjMrqUAQ6VXDGlFw7RuS3U5fmJyeX6Pq2ZE8EgD2cM/Sl9ORce9ADCWddnRmPrEEl2Yep5t0TNj88pjXcpGE/qdzym8pbn0bXHojzJD3YSqahz42xJDUmNTST7xR5z+XJrkkVb3zI7oGR+isqV6jcZ/qEjlOO/1g9ivIu0w/s/BpZ3/r1DvY5fBnpy9AGZg/OGS/9Vo+tyQGXU36C/x3e4+aKjL5v1vdqSbkHc953YcHhe0q3bcuccTJtR0Int3bw6xTjTUeHjq29EPZrVicM9Ac+TfQqkObb6oAZBCQNM9f6a0Wrfvh6iEcnt36l8p++Z0PqUj1lkrlIMDHEg+6hh4IjoeXBhg0b7O9f//VfD96bZ//6r//qfL9///5onvv27QMAHHbYYVOjUyAQCAQCgUAgEAgEpSUnVQAAnN5JREFUAoFAIBAIBAKBYEWhC/8EhwxWzINuy5YtAEYhK48++ujg/XHHHQcAuPfeewEAxx57LIBRaMx9+/ax99D96Ec/cvKeJnRfd2bunDXXRNZvHcGWFbEoLvG0cfIAnw9LR08nw/hFQ4GOIVii9fAscbJeRNmCCr4h1mBd889ZyE0DybzGEeq5NDFrrRJ6Yll61tyBBZ5vRTXhZJW0Vivs5y71ZMMArlJMizeLrGwj4zgFLrwcDT+X9RgaAzlPIPvdpPKvJExqqZU058EWa4fS8KzMGB2Lt8fsD9b7gD6PvOdCwRSXRfLQiffxjMqfx+qXzQtpS+2uMHTQcIQsSBg7tk8jeop9HZG1rQW4csZw4OzijfXcOC3VF8aSF4Xtn/S2TSDLu6XDkPOmok4U3dXlACwvGt1tqOCHlyvJb+J5yThKdJF1mXBVpR5+AO81OQnqvo7qwWlCJiu3k96TCVvLeeRyXp2sR+yYNOUQhHfzZFTW0ycTUj73npabKm9aHlhOmSRE5FhCIOfRQZ8th0qc03MI2DB8tk944rJe8CY17dspzstsWR3X8enMppDHMuZdzPOZNimVm46H8jgm8ROWP05ZWS/ocbPPRSPh0uT0dbTvp72nltMts+lrvi1z+hv1Umrr3+qxbPSDDFLRSNIJmzIj4bKTeZJDgXH5lW1/LsLAakbHtaLv+Vbqbew/y/FuKmqA/3/2SppDZ0vqoQ/pi4ccVsyD7ilPeQoA4MCBA1hYWAjeP/DAAwBaz7lNmzbZQ7pvfvObwfc//OEPsWPHDmzZsgUbN25cLrIFAoFAIBAIBAKBQCAQCAQCgUAgEAgOHvTosLXkTw7yDh2s2AHdsccei9NOOw1aaxvGksI8Mwd5AHDeeecBAK6++urge/PsggsuWA5yHaZ2mN2gXgEzjhLrSeX9OR+EacxdBPQSVicfjKyCnAua/QFfJWhTkfZrUBOfTucbjv5IOUHeXpxlR1Bp973SXv3M86FqLYhSbcrAWGFpVWCNploaWfoLyqIx22msc2XiRKfyo/0c6SOuzIOBXDlZqzzKe4wF21TrUSoPOlqSt/Qz/DhO/oyHT7IdEmObs7DL9hkdh2NAexeAl/ah7qO9cLxpPzpenDbNjJlkuTm5y433CKpFZS/nLqajaVv63PD/uF44wQXviXYyNOk5zcvlRNvqfpvGn4OAAsvJyHtOvjv0RPhRV2g9b31r3pkco4/+ovNrIVTNp9EzOk8DwnR+XmN70BXOhVAR+unY8xGZC505NTbOFFzvAzXSMepY7AhPPyntm6792AXOnRuZfmZ1l4weyNLty6gETWx/+uV18Vrg9Ey4fVLC6z4on3D0B+1AvrXvarT3bZTIbcJvuW978xOuIXw9NlGeoyeqtr+CdP79f17fxCz5UzIu4MdctWr3D/B0W38TYsx5LVcP7nl0Tvba05f7qbkSyMjDyHtalv1Okf9n2juIHACmbf15uNEHszLEf56gqVpSNppBqs04ncDNKD0OonxI5moDrm1Qtc+dNWVOX4zpHyl0laN+Ou5VhK856L4OoxyMU09651LHORbAaI0Rk1VevhxS+nOQD5U3TP/bZxXTPqpD/RJ1AZox3eielA5ObkfnsUi5UZ0pQWsRH8Z4rll3cPqXntNQA9XOKRPKyORalbYTo887cpvyATcOEoh5Tef4op4Z/Tk0eXdb2/Sp/TYnA5T1rf/Y8NUA9n7NUllE+8eun/xv/LEV6/OmzLrvyRaOT0plLNVpllGHd9YtVZ4+u7bh5h0/39ScnhojirzP9CW7bqfl9PO0CJYZuvBPcMhgxQ7oAOBNb3oTAOD3f//38ZOf/MQ+/9a3voUrrrgCAPDa177WPr/00kvR6/Xwt3/7t7jpppvs8zvvvBNvfvOb0e/3cemlly4LrRVZiLKCMxPGZlJwk0duIWrdlmODMxbqhUyktp5NeqXdze3UpjCfeXrioXWi37ATmtks8Yvw8nbS0c0uuuFK0zKKhO7pNlSFaQtDU2YUOZtVsXah771FdhehGmuz1Oaym0H7ni4AcmV2VW7GCWk3bigDh4fNM2ZTd6ohjgrDGXYNNWHHDsOPHGIb+mOVD0THHMDXORv+ZcI2pxuEsQUAC1oP037cM/83g6wSHSRof5oxWpK+ntX85dweHbnDLEce5ZD7rrBt1IJy5Yk/N3FFD9o0zlwWI7VQBnF8ojK0jBImaM2FreLkaoSWZDbeJg0tvyR0VnSTKfK+WK6XKv86LLOo7Wk5HpQJkZnQQ+xCVCM8nPY2B8y32Q0gQn9nGdABzkXymX5mdZeMHsjSrSO/m80gZ+MlVvdxFoZEzwxeUVkQMxLI5M3J26g+anQxboxyelqszIbfct8O52LCpf1jacnplhxZkfoGz2t37PgyLKbnxDaS7cFHB74w6w1ad249wBnBsWuMSduJPI/OyYQ/gnYg76NlkXHA0ptLz1YmTGMQPZhR4Z+TT6MzseXGNiBJerYfiW6bajN/8zyo0zDdJjE+LN3c5daq/jiJldt5rugqRwmNsX2CGF+z+dCDE5+m4GPEZZW/1k61BSPXdU+zY8uvQyxPZ6yWzhs51My8TNdqOf0uoz9kdV+NcIx672Pp/HdZ3i/lw5g8aXiCNZBdUNl1rbO/kTMyTK1VaT6MjOV09NEhFTMOUsh8GjtUzdUtN5fwiTLfRd5TPk7xcmwv0DFeGEPH9+edQK5zfEINcHNlMeuyScAaDRA+rAb5tqR5TXJomOIjZcaHWT958pCW67QL14+xNZjgoIEzpmINrASHDFbsDjoA+LVf+zV87nOfwz/8wz/g1FNPxdOf/nQcOHAAX/va17CwsIDXvOY1eOlLX2q/f/zjH4/LL78cl112GZ7xjGfguc99LmZnZ/G5z30OBw4cwDvf+U6ceOKJK1gjgUAgEAgEAoFAIBAIBAKBQCAQCASCKUMO3x5yWNEDOgB4//vfj7POOgvvec978KUvfQlKKZx++un4rd/6LVxyySXB92984xtx4okn4vLLL8dXvvIVAMAv/MIv4E1vehPOP//8ZaOz9sIXAAitj4glxrTBWop53l8sjJUQCdVgX9WtFQelnXpd+emo9RNryV+7liFcm5RecE4xSZsGXgMmNM6Qf5+FaQvz/3oKfW8yi/BWMZj0NOxNlM5IuhJS/LA6QXruXe7C4Qo2nALHJ1yZufHgh1maZD4bhVIaEWBkgz+eSnih9Luu3xokrVMB1lptXBTxGRBY8eTKTI4t4jlgx/OYfuEm/A21so1akZUiw2T0susccm1KLfW5uSnlKeXQ1NetlSgnD8ZpB0qTyacHVEvNb67P/LkqUy5nOezMawzfce9jfVLKhxxiabl5OYWoDKDj2dAfmZdy/Ujfl4znOLHNP8R6NNbOKTpz5VrvnCYvlhQvPJCTJ01T83ySQvQ96Qdbdhfdh0mfmqedfGkdpigjufRd+ClfQPMtSeeEMgIp35MnJZhYhnvtn5uTu8wfvu7Mlc+2JX3PtFWyTE3KVGQOpKGjxwh7O468iKap0+9TcyirJ05xfTaWfkDGI12LOPULp1+23Jhubb1QcmvVpnw9Zh+PU2cKbl0YhJsEomNgEt2xZOwm5x1mvaGG/BrArBV8j1OWPpOe5GXSUe+dkvbPebx06j+U6WH+WtZJ02XRFRnP9vUY842TPlW2x6/ZPGu+f/MFjcoK9IKOe0sp3ZZLF5trrGe8DnXTcdafMTpLdGj/26LyKZ95+mfSQ7EgT+3nXwCH96ncz9SJmwti5U6q042F0nUhldHjbriM0eZ+O0e/5dZ/kbViKXJ7tXRf1ZEbDO9m227cNuXozNQ5tz6GiUolWBlMwAuC1YkVP6BTSuE1r3kNXvOa1xSnueCCC5bvrjlgJGQqxDe4mAnfEbol+aNQ2TX5m0MAqqBz6elmDFHC2EmjB3YiS4WIcZQAjfCwq+CONWexktlIHgfZyXWSjdjE90l6qALObbyBvAezYGczbunhDtMMrxSFW2QUAU6x5ECVSGdseAs1N1GGHqKkWBIJb7BtksnTV8KSiNWZKOvaHOR03BTzaerybVcX9eh4KtxgLC3Dx3DDSCBU+8IYcX559dpRA1YHylcYHL9HZY/P21RG6vAzgMi1QXn7jLNR7cjNhDIe3bw1STIywl9c+pthNH3FhNDTypOXTJtym0ipua4akPcMH0YPdsbgV3azQhXyTwm4BStdlOXmktIFew982B6mfoopt4v8cOStGcaxcGIJPanutxv/qtLBojN7wFYrIDWHeUZDqXoA4bzvGxakwnL55dr3qY2TiG6V23y26WmfG5rJJjzlMyeNdtMA7b0mzoEllS10c5hpS+eO4JyBk8nTlJXRIyitTp+YV2QjUw3bF8WHz55OUnw4TvvJmzeyRnKEnxw+TB2+Jso3xiPV/CgD3dOubCN3DpfAN8KrmdDjyXZq1klcvobm9mH4HR1vWdbIjDc2zaQbRrHNUeU+0j2kQ5LR95rvHxq+MzdvJA+OZnR7qEn0C/Zgq1SeczqTp6M7zxGXQTFwsik1XhXhg3pOB6F/g7HpzzveWtZH8UEUfeSF/rVlmXGlI/1AyaRytfl23Y9HGR94ZNthY/E2palgrmwLG6Oowj2BoMyELqDN3gwQnX+yBx6Fz+qZVp91xoFPmwbqudp+69epXlej2h9vBDr/Gt277munLIfnmTr670d5kGd+GNkIOdVCm7HV50kZpXsSsb2pFBvl9sNy49XR3XQZ/43uNGz1KM6Aysmn4zjw9eGgHyvvfeU9J3PB6KP229GHcA5Vk6C8m9E9uyCpT9fl9EWNGDP9GKxlc2tpMm/YMJhM2HMAyb26FJIH7kTuK3o1Ep1f6RyKcGxw67hJdJ0iYw9mDWgN7Ss90vPHmCcEk0NFdMrYt4JDAyt+QCcQCAQCgUAgEAgEAoFAIBAIBAKBQCBIQA7eHnKQA7oYPAsXGtrRWurNtNbgNBxFFJ71y6TWnaxFj29N61tiEOvEei1j2aWA2Z2jZ4uH1+Fpu3+JsPeehntApUPrmhoYrnWtzRzyI23Ce5+opm4uEYE1mWf9WfveiJ4FvvU8m9FBeidfwiPG8s2m8ekhlpVs6EZDCgn5FrUCb8qmfGQ9HYzV31Chmh89G86g9eCLWU3mLoZOQOn2E8cTIWE5pfu69dShY42xGnToYCwFk+1Ey69IPTkLetrPsTpTPkiE4JzZU2Fxo9uovqcG0HiXmHwyVqQ0vR2jm2p2HDiWoGsaD7V5hlDPO8XyD7FGdr063LKo1d/MngpLTZ2N51yJZZZv+Tx62JRZAdV8U+YsKZPrMwJH3nEy0vTDQEE3VmyDdSPae/MV6qq1qrR1Hbr8GpTJVKO/XzV5k77hLCUT+QKuhTjbphVpR07uUmvzWW29FbnQzZr8dt5RC0n/dQ30mjyHa0hdqUcP+RYY1aHivHypdT+VJ1ydzCOm7atF1YYrJfT3Gm+u4Szf4KwnINDOa5w3V9VaH1tPCLT6gePZ4s/RpE7+nG4tqq1cJ2mpRTNpG1aeRMB5SmgyNtBzy6XeXJyHPuflrIZoPQhIXR1ra8bSlLUy5VCwMCm1oAdaWmg7sh5Tho97LU9Ta99ouMnmGRuJgZm3evMKwzmvkqRNqDykFsE2m5k2ZG3/AOF92nd+nahXHuO9y1lrBt/5Hi/UmrpG1jo50D0VbPvVs6087O9rZOx6nZ0Xsl4JlCZqkQ24UQ2oY3hKbs+QelCvRapvUp2Dk2PEq6LVWcmczek6tqA0fdCtp0U14GVsoMeR/OrDl9B7YMalw9OtTHo2nHGtmI4mIHrwmh2jhAtH1E4/GXldLbRehX545KDaCdno8DGdi9A+88duteTWy7egr5ZaD7FqQNYYdA5i+NB45kSjCzBjyPF+JfUcrm/0m31hXj7v+/NWtdRWlobPZL3BjFw5YgnqgZn2mR9pQrffqmGrTw/XMXVmdJ5qUbXzIuf1T9Z1QNhnzpqd0f0V0nwCT0aavH3Pf1OWocl6NkU8Li3PLCirj88fPUYMUsTnuNh30HDWqgDcuSSSDxuqlZmLqOd3bP0e6AcDN2Qr9USl37WFeO/pKzIH2eQMb+kZzUYt4ubyeqbVfU0aOn/SdVdO54hdpRLMO6RNqwHRcwlPGXnrhBmkYNqJenoG61pFPi2YV4IIOBE9zybxdF8zry9taHVQxxvdy6s+cgn9n84CGOk3vjc564FUtXOqJusmys8sH0f0SZOXlUtUR+6FvKtqIpf3VkTnaOa0oXL7ztNJoNs+U0sqrVMxbUvXpbnrFmib0Osg/PnV0WcjET+49ot6nJrnJNKOsydh2nyu5d3UHqsaqnYPMKKHUZ2PzSchY4Bw7NTrh+2eyNEL0PfNjb6L7EeZdbGVB7E1yiKRoQyfuImYOhh+YqLFxfR5J0saSrtgHSZYPohn3EMPckAnEAgEAoFAIBAIBAKBQCAQCAQCgUCwmiEHdA85yAEdB42A2R1r/cY6xrVcLTAh8C1QCyzcDayHlmcZysUe14x1B2etx1pl6pHnXJAXLZNa2nDWM9QizjeOIxYvHGIeN+yzzN1qXDtQLyALL75z7XvBMXUYEZtI44NaIXLGsDrxLkILax1ELEmHa41pWSbLnHXlGPdf5YZDtcTE9Uc6TepuRMDjHS4fz/OT9dDrgJT3i+89B/DWYrG7flgLQ/Ld4hG1/Y4bB/Rb1nPOfOdfWp/xag28VUnWS4fV8bGSQDAegdbCcEjGFmd9GumzrBeRGRrEYtXK2J5mPWrHgWOhyPE7Rcryl4xrtk611zeJflBLKmjT5G/v/7HxnrTgilplcnNp84/Pth15i/NmBuKec5bnIv2dnG/q9BwQvZ8o1WYKrucZEFgZx+a4GALL7ZQMjHnpUFoy+QOIW1fS8ZyYo4ru+crUWTG8HfPa42RHal5xvBpzHrGF/UTzGsa8bDL52DoT2eHwaKHXPK2TaRs1UAG/F+kupXejkT6jbW/nTVqnWZIwNf/Hhj1tPyY6hO+1GnyXQOwuYfZ3pB9rxvs0a+2esfD2abRpOH09db/j7hlWhrDeCmPo8KgB1aRb3BSuSxQinjgdxllAU8ZqnJM3Pu9zuqHDB4Xrhey9vBnvQEcPbLwSou2R8nSoAI6ZkrrrbrK9EIu6wng6pPRVCupVzN5hm7gTEIjL6uK7aDnPstj6lbnTKk5Ykz2V+5k7YKNZFa5xWBnUoaxsnUw5NFpHyd3oaHQgf+8EiX5KrZ0iMs6/U1INFB/ZgymHvQ+VlFEtKl7PJeWz7cfJMKZP6r5mxxH1NkvR79MySpT5PpNnVDfssEYYrPVe++t/73v1wExWV3LSN6gbj10uakyU1Eie/prDn0u49jUyXhMvY1WH+30mDx/Wi6rDPFdzO7+ZOjuewaSeyXIjMiSVxo/qZb6lc6HD51zUqoR+xN2z6+RZoFt33Ssy3nMARt5zfrt484b18MuMs+ias4Q+nf6u053sHdtDMH2IB91DD12G4MMTjBDTyt1AAMqVzdG3ZAE3AR0cTZ0Wo1x+Tb1M3ZQf0pI8M9/5gsF5ViHkMu6ZIckLE1N68SV113Z++59ptOGnFP9trk7+e6WRrBNFrt+5slN1Udp7X0hHFwShOBVTf4Y2///cJmkqTfhB+lunbRXatmB4o5ivUvDHJG2TTP6mPfwN+BxsO2bkgZsI6UMhyjOl+XLjJrHBk8wqMc51D2mepvKC8FhWFjL0G7mgFZHRKp0mB1o3Vq6Uyjg1CntRFRwWOnKZ43dVUG7T5izN3IEqgMFajcFa92PbpjG5VyugVoE8MTLV6dO+DsYLKy850Lbg5GYHsGVG2soWz9DO5uX9Dvg4QjPVKVL1i83Zfl2i4y0RjjCWp897fjmUT1nQ994cYGVoas6Plc/IOpYPc2Boss8LeSyrRxla6ZzG0IwazmadpvK9i14Q08+aPOnGQGk7BQt+QiufANG5yNks4sYWt4GWmWv9OpTo6VGeJ2OwntGoZ3RUbhqY8KNFZZG2KebTTJ7JzSBPx1VDNfrzx5aRr2Z+UbD9TNdN/vyUNVoB7FwBjA6KKxJinAsvOQkojdy4NGDn5Nj8yvF7pO5B+wDdZIs/10XSOPMrx++JMZpdA9Q8gU45hfWgv2nb2Dkgx0fI6IQRPTIFEzqepimaXzPrUxYRWdh17JesS7g+j+oXuc15f+8gUgeWliHYEPW2H5m1aA5V7NDNb18qw2Ig5dd9PdpUp3Q16YdzOs17/jgF2vCpjJ7ZdX2f5Hkvr6QektNlKtJmERqdecE+HP3Z9ZZ5V2k3lHeun2PrlcwYqRaVPeSyoLxLdaqMTsDuDRV8Z+hvP0Aoq5m5NkBOpzKfzWo3/L+Kr1EsEvpYkqbS9/Q7A65NUusEU6XCdTU7Lktp7LJ+JN9qbrz7eTVjk+XNTP5d3rP7HCYJ5eM6wrOC1QNd+Cc4ZCAedAKBQCAQCAQCgUAgEAgEAoFAIBAIBKsVXQ7f5JDukIEc0EVgrPKMxUBvn7LeAdqE1lhQ7vfGujRyKS0NawcwFsXmPXfBKRM74cSn/BB3fuuxNn+bJzcAw+To71MYrNduekW+MVZNlAwAa+4bEb5wVO18a2AvyB2okBaSP7XG4C7rpu1THz4yl612tSxLL61NCZ16be2EEZjZMyLWhEaiLu9Ke/kaOshlrqYrTGiE3t7KXnCuPUtMgFj7IbxM3n5LLLTM/9kwQWjzpG789j0J09drLLKHa9MX+Y7q2qQjl0SbPBW9hDphEebnay+YHyAIORB45XlZq6FyaPE/sB6QQBsOaKFNoxu6KWjoQlOGeW7q0MUDVXF1YZ6xbWP5QQXPAGL9pfh0ll9IO/g8A4zGDR0zQR1ql2aOVtpO/sXbTv0YK3aW16ks1AzvEdiLzuGOSzaECgkDZMdDrVA/YmGU7n4SC42RhxR27KzR6boQa9BcaKNo+L8GRq7OHxWa/qm6DXnEhUXl5IOlD6O2UcQidP2PR2XtfWzEzNCTR45VJrFqpe1IL/EO6OfCxAEYHjZs0/h8pN3f/njOjVc6F9JvTTtWiyTMbkVo5MY1zTdmjUgt/eGORzvOdZon/LHhy80jv6lw/2mMVbFn/W3K98vJhmAkz633FRANYWXbtBEx1SCsk+4RmgbK8o+5IL3uZ0LS0tA/pH1yMpSWT3WzYH7vub/NfNDfO0pUz2o+vB83L5G+o+3IydU2o/Zb+p2heWaPsuH9jFcKlYGUz7k5Vy3xIbZUjG4AiIRent01ypSGQefqwep5Hu+zbWL4af0QvT2hqbqVZ/22z+h4RkpWDxUro3zdy9BqeLK/b0TU0oa2wWojt/b1kmEGdUXkVqV5Od28HqzT7JjsNRbUg7WtUma98XQ79mg4zfU/bOT7sWkzcl2N9CYAqOeY8isApqyavDRzAO1nMoZyaxB/reR8RsKnqhrY9N1Rwt2Pa8jwPLJn9jbts47QZIrqa9v+HB8mxwBC+Wue+bznzMncGsKXiaasQatbcWOfrgUdWpj25epC62TalFrks/MfwrGp6tGYG1W6eUd1cCYtlSFUbjuyXjPpqNwk7RPUrfL4h5E3jm7m6XyO7saFq2RA69HbXwVypp7R6ZDoCvz6nlkDBfLUEtg+Syx7R/RsbNbNJtwoaXRHf2r+ndlZYamR7e4aIF6mM9dVpH/IGDB94rQP7RumIvUckXU29F9LmJ3TvD2P1LziRDqK1G9UOFDPhPRzYfDqPtDf36wXCM1Wjzig2iUsuR7FkUe+nl2T/YV9VRAy10cwdujY0ETnYurMrS+5fZi1P62w/zHN/MnoeWrQZqAGujhEnrOWJGPD2UdC2M/Wu8rwxpK3Z8CNKVMO0ZeqA27f+aBr3mzYXK9NaqKnVIsqDMdI1yhET4vR6uv2dR9274nKFpYeOj/1SdtS3Zz5llv/sntPXKhphximTglRCcBZy3O6K/WAo8/mdpA9SqYcqycvtjS3ewahfhDQaXUeps7eeiVIrzR0z2NOwq/QLV3DRl9VSyqYv0Yvmn/8/UKiJwdpTDXX1ujta/ZFe3EdBfDmevsQPE2CFUFuKAkOPcgBXQT+Ame4ppVAXGxyZ5OE2ySiwi6mbKXSM5OeOZxz0sQEJfN8uIa8ppuilL5AGwfmj66d//ug9+UFh5E6Mml5CmaQJ3PIUHqvj7NhrIHBBu8Db3EWKK7expKBndzo5MUskunz6Kavf4DlpWUVGa4scthk76Aj6enBj5NVr/3Wz1PPhBs3gVLv1V+r+ELclp/oc90Ly3Q/CNPVs24aWydf0XAyIf8rmeHIoiA15riNZA7Rzf7M4oYbo1xeqcM5m4aO/Yzs4cYctwlTemehVnDvrPS/ZcZW7I4YdsHZ01AP0EuKwnT0W8DdIIyN/RR9FLmFiJlL9KxmD+ZsOaqlOTe/sM+9zZh9xxTGAUtsPpt8Ddj7csgBD0djby+RoV6e3IZjlA4GdNFCZZBi7nOlizKfDptfaqFIF5IMfdzGjqERaMsO0ng00MM553s6hyX4lTvED75hDFNyacxBOn8wQuuvg7kueg+nQWbzNNio9csnRgimPId2stmkBrBh9GIbNilaYu+S9cvM74P12jmQAfjxAngy0G7sEPnNbGwU0wn+YI6dB5nf/sFNqs7V3l60rwE48y+9gyVJviYyymmf9j3XPkuH6eB9ZQ4PC8eSyTNGF+BtmBPYjcohXyBXRrF812QjnJMNNUJDKVIOK3dyfKC83z5J3v02O0/RQf6Uz42RIXuHGtl05GRwdgxwIoDqBDH5mJKbkbHJrhdytNDXzLeUn7l5j85f0bzgzgVOGQyd9ndk/sjqKvQZ5T1/Po7MZe1ayKPZX8N0DBXo08gZ2mTvKy5Yy3Fg5UjBpmi129X/6X1u3Ibu0uF1dq7yfzvrvwzNzt3jEfrtgUhsnyXoSO+/ifZ15iAqz/xQj3DXr1aP9A7eze/av9eYKK81Mfrh6FQMbwPkbjKi+8fgjDnAlVGRhubayTk09PYMDjyyTt5HOloLhXNF8I33npZD8618eaXhHhx5tPh7BlFDeJpnTfZJIm1cer9iLi1rAEPXDZk7VBVZo5jXDm0x2cLkWSKn7M+UQTnVI2PreG+uNOlKyo7SlJm3nT1KBq0xETMuc0a19Fuuzkw9Hd7WyjlUNd9xh9PGeCrqeBHRo+wBY2K8VgcqXmeLIVG+YBVA+uIhBzmgEwgEAoFAIBAIBAKBQCAQCAQCgUAgWMWQ+wEfepADukL41vbhB10ym5CYKSFmsZP1XMrlG7Ey5N6vCEqt0TLoYt16sMF6rWE8q9GyAiM0dPh+OWiYelmlhunLzePTrNOU8hq3ztl0hfRFrbQL0zvW9l3GSSL/XN0Cy9uDhIMmgzMeWKWeluOgc/ouHgrMs3HoXTZ5HCuvgMaxPAtKPWZjlqCTIGM5nLWCzegrK43YXD52XgcZnax0x8l/nDEY0Xdz3i3B+ynWqYtnUTavMeavnIffSoHzuEmtx1Z8fZHDQWzPLrJv6ijUlZOYZC06jfJXEjm5NGGezuODNGZKypnEYylPQOFnETpzob7t/6kXTiJCiE07Le+UCWXLOF710yq/KOpAhzFRPAceLDkx4diL7umshE7XpcwJ23c1yaZp5B29ooPzoJ9wPbLqdSHB9HCo6zuCAHJAJxAIBAKBQCAQCAQCgUAgEAgEAoFAsIohHnQPPaxCe+FVAC6musq8x2iALNsgydE0TpZDPgNaj1SdlIkJ3pEOVafvjOHo4D8Iy3ZoytAWy5+jz/m2Gv05z1Jl+bwzbaiWltK+m6Ss0jr49HQuqmNa59vKfe7nU8J/y4nl6Jux82v4uawQ7y/zvqSddV+3d42k8suRxrVpLD3z3Bk7hTJqtWPSMRjLM1XWsmFMed41n4OJiWTcagJtU2ZcJZOOqUdkZUNORtH/k9/LzsccSdwclRq7uTpH2n8cnYo+L5YnHeT2WIjI7SQic5xNX8BHZl4oLdNp+9x8UtJWnO7JpB+HviCvRP/Z/MdYQXYdW5xuW5z/MvHfQZfb3Ppz2nUrzTMh651nHeaAsejo+q2fNDbvTFJ+1T7Ljq1pYBXpMgG68FOH9P7YW5a5mq5hYnxCZXFizyC2z5ItP0WX/zuV1Rjtw81VWTlM54LYvJFZVyX3LyYZp+Oi4/pz4rLMzwg/+XSNu75LrpWngcK1WnYtGWl/NVTjjauGtiQvl6wBJmmrLmuM0jVv1zVQwXtu3ZDcCxasHHTHP8EhAfGg48AwsXPJpyLfmUd6md2JuUG1TAONhiTiLlbnLnt3QNpH0W/NO7PZwbSZ7msbak4nvovSrkLFxk8b0OTnkdt08MO2KQSXsroZkjIzWY8F7baVfbwcoaXGyItt5xjvJNBpjJVeSDxFOPRVSLqcO7xd2k+Et506mYujY5dcmzQcX1dwLhJmxwaVd+Y3VxZT55IQDdWiin6raqBuZqlc6JtOIbrMMyKPnH6gcbXI5myyHA7kgmzbtj3YdsrJIgqzIMmFzKGIXQQ/iWyIys1e8yPCG+ZibM0t5JkNFz8cYq79g+exMZipc7RPGFqDd5H3XfqZpukiG4q+KyjTgAtHGZPBTt90pCEX9tIpk+ZPxzCbMZM+S0z3+UIRfXHckDSp0FK6x8i+WHtl2n6scMJkHMXkSfCuw0HKOPMz1cep3EnOdZH50RH1mbr4+eZ4S3PzSwSx0EcOf3P9kOCdHCh9bNjPjKyd2VVh6bD0yaPfpoHuUho2nMrCVJ957ydBjkdL52RftzI8m9NpWH4es07ZujC8NdYYNXJ3qDrpKga9+VFhw7ncRI2J+ncqoQX9+XEY4XPvG6DjXBHj52VYUGbXAKWI6UkT1IHK20nmjxyqRWXzH85qnkY6XnJ7Q5nyOBnGreU5PYvVDwg6tU9mrkrlRXk/Oza59xUAUo9gPZHhm3HlDSdjo/s1iXEQnUtp2tTaoMroLxzNXfa2CG3cWi2Xp6PjknUtiwn0UFW7c31Qfs+db0vn0tLyS95PsnelhkgfmoHUadLreHJ0Jt4XhZwtpUOwrFAoPySVrjp0IAd0AoFAIBAIBAKBQCAQCAQCgUAgEAgEqxnL4v0hWEnIAV0hUpeSB+8PEUStjaZoVRi0S8bi2HjPOUm6eMakvi983zmNnsC7Z8roasW+nMi1Wdf00+zXSS/fzZZfahXexQo4NnZyVlaputZunrmLuZNljRkSMkWfVt0s48IMyt9TC0JHNk5iqct5co7p3TmOdSibz4SW51FkeIPt55glKvOuc/uPy48d55vsu0SeWavE0jaYUn/m5o+x65EsdDKaOrf9lL0QHMvpZcBE8g9T0D0mteKdcrogLZWnE87rXWkq+r6Qv6K0T2AFXVZw+5P1Rksg5z3n5MWhJAR2ysOPyX+aul1W3o3pzW7HdKadD0ZdUu+z+mDKAp96OXQYV1nPuYKyVwpRrznqPTlOnx7Eumb7fKLMJ/s2u1aZEuqZMSvMyKVo+H6aLDf2EmuQSfUDt9DEqy7yI+atnooIUZflES1/zPURNx7H1rPHWCPQdWGpHj2OHp5bx2fznKBvusDvj6B8xmu5WwETvh8DTt9xnpq+1+K06zQhDsX97YcbJMzoQw8rfgfdf/zHf+BlL3sZHvOYx2BmZgabNm3CM57xDLz//e+H1iHHDYdDvOMd78CTnvQkrF27FkcffTRe9rKX4bbbblsB6lvI4BAIBIJDF1qJIjoJxrkLQSAQCASrAzIHHmRorMqDJsF0MfG4WmYjEMHyQXRigUDAQfQtwdSgC/8EhwxW1IPummuuwctf/nIMh0OcfvrpeMYznoH77rsPX/nKV/Bv//Zv+MIXvoAPfehD9vu6rvHSl74Un/jEJ7Bp0yacd9552LFjB66++mpcf/31uOGGG3DGGWesYI0EAoFAIBAIBAKBQCAQCAQCgUAgEAimDDl8e8hhxTzoBoMBXve612E4HOJDH/oQvvGNb+CjH/0o/uVf/gW33norNm/ejA9/+MO44YYbbJr3ve99+MQnPoGTTjoJ3/3ud3H11VfjS1/6Ej72sY9h//79eMUrXoHBYLD8xDfWbM4lvqrcg0D39SjsgW85QfItycdcUE7TOu/pBbBMnrELSNVQjfJWo2+C75pLYqM00ktkufxrkq9Ht7281qeHa5NIOf63frp6TY16TdxHP6izH0Kn6b/eokJvUUHP6KyF4+hiW9h2i9Yl03ZsGsoGlHaOplj+pRaaDO+zn2n3z0eq/U05Nm2ENr+fgj7z0jk0DcMMUzzD0efn69CYu+SZtksH61hLf66fyfdZUJ7I0ZLiT8OLXttkaUqUqXvte5aXlPdbtWUoDehZnZanNF8i1xwZRevn153jgww4GRfN38u7aF4gdbHPvLlAz2noWCipTJv6vy1qNfoj0DPaykc7r5C8ewsKvQWVlOOl44TLn/2Om9MoCueV1JiL9ZMf7ihFi6rDvKK8Q0H5mJv3x/B0pDJu0vQBuQN+fk+lcTOIFUr+uHwbaBrOKjY/Z5DlqRR9EdrY/svJ5+Y95ROqZ7H5c/MmgcNzOd2PocWkCUJ9TTpvkW8tb/oyIPYthwgfOfpiQ3M1UKgGKisDqO4a08dz/MbWj7Yd174d9Mh6XY16XR3qFU0+qbmCbesx5GIqvdGduTHGzWV+OwfZTxqei6yLSuflgCZvbHQuP6Nfx9IlZWFufjU6VbP+4fQMJ08mlFeKNq5Odt3EJfHHQeH6mY4nm39E5+HoZPOiNMX0xCl70KXWRjH6cmMzqm9n1n1R3Zxbf/rI6LzZNu84VyTpMyQtKVSLo78SPSqQy1QNpdd2xMos7UdOH4+sQTrLB0IHl1b3kJ5fmfWfn3dufo6tN3Lfmu87rWWbb+waJVWGJ2PYT7m5hjyr14YTj88zSbnO8ZY/Zyv+u6BcTl/tMsYTc0mRHMjAyYebYwr3uWL0+PM3he5l+qFDnbi5MBbylqszpTOVJrouzPFrQv+oliaZsAQHFTrk2difHOQdOlixA7rvfve7uPfee/H4xz8ev/Zrv+a8e8ITnoBXvvKVAEYhMA3e/va3AwDe+ta34pGPfKR9ftFFF+GFL3wh7rrrLnzyk588CNQLBAKBQCAQCAQCgUAgEAgEAoFAIBAcJOjCP8EhgxULcTk3N1f03ZFHHgkA+P73v4/bbrsNa9euxXnnnRd8d/HFF+NTn/oUrr32Wlx00UVTpRVorBQag4JqcfSDXiRM35emDy6nNVaDhYYLekbby1qNNYl/GSm1RBwR3b4frq9RHWiIMFZLNTA8bHRDabW/cqyZDI2m3jFLKF017+vwvdJAPUNo9gRGtcRcEtsrb+dR+ST/hmbOAsq2TeXSYduKsdLTqrUsseUsqbb9IxfAVsSxk730lrMMVuFv55shyc/P0/mwbYs60TcBTfSZdp8Hl9p6z+l7zvqnt79q24ors6CffeheSCdAxkYfbTtTKyaGPp8nwg/Izwj9SVppf3GWdww/OelSl21735txGr1Em2urcY2nfGs0Jh//mR1HzMXIShM+Z8ZlNB9jRbqorAWrsZL0xxO9pNvmQy0Ftfve5OHXha1rj9DkpbX5w/2mpM040H62eVXuv6P/ANW8iuar6GXllM+pFSBTfj1r5oX2WbVgaEJbcZq3kXHDUC4GZeR43eNv3ddQ1ApQ+XSqlk/qRj4A0X4O2srrR9s+Nf9+VI6bSd14fijGWlETrxjDw9HLw5m5hJNhPu8HcobwO/3WzkWRi+wNn1cD5r1Kd11tvDH8PDm5yiH2nhtn3rwKAKBtT+pHxy433m3W4yx6InKfkyf1unqkixWkn3tg9N3CETX7rXr0/Cjvn64J9RDweboeIs1nmfmRWmoafbFaVK5+lJhrfHkV0KLbb6g+aq2TOdoUWn2ZlGvlJjSri3B6rpEhvt5mSTZJqI7OvQepa6Q9rbxaUkDFfGTXAG2m9dzoYTVfZccPrV+wJgGZKxg6WZ2C6tuatG+qbxIwuvPw8GZdsqfnzBG6YQpVtXMalWs+n3F17AQdyYPyjtFFBoSOoS94EOUf+0nN6G5UhlEvgkQ+Wo281YHReOH0dHbsG5or3epXg7adaTl+ntWiQh3xFmgTMnUyZRI9hktTz9XtWozIOKojczpVWw74OifkEp3//DUa4OlembaNwVmzm7FN+IjWI6evB2O2QHdW/rcZHTd43vzm5iJd6VZO+fpSJE9f5xhuJHIgQVMWiT4ZrtG2IRwPOJqctO1RW+8DANx3x1FB3kVzXmLedXTP3LoT4fv6sKFtqxz8dYODvoYetrSxdaD6NLO/lZsDaPsM1zdzGNF9HD3M120Hys4FxdB8/1K5N7uT0am4rBQC/YHqSZweofu6LV8RGcvoPBpg91nqdU07Haii+wY+rE5B657QcZ08ST/SdUtMXpoXsf0Hv07BGsXsnZE65+pny4rsc1nd8bBWltjlKbOGCfi9kM1Mmt4BhWGz1uvvq0byxcuHbXfTdpzeR+rhr7uyez5N2eweoiEtN3eD18cFKwNVj7MQFaxmrNgB3fHHH48TTjgBt99+Oz784Q87XnS33XYbPvjBD+KII47AhRdeCAD49re/DQB44hOfiJmZmSC/008/HQBw6623LjvtnODKHhpxC3IPqU0g9nuyqcmmoRtjzNit5sMdV62A/u6RMlf3NatkKbooYlAtmFmjLZTSqbgNTEIzB65tLB3+JOR/S9PUsIeSwUGeR19OAaaLT65Pc5v3OaWcTcssaGle9CDWLm6Jgm/6xm+z5EKvQO6neDa6oUuUef87RyHSzDOEba7Jgjwoy+STULKcTceCOqfGa29BYTg7/oSZU3bMxkTd19H2M89s+A5mfUE3Uh3FN7aA8sYc7ZPefGGdvcVV9NChoTlFE128V0vNZ1Xb56ODFzchVUbVMGzr2CFEkEeEJgq7KUYPIQYIN7t6Ol0msyANaCKL34A+kr6ecQ0GgnxiMjizGWTyjMor5oDV8Iuzro6M966o5ivXuMPOOyrIP3bgUDwXE3mbHLu+IUYmjEhgzBPhA9o3dEGfQ8xIyOQVfB/rezL2wkTeb26zqdDggCK1cZGkQbmPHOOpgSIHPiR5bn7rSn/Jd0au7qvCDXlCE6V/4cjRB45uSDZ68cO1o39NWO4CehSRV6Vyjx6MODpHTH8JMog8NmnIeI3pNFyerJ5GNw25sUXGaCADc3KZ8Hv0swzZziaYp7PGdCar25dMx8aIgtmkUQNebreJ+eeUPm5T1gmdxm30Epjx2H9wtFyt+9r5xtSV29AumcunBkancQ5USDtkD1bM84TuBvB9E8xrzXu7KciF74+NN2Zs0HbmDo4Mhmtq9jDPLYChOXeQa3kz07G5vlfovOYO1h/cWjO3rsu1ueFdKncjOkWO7iKdxKPFkbHNu6zhHq1Tk5eZiyi/VQutTpYySAtoI9/0druHTV3GtzMXJjAKC12e8f3/efToBxMqURN5xR5i0Hkpo3+xtJL03Pzc290rb6PEd2aeAHh9mcracTfrKZ39Xc0+FDWspDzptelwfe0Yk5Qam7LG7WTNSPk4l2cw3qhcXuQUarDr89j+D1cu3ccr1c+iTgIpMHOubzzv0zjKP02MXyd/jFoaIwf6rC6UWFfRNb+RJV32b8fBYF1ruDHYUGfXfdbIsvBA24c9QC01gmPKKKkzt8cqWBmMZSgqWNVYsQO6Xq+Hf/iHf8D555+PV7ziFbjiiitw0kkn4d5778VXvvIVnHrqqfj7v/97bN68GQBwzz33AACOOeYYNj/zfPv27UXlb926lX1+9913A5s2dayNQCAQCAQCgUAgEAgEAoFAIBAIBALBMkEO6B5yWLEDOgA466yz8K//+q+48MILccstt+CWW24BAMzOzuK5z30ujj/+ePvt3r17AQDr1q1j81q/fj0AYM+ePVOhLRoGDZNbU8SQyrc0pJqbYeR3BnUm/F9LVGihMbqwNG/JHH2Vsyilz3LlxPLxLXY4Cx6gyKLpoCNWZUMTZ3lEk4/RZsuCMSwFs5ZwGYvbfAGZ9xlreSozir3nxuwOx3sz5+mSsBjNWa8H3/thBMl4onVOWiYzlpfAeFaXUY9KrnzGytdJw3k6MJaaXcD2E6W5dDyWfJbz5DDFD8erS9bKMGcFnLMs71heDiyvAnGPlglEIzevTCPvnJWjU35Jfsulu4xpiZt81iDmSTmtOdmVIXq8vlqOaZXKqMJxwnolEPlvQzMSGUgRt4LuCCpjvfB8UwHNclKdpmTeb77rOl9Omy+48Pq2qIysjYKZF202HUJ9j0NLybg1nlK2zn5+MXluHma8qEqjJXTy1MnoPXS9FP0eyHv7MLRH257LK7eeSOWv0rKhWlLjzUulvKsmG15d1pr8h+EjNVCsPBqnfpxu2rXCE4Ud67JG49qCCy9K2mZcr65xvO0tCrznbN4dZLsJVT7OGNOpbwqgMzIsK8P9vICRt1KqrXJyp0O6GPw5zq+nX57yQmhPpBPT9RL1Bi9dA3GI9VOGJztFhyjEpOEIszIuB06eMXtYDiKe3111f1216XVhPpNGc3GufVgsnxdzyPHEiu5VCg4eyD5VybeCQwMrGjX2Ix/5CM444ww89rGPxb//+79j7969uOOOO/CqV70KV1xxBZ7znOdgYWFhWcretm0b+3fCCScsS3kCgUAgEAgEAoFAIBAIBAKBQCAQCARjQRf+CQ4ZrNgB3Z133olLLrkERx11FK677jqcccYZWL9+PU466SS85z3vwfnnn49bbrkF73vf+wAAGzZsAADs37+fzW/fvn0AgMMOO2xy4qhlYGMZ3yPxo9VQjS7hpJbLdfuXzd5857d+U5YqPA1fs8O9PDdIl8lLM1a4wOjSWXPxrF8nVQMzx+zDzDGj9vatiliPFb/cHlovNZ9+5iJdpYGZvQozexWbJkzglt3fp5z2qdfWqNfWwXdReO1o2mR2Z4XZnVU0L5rG1FkZi176B0D39eivwEqLra9B1fwpQM9q6Fnt8mmkzWy5tB7Nb7+vislKjIl6TX6gOOOkCumuFpW1KAeQvsvMz1uT9imFb12e4H1nzHSxbiHp/Xbzx7bfTxyqRZWWSx15zdaFyCpKs0GKj/161DOjP64cPaeT9XPyMrwP91lvXqE3r1wFhcmzWlKolpSVQ2qorOWcVrzcyo0NPaNDOcuNR+XK3QDG6i9j5Wfbw2sLn//qNXU4Bpk+5WiOobeg7J17NhlpR593AKBaUKj8NNxc5tePocnOyw38uwdtekZRrQaq+zgdY7xz/OM8y1noHrUY0uDNT3Ts5eqUej+6V4/oOj5tHbTHoO8JzdViJH+09UlnXkAAHQ9e/1M+qWe1rXN9xBLqI5ayWfv8zbVnip9zPKceOR/mE2l7K+uAtk2rNl29foh6/dB9z7RfbJwZuZHVcxUsI9p61vHy/Po5+nak7exvRh+P0dTfp9DfF9EjKzh8Yp8ftTD6I6gGysoMp85+kUQHivYz6ScOps9iuiNXfn34APXhg3ibkPRWH44gOS8V5O/3X67NfHm4ZsserNmyx52ziE4YzGVEtmR110k3LPz+MHUmfZ76Y/MCnDFjxlqn+YnTbxqe5fJSQ8XKyNi60niDpGRY3ffSlJJeWE/avsFzps2Sc53Hn6XfBahKOygOR16aZ8Pmr2P2/f0K/f1UUUH8twp1Jp+mONEkv+a3MxelEPvE8GPtflDP1ajn6rHWVTl5ZeXGjGb1ffs5mS+UBtTRC1BHLzjrb9sOC4pfmxA9hJMLvj4LwM0nM5ea9/VcKLuTOl8N1otJb15qaWT2aRyamLkstlaMPeuqu/ryoHQ/jlujUX2V8nGW57x+dnRLb46CavqY2Q+kstro8/7aydRRz2noOe3SldOzGN6K1o3Jy+xXOWB43Nd/km1n5hy/z8w4o/tJzNihdNI9Sg5GnhpZ4tfXpzV59yazDvVB9Vlosh8Y41FT54ZOp55c+T4NmbkZcPdFubo4NCs+T0dPmXzqE4wJhbB/on8rTaygGCsW4vIf//EfsbS0hOc///n28I3iZS97Ga677jp8+ctfxm//9m/j2GOPBQD86Ec/YvMzz7ds2TI5cUT4m3+Ha7RVWowg0lSAE7fpbNgQ82/BZacGZrFDDyEWjmgLyoXrMIqmc4gRSTN37+jjwXrdXkpv6lwBg3tG4UTB1Hk04TSKRU8H4SiUBrSV7ipwxVbMAker0SWrPsxEGFtUGAw2aIe+6oCvRbhtbzaMaYgF7uBxcdMoUW9/1dYjEmLB5t9r+8B5bxbOw7btWH7i8lTtN+0F3Ar9vSNinHCLZrHDXO7rF0XbbBxwF9gbVIvkcJnW2SShygKnzGsSrouWZ77leFu173Wv7Wcn9F0HJcOne8Tbo3e9AwqD9drJn743bV73gWoAlw6k2w4AZneN2m9xY215h72weUajXtPQRw5CbJ8P3ZA8rJzg6mwLaGmkdablBCFkvZArnGJn22lRsfVjaTPtWLX1qAYhn4w+bt4vtfUfNpuQ1YHKpnE3sZhsOHlK0szsHvXTYC1ZQNUIw/RoRi4RaAVUQ/f/DipAk/lJe2PG4a1ZjWqBi3/L1IUBx8dagQ+bUpEx4M8Vug0NRPO0eddw6Ajmktqrl8cf/f1VK/vaqcYufGn40brPaK5m0df8pnWldBja7IK6astsaWseeTwU8JS/5vVoVQ/MuukTGx264vUCp8/89948Z+rihKE11RvyfMCBDbdsnkUMhXL6FC2TC5PryBU651L+ApzQSNVCO/Z798+EtDNzBN3cickDv/6+MQOXxtJ5z9rAkMRve/N76bBGhi2pVrT02g9m7x0pkoN1vH7B8suwpd/M27oPVn916mnGgdE5fN0uUX49o8N5gbR9PaPtvGDmz2HBYcxgA5m3vDIRkzf3zbUk2Hm7+UFDgTF8Ws+QvmbmPId3Ff++t3vUZ7rXGhRRHdK2E81/bz9KE+Dpu/OVLdMP7+7oy0x/1XO1TR/L35d9lGYuJKAvEw/8eLQ+nNs5KmeJ6PMO/WQ9wMmbog03RHQWZgNs9CJMP+rTcI6xm5c1r/P68wmAtr8XImGxSJ6WNsLHNM1wfW3z8usxkhFMXUjfKU+3dtYdnHyfq1HtqwI6WBlHw/AanWAx0UkYtaHNl+qwhpYK7Rqd6VPlt7fVz5q1WKWTfKB7JI1unylurjHP6CZ2ZF1lyqgGdKCE5ZdguMYbW1Rn0u5zWwBdDyXodGg2ImLY5mvmIiOngVGdrC6TW79z5Sqgt88o9+V6A12ucHpokFOtWp5iNtwp79d9YOb7awCQfQqSZLChdtogILdu5lPAvaKCW89RQq0eplodxJcDGLVX1zDQnAxUD860z/q8vKBzHb8ecEh34H/faw6WzZqS6hfcmnG4eYD+vY3ONsvPAZZOOr8285tzHQFZf3N8zOm7Srey1ewdVZ5uGcwblcZwbbuPZMrlwjlzMlT3AJCDvOwepCm2kUdUZ4qGjrW0Nv/Wo7UVMNpbavcSQjodGUnJT+11rtXuHDVsnwPu/DV64S1C4O1R+mUPlZU5RpY4Y43yWYzORDtzY21p0xCVabNNA1R7Gp2OWa9oBbuXYfchBort/2qp1SO4/aOU0fpgA1kDmENPAtq3MYNrZ+9MsLLoJuIFhwBWzIPOHKgdfvjh7Hvz/MEHHwQAnHbaaQCA73znO1haCq2azf11T37yk6dOq0AgEAgEAoFAIBAIBAKBQCAQCAQCwUqh1INuteGrX/0qXvCCF2Dz5s3YsGEDzjjjDHzgAx8YO79rr70Wz3rWs7Bx40Zs3LgR55xzDq6//nr22+3bt+Nd73oXnv/85+NRj3oUZmZmcNRRR+H5z38+PvWpT0XLOO6446CUiv5997vfHZt+ihXzoHvUox4FALj55pvZ9//xH/8BYNQQAPC4xz0OT3jCE3Dbbbfh+uuvx4tf/GLn+6uvvhoAcMEFFywLvWrQWjvUnMVX7lJdwLH6ZF9nrDqD73MeJcRajvPSqeZ5q8yljaE1Fme9EfOO0ZVrjULTawX05kOLJbYcL18fbD8AcQsgEMso/zmxHrIeNxErWdaqMmdBkrDkdD7rk7JZMzOGDmopSKy8BsbiiHre9HWQJmuVOa5QT6WLWAoms1P8bwPqVZC07gOgayaPjvX06ab5Ua9N7r31fhhG2j9imW2wcOSoU9VS/lL6al4FzzjLY0NPCv6YM2EagHyd24fe7xT/6TR/OHWiIUKWyDMu/4ZmWh9jsa172mmHnHcQpdWHtail456zJo+kN6gGmfJjsprLa0ElrUupdXD0Pfeb48NIWEGTNtm2mfGZq6fjOUz4zMp3f74ulbcRWF7i5HKH8LtsnvaB+1/WqjHjfZush+b7hPUa7GI1meOnTNtH0zXfsR4cMfr8+nk8bPuqUC5TGa5j45njXW6uojxjdCdq2Z2yYkU7/3PPADcSgUp55NL8qZVvIo3TB7RPjD6oPX2wUMbQPA2qpVZ3Hc51mLhznyb6jKUro/PH5HZOl3GeUU+REprA6/vRNKRP/OgOfp/66B2o8vQzVugx/cMHjchB1yWcVz3nRRKlKYGS77JzAeMB0UnnbV5zuptbEJM0MrfTvEp1TqfvGA+LlB7R20d4IzaumLkm5zlnSZvVyfEX1a1N0RFZXextxJXtrytM/Sr3m6LsZ0N9uiv8/imSB8y6N6tTUn27Aec15ngpZcYrjQJCvZRLx47ziM71BeuyagDAeKfEeNyM0SUmwg+dx5fSjacL9o5S61rd48dBSsbm9lnYjdzM3Ojky80JJXtkBL63eyw6kUHvgb71vi3WHRHZR6KeiISPc/LED5XpeJAxew5Ktx58fmSZAIr36HLm+kJ5MeSiyuTSkjah6XPrCfZxIo2/L2nHZGQuZOVBqp+It3h07yS35u8ol3t7K8u/vZ191iOa84AzV7gEMsiITk7PAFidKwCzJg8+ScgdwSqDPvQ66ZprrsHLX/5y1HWNZz7zmTjqqKPwxS9+EZdccgluvfVWvO1tb+uU35VXXok3vvGN6Pf7OPfcczE3N4fPfe5zOP/88/Gud70Lr3/9653vX/GKV+CrX/0q5ubm8LSnPQ2PetSj8L3vfQ+f/exn8dnPfhZvfOMb8fa3vz1a3iWXXMI+jzmedYXSemV69ZZbbsFTn/pUAMDf/M3f4Ld/+7ftu5tuugnnnnsu9u3bh89//vM499xzAQDvfe978ZrXvAYnnXQS/u3f/g2PeMQjAAAf//jHcdFFF+HEE0/Ebbfdhn5//HPHrVu34s7778fP/dEfdErHHValvgXKv3/YIjc5jDl5ZNt/mco9GEiFC5x2OalFz3iZTpZPlzE4MXK0TqtNYtl3kCFTlTcp3l/mOmcxxXF50GR0F5rHqF8sFMtyyQSuLIPkIRy3xtJ8XpyxSBd+lPl3dYALSzlWPuMeGrKZ4eDLsIMwV6TaZRzdFeje1gdzfmZDdHUsv0hOTLHvaHi8ZHmYXplO1gV8AsA5XPbDDI8etu+7bMgWlWtAw0UebHlO+3yZ1wKxuuXqXNomTlitZWzHGG+ldIbc+ozlvci3yz5eIod94+YLrC79ZEVoWsXr7AApPdYP/ey/1wjCCx7UtSyD7HjNHSZ1KYuEzi3W7bl8YuuFWOjGMTCWvJqwnIM19jqVc5DG5kqPg0kxcd8dSjIQwI/fcjlOOvJIbNu2baVJedhg69at+P6P7sdp55WdWXz7+svxuGNWvo8eeOABPO5xj8Pu3btxzTXX4CUveQkA4Gc/+xnOPvts3HXXXbjhhhtwzjnnFOV3++23Y+vWrej3+7jhhhtw5plnAgDuuOMOPP3pT8euXbtw22234cQTT7RpfuVXfgVnn302LrnkEhx22GH2uXECGwwG+OxnP4vnPe95TlnHHXcctm/fjuU+PluxEJenn346fv/3fx8A8LrXvQ5PfOIT8bKXvQxnn302zjrrLOzbtw+/+Zu/aQ/nAODVr341LrzwQtx555045ZRT8NKXvhTPfvazcfHFF2Pt2rX44Ac/ONHhnEAgEAgEAoFAIBAIBAKBQCAQCAQCwWqDGpb9rRa8973vxe7du/GiF73IHs4BwCMf+Ui89a1vBQBcccUVxfldddVVGA6HeO1rX2sP5wDg5JNPxp/+6Z9iMBjgqquuctL84z/+I17/+tc7h3MAcN555+HVr341AOAjH/lI57pNCyt2QAcAl19+OT7+8Y/jec97Hn7605/iE5/4BP7rv/4Lz3rWs/DhD38Y73nPe5zvq6rCxz72MVxxxRV4zGMeg+uuuw7/+Z//iYsuugg333wzfumXfungEE4tRBt0sZCIhhjpSsaULPXSHyHpLp7Lc6K4tzkXbPKeljNxzN1MuaOwO+XZURq6pBsnf8Nby2V1ZNo0GsqrOCPm2YT87Fu1LWvMZSZ/p6wOVn9jFV+Ve57k+KFTO6XGRsyLKYNJ+iloc0b+FJfZUdZNA51kSU4ecklM33uhNJZjbHB8Nok8KkrTlR+xvPIxhtUa/53DctLqzFVEhqXKzNHSlceSZdU4+HKgo0dsSd/4OkEq3Ti6a27uWW5P/hKemDTPJE8ZHonRMQYPFS2ex5gDsuVyOh1XNG0PjTgtGiPPuSnpumw/TDH/zsjoHNMA7ROuX3LyrlQe0u+Wc15MhVbsXGbT5r4MYmVcbi03rXVCajx0zWoF9JMclntNyReK6Y+tjjpCii+c9x6tTroCueqnX+n+j45Xo7Mx7TLpWjYos3Ds2s8ja4+u+fjv2DwzzybFwZLLY6NwbE4qV5drrXqwwPUdJ1uWSwZmZVjHvUzBKoYu/FslMPfCXXzxxcG78847D2vWrMEXvvAFzM/PT5yfeXbttdcW03faaacBAP77v/+7OM20seLuZhdeeCEuvPDC4u97vR4uu+wyXHbZZctIlUAgEAgEAoFAIBAIBAKBQCAQCAQCwerAoXaI/e1vfxvAKJqij9nZWTzxiU/EzTffjDvuuANPfvKTk3nt3LkT99xzDwDgKU95SvD+sY99LI466ihs374du3fvxsaNG7P0fe973wMAPOpRj4p+c/nll+Puu+/G3Nwctm7digsvvBBHH310Nu9SrKgH3aGEnKfG8hOA0Gtvwt5L3V9g69vREni5PbdizziLo7Etj8axfk5YolAaJu2zXP6U9uW27OoEv03HGD9dJ6CsBfyE8PPn4tIvF9RAsZewd85HA2qooIZleU3i6TIuSsdWF5i+o/nScg6WxfBBs47saJ06dhnTqEskn2l6nid5NVL+snqDF+JQU8JLkerb0nuUplHW6AOsKktDiuIx4Fnwdx07WeveZfBg64LlyH+qeRoe8nWzCXGwPJ9pecuJLlE8VsSS/mB70zZg+2SFaLHFZ9ZgU0Opd4anE07kjbfaPFSWC1OeS6eBcdve8t4yzQXFXqqxb6rR32rlrbHGbqK+aqhaWR3zypmiDJvWemAS3XOaGNuTaUptulz7RdPIY5pzYa6dD4YMTO0djYNUdIykDCuNxCQnBasbGoDWhX8rTSywe/du7Nq1CwBwzDHHsN+Y59u3b8/mZw7njjjiCKxfv37i/Hbu3IkPfOADAIAXvehF0e/e9KY34T3veQ/e+c534rd+67dw3HHH4X3ve182/1KsuAfdqoQR/hpWMA1na1Tzo/+wF8UOFVDp4DmHqtlY1z3Nf2ueZQaSGsIJEWXSFB8CmTp6cGjiJjNKH0OrbR+qnHLvmc0LWif6jt0jyk1qpGzuIC8KejhJ/x8pV9WA7vHf0m+A5rtEeIjSzUjj+u+noQcMnZDguWpJoZ7JMKPpM3rgQenzkvcPKAzWpnzvw7p0qZOzcNIuPeTRZEi0me6hKATTuIpZPTsqNBsWKzLGbfkVoGnDMP3ofJ9S9kidbd9nyAOAupmFYnXJjS0OuXFi82ZkpdIA6lZGTw2Gj+kz2t6Fcr+oKHPgauYkb1zWM+a7kD5HVpsxWDCnzO4cfbR4eFtYtdS0Y7+d68yzeoYoi5RPc/MemV/8gwia3gnbRvKPzrkmXUSGTnLZPJ2nk2O+Qjv2qrZ/TJq6n++zSfnI0lozbdArkDk+SJ1ybT8OSmRotdjw3Kx2+QdePTXQa74dzhUSRXknwmPsXJ155tBpviF6BNsPqk1jNyKIzlMsFyObB6ZtegeUM85LUA0U6n7bpkauF81hgKNv0j6z/TWrnTz9+pXwSbEeRqqe+tbnrUkw2gA1jBbS4sjozDzu0xikb6CVJ6+DxPn8DS3KzAVkTo2FS+Pg8CTpZ0fGI5Ef/a6QdcfWpznQ+SOzxmCRW0MxKP0u+i3ho5ycCN7NaNvnXUD5MEs/Iw9i4PJi9ZrCPuHKU3U+zy59Ystq5GZnozxfJ8q0kxN2usM8T+tk5cVsuAiItg+bKSWML8t/l82S1N+Re55up3u61aG5/L31qdnTMfObJjJmZq/CYJ1uy+KyS+yJjDPu/byLx25KnwSc9Z2fvu7r5D6HyuwN+3ma9o+uwVL6PvM+hnrDiNGrvb32IbcuK8grIJHwu8P7XH+QNov2M1NnIxPqGV08dnsLaR23WDelsrTRD8aR+X7Z7N4gnYs5smN9Bvd7Xbl7QTZP5vdyoO4D1aClZfQj8u2aUaWr+ap4DeGgVCfSjJ5H27keT/YIDh66HCzffffd2Lp1K/tu27ZtU6Iojr1799rf69atY78xB2179uwpzi+WV9f8Xvva1+K+++7D0572NDbC4wtf+EI8+9nPxlOf+lQcffTR+N73vof3ve99uOqqq/Abv/EbOPLII5MHe6WQAzqBQCAQCAQCgUAgEAgEAoFAIBAIBILVjCnakpfgwgsvxG233dYpzQc+8AGcccYZy0TRdPDXf/3X+OhHP4rNmzfjQx/6EJQKT6Tf+c53Ov/funUrrrjiCpxyyin4zd/8TfzhH/6hHNAtG3RraaUaS4e53T0sHTYaAYd9b2T2sPcYjbqxQFEaGBoLEmPlRi2rh8pamPSaOw8H1BNTEWt5xsqqXtdY+ezrtVbS1OJm2FrE+GHbgNYqQg2UTX/mL30XN914ivOeektVi6qtH7Em00csjZ49MBNa9AwVhkctjup5/4y1LBvOtlYZ9ZGj99WOWeupYj09IlZEtnylgzKpRYxWrSWP9XgcEms+1dLfu3e2bTNjcaSA3rxrXaRnNfq7R3kN12hbp2qUDQZrdWuRNNtKSdfTo0mz0NJH+8S3cFdDhcFhI6J6+yvLE9bwVqNtu5621kPWQr0GZhqaBxs0KkPfhhFDqAVX6Jj6141Xm1powx4O19WBVRS14FdDhdn7R+8XjmwYrgK0IZZYvpsKDNYRPiXWxKae1ZIKLDCrJWKBTy01zTPqfaJarwlTzmBDbccmtcozz6jlFmt1SMqkVmTWm21J2XxnH6xaDwNq5etbas4QS00ybq3HxxwxO6za57oOLTJ9+n06zf/pezVQttx6VocecJ61nGMhaerU/J7dWWFx4+gDy9tDFSoOngWctSajlmHEutWMrcG6xppsqR0vrCcBeVYNFNb+dPRg32PNB0Dd5NXf2Wst/4Zt3w43NzJiZ996+Nm27xFLUeLRxFnDOR4dRBZzltA2/74O+aQH9Pd7lsm0TCq3ZrXlQ+t9WLVl1n3dlkX5x4znuh1nNH/Kk7bOpn4DYP7Ro46s9reZ9g+M0sw/aoj+rp5Nb9L0mvdLG4kMNrKO0Kw0aQvjcbmkHK91a2XejNHh4QNUu/ttOxoZOTfqiJmdPSvj+/sUFjcPnfJpPaDovN62DZUBzrwPOOPZ8JhWsLIYAPQjFkZJ7p0bPRi4fGT62qTRs61XgiYedq1HrbL9BBV62qoh8YbWjJUyoU/P6MBqsneglct0DjDjqbenZ/vEYDhXo1ogbUk8BE3Zjgz0dQrf8tjXeTYN0Nvdc+g06cx3dt5bauXd7K52TrfztlYYrB99oBs+qfb1HI9UyvOmHainKODxK/VKIHKNW09xlrzHPem/cc83fm7UPA2dvQOtbka9Mo66ZfTv/U9WrXV0T0M1PHHqL/wAALDtG8e5MsawzIDUk/SPqZ8hev6RA6tfGfhzpuHD4WFNO+5XGK5pdR4z9qk8czwADc8O2rY1dAzX1Db/pU2NbnygtTIO9I/mWTVPxobRx+38oltVZNjSUh/WyLXdfZv/0MxF85XricF0qrWCXqhCbzXd8kxvf+W2haHlkSMZoe+da2WHL5/hjpPjn/RjAMBddz46kGO0TXRfQ29qFNmdM62eOmiTDNd6esySsu0wu7PC4ibXHc3Xm/r7mn7aQGSRYTOiM1lPR+p5VQH12oYRmjFYzVetjNHK6Z9R/p7SY+jivBfob7LusLxvxu3AbWvjtWnS1LOajQRg5EY9qzGzd/S7N68wf3Sr25v3VBZT71eH9gZmfnZCkzO8N2xkWW8fETgkidVtBm46I5tN3w3X8HLP0FktKStDbdNTHYfqfJRk6vHRPF/349HD/cfUUPNk3tNumuG6utUpZnVQ/3rDEP0H+va91YXIvGDnb+LZy60RbLMstvOf0q3+Y3WjGe2uSzyPlpFnFVm3eVB0/qtb/caMYd1r6fLnGA6+bjpcW7ty2/BXU/9qUVkZ4+iZ1IPPkyGGrlGiVr87+ubRo/vIFTPDI5dQPWjCNzRJ5lyafE+OauDuSWzaej8AYOd/HdnSWRi2zomsYl+284LV03pAvdbMW5WjS5j2qM1ctqRw5yv/HwDg+E/8ln1msHT8PNTPRvqdI/+ZdSHVq/uNvBis16GeFlRw9I/De82zI05+AA/cuTko318vKd3q4zS6jaFt9tH7sHhf46GwSAdxU/aiwmDjqAIzu3vOXogpR5O1ulp0O2K4rsbs/SNGGqzTGB4+EjS9Pa1nG/UWC8b7Gj2Sc039LU8bfZbuLRH5esw/j9L85Kw2r96+th2O/sbo2Y6nhPJSKyIDPN0dGO2nmL2W4ZFL6O0Y7Tn5egQA9PdUVj8yY6iaVzb/3rwi81Krr//pC68BAPzV9S9px+4aQwBpIELb0lHNnLmn50TfAkbrKxpxJZiDati+qXb3Hb3AR0X0PKMnVANPTydjCmg8Mcm8ZnU+qs8SfZTKLqCRkWaP0sgaStOisu2z+cQHAAAP3LG51XHR6jx2/Rnx/I15tVldhokS0T9A5kq6RqQydC2ZtwFAuXnZMdu0k4I7h/kevdx+VkXGn55p97Lts4r0gwLqytM5NFlj1SoumwTLDgVW5Y1+e8IJJ0zsKff9738ft99+e6c0+/fvBwBs2LDBecbdCbdv3z4AwGGHHZbN1+Rn8udQkt8HP/hB/PEf/zHWr1+P66+/Hscff3y2bIr/8T/+B/7sz/4Mt99+O37wgx/guOOO65Teh9Jay6gi2Lp1K+68/3783B/9wUqT4iA3EaxKqFAJzCYhdcq5VE8cviX6cfNvZmRMkz6BoBMKeXTqxS4DT08rz0NlvMUOf1dDbPCHElb9/BjDCo1tW/yUx1G0HzrMswezHw+2jFvtfLra6SvGuDJ2SuMx144HtZ1pnXL1Y95Hwwz776fB75O0/wrNqyuui2Ta7GDQFwsx7bznQnytAtpXA7qshVN5jJNO0A3Jdh5HBq3y9UCnUKXZzJp/GfkAHEQZtdxY5X3aBQdNrjyE2sygE8+t8FqwK378lstx0pFHHpQwiYIRtm7dih/88H6c/pzfK/r+ln+5Asc9duX7aNOmTdi1axe2bduGU089NXj/i7/4i7j55pvx7W9/G09+8pOTee3cuRNHHHEEgFG4S+4euqOPPho7duzArl272APB6667DhdeeCGqqsK1116L5z3veWPV6+lPfzpuvPFGfO1rX8OZZ545Vh4G05piBQKBQCAQCAQCgUAgEAgEAoFAIBAIBMsApcv+VgtOO+00AMAtt9wSvFtaWsJ3vvMdrFmzBieffHI2r02bNuHYY48FAHzzm98M3v/whz/Ejh07sGXLFvZw7l//9V/x0pe+FFprfPjDHx77cA4AHnzwQQBgDwm7Qg7oCmFCdQEYtZrXctVAFQ8APadHfz3vea8JGzJUbugSYBTuonafUZdsjiZKi6pdd2pDM/3Whr3j6tGEhqgGCvXa2rpg+98q3X4LNCEAYh5mQ2XpNvnUJOiqSdtbVKiWlFtfwAkd5eTr1bXyQq3oGW3Dd1n6Cc3md6odoEYhJGd2V6iPWGLb39CvNJw2c8pqUK+pR6GQmHyiIN9ygliTMA85pIQ45dPYN36b+7wXlPeo+WJ6uPYCmnAsxE2fK2f0YvSnapIv4T02SSZkivmG/vn0+3XpBKbO0TzIt8F49OXIGKB5xsZckr7MdykZYfuoIm0dKcfJp2kTOs59GgI6mvHijAXCM0Y+O2m8/5fAD8OnathQHZqEcHXCd5YoWESWmjkmmiYlZ0i5tM7ZOYKBocORp0SG2jHM0cLwf6pMf47w+dQfr/44sX0RkTeldU59p72wP+y3evQ3CtXm0hKV59z8VRfQzPAZHUcx2VYynnJlm7yrRcWWY5Cz1M7Ws+LHroGdl71vjZ4wKgQsTwCwodKmvvhJlBl9zqSPzUWsTGHKdMJuGt21UK9w8mHCN8ZQMSGAKZx5jZMnXJp6lG9Fw7QBrSyM6Hz1piXUJgwkHVs0HF2MTjp/TDoV03FAdexm3qBzVxZGxuhRKMB6wygGEzuOHzk/+qN6gNFthwihPBni5alJf7HwdeN1Q9Trhm3fwR3PwbwdkSPmPZ0jsm3F0Un4JDYncnpSl02TrOxMyFujT1jvSEM/mfN9ncuf9wL6SJ2tntKFn2PTFpH7tryBcsLa+rKm5b12fgx0toSe6OghE6BkjcInTH/LyYu6766Ni2lcUk7ouYnbJDYHUn4sHNucDPX1Q04/ySE1dhw5lNAJqiVl5xJufvTHrh3v3BgaKLY+Kf5x1qqZ+XkcPi6aI3zdk6ylUmvAFB+xe1tGVnlgeT43zmj7JnRKW3RF1igk/bT0OSq76vU09vHor7eonDazfc61U5Wmy3me2tvw+a2wzqXzW5SmcaD5cRqbdw3Mfltq/RLQZtok085OmQw/ptaQQLsnWNI2vQNtiFAWZH/EJcIrn8ki0LG9ND26txajlVljCFYAuuPfKsB5550HALj66quDd9dddx3m5+dx7rnnYs2aNRPnZ55dcMEFwbtbbrkFL3zhC7GwsID3vve9uOiii4rr4GPbtm24/fbbsW7dOpxyyilj52MgB3QCgUAgEAgEAoFAIBAIBAKBQCAQCASrFhpqWPa3Wk7ofuM3fgMbN27EJz/5SXz84x+3z++991686U1vAgD83u+FYTtPOeUUnHLKKfjxj3/sPL/00kvR6/Xwt3/7t7jpppvs8zvvvBNvfvOb0e/3cemllzppbr/9djz/+c/H7t27cdVVV+FVr3pVlu5Pf/rT+Jd/+Zfg+a233mq98H7jN34Ds7Oz2bxyGMMm62EEBcvLg/Xa/raXZJPjTe7SWicrYhGnvAvG7Tf2guAwI2uJQ145l/AylkKUPu1b32J06SprAesQ1fxj7pLraXuxuZOvIY+0GYD2CJhYSdqLbnvaPm8vPSVFN/kMZ1trTTVMW9SPLDzdMp0LwGtCXvOjntOOlaG98JumI81vfi9tbC5yfaC9iNa/IxAY0dvb22vrzPCJuSTbuUQ6A3rxuOM5hMbabam1Ns72s6Gf9Ke1iiFp6eW3Jk81DC+IpZdxc3XRP22tInyrRgDuJcXUMpOyvH8JdQ9s2xprXFS6pUVpdszYvHJeYuQC95xlfOwellRa/3L35Hvafr5xYq8dO0V3emS+LUWqfv4YMVaSLI/W7jgKC4LLH4BjJaSWVJYGA2NZH7RDQjZhyNfVtygzVnP+d3UjOlRN+NSUS9OTS91jsBeEz7YXk7N3vABJ3nfypJehU/IKrXbpRfCa4y3jATxk8iR961jtkvRUXlp5YNJ4c0VgLe3IHd22b4T3k3VWhJaUdbzn3ZO6m43SZOfHmGcRw/tFc4j1SCGyjMzZrOwmFuDmtxnD9AJ4tr+o/G6801TOAjMnt2i/MmXp2opqtr17+6v2ObW4rePj0S0/LyzHukuIm/eIPubn6ZdjHtM5gvaJ07dmzNRhGgqrJ0XaI0nTksrPL5afwg9o/rTN6VxoZUCFQMbpCtANYbT9qL5n0lNaq50zJJM2Lz8fDroHx0Mo2e+1AhK8pIakfMLzdN7K3m3KOC9Ue5rMKk/vMt//9xq3TKIb6irkaTpPOpb2lI7UePKeW93ZDGIFVEtt/jZLTwe2+Xo6tZ5tB0d2HHKyi8hFh8+p7mzawPCryvMMbcegLow+EZDqeUQE6ehcat5X4WdOXUhadg5i5JGTR0I0agXroezoaXTdFSkL8OZsMG2m2rFB5QnVQ1K6W+7+RN2DlTG0nC55cs+p97rpB6qH5dJT1GvMItdkjrYdaV0K5XJUN6JjvHAu5+SCJnqUk4zS6dc10o8siIyNtSmAkedmSt/09CRKC+C17VxtvaFoesqvvp4XHTuc3OTGgPNBmC671mPSqCXlyjCvfRTt+xifmPmN5k/lJmkHMy+bNDR/ln6vHc3aipsrWoJg10o0r+i6KbGGcdaqjKys9pJJu/luONdm5MjQPtFTTAST2BqJI7Um7xNrSbpfF60zwvfc/kRsPHNIjldP362Y/VZnWovo9H4a53tGhlDeze7pJNrK0bdN39XuM1u9zL6cv8cVfsA/DvYRmWwcHZtZY9D9Zac9uLVmyXpGsLzIL0NXFTZv3oz3ve99eNnLXoaLL74Y55xzDo488kh84QtfwM6dO3HZZZfhnHPOCdLdfvvtAEZhMCke//jH4/LLL8dll12GZzzjGXjuc5+L2dlZfO5zn8OBAwfwzne+EyeeeKKT5ld+5Vdw33334eijj8Y3vvEN9oDulFNOwR/90R/Z/3/961/H//f//X/YsmULTjvtNKxbtw7f+973cMstt2AwGOCcc87BW97ylskbCHJAlwaVX0vtIUTdTKpqoFpBiPSkE4QTwUh5dpQDslkZpOE27JlNCJ9+bqPY1Ku/T2G4zmgKzfsZDdCQGP4GYd0q/WpRhUqcX55PElmwm4nXyZ857FG6VdbopOFsLpJ9I7OQV02mzkYtycsuVLxNTzNxORuAdCJtnveavls6rA7ypHWCAoabR8Kk2t3nJ09uocwsANhFPOEDq0AvtaGe6CQf49HSQwybD92EYTaVqgHpx5o5wKMHiLQcpq6xAxzjol9TZZa0ieWD2GGf+ZYLwRlR0uyzng54PbsBx4FZnMXgKH7kEJrbDKX5J4svUUapHDJjzoR+4zb00NLJ5c9uDERQHz5A78G+m7UnS7nNLMo7dGPN1snsSRIZZA8MBsqObWp4MWxC4VXzlctniQ0F7vCZLrQc3uQW/DZxZjxqWCOOalG1G82RxbvZMIgdLpi8qDwxeVVLyrZVbBPK1juzWLP5VOGmphq05ccOm+hz02d2jogtDpk2CcIIebTTbzi5y4X35RY1Tj1j3ybGjr/g9g8V/f7IbthXbiXpwU5ONhTrOX2wBgVGlszurKyxCw2Xwh0EdtlUdDY9nc0Fl/eH62pUB5oF/ZxGbQ9xGv1hUbk8QXSNKJiNEZ8m+iypO85oe7Cb2rwE+Hy4sdVlI9UZG2azpBe+p2XW/XZjzD4umR9JWb7+wtXTLYDQwrVTX1sjBk5PM+HlfVpM+Ee6wWZlOZk/2E3N2s3TzpsM/XpG84Z2ptiIrm/1I3q47G0QtZm5ZUK1IV57+6qgf9QQGG5o5r394Q6ZntWh0QFzuOPkSXmP6LGxzbrhEY3u/OCMzY+WYQ8oyVxB+d2np7egXONGrlhTP+ZQTYHZACOgh7KcXAf4+Tc5julaLyJbaFr7OjPe6VrT8EG1v4rOfYAvS71y4LZd7vDArAGcuYbo5Y5OERhzhusKwCuTHuRzsjE1h2XamcKsQeih3ehBWA57SJNBTAbk0ldUjwX49Sl5HjWAZXQOzsCDInZQ6fez7mlnDZ47nObK5PQDNgmRsY4xhv9dD60+vaBYGWXWl9VQhXVS7e+Z3RUG61x9OmesQPOi65nhejJGyXtfj83xhSMjYmUzc4Wju/qHikSPcQ7wmDo5fMYYb0Mj3DOg6wqA3xOjRRUYO0cNRE05nr7OtlnKAIfI7ZromYbfewuN8b8lyKNd8fWk83z0kCtGsPfM8HkQNtz71qGDKdMaKcy0iSgfjCP3oFz9xtJBeY6ZS80ebTXv6gaAq39wjgvZA+8O4IwA2LHhjbfWCMUQwuef3PtB22f1LJGxdAySA0R/PTWzp8Kg0f0cPbLDfpng4OFQDDN60UUX4ctf/jL+6q/+CjfddBMWFxdx6qmn4vWvfz0uueSSzvm98Y1vxIknnojLL78cX/nKVwAAv/ALv4A3velNOP/884PvzX1x9913H/7hH/6BzfNZz3qWc0D3y7/8y/jhD3+I//iP/8BXv/pV7Nq1Cxs3bsTZZ5+NV7ziFfj1X/919Hpj3H3DQA7oBAKBQCAQCAQCgUAgEAgEAoFAIBAIVis0AF14QrfKDvLOOussfOYznyn+XmfqecEFF7B3zXH4wQ9+UFyuwZlnnokzzzyzc7pxIAd0OTRWAbO7FBY2jxhjZvfI/KHut9Yi9bratTDlsmosPYaHjcwafItVznPOoF7bWLkQj7v+PoXBWmK5BzfEVDVoPR04S8elRyyh2t13ylRLCvrIxdHvB2ftmKdef5bemoRoI5YaRgBU5IJqG6ZgOKIbGFkO+d509awOLIbUMBJukgmDRMOatBYhCqqpSD2roY1FdFN33yLEt8hRA9eSxfw2bQ8Qr0pqjUcsVapd/faZ16ZOf9NwCQw/sB4fxCOHhnM485nbAABf++qpti0tnb7V81zIX5Qmn5bAq8Gz8KzpJfIklKlF3V6+zVmH6lkN3fRDRTyZqTVWEAZLo23zAYBe04+m6ZdUG+qs0m25jLVYNMQI+dZYM/aaEFGUdyktNimxnHJ42xtD9H2s/Hpdw8P7emx4mTYEY+v5RMcm59GjFeOFRPLSilrbNc9o+FRNrM/q0KrN1s2zqFWJUDP9+2ZsO1sLMNpPNMwQCVVmvlWVDsIcKQ0ow1PUwt4OImDpiBFR1b6e9dRkw2bUYK0e7Ws63pjwcTaMS1+3nsnMGHS8MDivN8JvVC7GwvQMmwvLTWhdwO1nX4bSUCs05HDAb4CTrl7X9N1Cxc5Bsw+Oyl/aWIeWlj244Wkz1oS+V0Q923rpqrqVA+2cqBxLyjZkb/vMqZMf0ofI3eAy++Z94L3p0R5Y73pW+9TrEgDquRo94+1FrLSpDLVeOkzIVOrRU/d16EldtXOYntFsOG8O1oO8avNsPR2IXGDmssUjarfOCWtw1ouKzj8R3nDk7RpXnvT29toxNVCobAVIOQlL1Zk9FZYO810beDocEJr89lV12w9qsfX4cepJ6Au8CQlvVEsROUDCblGrZ5ueWhmb342MqvZXoRwiaaqldl6D9cZtw6NHrW0NzYuqlWO0nUw66pVPLPyNhx9n4d070OrbNNS48aSollrCaN3M/K5VS0vvQDtGLOlc6MJaud4rnrzUtM0WVdgPcEN99bzw67pyQ0DZEOGM9wcFlXFm3aJ7YVgx3UNr9U/mL1POzM7KegBQS/xkmDrfuJR4INj3xDNZ7R4xUt3MWb29PfT3jmgabKgx3Dh6rho6KxIusZon3uSmv2kbe5bttn6W6PaZUw+qZzb5mDTVoF0nGQ8mNWj5WfcAPec2UCz0MdUPWK97hr56jug/hDesfjTHhE6uCZ/3NB/NhdNnGbnrRApIeFk53tDUs5p69c0RPvb0g+GGIfqEZp+/Rm3Qjr1cCDcDPzIHhe8R6c9RvXnFr0tJ+lJPg3pt653aRgNxE+c8ppzwgPDamZk3qKcFF/2hWlRhhBkFd81sZbQhktSJetM3tPXmqzDcGoCZPaMfi0fUrJcb5yHleFpyHkfUo4dZq9voQnsV+o1cnD+qDr5T9UinNfWw80XToDQyx+IjBtb72hl3Hi8F+Rs9babdJ6FjlNazJKQdRT2j2zWFtzYw9Pkhh+u5ur2KQ4c8N4rigZY+v88qIo/6CPYEaLST0fq++d3wC93P4nS+3oLCwKwVS8JpU8JHP8JHJKx0NHSweUSiQ7X7YW2bcfuDw7Xa7icuHjm0c227bvLGuxeSXmmyVqP6PO1TRkbQtrXpGd3SSU/ec96nNgS1Jxd9mn2afPpohBY1JOqdGQ/e1T62rUiUCbs0iK1bmP5zIn5E1mv+t7H1td3naYif3V1hsYkSoobEs3mOJmzzMnsea34y6qiljTXr8TzcNBpQak+fDZfpzoUerUT55Obpxc3Ddi6gujeh07bZDNjyBQcPh6IHnSANOaDLoWH6hc2tNjFsFl9UYOYO54B2gqoimxQpxd1fwFE6AG8h1aCeIUocs5Cq9vDdX+1oLzf0JxrnTqF+KJQdpaxWwd15uhopJeaZTzc9YKQ0sHeacAdczEKMLhqqJQW9swmZRwUaOVCwG3gkdA4Xs9sWGbsXyKPX0MQqDdZ9v6UjFbYj4BV/8gTw1RtPHT2ifGrupvJ153n+uaU5UUWtECjbJp3zLyXX2Wxj3lM+48IIMc/oXYLOAZvhg752SeE2y1JKvdf37cZWq0yyMGOIKTsIP5aaZMm7mQdGPDykh27ct5VGnQhj4G+kDb3NLH9DIQwz5C5K7LcF9zLRfPwyDeghC4tIOCN7MDYMN1qcDTayGKFKebXPNBq/UeNmGNJtwI03J4QGkY+cjLffMSFVnfcV3BCZ/reEt3QPqBbCAUR5g5PBzsYOl9YLWQOQeTEiVxY3hYPfoYPyUSTEWwzu4rH9v/LmNFuWIbVQ0eWMa6LfRLo2SOcvbv2wvguuUU8gQxXhuYhKwm2GUVlpeTIWpogBG2qMypDUAZrmx35uI9PKoFi+lBYamsk/6CdhPqulkNbAWMMrKzic48CMDSR0Cq0wuivLR+Lgw9IKb4xG5Lc5XNJEBrIbPDS8Hzm89e8P9g1crGGN2aBR/GGRW5nmH7rZxG0acvNLhl+pTuUeVLY0c4zk0GxoMYctGX7VfXcjNTVX0I199zCMzGte6GEnLzqXmY39WHuTRHQz0p/raN9zbWtD5AN8mDoONa/Pc/og3XBX+9qDUjrmzB16bEgzrp18JHjfIDh8Nd/4tDe//XmfHuSiTs/1FNyBLbdR59C6qNi2oHe0BsYwPu97bRLc/wTEQ3kyB4jsGoauqyL1c/jYK7+3v+J1s5iM8NdtDA8CBfoeyc+f3x0ZEymrFHRzPrbGzM6RjOFHMN7o+9gYpoaXDOyGfcbIkL7n7nofJRz9s7Rh9KN3wOtnOi8zSK0Xo/3gPR+s0zYsJfedVnDuSvQN2ugaw64lEOkvykdUTHh7J4DL05y+nguta1Byl3G4R1A5Rg5cHuYATg2ZuZ4e6HNrfrJu0SoMD1j3vYWHV6/hnLdWjITio3nSAySKqIzNrM8Dfd0zgGxlZ5uRmcuqhdboKLYfw4X91Mwc4X9DoYhRqfM+osayOhO3Jo/oQqyRLtmj8+mj/ezkQ9Yljk7rHWJqlZ9fuYO11B5kkD6hK3DyxxzOjQhGcF+6D7NuXjw81imjf8z1Hw5N4OsXZNHzxpPPJ0vEiJ7ZT6PllLSZYJkxho4hWN2QAzqBQCAQCAQCgUAgEAgEAoFAIBAIBIJVDFUa4lJwyEAO6DgY6ydi2U2tDk14m3qWWLRuXkJ1/4yTjW+FY6xreo/dDwAY/midY4Hne7M4Vn1NSBRqkdObb0NcGq+var5qvUPIJaB+mDcAqA8fWAsMailJ3d/9sGTDtbX1vlADYjlHwmJYr7teGK4SICEu1+ngAno/vI2h3XoLOh4zaJ9FrMQtTSZNBTzilPsAAPfefnSbPuPxQ9vEWLb39zTh2Ta1oQk4UKu9uu+GvzB0BmFRqDVMJF96wWsbh8KYl2rUh48K6u3sW0viunL7y9DXusI3eVPL4z7jxaRbGllLVGKZpn3LNwD1YUNr+cx6rRGPF+Ph4Hu4BaFIqcUP+dZa+Sy2FmrUioyzNmQvoNdoLbKJRbENO9Uj1mIxi2IOjHeFH9rOx9CERohYBbeWTYq19qft6Ix3LjwMaVPfApJaQTtjMGKd7EPpNhwUIqEd/TA9uTyBlo/rNdqGwzJtpmq046RqPQgUCQFVmzC4e3otn0c8du1P1pq8HY+xMEPmWRtSVwVjIhZOygFtH+/ib8fTsW7DLJvQOw5UWBfHgr4PJ7SXraoJS0L6cc2WPQCAhe9tbMcxKWfzf46+vf80HXpjaziezVxYzRhPB1XSbZs44d9Im7TzlqmQlwfXvzR/M2bo/OuLzTntWsZ7lqK0TtVAYeiFY9R93YY9Y8abJvRXS+3Yaa1kVXauc8LX+rKLsZ40tAIRC3sFvs/ov4SffCt3NeBD/znRASJ9buhMecDTucz+H+3YoWE7g3o19Fnejlmu+jKSzp++d46hveH9atCOLSrXubDfhk+Ga4hn1JJqrf0JHamQOHR+czy/aEhHP/yS17c23FXTNr2jFqB/smb0jMoYNqJAy6dcOEiQPqW80dvXhIhivA8dD7zK45/mPQ0RbklZ2+ioByo7V5vQvAtHtAOE8xJyvNqYfjbPR/+SuZobW6Se3HjTJLSzobm3v+J513iNqjZMUrWzH+hUuiJrC6rjcnqWkf8D5UTGYGmlc6pPXwU3xKUX2krVwNyOxsL8CI0Nj38QALD7+5ua8t0+NYjqD5x3q03EfK+Inszp4CoMz6aIvlr3tePZ55ftzCWcqI6tFak88cO315FxRMrkPJOcvjPzI1lDOHOe6TOTD1nLRT2HOE8Gcl1DfeRoQKoH2rV1zosn5q3vp6f6rNN/KRDedEL/c9/Sdia8kQVZt9C5GHBlNitv/Kz8JYzvDW6+M7J0rUZ1wDBSk8Zbd9nnpAwnTH3zwu45zPC6L50fnHCLzU8rY4+s23rTcUP4LdsOdL3itwmTtjevbLh/6q1L50e6NzJs1gsm7GW1pFqvw6OWoHbOtPWj/0agFfEwnG29sOkYza2HuHFI37Fjj+rDXp71uhozO1sv5sAbjS4lIrKUtlmg8wyU3edSA9WGCabeZDRai0c/FxbT1NWkD7wCBwrDtW7EJPMcGPGu7YeZcA0SVDGhzw8PH6C3y91yHUUaaGWs5VPzWWbvQA3buZbySSq9rshcveh5STF1y4XRDfR1Tsf3843IY5uGmQ8U0QmcIglP2DQZDzK/nRQJZUojg2W9whh+8K9/AQAakr1eo9Hb60ZfctYwqu2fTdtG3+0+UbN1aj2XFTjP9dwao2bqT+WGrtv3OZ6i5QlWAB0jDAlWPwoDGKXxjW98A295y1vwkpe8BMcccwyUUlAqP1L//u//HmeccQY2bNiAzZs34wUveAG+9rWvJdN89atfxQte8AJs3rwZGzZswBlnnIEPfOAD06iGQCAQCAQCgUAgEAgEAoFAIBAIBALBqoLSgNK68G+lqRWUYioedP/7f/9vfPKTn+yU5g1veAOuuuoqrF27Fs973vMwPz+Pz3/+8/jc5z6Hq6++Gi9+8YuDNNdccw1e/vKXo65rPPOZz8RRRx2FL37xi7jkkktw66234m1ve9s0qsODWoqSi3Dt650zodUETU4sRYc/XNfmaZCxhnLuDGreD9Zre2rea6xbnVj9xEp4SO5ksqBxwokliNq8MPp931xgEdE7UGG4rrXOtZZvhHbrobCoAutopUeXPo9oUYHnGb2E2pI5p/PWcpzFjalrrZx48T+793Cbl607+e3XCZ5ldUXbGqP/+5bHAByrffO+GujQYlwjvJjbs8JheYJaLVorNlL+/p6lmfNktMWT+teNoZ9/Z09gLeflw1qOk0uDffQf7Kfveqjb9I5VXzzFiO+Yy3BN39WzxKOzT+gilkO2fozXH7USdkg13g206FJLoohFq/WS8bxdWGs2Mna5eO7UkrUlkCR3LK9Ck7CU90jUCrqw/jrSpvb9DLlXkLmTgY5d1xOwGW/znswz+XL3OxivRAX0drdjh4uHz1mjc3WmFqF07AQ0KbherZ7lefS+FwLazxW9i5H8O/pAtZfNc5a3xNrOqSd1oMh58zVY+N7G0WfEuwOk7fceQ+SiNxdpoPX2Yqz+c/C91dr5uxmvA0VkAPHULLS+dcqid0XZxvOsmIHRpfMExkrY90YCRv1oPSFN29U8EY7XnxkPszocW0o7dQrmBYWkFbUilqDO2LOe9Pwcw8otajGauisqNk+kxpNDs1cn7/4l99u206zn0GI4/znpK237RSGcCxwwOgfn1aAVbL2XjmwjHZg61zPu/GTKW/+T0Y+9xxL5Te6a2PyE+wEA999+JD/2GcthXbW6hjpgZDFTNb9tTP2aK431f6/xPPTi5ug0ooDzlrEItve2LFatN7Zmsq0w8iYEANXKG3oXUvbOxSbPhSONewR5xbQn9SqILYzteqLmPxgYb+75ys5rnPeko/usGaXRCxVv8UxpNV7Wum1LalVvvBI5me94uJu29XSCwMNee7o3mN/m06rlheoR8wCA+qdrsLC5GWc9jZ33bxglN/pmj3jwUa8Ekz/1gvLqY+lI6IbQCKz6dUXoJ95qdl3Q11YPVTXae4E53UQRbwHGe9JZGzDRC6hOQutuI5esDz2CHS8pAuq1YGlmoqUo6q3NPGPXMBEZ2W9kzHBWQz04E7y3Og3Vc5h5Rfd4j1hL34BXKlhZb8cYyavm+4HjidK7XGleNE1vPuQDzovJR84T0/dmq+apokPmeZOEWRbF+nUdmYvouqq9J7wZ10uK7b+l5t6mymtz9aiRHADxxs7BuZvV0205DA6rMTis+Y5EZnHyJB4lr3raVwEA//Cvz2hpasrp/XQ2uCcs54000ieZ7zkZGVl32no+Yh76p2uYhEz5JlLDjA51gUo7UUjYNb/Jp0bI85q0GTM/O9EhKkCva5SOeSID6J6KJ9dHfMToEpxcszJCOzq18uQmtMvHOXARZmw9yX4b58mvhq0HPbsm5oqvNAabSLQXO9AStKp2TlWLKhjuQbF03mxoM2tltcDvIdqihu0cyEXB8OkCPN7RrX4WRKlo3pPllqWX9fqn9SFjAwD0T9eQOTvkmZJ9R5q/L5dn9yksbhw97O+u2nHk6RkAnH2u3ceny7LrDo9m4wHK3tVM0lZLpE3puhtwdVtfZzK06vbbmHe24CBB2v4hB6X15IFL//qv/xr79u3DL/7iL+IXf/EXcdxxx2FhYQGxrL/whS/guc99Lo488kjceOONOOmkkwAAN954I8455xysW7cO3//+97Fp0yab5oEHHsDjHvc47N69G9dccw1e8pKXAAB+9rOf4eyzz8Zdd92FG264Aeecc85Eddm6dSvuvP9+/Nwf/YHzvCSMwsHGctHEhQnklQMkN9dZJc3PP0kIUzZdfHL58Gsud3OACzdlflL3f+4i3BidmTJLkQz7VJKe2bTjFiWd+Iapn897nfONtdkE0L12wyM6NpiF4HKADfOQK69DmyRDyoHng2ixiQ0JTlGl8NOMI49SvNMlz2K5VZL/GGO3FOzGgM6P+ewYy4XqmlJdxmqzcWTklGWENbxQzEKPfjeOjJwiJi2/U/op8XkyfOsyt+M09aDoRvKkfLiM8oQrxz9k9zeSnQOFLjKS/psLv5SYK6aK0npE9FE/hDGQnwu78DY3NmLfTaSncWXGNoKZkG/R+dc70HY2uCJhO6cKn3djOglt5wZ+KFg27wnG5ErPFTFk1xG5McNttnYcBz4/c3I1Z3SUwjTk/tT6z9RzmGjzVYBl28fgdM8Mj5WudR2jmg7rGjavcdcTyQ/RHmxxG+JjyJiAzg5zNTCddVksb2BkDMmGpCeY2hphihh7/wOYio6cKjfbT1Nqn6LrEpYbK7g+n4aeNcl+16R7M8XlAZP17xj89uO3XI6TjjwS27Ztm6BgQRds3boV23+wA0/7hUuLvr/p5quw5bijpI8OAUxFlfzDP/xD/OVf/iUuuOACPOpRj8p+//a3vx0A8Gd/9mf2cA4AzjzzTLz2ta/Fzp078Xd/93dOmve+973YvXs3XvSiF9nDOQB45CMfibe+9a0AgCuuuGIa1REIBAKBQCAQCAQCgUAgEAgEAoFAIFg90LrsT3DI4KDbeh04cAD/8i//AgC4+OKLg/fm2bXXXus8v/7666NpzjvvPKxZswZf+MIXMD8/P22SAbTh0wCMWi3iGXYw47uWWl+Y38Flrn4oR8D1MMLI2saG5/Q/j3lqNH/GOtcpV2Hkll3zYSPgua+H8TTIzx6cMBDmPdsPJi9Cn3mmq/ZP6ZFLvXWr5+C3RazPKf1c+3FJSJ1sPTxrOJXKk5RpvvXDVUWtyDPW1k4xHXiPRe79GBZ/JgwXEFoQ0tAVqg6t/8aC32ZeO7f/QZklUu4bkmc9o5MhI7Rq/3JyieMJkzb4tuFP7r2TR9M2JfKQHcdMnk4/Mr8dmrjxnst/qKxMtHwSkycEyfe07Gbe0JXbP1oxbceUEesTixpOeC1KGx3DtE7ZOSsmYzhwfE5kbbIsTiZ1lBHsXEOz6+nRX0wGmu8y7yed5zl55JTP9HNRX5n0Xej3+qxLOWyZnnWoVgA4PcAvKzU/pp77Y4eiGW9qqMJyYnKbk0MFbZFts8I5gPIDN3ZLEcw/tgC09W1oCsYMJ9c9nckPA8vVnepV43iVUHnslzUOn0ZpMHqg4scefcbONak66NHaoe7n+9EPT+aXCaDlaU4P9HXSkjlfJcoCGdOR+XVsz4xCfZiWlf2M6PAGjvccp5NkeD+GLB94eQV8qsD3W6ScXP7+t9kxl2tT5r2tK2lHrhx23cHVMdJ+xfNcQftl8+DaaZx8zXioiD7rzzFmjDFzPovYuo+jtZTMCddt0feN7pkFsxbNlTlNj8TAK6207VLjLUK/naPoWrN0H8CXHYUycFqgax/2PRea11vXxNJF197+nO7LSD9PTsZm2ncSeZMqO/qZpw9T2RnMtZnx7lyBQNcQ4+hBHfip5WOij004dnLlO3MIN/YoH40hA7Nr9oI8ct6Qsfy76EzFa7mYvIitd0p419szyOkcghWEbvWK3J+Ewjx0cNAP6G6//XYsLCzg6KOPxjHHHBO8P/300wEAt956q/P829/+tvOeYnZ2Fk984hMxPz+PO+64YxmoFggEAoFAIBAIBAKBQCAQCAQCgUAgWAlolHvQyQndoYKDfkB3zz33AAB7OAcA69evx6ZNm/Dggw9iz549AIDdu3dj165dyXTm+fbt2ycnkloKNL+rJfJ7QaFacD3AdM+1CEtZi3NW0s5vzop5Ro/uRiPPzUWofv7tB+3HxqqPesMNN1Az6OZPA/UajXqNHnFHM56txQqxPukfUIH3i3OPB+dpo+F4MqS80Wx9qvbPsTykVnvGMql262rrS/Ifbl7CcPMS6hmgngH0nHY8g2Z3Vpjd2Q4N3YPrNWnapPmrlhRrFVkN1OiuMNWmcUYctapsrJNM2wCEn7zLWumdIKpu20XVsHWqZzTq2dGfntO2rHE86KJ9Qp7ZNmesZzkLm95CzMSOoSNhuUPz7h1QLn1+uoq0c4WAvwKPDp8Mf0z7Xla+pSTznf/M0OXztvOesb6t52rUcy7DOfQZby0il6J14to84iVsx5w3Ju09EWYcDpRj8ZdCPVujnq15udfXTv6Gd6l3oPldr61Rrx21Sf+AQv/AKDOOFiu3SJ8PNw0w3DSIynWnHwjv+WPKkfXE0k8NRn9QpP3MuJzRSQsnxzKbgwL6+xX6+5s6+23mjRfj1RGzKGX7n/CUX1cA6C0q9DzPY1P/mAefkbXcfZ+Vzz++xWyd6Qfvfb1hOPpbV6NeV8d5nhlv9LfjSUDqZMB6QdN3pM3tXOvX0cvfpqm99vV0CserIeMxBIzmroq0fWABHbO+VLzscDyezLuaz9Npp4RXge+tZX+b/pjRvM5lxl5FOyCUB8782uhZNM9qSbn5k/qXWpAGuozXz3beZGQc910MRjelqGfa9MO1NYZrR7xv+7522yLoB6K/mDS0/XxP7VYfcD1W2XxjU0QFRxcy35u86Jix9SBzEsezVn9WbhsaPU33RvON9u4so8+cseWPMaYuxnNu5LWfng+pXHX0dVp/06Z0PDFr7npOozZ6H1eWWU9oWHkYo6+3oEb6Gjf/UT0q1Z9w+SPpvenxkukfWye4csvM26YedD6rNw7cedXXTWl0EKYd7Tt/XddEAXHmZBXmZfXNMlXIwq5HOFlO12V0/iM84evjpq6GDwOvCNXyub9GBeiYpkR6TUV0f60SugrN0/AhzdbXtxidCxjpe8lIB1VYjxLEdG9DA83f/mz4sd4wRDVoonnE9H1mfmdBpy3qbUbTcjxLwfFOqkjj+eeNQWfcelDajTpjxy2Zi7h5y4zXWCQQw5tGpwcavbVZ19Lyi1F5dfFkF7f+48YDQMbYUGF4xBKGRyy5c4WZF4ft/gCnT7L6Nl0b0HUryd/pF7re88YdbatScDp+PdvOO3T9TtuMyp0AZP7mojL5azxo98/X5+nawqRl19WU9zz9y9EnIzKE42OOH7j0zly0tg7Ga2yfxPC4U2eixzqynraZv35gZIye0e3elFOByPdWnrf6WCzSg+VnMm/6awwgslakfER0X6uvm/9T3mH0PLYfvXpZPdLIEvq+ZLhwvBUpC4C7xuqRCEhU5yFtqueafTtGT2LXmt5cYfrXroEojYYHVViPmCzVMf0oUWfBQYYu/BMcMujnP5ku9u7dCwBYt25d9Jv169dj586d2LNnDw477DCbJpVu/fr1AGAP9XLYunUr+/zuu+8GNm1qGbn5t57VdlPCKjxUaDaHVwDYxRh74W8F6Jr5hlOCmQ3M3v7K2cwJQDemjOA3izsAvX2Vqww1tFXzqs3Tr5Nufw/W6lBpJRsxzmW0vrs0eFlhNkCdMunE5W2EO99577mLcJUGqp0zbqHehtbCEd7BRyYcou5rtv1peB1H4eXyMIoPVZqY+tl28DZo6KGd/YDhp1R4CYc4b8K3v0mfOM88WhWhievn6KKB0pEYDw4tDQbrdfttzfR/7fWJVxf/IDlXXjB2yNhSQ1eBDqDIvwxvsyHKSJv2DlQBTQ59TZ6VDukO4G0CRp8RBc6RWw1N1YDU2fBzgUJQLSYYheFtKsNo3aoDbaMN6aFHFWZsF3mkfaq9PZunIjzT5kNoSvCKM17pc0YW2u+WFHQ//h66DeEaCxc1WNe2OSfXSzcJ6Xh2YOR6Y0jg08JttDgLR9LmJu3Sxrr9zm/HnvfAkwfBwSIzjh3+aPqXmzdGFQjIj49JQoefF5WxzqEUQvlrFv6pud8hcVZHwwtx5bPI9Jnlc+6w2K+z36e0z6gs536btAPF6lQ2G27+Ib/VIr8BHtv88PmK8p4iuoCVZX4b+fzDjRfvGTdvxmj1aY7JO/qdPfAl7cjJMDpvmL6PbShZkPQmf27TO9BTvMxS48cH3VQIZBlG/WSzN/0Z24gnfMa1o9EJRvNX2AD04CpZb24TZeiNjcQ4r5bIXNeMYV8PMJvgsTnAlm/0yEgb0/VENW8mcEaP1cBwLhyblJ9LZVfOaCsmT+t1o0ZXZEOPyq1qd99NT8Zeb1c/DMFYt+3KzTssTd6cb2SIMyeT8ll5yOUbmZMDGU6z4daMkfT82ohk1iN9q8N1Hy0noEm544Hq/gHNDLTi17Wl+n5vfxV9BzQ6eOZAip2XEnQE5Ri52/Bjb3ePLZPVSQp0YzafDumU4U86p2cLIWXZjJp/agT6k1aAZtZzdI2nSXqrb+9Pd44dY0SvV0MAniFBp8NvKq+oiPTXr3VBvnbdp9Hb1Xdp8WQFyxOMzhY1uqjCTuHWexzYMZYDI4vVoDVKt3MCpaNmyeTzZPV99xm7v8PKs3QdUu2kFdoDT86gVHt8zOgiHG0GdK6pFkKCzaGrn6WhpSS8q6MrFMgGtaTgGz45/8Z0hkjfWv6g/ydzgJ2rSb5DRl7ExltgMMXtjeTyich2Z2x0kcvcnhHJ239eN8ZQQNOOvjGU8uTBotuolKTsuNdeaG+TD9MGdH0e6GS1avWDyLzTSfYKlhVK7pd7yOGgH9AJBAKBQCAQCAQCgUAgEAgEAoFAIBAIOkAO6B5yOOgHdBs2bAAA7N+/P/rNvn37AACHHXaYk8ak27hxYzZNDtu2bWOfb926FXfef79jlQyMLI9joRgMOAt++i7wBCmitEnDWRP10Xo1pCyXSt43Fh26rx1PD9aCh1qK+pa+vuWVX8kKwKD97bdZ1JK81JoNoaWna/naWiEbKzDq+QOEFlW+hXngzUdpi1gG57wKW+LDNNSyLGs5TS7OrZtyaP1ifJDkXcYiiHvWGb7lFvWy5CwYCT+xZXv8ZnnTtAOxUC+mDV6epPyA90n5dZ+3OA4sNSNWviYkQkzmmBBfioTaZb1Iqog1vnmvMnUCeJ42iNSZ83i1STzeMeORWn21HggqtCLz+sPKDsYSkspdx1qa8rux6DV0zLYWr2oAzOwd/Wfx8FgMvrYsm2eDuplhqUeIQxP1mGCs3g16i4r1ZDCoFlUb9o0Jtep7OxkLPWpJyckb+86xttOhtSGRN05IQmMpWLXzivW+jKQx0BXp0yUVyktvvAeWnT0E8yNAvICIRw3QhtTUjRmwDSkFl09cU9fmH9qnnDyLjEETsk0tMp0OuN5saGQY8QxOynUV8qSe0UDCalRpoudXCNrcGbu0/U2YRM5jhNDB0tzX7Ngx71lZypRB6xHQTL1fjcdU0w6OofqSckJUJYskVsKBt3ZHK2QOjjecp4sEluTceGXmmJwXE+dN51iBE8vewLJ5TrceiJrJX7eZ6p7Oz1sDt1NpP1L9wHoCZkIqR70iDG+QKBiuybJLl58nh7btW34KvCLMt0aux8J+e151Prp4o4WJ2591fyRfAK+dmUgUqbI5XZHm6VhuR3Qr+6wGMGh5zqRxdBYjt0k7x0KsOXRMAjOemLkAdTt2rccjXeuQeSm3CHTmTyI3U33KrjX9ucynWfO87PQdkScmn9QY7u9TGKzNrJlpP3qf5vKv59qoNjRihcHMngpLh9Vtem8u1T3iAUnLTazPqiXV6uRclAm4zwI9WDPPIqjX1KzXTScUrtuMB1w1T/Rtfxwizne0n5Lyhq7lDT8NXI8eTl7472LvHVq5pqM6jSmnx8hWquIycybdB+BA39VzOu357OmxlDZTpuZ0a072Rtosuq8So8mL7mTK7joXRMHU2Z+f/XndkZscyQsKw2a8qwWV5tnMXBCL8pQbB0GUD7JH5/C+5T1N6sc3IO0nq7MaHbXn0pjsU6LPOW2b0VktyFqPk5Gx8WjX1XQvI6WvD928fI/sVJjlrnDkFtGzSvX04u9Infp7FBaPHg3k3r4mogVgI3VUi6rtErM+9HTDlIwE+HmHA40AEHj/kshDRR7FgpWDRpl3vPlWcEjgoB/QHXvssQCAH/3oR+z7ffv2YefOnTjiiCPsYdvGjRtx+OGHY9euXfjRj36EU089NUhn8tuyZctU6Aw2++hmxBjKiaMITEvQlS7CY6BpbGyCeBlc+anwMeb/Dn2eklJCd9T1veNzU6Y5CODCwIyeuw8CXugQqsn/pkuIktzkyoIsrnIhIpyyEm1WxFvc4jqnzCbqxR6E5tohdn5CDi2LN0gjZTnpHKYG1ADBxkg2z1h/+CEK/GTevVFR0PHGleuPwUy7BGVRnZtT1gvALUy8pi2ih9ugZfkwIuPYjVQFDNZEZKOXngPLB5TM3BgxRWTCl+gZXRyqdUQDN4kl8qdyKyLD2DnQHHYp8KE9cwcvdYqoTNrI/BgL/8aF7cy2Zalcj21sxQ7mDDKHkpaMyKZrUJ4fWqWj3I62BxNG1iISki13WGSzHuTHR24uceSU2RDh6peRu0G+aGgvXBx10tPqdvO5a/oSme4flJakT7WPz8uBXuUVlp23Gp7KbYrm7nbLlkf5vePiNdsflc7mWSoDuEMMn4YuB8CjD8O0zuvcvDOB/l7ynWaMBPxvNQ3TmCpzmhsTGf3OGjxy8q50IwVw508/nxgII8fCA7cPI7/9NEQnytFing1nM3Rm0sfeG8RC1hosrfdCtvqbmMyhHved88qTf8FGvlceNXKwdBaOBxv6PUNTF0QP2Mx6QuX7pEu+wXec4Z+nOxSvRaO6bYIAbv6IhDbPydocnTaM/iBzB2VmHySmO5fyLns4FzO+TuiOuTVdJ92mgM6crhqk7+uorhkrf6Rb8etGNllXPdTvN05ecbpzZC4L9C8mZGcOWpXrTE66nNFupH9YI+MOayWtvMc5fapgX5E1BsqtRRkU8zz5briGD+/LrUtiumHRfkEBUjKky3iWA7yVh4S4fOjhoB/QPf7xj8fc3Bzuu+8+/PjHP8bP/dzPOe9vueUWAMCTn/xk5/lpp52GL3/5y7jllluCA7qlpSV85zvfwZo1a3DyySdPRN8999yDpQMH8OO3XD5RPgcFY2woFOeLZcx7JeVISfnLWf9c/tMse7nrMS0cKnR2wXLz+aRtthxtvtJje7lAD6a64lDk7XH6sWTjpGs54+RZkv9KyveDgdz8cijx4sMZpXy0Wvp0peg4mOPNnwMeCnrapHK2azkaK1/n5YZfv9UyRh9K6Mq3XeRpaZ7LiYM1LmlZK11nioNF08FYq62mdk0hx3MrMdd2kaGrkY8NVpoPpjn/ruZ2PlTBtWmunVdZPyzt2IF7mmh2goMMOaB7yOGgH9CtXbsWz3nOc/CZz3wGH/vYx/CGN7zBeX/11VcDAC644ALn+XnnnYcvf/nLuPrqq/HKV77SeXfddddhfn4e559/PtasWTMRfevXr8f+/ftR7dyJE044YaK8BAKB4FDE3XffDQAiAwUCwcMSIgMFAsHDGSIDBQLBwxkiAwWCMtyzbx/Wr1+/0mQ8PDGUA7qHGpTW0z92XbNmDRYWFhDL+gtf+AKe+9zn4sgjj8SNN96Ik046CQBw44034tnPfjbWrl2L73//+9i0aZNN88ADD+Bxj3scdu/ejWuuuQYveclLAAD33nsvzjrrLNx111244YYbcM4550xM/9atWwHE76kTCASChzJEBgoEgoczRAYKBIKHM0QGCgSChzNEBgoEgtWKrVu3Yvvd9+LsU3676Pt/++7/w5YTHiHy7BDANK7LxvXXX4+nPe1p9m9xcREAnGfXX3+9/f7cc8/FpZdeivvvvx8///M/jxe/+MV4wQtegGc+85kYDAZ4//vf7xzOAcDmzZvxvve9D1VV4eKLL8ZznvMcvPSlL8XjH/943HXXXbjsssumcjgnEAgEAoFAIBAIBAKBQCAQCAQCgUCwqqB12Z/gkMFUQlzed999+Pd///fgOX123333Oe+uvPJK/PzP/zze/e534/Of/zxmZ2dx7rnn4s///M/x9Kc/nS3noosuwpe//GX81V/9FW666SYsLi7i1FNPxetf/3pccskl06iKQCAQCAQCgUAgEAgEAoFAIBAIBALB6oIcvj3kMJUDule96lV41atedVDSnXXWWfjMZz7TuSyBQCAQCAQCgUAgEAgEAoFAIBAIBIJDEnJA95DDVA7oBAKBQCAQCAQCgUAgEAgEAoFAIBAIBMuEeqUJEEwbckAnEAgEAoFAIBAIBAKBQCAQCAQCgUCwSqE0oAo96JQ42h0ykAM6Btu2bVtpEgQCgWDFIDJQIBA8nCEyUCAQPJwhMlAgEDycITJQIBCsekiIy4cc5IBOIBAIBAKBQCAQCAQCgUAgEAgEAoFg1UIDdekBnRzkHSqQAzqBQCAQCAQCgUAgEAgEAoFAIBAIBILVjFouoXuoQQ7oBAKBQCAQCAQCgUAgEAgEAoFAIBAIVis0ykNcigPdIQM5oBMIBAKBQCAQCAQCgUAgEAgEAoFAIFjNKA5xKThUIAd0AoFAIBAIBAKBQCAQCAQCgUAgEAgEqxYa0KUhLuUg71CBHNAJBAKBQCAQCAQCgUAgEAgEAoFAIBCsZpSGuBQcMpADOoFAIBAIBAKBQCAQCAQCgUAgEAgEgtUKjfIQl3KOd8hADugEAoFAIBAIBAKBQCAQCAQCgUAgEAhWM8SD7iEHOaATCAQCgUAgEAgEAoFAIBAIBAKBQCBYzZADuocc5IBOIBAIBAKBQCAQCAQCgUAgEAgEAoFgNUMO6B5ykAM6gUAgEAgEAoFAIBAIBAKBQCAQCASCVQsN1HX5t4JDAnJAJxAIBAKBQCAQCAQCgUAgEAgEAoFAsFqhUX5AJ+dzhwzkgE4gEAgEAoFAIBAIBAKBQCAQCAQCgWA1o5aTt4ca5IBOIBAIBAKBQCAQCAQCgUAgEAgEAoFglUJDQ+syDzotLnSHDOSATiAQCAQCgUAgEAgEAoFAIBAIBAKBYLVCo9yDTs7nDhnIAZ1AIBAIBAKBQCAQCAQCgUAgEAgEAsFqhpaTt4ca5IBOIBAIBAKBQCAQCAQCgUAgEAgEAoFgNaMuC3EpOHRQrTQBAoFAIBAIBAKBQCAQCAQCgUAgEAgEggS0LvtbZfjqV7+KF7zgBdi8eTM2bNiAM844Ax/4wAfGzu/aa6/Fs571LGzcuBEbN27EOeecg+uvv5799gc/+AGUUtG/Rz3qUdFyHnzwQVx66aXYsmUL5ubmsGXLFrzhDW/Azp07x6bdh3jQCQQCgUAgEAgEAoFAIBAIBAKBQCAQrFZoDV3qQbeKDumuueYavPzlL0dd13jmM5+Jo446Cl/84hdxySWX4NZbb8Xb3va2TvldeeWVeOMb34h+v49zzz0Xc3Nz+NznPofzzz8f73rXu/D617+eTffIRz4Sz3/+84Pnhx9+OPv9jh07cOaZZ+Kuu+7C8ccfjxe/+MXYtm0brrrqKnzmM5/BjTfeiM2bN3einYPSehX1lkAgEAgEAoFAIBAIBAKBQCAQCAQCgQAAsHXrVmy/7cc4a+0Li77/6oFPYcsTfg7btm1bZsrSeOCBB/C4xz0Ou3fvxjXXXIOXvOQlAICf/exnOPvss3HXXXfhhhtuwDnnnFOU3+23346tW7ei3+/jhhtuwJlnngkAuOOOO/D0pz8du3btwm233YYTTzzRpvnBD36Axz3ucXjWs56FL33pS8W0v/KVr8SHPvQhvOQlL8FHP/pR9PsjX7ff/d3fxbve9S5ccskl+Pu///vi/GKQEJcCgUAgEAgEAoFAIBAIBAKBQCAQCASrGbUu+1sleO9734vdu3fjRS96kT2cA0bebG9961sBAFdccUVxfldddRWGwyFe+9rX2sM5ADj55JPxp3/6pxgMBrjqqqsmpvsnP/kJPvKRj2B2dhZ/8zd/Yw/nAODyyy/H0UcfjQ9+8IO49957Jy5LDugEAoFAIBAIBAKBQCAQCAQCgUAgEAhWK7SGHg6L/lZLiEtzL9zFF18cvDvvvPOwZs0afOELX8D8/PzE+Zln11577bjkWvzzP/8z6rrGM57xDDzykY903s3NzeGCCy7AcDjEpz/96YnLkgM6gUAgEAgEAoFAIBAIBAKBQCAQCASC1Qxdl/2tEnz7298GAJx++unBu9nZWTzxiU/E/Pw87rjjjmxeO3fuxD333AMAeMpTnhK8f+xjH4ujjjoK27dvx+7du4P3P/vZz/AXf/EX+M3f/E38wR/8Aa6++mosLi52pps+v/XWW7N059DPfyIQCAQCgUAgEAgEAoFAIBAIBAKBQCBYKehVFL4yh927d2PXrl0AgGOOOYb95phjjsHNN9+M7du348lPfnIyP3M4d8QRR2D9+vXR/Hbs2IHt27fjSU96kvPuu9/9Lv7yL//SeXbsscfiYx/7GM444wy2rBTdALB9+/YkzSWQAzqBQCAQCAQCgUAgEAgEAoFAIBAIBIJViv3Yixvrfy7+9u6778bWrVvZ99u2bZsmaSz27t1rf69bt479xhy07dmzpzi/WF6x/Obm5vDbv/3bePnLX44nPOEJWLt2LbZt24b//b//Nz796U/jl3/5l/Gtb30LW7ZsKS6rC905yAGdQCAQCAQCgUAgEAgEAoFAIBAIBALBKsQJJ5zQOY3xApsEF154IW677bZOaT7wgQ8EHmkriUc/+tH4m7/5G+fZ0572NFx//fV4xStegQ9/+MP4P//n/+A973nPitAnB3QCgUAgEAgEAoFAIBAIBAKBQCAQCASrEJ/61KdWpNzvf//7uP322zul2b9/PwBgw4YNzrONGzcG3+7btw8AcNhhh2XzNfmZ/Dl0yQ8A/uRP/gQf/vCH8dnPfrZTWV3LSUEO6AQCgUAgEAgEAoFAIBAIBAKBQCAQCAQW3/rWt8ZOu3HjRhx++OHYtWsXfvSjH+HUU08NvvnRj34EAE54yRiOPfZYAMCDDz6Iffv2sffQdckPAE466SQAwE9+8hO2LJPfpOWkUE2cg0AgEAgEAoFAIBAIBAKBQCAQCAQCgUDQ4LTTTgMA3HLLLcG7paUlfOc738GaNWtw8sknZ/PatGmTPTj75je/Gbz/4Q9/iB07dmDLli2stx6HBx98EACCw74U3fT5k5/85KJyUpADOoFAIBAIBAKBQCAQCAQCgUAgEAgEAsHUcN555wEArr766uDdddddh/n5eZx77rlYs2bNxPmZZxdccEExfddccw0A4PTTT3eeP//5z0dVVfjKV76Ce++913m3sLCAa6+9Fr1eDy94wQuKy4pBDugEAoFAIBAIBAKBQCAQCAQCgUAg+P/bu1+QutoAjuM/8WKUoajYBFEuC4JgMN6gzbIyGAgWyzZFULwYFbQoIpgEs1FR1sSyINxo2EQQMRm8DrEJIt6VF0HcXt6XjZM+n/j8Oec5T/2GA/w1k5OTaW1tzcHBQfb29p7H6/V6qtVqkmRubu7VvnK5nHK5nKurqxfjMzMzaW5uztbWVmq12vP4+fl5VlZWUiqVMjMz82LP9vZ2zs7OXr1jb28vCwsLSZLPnz+/mOvu7s6HDx/y8PCQT58+5fHx8XmuWq3m5uYm4+Pj6ezs/K9X8VtNjUaj8cdPAQAAAAAAgH/s7u7m/fv3aTQaqVQqaW9vz9HRUe7u7jI7O5v19fVXe5qampIkl5eX6enpeTG3sbGR2dnZlEqljI6OpqWlJYeHh7m/v8/m5mamp6dfrK9UKvn69WsGBgbS39+fp6ennJ6ePke7+fn5rK6uvjrDjx8/Mjw8nIuLi/T29mZoaCjfv3/Pt2/f0tfXl1qtlra2tj++H4EOAAAAAACAv+74+DjLy8up1Wp5eHjI27dvMzU1lYmJiV+u/7dAlyRfvnzJ2tra87/oBgcHU61WMzY29mrtzs5O9vf3c3Jyknq9nvv7+3R0dGR4eDgfP37MyMjIb899e3ubxcXF7O/v5/r6Ol1dXXn37l2Wlpby5s2b/38Rv/pWgQ4AAAAAAACK4x90AAAAAAAAUCCBDgAAAAAAAAok0AEAAAAAAECBBDoAAAAAAAAokEAHAAAAAAAABRLoAAAAAAAAoEACHQAAAAAAABRIoAMAAAAAAIACCXQAAAAAAABQIIEOAAAAAAAACiTQAQAAAAAAQIEEOgAAAAAAACiQQAcAAAAAAAAFEugAAAAAAACgQAIdAAAAAAAAFEigAwAAAAAAgAIJdAAAAAAAAFAggQ4AAAAAAAAK9BO2bzs9zDglVgAAAABJRU5ErkJggg==", + "image/png": "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", "text/plain": [ "
" ] @@ -919,7 +943,7 @@ } ], "source": [ - "plt.matshow(per_module_scores[\"6\"] / len(train_dataset))\n", + "plt.matshow(per_module_scores[\"6\"])\n", "plt.colorbar()" ] }, @@ -931,6 +955,62 @@ "Note that the scores were divided by `len(train_dataset)`, as `compute_pairwise_scores` does not normalize the scores by the total number of training dataset." ] }, + { + "cell_type": "markdown", + "id": "ef7883af-cb74-484f-b13b-bca1f6492c17", + "metadata": {}, + "source": [ + "Furthermore, we can compute self-influence scores." + ] + }, + { + "cell_type": "code", + "execution_count": 25, + "id": "a6222034-60e6-40fb-b0a1-14f6e7aa1a2b", + "metadata": {}, + "outputs": [], + "source": [ + "analyzer.compute_self_scores(\n", + " scores_name=\"tutorial_score\",\n", + " factors_name=\"tutorial_factor\",\n", + " train_dataset=train_dataset,\n", + " overwrite_output_dir=True,\n", + ")\n", + "self_scores = analyzer.load_self_scores(scores_name=\"tutorial_score\")" + ] + }, + { + "cell_type": "markdown", + "id": "c60deb7f-2ddf-4c0a-ba02-9e7678fbb1c9", + "metadata": {}, + "source": [ + "You can load the self-influence scores with `compute_self_scores`. The pairwise score will have the dimension `train_dataset_size`." + ] + }, + { + "cell_type": "code", + "execution_count": 26, + "id": "fb200b31-ff11-4ddd-964f-4868a6f82250", + "metadata": {}, + "outputs": [ + { + "data": { + "image/png": "iVBORw0KGgoAAAANSUhEUgAAA2EAAAJqCAYAAACivEhIAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjguMywgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/H5lhTAAAACXBIWXMAABcSAAAXEgFnn9JSAACCNElEQVR4nOzdeXxU9b3/8fdk3zNkAQIBEiCAgGETJCKCqCCiohasbbWmWFur9BrR+9PWpQrWpQvmXrWtXUTcvUJFAWsBWZQl7BDZBEIWAgRCyGRfJpnz+4NmICaBJJzMTJLX8/Hg8ZjzPefM+Qx+JfPO95zv12IYhiEAAAAAgEt4ubsAAAAAAOhMCGEAAAAA4EKEMAAAAABwIUIYAAAAALgQIQwAAAAAXIgQBgAAAAAuRAgDAAAAABcihAEAAACACxHCAAAAAMCFCGEAAAAA4EKEMAAAAABwIUIYAAAAALiQj7sLQNNsNptsNpskafTo0aqsrFTv3r3dWxQAAAAA5eTkKDg4WHl5eS0+l5EwD5aamqr4+HjFx8fr9OnTqqiocHdJKisrU1lZmbvLQDtHP4IZ6EcwA/0IZqAfdU52u73V/90thmEYJtcDk5w/EjZ58mR5e3tr//79bq1p9erVkqRJkya5tQ60b/QjmIF+BDPQj2AG+lHnNGTIEEnS3r17W3wutyN6MKvVKqvVKkny9fV1bzEAAAAATMHtiAAAAADgQoQwAAAAAHAhbkf0YOc/E2a32+Xt7e3eggAAAABcMkbCPNj5syMeOnRIBQUF7i4JAAAAwCUihHmwlJQUZWZmKjMzUwkJCYqMjHR3SQAAAAAuEbcjejBmRwQAAAA6HkbCAAAAAMCFCGEAAAAA4EKEMAAAAABwIUIYAAAAALgQE3N4MNYJAwAAADoeRsI8GOuEAQAAAB0PIcyDsU4YAAAA0PFwO6IHY50wAAAAoONhJAwAAAAAXIgQBgAAAAAuRAgDAAAAABcihAEAAACACxHCAAAAAMCFmB3Rg7FYMwAAANDxMBLmwVisGQAAAKjv718f0Quf73d3GZeEkTAPlpKSouTkZEnS5MmTGQkDAABAp/biv/brjXVHJEldQ/310/F93VxR6zAS5sGsVqvi4uIUFxcnX19feXnxnwsAAACd0ze5Rc4AJklvb8pWRXWtGytqPb7VAwAAAPB4W7LOOF8H+Xlr0S+SFOjXPu8UI4QBAAAA8HiHTpY4X995RS91DQ1wYzWXhhAGAAAAwON9e14IG9g91I2VXDpCGAAAAACPVlRh1ze5Rc7tAd1C3FjNpSOEAQAAAPBor355SDUOQ5Lk7+Olgd3D3FzRpSGEAQAAAPBo27ILna+/P7qXQvzb90pbhDAAAAAAHq3Sfm4q+qv7R7mxEnMQwgAAAAB4tIrzQliQX/seBZMIYQAAAAA83PmLMgf6tf8I0/4/AQAAAIAO7fyRsADf9rlA8/na/1heB2az2WSz2SRJdrtd3t7tv8MBAAAALVVvJKwDhDBGwjxYamqq4uPjFR8fr0OHDqmgoMDdJQEAAAAuZa91OKenl6RAP0IY2lBKSooyMzOVmZmphIQERUZGurskAAAAwKXOvxVR6hgjYdyO6MGsVqusVqskydfX173FAAAAAG5QWf2dEMZIGAAAAAC0nfNHwrwskp93+48w7f8TAAAAAOiwzg9hgb7eslgsbqzGHIQwAAAAAB6rvN4aYe3/VkSJEAYAAADAg53/TFhHWCNMIoQBAAAA8GDfvR2xIyCEAQAAAPBY9UIYtyMCAAAAQNuqqGYkDAAAAABcppKRMAAAAABwnXJGwgAAAADAdY7bKpyvg/x83FiJeQhhAAAAADySw2Hoi715zu2RfazuK8ZEhDATvfXWW7JYLA3+rF271t2lAQAAAO3O/rxinSyukiR5e1k0dWiMmysyR8cYz/Mw69evl7f3uftVBw8e7MZqAAAAgPbpdGm183WvLoGKCPZzYzXmIYS1gSuvvFI+PvzVAgAAAJeiuMLufB0W6OvGSszF7YgAAAAAPFJx5XkhLIAQ1u5s375dL730ku644w7FxsY6n9e6mIqKCj3zzDMaMGCAAgIC1KNHD82aNUvHjh1r8pyePXvKx8dHiYmJWrRokZkfAwAAAOg0iitqnK/DAjvOnWYd55NcxLx58/Tpp5+26JzKykpNmjRJaWlpiomJ0fTp05WVlaUFCxZo2bJlSktLU9++fZ3Hx8TE6Le//a2uvPJKVVRU6B//+IdmzpypJUuWaPr06WZ/JAAAAKBDK+mgI2GdJoQlJSUpMTFRo0eP1ujRoxUXF6eqqqoLnvP8888rLS1NSUlJWrFihUJCQiRJ8+fP16OPPqpZs2bVm/lwypQpmjJlinP75ptv1vjx4/XCCy8QwgAAAIAWOv92xNCAjhNdOs4nuYjHH3+8RcdXV1frtddekyS9/vrrzgAmSXPmzNHChQu1bt06bd++XaNGjWryfaZPn64nn3yydUUDAAAAnVi92xE70EhYp3kmrKU2bNigoqIi9evXTyNGjGiwf8aMGZKkpUuXuro0AAAAoFOoNzFHB5odsdOMhLXU7t27JUkjR45sdH9de3p6epPvYRiGPvnkk0ZDXFOGDBnSaHtGRoa6d++u1atXN/u92kJZWZkkub0OtG/0I5iBfgQz0I9gBvpR28k9ee7xoaNHDmp15RE3VlNfWVmZgoODW3UuIawJOTk5kqTY2NhG99e1Z2dnO9tmzJihMWPGKDExUVVVVfr73/+uTZs26bPPPmv7ggEAAIAOptxuOF8H+Vx8ZvP2ghDWhNLSUklSUFBQo/vrUm9JSYmzbcCAAfr73/+u3NxcSdKIESO0bNky3XTTTc2+7t69exttrxshmzRpUrPfqy3U/YbH3XWgfaMfwQz0I5iBfgQz0I/aTm3aKklnR8OuvnKUxsRHuLeg87R2FEwihJnqhRde0AsvvGDa+9lsNtlsNkmS3W6Xt7e3ae8NAAAAeLJKe60Ky849ExbegZ4JY2KOJtTNhlheXt7o/rp7f0NDQ9ushtTUVMXHxys+Pl6HDh1SQUFBm10LAAAA8CRpRwpUXeuQJAX7eSs+qvUjT56GENaE3r17S5Lz1sLvqmvv06dPm9WQkpKizMxMZWZmKiEhQZGRkW12LQAAAMBT1NQ69IcV3zq3r+ofJT+fjhNduB2xCcOGDZMk7dixo9H9de2JiYltVoPVapXVapUk+fp2nOFXAAAA4EI+2XlMe44VO7cnD+7mxmrM13HipMnGjRun8PBwZWRkaNeuXQ32L1q0SJJ0yy23tFkNNptNWVlZysrKkt1ul8PhaLNrAQAAAJ5izbennK+HxYbrjpGNz1jeXhHCmuDn56fZs2dLkh566CHnM2CSNH/+fKWnp2vChAkaNWpUm9XAM2EAAADobGodhjYcPve995eTEuTt1XGmp5c60e2Iy5cv17x585zb1dXVkqSxY8c6255++mlNmzbNuf3UU09p1apV2rhxoxISEjR+/HhlZ2dr8+bNio6O1ptvvtmmNaekpCg5OVmSNHnyZGZHBAAAQIeXXVCmooqzsyJ6WaQr+3rOtPRm6TQhLD8/X5s3b27Qfn5bfn5+vX0BAQFas2aNXnzxRb3//vtasmSJIiIilJycrHnz5jW5kLNZeCYMAAAAnc2pkirn6+hQf4UGdLzvwZ0mhCUnJztHlVoiMDBQc+fO1dy5c80vCgAAAEA9p0vPhbCoEH83VtJ2Ok0Ia49YrBkAAACdzemSjh/CmJjDgzExBwAAADqb06XVzteEMLgcizUDAACgs6l3O2KonxsraTvcjujBmJgDAAAAnc35ISyakTAAAAAAaFv535kdsSNiJMyDMTEHAAAAOhOHw1DOmXLnNs+EweWYmAMAAACdSfqxIhWWn12o2dvLoqE9wt1cUdsghHkwJuYAAABAZ7Ex47T+64Odzu0r+nRReFDHnBeB2xE9GBNzAAAAoDOoqK7Vz9/erpKqGmfbrcN7uLGitsVIGAAAAAC32pp1pl4A+9k1ffXDMb3dWFHbYiQMAAAAgFtU2mu1OfOMnlu619k2rn+kfn3TZW6squ0RwjwYsyMCAACgozIMQ/ct3KoNh+tPPjdpUDc3VeQ63I7owZgdEQAAAB3VWxuzGgSwUH8f3XR5dzdV5DqEMA/G7IgAAADoiDZlFGjesn312h6Y0E//fuQaxYQHuqkq1+F2RA/G7IgAAADoaMqra5Ty0U45jLPbPl4WLfrFVRrey+rWulyJkTAAAAAALnGiqEJ3/GmjThZXOduev21opwpgEiNhAAAAAFzgvc3ZeunzA/Wmoh8WG67vj+7lxqrcgxAGAAAAoE19seeEnvxkT722O6+I1eM3DpLFYnFTVe5DCAMAAADQplJXHXK+DvLz1rO3DNGdnXAErA4hzIOxThgAAADau7KqGn17ssS5/dZPxmhMfIQbK3I/JubwYKwTBgAAgPZu7/FiGf+ZCTHYz1tX9Oni3oI8ACHMg7FOGAAAANq79Fyb8/XQnuHy8up8z4B9F7cjejDWCQMAAEB7l3bk3N1cwzrZVPRNIYQBAAAAMF2lvVYfb8/Vqv2nnG3jE6LcWJHnIIQBAAAAMN3zy/fp3bQc53aQn3enn5CjDs+EAQAAADBVUYVdH2/Lrdc2fXhP+fsw27fESBgAAAAAk320NUdVNQ7n9h9mDtPtI3q6sSLPQggDAAAAYIpPdx3TJzuPae23+c62+8fHa8aoWDdW5XkIYR6MxZoBAADQXqzad1IPf7irXluAr5fuv6avewryYDwT5sFYrBkAAADtxT/WZ9bbDvH30dzpQ9U1NMBNFXkuRsI8WEpKipKTkyVJkydPZiQMAAAAHic916bnlu7T9uxCZ9sr3x+maZf3kJ8PYz6NIYR5MBZrBgAAgCdLO1Kge/6xWfZaw9k2srdVtw3vKYvF4sbKPBshDAAAAECLVNpr9ee1Gfrzuox6ASwqxF//c9cIAthFEMIAAAAANJvDYeiXH+zUyn0nnW2+3hY9d+tQTbs8RuFB3MF1MYQwAAAAAM22NP14vQDm5+2l/7lruKZeHuPGqtoXQhgAAACAZvnXNyfqTUM/oFuI3kwerdguQe4rqh0ihAEAAAC4oFqHoeeX79OCDVn12n8/YxgBrBUIYQAAAAAa9W1eid5Jy9LmI2d06FSps93by6JHJw/QsF5W9xXXjhHCAAAAANRTXGnX37/O1OtrDqvWYdTbFx8VrNd+OEJDeoS7qbr2jxAGAAAAdHJnyqq1LP24cgsr9G1eidYfPt0gfEnS1KHdNf/O4Qr083ZDlR0HIQwAAADoJBwOQ1uyziivqFIllXYdyCvRvhPF2nXUJqNh5pJ0dvbD+6+J102XxzD6ZRJCGAAAANBJPP3pHr23OadZxwb7eevusX10T1IfJt8wGSHMg9lsNtlsNkmS3W6XtzfDvgAAAGgZwzD0zbEifbrr+AUDmK+3RZOHdFe/qGAN723VNQnR8vH2cmGlnQchzIOlpqbqueeec25HR0e7sRoAAAC0JzW1Di3/5oT+vDZDB/JKGuwfnxClmPAAjejdRb26BGlozzBZg/zcUGnnQwjzYCkpKUpOTpYkTZ48mZEwAAAAXFRVTa0+3parv351RDlnyhvsj48K1ns/vVI9rIFuqA4SIcyjWa1WWa1WSZKvr697iwEAAIDHK62q0Y/+vlm7j9oa7IuPCtb1l3XV7GsTFB7Ed0t3IoQBAAAA7Ux1jUOHTpXoVEmVNmUUaN/xYp0srlTm6TLVnDe1vLeXRdOH9dADE/tpQLdQN1aM8xHCAAAAgHbkuK1Ct/9pg04WV13wuOsv66bf3DJYvSKY2dDTEMIAAACAdsJe69DLXxy4YADz8bLoB2N665lbBsuX2Q09EiEMAAAA8HAZ+aX6v21HtWz3CR2zVdTbNyY+QhMHRqt/dIi6hQUoLjKYZ748HCEMAAAA8AD7jhdrW/YZVdc4VFXjUGFZtfadKFZeUaWOnC5rcHz/riH618PjGe1qhwhhAAAAgBsYhqHM02X6dNdx7cgp1NeHTjfrPB8vi24b0VP/b8pAAlg7RQgDAAAAXGzp7uP689oM7TtR3OxzugT56gdjeutHY/uoJ2t8tWuEMAAAAMBFbOXVeuSjXVrzbX6j+6NC/DSydxf5+3orxN9HcZFB6t81RFEh/hoUEyp/H28XV4y2QAgDAAAA2tiJogp9/k2e/vpVRoOZDbuF+WvKkO66dlBXje8fJR9uMezwCGEAAABAG/n6UL4+2XlMy9NPqKrGUW9fdKi//nHvFUqMtbqnOLgNIQwAAAAwSU2tQ7tzi/T+5hztyClUZiOzGkrSlCHd9ORNg9U7koWUOyNCWBv55ptvNGLECHXv3l25ubnuLgcAAABtKL/codnv79Dab/NVWlXT6DGBvt66bURP3TW6l4b1srq2QHgUQlgbSUlJUWRkpLvLAAAAQBsrtRt6YUuViqpPNLr/spgw3T8+XpMGdZU1yM/F1cETEcLawJIlS3TkyBHNmjVL77zzjrvLAQAAQBs4eqZcr64+pP/bVlmv3cfLokExoZo6NEZX94/S5T3D5eVlcVOV8ESEMJNVV1frscce00svvaT9+/e7uxwAAACYrLSqRou2HdUfVx5USWX9Ww/vTeqjx6cOUpAfX7PRtE4z/+X27dv10ksv6Y477lBsbKwsFosslov/RqKiokLPPPOMBgwYoICAAPXo0UOzZs3SsWPHGj0+NTVV0dHR+v73v2/2RwAAAIAb2WsdWrE3T9f/cZ2eXbqvQQC7ql+kfnXTZQQwXFSn6SHz5s3Tp59+2qJzKisrNWnSJKWlpSkmJkbTp09XVlaWFixYoGXLliktLU19+/Z1Hn/y5En99re/1RdffGF2+QAAAHCTmlqH1h3M128+26vcwop6+/x8vHRXgrcSo731valXNuuX/ECnCWFJSUlKTEzU6NGjNXr0aMXFxamqquqC5zz//PNKS0tTUlKSVqxYoZCQEEnS/Pnz9eijj2rWrFlau3at8/hf//rXuvHGG5WUlNSWHwUAAABtzOEwtOd4kb7Yk6f3NueoqMJeb7/FIt0zto/uuzpeGbs3/6eNAIbm6TQh7PHHH2/R8dXV1XrttdckSa+//rozgEnSnDlztHDhQq1bt07bt2/XqFGjtGfPHr377rtKS0uTzWaTdHYkzTAM2Ww2BQUFyc+P2XAAAAA82dEz5Xp9zWGt/TZfecWVjR5z7cBo3X9NX13VL0qSlOHKAtEhdJoQ1lIbNmxQUVGR+vXrpxEjRjTYP2PGDKWnp2vp0qUaNWqUDh8+rOrqao0cObLBsV26dNGf//xnPfDAA64oHQAAAC1UVlWjf+/N09xl+2QrtzfY7+1l0eCYMD1yQ4ImDermhgrRkRDCmrB7925JajRUnd+enp4uSbr66qu1Zs2aese89dZbWr58uT7++GMNGDCgWdcdMmRIo+0ZGRnq3r27Vq9e3az3aStlZWdXfXd3HWjf6EcwA/0IZqAfdW5FVYYO2xzKLnZobW6NSr+TvSySeodZNLqbtybG+ijIt1o6vlerj++tdxz9qHMqKytTcHBwq84lhDUhJydHkhQbG9vo/rr27OxsSVJUVJQmTpxY75i1a9fK39+/QTsAAABco9xuKO1ErQoqDVU7DFXYpWOlDhVWGSqubvwcH4v0/YG+GtXNW1Z/nvOC+QhhTSgtLZUkBQUFNbq/LvWWlJSYet29e/c22l43QjZp0iRTr9dSdb/hcXcdaN/oRzAD/QhmoB91bIdPleiuv27W6e8OcTXBz8dLkwd30y8nJWhg99BmX4d+1Dm1dhRMIoS1qWeffVbPPvtsq8+32WzOST7sdru8vb3NKQwAAKCDqnUYemXlQX2xN0+HT5Ve9PioED9dFhOmq/tH6ZZhPdTDGuiCKtHZEcKaUDcbYnl5eaP76+79DQ1t/m9JWio1NVXPPfecczs6OrrNrgUAANCe1ToMFZRV6S9rj+jNDZkN9l83qKv6RAYr0M9LfSKC1Tc6WN3CAtTTGigvL245hGsRwprQu3dvSVJubm6j++va+/Tp02Y1pKSkKDk5WZI0efJkRsIAAAB0dvHkrw+d1qFTJTp4slRbs84ot7BCtQ6j0ePvHx+vJ6cNdnGVQNMIYU0YNmyYJGnHjh2N7q9rT0xMbLMarFarrFarJMnX17fNrgMAAODJNmUUaOHGLB3OL1VpZY3yS6uaDFzne//+KzW0Z7jCAvgeBc9CCGvCuHHjFB4eroyMDO3atUvDhw+vt3/RokWSpFtuuaXNauCZMAAA0NkYhqH8kioVV9q162iRvtiTpy8PnJRx8cwlSQr289awXlY9Oe0yDekR3rbFAq1ECGuCn5+fZs+erd/+9rd66KGHtGLFCucMKPPnz1d6eromTJigUaNGtVkNPBMGAAA6urKqGh23Veh0abW2Zp3RB1tydKKo8qLnWSzSiF5WDetlVXxUsK7qF6nu4YEK8efrLTxfp+mly5cv17x585zb1dVnF4YYO3ass+3pp5/WtGnTnNtPPfWUVq1apY0bNyohIUHjx49Xdna2Nm/erOjoaL355pttWjPPhAEAgI6m1mHIVl6tnDPl+utXR7Ry30nVNOPWwr5RwXr4+gR1DwtQSICPYsIDFRHs54KKAfN1mhCWn5+vzZs3N2g/vy0/P7/evoCAAK1Zs0Yvvvii3n//fS1ZskQRERFKTk7WvHnzmlzI2Sw8EwYAANqjkkq7sgvKtTvXpn3Hi1VSWaMDecU6XVotW3m1mpG55O1lUe+III3obdXgmDDdNaY3o1zoMDpNT05OTnaOKrVEYGCg5s6dq7lz55pfFAAAQAex51iRFu/I1eFTpdp85Iyqax3NPtfP20uRIX6KDPHT5T2tmj2pv3qyXhc6sE4TwtojJuYAAACeqKK6VmmZBfr64Gl9e7JYR89UKOdM42urNsXX26IhPcL1k3Fxujmxh7xZqwudCCHMgzExBwAAcLdKe63yS6p0qqRKBaVV2p1r07tpOSqqsF/03OhQf13Rp4vio4LVLzpEcVHBigz2U0SIn0L9fWSxELzQORHCPBgTcwAAAFcrr67Rum/ztWTXMe07cXaUq7lG9LZqypDu6h0RpIkDoxXkx1dNoDH8n+HBmJgDAAC4Qk2tQx9sPaov9pzQ1szCZj3PZbFI11/WTUl9I9UrIkiDuoeqV0SQC6oF2j9CGAAAQCeWW1iuxz7erbQjZy54XESwn6JC/NQlyE+9I4J0/zV9NaBbqIuqBDoWQpgHY2IOAADQVorK7Xrmsz36dNfxBvtC/X10zcBoTR7cTSN7d1H38AD5enu5oUqgYyKEeTAm5gAAAGYqqrAr7UiBvjqYryU7j6msurbe/smDu+n+a/pqeC8roQtoQ4QwD8bEHAAAoDX2HS/WpiMF/5nVsFK7jtp0qrhKpVU1jR4f7OetF+64XLcO68GMhYALEMI8GBNzAACA5jAMQwdPlirtSIFW7T+prw+dbva5M0fFat5tQxXgyy97AVchhAEAALQDDoehzZlntDHjtM6UVaum1lBJlV2HTpbqVElVs9btkqQuQb66OiFa4xOiND4hSjHhgW1cOYDvIoQBAAB4qLKqGn2xJ0+r9p/U7qM2HS+qbPa5g7qHangvq6xBfkroGqKB3UMVHeqv6BB/eXlxyyHgToQwD8bsiAAAdC5VNbU6cKJEu3Nt+vybE9qeXSh7rdGscy0WaUiPMI2Nj9S4hChdkxAtb8IW4JEIYR6M2REBAOj4Ku21emPdEa3af1IH8oovGLq8LFJSv0gNjgmTv4+3fL291K9rsLqFBWhA11CFB/EMOdAeEMI8GLMjAgDQ8dTUOrRq/ykdyCvWqZIqpWUU6MjpsiaPDwvw0bWDumpc/yhNHBitrqEBLqwWQFsghHkwZkcEAKDjqKqp1ac7j+v3K75VfklVk8f5+XgpoWuIhvWy6nsje2pYrFU+rNkFdCiEMAAAAJPVOgxlF5Tp27wSHcgr0d7jxVp/OF+Vdkejx1ss0k+uitf3RvXUgG6hLJQMdHCEMAAAgEtUWlWjtzZk6rPdx3W6tFrFFXbVOC48oUbfqGBdd1lXRYf6a3xCtC6LCXNRtQDcjRAGAABwETW1DqUfK9LRM+U6kl+myppaVdkdOlFUoX0ninX0TEWz32viwGglXxWnCQOiZbEweyHQGRHCAAAAdHaWwm1ZhSooq5Kt3K6TxZXKK6rUiaJK7c8rlq28eYsh17FYpLjIYA3sFqr+XUMU2yVQNyXGKCyA57yBzo4Q5sFYJwwAgLZVXGVob0GtPv94t9Z+e0qnS6tb/V5Bft76cVKcpgzpprBAX/UID1SgHz+7ATRECPNgrBMGAIA5HA5De44X6cCJEmXkl6q4ska7j9q070Tlf47Iveh7+Hl7KT4qWP27hSg80Ff+Pl4KC/DV4B5hzpEufx9CF4CLI4R5MNYJAwCgdQzD0L4TxVr7bb4OnSzR7twiZV5gLa46Pl4WDe4RJmuQn7qG+ismPEDdwwPUwxqo0XERCvHnqxOAS8e/JB6MdcIAALi4wrJqrTuYr8LyahWW27Ut64wy8kt1srjptbjO18VfuuHyWF3VL0pXJ0QpKsS/jSsG0NkRwgAAQLtT6zBkK6/Wou25Sl11SBX22madFxcZpFF9IhQe6Kv46GDp1CH1CLbouuuGtXHFAHAOIQwAALQrf16bodRVB1VV0/jCx+cb0duqpL6R6hsdosTYcA3oFlpv/+rVGW1VJgA0iRAGAAA8nmEY2ppVqI+3HdXH2xufRGNs3wh1CwtQt7AAjY6L0LDYcHUNC3BxpQBwcYQwAADgEapqalVUYVdJZY32Hi9WbmG5sk+X65tjRSooq2r0Ga9Qfx/dPKyHfjIursEoFwB4KtNDWE1NjZYvX64tW7bo9OnTuvLKKzVr1ixJ0vHjx3X69GkNHjxYPj7kPwAAOruicrtW7j+p//nyoI6eqWj2eeGBvvrrPaM0qk8X+Xh7tWGFAGA+U5PQ+vXrdffdd+vo0aMyDEMWi0V2u90ZwjZt2qQ777xTH3/8se644w4zL90hsVgzAKCjMAxDGfmlOnyqTNuzz+hkcZXSc23KKihv0fuMiYvQtMQY3Taip8IDmTkYQPtkWgjbt2+fbrzxRtntdv3yl7/U1VdfrTvvvLPeMbfccouCgoK0ePFiQlgzsFgzAKA9q7TX6os9edqZU6iNGQU6dKq02ef2jgjSZTGhigkP1OCYMPWKCFLvyCD1tAa2YcUA4BqmhbB58+apsrJSn3/+uSZPntzoMX5+fho5cqR27txp1mU7NBZrBgC0NyeLK/XOpmxtzy7UlqwzqnUYFz0nxN9H113WVbPGxSuhW4gCfb1lsVhcUC0AuIdpIWzNmjUaM2ZMkwGsTs+ePbV7926zLtuhsVgzAMBTGYahPceKdehUidYdzNfe48Uqq6rRyeJKXSh3xYQHaFSfLoqPClb/riEa2buLYrsEEroAdCqmhTCbzaZevXpd9LiysjLZ7XazLgsAAFzsdGmVHnxvh7ZknrnosRaLdNPQGI3tF6mkvpHq3zXEBRUCgGczLYR17dpVhw8fvuhx+/fvb1ZYAwAA7ldpr1VVjUMOh6GvDuXrq4OntXJfnoora5o8xxrkq3uT4jSyTxcN6h6qbqzVBQD1mBbCJk2apHfeeUdr1qzRtdde2+gxn3zyiQ4fPqyHHnrIrMsCAAATVVTXat3BfG3KOK0NGQU6fJHJNPpGBys+MlgTB3XVgK4hsgb5qW90sHyZNh4AmmRaCHviiSf00Ucf6bbbbtNLL72k22+/3bmvsLBQn3zyiR577DEFBwdrzpw5Zl0WAABcgkp7rfJLqrRk5zH9c+cxZZ4ua9Z5Ab5e+uWkBD04sR/PcwFAC5kWwgYNGqQPPvhA99xzj2bPnq3Zs2fLYrFo4cKFWrhwoSQpICBAH3zwgeLj4826LAAAaAHDMHSmrFo5Z8r1wZYcLdl5XNW1jmad62WRrugToVuH99Atw3qwThcAtJKpizXfdttt2rNnj1555RWtXLlSWVlZcjgcio2N1Q033KBHH31U/fr1M/OSAACgGZbsPKb3NmfrQF6JSi7wPJck+XpbNKpPF00Z0l0TBkQrMsRfFovk5+2lAF+WSwGAS2VaCPvss8/k6+urqVOnKjU11ay3BQAArVBSadcXe/L05f5Tyioo04G8kgseHxbgo/EDonX/+L4a0C1EQX6m/p4WAHAe0/6Fvf322zV58mRNnTrVrLcEAADN4HAY+teePK3Yl6fiCrsOnSpVbmHFBc8J8vNWv+gQXZ0QpQcn9lNoALcWAoCrmBbCoqOj1aVLF7PeDgAAXIRhGPps93G99K8DOlFUedHj+3cNUcr1Cbq6f5TCA32ZUAMA3MS0EDZx4kRt2bJFhmHwjzoAAG2oqMKub3KL9Nevj+irg/lNHhfs561bh/fQmPgI9Y0KUWJsOD+jAcADmBbC5s2bp9GjR+uRRx7RSy+9pIAAFma8VDabTTabTZJkt9vl7c3D0ADQmTgchtIyC7TveLGqahz6Nq9E3xwranQaeYtFurp/lCYMiFZPa6CG9gxXTHiAfFivCwA8jmkh7IMPPtBNN92kV199VR9++KGuv/569e7du9EwZrFY9PTTT5t16Q4rNTVVzz33nHM7OjrajdUAAFwh83SZ9hwr0vbsQn2y85iKKuwXPWd8QpR+e9vl6h0Z5IIKAQCXyrQQ9uyzz8piscgwDJ06dUrvv/9+k8cSwponJSVFycnJkqTJkyczEgYAHVBRuV2bMwu0O9emLZlntDWrsFnnhQX4KDHWqilDuulHV/aRlxe3GQJAe2FaCFuwYIFZb4X/sFqtslqtkiRfX2atAoCOYlNGgd5Jy9LhU6U6dKpUhtH0sRaLNKKXVd3CAtQ7IkhDe4YrMTZcvSOCeL4LANop00LYvffea9ZbAQDQ4ew/UawPtuTo82/ydLq06oLH9o4I0uWx4Roea9WMUbHqEuznoioBAK7ASowAALQBh8PQqv0ntf7waWWeLtPXh043eWxksJ9Gx0Xo8thwjerTRVfGRzDKBQAdWJuEsE2bNunrr7/WsWPHJEk9e/bU+PHjlZSU1BaXAwDAY1TXODRv2T79a8+FR7xG9rbqh1f20dCeYRrYLZTQBQCdiKkh7ODBg7rnnnu0bds2SWcXkZTk/MFyxRVX6N1331VCQoKZlwUAwC0Mw9Dp0mrlnCnX1qwzKq2s0b/2nFBGfsMp5CUpwNdL910dr++NjFXf6BAXVwsA8BSmhbATJ05owoQJOnnypHr06KGZM2cqLi5OFotFWVlZ+vjjj7V161ZNnDhR27ZtU0xMjFmXBgDAJQzDkK3crmO2Ch2zVeiFz/cru6D8gudMGBCtK/p0Ud/oEE0cGK1gf54EAIDOzrSfBM8//7xOnjypRx55RC+++KL8/Oo/RPzyyy/rV7/6lebPn68XXnhBr776qlmXBgCgzZwsrtTOnEItSz+hrw7mq7iyplnn9Y0O1ht3j1JCt9A2rhAA0N6YFsI+//xzDRw4UH/84x8b3e/r66vf//73Wr58uZYtW0YIAwB4rKqaWr2zKVtfHTqtDYdPq9ZxgTnk/+OKPl0UFxWsbmH+Gtc/SmPjI1m7CwDQKFNvR/ze9753wWMsFotGjhypxYsXm3VZAABMVV3j0C/e3aHVB041eUyQn7ciQ/zk6+WliQO76tc3DZKPt5cLqwQAtGemhbCwsDAdPXr0oscdPXpUYWFhZl0WAABTVNXU6pMdx/TXr4/oyHcm1ogK8dfwXuGaPKS7Jg3qqshgP2YzBAC0mmkhLCkpScuWLdPy5cs1bdq0Ro/5/PPPtWHDBt1yyy1mXRYAgEuWXVCmn729Xd+eLKnXfllMmH590yCN7RspX0a6AAAmMS2EPfHEE/r88891++236/vf/75++MMfKi4uTpKUnZ2tDz74QB9++KG8vLz0xBNPmHVZAABa5bitQl8fytfn3+Tpq0P5Mr7z2NfMUbGad9tQBfh6u6dAAECHZepI2IIFC/Tzn/9c7733nt5///16+w3DUGBgoN544w2NHTvWrMt6lIULF+rVV1/V4cOHZbfbNXDgQP2///f/dNddd7m7NADotArLqrUr1yZbebVOFFVq3/Fi7cyx6ZitotHjJw6M1t1X9tH1g7u5uFIAQGdh6mIld999tyZOnKi//e1vWr9+vY4fPy5J6tGjh8aPH6/77rtPvXr1MvOSHqWwsFC33Xabhg8froCAAC1ZskQ/+MEPFBAQoNtuu83d5QFAp5FTUK7Pdh/TsvQTOpBXcvETdPa5r4evT9A9Y/u0cXUAgM7O9BUjY2Nj9dxzz5n9tu1CSkpKve3rr79eu3bt0nvvvUcIA4A2Vmmv1a6jNq3Ye1Jvb8pSTTOmle8dEaRhvay6Y2RPTUiIZkp5AIBLmB7CUF9kZKTsdru7ywCADunrYzX6v3e263hRhfafKJa9tvHgFR7oq37RwQoP9NWA7qEaFmvViN5WxYQHurhiAABMDGGrV6/Wa6+9pkceeUTjx49v9JivvvpKqampSklJ0TXXXGPWpZtl+/btWrlypbZs2aItW7bo2LFjks4+q3YhFRUVevHFF/Xhhx8qJydHERERuvHGGzVv3jz17Nmz0XNqampUXl6uf/3rX1q5ciXrogGASRwOQwVl1Tp4skQvbKnSYZtDUl6jx0YG++mHV/bWlCHdNaRHGFPKAwA8hmkh7I033tDKlSu1cOHCJo8ZPny4VqxYoYCAAJeHsHnz5unTTz9t0TmVlZWaNGmS0tLSFBMTo+nTpysrK0sLFizQsmXLlJaWpr59+9Y7Jy8vTzExMZIkb29v/elPf9LUqVNN+xwA0BmcLq3S8vQTyi4o18niSuUVVyqvqFKnSiqbHO2Szi6iPLJ3F42Jj9APxvRWdKi/C6sGAKB5TAthW7Zs0YgRIxQaGtrkMWFhYRo5cqQ2b95s1mWbLSkpSYmJiRo9erRGjx6tuLg4VVVVXfCc559/XmlpaUpKStKKFSsUEhIiSZo/f74effRRzZo1S2vXrq13TlRUlLZu3aqSkhJ98cUXmj17tiIjI/W9732vrT4aALR75dU1entTtr46mK9v80p0pry6wZTxF/LoDQM0YWC0BseEyYf1vAAAHs60EJaXl6errrrqosf16tVL27ZtM+uyzfb444+36Pjq6mq99tprkqTXX3/dGcAkac6cOVq4cKHWrVun7du3a9SoUc59Pj4+uuKKKyRJ1157rc6cOaNf/epXhDAA+I/jtgrtzLHpm2NF2pFdqNNlVTqSX9ai9wj191FcqEMxwV76r1vHamjP8DaqFgAA85kWwoKDg3Xy5MmLHnfq1CkFBASYddk2s2HDBhUVFalfv34aMWJEg/0zZsxQenq6li5dWi+Efdfw4cO1YMGCZl93yJAhjbZnZGSoe/fuWr16dbPfqy2UlZ39ouTuOtC+0Y86n6paQxk2h1Zk1yj9tKNZ54T7S8OivBUZaFEXf4u6BFhk9T/7OtBHKi8vlySd+na7Vn/bltWjI+PfI5iBftQ5lZWVKTg4uFXnmhbCRowYofXr1ysnJ0e9e/du9JicnBx9/fXXSkpKMuuybWb37t2SpJEjRza6v649PT39gu+zceNGxcXFmVobAHiyWoehwipDpysMnSgztONUrQ7bHKqqvfB5PhbpyhhvXdXDW5EBFkUFWuTFZBoAgA7ItBA2a9Ysffnll7r55pv15ptvOm/Jq7Nt2zbdd999stvtmjVrllmXbTM5OTmSzq571pi69uzsbGfbtddeq+9973saNGiQKisr9emnn+r999/XX//612Zfd+/evY22142QTZo0qdnv1RbqfsPj7jrQvtGPOh7DMLQ1q1CvrTmsjYdPX3SNrp7WQCV0C1Fiz3AlxlrVKyJIfSKDFODr3exr0o9gBvoRzEA/6pxaOwommRjCfvCDH+iTTz7RokWLdOWVV2rYsGHq16+fpLO30u3evVuGYej222/XPffcY9Zl20xpaakkKSgoqNH9dX/pJSUlzrZhw4bp1Vdf1dGjRxUcHKzBgwdr6dKluvnmm9u+YABwMYfD0PacQn3+zQl9sSdPJ4oqL3i8v4+XRsdF6O6xZ6eNZ8p4AEBnZepizR9++KFeeOEFzZ8/X7t27dKuXbuc+6xWqx555BH9+te/NvOSHiU1NVWpqammvZ/NZpPNZpMk2e12eXs3/zfEANAWDMPQ0vQTWnvglL4+fFr5JU3PMts11F+xXQLVv2uIJg3qpqR+kQoP9HVhtQAAeCZTQ5iXl5eeeuopPf7449q2bZuOHj0q6eyMiKNGjZKfn5+Zl2tTdbMh1j34/V11D2BeaEr+S5WamqrnnnvOuR0dHd1m1wKAi8k8XaYXP9+vFfuanoSpd0SQHr9xkK67rGuLbi0EAKAzMTWE1fH19VVSUlK7mICjKXWTi+Tm5ja6v669T58+bVZDSkqKkpOTJUmTJ09mJAyAyzgchtYePKVvcot1sqRSh0+VakvmmQbHeXtZdFW/SE0dGqPrL+uq6FB/bjMEAOAi2iSE1SkvL9fGjRt1+vRp9ezZU0lJSfLxadNLmmbYsGGSpB07djS6v649MTGxzWqwWq2yWq2SzgZbAGhrxZV2ZZwq1WMf71bGBdbu6hcdrJ9f0083DO6mLsHt5y4HAAA8QasTUUZGhhYvXqzLLrtMt9xyS4P9ixYt0s9//nPnM02S1LNnT73zzjuaMGFCay/rMuPGjVN4eLgyMjK0a9cuDR8+vN7+RYsWSVKjn90sPBMGwBXWHczXO5uytP9EiY7ZKi54bNdQf828Ilb/dV2C/H34NwkAgNZodQj78MMP9cwzz+i9995rsG/79u364Q9/qJqaGgUFBemyyy5TZmamcnNzdcstt2jv3r3q1avXJRXe1vz8/DR79mz99re/1UMPPaQVK1Y4Z0ScP3++0tPTNWHChAsu1HypeCYMQFtwOAztOV6klftOamvWGaUdaXibYZ0uQb6aMSpW3cMDdVn3UI2Jj5CPt5cLqwUAoONpdQj76quvFBgYqOnTpzfYN3fuXNXU1GjIkCFatWqVunXrJofDoQceeEB///vf9eqrr+p3v/vdJRXeUsuXL9e8efOc29XV1ZKksWPHOtuefvppTZs2zbn91FNPadWqVdq4caMSEhI0fvx4ZWdna/PmzYqOjtabb77ZpjXzTBgAs5RX1+jDLUe1PbtQm44U6ExZdZPHBvl5q3tYgEb26aLnbh2iYP/2cRs5AADtRat/sh4+fFhXXHGFAgMD67VXVFToiy++kMVi0e9+9zt169ZN0tmZE1NTU7V48WKtXLny0qpuhfz8fG3evLlB+/lt+fn59fYFBARozZo1evHFF/X+++9ryZIlioiIUHJysubNm9fkQs5m4ZkwAJeqptahA3kl+tU/v9E3x4qaPK5/1xD9clJ/jezdRbFdAplcAwCANtTqEHbq1KlGZz/ctm2b7Ha7QkNDdcMNN9TbFxQUpCuuuKLRMNTWkpOTnaNKLREYGKi5c+dq7ty55hcFAG1oz7EiJS/YqtOlDdfyslikYbFWXTMgWhMHRmt4rFVeXgQvAABcodUhrKampt6kG3W2b98uSRo1alSjMyFGR0erouLCD37jLCbmANBaZVU1mv3+jgYB7Mr4CCVfFaer+kUpPIgRdgAA3KHVIaxXr17atWuXamtr64WDL7/8UhaLpd6zVucrLCxkgolmYmIOAC1VXGnX6v2n9L+rDymr4Nxi812CfPXQtf1139Xx3GoIAICbtXqKq0mTJunEiRP1QsKGDRv0xRdfSJJuvfXWRs/buXNnmz9L1VGkpKQoMzNTmZmZSkhIUGRkpLtLAuChDMPQ0t3HNf7lNUr5aJeOnLfG1wMT+mnnM5P10/F9CWAAAHiAVo+EPfbYY3rnnXf029/+Vm+99Zaio6P1zTffqLa2VuPGjWt0JGzTpk3Ky8vT97///UsqurNgYg4ATfkmt0irD5zS6dIqnSqp1PbsQp0ubTjj4TUDojXnhgFuqBAAADSl1SGsf//++uSTT3TPPfcoNzdXubm5kqTBgwc3unaYdPb2OkmaMmVKay8LAJ3W+kOn9eHWHH2bV6JDp0oveOx1g7rq3qvidHX/KCbcAADAw1zS4i+TJ09Wdna21q9fr/z8fPXq1UtXXXWVvLwav8vx7rvv1g9/+ENNmjTpUi7baTAxB4CSSrsKSqv1t6+P6L3NORc9/obB3fTrmy5TfFSwC6oDAACtcckrcAYEBOj6669v1rG33HLLpV6uU2FiDqDjMwxDWQXlOlFUoQMnSnSyuFJ5xZXKLihXzpnyCy6qPCw2XKP6RMga5KsB3UI0pEe4ekUEubB6AADQGpccwtB2UlJSnGubTZ48mZEwoAP5JrdIn+0+pq8PndaBvJJmn9c3Kli/mNhPl8eGa1D3sDasEAAAtBVCmAdjYg6gY6mucWjp7uP6cGuOtmYVtujc3hFBGp8QpZTrByg61L+NKgQAAK5ACAMAF6i01+qef2xuMnx1CfLVoO5hiosKVnSIn3pFBKlPZLD6RAYpMthPPt6tXlEEAAB4GEIYALSh8uoaHSus0Auf728QwHqEB+iuMb01ZUh3Dewe6qYKAQCAqxHCPBizIwLt14miCr22+rD+b9tR2WuNevvO3laYoGGxVka4AADohAhhHozZEYH24/CpEn2y85jyiqpUUFaljRkFqq5xNDguqW+k3kweLV/CFwAAnRYhzIMxOyLg+cqra/T2pmz94d/fqsZhNHqMxSL1CA/UDYO76dHJAwhgAAB0cm0Swnbv3q0tW7bo9OnTGjJkiG699VZJUlVVlaqqqhQWxrTKzcHsiIDn+nL/Sb38xQEdPFna5DHdwvw1e1KC7rwiVv4+/BIFAACcZWoI+/bbb/WTn/xEmzdvdrbde++9zhD2/vvv66c//amWL1+uG2+80cxLA0CbK62q0V+/OqKV+05q/4niBvvDA331gzG9FRHsqz6RwZo4MJrwBQAAGjAthB09elTXXHON8vPzdeutt2r8+PH67//+73rH3HnnnXrwwQe1ePFiQhiAduPbvBIt3JSlJTuPqby6tsH+UH8fXTMwWo9NHqj4qGA3VAgAANoT00LY3Llzdfr0af3973/XrFmzJKlBCAsODtbw4cPrjZQBgCcyDENf7j+lv68/orQjZxo9pm9UsH4/c5iGxYYzyyEAAGg200LYF198ocTERGcAa0pcXJxWrFhh1mU7NKaoB1wvPdembVmF+teeE40urBzq76MfXNlbkwd30/BeTDEPAABazrQQdurUKY0bN+6ix9ntdpWXl5t12Q6NKeoB1ziSX6qPth7VuoP5OpBX0ugxQ3qE6d6r4nTrsB4K8OUXIgAAoPVMC2GRkZHKycm56HEHDx5UTEyMWZft0JiiHmh73+QW6a6/blJZI896SdLlPcP17K2DNbJ3F1ksFhdXBwAAOiLTQti4ceO0ZMkS7dq1S8OHD2/0mHXr1mnPnj3OYIELY4p6oG3Yax3alFGgNd+e0sfbchsEsIHdQjW2b4TGJ0Rr4sBobjkEAACmMi2EPfbYY/rkk080ffp0/eUvf9HkyZPr7V+9erWSk5Pl4+OjlJQUsy4LAC1y6GSJHnxvhw6dari+1z1j+yh5XJz6RYe4oTIAANBZmBbCrrzySv3v//6vHn74Yd18880KCgqSxWLR4sWL9cknn6i4uFgWi0V/+tOflJiYaNZlAaBZKu21+ttXR/T62sOqtDsa7H9wYj/9vxsHuaEyAADQ2Zi6WPODDz6okSNH6qWXXtLq1atlGIZKSkoUEBCgKVOm6Mknn2zW5B0AYJa9x4u0PP2ElqWfUM6Z+pMCDetl1dSh3TVxYLQGdQ9zU4UAAKCzMTWESdLYsWO1ZMkSGYah06dPy+FwKCoqikklALiEYRjae7xY6blF2pJZoE93H5dh1D8mKsRfv5txuSYN6uaeIgEAQKdmegirY7FYmFIdgEvU1Dp0prxaH2/L1ae7jungyYbPe0mSt5dF9ybFKeWGBIUFMNkNAABwD9NC2OHDh/X5559r0qRJGjp0aKPH7NmzR6tXr9bNN9+svn37mnVpAJ1UUbldc5ft06e7jqnGYTR5XN+oYN0+oqduHd5DfSKDXVghAABAQ6aFsNTUVL3xxhvKyMho8pjQ0FDNmTNHR44cUWpqqlmX7rBsNptsNpuks4tcc0sncFZ1jUOf7T6u55buVUllTaPHDO0Zpst7WnVFny66dXgP+TLNPAAA8BCmhbAvv/xSw4cPV+/evZs8pk+fPho+fLhWrlxp1mU7tNTUVD333HPObW7vRGdkGIZ25xZpR3ahtmSe0cFTJcouKFdtIyNf/j5eunFodz0woZ8ui2GiDQAA4JlMC2FHjx7VtGnTLnpcv3799MUXX5h12Q4tJSXFubD15MmTGQlDp1LrMPTBlhz9Y32mMk+XNXmcj5dFD07sp/vG91VYgI8sFosLqwQAAGg500KYt7e3qqqqLnpcVVWVamtrzbpsh2a1WmW1WiVJvr5MIoDOY82BU3rh8/2NLqhcx9vLoiE9wvTUtMEaEx/hwuoAAAAujWkhbMCAAVq/fr3Ky8sVFBTU6DHl5eVav369EhISzLosgA4m7UiBfvr2tga3GyZ0DdHo+AiN7x+l/l1D1CcyWH4+POcFAADaH9NC2IwZM/SrX/1KP/3pT/W3v/1NwcH1ZyArLy/X/fffr8LCQj3yyCNmXRZAB5FTUK5/7szVX9Zl1Atg04f30CPXD1BcFLMaAgCAjsG0EPbLX/5S77zzjj766COtWbNGP/jBD9SvXz9JUkZGhj744AOdOnVKAwcOVEpKilmXBdDOORyGUr88pFdXH6q3qLKvt0V/uXuUrruMBZUBAEDHYloICwoK0qpVq3T33Xdr9erVSk1NdT4gb/znm9W1116rd955p8EoGYDO4WRxpbZlFWpTTo3Kagy9f3SrtmUXylZur3dckJ+3/nL3KF0zgBlBAQBAx2NaCJOk7t27a9WqVdq6datWrVqlo0ePSpJ69eql66+/XqNHjzbzcgDaifRcmz7aelQfb8tVda3jvD2n6h0XFuCj6y/rpp9N6KtB3ZliHgAAdEymhrA6o0ePJnABnVhxpV05BeU6fKpUS3cf15cHTl3weItFmnZ5jOZOH6qIYD8XVQkAAOAebRLCAHReb6zL0O///a1qGllMWTo72tUr2KFgX4sS+/fW1QlRGt7LKmsQ4QsAAHQOpoewzMxMff311zpx4kST64ZZLBY9/fTTZl8agBudKq7U7//9rT7entvo/tAAH90/vq/uuzpemzd8JUmaNGmwK0sEAADwCKaFsOrqav30pz/Ve++9J+ncZByNIYQBHUtpVY1u/9NGHbNV1GvvFRGopL6RumZAtKYM6S5fb9b1AgAAMC2EPfPMM3r33XdltVp19913a8CAAQoNDTXr7QF4oPRcm3bnFul3XxxQSWWNs71LkK8W/GSMhveyuq84AAAAD2VaCHv//fdltVq1c+dO9enTx6y37dRsNptsNpskyW63y9vb270FATq7rlf2mXKlrjqoT3cdb7B/XP9Ivf7DkTzjBQAA0ATTQtipU6c0ZcoUApiJUlNT9dxzzzm3o6NZMwmuZRiGzpRV6/CpUu09Xqy1B/O1NfOMKuy1jR4/qHuo/vSjUQoP9HVxpQAAAO2HaSGM8GW+lJQUJScnS5ImT57MSBhcotZhKD3XprXf5uudtGydKau+4PEh/j6aNKirZl4Rq6v6Rcnby+KiSgEAANon00LYrFmz9MILLyg/P58RG5NYrVZZrVZJkq8vIwtoe4Zh6MH3tuvfe09e9Ni4yCDdMTJWP7umrwJ8+QUBAABAc5kWwv77v/9bO3fu1LXXXqtXX31VEydOlMXCb8SB9qK40q7XVx9uNICFBfhocI8w9e8aousu66YxcREK9meZQQAAgNYw7VtU//79JUnZ2dm6/vrr5evrq+7du8vLq+GU1BaLRRkZGWZdGkArGYahXUfP3nr49qYsFZbb6+2fc8MA3TW6l7qGBbipQgAAgI7HtBCWlZVVb7u6ulo5OTlmvT0Ak1TXOJR2pECHTpXq421HdSCvpMExPa2B+viBJPWwBrqhQgAAgI7NtBDmcDjMeisAbcQwDP307W366mB+o/uD/bx1d1IfPTihv8KDeA4RAACgLfBQB9CJfHOsqNEANrK3VVfEReinV8dz6yEAAEAbI4QBnUBRuV1bs85o3vJ99dp/nNRHP7qyjwZ2D3VTZQAAAJ1Pw1kzLtGKFSt0++23q2fPnvL399d9993n3Pfvf/9bc+bM0fHjx82+LIBG1DoMvfD5fo18fqV++vY2ZReUO/fNv3OY5k4fSgADAABwMVNHwh5++GG99tprMgxDISEhstvtMgzDuT8mJkapqanq1auXHnnkETMvDeA7MvJL9cTidG3NKmyw7/rLumn68J5uqAoAAACmjYS9/fbbevXVVzVq1Cjt2LFDxcXFDY5JTExUr169tHTpUrMuC6ARGw6f1rT//bpBAJswIFq/n5Gov9w9Ut5erOMHAADgDqaNhP35z3+W1WrV8uXLFR0d3eRxiYmJ+uabb8y6LIDz5BaW6/ll+/XF3rx67cNiw/Xe/WMVwgLLAAAAbmfaN7I9e/ZowoQJFwxgkhQeHq6TJ0+adVkA//FNbpEeeHe7jtkqnG1+Pl567tYhmjEqVr7epj8CCgAAgFYw9VuZxXLx25uOHz+uwMCOuwDs//3f/2natGmKiYlReHi4rrnmGq1fv97dZaGD+2z3cd3x5w31AlhEsJ9e/+FI/WBMbwIYAACABzHtm1lCQoJ27Nghu93e5DElJSXatWuXhgwZYtZlPU5qaqqioqL0+uuv6+OPP1bPnj113XXXaffu3e4uDR1UXlGlHl+ULnvtuUlwHpjQT5t/fZ1uGNzNjZUBAACgMaaFsJkzZ+rEiRN64oknmjzmV7/6lYqKinTXXXeZdVmPs3TpUi1cuFB33HGHJk+erPfee0/9+/fX66+/7u7S0AF9m1eiWW9tVYW9VpIU6u+jhbPG6Impgxj9AgAA8FCmPROWkpKiDz/8UKmpqdq4caOmT58uScrIyNArr7yiTz75ROvXr9fIkSN1//33m3VZjxMZGVlv28vLS0OHDlVmZqabKkJHtO94sRZuzNJH247Wa//1tMs0YcCFn8sEAACAe5kWwgIDA7Vq1SolJyfrX//6l7Zs2SJJ+vrrr/X1119Lkm644Qa9++678vPzM+uyLbJ9+3atXLlSW7Zs0ZYtW3Ts2DFJqreWWWMqKir04osv6sMPP1ROTo4iIiJ04403at68eerZ88JrLdXW1mrr1q2aMmWKaZ8DnZdhGHpzQ5bmLdvXYN9do3vprtG93FAVAAAAWsLU+aqjo6O1fPly7d69WytWrFBWVpYcDodiY2N1ww03aMyYMWZersXmzZunTz/9tEXnVFZWatKkSUpLS1NMTIymT5+urKwsLViwQMuWLVNaWpr69u3b5PmvvfaacnJy9OCDD15q+eikbOXV2p1bpNX7T2rV/lP1Jt+o89S0y3Tf1fHNmhwHAAAA7tUmiwYNGzZMw4YNa4u3viRJSUlKTEzU6NGjNXr0aMXFxamqquqC5zz//PNKS0tTUlKSVqxYoZCQEEnS/Pnz9eijj2rWrFlau3Zto+du3rxZTzzxhJ566ildfvnlZn8cdFCGYWhTRoG2ZRdqxb487TnWcOHzOrcM66GpQ7vrpstjXFghAAAALkWnWrn18ccfb9Hx1dXVeu211yRJr7/+ujOASdKcOXO0cOFCrVu3Ttu3b9eoUaPqnZuVlaXp06frlltu0W9+85tLLx6dgsNh6L8+3Kll6ScueNw1A6L1+xmJ6hYW4KLKAAAAYBbTQtjbb7/douN//OMfm3XpNrNhwwYVFRWpX79+GjFiRIP9M2bMUHp6upYuXVovhNlsNk2bNk1xcXFauHBhi24Ra2r6/oyMDHXv3l2rV69u+QcxUVlZmSS5vY6OalterZalVzdo9/WShkZ6aVi0ty6P8laXgFLt3bZRe91QoxnoRzAD/QhmoB/BDPSjzqmsrEzBwcGtOte0EJacnNyssGEYhiwWS7sIYXVre40cObLR/XXt6enpzrbq6mrdcccdKi8v1+rVqzv0wtQwl8Mw9M/D59bZC/aVxvf0UWKUlwZ08ZIXz3sBAAB0CKaFsGeeeabREOZwOHT06FGtW7dOmZmZSk5OVp8+fcy6bJvKycmRJMXGxja6v649Ozvb2fbggw9q3bp1+tvf/qbMzEzn1PT+/v6NjqZ91969jY9t1I2QTZo0qfkfoA3U/YbH3XV0NIdOluixRenKK6+UJFks0j8fukYDu4e6ubK2QT+CGehHMAP9CGagH3VOrR0Fk0wMYc8+++wF99fU1GjOnDn66KOPtHXrVrMu26ZKS0slSUFBQY3ur/uLLykpcbatWrVKDodD9913X71j+/Tpo6ysrLYpFO1WWVWNvjxwSk9+8o1KKmuc7bcO69FhAxgAAEBn57KJOXx8fPTKK6/os88+0xNPPKH333/fVZd2KTODls1mk81mkyTZ7XZ5e3ub9t5wL8MwtHjHMf12+T4Vltvr7RvbN0LP3dr4s4EAAABo/1w6O6K3t7dGjRqllStXuvKyrVY3G2J5eXmj++sewgwNbZsRi9TUVD333HPO7ejo6Da5DlyrqqZWP39nu9Z+m99g37zbhuruK3uz3hcAAEAH5uXqC+bl5TnDi6fr3bu3JCk3N7fR/XXtbfWMW0pKivO5soSEBEVGRrbJdeBaCzdm1Qtgft5euv6ybnr3vit1z9g+BDAAAIAOzmUjYQ6HQ6+//ro2bdqkMWPGuOqyl6RuwekdO3Y0ur+uPTExsU2ub7VaZbVaJUm+vr5tcg24lsNh6N20HOf2uP6Revl7iYrt0vhzhwAAAOh4TAthF5oNprS0VJmZmTpz5oy8vLzazeLF48aNU3h4uDIyMrRr1y4NHz683v5FixZJkm655ZY2uT7PhHU8/96bp5wzZ29v9bJIv5sxTD2tLGMAAADQmZgWwtauXXvhC/n46Oqrr9Yzzzyj6667zqzLtik/Pz/Nnj1bv/3tb/XQQw9pxYoVzhkR58+fr/T0dE2YMKHeQs1m4pmwjqOqplZzl+7Te5vPjYJdd1k3AhgAAEAnZFoIq1sPqzF+fn6Kiopy+y11y5cv17x585zb1dXVkqSxY8c6255++mlNmzbNuf3UU09p1apV2rhxoxISEjR+/HhlZ2dr8+bNio6O1ptvvtlm9aakpCg5OVmSNHnyZEbC2qmiCrse/nBng4k4fpzUPtbLAwAAgLlMC2HtYQHm/Px8bd68uUH7+W35+fW/KAcEBGjNmjV68cUX9f7772vJkiWKiIhQcnKy5s2b1+RCzmbgmbD27aOtOXp19WHlFlY02PfI9QN0df8oN1QFAAAAd3PpFPXulpyc7BxZaonAwEDNnTtXc+fONb+oC+CZsPbr8KkSPfnJHtU4jHrtfSKDtPSXVyssgFANAADQWbU6hOXk5Fz8oAuom/4dTeOZsPbrf748XC+A+Xl76afj4/WLif0USgADAADo1FodwuLi4lq9npHFYlFNTU1rL91p8ExY+1RRXatV+046tx+Y0E+/mNhP4YGELwAAAFxCCLvmmmtYVLaN8UxY+1JSade2rEIt2JilCnutJCnU30eP3JAgfx8CNAAAAM5qdQi72JT0QGeyLP24frX4G5VU1R/hvXlYDAEMAAAA9bQ6hM2aNUtXX321Zs2aJensM2IhISGKiIgwrTjAkx2zVejTXceUduSMvjqY32D/kB5h+n9TBrmhMgAAAHiyVoewt956S5KcISw+Pl7Jycn6xz/+YUphYHZET1ZaVaM7/7JJx2z1p5/3skgTB3bV1KHddevwHoyCAQAAoIFWhzBfX19VVlY6tw3DkGEYFzgDLcXsiJ4rdeXBBgEsMTZcr/1gpHpHBrmpKgAAALQHXq09sVevXvr666+VnZ1tZj04T0pKijIzM5WZmamEhARFRka6uyRI+upgvv6+PtO53SXIVy/dcbkWPXAVAQwAAAAX1eoQ9sMf/lDHjh1T3759nbfJLVy4UN7e3hf94+PTqdaIbjWr1aq4uDjFxcXJ19dXXl6t/s8FkxSUVumnb29zbncPC9CmX12nu8b0lp8P/30AAABwca1OQ88++6ysVqs+/fRT5ebmKjMzU0FBQYqKijKzPsCj7D1erOoah3P72VsHK8CX574AAADQfK0OYV5eXpozZ47mzJnj3J45c6befPNN04oDPM3RwnLn637RwbpxaIwbqwEAAEB7ZNp9gffee6+uvvpqs94OYnZET3T0zLnJOIb36uLGSgAAANBemRbCFixYYNZb4T+YHdHznD8S1isi0I2VAAAAoL1qkxkyampqVFBQoKqqqiaP6d27d1tcukNJSUlRcnKyJGny5MmMhLnZiaIKLU8/4dzu1YWZEAEAANBypoawVatW6fnnn1daWprsdnuTx1ksFtXU1Jh56Q7JarXKarVKOrsuG9zricXf1NvuFUEIAwAAQMuZFsKWLVum22+/XbW1terSpYvi4+MVGhpq1tsDbpWea9O6g/nO7YhgPw3vZXVfQQAAAGi3TAthzz33nBwOh1555RXNnj2bW+fQoSzenltve+MTk1gXDAAAAK1iWgjbu3evkpKS9PDDD5v1loBHqKl1aPk3554Fe/l7l7M2GAAAAFrNtBAWEhLCZBsmY4p6z7A1q1CnS6slSb7eFk0Z0t3NFQEAAKA9M+1+quuvv17btm0z6+2gs1PUx8fHKz4+XocOHVJBQYG7S+qU/r03z/l6XP8oWYP83FgNAAAA2jvTQtjLL7+s4uJiPf7448x8aJKUlBRlZmYqMzNTCQkJioyMdHdJnY7DYeiLPedC2I2MggEAAOASmbpY89SpU/WHP/xBixcv1sSJExUbGysvr4Y5z2Kx6Omnnzbr0h0WU9S73+5cm/KKKyVJXhbp+sHd3FwRAAAA2jvTQtizzz4ri8UiwzB05MgRHTlypMljCWFoL74471bE0XERigrxd2M1AAAA6AhMHQkDOpJKe63eWHfulwlTh3IrIgAAAC6daSHs3nvvNeutALfbfKRA97y5pV7bZJ4HAwAAgAlYbRb4jvySKv3ivR2qrnE4267o00U9rIFurAoAAAAdhWkjYUBHsCOnUD9ZsFVFFXZn2+TB3fT/bhzkxqoAAADQkbQ6hE2aNKnVF7VYLPryyy9bfT7QFr7Yk6cH39suh3Gube70IfpxUpzbagIAAEDH0+oQtnbt2lZf1GKxtPrczsRms8lms0mS7Ha7vL293VtQB1ZT69BvP99XL4BNHBite8b2cV9RAAAA6JBaHcLWrFljZh1oRGpqqp577jnndnR0tBur6djWfJuvo2cqnNvP3zZUM0bF8gsDAAAAmK7VIWzChAlm1oFGpKSkKDk5WZI0efJkRsLa0PpD+c7XU4d2192MgAEAAKCNMDGHB7NarbJarZIkX19f9xbTgTkchhZuynZuj+sf5cZqAAAA0NG16RT1L7/88iVN4AG4QspHu+ptXxHXxT2FAAAAoFNo05GwAwcOaN26dW15CaDVHA5DTy7Zo892H3e2RYX4KaFrqBurAgAAQEfHYs3otF781359sCWnXtv8O4fL24vJOAAAANB2eCYMnY691qEnP/lG/7ct19kWHeqvJQ+NU09roBsrAwAAQGdACEOnYhiGHnpvh1bsO+ls8/W26J+/uIoABgAAAJdo9e2IkyZN0u9+9zvn9ldffaWDBw/WO2bQoEG65pprWl8dYLIdOYX1Alj3sAC9+oMR6hUR5MaqAAAA0Jm0OoStXbtWBw4ccG5PnDhRL7/8cr1jHn/8cRZ1hkf5Nq/U+Xpgt1Ct/e+JunFojBsrAgAAQGfT6hDm5+ensrKyem2GYVxyQUBbyj5zrs9eHhuuAF8WwAYAAIBrtfqZsP79++vLL7/UunXrFB8fL0kqLS1VTk7ORc48q3fv3q29NNBqOQXlztd9uAURAAAAbtDqEPazn/1MKSkp9RZjXrx4sRYvXnzRcy0Wi2pqalp7aaBVqmsc+teePOd270hCGAAAAFyv1SHsv/7rvxQbG6tPP/1Uubm5WrNmjbp27apBgwaZWV+nZrPZZLPZJEl2u13e3tw61xpF5Xb99esMLdqeW6+9NyNhAAAAcINLmqL+jjvu0B133CFJ8vLy0tSpU/Xmm2+aUhik1NRUPffcc87t6OhoN1bTfj3w7nZtOlJQry00wEcJ3ULdVBEAAAA6s1ZPzPFdv/nNb3TbbbeZ9XaQlJKSoszMTGVmZiohIUGRkZHuLqndqbTXanNm/QB2ZXyE3vrJaIX4s0weAAAAXM+0b6G/+c1vzHor/IfVapXVapUk+fr6ureYdurwqVI5zpu0c9kvr9bQnuHuKwgAAACdnmkjYYAnOniyxPn6spgwAhgAAADcrtUjYZcySQSzI8JVNmacuxVxUHeeAQMAAID7tTqE9erVSxaLxcxaAFPtPV6kf+44NyPikB5hbqwGAAAAOKvVISwrK8vEMgDzvbc5x/k8WI/wAM0c1cu9BQEAAADimTB0UFU1tVqefsK5/cgNAxQexOQmAAAAcD9CGDqkLZlnVFRhlyQF+Hrppstj3FwRAAAAcJbpIWzFihW6/fbb1bNnT/n7++u+++5z7vv3v/+tOXPm6Pjx42ZfFpAkFZRW6fU1h3XPP7Y4267uH61g1gQDAACAhzD1m+nDDz+s1157TYZhKCQkRHa7XYZxbpGmmJgYpaamqlevXnrkkUfMvDSg7IIyzfjLJuWXVNVrnzSoq5sqAgAAABoybSTs7bff1quvvqpRo0Zpx44dKi4ubnBMYmKievXqpaVLl5p1WUCSVGmv1c/f2d4ggHUPC9A0bkUEAACABzFtJOzPf/6zrFarli9frujo6CaPS0xM1DfffGPWZQFJ0j93HNOBvHMLM88cFaspQ7orqV8ktyICAADAo5j27XTPnj2aMGHCBQOYJIWHh+vkyZNmXRaQJH1zzOZ8fdPl3fX7mcPcVwwAAABwAaZOzNGcxZuPHz+uwMBAMy/rUbZt26Yf//jH6t+/vywWi5566il3l9QpZJ4uc75O6hflxkoAAACACzMthCUkJGjHjh2y2+1NHlNSUqJdu3ZpyJAhZl3W42zYsEFpaWm6+uqrFR4e7u5yOo3zQ1h8ZLAbKwEAAAAuzLQQNnPmTJ04cUJPPPFEk8f86le/UlFRke666y6zLutxfvnLX+rgwYN66623ZLVa3V1Op1BWVaOTxecm5IiLCnJjNQAAAMCFmfZMWEpKij788EOlpqZq48aNmj59uiQpIyNDr7zyij755BOtX79eI0eO1P3332/WZT2OlxfrX7vasvRz6875+XipR3jHvd0VAAAA7Z9piSEwMFCrVq3SjTfeqM2bN+vJJ5+UJH399dd69NFHtX79et1www3617/+JT8/P7Mu2yLbt2/XSy+9pDvuuEOxsbGyWCzNeo6toqJCzzzzjAYMGKCAgAD16NFDs2bN0rFjx1xQNS5ka9YZPbVkj3M7qW+kvLwu/t8UAAAAcBdT5+6Ojo7W8uXLtXv3bq1YsUJZWVlyOByKjY3VDTfcoDFjxph5uRabN2+ePv300xadU1lZqUmTJiktLU0xMTGaPn26srKytGDBAi1btkxpaWnq27dvG1WMC8krqtTP39kue+3ZBcEjg/30/G1D3VwVAAAAcGFtsoDSsGHDNGyY500RnpSUpMTERI0ePVqjR49WXFycqqqqLnjO888/r7S0NCUlJWnFihUKCQmRJM2fP1+PPvqoZs2apbVr17qgenzXy18c0JmyakmSn7eX3rhnlHpF8DwYAAAAPFubr2JbWVkpm82mqKgo+fi4d9Hcxx9/vEXHV1dX67XXXpMkvf76684AJklz5szRwoULtW7dOm3fvl2jRo0ytVZcWG5huT7Zee520KdvGawr4iLcWBEAAADQPK1ORSUlJdq/f7+sVqsGDBjQYP+hQ4c0e/ZsrVmzRrW1tfLz89P06dP1yiuvKCYm5pKKdpUNGzaoqKhI/fr104gRIxrsnzFjhtLT07V06VLTQlhT0/dnZGSoe/fuWr16tSnXaa2ysrNTwbu7jlU5Nc7XXQMtiik7rNWrM9xYEVrCU/oR2jf6EcxAP4IZ6EedU1lZmYKDW7c0Uqsn5nj77beVlJSkZcuWNdiXl5en8ePHa9WqVaqpqZFhGKqqqtLHH3+sa6+9VhUVFa29rEvt3r1bkjRy5MhG99e1p6enu6wmnLU7v9b5emQ3b3k1Y4IVAAAAwBO0eiRs3bp18vLy0t13391g37x583Tq1ClFRERowYIFmjRpkg4dOqSf//zn2rZtm/785z9rzpw5l1S4K+Tk5EiSYmNjG91f156dne1sy8/P17p16yRJ5eXlOnDggBYtWqTg4GBNnTr1otfcu3dvo+11I2STJk1q/gdoA3W/4XFnHSWVdh38cqVze9aU0RoTz62I7Ykn9CO0f/QjmIF+BDPQjzqn1o6CSZcQwtLT03X55Zera9eu9dodDoc++OADWSwWvfDCC7rlllskScOHD9c///lP9e3bV0uWLGkXIay0tFSSFBTU+GQPdX/xJSUlzra9e/dq5syZzu3Fixdr8eLF6tOnj7Kystqu2E7kq4OnnTMidgny1cjeVvcWBAAAALRAq0PYqVOnNGXKlAbt6enpstls8vX11V133VVvX8+ePXXllVdq//79rb2sx5s4caIMwzDlvWw2m2w2myTJbrfL29vblPdt7zZmnHa+vnZgV/l4s0A2AAAA2o9Wf3stKyuT3W5v0L59+3ZJUmJiosLCwhrsj42NVXFxcWsv61J1syGWl5c3ur/uIczQ0NA2uX5qaqri4+MVHx+vQ4cOqaCgoE2u096cKKp0vh7co2EfAwAAADxZq0NYdHS0Dhw40KB9/fr1slgsTS7MXFlZqfDw8NZe1qV69+4tScrNzW10f117nz592uT6KSkpyszMVGZmphISEhQZGdkm12lvTpWcC2HdwgLcWAkAAADQcq0OYXW3FS5dutTZlp+fr3/+85+SpMmTJzd63r59+9SjR4/WXtal6hac3rFjR6P769oTExPb5PpWq1VxcXGKi4uTr6+vvLy47U6SThafW2CbEAYAAID2ptXf6mfPni3DMDRz5kz9+Mc/1qOPPqrRo0erpKREvXr10rRp0xqcc+TIER08eNAZbjzduHHjFB4eroyMDO3atavB/kWLFkmSc/IRs9lsNmVlZSkrK0t2u10Oh6NNrtOe1NQ6dLr0/BDm78ZqAAAAgJZrdQi79tpr9eyzz8put+vdd99VamqqcnJyFBgYqAULFsjHp+GcH3/5y18kqdEJPTyRn5+fZs+eLUl66KGHnM+ASdL8+fOVnp6uCRMmmLZQ83fxTFhDp0urdf68J11DGQkDAABA+9Lq2REl6ZlnntHNN9+sf/7zn8rPz1evXr30ox/9SPHx8Y0e7+/vr4cfflg33njjpVy21ZYvX6558+Y5t6urqyVJY8eOdbY9/fTT9UbxnnrqKa1atUobN25UQkKCxo8fr+zsbG3evFnR0dF6880326zelJQUJScnSzp7eyezI0qHTp1bDiA0wEeBfvydAAAAoH25pBAmSSNHjtTIkSObdez5Acgd8vPztXnz5gbt57fl5+fX2xcQEKA1a9boxRdf1Pvvv68lS5YoIiJCycnJmjdvXpMLOZvBarXKarVKknx9fdvsOu1FWVWN7vnHFuc2z4MBAACgPbrkENaeJCcnO0eWWiIwMFBz587V3LlzzS/qAlgn7JzCsmo99emeem2j47q4qRoAAACg9TpVCGtvUlNT9dxzzzm3o6Oj3ViN+2SdLtPMNzYpv+TchBzjE6L0m1uGuLEqAAAAoHWY89yDsU7YWSkf7aoXwEbHddFf77lCAb6dd2QQAAAA7RcjYR6MZ8LOLsy866jNuf2za/rq0ckD5O9DAAMAAED7RAiDR9tzrMj5OjrUX7++6TI3VgMAAABcOm5HhEdLzz0XwobFWt1XCAAAAGASRsI8GLMjSnuPFztfX94z3I2VAAAAAOZgJMyDpaamKj4+XvHx8Tp06JAKCgrcXZLL5RVVOl/3jQ52YyUAAACAOQhhHozZEc9OzFEnOtTfjZUAAAAA5uB2RA/W2WdHdDgMnS6tdm4TwgAAANARMBIGj1VYXq1ah+Hc7koIAwAAQAdACIPHOnXeAs0Bvl4K8WfgFgAAAO0f32o9WGefHTH/vBDWNTRAFovFjdUAAAAA5mAkzIN19tkRd+bYnK95HgwAAAAdBSHMg3Xm2RE/2pqjV1YddG7Hdgl0YzUAAACAebgd0YN11tkRl6Uf1+OLv3Fu+3l76Sfj4t1YEQAAAGAeRsLgUfafKNbs93c6t/28vbTgJ6M1vJfVfUUBAAAAJiKEwaNszKj/3Ntvbx+qcf2j3FQNAAAAYD5CGDzKscIK5+suQb6aMSrWjdUAAAAA5uOZMA/WGaeoP2Yrd77+6fi+TEsPAACADoeRMA/WGaeoP2Y7NxLGjIgAAADoiAhhHqwzTlF//u2IPa2EMAAAAHQ83I7owTrbFPWnSipVWG53bvdkJAwAAAAdECNh8AjpuTb96G+bndvWIF91DQ1wY0UAAABA22AkDG539Ey5Zvx5k6prHc62X0zoJ28vJuUAAABAx8NIGNxu/eHT9QLYtMQY/WRcvBsrAgAAANoOI2Fwuz3Hipyvx/aN0Os/HOnGagAAAIC2xUgY3G7P8WLn61uH9XRjJQAAAEDbI4TBrapqanXgxLkQNrRnmBurAQAAANoetyN6MJvNJpvNJkmy2+3y9vZ2b0FtYEe2TVU1Z58HC/D10sDuoW6uCAAAAGhbjIR5sNTUVMXHxys+Pl6HDh1SQUGBu0sy3deH8p2vr4yPlL9PxwuaAAAAwPkIYR4sJSVFmZmZyszMVEJCgiIjI91dkqkMw9CKfSed2+MTotxYDQAAAOAa3I7owaxWq6xWqyTJ19fXvcW0gfTcIh0+Vercnjy4uxurAQAAAFyDkTC4zeIduc7XY+Ii1DsyyI3VAAAAAK5BCINbVNc49Nnu487t741ianoAAAB0DoQwuJxhGFqy85hs5XZJkr+Pl6ZeHuPmqgAAAADX4JkwuFRJpV3ffyNN+85bG2zKkO4KC+h4z7wBAAAAjWEkDC61PP1EvQAmSd8bFeumagAAAADXI4TBpY7bKuptT0uM0fj+TE0PAACAzoPbEeFS+aXVztc/TuqjudOHurEaAAAAwPUYCYNL5ZdUOV/36sKU9AAAAOh8CGFwqdOl50JYVKifGysBAAAA3IMQBpc6fyQsOiTAjZUAAAAA7sEzYR7MZrPJZrNJkux2u7y9vd1b0CUyDIORMAAAAHR6jIR5sNTUVMXHxys+Pl6HDh1SQUGBu0u6JCVVNaqqcTi3o0P83VgNAAAA4B6EMA+WkpKizMxMZWZmKiEhQZGRke4u6ZKcPu9WRC+L1CWIkTAAAAB0PtyO6MGsVqusVqskydfX173FmMBWYXe+tgb5ycvL4sZqAAAAAPdgJAwuYys/t0aYNaj9h0oAAACgNQhhcJnCsvNGwgIJYQAAAOicCGFwmfNvR+R5MAAAAHRWhDC4TP3bEQlhAAAA6JwIYXCZQp4JAwAAAAhhcB1b+fm3IxLCAAAA0DkRwuAy54cwbkcEAABAZ0UIg8ucfzsiE3MAAACgsyKEwSVqah06eqbcud0lmNsRAQAA0DkRwuASGzIKVFxZI0ny8/ZSYqzVvQUBAAAAbkIIQ5srqbTrV4vTndtj+0UqxN/HjRUBAAAA7kMIM9muXbs0fvx4BQYGKj4+Xq+99pq7S3K7tzZk6XhRpXN75qhYN1YDAAAAuBfDESbKz8/XDTfcoDFjxmjZsmXasWOHUlJSFB4ernvuucfd5bnFF3vy9MeVB53btw3voZsTY9xYEQAAAOBehDAT/eUvf5HFYtHHH3+soKAgXXfddcrMzNS8efM6ZQgzDEO//XxfvbY5NwyUxWJxU0UAAACA+3E7oon+/e9/66abblJQUJCzbebMmTp06JCOHDnixspc69u8Et3+pw2K/9XnOnqmwtk++9r+6h0ZdIEzAQAAgI6v04Sw7du366WXXtIdd9yh2NhYWSyWZo3IVFRU6JlnntGAAQMUEBCgHj16aNasWTp27FiDYw8ePKhBgwbVa6vb/vbbb835IO3Ac0v3ameOrV7b2L4RemzKQPcUBAAAAHiQTnM74rx58/Tpp5+26JzKykpNmjRJaWlpiomJ0fTp05WVlaUFCxZo2bJlSktLU9++fZ3HFxYWymq11nuPLl26OPd1BkfPlGvTkYJ6bX7eXnr4ugFuqggAAADwLJ0mhCUlJSkxMVGjR4/W6NGjFRcXp6qqqgue8/zzzystLU1JSUlasWKFQkJCJEnz58/Xo48+qlmzZmnt2rUuqL59qKl16Jcf7JRhnGv7/L/GKzYiUGEBLM4MAAAASJ0ohD3++OMtOr66uto5vfzrr7/uDGCSNGfOHC1cuFDr1q3T9u3bNWrUKElnR72KiorqvY/NZnPu6+j+vfekdh21Obd/NXWQBvcIc19BAAAAgAfqNCGspTZs2KCioiL169dPI0aMaLB/xowZSk9P19KlS50hbMCAATpw4EC94+q2Bw5s3vNQQ4YMabQ9IyND3bt31+rVq1vyMUxXVlYmSY3W8T9bz40sXh7lpX41WVq9OttltaH9uFA/ApqLfgQz0I9gBvpR51RWVqbg4OBWndtpJuZoqd27d0uSRo4c2ej+uvb09HRn25QpU/T555+rouLcjICLFi1SQkJCvWfHOqLSakOHCh3O7Vv7+sqLqegBAACABhgJa0JOTo4kKTY2ttH9de3Z2edGeh544AH97//+r+68806lpKRo586deuONN/Tmm282+7p79+5ttL1uhGzSpEnNfq+2UPcbnu/W8dnu4zK0U5IUHeqvn95+HeuBoUlN9SOgJehHMAP9CGagH3VOrR0FkxgJa1Jpaakk1Vvz63x1f+klJSXOtujoaK1cuVKFhYWaNm2aXn31Vc2fP79TLNT82a5zU/ZfkxBNAAMAAACawEiYyYYPH67169eb8l42m805sYfdbpe3t7cp72u2XUdtWrX/lHN7WmJ3N1YDAAAAeDZGwppQNxtieXl5o/vrHsAMDQ1tsxpSU1MVHx+v+Ph4HTp0SAUFBRc/yQ2e/ezcLZRDe4bp2oFd3VgNAAAA4NkIYU3o3bu3JCk3N7fR/XXtffr0abMaUlJSlJmZqczMTCUkJCgyMrLNrtVaeUWV9aalf/KmwdyKCAAAAFwAtyM2YdiwYZKkHTt2NLq/rj0xMbHNarBarbJarZIkX1/PXOz4eNG5mSDDAnyU1M/zgiIAAADgSRgJa8K4ceMUHh6ujIwM7dq1q8H+RYsWSZJuueWWNqvBZrMpKytLWVlZstvtcjgcFz/JxU4VVzpfx4QHurESAAAAoH0ghDXBz89Ps2fPliQ99NBDzmfAJGn+/PlKT0/XhAkTnAs1t4X28ExYXtG5ENYtPMCNlQAAAADtQ6e5HXH58uWaN2+ec7u6ulqSNHbsWGfb008/rWnTpjm3n3rqKa1atUobN25UQkKCxo8fr+zsbG3evFnR0dEtWv+rNVJSUpScnCxJmjx5skfOjniypMr5uluovxsrAQAAANqHThPC8vPztXnz5gbt57fl5+fX2xcQEKA1a9boxRdf1Pvvv68lS5YoIiJCycnJmjdvXpMLOZulPTwTdvK8kbDujIQBAAAAF9VpQlhycrJzVKklAgMDNXfuXM2dO9f8oi6iPawTdrLkvNsRwwhhAAAAwMXwTJgHaw/PhOWfdztiV25HBAAAAC6KEObB2sM6YbZyu/N1l2A/N1YCAAAAtA+d5nbE9qg9PBNWVHEuhIUHemaNAAAAgCdhJAytVmmvVVXNubXLCGEAAADAxTES5sE8fWKO80fBJEIYAAAA0ByMhHkwT5+Y4/wQ5ufjpQBfzwqJAAAAgCcihHkwT5+Yg+fBAAAAgJbjdkQP5ukTcxSVE8IAAACAlmIkDK12/kiYlRAGAAAANAshDK3G7YgAAABAyxHC0GqEMAAAAKDleCbMg3n6FPUFZVXO1+FBhDAAAACgORgJ82CePkV91uly5+veEUFurAQAAABoPwhhHszTp6jPKihzvo6LCnZjJQAAAED7we2IHsyTp6ivqqnVcVuFczsukhAGAAAANAcjYWiVo2cq5DDOvvb2sii2S6B7CwIAAADaCUIYWuVo4bnnwXpaA+XrTVcCAAAAmoNvzmiVM6XVztfRof5urAQAAABoX3gmzIN58hT1heXnQlgXpqcHAAAAmo2RMA/myVPUnx/CrEF+bqwEAAAAaF8IYR7Mk6eoLyy3O19HBBPCAAAAgObidkQP5slT1NvqjYR5Vm0AAACAJ2MkDK1ypuz8Z8IYCQMAAACaixCGVrGddzsiIQwAAABoPkIYWqy42tCBvBLnNrMjAgAAAM1HCEOLHCt16L+/qqzX1oWJOQAAAIBmI4ShRXadqpXdcW47yM9bPayB7isIAAAAaGcIYWiRakf97fl3DleIP5NsAgAAAM3Ft2cPZrPZZLPZJEl2u13e3t7uLUhSrXHu9Q/G9NaNQ7u7rxgAAACgHWIkzIOlpqYqPj5e8fHxOnTokAoKCtxdkmrOGwnz9ba4rxAAAACgnSKEebCUlBRlZmYqMzNTCQkJioyMdHdJqnWcGwrz9ab7AAAAAC3F7YgezGq1ymq1SpJ8fT1jGvjzb0ckhAEAAAAtx7dotAi3IwIAAACXhhCGFqlhJAwAAAC4JHyLRovU1hsJo/sAAAAALcW3aLRIjXH+xBzcjggAAAC0FCEMLcJIGAAAAHBp+BaNFmF2RAAAAODS8C0aLXL+7Ig+3I4IAAAAtBghDC1y/kiYHyNhAAAAQIvxLRotUuM4l8IYCQMAAABajhCGFqlhYg4AAADgkvi4uwA0zWazyWazSZLsdru8vb3dW5C4HREAAAC4VHyL9mCpqamKj49XfHy8Dh06pIKCAneXVG+Kem5HBAAAAFqOEObBUlJSlJmZqczMTCUkJCgyMtLdJamGKeoBAACAS8LtiB7MarXKarVKknx9fd1bzH/UnjcxByEMAAAAaDm+RaNF6o+EcTsiAAAA0FKEMLQIsyMCAAAAl4Zv0WiRWp4JAwAAAC4J36LRIrX1RsK4HREAAABoKUIYmq3WYei8gTBGwgAAAIBW4Fs0ms1+/jCYCGEAAABAa/AtGs1W3SCEcTsiAAAA0FKEMDRbzfmzcoiRMAAAAKA1+BaNZuN2RAAAAODS8S0azVZdw+2IAAAAwKUihKHZahznbkf08bLIYiGEAQAAAC1FCDPZtm3b9OMf/1j9+/eXxWLRU0895e6STHP+7YjciggAAAC0Dt+kTbZhwwalpaXp6quvVnh4uLvLMdX5tyP6cCsiAAAA0Co+7i6go/nlL3+phx9+WJIUFxfn3mJMVmGvdb4O8PV2YyUAAABA+8VImMm8vDruX+npkirn68hgPzdWAgAAALRf7TYxbN++XS+99JLuuOMOxcbGymJp3kQRFRUVeuaZZzRgwAAFBASoR48emjVrlo4dO+aCqtu3/NJzISw61N+NlQAAAADtV7u9HXHevHn69NNPW3ROZWWlJk2apLS0NMXExGj69OnKysrSggULtGzZMqWlpalv375tVHH7d/5IGCEMAAAAaJ12G8KSkpKUmJio0aNHa/To0YqLi1NVVdUFz3n++eeVlpampKQkrVixQiEhIZKk+fPn69FHH9WsWbO0du1a5/E2m015eXkXfM+goCD17t37kj9Pe1BvJCyEEAYAAAC0RrsNYY8//niLjq+urtZrr70mSXr99dedAUyS5syZo4ULF2rdunXavn27Ro0aJUn68MMP9Ytf/OKC7zthwoR6wa0jy2ckDAAAALhk7TaEtdSGDRtUVFSkfv36acSIEQ32z5gxQ+np6Vq6dKkzhD3wwAN64IEHXFrnkCFDGm3PyMhQ9+7dtXr1apfWU6+GY5XO16dyDmv16iy31YL2raysTJLc2p/R/tGPYAb6EcxAP+qcysrKFBwc3Kpz2+3EHC21e/duSdLIkSMb3V/Xnp6e7rKa2pvi6nOvw/xYJwwAAABojU4zEpaTkyNJio2NbXR/XXt2dvYlXSc/P1/r1q2TJJWXl+vAgQNatGiRgoODNXXq1Iuev3fv3kbb60bIJk2adEn1tZZhGCr58gtJhiRp8jVjNaBbqFtqQftX95tCd/VndAz0I5iBfgQz0I86p9aOgkmdKISVlpZKOjuRRmPq/hJLSkou6Tp79+7VzJkznduLFy/W4sWL1adPH2VlZV3Se7tTVY1D1w6K1uHcfBVVG0zMAQAAALRSpwlhrjJx4kQZhmHKe9lsNtlsNkmS3W6Xt7e3Ke/bGgG+3nrjniucv+npwmLNAAAAQKt0mmfC6mZDLC8vb3R/3QOVoaGec4tdamqq4uPjFR8fr0OHDqmgoMDdJQEAAAC4RJ0mhNWt5ZWbm9vo/rr2Pn36uKymi0lJSVFmZqYyMzOVkJCgyMhId5cEAAAA4BJ1mtsRhw0bJknasWNHo/vr2hMTE11W08VYrVZZrVZJkq+vr3uLAQAAAGCKTjMSNm7cOIWHhysjI0O7du1qsH/RokWSpFtuucXFlTXNZrMpKytLWVlZstvtcjgc7i4JAAAAwCXqNCHMz89Ps2fPliQ99NBDzmfAJGn+/PlKT0/XhAkTnAs1ewKeCQMAAAA6nnZ7O+Ly5cs1b94853Z19dmVhMeOHetse/rppzVt2jTn9lNPPaVVq1Zp48aNSkhI0Pjx45Wdna3NmzcrOjpab775pus+QDOkpKQoOTlZkjR58mS3zo4IAAAAwBztNoTl5+dr8+bNDdrPb8vPz6+3LyAgQGvWrNGLL76o999/X0uWLFFERISSk5M1b968JhdydheeCQMAAAA6nnYbwpKTk52jRC0RGBiouXPnau7cueYXZTJPWicMAAAAgDk6zTNh7RHPhAEAAAAdDyHMg7FOGAAAANDxtNvbETsDngkDAAAAOh5GwgAAAADAhRgJ82BMzAEAAAB0PIyEeTAm5gAAAAA6HkKYB2NiDgAAAKDj4XZED8bEHAAAAEDHw0gYAAAAALgQIQwAAAAAXIgQBgAAAAAuxDNhHowp6gEAAICOh5EwD8YU9QAAAEDHYzEMw3B3EWjc+SNhQ4cOld1uV//+/d1aU1lZmSQpODjYrXWgfaMfwQz0I5iBfgQz0I86p4yMDPn6+qqkpKTF5zIS5sGsVqvi4uIUFxenkJAQ+fn5ubsk5eXlKS8vz91loJ2jH8EM9COYgX4EM9CPOidfX99WB29GwtAiQ4YMkSTt3bvXzZWgPaMfwQz0I5iBfgQz0I/QUoyEAQAAAIALEcIAAAAAwIUIYQAAAADgQoQwAAAAAHAhQhgAAAAAuBCzIwIAAACACzESBgAAAAAuRAgDAAAAABcihAEAAACACxHCAAAAAMCFCGEAAAAA4EKEMAAAAABwIUIYAAAAALgQIQwAAAAAXIgQhouqqKjQM888owEDBiggIEA9evTQrFmzdOzYMXeXBhcrLy/XkiVLdN9992ngwIEKCAhQcHCwhg0bprlz56q0tLTJc9966y2NGTNGISEhioiI0E033aSNGzde8HobNmzQTTfdpIiICIWEhGjMmDF6++23zf5Y8AAFBQXq2rWrLBaL+vfvf8Fj6Uv4rvz8fD322GMaOHCgAgMDFRERoZEjR+q///u/Gz1+6dKlmjBhgsLCwhQWFqaJEydq+fLlF7zG3r17NXPmTEVHRyswMFCXX365UlNT5XA42uIjwQ22bt2qO++8Uz169JCvr6+sVqvGjx+vBQsWyDCMBsfX1tbqlVde0eWXX67AwEBFR0frzjvv1P79+y94ndb0P3RABnABFRUVxtixYw1JRkxMjHHnnXcaY8aMMSQZ0dHRRkZGhrtLhAv97W9/MyQZkozLLrvMmDlzpjFlyhQjNDTUkGQMGjTIOHnyZIPzHn74YUOSERgYaEyfPt2YMmWK4ePjY3h7exuffPJJo9datGiR4e3tbVgsFmPChAnG9773PcNqtRqSjEcffbSNPylc7d577zUsFoshyejXr1+Tx9GX8F3btm0zIiMjDUnGkCFDjO9///vG1KlTjT59+hje3t4Njn/llVcMSYaPj49x4403GtOnTzcCAwMNScarr77a6DU2btzoPGbMmDHGnXfeaXTv3t2QZMycOdNwOBxt/THRxur+nZBkjBw50rjzzjuNa6+91vDx8TEkGT/84Q/rHV9bW2vcfvvthiTDarUa3/ve94wJEyYYFovFCAoKMjZv3tzodVrT/9AxEcJwQU8++aQhyUhKSjJKSkqc7X/84x8NScaECRPcVxxc7q233jJ+9rOfGfv27avXfvz4cWPEiBGGJOMHP/hBvX0rV640JBmRkZHGwYMHne0bN240/Pz8DKvVahQWFtY7p6CgwAgLCzMkGYsXL3a25+XlGf379zckGWvWrDH988E9Vq1aZUgyfvazn10whNGX8F2nTp0yoqKijKCgIOPTTz9tsP+7X4QPHDhgeHt7G/7+/sbGjRud7d9++60RGRlp+Pj4GIcOHap3TnV1tREfH29IMubPn+9sLykpMZKSkgxJxoIFC8z9YHApu91udO3a1ZBkvPfee/X27du3z4iIiDAkGatXr3a21/1SMiEhwcjLy3O2L1q0yJBk9O/f37Db7fXeqzX9Dx0XIQxNqqqqMsLDww1Jxo4dOxrsT0xMNCQZ27Ztc0N18DQbN240JBn+/v5GVVWVs33q1KmGJOOVV15pcM5//dd/GZKMP/zhD/XaX375ZUOSMX369Abn/POf/zQkGTfffLPZHwFuUF5ebvTr188YPHiwcfDgwQuGMPoSvusXv/iFIcl4/fXXW3T8ww8/3GDf/PnzDUnG7Nmz67V/9NFHhiRj2LBhDc7Zvn27IckYOnRoa8qHh/jmm28MScbAgQMb3V/378vLL7/sbLvssssMSY2OwN96662GJGPRokX12lvT/9Bx8UwYmrRhwwYVFRWpX79+GjFiRIP9M2bMkHT23mZg2LBhkqSqqioVFBRIOvs84erVqyWd6y/na6oP1d0b39g506ZNU0BAgFatWqXKykrzPgDc4rnnntORI0f0l7/8Rb6+vk0eR1/Cd1VUVOjdd99VcHCwfvKTnzTrnAv1h9b0oZEjR6pv377as2ePsrKyWlI+PIi/v3+zjouMjJQkZWZmav/+/QoMDNS0adMaHNeavsR3qs6HEIYm7d69W9LZHzKNqWtPT093WU3wXEeOHJEk+fr6KiIiQpL07bffqqqqStHR0YqNjW1wTlN96EJ9z8/PT0OHDlVlZaUOHjxo6meAa6Wnp+uPf/yjfvKTn2j8+PEXPJa+hO/atm2bSkpKNGLECAUGBupf//qX5syZowcffFCpqak6fvx4veNtNptycnIkqdFfLPbq1UtRUVHKzs5WcXGxs52fhR1f37591a9fP3377bd6//336+3bv3+/3n33XXXp0kW33367pHN9YujQoY3+8qixPtHa/oeOixCGJtX9Y9HYF57z27Ozs11WEzzX//zP/0iSbrzxRudvFS/Wh4KDg2W1WlVYWKiSkhJJUnFxsYqKii54Hn2v/XM4HPrpT38qq9Wq3/3udxc9nr6E79q3b58kqWvXrrrtttt000036ZVXXtGf//xnPfLII+rfv78++OAD5/F1fahLly4KDg5u9D0b6w/8LOz4vL29tXDhQlmtVv3oRz/SqFGjdNddd2nSpElKTExUbGysvvzyS+cvGFvTJ1rb/9BxEcLQpLrpxoOCghrdX/ePSN0XHnRen3/+uf7xj3/I19dX8+bNc7ZfrA9JDfvR+dPc0/c6rldffVVbt27V73//e+ctPhdCX8J3FRYWSpI+++wzffHFF3r99dd16tQpZWVl6bHHHlNFRYXuvfde7dq1S1Lr+lBzzqMPdQzjxo3TunXr1LdvX+3YsUMfffSR1qxZIy8vL91www3q27ev89jW9InW9j90XIQwAJfkwIEDuvvuu2UYhn7/+987nw0DmpKTk6OnnnpKEyZMUHJysrvLQTtVtz5XTU2N5s6dqwcffFDR0dHq06ePfv/732vmzJmy2+36/e9/7+ZK0R588MEHGjNmjHr16qXNmzertLRUBw8eVHJysv74xz9q0qRJqqqqcneZ6EAIYWhSSEiIpLML9DamrKxMkhQaGuqymuBZjh07phtvvFGFhYWaM2eOHn744Xr7L9aHpIb9qO6cC51H32vfHnroIVVXV+svf/lLs8+hL+G7zv/v29jEHHVt69atq3d8S/pQc86jD7V/hw4d0r333quoqCgtW7ZMY8aMUXBwsBISEvTGG2/o5ptv1o4dO/Tmm29Kal2faG3/Q8dFCEOTevfuLUnKzc1tdH9de58+fVxWEzzHmTNnNHnyZGVnZ+snP/mJ/vCHPzQ45mJ9qKysTDabTV26dHH+0AkLC1N4ePgFz6PvtW/Lli1TUFCQHnjgAU2cONH556677pJ0NtzXteXl5UmiL6Ghuv9mQUFBio6ObrA/Li5OknTq1ClJ5/pQYWGh88vudzXWH/hZ2PF9+OGHstvtuvHGG+uF+zp33nmnJOmrr76S1Lo+0dr+h46LEIYm1d1WtmPHjkb317UnJia6rCZ4htLSUk2dOlX79u3THXfcob/97W+yWCwNjhs4cKD8/f2Vn5+vY8eONdjfVB+6UN+z2+3as2ePAgICNGDAADM+DtzAZrNp3bp19f5s3rxZklRZWelsq5s6nr6E76qbYa6ioqLR28TOnDkj6dwIhNVqdX4R3rlzZ4Pjjx49qtOnT6tPnz4KCwtztvOzsOOrCz91v7T5rrr2uucQ6/rEnj17ZLfbGxzfWJ9obf9Dx0UIQ5PGjRun8PBwZWRkOB9sPt+iRYskSbfccouLK4M7VVVVafr06dqyZYumTJmiDz74QN7e3o0eGxgYqEmTJkmSPv744wb7m+pDdeuu1O0/37Jly1RZWanrr79eAQEBl/RZ4B6GYTT6JzMzU5LUr18/Z1vdaAZ9Cd/Vu3dvDRs2TIZhOG85PF9d2/nTgV+oP7SmD+3cuVNHjhzR0KFDnX0V7U/37t0lnV32oDFbt26VdG50NT4+XpdddpkqKiqca3+drzV9ie9UnZC7VolG+/Dkk08akoyrrrrKKC0tdbb/8Y9/NCQZEyZMcF9xcLmamhrj9ttvNyQZ48ePN8rKyi56zsqVKw1JRmRkpHHw4EFn+8aNGw1/f3/DarUahYWF9c4pKCgwwsLCDEnG4sWLne0nT540+vfvb0gy1qxZY9bHgofIzMw0JBn9+vVrdD99Cd/13nvvGZKMyy+/3Dh+/LizfefOnUZERIQhyfi///s/Z/uBAwcMb29vw9/f39i0aZOz/eDBg0ZkZKTh4+NjHDp0qN41qqurjfj4eEOSMX/+fGd7aWmpkZSUZEgyFixY0HYfEm1u+/bthiRDkvGnP/2p3r5NmzYZwcHBhiRj5cqVzva//e1vhiQjISHBOHnypLN98eLFhiSjf//+ht1ur/derel/6LgIYbigiooK48orrzQkGTExMcadd97p3I6OjjYyMjLcXSJcKDU11fmD6vbbbzfuvffeRv/k5+fXO+/hhx82JBlBQUHG9OnTjalTpxo+Pj6Gt7e38cknnzR6rUWLFhleXl6GxWIxrr32WmPGjBmG1Wo1JBlz5sxxwaeFq10shBkGfQkN3XvvvYYkw2q1GjfddJNx7bXXGv7+/oYk4/77729w/Pz58w1Jho+PjzF16lRj+vTpRmBgoCHJ+N///d9Gr7FhwwbnMVdeeaVx5513GjExMYYkY8aMGYbD4Wjrj4k29thjjzl/vg0ZMsSYOXOmMW7cOMPLy8uQZPzsZz+rd3xtba3zl5JdunQxZsyYYUycONGwWCxGYGCgkZaW1uh1WtP/0DERwnBR5eXlxtNPP23069fP8PPzM7p3724kJycbR48edXdpcLHf/OY3zh9SF/qTmZnZ4NwFCxYYo0aNMoKCggyr1WrceOONxoYNGy54vfXr1xs33nijYbVajf/fzh2jJhCEYRgm3aLbyB5jWxuLLWy8gF5LD2BlZWmhV9HbqF+6QEAkBPML4XnqhWHYv3mZYUajUabTaXa73R/tjnf7SYQlZonv7vd7ttvt10yMx+PMZrOn//d4PGYYhrRtm7ZtMwxDTqfT03XO53OWy2W6rkvTNOn7PpvNJrfb7dVb4k0Oh0MWi8XXqdRkMsl8Ps9+v3/4/fV6zXq9Tt/3aZomXddltVrlcrk8Xec388f/85Ekr77iCAAAwGMe5gAAACgkwgAAAAqJMAAAgEIiDAAAoJAIAwAAKCTCAAAACokwAACAQiIMAACgkAgDAAAoJMIAAAAKiTAAAIBCIgwAAKCQCAMAACgkwgAAAAqJMAAAgEIiDAAAoJAIAwAAKPQJ90AoQIK+JcIAAAAASUVORK5CYII=", + "text/plain": [ + "
" + ] + }, + "metadata": {}, + "output_type": "display_data" + } + ], + "source": [ + "plt.plot(torch.sort(self_scores[\"all_modules\"]).values)\n", + "plt.ylabel(\"Self-Influence Score\")\n", + "plt.yscale(\"log\")\n", + "plt.grid()" + ] + }, { "cell_type": "markdown", "id": "261a508a6c02267d", @@ -949,17 +1029,17 @@ "id": "35970418-4649-4a8f-886f-ecd4f8e7dbad", "metadata": {}, "source": [ - "How would the model's behaviors on some query data points change if one or more data points were removed from the training dataset? We can use influence functions to identify influential training data points for a randomly selected query dataset." + "How would the model's behaviors on some query data points change if one or more data points were removed from the training dataset? We can use influence functions to identify influential training data points for randomly selected query data points." ] }, { "cell_type": "code", - "execution_count": 24, + "execution_count": 27, "id": "4522bd15c0f00882", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T21:05:54.279349Z", - "start_time": "2024-03-12T21:05:54.269910Z" + "end_time": "2024-06-24T17:27:57.312028Z", + "start_time": "2024-06-24T17:27:57.306719Z" }, "collapsed": false, "jupyter": { @@ -970,16 +1050,16 @@ { "data": { "text/plain": [ - "1" + "10" ] }, - "execution_count": 24, + "execution_count": 27, "metadata": {}, "output_type": "execute_result" } ], "source": [ - "single_query_dataset = torch.utils.data.Subset(query_dataset, list(range(1)))\n", + "single_query_dataset = torch.utils.data.Subset(query_dataset, list(range(10)))\n", "len(single_query_dataset)" ] }, @@ -993,12 +1073,12 @@ }, { "cell_type": "code", - "execution_count": 25, + "execution_count": 28, "id": "d705577ebd58b8c8", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T21:13:13.194949Z", - "start_time": "2024-03-12T21:13:13.178273Z" + "end_time": "2024-06-24T17:28:01.446294Z", + "start_time": "2024-06-24T17:28:01.437385Z" }, "collapsed": false, "jupyter": { @@ -1007,40 +1087,47 @@ }, "outputs": [], "source": [ - "def train_and_evaluate(modified_train_dataset: data.Dataset, query_dataset: data.Dataset) -> float:\n", + "def train_and_evaluate(current_train_dataset: data.Dataset, current_query_dataset: data.Dataset, seed: int) -> float:\n", " current_model = train(\n", - " dataset=modified_train_dataset,\n", + " dataset=current_train_dataset,\n", " batch_size=train_batch_size,\n", " num_train_epochs=num_train_epochs,\n", " learning_rate=learning_rate,\n", " weight_decay=weight_decay,\n", + " seed=seed,\n", " disable_tqdm=True,\n", " )\n", - " return evaluate(model=current_model, dataset=query_dataset, batch_size=len(query_dataset))" + " return evaluate(model=current_model, dataset=current_query_dataset, batch_size=len(current_query_dataset))" ] }, { "cell_type": "code", - "execution_count": 26, + "execution_count": 29, "id": "be1ab297-9260-47a9-82ec-56d1349fa862", - "metadata": {}, + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:28:15.326422Z", + "start_time": "2024-06-24T17:28:04.327587Z" + } + }, "outputs": [ { "data": { "text/plain": [ - "0.02686239752540132" + "0.08797347468137742" ] }, - "execution_count": 26, + "execution_count": 29, "metadata": {}, "output_type": "execute_result" } ], "source": [ - "num_iter = 25\n", + "num_iter = 50\n", + "seed_lst = [i + 1004 for i in range(num_iter)]\n", "base_loss = 0.0\n", - "for _ in range(num_iter):\n", - " base_loss += train_and_evaluate(modified_train_dataset=train_dataset, query_dataset=single_query_dataset)\n", + "for i in range(num_iter):\n", + " base_loss += train_and_evaluate(current_train_dataset=train_dataset, current_query_dataset=single_query_dataset, seed=seed_lst[i])\n", "base_loss /= num_iter\n", "base_loss" ] @@ -1055,38 +1142,37 @@ }, { "cell_type": "code", - "execution_count": 27, + "execution_count": 30, "id": "8b3eb92fd287b0cc", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T21:06:42.141007Z", - "start_time": "2024-03-12T21:06:42.067707Z" + "end_time": "2024-06-24T17:29:09.085348Z", + "start_time": "2024-06-24T17:29:08.178507Z" }, "collapsed": false, "jupyter": { "outputs_hidden": false } }, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "Fitting covariance matrices [1/1] 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n", - "Fitting covariance matrices [1/1] 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n", - "Performing Eigendecomposition [4/4] 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n", - "Fitting Lambda matrices [1/1] 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n", - "Fitting Lambda matrices [1/1] 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n", - "Computing pairwise scores (query gradient) [0/1] 0%| [time left: ?, time spent: 00:00]\n", - "Computing pairwise scores (training gradient) [1/1] 100%|███████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\u001b[A\n", - "Computing pairwise scores (query gradient) [1/1] 100%|██████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n", - "Computing pairwise scores (query gradient) [0/1] 0%| [time left: ?, time spent: 00:00]\n", - "Computing pairwise scores (training gradient) [1/1] 100%|███████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\u001b[A\n", - "Computing pairwise scores (query gradient) [1/1] 100%|██████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n" - ] - } - ], + "outputs": [], "source": [ + "torch.manual_seed(0)\n", + "model = train(\n", + " dataset=train_dataset,\n", + " batch_size=train_batch_size,\n", + " num_train_epochs=num_train_epochs,\n", + " learning_rate=learning_rate,\n", + " weight_decay=weight_decay,\n", + ")\n", + "model = prepare_model(model, task)\n", + "analyzer = Analyzer(\n", + " analysis_name=\"counterfactual\",\n", + " model=model,\n", + " task=task,\n", + " cpu=True,\n", + " disable_tqdm=True,\n", + ")\n", + "\n", "analyzer.fit_all_factors(\n", " factors_name=\"counterfactual_factors\",\n", " dataset=train_dataset,\n", @@ -1114,12 +1200,12 @@ }, { "cell_type": "code", - "execution_count": 28, + "execution_count": 31, "id": "86fd82e97c1c0af9", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T21:08:01.250153Z", - "start_time": "2024-03-12T21:08:01.244362Z" + "end_time": "2024-06-24T17:29:13.536671Z", + "start_time": "2024-06-24T17:29:13.442014Z" }, "collapsed": false, "jupyter": { @@ -1129,7 +1215,7 @@ "outputs": [ { "data": { - "image/png": "iVBORw0KGgoAAAANSUhEUgAAA2MAAAJqCAYAAACmSZh1AAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjguMywgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/H5lhTAAAACXBIWXMAABcSAAAXEgFnn9JSAABc4ElEQVR4nO3deXxU1f3/8fdk30hCWAz7vlQgLCoWEIH8rCKUIghYq19BqtYq1m+xtbUuXwr67dcFXIqtVgXqRqkBUVBqiSxWEGQRIoigAYIgCMiWfZvz+wNmmJsMJDOZZLgzr+fjwYOZc+8990x6aubN595zHcYYIwAAAABAg4oI9gAAAAAAIBwRxgAAAAAgCAhjAAAAABAEhDEAAAAACALCGAAAAAAEAWEMAAAAAIKAMAYAAAAAQUAYAwAAAIAgIIwBAAAAQBAQxgAAAAAgCAhjAAAAABAEhDEAAAAACIKoYA8AvklPT1dhYaHatm0b7KEAAAAAYW/fvn1KTEzUoUOHfD6WypjNFBYWqry8POhjKCwsDOoYEBqYSwgE5hECgXmEQGAehafy8nK//3enMmYzrorY9u3bgzaGFStWSJIyMzODNgaEBuYSAoF5hEBgHiEQmEfhqUePHn4fS2UMAAAAAIKAMAYAAAAAQUAYAwAAAIAgIIwBAAAAQBAQxgAAAAAgCAhjAAAAABAEhDEAAAAACALCGAAAAAAEAWEMAAAAAIKAMAYAAAAAQUAYAwAAAIAgIIwBAAAAQBAQxgAAAAAgCAhjAAAAABAEhDEAAAAACALCGAAAAAAEAWEMAAAAAIKAMAYAAAAAQUAYAwAAAIAgIIwBAAAAQBAQxgAAAAAgCKKCPQAAAAAA8MUnud/rcH6JJCmjdao6NE0M8oj8QxgDAAAAYCsvrM7V6l1HJEkzRvewbRjjMkUAAAAAtmI83zgcwRpGnRHGAAAAANiKMWfjmH2jGGEMAAAAgI3ZuDBGGAMAAABgLx6FMTlsXBsjjAEAAACwFeNx1xiVMQAAAABoINbKmH0RxgAAAADYiiWM2TiNEcYAAAAA2IrlMkUb18YIYwAAAABsxbMyZuMsRhgDAAAAYC8hksUIYwAAAABsxnLPmH3jGGEMAAAAgK1Y7xmzL8IYAAAAAFthNUUAAAAACALLPWOEMQAAAABoGMawtD0AAAAANDgqYwAAAAAQBJbnjNkYYQwAAACArVgrY/YtjRHGAAAAANiLYWl7AAAAAGhw3DMWIoqLi/XII4+oa9euiouLU8uWLTV58mQdOHDA576WL1+ukSNHqlmzZoqOjlaTJk109dVX6+23366HkQMAAADhyfKcMRvXxsI6jJWUlCgzM1MzZsxQQUGBRo8erTZt2mju3Lnq27evdu/eXeu+nnnmGV199dVatmyZunbtquuvv17du3dXdna2xo4dqwcffLAePwkAAAAQPoxHbYzKmE09+uijWrdunQYMGKBdu3ZpwYIFWr9+vWbOnKkjR45o8uTJternyJEj+v3vf6/o6GitXLlSa9as0T/+8Q+tWbNGq1atUmxsrP70pz/5FO4AAAAAeGetjNlX2IaxsrIyzZ49W5L0/PPPKykpyb1t6tSpysjI0OrVq7Vp06Ya+1q/fr1KS0uVmZmpIUOGWLZdeeWVuuaaa2SM0caNGwP7IQAAAIAwZAljNk5jYRvG1qxZo5MnT6pTp07q27dvte3jxo2TJC1ZsqTGvmJjY2t1ziZNmvg2SAAAAADVWB8zZt80FrZhbOvWrZKkfv36ed3uas/Jyamxr/79+ys1NVUrVqzQ6tWrLds++ugjffDBB+rSpYsGDx5cx1EDAAAAMCY07hmLCvYAgmXfvn2SpNatW3vd7mrPy8ursa+UlBS98sor+tnPfqZhw4Zp4MCBat26tfbv36+1a9dq0KBBevXVVxUTE1Pr8fXo0cNre25urtLT07VixYpa9xVohYWFkhTUMSA0MJcQCMwjBALzCIHAPGo4BQUl7tfbPv9c0d99EbSxFBYWKjEx0a9jwzaMFRQUSJISEhK8bnf9QPPz82vV39ixY7Vs2TJNmDBBa9ascbcnJyfr6quvVqtWreo4YgAAAABS1csU7Stsw1igzZw5U/fff7+uu+46TZs2TR07dtTu3bv1yCOP6JFHHtH69eu1dOnSWve3fft2r+2uillmZmZAxu0P17/2BHMMCA3MJQQC8wiBwDxCIDCPGs7/bV0tnSmu9M7IUObFFwVtLP5WxaQwvmfMtXpiUVGR1+2uMnOjRo1q7GvVqlX6zW9+oz59+uitt95Sr169lJiYqF69eikrK0t9+vTRe++9p2XLlgXuAwAAAABhitUUba5t27aSpP3793vd7mpv165djX299tprkqQxY8YoIsL6I42MjNTYsWMlnV7MAwAAAEDdeF6mSBizod69e0uSNm/e7HW7qz0jI6PGvlzBLSUlxet2V/vx48d9HicAAAAAK8tqiixtbz+DBg1SSkqKcnNztWXLlmrbs7KyJEmjRo2qsa/09HRJOudDnTds2CBJat++vX+DBQAAAOBmWcDDvlksfMNYTEyMpkyZIkm6++673feISdKsWbOUk5OjIUOG6JJLLnG3z549W927d9cDDzxg6eu6666TJL3xxhvVFul455139OabbyoiIkJjxoypp08DAAAAhA/Pe8YibHydYlivpvjQQw8pOztba9eudT+UOS8vT+vXr1ezZs00Z84cy/5Hjx7Vzp07dfDgQUv7ddddp/Hjx+utt97SqFGjdOmll6pDhw7as2ePu1r22GOPqVu3bg322QAAAIBQZb1M0b7CtjImSXFxcVq5cqUefvhhJSQkaPHixcrLy9OkSZO0efNmdezYsVb9OBwOLViwQK+88oquvPJKff3113r77be1d+9ejRgxQsuWLdMf/vCHev40AAAAQHgIlQU8wroyJknx8fGaPn26pk+fXuO+06ZN07Rp07xuczgcmjx5siZPnhzgEQIAAADwZFna3sa1sbCujAEAAACwH+NRG7NzZYwwBgAAAMBWrJUx+yKMAQAAALAVY7lpLGjDqDPCGAAAAADb4p4xAAAAAGgglqXt7ZvFCGMAAAAA7CVErlIkjAEAAACwF8sCHjYujRHGAAAAANgKS9sDAAAAQBCwtD0AAAAABIHlnjEbpzHCGAAAAABbsTxnzMa1McIYAAAAAJvhnjEAAAAAaHDcMwYAAAAAQWC9Z8y+cYwwBgAAAMBWjEdpzL5RjDAGAAAAwGZYTREAAAAAgsB6z5h90xhhDAAAAICtWC5TtG8WI4wBAAAAsBdT8y62QBgDAAAAYC+elylSGQMAAACAhmFZwIN7xgAAAACgYXDPGAAAAAAEAUvbAwAAAEAQsLQ9AAAAAASBEZcpAgAAAECDs1bG7IswBgAAAMBWuGcMAAAAAILB8tRn+6YxwhgAAAAAW+GeMQAAAAAIAu4ZAwAAAIAgsN4zZt84RhgDAAAAYCvGozRm3yhGGAMAAABgM6ymCAAAAABBYL1nzL5pjDAGAAAAwLaojAEAAABAA/C8X8zuCGMAAAAAbKNqFqMyBgAAAAANoGpdjKXtAQAAAKABVL1M0b5RjDAGAAAAwEaqV8aCMoyAIIwBAAAAsI1q94zZuDZGGAMAAABgG6ZKbYzKGAAAAAA0gOqVMfsijAEAAACwLxunMcIYAAAAANvgnjEAAAAACALuGQMAAACAIOCeMQAAAAAIgurPGbNvHCOMAQAAALANZ5XSWIR9sxhhDAAAAIB9sIAHAAAAAARDtesUgzKKgCCMAQAAALANVlMEAAAAgCBgNcUQUlxcrEceeURdu3ZVXFycWrZsqcmTJ+vAgQN+9bd3717deeed6tChg2JjY9W0aVMNGDBATz75ZIBHDgAAAIQfVlMMESUlJcrMzNSMGTNUUFCg0aNHq02bNpo7d6769u2r3bt3+9TfsmXL1KNHD/3tb39TkyZNNHbsWPXr10979+7Viy++WE+fAgAAAAgfpkppzL5RTIoK9gCC6dFHH9W6des0YMAA/fvf/1ZSUpIkadasWbrvvvs0efJkrVq1qlZ9ffnllxo7dqwaNWqk5cuXa+DAge5tTqdTmzdvro+PAAAAAISV6pWxoAwjIMK2MlZWVqbZs2dLkp5//nl3EJOkqVOnKiMjQ6tXr9amTZtq1d/UqVNVUlKiefPmWYKYJEVEROjSSy8N3OABAACAMMXS9iFgzZo1OnnypDp16qS+fftW2z5u3DhJ0pIlS2rs65tvvtEHH3ygjh07asSIEQEfKwAAAIDTQmk1xbC9THHr1q2SpH79+nnd7mrPycmpsa9Vq1bJ6XRq4MCBqqio0KJFi7RmzRpVVlaqZ8+euuGGG9S4cePADR4AAAAIV1WvU7SxsA1j+/btkyS1bt3a63ZXe15eXo19ffHFF5KkpKQkDR48WOvWrbNsf/DBB5WVlaVhw4bVenw9evTw2p6bm6v09HStWLGi1n0FWmFhoSQFdQwIDcwlBALzCIHAPEIgMI8axvESaxpbvXqVoiOCVx4rLCxUYmKiX8eG7WWKBQUFkqSEhASv210/0Pz8/Br7On78uCTp5Zdf1pdffqk333xTx44d086dO3XzzTfr2LFjGjNmjN/L5QMAAADwzsZXKYZvZSyQnE6nJKmiokIvvviiJkyYIElq3LixXnvtNe3cuVMbNmzQX/7yFz322GO16nP79u1e210Vs8zMzACM3D+uf+0J5hgQGphLCATmEQKBeYRAYB41jIMni6WPzlYfhw0bpujI4NWY/K2KSWFcGXOtnlhUVOR1u6vM3KhRo1r3lZSUpPHjx1fbfuutt0qSVq9e7ddYAQAAAJxWfTVF+wrbMNa2bVtJ0v79+71ud7W3a9euxr5c+7Rt29brE8Dbt28vSTp8+LA/QwUAAABwRvXnjNk3joVtGOvdu7cknfNhzK72jIyMGvtyLY3vunesqmPHjkmS5VlmAAAAAHxnqpTG7BvFwjiMDRo0SCkpKcrNzdWWLVuqbc/KypIkjRo1qsa+Bg4cqCZNmujQoUPauXNnte2uyxO9Pc8MAAAAQO1Vu0zRxmksbMNYTEyMpkyZIkm6++673feISdKsWbOUk5OjIUOG6JJLLnG3z549W927d9cDDzxg6SsqKkpTp06VMUZ33323Tp065d6WnZ2tefPmyeFw6Be/+EU9fyoAAAAgvNj5MsWwXk3xoYceUnZ2ttauXasuXbpo8ODBysvL0/r169WsWTPNmTPHsv/Ro0e1c+dOHTx4sFpfv/3tb7Vy5UplZ2era9eu+uEPf6ijR49q3bp1qqys1GOPPab+/fs31EcDAAAAQlLVypidhW1lTJLi4uK0cuVKPfzww0pISNDixYuVl5enSZMmafPmzerYsWOt+4qOjtb777+vxx9/XE2bNtUHH3ygzz//XEOGDNGSJUv0hz/8oR4/CQAAABAejMcSHjYuikkK88qYJMXHx2v69OmaPn16jftOmzZN06ZNO+f26Oho3X///br//vsDOEIAAAAALp6VMZtnsfCujAEAAACwF8+rFO18v5hEGAMAAABgI55L29s7ihHGAAAAANiItTIWtGEEBGEMAAAAgG1Y7xmzdxojjAEAAACwkdBZwYMwBgAAAMA2WE0RAAAAAIKAe8YAAAAAIAi4ZwwAAAAAgsB41MaojAEAAABAA+GeMQAAAAAIAksYs3lpjDAGAAAAwDYslykGcRyBQBgDAAAAYBvGspxi0IYREIQxAAAAALZk8yxGGAMAAABgH9wzBgAAAABBwNL2AAAAABAELG0PAAAAAEFgWb/D5qUxwhgAAAAA2zCGpe0BAAAAoMFZK2NBG0ZAEMYAAAAA2IblOWM2r40RxgAAAADYCKspAgAAAECDYzVFAAAAAAgC7hkDAAAAgCCwVsbsncYIYwAAAABsw7K0vb2zGGEMAAAAgH1YLlMM2igCgzAGAAAAwDYslynavDRGGAMAAABgG8ZSG7M3whgAAAAA+7BUxoI3jEAgjAEAAACwDSdhDAAAAAAanudlihE2T2OEMQAAAAC2YX3OmL0RxgAAAADYhmVpeypjAAAAANAwLA99DuI4AoEwBgAAAMA2LAvb2zyNEcYAAAAA2Af3jAEAAABAw/NcTZF7xgAAAACggbCaIgAAAAAEgeGhzwAAAADQ8CxL29u8NkYYAwAAAGAblqXt7Z3FCGMAAAAA7MPUvIttEMYAAAAA2Ib1njF7l8YIYwAAAABsxOMyxSCOIhAIYwAAAABsg9UUAQAAACAILKspEsYAAAAAoGFYH/ps7zRGGAMAAABgG0YsbQ8AAAAADc5aGbM3whgAAAAA27A8Z8zmpTHCGAAAAADbMIal7UNGcXGxHnnkEXXt2lVxcXFq2bKlJk+erAMHDtSp36+++krx8fFyOBy66qqrAjRaAAAAAC42L4yFdxgrKSlRZmamZsyYoYKCAo0ePVpt2rTR3Llz1bdvX+3evdvvvu+44w6VlpYGcLQAAAAAuGcsRDz66KNat26dBgwYoF27dmnBggVav369Zs6cqSNHjmjy5Ml+9fvKK69o1apVuv322wM8YgAAACC8WVdTtHccC9swVlZWptmzZ0uSnn/+eSUlJbm3TZ06VRkZGVq9erU2bdrkU7/fffedfvvb3+pHP/qRbrzxxoCOGQAAAAh3VMZCwJo1a3Ty5El16tRJffv2rbZ93LhxkqQlS5b41O+9996r4uJi/eUvfwnIOAEAAACcZQljNk9jYRvGtm7dKknq16+f1+2u9pycnFr3+f7772vBggX6wx/+oM6dO9d9kAAAAAAsPJe2d9i8NhYV7AEEy759+yRJrVu39rrd1Z6Xl1er/goLC3XXXXepW7du+t3vflfn8fXo0cNre25urtLT07VixYo6n8NfhYWFkhTUMSA0MJcQCMwjBALzCIHAPGoYXxyocL8+cfJE0H/ehYWFSkxM9OvYsA1jBQUFkqSEhASv210/0Pz8/Fr199BDDykvL08rV65UTExMYAYJAAAAwMJaGbO3sA1jgbRx40Y999xzuuWWWzR06NCA9Ll9+3av7a6KWWZmZkDO4w/Xvz4EcwwIDcwlBALzCIHAPEIgMI8axtEN30jbT99K1LhxqjIzBwR1PP5WxaQwvmfMtXpiUVGR1+2uMnOjRo3O209FRYVuv/12paam6qmnngrsIAEAAABYWJa2t3ltLGwrY23btpUk7d+/3+t2V3u7du3O28/+/fu1ZcsWpaena/z48ZZtJ06ckCRt2rTJXTFbtWqV/4MGAAAAwlworaYYtmGsd+/ekqTNmzd73e5qz8jIqFV/hw4d0qFDh7xuO3HihFavXu3HKAEAAAB4stwzZvMwFraXKQ4aNEgpKSnKzc3Vli1bqm3PysqSJI0aNeq8/bRv317GGK9/Vq5cKUn6f//v/7nbAAAAAPjP+tBne6exsA1jMTExmjJliiTp7rvvdt8jJkmzZs1STk6OhgwZoksuucTdPnv2bHXv3l0PPPBAg48XAAAAQJV7xuydxcL3MkXp9HL02dnZWrt2rbp06aLBgwcrLy9P69evV7NmzTRnzhzL/kePHtXOnTt18ODBII0YAAAACG+hdLFZ2FbGJCkuLk4rV67Uww8/rISEBC1evFh5eXmaNGmSNm/erI4dOwZ7iAAAAAA8WO8Zs3dpLKwrY5IUHx+v6dOna/r06TXuO23aNE2bNq3WfQ8dOpT7xAAAAIBAMp5L29tbWFfGAAAAANgLqykCAAAAQBBYV1O0N8IYAAAAANvwvA3I7veMEcYAAAAA2IblMsWgjSIwCGMAAAAAbMNymaLN0xhhDAAAAIBtWNcqt3caI4wBAAAAsA3rPWNBHEgAEMYAAAAA2JLNsxhhDAAAAIB9VDrPVsYiI+wdxwhjAAAAAGyjwiOMRUXaO87Ye/QAAAAAwkp5pdP9OprKGAAAAAA0jIpKz8oYYQwAAAAAGoSlMsZligAAAADQMMo9KmOEMQAAAABoIBXOs5WxKO4ZAwAAAICGUV7JaooAAAAA0OAqLPeMURkDAAAAgAZhec5YhL3jjL1HDwAAACCslHlWxqKojAEAAABAg7BcpkhlDAAAAAAaBg99BgAAAIAgKHeymiIAAAAANDjrZYpUxgAAAACgQVTwnDEAAAAAaHjlTp4zBgAAAAANrtzy0Gd7xxl7jx4AAABAWLFcpsg9YwAAAADQMKiMAQAAAEAQVDh5zhgAAAAANDjrZYr2jjP2Hj0AAACAsGK9TJHKGAAAAAA0CM/LFLlnDAAAAAAaSHnF2coY94wBAAAAQAOxPvTZ3nHG3qMHAAAAEFZ4zhgAAAAANDBjDPeMAQAAAEBD8wxiEveMAQAAAECD8LxEUeI5YwAAAADQIDwX75CkGC5TBAAAAID657msvcRligAAAADQILhnDAAAAACCoLzSWhmL5p4xAAAAAKh/ngt4RDikCJ4zBgAAAAD1r8JjAY8omy/eIRHGAAAAANhEuUdlzO4rKUqEMQAAAAA24XnPmN0X75CkqIY60Y4dO7R9+3a1adNGl19+eUOdFgAAAECI8KyM2f2Bz1KAK2MLFixQZmam1q9fb2n/7W9/q549e+qGG27QwIEDNWbMGFVWVgby1AAAAABCXIVHZSw6BCpjAQ1jr7/+urZs2aK+ffu629auXauZM2eqUaNG+ulPf6r27dvr3Xff1RtvvBHIUwMAAAAIcZ7PGQuFyxQDGsa2bdumjIwMxcTEuNtee+01ORwO/fOf/9Qbb7yhDRs2KCkpSS+//HIgTw0AAAAgxHneM2b3Z4xJAQ5jhw8fVqtWrSxtK1euVPPmzXX11VdLktLS0nTllVfq66+/DuSpAQAAAIQ4z+eMURmrIj4+XqdOnXK/P3jwoHbt2qUhQ4ZY9ktNTdXx48cDeWoAAAAAIc7zOWPRLG1v1bFjR/3nP//RiRMnJElvvPGGHA6HuyrmcujQITVv3jyQpwYAAAAQ4soslTHCmMWkSZN06tQpXXLJJbr++uv10EMPKSkpSaNHj3bvU15ero0bN6pr166BPLXfiouL9cgjj6hr166Ki4tTy5YtNXnyZB04cKDWfZw4cUJvvvmmbrzxRnXo0EExMTFq1KiRLr/8cj377LMqLy+vx08AAAAAhAfLaooR9r9MMaDPGbv99tu1cuVKLVy4UHv27FFiYqJefPFFNWnSxL3P0qVLdfLkSWVmZgby1H4pKSlRZmam1q1bpxYtWmj06NHau3ev5s6dq6VLl2rdunXq2LFjjf089dRTeuyxx+RwONSnTx9dfvnlOnLkiNasWaNPP/1UWVlZ+uCDD5SQkNAAnwoAAAAITaF2z1hAw1h0dLTeeust7d27V0eOHFH37t3VqFEjyz4dOnTQ22+/rR/+8IeBPLVfHn30Ua1bt04DBgzQv//9byUlJUmSZs2apfvuu0+TJ0/WqlWrauwnMTFR999/v+6++261bdvW3f7VV1/pqquu0scff6xHH31U//u//1tfHwUAAAAIeeXcM1az9u3b67LLLqsWxCSpT58+Gj16tC666KL6OHWtlZWVafbs2ZKk559/3h3EJGnq1KnKyMjQ6tWrtWnTphr7euCBB/T4449bgpgkdenSRf/3f/8nSZo/f34ARw8AAACEH0tlLAQuU6y3OHns2DEtX75c8+fP19q1a+vrNH5bs2aNTp48qU6dOlkeUu0ybtw4SdKSJUvqdJ7evXtLkr799ts69QMAAACEO8tzxqiMVXfkyBH97Gc/U3p6uoYPH66bb77Z8oDnl19+WWlpafr4448DfWqfbN26VZLUr18/r9td7Tk5OXU6z+7duyVJ6enpdeoHAAAACHcVzrOVsVAIYwG9Z+zYsWMaOHCgcnNz1adPHw0aNEjPP/+8ZZ+xY8fql7/8pbKysnTFFVcE8vQ+2bdvnySpdevWXre72vPy8up0nmeffVaSLCtK1kaPHj28tufm5io9PV0rVqyo07jqorCwUJKCOgaEBuYSAoF5hEBgHiEQmEf1b+fus6uUf3/k8AXxsy4sLFRiYqJfxwY0Tj722GPKzc3VI488os2bN+vPf/5ztX3S0tLc92MFU0FBgSSdc4VD1w80Pz/f73O88MILys7OVmpqqn7/+9/73Q8AAAAAyeMqRYVAYSywlbHFixera9eumjZt2nn369SpU61WKbSz//znP7r33nvlcDg0Z84ctWzZ0qfjt2/f7rXdVTEL5qMBXP8CcSE8ngD2xlxCIDCPEAjMIwQC86j+bSz7UtqdK0lq06qlMjMzgjwi+V0VkwJcGTtw4IB7wYrzcTgcOnXqVCBP7TPX6olFRUVet7vKzN5WhKzJtm3bNHr0aJWVlenZZ5/VmDFj/B8oAAAAAEnWe8ZC4TljAQ1jycnJOnjwYI375ebmqlmzZoE8tc9cy9Dv37/f63ZXe7t27Xzqd8+ePbr66qt1/PhxTZs2Tffcc0/dBgoAAABAknU1xagI+1+nGNBPcNlll2nDhg3as2fPOffZunWrtmzZokGDBgXy1D5zVfA2b97sdburPSOj9qXPgwcP6kc/+pEOHjyoe++9V//zP/9T94ECAAAAkGR9zlhMFGHM4p577lFpaanGjBmjHTt2VNv+9ddf67/+679kjNGUKVMCeWqfDRo0SCkpKcrNzdWWLVuqbc/KypIkjRo1qlb9HT9+XNdcc41yc3N166236umnnw7kcAEAAICwZ62McZmixfDhw3X//fcrJydHPXv2VPfu3eVwOPTBBx+od+/e+sEPfqBt27bpD3/4Q1CXtZekmJgYdyC8++673feISdKsWbOUk5OjIUOG6JJLLnG3z549W927d9cDDzxg6auoqEgjR47U559/rgkTJuill16Sw2H/yQEAAABcSMorPe8Zs39lLKCrKUrS//3f/+mSSy7RY4895n5g8sGDB3Xw4EF1795dDz/8sG688cZAn9YvDz30kLKzs7V27Vp16dJFgwcPVl5entavX69mzZppzpw5lv2PHj2qnTt3Vrsv7sEHH9Qnn3yiyMhIRUVF6ec//7nX882bN6++PgoAAAAQ8iqcZytj0SFQGQtoGDt16pQcDofGjx+v8ePH68iRI9q7d6+cTqdat26tVq1aBfJ0dRYXF6eVK1fqT3/6k958800tXrxYaWlpmjRpkmbMmHHOB0JXdfz4cUlSZWWl3nzzzXPuRxgDAAAA/FdBZezcUlNTdfnll+uTTz6RJDVr1izoqybWJD4+XtOnT9f06dNr3HfatGlen6E2b948ghYAAABQzzzvGYtmaXurlJQUdezYMZBdAgAAAIAk63PGokOgMhbQT9C3b1/l5uYGsksAAAAAkFRlNUUqY1a/+93vtGHDBvey8AAAAAAQKJbLFEPgoc8BvWcsPj5et912m2644Qb9+Mc/1qhRo9S2bVvFxcV53f/KK68M5OkBAAAAhDDrAh72r4wFNIwNHTpUDodDxhgtWbJES5cuPe/+lZWVgTw9AAAAgBBW7mQ1xXO65ZZbeNgxAAAAgHpRUclzxs6J5d0BAAAA1JdQe86Y/T8BAAAAgLBQ7gyt54wFtDLmqaysTFu2bNGBAwckSa1atVKfPn0UExNTX6cEAAAAEMI8K2Oh8JyxgIexkpISPfLII3rxxRdVUFBg2ZaUlKQ777xTf/zjH8+5wiIAAAAAeGN5zhj3jFmVlpbqqquu0ieffCJJysjIUPv27eVwOLR3715t3bpVTz31lNasWaMPP/xQsbGxgTw9AAAAgBBWWuH50Gf7V8YC+gmefvpprV27VoMGDdKWLVv02Wef6e2339aiRYu0efNmbd26VYMHD9Ynn3yiZ555JpCnBgAAABDCCkordKywzP2+RYr9r7QLaBibP3++mjVrpvfee0+9evWqtr1nz55aunSpmjZtqjfeeCOQpwYAAAAQwvYeLXS/jo+OVHoyYczi66+/1tChQ9WoUaNz7pOUlKShQ4cqNzc3kKcGAAAAEMJyj5xdj6J900RFhMA9YwENY1FRUSoqKqpxv6KiIkVF1dtCjgAAAABCzB6PyljHpolBHEngBDSM9erVSytWrNDu3bvPuc+ePXu0YsUKZWRkBPLUAAAAAEKYZxjrQBir7he/+IWKi4s1dOhQvfLKKyouLnZvKy4u1ty5czV06FCVlJTozjvvDOSpAQAAAISwUAxjAb1W8L/+67/08ccf66WXXtIdd9yhO+64Q02bNpUkHT16VJJkjNEvfvEL3XTTTYE8NQAAAIAQZYzRniMeYaxZaISxgC/O/+KLL+qtt97SFVdcoejoaB05ckRHjhxRdHS0Bg8erLfeekt//etfA31aAAAAACHqaEGZ8ksr3O9D5Z6xellF4/rrr9f111+viooKff/995KkJk2asGgHAAAAAJ8UlVXovxd85n6flhij1ISYII4ocOo1HUVFRemiiy6qz1MAAAAACGF/WPS51nz9vft952ZJQRxNYAX8OWPPPfectm3bds59tm3bpueee+68Ky4CAAAAgCRt3nfC8v7OoR2DM5B6ENAw9swzz+i+++5TcnLyOfdp1KiRpk6dqueeey6QpwYAAAAQgo4Xlblfv/bz/srsHjpX3gU0jH344Yfq06eP2rZte8592rVrpz59+mj58uWBPDUAAACAEFNR6VR+ydmFO9o3CY2FO1wCGsa++eYbdexYc9mwU6dO2r9/fyBPDQAAACDEnPIIYpKUmhAdpJHUj4CGscjISJWWlta4X2lpqSorKwN5agAAAAAh5oTHJYpREQ4lxYbW6uwBDWNdu3bVxx9/rKKionPuU1RUpI8//lhdunQJ5KkBAAAAhJjjReXu16kJ0XI4HEEcTeAFNIyNGzdOx44d02233abCwsJq24uKinT77bfr+PHjGjduXCBPDQAAACDEnCw+WxlLiQ+tSxSlAD9n7J577tFrr72mBQsWaOXKlbrxxhvVqVMnSVJubq7mz5+vw4cPq1u3bvrv//7vQJ4aAAAAQIg5YamMhcaDnj0FNIwlJCQoOztbN998s1asWKFnnnnGXUo0xkiShg0bptdee02JiaG1EgoAAACAwPK8TLFxiC3eIQU4jElSenq6srOztWHDBmVnZ+ubb76RJLVp00ZXXXWVLrvsskCfEgAAAEAIOlnkeZkilbFau+yyy9zBq6SkRCdOnFDTpk3r63QAAAAAQswxjzAWasvaS3VcwCM/P1+ffvqpdu3a5XX7V199pWuuuUbJyclq1aqVGjVqpJ/+9Kc6ePBgXU4LAAAAIAwczT8bxpomxQZxJPWjTmHs1Vdf1YABA7R06dJq2w4dOqTBgwcrOztbFRUVMsaotLRUb731loYNG6bi4uK6nBoAAABAiDtacPYZxk2TQu8yxTqFsdWrVysiIkI333xztW0zZszQ4cOH1bhxY73zzjvKz8/X5s2bdemll+qrr77SX//617qcGgAAAECIs4SxRlTGLHJyctSrVy81b97c0u50OjV//nw5HA797//+r0aNGqXExET16dNHixYtUlRUlBYvXlyXUwMAAAAIcUcLzl6m2IzLFK1czwyrKicnRydOnFBUVJR++tOfWra1atVKl19+uXbs2FGXUwMAAAAIYSXllSoorXC/556xKgoLC1VeXl6tfdOmTZKkjIwMJScnV9veunVrnTp1qi6nBgAAABDC9h0rsrxPS+SeMYtmzZrpyy+/rNb+8ccfy+FwqH///l6PKykpUUpKSl1ODQAAACBE/Xv7IV399Efu9ynx0YqJqlN0uSDV6RO5LjdcsmSJu+3IkSNatGiRJOnqq6/2etwXX3yhli1b1uXUAAAAAELUvLV7Le9bN44PzkDqWZ3C2JQpU2SM0fjx43XLLbfovvvu02WXXab8/Hy1adNGI0eOrHbM7t27tWvXLvXu3bsupwYAAAAQog7nl1re//qqrkEaSf2KqsvBw4YN07Rp0/THP/5Rr7/+uhwOh4wxio+P19y5cxUVVb37F154QZJ0zTXX1OXUAAAAAELU8cKzqyguumug+rVtHMTR1J86hTFJeuSRR/TjH/9YixYt0pEjR9SmTRvddNNN6tChg9f9Y2Njde+992r48OF1PTUAAACAEON0Gh0vOhvG0hJCb+EOlzqHMUnq16+f+vXrV6t9Z8yYEYhTAgAAAAhBp0rK5TRn3zcOwVUUXUJvSRIAAAAAtnXM4xLFyAiHkuMCUj+6IBHGAAAAAFwwPC9RbJwQI4fDEcTR1C/CGAAAAIALxrHCcvfrtMToII6k/hHGAAAAAFww3tlywP26cQgv3iEFaAEPAAAAAKiL4rJKzVmzR0tzDrrb0kJ48Q6JMAYAAAAgCPYcLdTBk8XK2X9Sy7Yd0hffnlR5pbHsM6x78yCNrmEQxgAAAAA0qMf/9aX+uir3nNtjoyL05Pje+knvlg04qoZHGAMAAABQrw6fKtG7W7/VF9+e0odfHtbJ4nKv+yXGRGpsv9b6xZCOat04oYFH2fAIYwAAAADqRWFphd7Z8q0ee+8LFZZVet2nc/MkdUtvpEkD26tPm1RFR4bPGoNhH8aKi4v1pz/9Sf/4xz+0b98+paWlafjw4ZoxY4ZatWrlU1/Hjx/XtGnTtHjxYh06dEjp6ekaM2aMpk2bptTU1Pr5AAAAAMAF4lhhmb76Ll85+09qSc632nkoX6UVTq/7/rBjmp65oa/SU+IaeJQXjrAOYyUlJcrMzNS6devUokULjR49Wnv37tXcuXO1dOlSrVu3Th07dqxVX0ePHtWAAQP09ddfq2PHjrruuuu0fft2Pfvss1q2bJk++eQTpaWl1fMnAgAAAAKrvNKpwtIKFZRWqKS8UiXlTuWXVOhwfom+O1Wiw6dKdbyoXPuPF2nD3mNymnP39ZPeLTW4S1MN6NQkLC5DrElYh7FHH31U69at04ABA/Tvf/9bSUlJkqRZs2bpvvvu0+TJk7Vq1apa9fXf//3f+vrrrzV27FgtWLBAUVGnf7S/+tWv9Oc//1lTp07VvHnz6umTAAAAAOdnjFFZpVPFZZU6WVyuk8XlOlFUriP5pTp0qkSHT5Vo27endLK4XMVllSoqq1BhWaXKzlHZqq3kuChd27OFJg1qrx+0SA7QpwkNYRvGysrKNHv2bEnS888/7w5ikjR16lT9/e9/1+rVq7Vp0yZdcskl5+3r4MGDmj9/vmJiYvSXv/zFHcQk6cknn9Q//vEPvf7663riiSfUvHloL88JAAAQiowxqnAaVTqNnOb03+4/Z94fLXbKaaSvDxeowulURaVReaVTFc4zf1caVTidKq80VV47Ve48/XdFpVH5mWM928vP7H+6T+O1/0rn2TGe/tupknKnjhWWqbC04pyXCwZaUmyUOjdPUq9WKRrRq4X6tUtVbFRkg5zbbsI2jK1Zs0YnT55Up06d1Ldv32rbx40bp5ycHC1ZsqTGMPavf/1LTqdTgwcP1kUXXWTZFhsbq1GjRmnOnDl6//33NWnSpEB+DAAAqjHGeLz2aD/ffpb2KvvJe3/Vz+v9mPONw1Tp8Hzj0Dn6rz7e6v0bj/3cx3r8dbzk9JuDJ4stfRrPPqqex5ztq+r+xrK/ObNv9WNc7y0/H482427z7NPjeHc/xmPfM+fw2O763GeP8fy5GDmdp187jTndl5Gc5vR755l9nTXsd7Yvc2YfeexjTr8/Mwan8+x7136ufZxn+nD3a3Q28FSe/tt55n2F88zrKqHIaU6HHUtoMkaVztPnrnCeDk1VA5XT6dGnsb4+39yv5uPVPuxsLzFREYqLilBCTJSaJ8eqeaM4XZQcq7TEGCXHRatdkwRldm+uqDBahKMuwjaMbd26VZLUr18/r9td7Tk5OQHpa86cObXqC/CH0+MXUIXHL5YKp9Pyi8b1S6WiSptrX6dT3o9x/ZJ06swvpDO/0NztRpVnfplWOq2/cF2Mxy8y15eD3XtOL2u7tXLX6S8RXr5MVD3G9QXjbNs59pHHF5Ba9Gv5cnO+c59pqP5lpnqbavjSde4vVufu1/MzVO3Xs49q5z5Hv6rWVrXfsz9Dr+f2+pmtP3dv/crLPqcPt35pdbd69Ot5rCSVl5fLSIr++N/uc3jb7+xP2nOj15e1Dgi1DRzVz2X93/Pc5zr32H36Yoja+2hFsEcANAiHQ0qOi1ZKfLTSEmPUMjVOyXHR6tEqRa1T4xUXHam46AglxUYp0fUnJpKQFWBhG8b27dsnSWrdurXX7a72vLy8Bu3LpUePHl7bc3NzlZ6erhUrgvfLorCwUJKCOoa6OH2ZgVTulCqcUoUxZ187pXKnUVmlVFoplVaefl3mlPtf0JxyvT7zr4GSO6ScfX223WnOvD/zL36nX5/9c/a9sWyz7ld9m/HYx/bfyXK/CvYIEArKvT+zBgAakkNGkQ6HIiOkSMeZPxHy2hbl8N7uvU2KcjgU4aU90uFQ1JnXEQ7Pv0/3kxzjUFyUFBMhRUc6FBMhxUVJEQ6HTn+LKD3zR1LRUZkiqVin/xwP2k/SPgoLC5WYmOjXsWEbxgoKCiRJCQneV3Fx/UDz8/MbtC/UXbnT6HiJ0fclRseKz/xdYnSoyGjvSafKGuZyaQBAGHB4vnZY2xwebxye7R4HOc5znOV91f2q9euQw1FlPN768dJvhOPsOFx9uN+7/3ZY3rtqI44zX/pd/bleu/rUmX0dZ95EnLP/s/tGOM6ew9W/48yxXv/I9frsPpGOs8e6w8mZzxDp0V71vbsPneNcntsjqu9XXFSkCIeUmMgqgaidsA1jF7rt27d7bXdVzDIzMxtyOBauiliwxmCM0Z6jhcr7vkiffXNCe48W6nB+iQpLK/XdqRIdKSjl8h0vHA4pKsKhyAiHoiIiFOGQoiIjFBnhOPMvZw5FRTqqvT/9i+f0e9cvotOvHYo40xbpcMhx5l/fXPu7fuE7HK6vCbJ8UXA4HDp06JAcktJbpFf7RX/6GIf7y4G75Vz7ePSrc2x3tzkcVb7MVO2jepvOHHPu857/3FX7Pdtt1X7PtFm+dNXi3B7fsBze+j3Puc/1mV1fnqqNzUu/3s597p+ldSzePvP5zu3J4ZA2bNgoSep/2WWW9nOpus2zV+vPp5b7nbd/xzm3OSztVfY75zHn7q+qqv9bex/Rufuv/vnP9caXn433cZzvZ13tXHX4GVb9/4DnttVnVk8eNmxY9f+eePZxvh86wl6wvyMhOPytiklhHMZcqycWFRV53e66FK9Ro0YN2hfO7ZtjRfok93v97T+79fXhgoD2HeGQYqMiFRMVoejICMVFRyghJlLxMVFKiI5UfEzkmRBzJqy4Q41DkRERpy81iIjwsu1MsIk8/T7CcaY9MuL031VCjzssud+fo0+vY4lQhMc4LNvPBKcLzdlfWn2COxDY2tFdp/+NvlfrlCCPBHYWdea/kdHcDwOgAYVtGGvbtq0kaf/+/V63u9rbtWvXoH3hdOVr//FifXkoX0fyS/XtiWKt2nVY2w6cqnUfDofUvFGsWqbGq2VqvFqnxqtFSpwubpmiFilxio2KUIzrT2QEN6MCAACgwYVtGOvdu7ckafPmzV63u9ozMjIatK9wZIzRut3HtPizA/ryu3ztPlyg/NKKGo+7KDlWvVql6uIWjdQiNV5JsVFq1ihWrVLjdVFynGKiCFgAAAC4cIVtGBs0aJBSUlKUm5urLVu2qE+fPpbtWVlZkqRRo0bV2Nfw4cMVERGh//znPzp8+LDlwc6lpaVasmSJIiMjNWLEiIB+Brs7dLJEc9fu0dKtB3XgRHGtjkmMidRlHdJ07//ror5tG9fzCAEAAID6E7alg5iYGE2ZMkWSdPfdd7vv65KkWbNmKScnR0OGDLE88Hn27Nnq3r27HnjgAUtfLVq00I033qiysjLdddddqqg4W9W5//77deTIEd18882WkBbO3t36rUbP/lhXPrFSL67efc4g1iYtXj/smKYxfVvpziGdNPtnfZUz7RrNu7U/QQwAAAC2F7aVMUl66KGHlJ2drbVr16pLly4aPHiw8vLytH79ejVr1kxz5syx7H/06FHt3LlTBw8erNbXM888o3Xr1mnhwoXq3r27Lr30Um3fvl3btm1Tly5dNGvWrIb6WBe0zfuO69cLtqjSaV3u0OGQrujcVCN6tVDXi5LUqVmSUhNigjRKAAAAoP6FbWVMkuLi4rRy5Uo9/PDDSkhI0OLFi5WXl6dJkyZp8+bN6tixY637atq0qT799FPdc889Kisr09tvv62TJ0/qV7/6lT799FOlpaXV4yexB2OMpi/5whLEmibF6qGRP9D6P/w/vfbzy3Vj/7a6pF0aQQwAAAAhL6wrY5IUHx+v6dOna/r06TXuO23aNE2bNu2c29PS0vTcc8/pueeeC+AIQ8fGvOPa8s0J9/vfXN1Vd1zZiYU2AAAAEJbCPoyhYRhj9PTyXe73P+yYpimZXYI4IgAAACC4KEmgQUxf+oXW5n7vfn/HlbW/BBQAAAAIRYQx1LttB05q7pq97vc9WiZraFdWlgQAAEB4I4yh3r2wOtf9Oj05Tq9MvEwREY4gjggAAAAIPsIY6lV+SbmWf/Gd+/20n1ys9JS4II4IAAAAuDAQxlCvPtj+nUornJKk1IRoZXa/KMgjAgAAAC4MhDHUG6fT6PV1ee73I3q1YBl7AAAA4AyWtke9+Oq7fP3s5fU6kl/qbhvdu2UQRwQAAABcWChToF48t+JrSxDrdlEjXdY+LYgjAgAAAC4shDEE3BffntKSrd9a2l6eeCkrKAIAAAAeuEwRAVVW4dSEFz+xtH0x/RolxDDVAAAAAE9UxhBQ2Tu+U0Fphfv92L6tCGIAAACAF4QxBEzO/hO6643NlrYZ1/UM0mgAAACACxthDAFhjNH0JV9Y2p4Yl6HEWKpiAAAAgDeEMQTE9m9PaWPecff7jNYpGtO3VRBHBAAAAFzYKFsgILbuP+F+3atVit6dckXwBgMAAADYAJUxBMT2b0+5X2e0TgniSAAAAAB7oDKGOjmcX6K739isDXvPXqLYoyVhDAAAAKgJlTHUyb3zt1iCmCT1aJkcpNEAAAAA9kEYg9++OVakT3Z/b2kb3KUplykCAAAAtcBlivDbyeJyy/uPfjtMbdLi5XA4gjQiAAAAwD4IY/BbSXml+3WzRrFq2yQhiKMBAAAA7IXLFOG3knKn+3VcNFMJAAAA8AXfoOE3z8pYXFRkEEcCAAAA2A9hDH4rqfAIY9GEMQAAAMAXhDH4jcsUAQAAAP/xDRp+s1ymSGUMAAAA8AlhDH7zDGOx3DMGAAAA+IQwBr+VVnCZIgAAAOAvvkHDb56VsXguUwQAAAB8QhiD37hnDAAAAPAfYQx+YzVFAAAAwH98g4bfqIwBAAAA/iOMwW8llgU8CGMAAACALwhj8FtxmefS9kwlAAAAwBd8g4bfSiu4TBEAAADwF2EMfuOeMQAAAMB/hDH4jdUUAQAAAP/xDRp+s1TGoqiMAQAAAL4gjMFvJdwzBgAAAPiNMAa/lXpcphjLZYoAAACAT/gGDb+VVXqEMZa2BwAAAHzCN2j4zbMyFkMYAwAAAHzCN2j4zVoZ454xAAAAwBeEMfil0mlU6TTu91TGAAAAAN/wDRp+qXBa38dEMpUAAAAAX/ANGn4pN9b3rKYIAAAA+IZv0PCLx/OeJVEZAwAAAHzFN2j4pcJpLY2xtD0AAADgG75Bwy8VHlksJjJCDocjeIMBAAAAbIgwBr94XqbISooAAACA7/gWDb94PO+ZSxQBAAAAP4T9t+g1a9ZoxIgRSktLU1JSkvr3769XX33V5342bdqkadOmaeDAgUpNTVVMTIzatGmjm2++WTk5OfUw8uCq4BljAAAAQJ1EBXsAwbRw4ULdcMMNcjqduvLKK9W0aVN9+OGHmjhxonJycvTUU0/Vqp+KigpdeumlkqS0tDQNHDhQiYmJ+uyzz/TGG2/orbfe0htvvKFx48bV58dpUJ6VMcIYAAAA4Luw/RZ97NgxTZ48WZWVlcrKytKqVauUlZWlL7/8Up07d9bMmTO1atWqWvd32WWXafHixTp8+LDef/99vfXWW9q1a5cefPBBlZWVafLkyTp69Gj9faAGVsFligAAAECdhO236JdfflmnTp3S6NGjNXbsWHf7RRddpCeeeEKSNHPmzFr1FRUVpU8//VSjR49WZGSkuz0iIkIzZsxQt27dlJ+fr/feey+wHyKIqIwBAAAAdRO236JdwcjbpYMjR45UXFycsrOzVVJSUqfzOBwOZWRkSJK+/fbbOvV1IbHcM8YDnwEAAACfhe236K1bt0qS+vXrV21bTEyMevbsqZKSEu3atavO59q9e7ckKT09vc59XSisqylGnntHAAAAAF6F5QIep06d0smTJyVJrVu39rpP69attXHjRuXl5bkrW/74+OOPtWnTJsXExGj48OG1Pq5Hjx5e23Nzc5Wenq4VK1b4Paa6KiwsVGGJ5Mry+SePB3U8sK/CwkJJYv6gTphHCATmEQKBeRSeCgsLlZiY6NexYVkZKygocL9OSEjwuo/rB5qfn+/3eU6dOqXJkydLkn7961+rRYsWfvd1ofFcwCOawhgAAADgM9tWxsaMGaMdO3b4dMyrr76q/v3719OIrCorK3XTTTfpq6++Uv/+/TV9+nSfjt++fbvXdlfFLDMzs85j9NeKFSsUEVUuqUKS1Cr9ImVmVr/cE6iJ618OgzmfYX/MIwQC8wiBwDwKT/5WxSQbh7E9e/Zo586dPh1TVFQkSUpKSrK0JScnV9vXVWZu1KiRX+P75S9/qaVLl6pbt2567733FBMT41c/FypWUwQAAADqxrZhbMuWLX4fm5ycrJSUFJ08eVL79+/XxRdfXG2f/fv3S5LatWvnc/+///3v9dJLL6lNmzZavny5mjZt6vdYL1Qs4AEAAADUTdiWNHr37i1J2rx5c7Vt5eXl2rZtm+Li4tS1a1ef+n3iiSf0+OOPq3nz5lq+fLnatGkTkPFeaMo9lrbnoc8AAACA78L2W/TIkSMlSVlZWdW2LV26VCUlJbrqqqsUFxdX6z5feukl/e53v1Nqaqo++OADdevWLWDjvdBUcJkiAAAAUCdh+y36tttuU3Jyst555x0tWrTI3X748GHdf//9kqT77ruv2nHdu3dX9+7ddeDAAUt7VlaW7rzzTiUlJen9999Xnz596nX8wWa5Z4yHPgMAAAA+s+09Y3WVlpamOXPmaMKECRo3bpyGDh2qJk2aKDs7WydOnNDUqVM1dOjQase5Fg0pLy93tx0+fFg33XSTnE6nOnTooBdffFEvvvhitWOvu+46XXfddfX1kRpUaeXZ1wmx3DMGAAAA+Cpsw5gkXX/99froo4/06KOPat26dSorK9PFF1+sKVOmaOLEibXup6ioSGVlZZKkzz//XJ9//rnX/dq3bx9CYezsPWOJMWE9jQAAAAC/hP236EGDBmnZsmW13t8YU62tffv2XttDWWnF2dcJMVTGAAAAAF9xsw/8UuJZGYsN+0wPAAAA+IwwBr9Y7hmjMgYAAAD4jDAGv3hWxpKojAEAAAA+I4zBL9Z7xghjAAAAgK8IY/CZMcZymWIiS9sDAAAAPiOMwWdlTslz7UgqYwAAAIDvCGPwmeclihKVMQAAAMAfhDH4zHPxDodDiosijAEAAAC+IozBZ5Zl7aMjFRHhCN5gAAAAAJsijMFnpR6VsQSWtQcAAAD8QhiDz0o87hlL5IHPAAAAgF8IY/CZ5TJFVlIEAAAA/EIYg8/KnWcvU4yNZgoBAAAA/uCbNHxmPB4yFulg8Q4AAADAH4Qx+Mzp8ZosBgAAAPiHMAafeVbGHKQxAAAAwC+EMfjMI4uJR4wBAAAA/iGMwWeelbEIKmMAAACAXwhj8JlnZYwsBgAAAPiHMAafURkDAAAA6o4wBp9ZK2OEMQAAAMAfhDH4zFoZC944AAAAADsjjMFnlspY0EYBAAAA2BthDD5zcs8YAAAAUGeEMdQJ94wBAAAA/iGMwWc89BkAAACoO8IYfOb0WMGDwhgAAADgH8IYfMZzxgAAAIC6I4yhTghjAAAAgH8IY/CZ0+M1WQwAAADwD2EMPvO8TJHVFAEAAAD/EMbgM1ZTBAAAAOqOMAafsYAHAAAAUHeEMfjMszJGFgMAAAD8QxiDzyz3jIk0BgAAAPiDMAafea6myD1jAAAAgH8IY/Ad94wBAAAAdUYYg88sqykygwAAAAC/8FUaPnN6pjHuGQMAAAD8QhiDz3jOGAAAAFB3hDH4jnvGAAAAgDojjMFnrKYIAAAA1B1hDD4zHg8ac1AZAwAAAPxCGIPPPO8ZI4sBAAAA/iGMwWfWBTxIYwAAAIA/CGPwmbEs4BG8cQAAAAB2RhiDzzzDGPeMAQAAAP4hjMFn3DMGAAAA1B1hDD7jnjEAAACg7ghj8Bn3jAEAAAB1RxiDzzwf+uwQaQwAAADwB2EMvqMyBgAAANQZYQw+sy7gQRoDAAAA/BH2YWzNmjUaMWKE0tLSlJSUpP79++vVV18NSN8///nP5XA45HA49PHHHwekzwuB01IZI4wBAAAA/ogK9gCCaeHChbrhhhvkdDp15ZVXqmnTpvrwww81ceJE5eTk6KmnnvK775UrV2rOnDlyOBwynitehACWtgcAAADqLmwrY8eOHdPkyZNVWVmprKwsrVq1SllZWfryyy/VuXNnzZw5U6tWrfKr75KSEv3iF79Qjx49NGDAgMAO/ELAPWMAAABAnYVtGHv55Zd16tQpjR49WmPHjnW3X3TRRXriiSckSTNnzvSr7xkzZujrr7/WCy+8oOjo6ICM90JiWU2R0hgAAADgl7ANY++9954kady4cdW2jRw5UnFxccrOzlZJSYlP/X7++ed68sknNXnyZF1xxRUBGeuFxnDPGAAAAFBnYRvGtm7dKknq169ftW0xMTHq2bOnSkpKtGvXrlr36XQ6dccddyg1NdVdXQtFxuM6RbIYAAAA4J+wXMDj1KlTOnnypCSpdevWXvdp3bq1Nm7cqLy8PGVkZNSq3+eff17r1q3T3//+d6WlpdVpjD169PDanpubq/T0dK1YsaJO/ddFRUWldOZhz7tzv9aKyrygjQX2VlhYKElBnc+wP+YRAoF5hEBgHoWnwsJCJSYm+nVsWFbGCgoK3K8TEhK87uP6gebn59eqz/379+vBBx/U0KFDdcstt9R9kBcwz6XtKYwBAAAA/rFtZWzMmDHasWOHT8e8+uqr6t+/f72M5+6771Zpaan++te/BqS/7du3e213VcwyMzMDch5//PmzZXIt49G1a1dlXtEhaGOBvbn+5TCY8xn2xzxCIDCPEAjMo/Dkb1VMsnEY27Nnj3bu3OnTMUVFRZKkpKQkS1tycnK1fV1l5kaNGtXY78KFC/Xuu+/q4YcfVvfu3X0akx1ZnjMWtFEAAAAA9mbbMLZlyxa/j01OTlZKSopOnjyp/fv36+KLL662z/79+yVJ7dq1q7G/JUuWSJKWL1+ujz76yOs477nnHqWkpGjSpEmaNGmS32O/EBieMwYAAADUmW3DWF317t1bH330kTZv3lwtjJWXl2vbtm2Ki4tT165da93nunXrzrnNFcqGDh3qz3AvKJ6VsQjSGAAAAOCXsFzAQzr9LDFJysrKqrZt6dKlKikp0VVXXaW4uLga+5o3b56MMV7/DBkyRJL0n//8R8YYTZs2LaCfIxgsC3iwtj0AAADgl7ANY7fddpuSk5P1zjvvaNGiRe72w4cP6/7775ck3XfffdWO6969u7p3764DBw402FgvNNwzBgAAANRd2F6mmJaWpjlz5mjChAkaN26chg4dqiZNmig7O1snTpzQ1KlTvV5S6Fo0pLy8vIFHfGGKoDIGAAAA+CVsw5gkXX/99froo4/06KOPat26dSorK9PFF1+sKVOmaOLEicEe3gXLyQIeAAAAQJ2FdRiTpEGDBmnZsmW13t94LiVYC6tWrfJxRBc+62qKpDEAAADAH2F7zxj8Z4mjZDEAAADAL4Qx+MyytD2VMQAAAMAvhDH4jIc+AwAAAHVHGIPPqIwBAAAAdUcYg888FzEhiwEAAAD+IYzBZ5aHPpPGAAAAAL8QxuAz7hkDAAAA6o4wBp85PV5zzxgAAADgH8IYfOZZGSOKAQAAAP4hjKFOuGcMAAAA8A9hDD5zcs8YAAAAUGeEMfiM54wBAAAAdUcYg88s94yRxQAAAAC/EMbgMypjAAAAQN0RxuAzKmMAAABA3RHG4DMqYwAAAEDdEcbgMypjAAAAQN0RxuAzKmMAAABA3RHG4DMqYwAAAEDdEcbgMypjAAAAQN0RxuAzp2dlLHjDAAAAAGyNMAY/nE1jERHEMQAAAMAfhDH4zOnxmiwGAAAA+IcwBp9ZF/AgjQEAAAD+IIzBZ4Z7xgAAAIA6I4zBZ6ymCAAAANQdYQw+I4wBAAAAdUcYg8946DMAAABQd4Qx+MyzMkYYAwAAAPxDGIPPPCtjXKYIAAAA+IcwBp9xzxgAAABQd4Qx+MxaGQveOAAAAAA7I4zBZ06P1xTGAAAAAP8QxuA7y2qKpDEAAADAH4Qx+Ix7xgAAAIC6I4zBZ07uGQMAAADqjDAGn1meMybSGAAAAOAPwhh8YjyXUhQLeAAAAAD+IozBJ1WymCK4ThEAAADwC2EMPnFWSWNkMQAAAMA/hDH4xFmlMsY9YwAAAIB/CGPwiRGVMQAAACAQCGPwSdV7xnjoMwAAAOAfwhh8wj1jAAAAQGAQxuCTaveMURkDAAAA/EIYg0+qPmeMyhgAAADgH8IYfEJlDAAAAAgMwhh8QmUMAAAACAzCGHxCZQwAAAAIDMIYfEJlDAAAAAiMqGAPAPYSGeFQp5QISUZJySmKoDIGAAAA+IUwBp+kJsTowctjJUmZmYOCPBoAAADAvsL+MsU1a9ZoxIgRSktLU1JSkvr3769XX33V7/6MMZo3b56uvPJKpaWlKT4+Xh07dtTPfvYzbd++PYAjBwAAAGBnYV0ZW7hwoW644QY5nU5deeWVatq0qT788ENNnDhROTk5euqpp3zqr6SkRGPGjNG//vUvpaWladCgQYqPj9fu3bv1z3/+UyNGjFCPHj3q6dMAAAAAsJOwDWPHjh3T5MmTVVlZqYULF2rs2LGSpO+++05XXHGFZs6cqR//+McaOnRorfu888479a9//Uu33367nn32WcXHx7u3HTx4UOXl5YH+GAAAAABsKmwvU3z55Zd16tQpjR492h3EJOmiiy7SE088IUmaOXNmrfv79NNP9fe//139+/fXiy++aAliktSiRQu1bds2MIMHAAAAYHthG8bee+89SdK4ceOqbRs5cqTi4uKUnZ2tkpKSWvX30ksvSZKmTJnCs7cAAAAA1ChsL1PcunWrJKlfv37VtsXExKhnz57auHGjdu3apYyMjBr7W7FihSRp4MCBys3N1fz58/XNN9+oWbNmGj58uK644orAfgAAAAAAtuYwVZ/iGwZOnTqllJQUSdLJkyeVnJxcbZ8xY8Zo8eLFevfddzVq1Kjz9ldSUuK+LPFvf/ub7rnnHpWWllr2ueGGG/Tqq68qJiamVmM810Ifubm5Sk9P15w5c2rVT30oLCyUJCUmJgZtDAgNzCUEAvMIgcA8QiAwj8LT5MmTlZiY6NfK6WF5mWJBQYH7dUJCgtd9XP8nys/Pr7G/EydOuF/fddddGjlypHbs2KETJ05o0aJFatq0qRYsWKAHH3ywbgMHAAAAEDJse5nimDFjtGPHDp+OefXVV9W/f/+Aj8XpdLpfd+/eXW+99ZYiIk7n3DFjxig2NlYjR47U7Nmz9fDDD3utxFV1rmTtqphlZmYGYOT+cV2SGcwxIDQwlxAIzCMEAvMIgcA8Ck91qYTaNozt2bNHO3fu9OmYoqIiSVJSUpKlzVs4cpWZGzVqVGO/nv3dcsst7iDmMmLECDVv3lyHDx/Wp59+qquuusqncQMAAAAIPba9THHLli0yxvj0x/XMsOTkZPc9Y/v37/fav6u9Xbt2NY4lOTlZjRs3liS1b9/e6z6u9sOHD/vwKQEAAACEKtuGsbrq3bu3JGnz5s3VtpWXl2vbtm2Ki4tT165da9Vfnz59JEnHjx/3uv3YsWOSrFU0AAAAAOErbMPYyJEjJUlZWVnVti1dulQlJSW66qqrFBcXV6v+fvKTn0iSVq1aVW3bvn37tHfvXklS3759/RswAAAAgJAStmHstttuU3Jyst555x0tWrTI3X748GHdf//9kqT77ruv2nHdu3dX9+7ddeDAAUv75MmT3asmvvvuu+72oqIi3XXXXaqoqNCIESPUpk2bevpEAAAAAOzEtgt41FVaWprmzJmjCRMmaNy4cRo6dKiaNGmi7OxsnThxQlOnTnXfY+bJtWhIeXm5pT05OVmvv/66Ro0apeuuu06XX365WrRoofXr1+vbb79V+/bt9be//a0hPhoAAAAAGwjbypgkXX/99froo490zTXX6LPPPtP777+vzp07a968eZo5c6bP/V1zzTXasGGDxowZo6+//lpLly5VXFycfv3rX2vDhg1q1apVPXwKAAAAAHYUtpUxl0GDBmnZsmW13t8Yc97tvXv31sKFC+s6LAAAAAAhLqwrYwAAAAAQLIQxAAAAAAgCwhgAAAAABAFhDAAAAACCwGFqWpECF5RGjRqpvLxcnTp1CtoYCgsLJUmJiYlBGwNCA3MJgcA8QiAwjxAIzKPwlJubq+joaOXn5/t8LJUxm0lMTFR0dHRQx3Do0CEdOnQoqGNAaGAuIRCYRwgE5hECgXkUnqKjo/0O4FTG4LMePXpIkrZv3x7kkcDumEsIBOYRAoF5hEBgHsFXVMYAAAAAIAgIYwAAAAAQBIQxAAAAAAgCwhgAAAAABAFhDAAAAACCgNUUAQAAACAIqIwBAAAAQBAQxgAAAAAgCAhjAAAAABAEhDEAAAAACALCGAAAAAAEAWEMAAAAAIKAMAYAAAAAQUAYAwAAAIAgIIyh1oqLi/XII4+oa9euiouLU8uWLTV58mQdOHAg2ENDAysqKtLixYv185//XN26dVNcXJwSExPVu3dvTZ8+XQUFBec8dt68eerfv7+SkpKUlpamESNGaO3atec935o1azRixAilpaUpKSlJ/fv316uvvhroj4ULwPfff6/mzZvL4XCoc+fO592XuYSqjhw5ot/85jfq1q2b4uPjlZaWpn79+um3v/2t1/2XLFmiIUOGKDk5WcnJyRo6dKjee++9855j+/btGj9+vJo1a6b4+Hj16tVLzzzzjJxOZ318JDSwDRs2aMKECWrZsqWio6OVmpqqwYMHa+7cuTLGVNu/srJSTz/9tHr16qX4+Hg1a9ZMEyZM0I4dO857Hn/mHkKUAWqhuLjY/PCHPzSSTIsWLcyECRNM//79jSTTrFkzk5ubG+whogG99NJLRpKRZH7wgx+Y8ePHm2uuucY0atTISDLdu3c33333XbXj7r33XiPJxMfHm9GjR5trrrnGREVFmcjISPP22297PVdWVpaJjIw0DofDDBkyxFx//fUmNTXVSDL33XdfPX9SNLSJEycah8NhJJlOnTqdcz/mEqrauHGjadKkiZFkevToYW644QZz7bXXmnbt2pnIyMhq+z/99NNGkomKijLDhw83o0ePNvHx8UaS+fOf/+z1HGvXrnXv079/fzNhwgSTnp5uJJnx48cbp9NZ3x8T9cj13whJpl+/fmbChAlm2LBhJioqykgyP/vZzyz7V1ZWmjFjxhhJJjU11Vx//fVmyJAhxuFwmISEBLN+/Xqv5/Fn7iF0EcZQKw8++KCRZAYMGGDy8/Pd7TNnzjSSzJAhQ4I3ODS4efPmmTvuuMN88cUXlvZvv/3W9O3b10gyN954o2Xb8uXLjSTTpEkTs2vXLnf72rVrTUxMjElNTTXHjx+3HPP999+b5ORkI8ksXLjQ3X7o0CHTuXNnI8msXLky4J8PwZGdnW0kmTvuuOO8YYy5hKoOHz5smjZtahISEsw777xTbXvVL8VffvmliYyMNLGxsWbt2rXu9p07d5omTZqYqKgo89VXX1mOKSsrMx06dDCSzKxZs9zt+fn5ZsCAAUaSmTt3bmA/GBpMeXm5ad68uZFk3njjDcu2L774wqSlpRlJZsWKFe521z9MdunSxRw6dMjdnpWVZSSZzp07m/Lycktf/sw9hDbCGGpUWlpqUlJSjCSzefPmatszMjKMJLNx48YgjA4XmrVr1xpJJjY21pSWlrrbr732WiPJPP3009WO+dWvfmUkmaeeesrS/vjjjxtJZvTo0dWOWbRokZFkfvzjHwf6IyAIioqKTKdOnczFF19sdu3add4wxlxCVb/85S+NJPP888/7tP+9995bbdusWbOMJDNlyhRL+4IFC4wk07t372rHbNq0yUgyPXv29Gf4uAB8/vnnRpLp1q2b1+2u/7Y8/vjj7rYf/OAHRpLXavxPfvITI8lkZWVZ2v2Zewht3DOGGq1Zs0YnT55Up06d1Ldv32rbx40bJ+n09c9A7969JUmlpaX6/vvvJZ2+33DFihWSzs4XT+eaQ67r570dM3LkSMXFxSk7O1slJSWB+wAIij/+8Y/avXu3XnjhBUVHR59zP+YSqiouLtbrr7+uxMRE3XrrrbU65nzzwZ851K9fP3Xs2FHbtm3T3r17fRk+LhCxsbG12q9JkyaSpD179mjHjh2Kj4/XyJEjq+3nzzzi+1R4IoyhRlu3bpV0+peNN672nJycBhsTLly7d++WJEVHRystLU2StHPnTpWWlqpZs2Zq3bp1tWPONYfON/diYmLUs2dPlZSUaNeuXQH9DGhYOTk5mjlzpm699VYNHjz4vPsyl1DVxo0blZ+fr759+yo+Pl7Lli3T1KlTddddd+mZZ57Rt99+a9n/xIkT2rdvnyR5/QfGNm3aqGnTpsrLy9OpU6fc7fwuDG0dO3ZUp06dtHPnTr355puWbTt27NDrr7+uxo0ba8yYMZLOzoeePXt6/Qckb/PB37mH0EYYQ41c/+Hw9sXHsz0vL6/BxoQL17PPPitJGj58uPtfGmuaQ4mJiUpNTdXx48eVn58vSTp16pROnjx53uOYe/bndDp12223KTU1VU888USN+zOXUNUXX3whSWrevLmuu+46jRgxQk8//bT++te/6te//rU6d+6s+fPnu/d3zaHGjRsrMTHRa5/e5gO/C0NbZGSk/v73vys1NVU33XSTLrnkEv30pz9VZmamMjIy1Lp1a3344Yfuf2T0Zz74O/cQ2ghjqJFrmfKEhASv213/QXF98UH4ev/99/XKK68oOjpaM2bMcLfXNIek6vPIc3l85l7o+vOf/6wNGzboySefdF/+cz7MJVR1/PhxSdK7776rf/3rX3r++ed1+PBh7d27V7/5zW9UXFysiRMnasuWLZL8m0O1OY45ZH+DBg3S6tWr1bFjR23evFkLFizQypUrFRERoR/96Efq2LGje19/5oO/cw+hjTAGICC+/PJL3XzzzTLG6Mknn3TfOwacy759+/TQQw9pyJAhmjRpUrCHA5tyPd+roqJC06dP11133aVmzZqpXbt2evLJJzV+/HiVl5frySefDPJIcaGbP3+++vfvrzZt2mj9+vUqKCjQrl27NGnSJM2cOVOZmZkqLS0N9jARYghjqFFSUpKk0w/69aawsFCS1KhRowYbEy4sBw4c0PDhw3X8+HFNnTpV9957r2V7TXNIqj6PXMec7zjmnr3dfffdKisr0wsvvFDrY5hLqMrzf19vC3i42lavXm3Z35c5VJvjmEP29tVXX2nixIlq2rSpli5dqv79+ysxMVFdunTRiy++qB//+MfavHmz5syZI8m/+eDv3ENoI4yhRm3btpUk7d+/3+t2V3u7du0abEy4cBw7dkxXX3218vLydOutt+qpp56qtk9Nc6iwsFAnTpxQ48aN3b+AkpOTlZKSct7jmHv2tnTpUiUkJOjOO+/U0KFD3X9++tOfSjod8l1thw4dksRcQnWu/80SEhLUrFmzatvbt28vSTp8+LCks3Po+PHj7i++VXmbD/wuDG3/+Mc/VF5eruHDh1sCvsuECRMkSR999JEk/+aDv3MPoY0whhq5LjfbvHmz1+2u9oyMjAYbEy4MBQUFuvbaa/XFF19o7Nixeumll+RwOKrt161bN8XGxurIkSM6cOBAte3nmkPnm3vl5eXatm2b4uLi1LVr10B8HATBiRMntHr1asuf9evXS5JKSkrcba4l55lLqMq1Kl1xcbHXS8iOHTsm6WxVIjU11f2l+LPPPqu2/zfffKOjR4+qXbt2Sk5OdrfzuzC0uUKQ6x9uqnK1u+5RdM2Hbdu2qby8vNr+3uaDv3MPoY0whhoNGjRIKSkpys3Ndd8A7SkrK0uSNGrUqAYeGYKptLRUo0eP1qeffqprrrlG8+fPV2RkpNd94+PjlZmZKUl66623qm0/1xxyPbvFtd3T0qVLVVJSoquuukpxcXF1+iwIDmOM1z979uyRJHXq1Mnd5qpuMJdQVdu2bdW7d28ZY9yXInpytXkuJX6++eDPHPrss8+0e/du9ezZ0z1XYS/p6emSTj8qwZsNGzZIOltp7dChg37wgx+ouLjY/ewwT/7MI75PhalgPW0a9vLggw8aSWbgwIGmoKDA3T5z5kwjyQwZMiR4g0ODq6ioMGPGjDGSzODBg01hYWGNxyxfvtxIMk2aNDG7du1yt69du9bExsaa1NRUc/z4ccsx33//vUlOTjaSzMKFC93t3333nencubORZFauXBmoj4ULxJ49e4wk06lTJ6/bmUuo6o033jCSTK9evcy3337rbv/ss89MWlqakWT++c9/utu//PJLExkZaWJjY80nn3zibt+1a5dp0qSJiYqKMl999ZXlHGVlZaZDhw5Gkpk1a5a7vaCgwAwYMMBIMnPnzq2/D4l6tWnTJiPJSDJ/+ctfLNs++eQTk5iYaCSZ5cuXu9tfeuklI8l06dLFfPfdd+72hQsXGkmmc+fOpry83NKXP3MPoY0whlopLi42l19+uZFkWrRoYSZMmOB+36xZM5ObmxvsIaIBPfPMM+5fWmPGjDETJ070+ufIkSOW4+69914jySQkJJjRo0eba6+91kRFRZnIyEjz9ttvez1XVlaWiYiIMA6HwwwbNsyMGzfOpKamGklm6tSpDfBp0dBqCmPGMJdQ3cSJE40kk5qaakaMGGGGDRtmYmNjjSRz++23V9t/1qxZRpKJiooy1157rRk9erSJj483ksxzzz3n9Rxr1qxx73P55ZebCRMmmBYtWhhJZty4ccbpdNb3x0Q9+s1vfuP+3dajRw8zfvx4M2jQIBMREWEkmTvuuMOyf2VlpfsfJhs3bmzGjRtnhg4dahwOh4mPjzfr1q3zeh5/5h5CF2EMtVZUVGQefvhh06lTJxMTE2PS09PNpEmTzDfffBPsoaGB/c///I/7F9b5/uzZs6fasXPnzjWXXHKJSUhIMKmpqWb48OFmzZo15z3fxx9/bIYPH25SU1NNQkKCufTSS828efPq6dMh2GoTxoxhLsHK6XSav/3tb+45kZiYaAYMGHDe/33fffddM3jwYJOUlGSSkpLM4MGDzZIlS857nm3btpnrr7/eNGnSxMTFxZkePXqYWbNmmcrKykB/JATBokWLzNVXX+2uUjVu3NgMGzbMvPnmm173r6ioMDNnzjQ9evQwcXFxpkmTJmbcuHFm+/bt5z2PP3MPoclhjDGBvvQRAAAAAHB+LOABAAAAAEFAGAMAAACAICCMAQAAAEAQEMYAAAAAIAgIYwAAAAAQBIQxAAAAAAgCwhgAAAAABAFhDAAAAACCgDAGAAAAAEFAGAMAAACAICCMAQAAAEAQEMYAAAAAIAgIYwAAAAAQBIQxAAAAAAgCwhgAAAAABAFhDAAAAACCgDAGAAAAAEHw/wG0DRcNWqMngQAAAABJRU5ErkJggg==", + "image/png": "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", "text/plain": [ "
" ] @@ -1154,9 +1240,14 @@ }, { "cell_type": "code", - "execution_count": 29, + "execution_count": 32, "id": "b07f18b5-7009-4458-aec2-4a49fd595348", - "metadata": {}, + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:29:22.452067Z", + "start_time": "2024-06-24T17:29:22.432614Z" + } + }, "outputs": [], "source": [ "def get_topk_indices(current_score: torch.Tensor, topk: int = 1) -> torch.Tensor:\n", @@ -1178,9 +1269,14 @@ }, { "cell_type": "code", - "execution_count": 30, + "execution_count": 33, "id": "1a22d21e-3d87-4a88-9665-517487f71ef6", - "metadata": {}, + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:29:25.919010Z", + "start_time": "2024-06-24T17:29:25.909034Z" + } + }, "outputs": [ { "data": { @@ -1188,7 +1284,7 @@ "(917, 907)" ] }, - "execution_count": 30, + "execution_count": 33, "metadata": {}, "output_type": "execute_result" } @@ -1199,9 +1295,14 @@ }, { "cell_type": "code", - "execution_count": 31, + "execution_count": 34, "id": "b90e2a6a-e6bc-4865-a0f8-4e8d6d8012c6", - "metadata": {}, + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:30:37.631071Z", + "start_time": "2024-06-24T17:29:31.948220Z" + } + }, "outputs": [], "source": [ "topk_lst = [5, 10, 15, 20, 25, 30] \n", @@ -1211,8 +1312,8 @@ " keep_indices = get_topk_keep_indices(scores[\"all_modules\"].sum(dim=0), topk=topk)\n", " \n", " new_loss = 0.\n", - " for _ in range(num_iter):\n", - " new_loss += train_and_evaluate(modified_train_dataset=torch.utils.data.Subset(train_dataset, keep_indices), query_dataset=single_query_dataset)\n", + " for i in range(num_iter):\n", + " new_loss += train_and_evaluate(current_train_dataset=torch.utils.data.Subset(train_dataset, keep_indices), current_query_dataset=single_query_dataset, seed=seed_lst[i])\n", " new_loss /= num_iter\n", " if_removed_loss_lst.append(new_loss)" ] @@ -1222,17 +1323,17 @@ "id": "6194cf30-e728-4484-b4a0-60736016feb6", "metadata": {}, "source": [ - "We compare the results with the random baseline, where the same number of data points are removed." + "We compare the results with the random baseline, where the same number of (randomly selected) data points are removed." ] }, { "cell_type": "code", - "execution_count": 32, + "execution_count": 35, "id": "868b1ff7796cfafa", "metadata": { "ExecuteTime": { - "end_time": "2024-03-12T21:16:31.778208Z", - "start_time": "2024-03-12T21:16:22.922187Z" + "end_time": "2024-06-24T17:31:48.238509Z", + "start_time": "2024-06-24T17:30:42.784748Z" }, "collapsed": false, "jupyter": { @@ -1249,31 +1350,36 @@ " keep_indices = random_indices[topk:]\n", " \n", " new_loss = 0\n", - " for _ in range(num_iter):\n", - " new_loss += train_and_evaluate(modified_train_dataset=torch.utils.data.Subset(train_dataset, keep_indices), query_dataset=single_query_dataset)\n", + " for i in range(num_iter):\n", + " new_loss += train_and_evaluate(current_train_dataset=torch.utils.data.Subset(train_dataset, keep_indices), current_query_dataset=single_query_dataset, seed=seed_lst[i])\n", " new_loss /= num_iter\n", " random_removed_loss_lst.append(new_loss)" ] }, + { + "cell_type": "markdown", + "id": "93d0921c53c621b4", + "metadata": { + "collapsed": false, + "jupyter": { + "outputs_hidden": false + } + }, + "source": [ + "We also compare the results with the gradient dot product baseline (where the Hessian is approximated as identity)." + ] + }, { "cell_type": "code", - "execution_count": 33, + "execution_count": 36, "id": "9e6fd656-c1e5-42cb-87f4-d7690f473fe9", - "metadata": {}, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "Computing pairwise scores (query gradient) [0/1] 0%| [time left: ?, time spent: 00:00]\n", - "Computing pairwise scores (training gradient) [1/1] 100%|███████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\u001b[A\n", - "Computing pairwise scores (query gradient) [1/1] 100%|██████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n", - "Computing pairwise scores (query gradient) [0/1] 0%| [time left: ?, time spent: 00:00]\n", - "Computing pairwise scores (training gradient) [1/1] 100%|███████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\u001b[A\n", - "Computing pairwise scores (query gradient) [1/1] 100%|██████████████████████████████████████████████████████████████████████████████████████████████████ [time left: 00:00, time spent: 00:00]\n" - ] + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:32:09.792089Z", + "start_time": "2024-06-24T17:32:09.360547Z" } - ], + }, + "outputs": [], "source": [ "from kronfluence import FactorArguments\n", "\n", @@ -1298,9 +1404,14 @@ }, { "cell_type": "code", - "execution_count": 34, + "execution_count": 37, "id": "220ea9ef-778c-48e5-8c65-6f1efac05233", - "metadata": {}, + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:33:28.605345Z", + "start_time": "2024-06-24T17:32:23.248126Z" + } + }, "outputs": [], "source": [ "id_removed_loss_lst = []\n", @@ -1309,17 +1420,22 @@ " keep_indices = get_topk_keep_indices(identity_scores, topk=topk)\n", " \n", " new_loss = 0.\n", - " for _ in range(num_iter):\n", - " new_loss += train_and_evaluate(modified_train_dataset=torch.utils.data.Subset(train_dataset, keep_indices), query_dataset=single_query_dataset)\n", + " for i in range(num_iter):\n", + " new_loss += train_and_evaluate(current_train_dataset=torch.utils.data.Subset(train_dataset, keep_indices), current_query_dataset=single_query_dataset, seed=seed_lst[i])\n", " new_loss /= num_iter\n", " id_removed_loss_lst.append(new_loss)" ] }, { "cell_type": "code", - "execution_count": 35, + "execution_count": 38, "id": "6af9e2b1-b8c1-4825-97ef-e9accc6c2e91", - "metadata": {}, + "metadata": { + "ExecuteTime": { + "end_time": "2024-06-24T17:33:33.714869Z", + "start_time": "2024-06-24T17:33:33.597472Z" + } + }, "outputs": [ { "data": { @@ -1327,13 +1443,13 @@ "Text(0.5, 0, 'Number of Training Samples Removed')" ] }, - "execution_count": 35, + "execution_count": 38, "metadata": {}, "output_type": "execute_result" }, { "data": { - "image/png": "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", + "image/png": "iVBORw0KGgoAAAANSUhEUgAAA18AAAKHCAYAAAB+XPKfAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjguMywgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/H5lhTAAAACXBIWXMAABcSAAAXEgFnn9JSAADvzUlEQVR4nOzdd1gUV9sG8HuBpRdFUewIClixixXsGmKLMbElUWOMBXs0thh719ijiS2vJmqisaKxg4oBK2BBVAQURQVBell25/uDj1ECSFuYXbh/7+X1es6U8ywT1r13Zs7IBEEQQERERERERMVKR+oCiIiIiIiIygKGLyIiIiIiohLA8EVERERERFQCGL6IiIiIiIhKAMMXERERERFRCWD4IiIiIiIiKgEMX0RERERERCWA4YuIiIiIiKgEMHwRERERERGVAIYvIiIiIiKiEsDwRUREREREVAIYvoiIiIiIiEoAwxcREREREVEJ0JO6AMqdtbU1EhMTUbNmTalLISIiIiIq854+fQoTExO8fPmyUNvzzJcGS0xMhEKhkLoMJCYmIjExUeoyqAh4DLUfj6F24/HTfjyG2o/HUPtpwjFUKBRFqoFnvjRY5hmve/fuSVrHhQsXAACdO3eWtA4qPB5D7cdjqN14/LQfj6H24zHUfppwDBs0aFCk7Xnmi4iIiIiIqAQwfBEREREREZUAhi8iIiIiIqISwPBFRERERERUAhi+iIiIiIiISgDDFxERERERUQngVPOllCAIEARBrftUqVRq3R+VPB5DzSCTySCTyaQug4iIiEoYw1cpolQq8ebNG8THxyMtLU1t+zUxMQEABAUFqW2fVLJ4DDWPvr4+zMzMUKFCBejq6kpdDhEREZUAhq9SQqlU4unTp0hJSVH7vjM/uJP24jHUPGlpaXjz5g0SExNRs2ZNBjAiIqIygOGrlHjz5g1SUlKgq6uLypUrw8TEBDo66rmlLy4uDgBgbm6ulv1RyeMx1CwqlQqJiYl49eoVUlJS8ObNG1SqVEnqsoiIiKiYMXyVEvHx8QCAypUrw8LCQq37zgxx6gpzVPJ4DDWLjo6O+Hv64sULxMfHM3wRERGVAfwkVgoIgiDe48XLy4i0R+bva1pamtonyCEiIiLNw/BVCrz/oY1nNoi0x/u/rwxfREREpR8/qRMREREREZUAhi8iIiIiIqISwPBFRERERERUAjjbIRERERERaTSFQoGgoCAkJSXB3NwcTk5OkMvlUpdVYDzzRaWaTCbL8idzim9nZ2esW7cOCoVC6hJz5erqCplMhtDQUKlLISIiIpJEZGQkfpj3A6yrW2PG/BlY8vMSdOnTBdbVrfHDvB8QGRkpdYkFwjNfVCgKhQL+/v6Ii4vTim8fvvrqKwCAUqlEaGgorl69Cl9fX5w4cQL//PMP9PT4q0BERESkSQIDA+Ha1RWCjYByY8rBqLYRZDIZBEFA8pNkbDu/Db/s+AVe573g6Ogodbn5wk+cVCCRkZHYsHETNv+8DekyPchNzKFIjIOekI7xY7/FxAnusLKykrrMbHbv3p2l7evrC1dXV5w/fx779+/HsGHDpCmMiIiIiLKJjIyEa1dX6HXSg2U3yyzLZDIZjO2MYWxnjOiz0XDt6oo7t+9o5GfQ/+Jlh5RvgYGBaNC4KTYfuQyDXjNQfsRWmH2+CuVHbIVBrxnYfOQyGjRuigcPHkhdap5at26N4cOHAwBOnz4tbTFERERElMWGjRsg2AjZgtd/WXazhKqWChs3bSyhyoqG4YvyJSoqCi6du0HV0A3mH02DQVUHyGQyABnfPhhUdYD5R9OgaugGl85dteL62wYNGgAAXr9+naVfEATs27cPgwYNgr29PUxMTGBmZoZWrVphy5YtUKlU2fY1f/58yGQy7N69G3fu3EGfPn1Qvnx5mJiYwMXFBVevXs2xBqVSidWrV8PR0RGGhoaoUaMGJk2ahLi4uA/W/u+//6Jv376wsrKCgYEBbGxsMG7cOLx48SLburt374aFhQWWLVuG4OBgfPbZZ6hYsSLMzc3Rq1cv3L9/HwCQnp6OpUuXwt7eHoaGhqhTpw42b96cr58lERERkbooFAps2bYFZl3N8rW+WRczbNm2RaPv5c/E8EX5snXbL0i3sodp8z4fXM+0eR8oKtpj4ybN/9AeHx8PAKhUqVKW/tTUVAwZMgTnzp2DtbU1evfuDWdnZ9y7dw/jx4/HyJEjc93njRs34OzsjNDQUPTo0QN169bFpUuX0KVLF9y9ezfb+sOGDcP06dPx7NkzdO/eHS1btsRvv/2Gzp07IzU1Nccx9u7diw4dOuDYsWNwcHDAJ598AgMDA/z8889o1qxZrmcew8LC0KpVK9y9exddu3aFjY0N/vnnH7i6uuLly5f49NNPsXLlSjRo0ACurq549uwZ3N3d8euvv+b3R0pERERUZP7+/kjXTYdRbaN8rW9kawSFTsZ8BJqO93yVAYIgIC4lvdDbR8cn45cdO2H00cx8rW/Q5GNs+nklJk6bUaRJOMwN9cSza8Xhn3/+AQD07NkzS7+enh4OHz4MNze3LPVHRkbio48+wm+//YaRI0eiY8eO2fa5efNmrF+/HhMnThT7pkyZgnXr1mHlypX43//+J/YfOHAA+/fvR82aNeHl5QUbGxsAGWfiunTpgps3b2bb/7NnzzB69GgAwNGjR9GnT0YYVqlUmDZtGtatW4cvvvgC169fz7btvn37MHPmTCxdulS8WXXkyJHYvXs3unTpAh0dHTx69Ei8Xvr8+fPo2rUrlixZgm+++SZfP1MiIiKiooqLi4O+mX6+PwfKZDLom+nneeWQJmD4KgPiUtLhtOBMobdPjXiExHQdWFSxz9f6+lXsEZUGNJrwCwyq1C30uP4/doeFkXpnUFSpVAgJCcHq1atx6dIl9O3bF59//nmWdfT09NCvX79s21pZWWHZsmXo1q0bjh49mmP4ateuXZbgBQBz587FunXrcOnSpSz9W7ZsAZBxyWJm8AIyzsStWrUKvXr1yrb/7du3Izk5GYMHDxaDFwDo6Ohg+fLl+PPPP3Hjxg14e3ujXbt2Wba1sbHBwoULs1wuOmXKFOzevRv379/HuXPnstyo2qVLFzRt2hS3b99GaGholhqJiIiIiou5uTnS4tIgCEK+ApggCEiLT4O5uXkJVFc0DF+UJ1VaEnSMzQv07YOOkTlUaUnFXFn+5VT7N998g23btuX6uvz8/HDmzBmEhYUhKSkJgiCIlyo+evQox226d++era9ChQqwtLRERESE2KdQKODj4wMA2cIfkHE2rnz58oiJicnSf/nyZQDA0KFDs21jYGCAgQMHYv369bh8+XK28NW+fftsZyJtbW0BAHK5HK6urtn2aWtri9u3byMiIoLhi4iIiEqEXjU9JKckIzkkGca2xnmun/wkGXKVHE5OTiVQXdEwfFGedPSNoUqKK9C3D6rkOOjo5/3LUlIyn/OVkpICf39/PHjwAL/++ivatm0rznqYKS0tDcOHD8e+ffty3V9mCPuv6tWr59hvZmaG6Ohosf3mzRukpaXBysoKxsY5/5xq1aqVLXxlTqiRWxDK7H/+/Hm2ZVWrVs3WZ2pqCgCwtraGrq5urstzu/+MiIiISJ1OPDmB+Vfno5xLObw58wbGY/L+PBl/Ph7jvh2n0c+czcTwVQaYG+rB/8fsZ2TyKzq6BZp6LEVaxEMYVHXIc/20iIcw0wfubBxd5Hu+1OW/z/latWoVZsyYgfHjx6NTp06oVauWuGzt2rXYt28fGjVqhJUrV6JZs2YoX7485HI5Hj58CAcHBwiCkOM4OjrSzmHzoXD8odqkrpuIiIjKNqVKifW31mPXvV0AgApdKiB4QTDenHmDCt0r5Lpd9Nlo6ITpYIL7hJIqtUj4iasMkMlksDCSF/pPBTMjjP56JFL9TuRrvFS/E3AfOwYVzY2LNG5xTrYxffp0dO/eHUlJSViwYEGWZYcPHwaQMUFFz549UalSJTFEPnnyRC3jV6hQAfr6+oiMjERycnKO6zx9+jRbX+bZq7CwsBy3CQ0NBQBUq1ZNLXUSERERFbe4tDi4X3AXgxcAVLCqgD1/74HKS4VXv7xCUnCS+OW3IAhIepyEV7+8gtJTCc9znlrxgGWA4Yvyacy3oyGPeoiEm8c+uF7CzWOQRz3EBPfxJVRZ4S1fvhwAsGfPnixhJvNSv5wuIfzzzz/VMrZcLkfr1q1z3eeZM2eyXKaYqUOHDgCQ4yWRaWlp+Ouvv7KsR0RERKTJQmJDMNRjKK48vyL21SlXB/vc9uHzjp/jzu07GNN1DN5ue4uw78MQvjAcz2Y/Q+yvsRjTdQzu3L4DR0dHCV9BwTB8Ub5UrFgRXhfOQfeeB+JOrkHqi6As3z6kPn+AuJNroHvPA14XzmnFtw9NmzZFv379kJ6ejpUrV4r99vYZszpu3bo1y/oHDx7MMlV8UY0dOxYA8OOPP2Y5yxUVFYXp06fnuM3XX38NIyMj7N+/Hx4eHmK/SqXC7Nmz8fz5czRv3jzbZBtEREREmuZS+CUM8RiC0LhQsa9LzS74/aPfUcOsBoCM2aYXLliIl+EvsXL+SswdNxfnj51HxLMILFywUCs+c76P4YvyzdHREXf9b8O9f0eknlqJmF1jEL9/OmJ2jUHa6VVw798Rd/1va9W3D/Pnz4dMJsPOnTvx8uVLAMCMGTOgq6uLmTNnokWLFhgyZAhatmyJgQMHYsqUKWobe/DgwRg4cCDCwsJQv3599O3bFwMGDEDdunWhp6cHZ2fnbNvUrFkT27Ztg0qlQu/evdGhQwcMGTIE9evXx5o1a1C5cmXs3btXbTUSERERqZsgCNh5dyfcz7sjQZEg9o91Gou1rmthLM8+yYZcLoeDgwOaNm2KFi1aaMXkGjlh+KICyfj2YT5evXiGCyeP4K+dm3Dh5BG8fP4MCxfM17pvH5ycnNC/f3+kpKRg7dq1AICOHTviypUr6Ny5M548eYITJ05AX18fhw4dwvjx6r2c8o8//sCKFStQrVo1/PPPP/Dx8cGQIUNw4cIFGBgY5LjNF198gcuXL+Pjjz9GYGAgDh48iOTkZIwdOxY3b97UqvBLREREZUtKegpmXp6Jn27+BAEZV1EZ6RlhretajGsyDjqy0h1PZEJu07aR5Bo0aAAAuHfv3gfXU6lUCAoKAgA4ODiofea6zKeFa8OD6yhnPIaaqSC/uxcuXAAAdO7cuURqI/Xi8dN+PIbaj8dQei8TX2LSxUm4/+a+2FfNtBrWd1oPB8u8Z9TWhGOY38/nueFU80REREREVKxuv76NKRen4E3KG7GvlXUrrHZZjfKG5SWsrGQxfBERERERUbE59PAQFvsuRroqXewb7DgY01tOh1xHO+/dKiyGLyIiIiIiUjuFSoFV11dh34N3j8jR09HD3NZzMcB+gISVSYfhi4iIiIiI1OptyltM85qGay+viX2WhpZY12kdmlZqKmFl0mL4IiIiIiIitXkY8xATL0zE84TnYl89y3rY0HkDrE2sJaxMeqV7LkciIiIiIiox58POY9jJYVmCV6/avfBbr9/KfPACeOaLiIiIiIiKSCWosM1/G7b4bxH7ZJBhUrNJGNlwJGQymYTVaQ6GLyIiIiIiKrQkRRLmXJmDc0/PiX2mclOs6LgCHat3lLAyzcPwRUREREREhfIs/hkmXpiIx28fi3025jZY33k9bC1sJaxMMzF8ERERERFRgflG+GKa1zTEpsaKfe2qtcPKjithrm8uYWWai+GLiIiIiIjyTRAE7HuwDyuvr4RSUIr9IxqMwKRmk6CroythdZqN4YuIiIiIiPIlTZmGJb5L8Pejv8U+A10DzG87Hx/bfixhZdqB4YuIiIiIiPIUlRyFKRenwC/ST+yrZFwJGzptQIOKDaQrTIvwOV9UKslksgL9sbGxkbTe3bt3QyaTYfjw4SUy3uPHj6Gvr4/p06dn6Z8/fz5kMhnmz59foP1pws8wN6GhoZDJZHB1dS3WcQRBQNOmTdGoUSOoVKpiHYuIiKik3XtzD4NODMoSvJysnLDfbT+DVwHwzBeVSl999VW2vitXriA4OBhOTk5o0qRJlmUVK1Ysoco0w6xZs6Cvr48ZM2ZIXUqR7d69GyNGjMCPP/5YoNBY2O1yI5PJMG/ePHzyySfYvXs3Ro4cWeR9EhERaYKTT05i3tV5SFWmin396/THXOe50NfVl7Ay7aPV4Ss5ORnLli3D/v378fTpU1haWqJnz55YtGgRqlWrlq99vH37FidPnsTx48fh4+OD58+fw8DAAPXr18eQIUMwbtw4yOXyLNsoFApcvHgRx44dg6enJ548eQJBEGBjYwM3Nzd8//33sLKyKo6XrDEUCgX8/f0RFxcHc3NzODk5Zfs5SWn37t3Z+oYPH47g4GD069dPLR+2tdWtW7dw8OBBTJw4sdT/dwoA1apVQ2BgIIyNjYt9rH79+sHR0RHz5s3Dl19+CT09rX6LJSKiMk6pUmLD7Q3YeXen2Kcr08X0ltMxxHEIH5xcCFr7ySAlJQWdO3eGj48PqlSpgr59+yI0NBS7du3CiRMn4OPjA1vbvJ8tsHr1aixZsgQymQxNmjRB69atERkZCW9vb1y7dg0HDx7E6dOns3xw8/LyQo8ePQAANjY26NWrFxQKBf7991+sWbMGv//+Ozw9PeHg4FBsr18qkZGR2LxxA37ZuhmGMgUqmughKjEdKYIco8eMx/gJZeMDvTb7+eefAQBffvmlxJWUDLlcDkdHxxIZSyaTYejQofjhhx9w7NgxfPLJJyUyLhERkbrFpcXh+0vf48rzK2KfhYEFVrushnMVZwkr025ae8/X4sWL4ePjgzZt2uDhw4c4cOAAfH19sWbNGkRGRub7kh8TExPMmDEDoaGhuHXrFvbv34/z58/jzp07qFmzJq5cuYLFixdn2UZHRwefffYZfH19ERISgkOHDuHYsWN4/PgxevTogZcvX2LEiBHF8bIlFRgYiOZODfDw+Doc6ZuG4DEyXPtSheAxMhzuk4qHx9ehuVMDPHjwQOpSCyTzfqv58+fj4cOHGDRoECpXrgwdHR0cOXIEQMY9UvPnz0ebNm1gbW0NfX19VK9eHV9++SUePnyY677fvHmDOXPmoFGjRjAxMYG5uTkaNWqEGTNmICIiIs/aMu/B2r17N+7cuYM+ffqgfPnyMDExgYuLC65evVqg15qQkID9+/ejbt26aN68eYG2jY6Ohru7O6pWrQpDQ0PUr18f69evhyAIH9zO19cXAwcORJUqVcSf26hRo/D06dMiv15XV1fxd23BggVZ7uPLPPuZ0z1feW138OBByGQyDBkyJNfXNXr0aMhkMuzatStLf+Y2v/766wd/LkRERJoqJDYEQz2GZgledcrVwT63fQxeRaSVZ77S0tKwadMmAMDmzZthamoqLps6dSp+++03eHl54ebNm3l+wJw1a1aO/XXr1sXy5csxZMgQ7Nu3D0uXLhWXde7cGZ07d862jYWFBXbu3Ilq1arh33//RVhYGGrVqlWYl6hxoqKi0KOLC6Y5JWBS66yXF8pkMrSuroc/qgPrfRPQo4sLbvjd1bozYEFBQWjZsiUqVKiATp06ISYmRryUcvv27Vi5ciUaNmyIli1bwsDAAPfv38eePXtw9OhRXL58GY0bN86yv8DAQHTv3h3h4eGwtrYWz5Y+fPgQq1atQtu2bdGvX7981Xbjxg2MHz8ednZ26NGjBx48eIBLly6hS5cuuH79Oho2bJiv/Xh5eSEhIaHAk0/ExMSgffv2CAwMhLW1Nfr27Yvo6Gh89913ePz4ca7bbdmyBRMmTAAAtGzZEh06dEBQUBB27NiBY8eOwcvLC/Xq1Sv06+3ZsyfS09Ph7e2d7V6+OnXq5FpXXtu1bt0a1tbW+Pvvv/HmzRtUqFAhy/YJCQnYt28fzM3N8fnnn2dZZmtrixo1auDChQtITk6GkZFRrnUQERFpmsvhl/H9pe8Rr4gX+zrX6IylHZbCRG4iYWWlg1aGL29vb8TGxsLOzg5NmzbNtvzTTz9FQEAAjh8/XuBv99/n5OQEAHjx4kW+t6latSqsrKwQGRmJFy9elJrw9eu2rWhfOTlb8PqvSa3l8H2ZhC2bNuLHBQtLqDr12L9/P9zd3bFu3Tro6mZ9OGC/fv3w7bffonbt2ln6d+3ahZEjR2Ly5Mm4cOGC2J+eno7+/fsjPDwckydPxooVK6Cv/+6G1Hv37sHQ0DDftW3evBnr16/HxIkTxb4pU6Zg3bp1WLlyJf73v//laz+XL18GkBGECmL27NkIDAxEz549cejQIfEy3GvXrqFLly45buPj44OJEyeiSpUqOHr0aJbfxR07dmDUqFEYMWIEfHx8sm2b39c7c+ZMWFtbw9vbu0D38uVnu5EjR2Lp0qXYs2cPJk+enGXZ/v37kZCQgLFjx+Z4L1mrVq1w6NAh+Pj4oFOnTvmqiYiISEqCIGD3vd346eZPEPDuqpYxTmMw1mksdGRae8GcRtHKn6K/vz8AoFmzZjkuz+wPCAgo0jhPnjwBAFhbW+d7m7dv3yImJqbA2xUrQQCS3xb6jyI+Crt3/ILJLT58eVmmSc1V+OXnjVDERRZpXORxOZu6WVlZYcWKFdmCFwA4OztnC14AMGLECLRr1w6enp6IjY0V+//++28EBQWhQYMGWL16dZbgBQANGjSAnZ1dvmtr165dliACAHPnzgUAXLp0Kd/7yfydKMj9iImJifjtt9+go6ODTZs2ZQkbrVq1wvjx43Pcbvny5VAqldi6dWu2L0G+/vpr9OnTB76+vrh9+3a2bdX1eoti9OjR0NHRyfHywe3btwMAvvnmmxy3zbzHzM/Pr9jqIyIiUpeU9BTMujILa2+uFYOXkZ4R1rquxfgm4xm81Egrz3xl3itSvXr1HJdn9oeFhRVpnPXr1wMA+vbtm+9tNm/ejPT0dDRq1CjHD+uSSIkFVhT+DFzICyWMlEloWdU075UBtKqmCwNFLPy/q40WVbMHmXz7PgwwKlf47Quoa9euH5wRLyEhAcePH4efnx+io6OhUCgAABERERAEAcHBwWLwP3fuHABg1KhROYa5gurevXu2vgoVKsDS0jJf945lev36NQCgfPny+d7m5s2bSE5ORqtWrXIMjIMHD8aKFSuy9KlUKpw/fx7Gxsbi5Zb/1aFDBxw7dgzXrl3LdgZbXa+3KGrVqoWePXvi5MmTuHr1Ktq2bQsAuHPnDnx9fdGiRYscz7wDgKWlJYCMCWqIiIg02cvEl5h8cTLuvbkn9lUzrYb1ndbDwbL0TR4nNa0MXwkJCQCQ6wdlE5OM61Hj4+NzXJ4fW7duxblz51CuXDnMnDkzX9vcvn1bnJzjvx9GP6RBg5wfTBccHAxra+ssl7PlxsTEBCYmJoiLi4OOzn++nUiJh3m+q8kuLlVARWNZvqcTlclkqGAsQ1xq0c5cxcXHAwr1fdOSGZZSU1MRFxcn9icnJwMAKleunKX/fV5eXhg5ciSioqJy3f/Lly/F7UNCQgAAVapUyXWf78usQaFQZFk/NTXjeRoVKlTIcT+mpqaIjo7Oc4zMh/5mnpUFkOM2meO9/zPKvKeratWqOW6TGTQEQRCXR0ZGir+n/z3r91/h4eHidoV5vZk/u/8e10yZ7wNKpbJA2wEZM0KePHkSW7ZsEe8z27x5MwBg2LBhuW6Xea/g69evP3hsVCoVlEolEhMT4enpmet6QMYZSAD5ej8gzcPjp/14DLUfj2F2Iakh2BG1A/Gqd5+Z6xjUwQjzEXju9xzP8VzC6rLThGOYmJgoZo3C0MrwVdwuX76MSZMmQSaTYefOnahatWqe27x69QqffPIJUlJSMHnyZPTq1asEKi0Z5gYyRCUJEAQhXwFMEAS8SRJgbqBdz37I7R6shIQEDB8+HDExMfj+++8xYMAA1KhRA0ZGRpDJZPj6669x8ODBPGf9K4psgbqQzM0zYnhmMCoumWHP1NQUffr0+eC6OU24oa7XW1Tdu3dH9erVcfjwYSxfvhz6+vo4cOAATE1N8emnn+a6XWbgsrCwKKlSiYiICsQnwQd/xvwJJZRiXwfTDuhfrj90ZUW/aodyppXhK3N2w6SkpByXZ6ZiMzOzAu/77t276Nu3L9LS0rBhwwb0798/z23i4+Px0UcfITQ0FAMHDsSaNWsKNOa9e/dy7M88I5bTzIrvU6lUCAoKApDx4TrbB1czs4xL+AqpdnQ0kv9uhusv0tGqWt6/jNeeK5Eqt4DT6odAER68bG5oAajx4X2ZZyMMDAzEEAJAnI3uv/2ZvL29ER0djU8//RTLly/PtvzZs2cAIE4lD0C85DQiIiLHff5XZg1yuTzL+gYGBuLynPaTGYbzGiMzDFSpUgV+fn5IS0vLcZvM8d7/WWReavjixYsctwkNDRVryVxubGwMQ0ND6OjoYO/evfk+a1qY15vX8ct8H9DV1S3QdplGjx6NefPm4fjx4zA3N8fbt28xatSoDz7IPfOsWvXq1T+4b5VKJdbVsmXLD4bOzG/58no/IM3E46f9eAy1H49hBoVKgdXXV2Pfs31in56OHua0noNP7XP/YlETaMIxLMpZL0BLJ9yoWbMmgIzLlXKS2V/QmQZDQkLQvXt3xMTEYP78+eIU2R+SkpKCPn364NatW+jevTv27t2rMd/ai2SyjHunCvlHblYRw78ejXU38vcBev1NHYweOwFyc6sijavO4FUUmZfq5XSP4ePHj3Hr1q1s/V27dgWQMatf5lkgTZA5g2dmWM+P5s2bw8jICDdv3hQnoXnf/v37s/Xp6enB1dUVcXFxOH/+fOELzofMyxrT09OLZbtRo0ZBT08Pv/76qzj5Rm4TbWQKDAwEgCxT2BMREUntbcpbjD07Fn88+EPsszS0xI7uOzQ+eJUWGpYS8ifzA2ROH3rf7//vc5c+JCIiAt26dUNERAQmTZqEH3/8Mc9t0tPT8fnnn8PT0xNt27bF33//nef9Ldrqm2/HwPu1Mdb7Kj643npfBbxfG2Oce97BVVvY29sDyJjB8P0JFN6+fYuvv/5avJfsfZ988gns7e1x9+5dzJgxI9s69+7dyzHIFLcOHToAAK5fv57vbUxNTfHFF19AqVRiwoQJ4lkdION5XJnP3PuvOXPmQEdHByNGjMjxfqaEhATs3Lkzy/4KI/Oy4IIEyoJsV6VKFfTp0we3b9+Gl5cXGjdujFatWn1wm2vXrkFfXx/OznwQJRERaYaHMQ8xyGMQfF/6in31LOthv9t+NKuc8wzipH5aGb7atWsHCwsLBAcH5ziV88GDBwEAvXv3ztf+YmJi0KNHDwQHB2PEiBH46aef8txGEASMGDECx44dQ5MmTeDh4VHk05CarGLFijh93gtrA8ww5KgKvuHp4j1OgiDAJzwdQ46qsDbADKfPe2ndA5Y/pEWLFujWrRuePn0Ke3t79O/fH/3790ft2rXx4sWLHGfD1NPTw6FDh2BtbY01a9agVq1aGDBgAD755BM0atQIDRs2LPKjEAqjY8eOMDU1zXNyh/9atmwZHBwccPLkSdjZ2WHQoEHo0aMH2rRpgy+//DLHbdq3b4/NmzcjIiICnTp1QqNGjTBgwAAMGjQIzs7OqFixIr7++mtxko3CcnZ2RqVKlXDw4EG4urpi5MiRGDVqFK5evaq27caMGSP+ffTo0R/cb3BwMMLDw9G5c2c+YJmIiDTC+bDzGHZyGJ4nvJtAo5dNL/zW6zdUMa0iYWVlj1aGL319fbi7uwMAxo8fL97jBQBr165FQEAAXFxcsjxbaNOmTXB0dMSsWbOy7CspKQlubm64c+cOPvvsM/z666/5uj9l8uTJ2Lt3LxwdHXHmzBmUK1dOPS9Ogzk6OuKG31049JmC/scMYLdVQMv/6cBuq4ABxw3g0GcKbvjdFZ9xVJocPXoUc+bMgZWVFU6dOoWbN29i0KBB8PHxyfXYN2zYEP7+/vjuu+9gZmaGkydP4sKFC5DJZPj+++8lOStiamqKwYMH4/HjxwU6+2VpaQlvb2+MHTsWgiDgyJEjePr0KZYvX46NGzfmut2YMWNw48YNfPXVV4iPj8eJEydw+vRpJCQkYOjQoThx4kSRJ6UwNDSEh4cHunXrBj8/P+zevRs7duzAw4cP1bZdhw4dIJfLYWRkhKFDh35wv3/8kXEpR16XJhIRERU3laDCz/4/Y7LnZCSnZ1xpIoMMk5pNwoqOK2Ckxy8JS5pMKM4p2opRSkoKXF1d4evriypVqqBDhw4ICwuDr68vrKys4OPjA1tbW3H9+fPnY8GCBfjqq6+we/dusX/KlClYt24ddHV18fnnn4uTMvzX+9scPXoU/fr1AwB069Yt19kQZ86cWaQgkjnhRm4TcmR6f8INBwcHtd9zljlZw/sTBygUCvj7+yMuLg7m5uZwcnLK9WdH0nv/GPr5+aFp06Zwd3f/YHCid/bt24chQ4Zke//4L0EQUK9ePSQkJCA0NBR6eh+e06ggv7uacJMxFR6Pn/bjMdR+Ze0YJimSMNd7Ls6GnRX7TOWmWNFxBTpW7yhhZYWnCccwv5/Pc6OVsx0CGd9aX7x4EcuWLcMff/yBI0eOwNLSEsOHD8eiRYtyfQDzf2VOpqBUKsVvrHPy/geu95+VdPbs2RzWzjB8+PBSeRYIyJiVr0WLFlKXQYXQpEkTDBw4EDt37sQPP/yASpUqSV2SRlMoFOJz+8aPH//BdY8cOYKgoCDs2LEjz+BFRERUXMLjwzHp4iQ8jHl3NUct81rY0GkDbMvZfmBLKm5aedlhJiMjIyxcuBCPHz9GamoqIiIisGvXrhyD1/z58yEIQrZvrXfv3g1BEPL8877hw4fnaxtXV9difPVEhbds2TIoFAqsXLlS6lI01rFjxzBy5Eg0bdoU/v7+6NevH1q2bJnr+oIgYOHChWjYsCGGDx9ecoUSERG951rENQz2GJwleLWr2g6/f/Q7g5cG4FezRGWQnZ0d0tLSpC5Do926dQu7du1C+fLlMWTIkDwv0ZTJZLh9+3YJVUdERJSVIAjY92AfVl5fCaXw7sHJIxqMwKRmk6CrwwcnawKGLyKiHMyfPx/z58+XugwiIqI8KZQKLPFdgkOPDol9+jr6mN92Pnrb5W/2byoZDF9ERERERFoqKjkKUz2n4vbrd1dfVDKqhPWd16NhxYYSVkY5YfgiIiIiItJC997cw6QLk/Aq6ZXY19iqMda5roOVcel55mppotUTbhARERERlUUnn5zEV6e+yhK8+tXph109djF4aTCe+SIiIiIi0hJKlRIbb2/Ejrs7xD5dmS6+a/EdhtYbCplMJmF1lBeGLyIiIiIiLRCfFo+Zl2fiUvglsc9c3xyrXVajTdU2ElZG+cXwRURERESk4UJjQzHhwgSExoWKfXXK1cGGThtQw7yGdIVRgTB8ERERERFpsCvPr2CG1wzEK+LFvk41OmFZh2UwkZtIWBkVFMMXEREREZEGEgQBu+/txrpb66ASVGL/GKcxGOs0Fjoyzp2nbRi+iIiIiIg0TEp6Chb8uwAnnpwQ+4z0jLC43WJ0t+kuYWVUFAxfREREREQa5FXiK0y6OAn33twT+6qaVMWGzhvgYOkgYWVUVDxXSaWaTCbLccrVzP7c/ri6uhZonPT0dDg4OKBVq1ZZ+j09PfMcSyaTYffu3eI2oaGhkMlksLGxyXGshIQEuLi4QCaToXXr1nj79i0AYP78+bnu38LCAhYWFvD09My2v3HjxkEmk0FHRwdhYWH5er0+Pj4YNWoU7O3tYWZmBkNDQ9jY2OCzzz7D4cOHoVK9uzQiIiICRkZGGDduXL72TUREVJb5vfbDII9BWYJXi8otsO/jfQxepQDPfFGhKBQK+Pv7Iy4uDubm5nBycoJcLpe6rAL76quvcux3dHQs0H62bduGhw8fwsPDI8fllStXRs+ePXPdvk6dOvkaJyEhAb169cKVK1fg7OyM06dPw9zcPMs6Tk5OaNKkSZY+hUIBALC2ts7Sn5aWhgMHDgDIuK78999/x+zZs3MdX6FQYOzYsdixI+PZIg4ODujatSv09fUREhKCQ4cO4a+//kLnzp1x/vx5AECVKlUwevRobNmyBZMnT4a9vX2+XitRaaFQKBAUFISkpCStfr8kouJ3+NFhLPJZBIVKIfZ97vA5vm/1PeQ6fN8oDRi+qEAiIyOxYeMGbNm2Bem66dA300dafBr0lHoY9+04TJwwEVZW2vNU9ffPOBVWamoqFi5ciMaNG+Ojjz7KcR1HR8cijxUfH49evXrB29sb7dq1w6lTp2BmZpZtvX79+mH+/PlZ+uLi4gAgW1A7efIkoqOjUaVKFURERGDPnj0fDF9ffvkl9u/fD3t7e+zatQtt27bNsvzFixdYuHAhzpw5k6V/xowZ2LRpE3744Qcx7BGVdu+/X6YhDXIzORauWai175dEVHzSVelYc2MN9gbuFfv0ZHqY7TwbA+0HSlgZqRvDF+VbYGAgXLu6QrARUG5MORjVNoJMJoMgCEh+koxt57fhlx2/wOu8V4HPHGmzgwcP4vXr15gxY0axjREXF4eePXvi33//RceOHeHh4QFTU9Mi73fPnj0AgAULFmDFihV48OABbty4gRYtWmRb96+//sL+/ftRuXJlXLp0CZUrV862TtWqVbF161Z4e3tn6a9WrRo6deqEw4cP49WrVzluS1Sa8P2SiPLrbcpbfHfpO/hG+Ip9loaWWOu6Fs0rN5ewMioOvOeL8iUqKgquXV2h10kPlUdXhrGtsXgvlUwmg7GdMSqPrgy9Tnpw7eqKyMhIiSsuOdu3b4dMJsOgQYOKZf9xcXHo0aMH/v33X7i6uuLkyZNqCV5v376Fh4cHDAwM8Nlnn2Ho0KEA3gWy/1q9ejWAjHvL8gpP7dq1y9Y3ZMgQKBQKtZxtJNJkkZGRfL8konx5FPMIgz0GZwle9SzrYb/bfgavUorhi/Jl6y9bIdgIsOxm+cH1LLtZQlVLhY2bNpZQZdKKi4vD5cuXUadOHVSrVk3t+4+NjUW3bt3g4+ODLl26wMPDAyYm6nmY4p9//onU1FR8/PHHsLCwwLBhwwAA+/fvR3p6epZ1o6KicO3atSKFzMxJTHK7L46otNiwcQPfL4koT+efnsewk8MQnhAu9vW06Ynfev2GKqZVJKyMihPDVxkgCALi0uIK/Sc6MRq/7vwVZl2z31+UE7MuZti8dTPeJL4p0riCIBTzT6borl69CqVSiZYtW6p938nJyejWrRuuXbuG7t274/jx4zA2Nlbb/jPPcGWGrrp166JVq1Z4/fp1tnu2/Pz8AAC2trYoV65cocaztbVFxYoVce3aNaSkpBS6biJNplAosGXblgK9X27ZtkWcFIeISj+VoMLP/j9j8sXJSEpPAgDIIMOkZpOwsuNKGOkZSVwhFSfe81UGxCvi0W5f9svA8is5JBmJSES12vk7s2Nka4RnwjO0XtUaRrUL/wbiPdgb5vrmea9YBDlNQw8AISEhuU71/r6AgAAAGbP+fYiXl1euYwFATExMtlDz+vVrvH79GsbGxti7dy+MjPL3s1ywYAEWLFiQrX/IkCH4/fffAWS8Pm9vb1haWmaZJGTYsGG4du0a9uzZk6X/zZs3AFDkyQEcHBzg7e2NwMBANG3atEj7ItJE/v7+SNdNz/d7n5GtEaJ0ouDv75/jvZZEVLokKZIw13suzoadFftM5CZY0WEFXGq4SFgZlRSGL8qTMlkJPVO9D4aH98lkMuiZ6kGZrCzmyoout6nm83tP1evXrwEA5cuX/+B6eU01r6+vn62vYsWKqFChAoKCgjBw4ECcOnUqXwEst6nmnZ2dxfbvv/8OQRDw2WefZRl70KBBmDp1Ko4ePYr4+PgcZ1MsCkvLjMuweI8LlVZxcXHQN9Mv0PtlumE69vntQ1XHqqhqWrWYKyQiqTxPeI6JFybiYcxDsa+mWU1s7LwRtuVsJayMShLDF+VJ10gX6QnpEAQhXx8oBEFAekI6dI10S6C6oinq5A+xsbEAkGdIKcxU8yYmJjh37hzat28PLy8vDBgwAEePHs3z+UAfmmo+038vOcxkZWWFHj16wMPDA4cOHcLw4cMBABUqVABQ9NCUOdV95oOhiUobc3NzpMWnFej9UpGgwNHwozhz6AyaVWoGN1s39LDpAQsDixKomIhKwvWX1zHVcyrepr4V+9pWbYuVHVfyd72MYfgqA8zkZvAe7J33irmIfhONpluaIjkkGca2ed9zlPwkGWYyM/hO9y3Sg0TN5Oo961IcLCwy3jDj4+OLZf/Vq1fH+fPn0aFDB5w6dQpDhw7Fvn37oKtb+GB77do1PHyY8a3b999/n215eHjGjb979uwRw5eTkxMA4MmTJ3j79m2h7/vKDKuF3Z5I0zk5OUFPqVeg90tBIcCwhiEA4NbrW7j1+haWXVuG9tXaw83WDa7VXWGoZ1jcpRNRMRAEAQeCDmDFtRVIF95NZvVV/a8wuflk6Onwo3hZwyNeBshksqLdO2UCfDPyG+w+txvGo/P+MBF/Ph7jx4xHBZMKhR9TS1SqVAkAEB0dXWxj2NnZ4dy5c3BxccFff/0FU1NT7NixI9+XNf3X+1PJ//d5XO/z9PREeHg4qlevDisrK7Rq1QrXrl3D/v37MWbMmEKNHRMTA6Do944RaSq5XI6BXw7E/jP7YTwmf++X3T7vBh1LHYTFhYn96ap0eD7zhOczT5jITdClZhe42bqhtXVr6Opo/lUFRAQolAosvbYUBx8eFPv0dfTxY9sf0ceuj4SVkZQ42yHly5jRY6ATpoPosx8OGdFno6ETpoMJ7hNKqDJpZZ4RCgoKKtZx6tevj9OnT8PCwgK7du3CpEmTCrWf9PR0HDhwAABw9+5dCIKQ45/hw4dDpVKJE3QAwHfffQcg4zlfmfe65ebq1as59j948AAGBgaoV69eoeon0nTP4p7B394fSY+SEHUm6oPrZr5f/rboNxzvdxz73PZhWL1hqGCY9YurREUijgUfw7dnv0XXg12x4toK3Iu6pxUzwhKVVVHJUfj6zNdZgpeVkRV299zN4FXGMXxRvlSsWBGe5zyh9FTi1S+vkBScJP7DLwgCkh4n4dUvr6D0VMLznGeZObPRtm1b6Orq4vr168U+VrNmzeDh4QFjY2Ns3LgRc+bMKfA+/vnnH0RGRqJRo0Zo0KBBrusNHjwYALB3716xb+DAgRg0aBBevXqFjh074t9//8223cuXL+Hu7p7tXjIACA4Oxps3b9CqVSsYGvISKip9opKjMPrsaCToJ8Bmmg3enH6DiG0R+Xq/lMlkaFixIb5v9T3ODTyHbV23oY9dHxjrGWcbY2/gXgzyGIQ+R/rgZ/+f8SzumRQvl4hycf/NfQz2GIzbr2+LfY0rNsb+j/ejkVUjCSsjTcDLDinfHB0dcef2HWzctBGbt25GlG4U9M30kRafBrlKjnHfjsME9wllJngBGRNtdOjQIcslejl58OCBeP9UTrp3744hQ4bkOV67du1w9OhRfPzxx1i6dCnMzc1zvG8rN5mXHGaGq9x06dIFlSpVwt27d+Hn5yfOnvi///0PxsbG2LlzJ9q2bQtHR0fUr18fcrkcoaGhuHHjBpRKJbp165Ztn56engAANze3fNdLpC0SFYkYd26c+LBUg6oG2HFyBx55PMLmrZvxGq+hZ6oHVbIqz/dLPR09tK3WFm2rtcVc57nweuYFjyceuPL8SpZ7RkLjQrHFbwu2+G1B44qN8ZHtR+hp0xMVjEr/Jd9EmupUyCnM856HFOW751n2teuLH9r8AANdAwkrI03B8EUFYmVlhYULFuKHuT/A398fcXFxMDc3h5OTU5Em19Bm33zzDTw9PbFv3z5Mnz49x3VevXqF3377Ldd9lCtXLl/hCwC6du2KP//8EwMGDMDMmTNhZmaGcePG5bldXFwcjh07BiBjSvkP0dXVxcCBA7F582bs2bNHDF9yuRw7duzAqFGjsH37dly6dAn//PMPlEolrK2tMWDAAAwdOhS9e/fOts8//vgDcrn8gyGUSBsplApMvjgZgdGBYt+U5lPwRcMvgFbAD3N/wPbt25GUlAQXF5cCvV8a6RmhZ+2e6Fm7J96mvMWZsDPweOKBW69vZVkvICoAAVEBWHV9FZyrOsOtthu61OwCY7n6HsxORLlTqpTYeHsjdtzdIfbpyHTwXYvvMKzesELfp02lj0zgReMaK/OysHv37n1wPZVKJd5z5ODgAB0d9V5NmjlNeeY04ZRVamoqatWqhUqVKokPXdY0Uh/D8PBw1KpVC59++ql4zxkV7Hf3woULAIDOnTuXSG2UPypBhZmXZ+JUyCmxb1i9YZjRckaWD1vqPn7PE57jVMgpeDzxwOO3j3Ncx0jPCK41XPGx7cdoU7UN5Dpl8wsydeHvoPYrrmMYnxaPmZdn4lL4JbHPXN8cq11Wo03VNmodq6zThN/D/H4+zw3v+SIqIgMDA8ybNw937tzBiRMnpC5HI61atQo6OjpYuHCh1KUQqY0gCFh1fVWW4NWrdi9Mbzm92L/lrmZaDaMajcLfff7Gwd4HMaLBCFQyrpRlneT0ZJwKOYXx58ejy59dsNhnMfxe+3GiDiI1CosLw9CTQ7MELzsLO+xz28fgRTli+CJSg9GjR8Pe3h4LFiyQuhSNExERgV9++QXffPMNHBwcpC6HSG123duFvYHvJqVxruKMJe2WQEdWcv+0ymQyOFg6YGqLqTj76Vns7LETA+oOyPacxJjUGBwIOoAvTn2Bj/7+CBtvb8ST2CclVidRaeT93BuDPQYjJDZE7HOt4Yq9H+1FTfOaElZGmoz3fBGpgZ6eXrFPN6+tqlSpguTkZKnLIFKrY8HH8NPNn8R2Pct6WNdpHeS60l3apyPTQUvrlmhp3RKzW8/G5fDL8AjxgNczL6Sp0sT1whPC8UvAL/gl4BfUs6wHN1s39KrdK9uZMyLKmSAI+N/9/2HtzbVQCSqxf3Tj0RjfZHyJfgFD2ofhi4iIqAAuh1/GPO95YruGWQ1s6boFJnITCavKSl9XH11qdUGXWl0QlxaH82HnceLJCVx/eR0C3l12GBgdiMDoQKy5sQatqrSCW203dK3VFWb6Zh/YO1HZlZKeggX/LsCJJ+9uMzDSM8KidovQw6aHhJWRtmD4IiIiyqeAyABM85oGpaAEAFgaWmJb122oaFRR4spyZ65vjv51+6N/3f54lfgqY6KOEA88iH4griNAgG+EL3wjfLHYZzFcarjAzdYNHap1gL6uvoTVE2mOV4mvMPniZNx9c1fsq2pSFes7r4ejpaOElZE2YfgiIiLKh5DYEIw/Px7J6RmX0RrrGePnrj+jhnkNiSvLv8omlTG84XAMbzgcwW+D4fHEAydDTuJ5wnNxnTRVGs6GncXZsLMw0zdD91rd4WbrhuaVm/NyKiqz/CP9MfniZEQlR4l9LSq3wBrXNbA0tJSwMtI2DF+lwPuzanEWKyLt8f7vK58Bo9leJ73GmLNj8Db1LYCMByGv67QO9SvUl7awIrArZ4eJzSZiQtMJ8Iv0g8cTD5wOPS2+RiBjCu1Djw7h0KNDsDaxRq/aveBW2w0Olpw8h8qOI4+PYOG/C6FQKcS+zx0+x/etvucjHKjAGL5KAZlMBl1dXSiVSqSmpsLYmA/VJNIGqampADIeas3wpbni0+Ix9txYvEh8IfYtbb+01EwjLZPJ0LRSUzSt1BTft/weV19chccTD1x8dhEpyhRxvZeJL7Hr7i7sursLdcrVgZutG9xqu6GKaRUJqycqPumqdKy5sSbLrKZ6Mj3Maj0Lnzl8JmFlpM0YvkoJY2NjxMfHIz4+nuGLSEvEx8cDAExMNGeiBsoqVZmKiRcm4mHMQ7FvRssZ6FW7l4RVFR+5rhwuNVzgUsMFiYpEXHh6AR5PPPBvxL9ZZnV7/PYx1t9aj/W31qNZpWZws3VDD5sesDCwkLB6IvWJTY3Fd17fwSfCR+yzNLTEWte1aF65uYSVkbZj+ColzM3NER8fj+joaOjp6cHc3By6urpq2bdKpcry/6R9eAw1i1KpRFxcHKKjowEAZmacWU4TKVVKzLo8Czde3RD7RjQcgS/qfyFhVSXHRG6C3na90duuN6KSo3A69DQ8nnjgTtSdLOvden0Lt17fwrJry9C+Wnu42brBtborDPUMJaqcqGgexzzGxIsT8Sz+mdjnaOmI9Z3Wo6ppVQkro9KA4auUMDMzg4WFBWJjY/H69Wu8fv1abftWKjNm9VJXmKOSx2OoucqVK8fwpYEEQcCya8twNuys2NfHrg+mNJsiYVXSqWhUEUPrDcXQekMRFheGk09OwiPEA2FxYeI66ap0eD7zhOczT5jITdClZhe42bqhtXVr6OrwvYe0w4WnFzDr8iwkpSeJfT1semBh24UwlvPKIio6hq9SQiaTwdraGkZGRoiJiRHvJVGHxMREABln10g78RhqHgMDA5QvXx4WFha830sD/RLwCw4EHRDb7au1x/y283msANQyr4WxTcZijNMY3HtzDx5PPHAq5BTepLwR10lUJOJY8DEcCz4GKyMr9KzdE262bqhvWZ8/Q9JIgiBgW8A2bPbbLPbJIMOEphMwqtEo/ndLasPwVYro6OigfPnyKF++PARBUNvMh56engCAli1bqmV/VPJ4DDWLTCbjP+Qa7NDDQ9jkt0lsN6rYCGtc1nBWs/+QyWRoWLEhGlZsiGktpuFaxDV4hHjgXNi5LGcNIpMjsef+Huy5vwc25jbiRB3aNEU/lW5JiiTM9Z6b5Uy3idwEyzssh2sNV+kKo1KJ4auUKo4Pdzo6fL6LtuMxJPqwi08vYqHPQrFtY26DzV0283KjPOjp6KFttbZoW60t5jrPhdczL3g88cCV51eQLqSL64XGhWKz32Zs9tuMxlaN4VbbDT1r9+RzkqjYKRQKBAUFISkpCebm5nBycoJcLsfzhOfZJtWpaVYTGzpvgF05OwkrptKK4YuIiAjA7de3Mf3SdHFWPysjK2ztthXlDctLXJl2MdIzQs/aPdGzdk/EpMTgTOgZeIR44Pbr21nWC4gMQEBkAFZeX4k2VdvAzdYNnWt0ZtAltYqMjMSGjRuwZdsWpCENcjM5Fq5ZCD2lHvoN64d7DveQaJAort+2alus7LiSM3dSsWH4IiKiMu9xzGOMPz8eqcqM+2VN5ab4uevPqGZaTeLKtFt5w/L43PFzfO74OZ4nPM+YqOOJB4Jjg8V1lIISV55fwZXnV2CkZ4RONTrBzdYNbaq24aWeVCSBgYFw7eoKwUZAuTHlYFTbCDKZDIIgIPlJMg6dPYSk3Umw+c4GBlUN8GX9LzGl+RTo6fDjMRUf/tdFRERl2svElxhzbgzi0zKeuybXkWND5w1wsHSQuLLSpZppNXzT+BuMajQKQTFB8HjigZMhJ/E66d3svMnpyTgZchInQ07C0tAS3Wt1h5utG5ysnHifJBVIZGQkXLu6Qq+THiy7Zb2sVSaTwdjOGMZ2xog6E4XQNaHYcXIHvmhZNh4jQdJi+CIiojIrNjUW3579Fq+SXgHImN1seYflaGnNyWmKi0wmg6OlIxwtHTG52WTcfHUTHiEeOBt6FvGKeHG96JRo7A/aj/1B+1HdtDo+sv0IbrZusLWwlbB60hYbNm6AYCNkC17/VbF7RaSHpuORxyOgVQkVR2Ua774nIqIyKTk9Ge7n3fEk9onYN7v1bHS36S5hVWWLro4uWlVphQVtF+Di5xfxk+tP6Fqza7bLDcMTwvFLwC/oe6QvPjv+Gf5373+ITIqUqGrSVOmqdLxOeg2/l37Y8PMGmHXN3zMUzbuYY8u2LVAoFMVcIRHPfBERURmUrkrHDK8Z8Iv0E/u+bfwtBjkOkq6oMs5A1wBda3VF11pdEZcWh3Nh5+DxxAPXX16HgHePTgmMDkRgdCDW3FyDVtat4Gbrhq41u8JU31TC6qk4pSpTEZUchcikyIz/T47M8vfMZdEp0RAgIDkkGcmyZBjVNsrX/o1sjRClEwV/f3+0aNGimF8NlXUMX0REVKYIgoBFPovgGe4p9g2oOwDjm4yXrijKwlzfHJ/U/QSf1P0ELxNf4p+Qf+AR4oEH0Q/EdVSCCj4RPvCJ8MFin8Vwqe4CN1s3dKjWAXJdTtShDRIViYhMiswSoMRwlRyJqKSMv8elxRVov8pkJfRM9fJ9n6BMJoO+mT7i4go2DlFhMHwREVGZsslvE/5+9LfYdq3hirnOczmhg4ayNrHG8IbDMbzhcAS/DRYn6nie8FxcJ1WZijNhZ3Am7AzM9c3R3aY73Gq7oVnlZtCR8Q6LkiQIAmJTY98FqA+csUpOTy6WGkzMTKBKUEEQhHz9XguCgLT4NJibmxdLPUTv0+rwlZycjGXLlmH//v14+vQpLC0t0bNnTyxatAjVquVveuC3b9/i5MmTOH78OHx8fPD8+XMYGBigfv36GDJkCMaNGwe5POdv0JRKJTZs2ICdO3fi8ePHMDU1RadOnbBgwQLUq1dPnS+ViIjUYP+D/fgl4Bex3bRSU6zquIpTS2sJu3J2mNhsIiY0nQC/SD94PPHA6dDTeJv6VlwnLi0OBx8exMGHB1HFpAp61e4FN1s32Je3l67wUkCpUiI6JTpLoMopXEUlR0GhKp57p8oZlENFo4qwMrKClbGV+PeKxv/fZ5TRJ4cc1hutkRySDGPbvJ8bl/wkGXKVHE5OTsVSN9H7tPZfm5SUFHTu3Bk+Pj6oUqUK+vbti9DQUOzatQsnTpyAj48PbG3znhFp9erVWLJkCWQyGZo0aYLWrVsjMjIS3t7euHbtGg4ePIjTp0/D2DjrL69KpcLAgQNx+PBhlCtXDm5uboiKisLBgwfh4eGBixcvolUrTptDRKQpzoSewVLfpWLbzsIOGztvhKGeoYRVUWHIZDI0rdQUTSs1xfctv8fVF1fh8cQDF59dRIoyRVwvIjECO+/uxM67O1G3fF241XbDR7U/QhXTKhJWr1nSlGnvglNSVK5nrKJTosUHkKuTjkwHFQwrZAQp43cBKsdQVYDLScd9Ow7bzm2D8ei8w1f8+XiM+zb3L9uJ1Elrw9fixYvh4+ODNm3a4MyZMzA1zbjRdu3atZg2bRpGjhwJT0/PPPdjYmKCGTNmYPz48ahZs6bY/+jRI3Tt2hVXrlzB4sWLsXTp0izb7dy5E4cPH0bdunVx+fJlVK5cGQBw6NAhfPrppxg6dCgCAwOhp6e1P2IiolLj+svrmHl5pjhxQ2XjytjabSssDCwkroyKSq4rh0sNF7jUcEGiIhHnn56HxxMP+ET4ZAkLj2IeYV3MOqy7tQ7NKzeHm60butfqnq//BhQKBYKCgpCUlARzc3M4OTlp/Af1JEVStsv83g9YmX2xqbHFMr5cR54lQOV0xsrK2ArlDcpDV0dX7eNPnDARv+78FdFnoz843Xz02WjohOlggvsEtddAlBOZIAhC3qtplrS0NFSqVAmxsbG4desWmjZtmmW5k5MTAgICcOPGDTRv3rzQ4+zbtw9DhgyBjY0NQkJCsiyrX78+AgMDcfjwYfTr1y/Lsr59++LYsWM4ePAgBgwYUOjxGzRoAAC4d+9eofehDhcuXAAAdO7cWdI6qPB4DLUfj2HhBUUHYfg/w5GgSACQMZnD/3r9D3bl7EqsBh6/kheVHJUxUccTD9x9czfHdfR09NChWge42brBpbpLtrOgkZGR2LBxA7Zs24I0pEFuJocySQk9pR7GfTsOEydMhJWVVUm8HAAZ9ybFpcVlveQvl5n/ktKTiqUGYz3jrJf85XDGysrYCub65pLfR/ngwQO4dnWFqpYKZl3MYGRrBJlMBkEQkBycjPgL8dAJ04HnOU84OjpKWivljya8lxb187lWnpbx9vZGbGws7OzssgUvAPj0008REBCA48ePFyl8ZV77++LFiyz9ISEhCAwMhJGREdzc3HIc/9ixYzh+/HiRwhcRERXN84TnGHNujBi8DHQNsLnL5hINXiSNikYVMaz+MAyrPwxhcWHweOIBjyceeBr/VFwnXZWOi88u4uKzizCVm6Jrra5ws3VDy8ot8TDoIVy7ukKwEVBuTDkY1X7vg/uTZGw7vw2/7PgFXue9ivzBXalSIiY1Jtt9VO8HrKikKEQlRyFNlVbUH02Ocr2fyqhiloBlLM/7Mj5N4ejoiDu372Djpo3YvHUzXuM19Ez1oEpWQa6SY9y34zDBfUKJBmgirQxf/v7+AIBmzZrluDyzPyAgoEjjPHmS8eBNa2vrHMdv2LBhjpcdqGt8IiIqvOiUaIw5OwZRyVEAMu4tWdVxFZpUaiJtYVTiapnXwrgm4zDWaSzuRt2FR4gHToWcQnRKtLhOgiIBRx4fwZHHR2ChsID/HH+YdzPPdsmaTCaDsZ0xjO2MEX02Gq5dXXHn9p0cP8ArlIoPXvKXecYqOiUaSkGp9ted3/upKhhVgL6uvtrH1wRWVlZYuGAhfpj7A7Zv346kpCS4uLhoxaWjVDppZfh6+jTjW6vq1avnuDyzPywsrEjjrF+/HkDGZYRSjE9ERIWTpEiC+3l3hMaFin3znOehU81O0hVFkpPJZGhk1QiNrBrhuxbfwTfCFx5PPHD+6fksl+k9PPEQura6H7xXCAAsu1ki4kkERs8fjY4jO2Y7Y/X+LIzqJPX9VNpILpfDwcEBAPggZZKUVoavhISMy0f+OwNhJhMTEwBAfHx8ocfYunUrzp07h3LlymHmzJnFOn7mtaP/FRwcDGtra/H6VqkkJiYCgOR1UOHxGGo/HsP8UwpK/Br1KwJTAsW+j8w/Qvnw8rgQLs3Pj8dPc3VFV3Ss3BF3U+7iRuIN3E+4jxivGNScWDPvjQFYdLXAiY0n8KjZI8j0inaPk4HMAOa65hl/dDL+30LX4l3f//cb6xi/u59KAJD0/3/eAOlIR8T//4+y4u+h9tOEY5iYmCh+1i8MrQxfxe3y5cuYNGkSZDIZdu7ciapVq0pdEhER5YMgCNgXvS9L8Opg2gHdzbtLWBVpOn0dfTQzboZmxs1wO+o2fpD/AKPaRvna1sjWCDK5DCnPUnLdxljH+INhKrPfQMdAnS+LiDSQVoavzGnlk5JynsknMxWbmZkVeN93795F3759kZaWhg0bNqB///7FPn5us6VknhGTenYsTZhZhoqGx1D78Rjmz9qba3E9/LrY7larG1Z1XCX5pVc8ftrFpLxJvmfqk8lkMDAzQNsKbdGycct3z6V675LA0no/lbbh76H204RjWJSzXoCWhq/M53GFh4fnuDyzv1atWgXab0hICLp3746YmBjMnz8fEybk/MyH4hqfiIgKb8/9Pdh1d5fYblG5BZZ1WCZ58CLtYm5ujrT4NAiCkK8AJggCdJJ1MLHNRLRoynuJiOjDdKQuoDAyp4C/detWjssz+xs3bpzvfUZERKBbt26IiIjApEmT8OOPP+Y5/t27d6FQKNQyPhERFd7JJyex8vpKsW1f3h4bOm+AgS4v46KCcXJygp5SD8khyflaP/lJMuQqufjZgIjoQ7QyfLVr1w4WFhYIDg6Gn59ftuUHDx4EAPTu3Ttf+4uJiUGPHj0QHByMESNG4Keffvrg+rVr10a9evWQnJwMDw+PIo9PRESFd/XFVczxniO2q5pUxc9df4aZfsEvPSeSyzOe/xR/Ln+TZsWfj8e4b8dx2nIiyhetDF/6+vpwd3cHAIwfP168xwoA1q5di4CAALi4uGR5wPKmTZvg6OiIWbNmZdlXUlIS3NzccOfOHXz22Wf49ddf83WZwdSpUwEAM2bMwOvXr8X+v//+G8eOHUOdOnWyTVFPRETqde/NPUy5OAXpqnQAGQ+K3dptKyoZV5K4MtJmEydMhE6YDqLPRn9wveiz0dAJ08EE95xvUyAi+i+tvOcLAObOnYtz587h6tWrqFu3Ljp06ICwsDD4+vrCysoKO3fuzLJ+VFQUgoKCEBGRderVOXPm4N9//4Wuri709PTw9ddf5zje7t27s7RHjhyJkydP4vDhw3B0dESXLl0QFRUFLy8vGBkZYe/evdDT09ofLxGRxnsW9wzjzo0Tn89kpGeEzV02o7ZFbYkrI21nZWUFz3OecO3qilchr2DWxSxjVkOZDIIgIDk4GfEX4qETpgPPc545PmCZiCgnWpsODA0NcfHiRSxbtgx//PEHjhw5AktLSwwfPhyLFi3K9QHI/xUTEwMAUCqV+OOPP3Jd77/hS0dHB3/99RfWr1+PnTt34sSJEzAxMcGAAQOwYMEC1K9fv9CvjYiIPiwqOQqjz45GdErGmQk9mR7WuKxBYyvea0vq4ejoiDu372Djpo3YvHUzXuM19Ez1oEpWQa7KuDRxgvsEBi8iKhCZIAiC1EVQzjKnms9tKvqSognTelLR8BhqPx7DdxIViRjxzwgERr97lteS9kvQx66PhFV9GI+fdlMoFNi+fTuSkpLg4uICJycn3uOlhfh7qP004RgW9fO51p75IiKiskehVGDyxclZgteU5lM0OniR9pPL5XBwcAAAtGjB6eSJqPC0csINIiIqe1SCCnO858AnwkfsG1ZvGEY0GCFhVURERPnH8EVERBpPEASsur4Kp0JOiX29avfC9JbT8zVDLRERkSZg+CIiIo23694u7A3cK7adqzhjSbsl0JHxnzEiItIe/FeLiIg02rHgY/jp5k9iu55lPazrtA5yXU54QERE2oXhi4iINNbl8MuY5z1PbNcwq4EtXbfARG4iYVVERESFw/BFREQaKSAyANO8pkEpKAEAloaW2NZ1GyoaVZS4MiIiosJh+CIiIo0TEhuC8efHIzk9GQBgrGeMn7v+jBrmNSSujIiIqPAYvoiISKO8TnqNMWfH4G3qWwCAno4e1nVah/oV6ktbGBERURExfBERkcaIT4vH2HNj8SLxhdi3tP1StKnaRsKqiIiI1IPhi4iINEKqMhUTL0zEw5iHYt+MljPQq3YvCasiIiJSH4YvIiKSnFKlxKzLs3Dj1Q2xb0TDEfii/hcSVkVERKReDF9ERCQpQRCw7NoynA07K/b1seuDKc2mSFgVERGR+jF8ERGRpH4J+AUHgg6I7fbV2mN+2/mQyWQSVkVERKR+DF9ERCSZQw8PYZPfJrHdqGIjrHFZA7mOXMKqiIiIigfDFxERSeLi04tY6LNQbNuY22Bzl80wlhtLWBUREVHxYfgiIqISd/v1bUy/NB0qQQUAsDKywtZuW1HesLzElRERERUfhi8iIipRj2Mew/28O1KVqQAAU7kpfu76M6qZVpO4MiIiouLF8EVERCXmZeJLjDk3BnFpcQAAuY4cGzpvgIOlg8SVERERFT+GLyIiKhGxqbEYc3YMXiW9AgDIIMPyDsvR0rqlxJURERGVDIYvIiIqdinpKZhwYQKCY4PFvtmtZ6O7TXcJqyIiIipZDF9ERFSs0lXpmH5pOm6/vi32jW48GoMcB0lYFRERUclj+CIiomIjCAIW+yyG5zNPsW9A3QFwb+IuWU1ERERSYfgiIqJis9lvMw49OiS2XWu4Yq7zXMhkMgmrIiIikgbDFxERFYsDDw5gW8A2sd20UlOs7LgSejp6ElZFREQkHYYvIiJSu7NhZ7HEd4nYtrOww8bOG2GkZyRhVURERNJi+CIiIrW6/vI6vr/0PQQIAIDKxpWxtdtWWBhYSFwZERGRtBi+iIhIbYKigzDxwkQoVAoAgLm+ObZ12wZrE2uJKyMiIpIewxcREanF84TnGHtuLBIUCQAAA10DbO6yGXbl7CSujIiISDMwfBERUZHFpMRgzNkxiEyOBADoyHSwquMqNKnURNrCiIiINAjDFxERFUmSIgnjz49HaFyo2DfPeR461ewkXVFEREQaiOGLiIgKTaFSYJrXNNyJuiP2uTdxxwD7ARJWRUREpJkYvoiIqFAEQcD8q/Nx5fkVsW+QwyCMbjxawqqIiIg0F8MXEREVyrpb63As+JjY7larG2a2mgmZTCZhVURERJqL4YuIiAps7/292Hl3p9huUbkFlnVYBl0dXQmrIiIi0mwMX0REVCCnQk5hxfUVYtu+vD02dN4AA10DCasiIiLSfAxfRESUb/+++Bezr8wW21VNquLnrj/DTN9MwqqIiIi0A8MXERHly/039zH54mSkq9IBAOUMymFrt62oZFxJ4sqIiIi0A8MXERHl6VncM4w9NxZJ6UkAACM9I2zushm1LWpLXBkREZH2YPgiIqIPikqOwuizoxGdEg0A0JPpYY3LGjS2aixxZURERNqF4YuIiHKVqEjEuHPjEJ4QLvYtaLcAHap3kLAqIiIi7cTwRUREOVIoFZh8cTICowPFvinNp6CPXR8JqyIiItJeDF9ERJSNSlBhjvcc+ET4iH3D6g3DiAYjJKyKiIhIuzF8ERFRFoIgYNX1VTgVckrs61W7F6a3nA6ZTCZhZURERNqN4YuIiLLYdW8X9gbuFdvOVZyxpN0S6Mj4TwYREVFR8F9SIiISHQs+hp9u/iS261nWw7pO6yDXlUtYFRERUenA8EVERACAy+GXMc97ntiuYVYDW7pugYncRMKqiIiISg+GLyIiQkBkAKZ5TYNSUAIALA0tsa3rNlQ0qihxZURERKUHwxcRURkXEhuC8efHIzk9GQBgrGeMn7v+jBrmNSSujIiIqHRh+CIiKsNeJ73GmLNj8Db1LQBAT0cP6zqtQ/0K9aUtjIiIqBRi+CIiKqPi0+Ix9txYvEh8IfYtbb8Ubaq2kbAqIiKi0kurw1dycjLmzZsHe3t7GBoaomrVqhg5ciSeP39eoP14eXlhwYIFcHNzg5WVFWQyGWxsbPLc7uHDhxgxYgRq1aoFfX19mJmZoWXLlvjpp5+QlpZWyFdFRFT8UpWpmHhhIh7GPBT7ZrScgV61e0lYFRERUemmJ3UBhZWSkoLOnTvDx8cHVapUQd++fREaGopdu3bhxIkT8PHxga2tbb72NWnSJPj7+xdo/KtXr6Jbt25ISkpCvXr10K9fP8TGxuLy5cuYOnUqjh49inPnzkFPT2t/xERUSilVSsy6PAs3Xt0Q+0Y0HIEv6n8hYVVERESln9ae+Vq8eDF8fHzQpk0bPHz4EAcOHICvry/WrFmDyMhIjBw5Mt/76t69OxYvXozTp0/j3r17+drG3d0dSUlJWLZsGe7fv48///wTp0+fxuPHj2FrawsvLy/s2bOnsC+PiKhYCIKAZdeW4WzYWbGvj10fTGk2RcKqiIiIygatDF9paWnYtGkTAGDz5s0wNTUVl02dOhWNGzeGl5cXbt68ma/9rVy5EnPmzEH37t1haWmZ5/oJCQm4ffs2jI2NMWPGjCzLqlatCnd3dwDA9evX8/uSiIhKxC8Bv+BA0AGx3b5ae8xvOx8ymUzCqoiIiMoGrQxf3t7eiI2NhZ2dHZo2bZpt+aeffgoAOH78eLGML5fLoaOT94+uQoUKxTI+EVFhHHp4CJv8NontRhUbYY3LGsh15BJWRUREVHZoZfjKvD+rWbNmOS7P7A8ICCiW8Q0MDNCxY0ckJSVh5cqVWZa9ePECmzdvhlwuxxdf8P4JItIMF59exEKfhWLbxtwGm7tshrHcWMKqiIiIyhatDF9Pnz4FAFSvXj3H5Zn9YWFhxVbD1q1bUaNGDcyaNQv169fHZ599hp49e6JOnToQBAEeHh6wt7cvtvGJiPLr9uvbmH5pOlSCCgBgZWSFrd22orxheYkrIyIiKlu0ciq+hIQEAICxcc7f2JqYmAAA4uPji60GBwcHXLlyBf3798etW7cQGBgIAJDJZOjUqRMaNGiQ733ltm5wcDCsra1x4cIFtdRcWImJiQAgeR1UeDyG2q+wxzBCEYENrzcgVZUKADCUGWKE+QgEXQtCEILUXifljL+D2o/HUPvxGGo/TTiGiYmJYtYoDK0886UJLly4ACcnJygUCly4cAFxcXEICQnBnDlzsGvXLrRr1w6RkZFSl0lEZVhMegy2Rm5FkioJAKALXYyqOArV9KtJXBkREVHZpJVnvjJnN0xKSspxeWYqNjMzK5bxo6OjMXDgQCgUCpw6dQrVqlUTx1u0aBFiY2OxceNGrF69GitWrMhzf7lNb595Rqxz587qK74QMr9dkLoOKjweQ+1X0GMYmxqLr059hbfKtwAAGWRY6bIS3W26F1eJ9AH8HdR+PIbaj8dQ+2nCMSzKWS9AS8981axZEwAQHh6e4/LM/lq1ahXL+B4eHoiOjoazs7MYvN43cOBAAMClS5eKZXwiog9JSU/BhAsTEBwbLPbNbj2bwYuIiEhiWhm+nJycAAC3bt3KcXlmf+PGjYtl/MxwZ2FhkePyzP6YmJhiGZ+IKDfpqnRMvzQdt1/fFvtGNx6NQY6DJKyKiIiIAC0NX+3atYOFhQWCg4Ph5+eXbfnBgwcBAL179y6W8a2trQEAt2/fhlKpzLY88+HKNjY2xTI+EVFOBEHAYp/F8HzmKfYNqDsA7k3cJauJiIiI3tHK8KWvrw9394wPE+PHjxfv8QKAtWvXIiAgAC4uLmjevLnYv2nTJjg6OmLWrFlFHr9nz54wMDBASEgIfvjhB6hUKnFZUFAQ5s2bB+Ddw56JiErCZr/NOPTokNh2reGKuc5zIZPJJKyKiIiIMmnlhBsAMHfuXJw7dw5Xr15F3bp10aFDB4SFhcHX1xdWVlbYuXNnlvWjoqIQFBSEiIiIbPvavn07tm/fDgBQKBQAgIiICDg7O4vrbNmyRXx4c5UqVbB69WpMnDgRy5Ytw4EDB9C0aVO8efMG//77L1JTU/HRRx9h+PDhxfTqiYiyOvDgALYFbBPbTayaYGXHldDT0dq3eSIiolJHa/9VNjQ0xMWLF7Fs2TL88ccfOHLkCCwtLTF8+HAsWrQo1wcw5yQ8PBy+vr5Z+tLS0rL0xcXFZVnu7u6Ohg0bYv369fDx8cHRo0dhbGyMpk2b4osvvsC3334LXV3dor1IIqJ8OBt2Fkt8l4htOws7bOqyCUZ6RhJWRURERP+lteELAIyMjLBw4UIsXLgwz3Xnz5+P+fPnF3jZh7i6usLV1bXA2xERqcv1l9fx/aXvIUAAAFQ2royt3bbCwiDnCYGIiIhIOlp5zxcREQFB0UGYeGEiFKqMy6XN9M2wtetWWJtYS1wZERER5YThi4hICz1PeI6x58YiQZEAADDQNcDmLptRp3wdiSsjIiKi3Gj1ZYdERGWBQqFAUFAQkpKSYG5ujpoONTHm3BhEJkcCAHRkOljVcRWaVmoqcaVERET0IQxfREQaKjIyEhs2bsCWbVuQhjTIzeRYuGYhUlNSYd7RHBW6VICeuR7mOc9Dp5qdpC6XiIiI8sDwRUSkgQIDA+Ha1RWCjYByY8rBqLYRZDIZBEFA8pNkvDn7BsHzgzFnxxwMsB8gdblERESUD7zni4hIw0RGRsK1qyv0Oumh8ujKMLY1Fh+ULJPJYGxnjBpjaqBiz4r42f1nREZGSlwxERER5QfDFxGRhtmwcQMEGwGW3Sw/uF6F7hWgqqXCxk0bS6gyIiIiKgqGLyIiDaJQKLBl2xaYdTXL1/pmXcywZdsWKBSKYq6MiIiIiorhi4hIg/j7+yNdNx1GtY3ytb6RrREUOgr4+/sXc2VERERUVAxfREQaJC4uDvpm+uI9XnmRyWTQN9NHXFxcMVdGRERERcXwRUSkQczNzZEWnwZBEPK1viAISItPg7m5eTFXRkREREXF8EVEpEGcnJygp9RDckhyvtZPfpIMuUoOJyenYq6MiIiIiorhi4hIg8jlcoz7dhzenn2br/Xjz8dj3LfjIJfLi7cwIiIiKjKGLyIiDdP3q76IfxiPqDNRH1wv+mw0dMJ0MMF9QglVRkREREWhJ3UBRET0jiAI2Ba8DTbTbBC6JhQpT1Jg2c0SRrZGkMlkEAQBycHJiL8QD50wHXie84SVlZXUZRMREVE+MHwREWmQ02Gn4RPhA4OqBrD70Q4tglvg5LaTeI3X0DPVgypZBbkq49LECe4TGLyIiIi0CMMXEZGGSFQkYtW1VWLbpZ4Lfnb/Gemr0rF9+3YkJSXBxcUFTk5OvMeLiIhIC6ktfCUlJSEqKgoVKlSAiYmJ2B8TE4MVK1bg7t27qFmzJqZNmwY7Ozt1DUtEVGps8duC18mvAQD6OvqY3Wo2ZDIZ5HI5HBwcAAAtWrSQskQiIiIqArWFr0WLFmHlypW4du0amjdvDgBITU2Fs7MzHj9+LD6z5uDBg/D390eVKlXUNTQRkdZ7GPMQvwf+Lra/bvQ1aprXlLAiIiIiUje1zXZ44cIF2NnZicELAPbu3YtHjx6hU6dOOH36NCZOnIioqCj89NNP6hqWiEjrqQQVFvsshlJQAgBqmNXAyIYjJa6KiIiI1E1t4evp06eoW7dulr5jx45BJpNh165d6NatG9atWwd7e3ucOnVKXcMSEWm9Y8HHcPv1bbE9q9UsGOoZSlgRERERFQe1ha+YmBiUK1dObAuCgCtXrqBx48aoUaOG2O/k5IRnz56pa1giIq0WmxqLn26+uxqga82u6FC9g4QVERERUXFRW/iytrZGSEiI2L558yZiYmLg4uKSZT2ZTKauIYmItN6GWxsQnRINADDSM8L3rb6XuCIiIiIqLmoLX02aNMG1a9dw5MgRxMfHY9GiRZDJZPj444+zrPfo0SNUrVpVXcMSEWmtu1F38dfDv8T2GKcxsDaxlrAiIiIiKk5qC18zZswAAAwYMADlypXD8ePH4eTkhM6dO4vrvHr1Cv7+/lkm5SAiKouUKiUW+SyCgIyZYG0tbPFFvS8kroqIiIiKk9rCV9u2bXH48GG0b98ejo6OGDZsGI4dOwYdnXdD7Nu3D2ZmZujZs6e6hiUi0koHHx7E/Tf3xfZc57mQ6/LByURERKWZ2p7zBQC9e/dG7969c10+efJkTJ48WZ1DEhFpnTfJb7D+9nqx7WbrhpbWLSWsiIiIiEqC2s58ERFR/qy9uRbxafEAAFO5Kb5r8Z3EFREREVFJUFv4evXqFS5duoRXr15l6Q8ODsagQYPQsGFDfPTRR/j333/VNSQRkda5+eomjgUfE9vuTd1R0aiihBURERFRSVFb+Fq+fDk6deqE2NhYsS8uLg7t27fHX3/9hfv37+Off/5B165d8ejRI3UNS0SkNRQqBRb7LBbbjpaO+NzhcwkrIiIiopKktvDl6emJ+vXrw97eXuzbvXs3Xr16hcGDByMoKAhr165FcnIy1qxZo65hiYi0xh+Bf+Dx28die67zXOjpqPXWWyIiItJgagtfz58/h62tbZY+Dw8P6OnpYd26dahbty4mT54MJycneHl5qWtYIiKt8CrxFbb4bRHbA+oOgJOVk4QVERERUUlTW/iKj4+HsbGx2FYqlfj333/RvHlzVKz47n4GR0dHhIeHq2tYIiKtsOrGKiSlJwEAyhmUw+Rmk6UtiIiIiEqc2sJX1apV8eDBA7F95coVJCQkwNXVNct66enp0NfXV9ewREQa7+rzqzgdelpsT242GeUMy0lXEBEREUlCbeGrTZs2CAgIwLp163Dnzh3MnTsXMpks23O/AgMDUa1aNXUNS0Sk0dKUaVh6banYbmzVGP3r9pewIiIiIpKK2sLXrFmzYGBggGnTpqFJkybw9vaGq6sr2rZtK64TGhqK+/fvo3Xr1uoalohIo+26uwthcWEAAB2ZDn5w/gE6Mj5ikYiIqCxS2zRbDRo0wJUrV7B+/XpERUWhefPmmD59epZ1Tp8+DScnJ/Tr109dwxIRaazw+HD8eudXsT3YcTAcLR0lrIiIiIikpNY5jps1a4bffvst1+Xffvstvv32W3UOSUSksZZfW45UZSoAoKJRRYxvMl7iioiIiEhKvPaFiKgYXHx6EV7h7x6r8V2L72CmbyZhRURERCQ1tT/d89WrV9i5cycuX76M58+fAwCqVauGjh07YsSIEahcubK6hyQi0ijJ6clYfm252G5l3Qof1f5IwoqIiIhIE6g1fB06dAgjR45EQkICBEEQ++/cuYPTp09j+fLl2LFjBwYMGKDOYYmINMqvAb/iReILAICeTA9zWs+BTCaTuCoiIiKSmtouO7xx4wYGDx6MxMRE9O/fH4cPH8bt27fh5+eHI0eO4JNPPkFCQgKGDBmCGzduqGtYIiKN8iT2CXbd2yW2v2zwJWzL2UpYEREREWkKtZ35WrZsGZRKJQ4ePIj+/bM+w6Zx48bo06cPDh8+jAEDBmD58uU4ePCguoYmItIIgiBgqe9SpKvSAQDWJtb4tjEnGSIiIqIMajvzdeXKFbRt2zZb8Hpf//790a5dO1y+fFldwxIRaYzToafhG+Ertme2nAljubGEFREREZEmUVv4io2NRc2aNfNcr2bNmoiNjVXXsEREGiEhLQErr68U2+2rtUfnmp0lrIiIiIg0jdrCl7W1NW7fvp3nen5+frC2tlbXsEREGmGL/xZEJkcCAPR19DG71WxOskFERERZqC189ejRA0FBQZg9ezaUSmW25YIgYO7cuXjw4AF69uyprmGJiCQXFB2EPwL/ENujGo1CDfMaElZEREREmkhtE2788MMP+Pvvv7FixQrs27cPn332GWxsbAAAYWFh+OuvvxAaGooKFSpg7ty56hqWiEhSKkGFxT6LoRQyvnSqYVYDIxuNlLgqIiIi0kRqC1/Vq1fHhQsXMHToUNy9exerVq0SL7nJfOZXo0aN8Pvvv6N69erqGpaISFJHHx+FX6Sf2J7dejYMdA2kK4iIiIg0llofstyoUSMEBATA09MTly9fxosXGQ8ZrVq1Kjp06ABXV1d1DkdEJKnY1Fj8dPMnsd2tVje0r9ZewoqIiIhIk6k1fGVydXXNNWjt3LkT4eHhmDdvXnEMTURUYtbfWo+Y1BgAgJGeEWa0nCFxRURERKTJ1DbhRn79+uuvWLBgQUkPS0SkVnci7+Dgw3cPix/rNBbWJpzJlYiIiHJX4uFLnZKTkzFv3jzY29vD0NAQVatWxciRI/H8+fMC7cfLywsLFiyAm5sbrKysIJPJxMlC8pKQkIAFCxagcePGMDU1hYWFBRo2bIjx48cjISGhEK+KiDSdUqXEIp9FEJBxP6udhR2G1R8mcVVERESk6YrlssOSkJKSgs6dO8PHxwdVqlRB3759ERoail27duHEiRPw8fGBra1tvvY1adIk+Pv7F7iGkJAQdOnSBSEhIbC1tUWvXr2QmpqKoKAgbNmyBbNmzYKpqWmB90tEmu2vh38hMDpQbM9xngO5jlzCioiIiEgbaG34Wrx4MXx8fNCmTRucOXNGDDlr167FtGnTMHLkSHh6euZrX927d8fAgQPRsmVLVK9eHQ0aNMhzm9TUVPTq1QtPnz7F1q1b8e2332ZZfvfuXVhaWhb4dRGRZotKjsKGWxvE9se2H6OldUsJKyIiIiJtoZXhKy0tDZs2bQIAbN68OcvZpalTp+K3336Dl5cXbt68iebNm+e5v5UrV4p/f/nyZb5qWL9+PYKCgjB9+vRswQsAGjZsmK/9EJF2+enmT4hXxAMAzORmmNZimsQVERERkbbQynu+vL29ERsbCzs7OzRt2jTb8k8//RQAcPz48WKr4ddffwUATJgwodjGICLNcuPlDRwLPia23Zu6o6JRRQkrIiIiIm2ilWe+Mu/PatasWY7LM/sDAgKKZfxnz57h8ePHqF69OmrUqAFvb28cO3YMsbGxqF27NgYMGIA6deoUy9hEJA2FSoElvkvEdj3Levjc4XMJKyIiIiJtU+jwpaurq846CuTp06cAgOrVq+e4PLM/LCysWMa/f/8+gIyHR48fPx5btmzJsnzu3LlYvnw5pk3L3+VIud1jFhwcDGtra1y4cKFoBRdRYmIiAEheBxUej2HRXYi7gMexjwEAMsjQU68nvDy9Smx8HkPtxuOn/XgMtR+PofbThGOYmJgIExOTQm9f6MsOBUEo9J+iypzC3djYOMflmT+Q+Pj4Io+Vk5iYjIeq3rp1C1u3bsX8+fPx7NkzREREYMWKFQCA7777Dh4eHsUyPhGVrLfpb3Eq7pTYdjZxho2BjXQFERERkVYq9JkvlUqlzjq0SuZrT09Px9ixY/Hjjz+Ky2bMmIGoqCisWrUKS5cuhZubW577u3fvXo79mWfEOnfurIaqCy/z2wWp66DC4zEsmmme05AmpAEAyhmUw8qPV6KcYbkSrYHHULvx+Gk/HkPtx2Oo/TThGBblrBegpRNuZM5umJSUlOPyzFOSZmZmxTo+AIwYMSLb8sw+X19fpKSkFEsNRFQyrj6/ijNhZ8T2lOZTSjx4ERERUemgleGrZs2aAIDw8PAcl2f216pVq1jGf3+/NjY22ZZn9imVSkRHRxdLDURU/FKVqVkm2XCyckK/Ov2kK4iIiIi0mlaGLycnJwAZ91zlJLO/cePGxTK+o6MjDA0NAby7/+t97weu98+SEZF22XV3F57GZ0zwoyPTwVznudCRaeXbJhEREWkArfwU0a5dO1hYWCA4OBh+fn7Zlh88eBAA0Lt372IZ38DAAD169AAAeHp6Zlvu5ZUxA5qtrS3Mzc2LpQYiKl7P4p9h+53tYnuI4xA4WjpKWBERERFpO60MX/r6+nB3dwcAjB8/XrzHCwDWrl2LgIAAuLi4oHnz5mL/pk2b4OjoiFmzZqmlhhkzZgAAFi1ahIcPH4r9ISEh+OGHHwAAY8aMUctYRFSyBEHA8mvLkapMBQBUNKqIcU3GSVwVERERaTutfMgykPEsrXPnzuHq1auoW7cuOnTogLCwMPj6+sLKygo7d+7Msn5UVBSCgoIQERGRbV/bt2/H9u0Z33ArFAoAQEREBJydncV1tmzZkuWhzm3btsW8efOwcOFCNG3aFO3atYOuri68vb0RHx+PXr16YerUqcXx0omomF18dhGXwi+J7ektpsNMv3gm8CEiIqKyQ2vDl6GhIS5evIhly5bhjz/+wJEjR2BpaYnhw4dj0aJFuT6AOSfh4eHw9fXN0peWlpalLy4uLtt2CxYsgJOTE9atWwcfHx+kp6fDwcEBX331Fdzd3SV9EDURFU6SIgnLry0X262sW6FX7V4SVkRERESlhdaGLwAwMjLCwoULsXDhwjzXnT9/PubPn1/gZXn55JNP8MknnxRqWyLSPL/e+RURiRlnyPV09DCn9RzIZDKJqyIiIqLSQG33fC1YsCDXqd+JiLTBk9gn2H1vt9j+qv5XsC1nK11BREREVKqoNXzVrl0bvXv3xrFjx6BSqdS1ayKiYicIApb6LEW6Kh0AUMWkCkY3Hi1xVURERFSaqC18LV68GDVr1oSHhwf69++PGjVq4IcffkBoaKi6hiAiKjanQk7B9+W7+zy/b/U9jOXGElZEREREpY3awtfs2bMRHByMM2fOYODAgXjz5g2WLFmCOnXqoGfPnjh06BDS09PVNRwRkdokpCVg9Y3VYrtj9Y7oXKOzhBURERFRaaT253x17doV+/fvx/Pnz7F69Wo4ODjgzJkz+Oyzz1C9enXMnDkTjx49UvewRESFttlvMyKTIwEABroGmNlqJifZICIiIrUrtocsV6hQAVOnTsW9e/dw5coVDB48GK9fv8aqVavg6OiILl264PDhw8U1PBFRvgRFB+GPB3+I7VGNRqGGWQ0JKyIiIqLSqtjCV6bg4GAcP34c58+fF/uqV6+Oixcv4tNPP0WrVq3w7Nmz4i6DiCgblaDCYp/FUAkZEwTVNKuJEQ1HSFwVERERlVbFEr4UCgX279+PLl26wN7eHitWrEB6ejqmTp2KBw8eICwsDN7e3ujVqxdu3LgBd3f34iiDiOiDjj4+Cr9IP7E9u/VsGOgaSFcQERERlWpqfchyYGAgfv31V+zZswfR0dEQBAFt27bFmDFjMHDgQBgYvPtQ06ZNG5w4cQLOzs7w8vJSZxlERHl6m/IWa2+uFdvdanVDu2rtJKyIiIiISju1ha/27dvj33//hSAIMDc3x9ixYzFmzBg0bNjwg9s1aNAA169fV1cZRET5sv72erxNfQsAMNIzwoyWM6QtiIiIiEo9tYWvq1evolmzZhgzZgyGDBkCY+P8PR9n1KhR6Nixo7rKICLKU0BkAA49PCS2xzmNg7WJtYQVERERUVmgtvB1/fp1NG/evMDbtWnTBm3atFFXGUREH6RUKbHYZzEECACAOuXqYGj9oRJXRURERGWB2ibc6NatG1xcXNS1OyKiYvHnwz8RGB0otue0ngO5jlzCioiIiKisUFv4Sk9PR/Xq1dW1OyIitYtKjsLGWxvFdm/b3mhh3ULCioiIiKgsUVv4atCgAZ4/f66u3RERqd3aG2sRr4gHAJjJzTC1xVSJKyIiIqKyRG3ha8KECbhy5QquXLmirl0SEanN9ZfXcfzJcbE9odkEVDSqKGFFREREVNaodar5UaNGoUePHhg1ahR69+6NmjVrwtDQMMf1a9asqa6hiYg+SKFSYInPErFdz7IePrP/TMKKiIiIqCxSW/iysbGBTCaDIAjYtGkTNm3alOu6MpkM6enp6hqaiOiD9t7fi+DYYACADDL84PwDdHV0Ja6KiIiIyhq1ha+OHTtCJpOpa3dERGrxMvElfvb/WWx/av8pGlk1krAiIiIiKqvUFr48PT3VtSsiIrVZeX0lktOTAQDlDcpjUrNJEldEREREZZXaJtwgItI03s+9cTbsrNie0nwKLAwsJKyIiIiIyjK1nfn6r0ePHiEqKgoVKlSAvb19cQ1DRJSjVGUqlvouFdtNrJqgb52+ElZEREREZZ1az3ylpqZi9uzZqFixIhwdHdG+fXssX75cXL537140a9YMfn5+6hyWiCibnXd34mn8UwCAjkwHc53nQkfGk/1EREQkHbV9EklOToarqytWrFgBfX19fPTRRxAEIcs6nTt3hr+/P/788091DUtElM2zuGfYHrBdbA9xHAIHSwcJKyIiIiJSY/hauXIlfH19MXLkSDx58gTHjx/Ptk7VqlVRv359nDt3Tl3DEhFlIQgCll1bhjRVGgDAysgK45uMl7gqIiIiIjWGrwMHDqBmzZr4+eefc32wMgA4ODjg2bNn6hqWiCiLC88u4PLzy2J7esvpMNU3lbAiIiIiogxqC18hISFo0aIF9PQ+PIeHvr4+YmJi1DUsEZEoSZGE5dfe3Wfa2ro1etr0lLAiIiIionfUFr6MjIzyFapCQkJQvnx5dQ1LRCT6JeAXvEx8CQDQ09HDbOfZfPg7ERERaQy1ha8mTZrgxo0biIyMzHWdkJAQ3L59Gy1btlTXsEREAIAnb5/gt3u/ie3hDYbD1sJWwoqIiIiIslJb+Prmm28QHx+PwYMHIyoqKtvyt2/fYuTIkVAoFBg9erS6hiUigiAIWOK7BOlCOgCgqklVjG7M9xkiIiLSLGp7yPLgwYNx/Phx7N+/H7a2tmjbti0AwNvbG3379oWXlxfi4uLw5Zdf4uOPP1bXsEREOBlyEtdeXhPb37f6HkZ6RhJWRERERJSdWp84+vvvv2PFihUwNDTEmTNnAACPHj3C8ePHIZPJsGTJEuzatUudQxJRGRefFo/VN1aLbZfqLuhUo5OEFRERERHlTG1nvgBAJpNh+vTpmDp1Km7duoXQ0FCoVCpUr14dLVu2hL6+vjqHIyLCZr/NiErOuNTZQNcAM1vN5CQbREREpJHUGr4y6erqomXLlpxYg4iK1YPoB9j3YJ/Y/qbRN6huVl3CioiIiIhyp9bLDomISopKUGGxz2KoBBUAoJZ5LYxoOELiqoiIiIhyp7YzXyNHjsz3ujKZDDt27FDX0ERUBh15fAT+kf5ie3ar2dDX5aXNREREpLnUFr52796d5zoymQyCIDB8EVGRvE15i59u/iS2u9fqjrbV2kpYEREREVHe1Ba+Ll68mGO/SqXCs2fPcObMGezfvx9TpkxB79691TUsEZVB626tw9vUtwAAYz1jzGg5Q9qCiIiIiPJBbeHLxcXlg8u//PJLuLm54auvvkKfPn3UNSwRlTH+kf74+9HfYntck3GobFJZwoqIiIiI8qdEJ9wYPHgwGjRogPnz55fksERUSihVSizxWQIBAgCgTrk6GFJviMRVEREREeVPic92WLduXdy4caOkhyWiUuBA0AEERgeK7bnOcyHXkUtYEREREVH+lWj4UqlUCAgIgI4OZ7gnooKJSo7CxtsbxXYfuz5oXrm5hBURERERFUyJpKCkpCT4+flh8ODBePToUZ73hxER/deaG2uQoEgAAJjpm2FK8ykSV0RERERUMGqbcENXVzfPdQRBgJWVFVatWqWuYYmoDLj+8jpOPDkhtic2nYiKRhUlrIiIiIio4NQWvmrUqAGZTJbjMn19fVSpUgUuLi4YP348KlWqpK5hiaiUUygVWOKzRGzXr1AfA+0HSlgRERERUeGoLXyFhoaqa1dERKI9gXsQHBsMAJBBhh+cf4CuTt5n2omIiIg0DWe+ICKN9TLxJbb6bxXbA+0HomHFhhJWRERERFR4ajvzlSk0NBRRUVFQKpWoWLEiateuzdkNiahQVlxbgeT0ZABAeYPymNhsosQVERERERWeWlLRjRs3MGTIEFSoUAF2dnZo3bo12rZtC3t7e1hYWGDAgAG4cuWKOoYiojLiyvMrOPf0nNie2mIqLAwsJKyIiIiIqGiKHL6+//57ODs7Y//+/YiJiYFcLkelSpVQqVIlyOVyJCYm4vDhw3BxccHs2bMhCIK47Zs3b3D8+PGilkBEpUyqMhVLfZeK7aaVmqKPXR8JKyIiIiIquiKFr++++w6rVq2Cvr4+pk2bhlu3biExMRERERGIiIhAYmIibt++je+++w6GhoZYsWIFJk+eDCDj8sR27drh9u3b6ngdRFSK7LyzE8/inwEAdGW6mNN6DnRkvHyZiIiItFuhP834+vrip59+Qo0aNXDr1i2sWrUKTZo0yfK8L11dXTg5OWHlypW4desWqlWrhs2bN2PLli1o27YtHj16hPLlyxe6+OTkZMybNw/29vYwNDRE1apVMXLkSDx//rxA+/Hy8sKCBQvg5uYGKysryGQy2NjYFGgfaWlpqF+/PmQyGfT01H4rHVGZ8SzuGbbf2S62h9QbAgdLBwkrIiIiIlKPQqeErVszZiDbt28fHB0d81zfwcEB+/fvR/v27TFhwgQIgoCJEydiwoQJhRo/JSUFnTt3ho+PD6pUqYK+ffsiNDQUu3btwokTJ+Dj4wNbW9t87WvSpEnw9/cvVB2Zli5digcPHhRpH0RlnSAIWHptKdJUaQCASkaVMM5pnMRVEREREalHoc98eXl5oXHjxmjbtm2+t2nbti2cnJwAAEuWLMG6desKOzwWL14MHx8ftGnTBg8fPsSBAwfg6+uLNWvWIDIyEiNHjsz3vrp3747Fixfj9OnTuHfvXoFrCQwMxLJly/DNN98UeFsieufC0wu48vzd5DzTW06Hqb6phBURERERqU+hw9fLly9Rr169Am+XeZZs1qxZhR0aaWlp2LRpEwBg8+bNMDV99+Fs6tSpaNy4Mby8vHDz5s187W/lypWYM2cOunfvDktLywLVIggCRo8ejXLlymH58uUF2paI3klSJGH59Xe/Q85VnNHDpoeEFRERERGpV6HDl6GhIZKSkgq8XXJyMszNzQs7LADA29sbsbGxsLOzQ9OmTbMt//TTTwGgRGZS3LZtG65cuYI1a9YU6f41orJuW8A2vEx8CQDQ09HD7NazIZPJJK6KiIiISH0KHb7s7Oxw9epVKJXKfG+jVCpx9epV2NnZFXZYABDvz2rWrFmOyzP7AwICijROXiIiIjBz5kx06dIFw4YNK9axiEqz4LfB+N+9/4ntEQ1GoLZFbQkrIiIiIlK/QocvNzc3vHnzBitWrMj3NitXrsSbN2/Qu3fvwg4LAHj69CkAoHr16jkuz+wPCwsr0jh5cXd3R0pKCrZs2VKs4xCVZoIgYInvEqQL6QCAqiZV8U1j3j9JREREpU+hZzucNGkSNmzYgHnz5kGlUmHWrFlZppl/n1KpxPLlyzFv3jyUL18eEydOLHTBAJCQkAAAMDY2znG5iYkJACA+Pr5I43zI0aNH8ffff+PHH3+Evb19kfbVoEGDHPuDg4NhbW2NCxcuFGn/RZWYmAgAktdBhafJx/BG4g1cj74utj8y+gj/XvpXwoo0kyYfQ8obj5/24zHUfjyG2k8TjmFiYqKYNQqj0OGrfPny+Ouvv+Dm5oYff/wR27Ztw8CBA9GiRQtYWVkBACIjI3Hjxg0cPHgQz58/h1wux59//qn190bFx8fD3d0d9vb2RZo4hKisS1Yl48jbI2K7oWFDNDJqJF1BRERERMWoSE8D7tKlCy5fvoyvvvoKDx48wPr167OtIwgCgIxZDv/3v/+hRYsWRRkSAMTZDXOb8CMzFZuZmRV5rJzMnj0b4eHhOHfuHAwMDIq8v9ymt888I9a5c+cij1EUmd8uSF0HFZ6mHsNlvssQ/zzjDLWBrgFWfbQK1c1yvpy4rNPUY0j5w+On/XgMtR+PofbThGNYlLNeQBHDFwC0bNkS9+/fx6lTp+Dh4QF/f3+8efMGAFChQgU4OTnho48+Qq9evdQ2c1nNmjUBAOHh4Tkuz+yvVauWWsb7r+PHj8PQ0BCLFi3CokWLsi1XKpVwdXUFAKxbtw5NmjQpljqItFngm0DsD9ovtkc3Hs3gRURERKVakcNXpl69eqFXr17q2t0HZT6o+datWzkuz+xv3LhxsdWQkpICLy+vXJdnLnv79m2x1UCkrVSCCot9FkMlqAAANuY2GN5guLRFERERERUztYWvktSuXTtYWFggODgYfn5+2c4sHTx4EACKPKtibkJDQ3NdJpPJoKuri/T09GIZm6g0OPzoMAKi3j0KYlbrWdDX1ZewIiIiIqLiV+ip5qWkr68Pd3d3AMD48ePFe7wAYO3atQgICICLiwuaN28u9m/atAmOjo6cIINIYjEpMfjp1k9iu4dND7St2lbCioiIiIhKhlae+QKAuXPn4ty5c7h69Srq1q2LDh06ICwsDL6+vrCyssLOnTuzrB8VFYWgoCBERERk29f27duxfft2AIBCoQCQ8QBlZ2dncZ0tW7bk+lBnIsq/9bfWIzY1FgBgrGeM6S2mS1wRERERUcnQ2vBlaGiIixcvYtmyZfjjjz9w5MgRWFpaYvjw4Vi0aFGuD2DOSXh4OHx9fbP0paWlZemLi4tTW+1EZZV/pD8OPToktsc1GYfKJpUlrIiIiIio5GjlZYeZjIyMsHDhQjx+/BipqamIiIjArl27cgxe8+fPhyAI2L17d67LPvQnc/bCvAiCwPu9iHKQrkrHYp/FYrtu+boYUm+IhBURERERlSytDl9EpD0OBB3Ag+gHYntu67mQ68glrIiIiIioZDF8EVGxi0yKxKbbm8R2X7u+aFaZ91ASERFR2cLwRUTFbs3NNUhQJAAAzPTNMKX5FIkrIiIiIip5DF9EVKyuRVyDxxMPsT2p6SRUMKogYUVERERE0mD4IqJio1AqsMR3idhuUKEBPrX/VMKKiIiIiKTD8EVExeZ/9/+HJ7FPAAAyyPCD8w/Q1dGVuCoiIiIiaTB8EVGxiEiIwLaAbWL7M4fP0KBiAwkrIiIiIpIWwxcRFYsV11cgOT0ZAGBpaIkJTSdIXBERERGRtBi+iEjtLoVfwvmn58X21OZTYWFgIWFFRERERNJj+CIitUpJT8Ey32Viu1mlZuhj10fCioiIiIg0A8MXEanVzrs7EZ4QDgDQlelijvMcyGQyiasiIiIikh7DFxGpzdO4p9hxZ4fYHlpvKOzL20tYEREREZHmYPgiIrUQBAFLry1FmioNAFDJqBLGNRkncVVEREREmoPhi4jU4vzT8/B+7i22p7eaDhO5iYQVEREREWkWhi8iKrIkRRKWX1sutttUaYMetXpIWBERERGR5mH4IqIi2xqwFa+SXgEA5DpyzG49m5NsEBEREf0HwxcRFcnjmMfYc2+P2B7RcARsLGykK4iIiIhIQzF8EVGhCYKAJb5LkC6kAwCqmVbDqEajJK6KiIiISDMxfBFRoZ14cgI3Xt0Q27NazYKRnpGEFRERERFpLoYvIiqUuLQ4rLmxRmy71nCFSw0XCSsiIiIi0mwMX0RUKJtub8KblDcAAENdQ8xsNVPiioiIiIg0G8MXERXY/Tf3cSDogNge3Xg0qplWk7AiIiIiIs3H8EVEBaISVFjiswQqQQUAsDG3wVcNvpK4KiIiIiLNx/BFRAXy96O/ERAVILZnt54NfV19CSsiIiIi0g4MX0SUbzEpMVh3a53Y7mnTE22qtpGuICIiIiItwvBFRPm27tY6xKbGAgBM5CaY3nK6xBURERERaQ+GLyLKF7/Xfvj70d9ie5zTOFQyriRhRURERETaheGLiPKUrkrHYp/FYtu+vD2G1BsiYUVERERE2ofhi4jydCDoAIJigsT2XOe50NPRk7AiIiIiIu3D8EVEHxSZFImNtzeK7X51+qFppaYSVkRERESknRi+iOiDVt9YjURFIgDAXN8cU5pPkbgiIiIiIu3E8EVEufKN8MXJkJNie1KzSbA0tJSwIiIiIiLtxfBFRDlSKBVY4rtEbDes0BAD6g6QsCIiIiIi7cbwRUQ5+u3+bwiJDQEAyCDD3DZzoaujK3FVRERERNqL4YuIsnmR8AK/BPwitj9z+AwNKjSQsCIiIiIi7cfwRUTZrLi2AsnpyQAAS0NLTGg6QeKKiIiIiLQfwxcRZXEp/BIuPLsgtqe1mAYLAwsJKyIiIiIqHRi+iEiUkp6Cpb5LxXazSs3Q27a3hBURERERlR4MX0Qk2nF3B54nPAcA6Mp0Mdd5LmQymcRVEREREZUODF9EBAAIiwvDjjs7xPawesNQt3xdCSsiIiIiKl0YvogIgiBgme8yKFQKAEAl40oY22SsxFURERERlS4MX0SEc0/PwfuFt9ie0XIGTOQmElZEREREVPowfBGVcUmKJCy/tlxst63aFt1rdZewIiIiIqLSieGLqIzb6r8Vr5NeAwDkOnLMbj2bk2wQERERFQOGL6Iy7HHMY+y5v0dsj2w4ErXMa0lYEREREVHpxfBFVEYJgoDFvouRLqT/X3v3Hd5k2bYB/MxqmnS3lC5my15lDwFZMhSRpQiCiqCIMl9QUBFlvfKJgGxeleGeTEWUDSLIXsoossoqdM+0aZrc3x9pQtOkbVrapGnP33HkaHo/I1fy5EmeK/cCAIR5huHlpi87OSoiIiKiiovJF1Elte3aNpy8f9L8/9tt34a73N2JERERERFVbEy+iCqh1OxULDyx0Px/t+rd0KV6FydGRERERFTxMfkiqoSWn1qOxKxEAIC7zB1vtX3LyRERERERVXxMvogqmfMJ5/Hj5R/N/78a+SpCPUOdGBERERFR5SB3dgBEVLZ0Oh2ioqKg0Wjg6eWJxXcXwyAMAIDaPrXxYqMXnRwhERERUeXg0slXZmYm5s+fj++//x43b96Ev78/+vTpg7lz5yIsLMzu/Rw4cAD79+/HsWPHcOzYMcTHx6NmzZq4ceOGzfV1Oh327duHn3/+Gfv378e1a9cghECtWrXQt29fTJ8+HYGBgaX0LIlKJi4uDsuWL8OqT1YhG9lQeCmQ/VE2tFot/Lr4IaBHAGb0mgGFTOHsUImIiIgqBZdNvrKystC9e3ccOXIEISEh6N+/P27cuIH169dj27ZtOHLkCMLDw+3a16RJk3D27Fm7H/vAgQPo3bs3AKBWrVp4/PHHodPp8Ndff2HRokX45ptvsH//ftSvX79Ez43oYV28eBFdH+sKUUvAd6wvVLVVkEgkEEIg81omEnYl4NbcW/Dp6QOEODtaIiIiosrBZft8zZs3D0eOHEGHDh1w+fJl/PDDDzh69CgWLVqEuLg4jBo1yu599erVC/PmzcOOHTtw/vz5IteXSqUYMmQIjh49iuvXr2Pjxo34+eefceXKFfTu3Rv37t3DSy+99DBPj6jE4uLi0PWxrpB3kyNoTBDU4WpIJBIAgEQigTpCjepjq8Onlw+6PtYVcXFxTo6YiIiIqHJwyeQrOzsbK1asAACsXLkSnp6e5mVTpkxBs2bNcODAAZw8ebKgXVhYsGABZsyYgV69esHf37/I9bt3744ffvgBbdu2tSj38fHBunXrAAB//fUXoqOj7X1KRKVm2fJlELUE/HsW/l727+kPQ00Dlq9Y7qDIiIiIiCo3l0y+Dh06hJSUFERERKBFixZWy59++mkAwC+//OLo0BAaGmru73X37l2HPz5VbjqdDqs+WQWvx7zsWt+rhxdWfbIKOp2ujCMjIiIiIpdMvkz9s1q2bGlzuan83LlzDovJJDk5GUlJSQCA4OBghz8+VW5nz55FjiwHqtoqu9ZXhaugk+qK1eeRiIiIiErGJZOvmzdvAgCqVatmc7mp3BnN/lauXImcnBw0bdoUtWvXdvjjU+WWmpoKNy83cx+vokgkErh5uSE1NbWMIyMiIiIilxztMD09HQCgVqttLvfw8AAApKWlOSwmADh9+jTmzZsHAPjwww/t3q5x48Y2y69evYrg4GDs3bu3VOIrqYyMDABwehxUtKioKGQlZ0EIYVcCJoRAVnIWoqKiHBAdPQyeh66Nx8/18Ri6Ph5D11cejmFGRoY51ygJl6z5Ko/u37+PQYMGISsrC5MnT8bjjz/u7JCoEoqIiIAkR4LM65l2rZ95LRPSHCkiIiLKODIiIiIicsmaL9PohhqNxuZyU1bs5WXfoAMPKy0tDU888QRu3LiBZ555BosWLSrW9gUNb2+qEevevftDx/gwTL8uODsOKppBGNCifwuc3XkW6rG2a4bzStuThknjJ6FXr14OiI4eBs9D18bj5/p4DF0fj6HrKw/H8GFqvQAXrfmqUaMGAOD27ds2l5vKa9asWeaxZGVl4amnnsKpU6fQq1cvfP3115BKXfJlJRenN+jx3qH3ENsiFpp/NYjfGV/o+om7EiGNlmLC+AkOipCIiIiocnPJmq/IyEgAwKlTp2wuN5U3a9asTOPIycnBs88+i/379+ORRx7Bpk2b4ObmVqaPSWRLjiEHMw/NxLZr2yD3lqPW1Fq4/fFt6G/o4dXDC6pwFSQSCYQQyLyaibS9aZBGS7F/937z1AhEREREVLZcsoqmY8eO8PHxwdWrV3HmzBmr5Rs2bAAA9OvXr8xiEELgpZdews8//4zmzZvj119/fehqSKKSyDHk4J2D72DbtW3mso4tOuLimYsY+9hYJH+SjOjp0bg95zZuvXMLKZ+lYOxjY/H36b/RoEEDJ0ZOREREVLm4ZPLl5uaG8ePHAwDGjRtn7uMFAIsXL8a5c+fQpUsXtGrVyly+YsUKNGjQAG+//XapxDB58mR8/fXXaNCgAXbu3AlfX99S2S9RcegMOkz/Yzp+u/GbuaxdcDus7LESNUNrYs7sObh3+x4WzFqAd19/F3t+3oOYWzGYM3sOa7yIiIiIHMwlmx0CwLvvvovdu3fj8OHDqFu3Ljp37ozo6GgcPXoUgYGBWLduncX68fHxiIqKQkxMjNW+1qxZgzVr1gAAdDodACAmJgbt27c3r7Nq1Srz5M1bt27FsmXLAADVq1fHm2++aTPGt956izULVGZ0eh3e/ONN7Lm5x1zWIaQDlnZfCpX8wSTLCoUC9evXBwC0bt3a4XESERERkZHLJl/u7u7Yt28f5s+fj2+//RZbtmyBv78/Ro4ciblz5xY4AbMtt2/fxtGjRy3KsrOzLcryTkKblJRkvr9r164C9zty5EgmX1QmsvXZmLp/Kvbf3m8u6xjWEUu7LYVSpnReYERERERUIJdsdmiiUqkwZ84cXLlyBVqtFjExMVi/fr3NxGvWrFkQQuDzzz8vcFlht65du5rXHzlyZJHr59+GqLRo9VpM3jfZIvF6tNqjWNZtGRMvIiIionLMZWu+iCqjrJwsTN43GYfuHjKXdaveDYu6LIJCpnBiZERERERUFJeu+SKqTDJzMjFh7wSLxKtnzZ5Y1JWJFxEREZErYM0XkQvQ6DQYv3c8jt87bi7rXas35neeD4WUiRcRERGRK2DyRVTOZegy8Pru13Eq9sGk4k/UfgL/7fRfyKU8hYmIiIhcBZsdEpVj6dnpGLtrrEXi9VTEU/ig0wdMvIiIiIhcDK/eiMqptOw0jN01Fufiz5nLBtYZiPc7vA+ZVObEyIiIiIioJJh8EZVDKdoUvLrrVZxPOG8ue7re05jZfiakElZYExEREbkiJl9E5UxyVjLG7BqDi4kXzWXP1n8W77R7h4kXERERkQtj8kVUjiRmJWLMzjGISooylw1vOBzT20yHRCJxYmRERERE9LCYfBGVEwmZCXh558u4knzFXPZCoxfwRus3mHgRERERVQBMvojKgfjMeLy842VcTblqLnupyUv4T8v/MPEiIiIiqiCYfBE5WawmFqN3jMaN1BvmsleavoIJLSYw8SIiIiKqQJh8ETnRvYx7eHnny4hOjTaXvRb5Gl6LfI2JFxEREVEFw+SLyEli0mMwasco3E6/bS4b33w8Xo181YlREREREVFZYfJF5AR30u9g9I7RuJN+x1w2qeUkvNz0ZSdGRURERERlickXkYPdSr2F0TtHIyYjxlz2Rus38GLjF50YFRERERGVNSZfRA4UnRqN0TtG477mvrlsepvpGNFohBOjIiIiIiJHYPJF5CDXU65j9I7RiMuMM5e90+4dDGswzIlREREREZGjMPkicoCryVcxesdoJGQlmMtmtp+JIfWHODEqIiIiInIkJl9EZezfpH/x8s6XkZiVCACQQIJZj8zCoLqDnBwZERERETkSky+iMhSVGIVXdr6CJG0SAGPiNbfjXPSv09/JkRERERGRozH5IiojFxIuYMyuMUjRpgAApBIp5nWch34R/ZwcGRERERE5A5MvojJwPv48Xtn1CtKy0wAAMokM8zvPx+O1H3dyZERERETkLEy+iErZubhzGLtrLNJ0xsRLLpHjw0c/RK9avZwcGRERERE5E5MvolJ0JvYMxu4eiwxdBgBj4rWwy0L0qNnDyZERERERkbMx+SIqJSfvn8Tru1+HJkcDAJBL5VjcZTG61ejm5MiIiIiIqDxg8kVUCo7fO45xe8YhMycTAOAmdcPH3T7Go9UedXJkRERERFReMPkiekhHYo5gwp4JyNJnAQCUMiWWdluKjmEdnRwZEREREZUnTL6IHsKhO4cwad8kaPVaAIC7zB3Lui9Dh9AOTo6MiIiIiMobJl9EJfTH7T/wn33/QbYhGwCgkquwssdKtAlu4+TIiIiIiKg8YvJFVAL7b+3HlP1ToDPoABgTr9WPrUaroFbODYyIiIiIyi0mX0TFtCd6D9744w3kGHIAAB4KD/zvsf+hedXmzg2MiIiIiMo1qbMDIHIlO2/sxBsHHiRengpPfNrzUyZeRERERFQk1nwR2em367/h7YNvQy/0AAAvNy982vNTNKnSxMmREREREZErYM0XkR22XduGtw6+ZU68fJQ+WNNrDRMvIiIiIrIba76IirD1ylbMPDQTAgIA4Kv0xZpea1Dfv76TIyMiIiIiV8KaL6JCbPp3k0Xi5e/uj7W91zLxIiIiIqJiY80XUQF+jPoRc4/MNf8f4B6Atb3XIsI3wolREREREZGrYvJFZMN3l77DB0c/MP8fqArEmt5rEO4T7sSoiIiIiMiVMfkiyufrC1/jw+Mfmv+vqq6Kdb3XoaZ3TSdGRURERESujskXUR5fnP8CC08sNP8f4hGCtb3Worp3dSdGRUREREQVAZMvolxr/l6DpaeWmv8P8wzD2t5rEeYZ5sSoiIiIiKiiYPJFBOCTs59gxZkV5v+reVbD2t5rEeoZ6sSoiIiIiKgiYfJFlZoQAqvPrsbqs6vNZTW8amBt77UI9gh2YmREREREVNEw+aJKSwiB5aeX47O/PzOX1fKuhbW916KquqoTIyMiIiKiiojJF1VKQggsObUE6/5ZZy4L9wnH2t5rUUVVxYmREREREVFFxeSLKh0hBBaeWIgvL3xpLqvjWwdreq1BgCrAiZERERERUUXG5IsqFSEEPjz+Ib65+I25rJ5fPXzW6zP4u/s7MTIiIiIiquiYfFGlYRAGfHD0A/wQ9YO5rKF/Q3za81P4uvs6LzAiIiIiqhSYfFGlYBAGzD0yFxsubzCXNQ5ojE96fgIfpY8TIyMiIiKiyoLJF1V4eoMes/+ajc1XNpvLmlVphtU9V8PbzduJkRERERFRZSJ1dgAPIzMzE++99x7q1asHd3d3hIaGYtSoUbhz506x9nPgwAHMnj0bffv2RWBgICQSCWrVqlXkdnq9Hh9//DGaNm0KlUqFwMBADBkyBBcvXizhM6LSpjfo8d7h9ywSr+aBzfFJz0+YeBERERGRQ7lszVdWVha6d++OI0eOICQkBP3798eNGzewfv16bNu2DUeOHEF4eLhd+5o0aRLOnj1brMc3GAx45plnsHnzZvj6+qJv376Ij4/Hhg0b8Ouvv2Lfvn1o27ZtSZ4alZIcQw5m/DkD269vN5e1rNoSqx5bBQ+FhxMjIyIiIqLKyGVrvubNm4cjR46gQ4cOuHz5Mn744QccPXoUixYtQlxcHEaNGmX3vnr16oV58+Zhx44dOH/+vF3brFu3Dps3b0bdunVx6dIlbNiwAfv378dPP/0EjUaD4cOHIycnp6RPjx6SzqDD2wfftki82gS3werHVjPxIiIiIiKncMnkKzs7GytWrAAArFy5Ep6enuZlU6ZMQbNmzXDgwAGcPHnSrv0tWLAAM2bMQK9eveDvb99w44sXLzZvGxQUZC4fPHgwnnrqKVy5cgVbt2619ylRKdIZdJj+x3T8fuN3c1m7kHZY2WMl1Aq1EyMjIiIiopLQ6XSIiorC6dOnceLECeh0OmeHVCIumXwdOnQIKSkpiIiIQIsWLayWP/300wCAX375pUwe//r167h48SJUKhX69u3r8Mengun0Oryx/w3sit5lLnsk9BGs6L4CKrnKiZERERERUXHFxcVh1nszUTMsCP83cyq+WTYLQ/p2Q82wIMx6bybi4uKcHWKxuGSfL1P/rJYtW9pcbio/d+5cmT5+kyZNoFAoHP74ZFu2PhtT9k/BgdsHzGWdwzrj424fQylTOjEyIiIiIiquixcvonePLugUlIkt/QXahMohkUgghMCxO1os/WUJWq35BDv3/oEGDRo4O1y7uGTN182bNwEA1apVs7ncVB4dHV0hH5+safVaTNo3ySLx6lqtK5Z0W8LEi4iIiMjFxMXFoXePLpgamY5v+0vRNkwGiUQCAJBIJGhXTY5v+0sxNTIdvXt0cZkaMJes+UpPTwcAqNW2++94eBgHVEhLS3OJx2/cuLHN8qtXryI4OBh79+4tQZSlJyMjAwCcHkdBsg3ZWBO/BlHaKHNZU1VT9EM//HngTydGVn6U92NIReMxdG08fq6Px9D18Ri6li8+X4dHAjMwqZ11K7O8JrVT4K+76XjzjSl44cWXyjyujIwM87V+SbhkzReRidagxafxn1okXs1VzfFSwEuQS1zytwUiIiKiSkli0EGpjYcqKQq//7wJU9pK7NruP22A337Z6hIjjbvk1alpdEONRmNzuemXDS8vL5d4/IKGtzfViHXv3r24IZYq0y9Ezo4jP41Og3F7xuFf7b/mssdrPY4POn8AudQl39plprweQ7Ifj6Fr4/FzbTqdDmvWrIFGo0GXLl0QGRlps883lW88D50kRwtkxAHpsQ/+pt+3LsuIBTKTAAAn7urhIclCm1DPInZu1DZMBg+FgL+/P1q3bl2Wz+ahar0AF02+atSoAQC4ffu2zeWm8po1a1bIxycgQ5eB13e/jlOxp8xlfcP7Yl7HeUy8iIioVMTFxWHl8mX49H8roTBkIcBDhpULgCyhwJix4zBuwkQEBgY6O0wix8vRPkiY0uNyk6nc+xZ/7wNZKcXefapWoIpaYu7jVRSJRIIADxlSU1OL/ViO5pJXqZGRkQCAU6dO2VxuKm/WrFmZPv4///wDnU5n9etXWT9+ZZeWnYbXdr+Gs3FnzWVPRTyFOY/MgUwqc2JkRERUUVTEUdaICqXLKiSZMtVQ3TeWaYufUBWHt4cH4jO1EELYlYAJIZCQkQNvb+8yjas0uGTy1bFjR/j4+ODq1as4c+YMmjdvbrF8w4YNAIB+/fqVyePXrl0bDRs2xMWLF/Hrr79iwIABDn38yiw1OxVjd43F3/F/m8sG1R2E9zu8D6mEXRiJiOjh5R1lLX9nf/Moa9WApUeNo6ydOPMPa8CofNJl5jbzi82XROX531SmLeNaI6UP4BkIeFTN8zf3ZrrvEQh4VkUk5MjaEITjd7PRNqzoH9aP3dFDC6W5gqQ8c8nky83NDePHj8d///tfjBs3Djt37jS3v1y8eDHOnTuHLl26oFWrVuZtVqxYgRUrVmDgwIGYP3/+Q8cwZcoUvPLKK5g2bRoeeeQRVK1aFQCwadMm/Pzzz6hTpw769+//0I9DD6RoUzBm1xhcSLhgLnum3jN4t/27TLyIiKjUrFy+DJ2CMu0aZe3oPQ1WrViO92fPcVB0VBI6nQ5RUVHQaDTw9vZ27X572Rk2kqgCmv5ll83I32buPlaJk2UylSfRUrjbvVsFgDFjx2HJL0vwbVjR6y89KcWYseNd4pi6ZPIFAO+++y52796Nw4cPo27duujcuTOio6Nx9OhRBAYGYt26dRbrx8fHIyoqCjExMVb7WrNmDdasWQPAeHICQExMDNq3b29eZ9WqVRaTOo8aNQrbt2/H5s2b0aBBA/To0QPx8fE4cOAAVCoVvv76a8jlLvvyljvJWcl4ZdcruJR4yVw2tP5QvNPuHbvbAxMRERVFp9Ph0/+txJb+wq71J7UyYNDyhXinmw8UKk9AoQLkKuNfhTugUAPy3L/5/5fxOqGsuUy/PW26dTM/m7VVcUB2etnG4u5rmTh5Bj1IrPImUx6BxUqoimvchIlovfZTLD2aVugPIUuP6nAo1gtLx08os1hKk8ue9e7u7ti3bx/mz5+Pb7/9Flu2bIG/vz9GjhyJuXPnFjgBsi23b9/G0aNHLcqys7MtyvJ34JNKpfjpp5+wdOlSrFu3Dtu2bYOHhwcGDx6M2bNno1GjRg/3BMksMSsRL+98Gf8mPRjVcETDEZjWZhoTLyIiKlVnz56Fu0SHNqH2tahoGyaDUqTj7Lfvo3VoMfsdS+W5SZmqgARN9eAmVxXxfwEJnun/Stgn2qn99oQwJkn2jPCXHgvobI+gXWpUfvlqqIIsm/6Zk6tAQK4s21jsFBgYiB17DqB3jy44ek+DSa0M5omWhRA4ekePZSelOBTrhR17DpSPJNoOEiGEfT/tkMOZhpovaCh6R3Hm0KzxmfF4ZecruJJ8xVz2YqMXMbX1VCZexcDhdV0fj6Fr4/FzHXs3rMFbUybg2Cg3u7dp81k6PnzMHd1rl+PftKWKB4mezQStFBI8U8JYDhK9uLg4tIpsbLPfXl5Lj+qw+JyXff32hAC0aQUkU/lG+EuPA3IyS/lZ5aPyz5dMVbWuofIMAtRVALn97+fyJi4uDqtWLMcnq1dAYciCv1qGFC2ghQJjxo7H6+MnODTxetjr83L8KUGVXZwmDqN3jsb1lOvmstFNRmNSy0lMvIiIqPRka4Dzm4ET6+B9/Cji07QQQmH/KGtZMng36ArU8DBecOsyjSPH6TRATu5fXVbZX4wXxqAzjlBXxqPUAQBkbrkJWyEJ2kMleLn/y1WA1HYNZfH67WVg1YLZeH/csMJH+MuINR7PMiMB1P42BqKwMTiFRxVAVv77N5WGwMBAvD97Dt55d2aFmG+PyReVS/cz7uPlnS/jRuoNc9mYZmMwvvl4Jl5ERFQ6Yi8CJ9YDZ783JyWRQVJk5QDH7xrsH2VN7onIN7YARV0IGgzGi/e8CVn+BM3q/8w8CZ3pZlonT1n+dfTaUniBSkifbbyVYH6nYpMpLZM3hRo6uOHTpYewZYh9tT2TWgkM+mQV3nH/EgpZaV9jSAB1QMHN/PIOTqGuwn6AhVAoFKhfvz4AlPlEymWJR5jKnXsZ9zBqxyjcSrtlLnu9+et4LfI1J0ZFREQVgi4LuLAVOLEOuHXEarFCJsGYrjWw5Fgsvh1YdPJVrFHWpFLATW28wb8EwReDwZCbkBWV4OVN4IqbBOaW6bPL9rkURq813rKSzUVn7+rhLs1Bm1D7BoNoGyaDUgacvW+wr9+eRGpMlPL3lfKsaj04hTqACRVZ4LuBypW76Xcxasco3Em/Yy6b0GICxjQb48SoiIjI5cVdBk5+Dpz9FshMsl7u5gk0fQZo/RLGyUPRunkT1x5lTSoF3DyMNwSU7WMZ9AXXwFn9X9IkMHeZQVdkOKlagSpqid0tZSQSCQLUEqRKfYGgWjbmospXa6UOKBf92sg1MfmicuN22m2M3jEadzPumsv+0+o/GNVklBOjIiIil5WjBS7+YmxaGP2n7XWCmwGtRwFNnwaUXgCAQKBCjrJWZqQyQOlpvJU1fU4hCZ4xYfM+8w/it8+GEML+fns5ani/vAVw4eZs5BqYfFG5cDP1JkbvHI17GffMZW+0fgMvNn7RiVEREZFLSrhqrOU68w2gSbBerlADTQYDrV8CQlsCNi7QGzRogBNn/sGqFcsx0GqUNSXGjB2PpQ4eZY1gbMIn8zInyrZE1umNrEkLcPxutv399qBEZGRkaUZKZBOTL3K6Gyk3MHrnaMRqYs1lb7V9C8MbDndiVERE5FJysoGoX421XNcP2F6namNjwtVsCODuU+QuK9ooa5WFQmGcQHnJL0vwbVjR6xer3x7RQ2LyRU51LeUaXt7xMuIy48xlM9rNwNAGQ50YFRERuYykG8DJL4DTXxuHAs9P7g40HmRMuqq1sVnLVZSKMspaZTJuwkS0Xvupa/fbowqJyRc5zZWkK3h558tIyHrQJOS9Du/hmXrPODEqIiIq9/Q5wOXfjLVcV/cCENbrBDYAWr0ERD4LqPwcHiI5V2BgIPvtUbnE5IucIioxCmN2jUFiViIAQAIJZj8yGwPrDnRyZEREVG4l3wJOfWm8pd+zXi5TAo36GwfQqNG+RLVcVHGw3x6VR0y+yOEuJV7CKztfQbI2GYAx8ZrXaR6einjKuYEREVH5Y9AD/+401nJd2QUIg/U6AXWMtVzNnwPUZTx/FrkU9tuj8obJFznU+YTzGLNzDFKzUwEAUokUH3T6AH3D+zo5MiIiKldS7wKnvjLWcqXetl4uVQAN+xlruWp1Yi0XFYr99qi8YPJFDvN33N94dderSNOlAQBkEhn+r/P/oU/tPk6OjIiIygWD3tiH68R64PLvgNBbr+NXG2g1Emg+3DjxLRGRC2HyRQ5xNu4sxu4ai3RdOgBALpFjQZcF6Fmzp5MjIyIip0u7D5z+Cjj1BZB803q5VA7Uf8JYy1W7CyCVOj5GIqJSwOSLytzp2NN4bfdryNBlAADkUjkWdVmE7jW6OzkyIiJyGoMBuL7fWMsVtR0w5Fiv41sDaPki0OJ5wCvI4SESEZU2Jl9Upk7cO4HX97yOzJxMAIBCqsDHXT9Gl+pdnBwZERE5RXoccOYb4OTnQNJ16+USGVCvj7GWK6IbIJU5PEQiorLC5IvKzNGYo5iwd4I58XKTumFJtyXoXK2zkyMjIiKHEgK4cdBYy3XxF8Cgs17HO8xYy9XyecA71PExEhE5AJMvKhOH7x7GxL0TodVrAQBKmRLLui3DI2GPODkyIiJyGE0icOZbYy1Xwr82VpAAdXsZa7nq9mQtFxFVeEy+qNT9eedPTNo7CdmGbACAu8wdy3ssR/uQ9k6OjIiIypwQwM2/jLVcF7YCuT/CWfAMBlq+YLz5Vnd8jERETsLki0rVH7f/wOR9k6HLbVKikquwssdKtAlu4+TIiIioTGUmAWd/AE6uB+Iu2VhBAkR0B1q/ZOzTJeMEt0RU+TD5olKz9+ZeTD0wFTm5I1ap5Wqsfmw1Wga1dHJkRERUJoQAbh831nKd3wTkZFmv41EVaDECaPUi4FfL4SESEZUnTL6oVOyO3o03D7yJHGFMvDwUHvjfY/9D86rNnRsYERGVvqwU4NyPxr5c9/+xvU7tLsZarvp9AbmbQ8MjIiqvmHzRQ9txYwem/zEdeqEHAHgpvPBJz0/QNLCpkyMjqhh0Oh2ioqKg0Wjg7e2NyMhIKBRsskUOJgRw95SxluufjYBOY72OOgBoPhxoNRIIiHB4iERE5R2TL3oo269txzt/vmNOvLzdvPFpr0/ROKCxkyMjcn1xcXFYuXwZPv3fSigMWQjwkGHlAiBLKDBm7DiMmzARgYGBzg6TKjptGvD3BuDEOuDeOdvr1OpsTLga9gPkSoeGR0TkSph8UYn9cvUXvHvoXRiEAQDgq/TFpz0/RcOAhk6OjMj1Xbx4Eb17dEGnoExs6S/QJlQOiUQCIQSO3dFi6S9L0GrNJ9i59w80aNDA2eFSRRRz1ljL9fdPQHa69XKVHxD5nDHpCqzn8PCIiFwRky8qVEpKCr744gskJibi4MGDmDx5Mnx8fLD53814//D7EBAAAD+lHz7r9Rnq+9d3csREri8uLg69e3TB1Mh0TGpn2bxQIpGgXTU5vq0GLD2ajt49uuDEmX9YA0alIzsD+GeTsZbr7inb69ToALR6CWjUH1C4OzY+IiIXx+SLbDp06BCeGz4Mt+/ehlQlhdxDht//+A1z/jsbwU1D4D3SG3Jv49vH390fa3qtQV2/uk6OmqhiWLl8GToFZVolXvlNaqfA0XsarFqxHO/PnuOg6KhCun/eWMt17gdAm2q9XOkDRA41DqBRla0biIhKiskXWVm9ejXGTx4H72ZeqPVibahqq8zNnTKvZSJhVwLiZsWh1hu1EBYRhrW91iLcN9zZYRNVCDqdDp/+byW29Bd2rT+plQGDVq/AO+/O5CAcVDy6TOD8FmMt1+1jttep1sZYy9V4IOCmdmh45QkHvSGi0sLkiywcOnQI4yePQ9DgIAT0rmKxTCKRQB2hhjpCjfid8bj+0XXM/2k+Ey+ih5WVauxfc/c0zh7cBfecVLQJte9Ct22YDEpdCs7ObI3WLZoCfrUB/9qAf7jxvk81QCor4ydALiUuyljLdfY7ICvZermbF9BsiLGWK7hyj1obFxeHZctXYOXqT6A1SKBQe2P2hx9DLnIw7rVXMXHCeDb5JaJiYfJFFp4bPgzezbysEq/8qvSqgqzLGkx9eQoGXx/soOiIKgBtunHEuLtngLunjbeEf82LU6/noIpKQCKR2LU7iUSCADWQGnMFcL9hvYJUAfjVzE3Kwo2Jmem+X02OTFdZ5GiBCz8ba7luHra9TmgLYy1Xk8GA0tOx8ZVDFy9eRJfuPZETWA/Kx6fBK6SeuRVIdsxlrNyyDZ98thZ/7NvNQW+IyG5MvsgsJSUFt+/eRq0Xa9u1vv/jVXDjw+tISUmBj49PGUdH5IKyNcYJaE1J1t3TxloHFNyk0FspQbxGQAj7EjAhBBI0At7KAtY16ICEK8abFYmxZsyvlmVtmSlJU3rZ9TSpHIu/ApxcD5z5FshMtF6u8ACaPm2s5Qpt4fj4yqm4uDh06d4ThiZ94d3qKYtlEokEytD6UIbWR/rJn9Gl+2P45+xp1oARkV2YfJHZkiVLIFVJoaqtsmt9VbgKMncpRgwfgqYDx0HuXwtBvmoEebsjyFuJYG93BPm4w0spt/tXfCKXpcsyDlpw95SxVivmDBB7EcidA69ISh8gtDki2zVD1sYlOH7XgLZhRTcXPHZHD63cC5FjPwHSbwGJN4DEa0DSdSAtpoitBZByy3i7cdB6sbrKg0TMnJjl3lcHADyvy6ecbODSNmMtl63jCgBBTY0JV9NnAHdvx8bnApYtX4GcwHpWiVd+nq2eQur9f7F8xUrMmT3LMcERkUtj8kVmMTExkHvIitXcSe4hQ9fsA5h66wg0N5W4LMJw2VAd+0U1RIkaiDJUQ4ZbAIK9VeakLMjHHcHe7ubkLMjbHVW9lFDIpGX8DIlKSU42EHvBskYr9gJgyLFvezcvICQSCG1urG0IbWFMaCQSKACMGSfBkl+W4Nuwone19KQUY16fBEXLodYLszVA0g1jIpZ4DUi8/uB+8q2iE0NNvPFmazAGN6/cRKy2dZNG7zBAyvPZ4RKvAyc/B858A2TEWS+Xq4xNClu/BIS1YvJcAJ1Oh5WrP4Hy8Wl2ra9s/iRWrl6Ame/O4CAcRFQkJl9kFhISgpwMfbGaO+Vk6BHoYfyyUUu0aC65hubSaxbrJQlPRKVWR1RyNVwW1XHSYPybCg/zOhIJEOChRLCPscasqrdlgma6761iLRo5mF4HxF2yTLTunwf02fZtr1DnJlot8iRaEYUmJ+MmTETrtZ9i6dG0QoebX3pUh0OxXlg6foLtFdzUQFAj483W80q+mZuMXbdMzJJuADlZhT+v7DRj37V756yXyZTG/mT5a8v8agO+NQC5W+H7JvvpdUDUb8Zarmv7bK8T2BBoPco4iIbK16HhlUdCCKRrc5Cs0SExIxtJmmzz/WRNNv45dxoavQReIfZNHO0WUg8pBilWbtiDTh3aGn9Q9FZCKedAN0RkjckXmU2ePBlz/jsbmdczoQ4veqS1zGuZMGQZ0KdVLSDnboHr+UnS0V5yEe2lFy3K7wp/XDZUxyVRHZcN1RCVUR3/pofhnzsFX5i5K6S5NWi5CZmPsdYsODdBMy1zk/NXdyoBg97YJ+vuaWOzwbungXt/F52ImMjdgeBmeRKt5kCVesUebTAwMBA79hxA7x5dcPSeBpNaGZsgmjr7H72jx7KTUhyK9cKOPQdK1tdEpgACIoy3/AwGIP2edW2ZKUnTphS+b70WiL9svOUnkeb2M6ttu0mjm4f1NmQt+SZw8gvg9FdA+n3r5TKlcXj41i8B1dtV2Foug0EgNcuUROmQrMnOTaJ0SNQYk6mkjDz3c9fR6Qvud5kZfQ3C3atYrUD0Ck/M//kUVGcf1Cb7qRXm7yRTU/yqeb6/gryVCPBUQiatmMeGqLRVlCkfmHyRmY+PD6qFVkPi9niox9cocv3E3+JRLbQ6qr570TiCW3wUcP+CsZ9L7AWI2AuQ2LooyBUqSUSoLBFdcdZcphcS3BDBuCyqIUpUR5ShOi6LarghgqGHDFk6A6ITNIhO0BQaW4CHW27tmTExe/Bll3vfxx1+agVr0Sozg8E4CEXeGq175wBd4e8tM5kbENTEskYrsAEgK52P1QYNGuDEmX+wasVyDFy9AgpDFvzVMqRoAS2UGDN2PJaOn1A2nfylUsA71Hir1clymRBAZlJuInbNukljIee8cXuDMXFIvglcP2C93DPI9siM/rUBlV+FTSLsos8B/t1prOW6shs2B24JqGus5YocCqj9HR7iw9DpDUjOTY6S8tREJebWTCXl1lIlaXTGvxnZSMnUwWDflHh2k7qpYdCkFqsViCEzFdJ886AZ49Th0r20AreVSSUI9FQiyFtpbvFhed/4v4+K31dUeVW0KR8kQohS/tii0tK4cWMAwPnz5x32mIcOHcKj3TvbnOcrr4Qd8bi/8T7+2HsQHTt2LHiHGQlA3EVjQnb/fG5idrHoX87z0QoFrohQROXWkhlry6rjLgIAlOwLyU0uNfZB83K36odm+gIM8naHu8L1m47s3bsXANC9e3cnR+IkBoMxMcibaMWcBbLT7dteKgeCGudLtBo6rPmcTqfDmjVroNFo0KVLl/L9a5823dhs0ZyY5UnSUm4bk6+SUvo86GeWv0mjZ3C57Wf20Mcv5Q5w6kvjLc1GKwOZG9DwKWMtV82O5SJBzdLpjQlTbpKUZKqJysiTPJkSqdykKi3Lzj6TpchHpYC/hxt81Qr4qd3gp3aDtxuw4KXH4N3vbShD6xe5D+3dKCT9PB9d3/sR8ZocxKfb2SS5GJRy6YN+03mSsqC8Pyp6u0Pl5vrfV2Wp0n8XuiCLKR+aPwm3fFM+aM9sgzzuskOnfHjY63PWfJGFjh07YtWy1Xh94mvI/FcD/8erQBWuMr/RM69mIvH3eKSeS8OqZasLT7wAwCMA8Ohk+eu5EEDqHXMNmbG27IKxuZdea3M3SokOjSXRaIxoIM93S5bUA7cUtfCvqI5/dKE4rQ3BJUN1JKHo0buycwy4lZiJW4mZha7nq1aYf4E0J2V5mjkG+7jDX+0GaTltOlJRquntJoTx4t/UbPDuaeDuWfsTfokMqNrQcjCMqo0BhXsZBl04hUKB+vWNF4GtW7d2Whx2UXoCwU2Mt/xysvP0M7tm3c+sqH502hTjcY05Y71M7v4gGTMnZbUf9DOTOf49b/q1dsWq/yFTp4fcTQUx70O4ySQY//rYwn+tNeiNtVsn1gP/7rCdtPpHAK1GAs2HGz9ry4AQAppsvbkpnzlpskiiLJv7JWmyocm2c5TPUiKTSuCnVsBX7Qa/PImUr4cC/rn3/TzczOv4e7jBR6UosMlfzqnXsXLLNvuSrzPbMG3SOMyZ0g2A8bslLl2L+6lZuJ+SZfybpjXeT8vCvZQsxKZqkaa1P9nU5hhwM1GDm4mF18x7ucvN31VVc5Oz/LVpgRzgilxERZ3ygTVf5Zgzar5MDh06hOEjnsOtO7cgVUkh95AhJ0MPQ5YB1UOr45uvvy068Soug954MRZ73jIxS7xa7F/LdaoqSPGsi3vu4YiW18QlfTWc1YbgZroE91KzkKV7iF/fbVDIJKjqlduuv4Bmjo7+VdJWNb1Bm+Gy1fQ2CWGsTclbo3X3NJCVbN/2EilQpb5ljVZwE0Bh33QLjlThf7E16IHUu9a1ZYnXjMPnZxfcdKtIEhngW932ACB+tYwDk5SyixcvonPX7siEEkiPhZebgL9ahkSNHmnZEsCzKlTQ4s8D+yx/rU2NAU5/DZz6wjgFQH5SOdDgSWPTwlqdi1XbZzAIpGXl5Kl1MvaHylsLlT+JSsrQIVtfup+XRXGTSy0SKD+PPMmU2lhTlfe+r9oNXkp5qf4AFhcXhyaRLaBv3BeehQw3n37yZ8jO/1qii74MbY4xMUvVIjY3KbufqsX9tCxzonY/VYvsnNJ9/Y0DXLkVUIP24H55/lGxpCr852gFM/O997Fyy0F4PzG1yHVTty/C+IGPOmTKh4e9PmfyVY45M/kySUlJwcSJE5GYmIjWrVtj8uTJjp9QWZdl7Lgfm1tDFnvRmJSl3i7+vnxrQlRtCK1/QyR6ROCusjZuIBQxaXrcSzX+Qnkv98swPl2L0j47vE2/SuZN0HzcEZRn0JDS6IBdHqvpH5oQxnmr7p42zqNlSrQ08XbuQAJUqWtMsEKa5yZaTY01NS6gUl80CAFoEvIM+pGvSaPd74ECeIXk6VtWyzJJU/kVe3dxcXGo36gJtBmp6BMhwfT2UrQJlZrPwWN39FhwVOD3qwLunt649M85BKb+bazlivrN9hQAvjWNtVwtRgCeVaE3CHPfKFNNlGmQiSRNNpLzDTKRlJGN5Ewd9KXdQaoIajeZVQKVt/bJ1Nwv7321m/1TnpSlS5cuoUv3x6CrYuNz9G4UtGd/hSL+Mg7sLbvPUSEEkjU6cyL2oCYtC/dSjEnb/dQsxKVpS73vm+lHxaq5zfODfYz3g/Mlap4uMpenSzXfJuh0OgSFVofy8Wl2N//N/n0B7t25VebHlclXBVYeki+gHF/0ZaUAsZfyJWXngczE4u1HKjd2Uq/aEKiaOyx31YbQeddAXLrOmIylPEjK7qeafqE0lpV28xqZVIKqXkrkHSHL1txonkrbrYbj4uLQuFkLGJqU3S+2DpF2P1/TwdNFD+aQl394vhqtZi49mWy5PQ/Lg6zUfP3MTE0abxhrRm0NTGEvlV/BA4B4BtnsXzVl6hv4ZNUyzO0ix5T2BV8ELD6iwwd/ZuPQ2CDU97Tuf2iADJd9O2G/dz8clzRDgibHnEylZOpK/pxKyNtdDr/cmib/3ETJmESZmvtZ11K5ep/ZuLg4LF+xEitW/S+3BYEXDFoNFMjBuNfGYsL4ceXi8zNHb0BCRvaD76c0LWLz3DfVpCVrSv99o3aT2eyDFpSn2WOgl9Jp74VK0QrEyQwGgUyd3njLNv7VZJvu5yAz2wBNdg6yTOX51zP9n62HRqdHVrYe96+fx/mv5iB0zGd2D3yTtH4s9m7fUubN85l8VWBMvkpACCA99kEyZm7CeAnQZRRvXwq1cfS6qsZkzJiUNbK44DLNF2P8wtM+qD0z/TqZm6CVxa+Snkq5RTNH05fcji+X47dDp+DT940i9+HIavpCZcRb1mbdPW17YIGC+Na0TLRCIivcfEYudR6WJ7osYz8zWwOAJEUDhoe4GFWorfqXZXpUQ3ingXgkVGDj00UPyPLcRg3qB8jwfleluey2qILvc7rhR31XxKL4NW9FkUoAX1OzvdwEyi9PEz4/tSK3f9SD+74qBeSVuJ9QRak1ydLpEWtq2pj7XRWbprW4fy8lC5m60u+zZ+o/XdX7QWuPvPeDvN0R4OFWqu+zCtkKpAR0eoNlgmOVANlIjPIkQrbXNyZUmTp9qXflAIDM6LNIPvA5Ql742O5t0r5/Ez+tW1Hm35MccIMoL4kE8Aoy3iK6PSg3GIDk6Ad9yUzJWfxlwFBAp2edBrh7ynjLS+VnHIChakNIqjaEV9VG8KraEHWqFjw6ZI7egPj0bHNyZvqiy5usFbcDdro2B+lxObga9yCpFPoc3PnmcwQOnGHXPpTNn8TSlR9iwMjxCPRRw0etgFdZNyHJTMqXaJ0BUm7av713NcvBMEJbuNyQ2uRACncgsJ7xlp9Bb6wZsxoAJPdW1A82Ok3uDzwPvoDlegGJLgvTO9jXj2xSOzcM+jET0zspcVDSEt/oe+APQyQMsO8CVCGT5NZE5RmxzyPPoBP5BpnwUyvg7a6ocH15yppLDXpTCHeFDDUC1KgRUPD7UwiBNG0OYnNbe9zLrTWLzXf/fmoWcorxq6JxGoHCh96XSoDA3JYfVb3cEeyTOyKxudWH8X9fO6aKcZXBGoQQ0OYYLGqCLBOhHMuapALXy1+eY75f2Lx25VVJpnzQaVLh7V3+W7gw+aLKQSp98Ot0gycelOdkGwf0yDsMfux5Y3OlgmQmAdF/Gm95eYc9aLpoqi0LrA8oVJDLpMY+XT6Fj5iXoc0psJmjKVGLTdMW2G8jO/Y6JDIF3EJsXGja4BZSD/E5EvSd9TWUIXUBGJs9+qoU8DE1K8q976syXcQp4GPqt6Fyy/2/gKQtK8U4pHveGq3CXtv8vELy1Wg1BzzZPISKTwiBjGw9kjKMA0ok5vaTSszQIzEjGEkafyRmNDNOyJuZjSS9FjJtHKrjPmpK7qOm1Pi3luQ+akjuw19ie5qCs/cNcJcDbULtS57ahskglUgQeWoAJK2Gwt/LDQ3y95GykUyZ7nuUk/5RVHFIJBJ4uxuT9DpVvQpcz2AQSNRkm39QNH1f5b9fnKH3DQK522oBFDxCbf6pYoLyDHhlGvzq0yXLkBNYzyrxys+z1VNIvf8vlq9YabMVSGk3qTMlU3kTKQd3xSxzSrkUajcZVAoZVG7Gm1ohh7ubDOo8ZSqFDGo3Gdxz/5rWV7vJoUBL9N/2AbJjLtvV5ys75jIUyEFkZKQDnuHDYfJFlZvcLTdhamhZbmPSaMReKLzPUeod4+3K7gdlEqmxf0jVhubaMlRtZCyzMRmvh1KOiEBPRAQWPAiE3iCQkKHF/dxmjqZk7X5qFs4cvYE/PXzsvhiTSCSQqrxhyH4wfLFx/9lIyMgGYH9TTW9pFtq630ZLxQ00lVxD3ZwrCM4pxqAoHoHWiZZ3iP3bU6WizdEbE6XcOaJMfxPSLf9PzDAONpGoyS7BqHG+iIUvTor6QL5NvZGBGpL7qCmJNSZnuQnaTc1VVFFnFusc9HCT4PUn2mLi+MeLGR+R80ilElTxVKKKpxKNQwseiCvv0PsW/dDyJWrFmeetqKlihD4Hd/63qlitQD5cMh8n/LpAq5dYJE/aUh5t0tkkElglQCo3OVQKKdRu8gcJkykhskqS5FC5SaFSyHMTJZnFNiqFrNRq1ieOe61YUz6Mf22sSzQHZvJFZIvSEwhrZbzlVdxJo4UBSLhivF385UG5TGlsBmWqITMlZj7Vipwg1Tggh/HXvaaw/MI7UVuH7l9mWFTTC30OsmOvw5CtgdRNDbeqtSHJTfyEEBBZafDy8kJxpjd1hxaNJNFoJr2GptJraCa5jgjJXUgNArA9VZuFROGJvw3hOCfCcV6EI1pZH1pJEPxSlfDVKeATp4Dv5ST4qdOLV9NWQVX0udoejNxnTJbyJlSJGdnm5OnBXx3Si9FE92HJpRL4eRib9vl5GPtG+Xu4wV8daSzP7R/l6eEG1YXTuLm1R7GayiRnCbRt3arIdYlckZtcijBfFcJ8C5/CQ5Odk6/2LLfZY27SZrpvz48oJWkFYpDIcPr0WXMrEGdRyCT5aoLyJEaF1BZZJkLyAtdTyqUu8905ccJ4fLpmLdJP/lzkAGKK+MuYMP4HB0ZXcky+iIqjlCaNhl4L3PvbeMtL6f2gJi5vYmbnBKqRkZGQixxkx1yG3DcYGad+RvbZX+Ep1yNALUWixoCEHBncIvvCo+VTyEm+By+FwMWVYyGVyZGSaZznJ9n0V6NDWlo6FAkX4J34N6qkXkBY5iWE6m5Clr8qoAApQo1zhnD8LcLxt6E2/hbhuC2qAMjz4a8BoNEA8YVPIJpXqTaPLKdsjdI1+8OPy/UoXab+IqbmfXlrnxJsJlLG95sjh34yDTRhGlTC30MBfw8l/D0eDHtuSrb8Pd2K9Z6pX7UjsuGG43cNaBtW9Ohux+7okS1RolUrJl9Uuand5KhdRY7aVTwKXEcIgZRMnTkRM9emmeZLy71/M1oDqdr7oVqBFMTUpE7tJod7AbVFhTevK6B2Kfd/ToD9QGBgIA7s3Y0u3R9D6v1/i5zyobx9HxaEyRfRw5JIjDVWPtWAuj0flOtzjJ338w6DH3ux8EmjtanAraPGW14eVXNHXMzTdDGwgdUcVQqFAuNeexXLvv8B+vtR6F0jG9OHStEmVJ77YSXNnWPoZ+z44nfIgupjUp5qen+lgH/yZeD+6QfDvMdeLHhQknz0Ck+k+jVGnFcj3FE1wDW3eog2BCI5M8ec0Ek1OnhrspFajCYmNh+rhM0jC0raTPdtJW2+aoXD57KxGKXr8WnwyjdK18ot2/DJZ2vLfJSuLJ3+QQ2UJk9NVJ5aqPxN/xzZudvDTWZR+2SulcqTWOUt9ynjUfsUCgVGv/oaPty+AhuHFD1Z94LDOrz86oQKVZNJVFYkEknuSJ1uqB9ccH+0o8d88NjulcWqgZbr0vH2Uy3RrEULhzSpI/s0aNAA/5w9nTvlwwKk5pvyYfxrYzFh/A8uk3gBHGq+XONQ8xVUKU8a/WBuMmNN2cU4PVq3aoG53dwwpX3BQ10vPpKN2Qeyceb7eagtvW9MtO6ft3/obYWHcUj3vP20/MONg5vYQW8QNmvakjU6y/8tlj180lZSjkzaymquthy9wTwpb97mfInpeWujdBa1VqU9j11hFDKJRRJlrn3ysPzf1PTPT+1WLueRiouLQ4smDTCleQamdFAWuN7iv7RYfMYDp/+55FIXDpUVvwtdR3meoJdKrrxM+cCh5olcjcIdCGlmvOVlnjQ6T1+yoiaNTo423i7/Zi76YX82+teTFZp4AcCU9m44cScHX34822KOIZvkKiC4qWWiVaUuIC35ha9MKjFfVBeHs5I2R9a0rfh4qd2jdKXcv4x3/rsQz46ZYkyk8tRMPaiNMtZOOXJiXokE5nmi8idUATYSLD8Px9culpXAwEDsPnAIPbt1xpG76ZjaToq2YTJzzeXRO3osPmrAX/Fe2H3gIBMvolJmagVSEQdrqMwqypQPTL6Iygt3H6BGO+PNxGLS6Dw1ZQVMGq3TC3x6UostQ4s3x9A7nd2gkOVe9MrcbCRa9W2OzugMFT1pM47StdruUbrcm/fD+rX/xU7FI+aBVMqCp1KeJ1lS5EmacpOpfE3/fFQKyCpx85wGDRrg1LkLWLViOQauWg65PhN+KilSsiXQSZQY89p4rBw/gYkXURmpqIM1kOsrH1dTRGRbMSeNPnv2QrHnGFLKpTjr9wRad+ljTLQCGxqH4K9gSpq05egNSM3KsZ20mctKL2kryShdEpkC2bHX7R6ly00utdGcT2FOoPzzJFb+HsbaOKW8/DXvK+8CAwPx/uw5eOfdmeWiqQxRZVJRB2sg1+fSyVdmZibmz5+P77//Hjdv3oS/vz/69OmDuXPnIiwsrFj7SkpKwqxZs7Blyxbcu3cPwcHBGDhwIGbNmgVfX1+b21y+fBnz58/H3r17ERMTA6VSiQYNGuC5557DuHHj4OZW8S5gqZwoYNLo1F07UGXnM8Ua4SnA1wupzccArdiPwRa5TOrQpO1+dvFH6ZKpvBGiFqhT2z9P0qQoYAAKN6g5Ma9DVZSmMkSupiIO1kCuz2WTr6ysLHTv3h1HjhxBSEgI+vfvjxs3bmD9+vXYtm0bjhw5gvDwcLv2FR8fjw4dOuDKlSsIDw/HgAEDcP78eSxduhS//fYb/vrrL/j7+1tsc/jwYfTs2RMajQYNGzbEgAEDkJKSgoMHD2LKlCnYunUrdu/eDbncZV9ickHefgGI1wgIAbtHeErIyIG3t7cDoqtcSpq0HTnqh557ijdKl5teg09Hd+aFPRFRPoGBgZgzexZmvjuDNdBULrjsZALz5s3DkSNH0KFDB1y+fBk//PADjh49ikWLFiEuLg6jRo2ye1+TJ0/GlStXMGjQIERFReGHH37AP//8gwkTJuDy5cuYMmWK1Tbjx4+HRqPB/PnzceHCBfz444/YsWOHOYE7cOAAvvrqq9J8ykRFioyMRJZQ4Phd++bgOnZHDy0UiIyMLOPIyF6tWraAQuiRHXPZrvWzYy5DgRweQyKiQphqoFu0aIHWrVsz8SKnccnkKzs7GytWrAAArFy5Ep6eD+Y6mjJlCpo1a4YDBw7g5MmTRe4rJiYG3333Hdzc3LBq1SqLmqqPPvoIgYGB+PrrrxEbG2suT09Px+nTp6FWqzFt2jSL/YWGhmL8+PEAgOPHjz/U8yQqLoVCgTFjx2HJCfualC09KcWYseP5JVSOmEbp0p7ZZtf62jPbMI6jdBEREbkEl0y+Dh06hJSUFERERKBFixZWy59++mkAwC+//FLkvn7//XcYDAZ07twZQUFBFsuUSiX69esHvV6P7du3m8sVCgWkdsxlFBAQUOQ6RKVt3ISJOBSrxtKjhQ8rvvSoDodi1Xh9/AQHRUb2mjhhPBTxl5F+8udC13swStc4B0VGRERED8Mlk6+zZ88CAFq2bGlzuan83LlzZbIvpVKJRx99FBqNBgsWLLBY/+7du1i5ciUUCgWef/75Ih+fqLQFBgZix54DWHzOC89tNeDo7RyY5lIXQuDI7Rw8t9WAxee8sGPPAXY0LodMo3TJzv+K1O2LoL0bZXEMtXcuIXX7IsjO/8pRuoiIiFyIS44GcfPmTQBAtWrVbC43lUdHR5fZvv73v/+hZ8+eePvtt/Hll1+iSZMmSE1NxR9//IGQkBD8+uuvqFfPvqGiTTNl53f16lUEBwdj7969du2nrGRkGOcmcnYcVDwfL1+Nn7duRr+NW6CS5sBfLUVipgFZBjke79cfH/cfiLt37+Lu3bvODpUKsHrFMmzZuhVbfv4vUqUKyFVeyMlMh0zoMKBfXwzoP57H0EXwc9T18Ri6Ph5D11cejmFGRgY8PDxKvL1LJl/p6ekAALXa9kSyphckLS2tzPZVv359/Pnnnxg4cCBOnTqFixcvAjCOMNetW7cCEyoiR/H19cULL76E54Y/j3/++QcajQYBAQGIiIjgKJwuwtfXFyNffBEjhg/nMSQiIqoA+O1dQnv37sXgwYNRvXp17N27F61bt0ZCQgLWrl2LDz74AHv27MGxY8fsag50/vx5m+WmBK57d+fOv2T6dcHZcVDJmS7UeQxdF4+ha+PnqOvjMXR9PIaurzwcw4ep9QJctM+XaXRDjUZjc7mpStLLy6tM9pWYmIhnnnkGOp0Ov/32G7p16wYvLy/UqlULc+fOxbhx43Djxg0sXLjQ/idFREREREQVmksmXzVq1AAA3L592+ZyU3nNmjXLZF+//vorEhMT0b59e4SFhVlt88wzzwAA/vjjjyIfn4iIiIiIKgeXTL5Mk4meOnXK5nJTebNmzcpkX6aEzMfHx+Y2pvKkpKQiH5+IiIiIiCoHl0y+OnbsCB8fH1y9ehVnzpyxWr5hwwYAQL9+/YrcV58+fSCVSnHw4EGLiZQBQKvV4pdffoFMJsMTTzxhLg8ODgYAnD59Gnq93mqfpsmVa9WqZe9TIiIiIiKiCs4lky83NzeMHz8eADBu3DhzvywAWLx4Mc6dO4cuXbqgVatW5vIVK1agQYMGePvtty32FRISgmHDhiE7Oxuvv/46cnJyzMumTZuGuLg4jBgxAlWrVjWX9+nTB0qlEtevX8fMmTNhMBjMy6KiovDee+8BeDDZMxERERERkcuOdvjuu+9i9+7dOHz4MOrWrYvOnTsjOjoaR48eRWBgINatW2exfnx8PKKiohATE2O1ryVLluDIkSPYuHEjGjRogNatW+P8+fP4559/ULduXSxevNhi/ZCQECxcuBATJ07E/Pnz8cMPP6BFixZISEjAX3/9Ba1WiyeeeAIjR44sy5eAiIiIiIhciEvWfAGAu7s79u3bh5kzZ0KtVmPLli2Ijo7GyJEjcerUKYSHh9u9rypVquDYsWOYMGECsrOzsXnzZqSkpGDixIk4duwY/P39rbYZP3489u7diwEDBkCj0WDr1q04deoUWrRogZUrV+Lnn3/mPDxERERERGTm0tmBSqXCnDlzMGfOnCLXnTVrFmbNmlXgcn9/fyxbtgzLli2z+/G7du2Krl272r0+ERERERFVXi5b80VERERERORKmHwRERERERE5AJMvIiIiIiIiB2DyRURERERE5ABMvoiIiIiIiByAyRcREREREZEDMPkiIiIiIiJyAIkQQjg7CLLNy8sLOp0OERERTo0jIyMDAODh4eHUOKjkeAxdH4+ha+Pxc308hq6Px9D1lYdjePXqVSgUCqSlpZVoe9Z8lWMeHh5QKBTODgP37t3DvXv3nB0GPQQeQ9fHY+jaePxcH4+h6+MxdH3l4RgqFIqHSv5Y80VFaty4MQDg/PnzTo6ESorH0PXxGLo2Hj/Xx2Po+ngMXV9FOIas+SIiIiIiInIAJl9EREREREQOwOSLiIiIiIjIAZh8EREREREROQCTLyIiIiIiIgfgaIdEREREREQOwJovIiIiIiIiB2DyRURERERE5ABMvoiIiIiIiByAyRcREREREZEDMPkiIiIiIiJyACZfREREREREDsDki4iIiIiIyAGYfFGBMjMz8d5776FevXpwd3dHaGgoRo0ahTt37jg7NLJD165dIZFICrz9/vvvzg6RAJw8eRL/93//h0GDBqFatWrm41OUzz//HG3btoWnpyf8/f3xxBNP4PDhww6ImPIq7vGbNWtWoeflW2+95cDoSaPRYMuWLRg9ejTq168Pd3d3eHh4IDIyEnPmzEF6enqB2/IcLB9Kcgx5HpY/ixcvxqBBg1C3bl34+PhAqVSiZs2aeOGFF/D3338XuJ0rnoecZJlsysrKQrdu3XDkyBGEhISgc+fOuHHjBo4dO4bAwEAcOXIE4eHhzg6TCtG1a1ccOHAAgwcPhqenp9XyqVOnomnTpk6IjPIaMGAAtm7dalVe2Efz5MmTsXTpUqhUKvTq1QtZWVnYs2cPhBDYsGEDBgwYUIYRU17FPX6zZs3C7Nmz0bFjR9SpU8dqed++ffHMM8+Uepxk25o1a/DKK68AABo2bIgmTZogNTUVhw8fRlpaGho0aIADBw6gatWqFtvxHCw/SnIMeR6WP1WqVEFGRgaaNWuGsLAwAMD58+dx+fJlKBQKbNq0CU8++aTFNi57HgoiG2bMmCEAiA4dOoi0tDRz+aJFiwQA0aVLF+cFR3bp0qWLACCuX7/u7FCoEP/3f/8nZs6cKX7++WcRExMjlEqlKOyjedeuXQKACAgIEJcvXzaXHz58WLi5uQlfX1+RlJTkgMhJiOIfv/fff18AEOvXr3dckFSgzz//XIwZM0ZcuHDBovzu3buiRYsWAoAYNmyYxTKeg+VLSY4hz8Py588//xSZmZlW5StXrhQARFBQkNDpdOZyVz4PmXyRFa1WK3x8fAQAcerUKavlzZo1EwDEiRMnnBAd2YvJl2sq6uL98ccfFwDExx9/bLVs4sSJAoBYuHBhGUZIhWHyVXEcPnxYABBKpVJotVpzOc9B11HQMeR56FoiIiIEAHH27FlzmSufh+zzRVYOHTqElJQUREREoEWLFlbLn376aQDAL7/84ujQiCq1zMxM7N27F8CD8zAvnptEpScyMhIAoNVqkZCQAIDnoKuxdQzJ9SgUCgCAm5sbANc/D+XODoDKn7NnzwIAWrZsaXO5qfzcuXMOi4lKbu3atUhISIBUKkW9evUwYMAA1KhRw9lhUQlERUVBq9UiMDAQ1apVs1rOc9N17N27F2fOnEFWVhaqVauGxx9/HK1atXJ2WJTHtWvXABgv/Pz9/QHwHHQ1to5hXjwPy7+vvvoKUVFRqFu3LurWrQvA9c9DJl9k5ebNmwBg8w2dtzw6OtphMVHJzZs3z+L/N954AzNnzsTMmTOdFBGVVFHnpoeHB3x9fZGUlIS0tDR4eXk5Mjwqhq+++sri/5kzZ2Lw4MH4/PPPbQ6QQ463dOlSAECfPn2gVCoB8Bx0NbaOYV48D8ufjz76COfPn0dGRgYuXryI8+fPIzQ0FN999x1kMhkA1z8P2eyQrJiGZVWr1TaXe3h4AADS0tIcFhMV36OPPoqvvvoKV69ehUajQVRUFP773/9CLpfjvffeM38pkeso6twEeH6Wd3Xq1MHChQtx/vx5pKen49atW/jmm28QFhaGjRs34vnnn3d2iARg+/btWLt2LRQKBebOnWsu5znoOgo6hgDPw/Jsx44d+OKLL7BhwwacP38eNWvWxHfffWdRI+nq5yGTL6IKas6cORgxYgTCw8OhUqlQr149vPPOO9iyZQsA41C7mZmZzg2SqJIZMWIEpk6dikaNGsHDwwPVqlXDc889h+PHjyMgIABbtmzBkSNHnB1mpXbp0iWMGDECQgh89NFH5n5D5DqKOoY8D8uv3bt3QwiBpKQk/PHHH6hbty66dOmC//73v84OrdQw+SIrpqp2jUZjc3lGRgYAlLtqXLJPr1690Lp1ayQnJ+Po0aPODoeKoahzE+D56apCQkLw0ksvAQAnQHeiO3fuoE+fPkhKSsKUKVMwadIki+U8B8u/oo5hYXgelh++vr7o3Lkztm/fjlatWmHmzJk4fvw4ANc/D5l8kRXTYAy3b9+2udxUXrNmTYfFRKXL1Gk1JibGyZFQcRR1bmZkZCA5ORl+fn7l8guHCsfz0rkSExPRq1cvREdH46WXXsLChQut1uE5WL7ZcwyLwvOwfFEoFHj22WchhDCPXujq5yGTL7Jiqp4/deqUzeWm8mbNmjksJipdSUlJAB60iSbXUL9+fSiVSsTFxeHOnTtWy3luujael86Tnp6Oxx9/HBcuXMCgQYPw2WefQSKRWK3Hc7D8svcYFoXnYflTpUoVAEBcXBwA1z8PmXyRlY4dO8LHxwdXr17FmTNnrJZv2LABANCvXz8HR0alIS4uDgcPHgRQ8HQCVD6pVCp0794dAPDTTz9ZLee56bqEENi8eTMAnpeOptVq0b9/fxw7dgy9e/e2GFUtP56D5VNxjmFheB6WTwcOHAAAREREAKgA56EzZ3im8mvGjBkCgHjkkUdEenq6uXzRokUCgOjSpYvzgqMiHTp0SGzevFnk5ORYlF+/fl107NhRABBPPfWUk6KjwiiVSlHYR/OuXbsEABEQECAuX75sLj98+LBQKpXC19dXJCUlOSBSsqWw4xcbGytWrFghUlNTLcrT0tLEq6++KgCI4OBgkZGR4YhQSQiRk5MjBg4cKACIzp072/Xa8xwsX4p7DHkelj9//vmn+O2334Rer7coz87OFsuWLRNSqVSoVCpx8+ZN8zJXPg8lQgjhnLSPyrOsrCx07doVR48eRUhICDp37ozo6GgcPXoUgYGBOHLkCMLDw50dJhXg888/x0svvYTg4GC0bNkSvr6+iI6OxsmTJ5GVlYXGjRtj7969qFq1qrNDrfR+/fVXi2GQjx07BiEE2rVrZy6bOXMm+vbta/5/8uTJWLp0KdRqNXr27Ins7Gzs2rULQghs2LABAwYMcORTqNSKc/xu3LiB2rVrw9PTE23atEFISAji4uJw6tQpJCQkwNfXF9u2bUPHjh2d8VQqpaVLl2Ly5MkAgIEDB8Lb29vmegsXLjQ3fQJ4DpYnxT2GPA/LH9M1S5UqVdCqVSsEBAQgPj4ef//9N2JiYuDu7o4vvvgCQ4YMsdjOZc9DJyZ+VM5pNBoxc+ZMERERIdzc3ERwcLAYOXKkuHXrlrNDoyJcuHBBvPbaa6Jly5YiMDBQyOVy4ePjI9q3by8WLVokNBqNs0OkXOvXrxcACr2tX7/e5natWrUSarVa+Pr6ij59+ohDhw45/glUcsU5fqmpqWL69OmiS5cuIiwsTCiVSqFWq0Xjxo3F1KlTxe3bt537ZCqh999/v8jjB0Bcv37dalueg+VDcY8hz8Py59q1a+Kdd94RHTt2FCEhIUKhUAgPDw/RuHFjMWHCBPHvv/8WuK0rnoes+SIiIiIiInIADrhBRERERETkAEy+iIiIiIiIHIDJFxERERERkQMw+SIiIiIiInIAJl9EREREREQOwOSLiIiIiIjIAZh8EREREREROQCTLyIiIiIiIgdg8kVEREREROQATL6IiIiIiIgcgMkXERERERGRAzD5IqIyIZFIIJFI4Ovri+TkZJvr/N///R8kEglmzZrl0NjssX//fkgkEowcOdLZoZSJZcuWoXHjxlAqlZBIJOjatWuB65qOpb23WrVqlXq8tWrVgkQiKXf7Kgv79u3D4MGDERYWBjc3N/j5+aF+/fp45plnsGLFCqSkpDg7xFIxa9YsSCQSfP75584OBcCDcz7vTS6Xo2rVqujTpw+2bt3q7BArlBs3bhT52UNUEcmdHQARVWwpKSlYvHgx5syZ4+xQKNemTZswadIk+Pn54amnnoKHhwcaNGhQ4PovvviiVdmff/6Jq1evIjIyEs2bN7dYVqVKldIOudKYM2cO3n//fQBAw4YN0a5dOygUCkRFRWHTpk3YsGEDWrdujfbt2zs50oorKCgIffr0AQBkZWXh/Pnz2LFjB3bs2IH333+/XP5YRESug8kXEZUZiUQCpVKJpUuX4j//+Q/8/PycHRIB2LJlCwBgw4YN6N69e5Hr26qZGDlyJK5evYoBAwY45GJ0z5490Ol05W5fpenkyZOYNWsWFAoFfvzxRwwYMMBi+b179/D111/D19fXKfFVFg0aNLB6zy9duhSTJ0/GvHnzMHLkyDKp3SWiyoHNDomozEilUowZMwapqalYuHChs8OhXLdv3wYAhIeHOzkS+0VERBRaO+esfZWmTZs2QQiBIUOGWCVeABAcHIw33nijXMZe0U2aNAkNGjSAXq/Hnj17nB0OEbkwJl9EVKbeeustqFQqLF++HAkJCXZt07VrV0gkEty4ccNqWUH9BPL2Hzl58iQef/xx+Pr6wt/fH0OGDDEnHBkZGZg2bRpq1aoFd3d3NGnSBBs2bCg0npiYGIwcORJBQUFQqVRo2bIlvvzyywLXT0xMxNtvv41GjRpBpVLBx8cH3bt3x7Zt2wp9PqmpqZgyZQpq164NhUKByZMnF/laAcCtW7fw6quvombNmlAqlahatSoGDRqE48eP23yN9u3bBwCoXbu2uW/L/v377Xqsonz++efmfnyXL1/G0KFDERQUBKlUaq5xu3LlCmbNmoUOHTogODgYbm5uqFatGl544QVcvnzZ5n5t9dPK+9plZmbirbfeMr8GderUwYcffgghRJnuCwAOHDiA7t27w8vLC35+fnjiiSdw4sQJi9fCHnFxcQCAwMBAu9Y3OXPmDKZNm4ZWrVohMDAQSqUS4eHheP3113H37l2r9fM+14yMDEyZMgXVq1c3v7d/+eUX87o//fQT2rVrBw8PDwQFBWHixInIzMy02qfpNRVCYOnSpWjUqBHc3d0RFhaGiRMnFtjvsyA5OTlYvXo1OnToAG9vb6hUKjRv3hxLlixBTk6O1fpxcXF466230KhRI3h6esLHxwf16tXDCy+8gGPHjhXrsQvSuHFjAEBsbKzVMo1Gg/nz56NFixbw9PSEp6cn2rdvjy+++MLmvkx9I3NycjB37lzUqVMHKpUKDRs2xPr1683r7d27F926dYO3tzf8/PzwwgsvFPg5mpCQgDfffBN169aFu7s7/P390adPH+zcudNivdjYWMjlcoSGhsJgMNjc1y+//AKJRIJBgwZZlAsh8N1336F79+7w8/ODu7s7GjZsiFmzZkGj0djc161bt/D8888jMDAQarUarVq1wtdff21zXaJKQRARlQEAQiaTCSGE+M9//iMAiOnTp1usM3/+fAFAvP/++xblXbp0EQDE9evXrfZ7/fp1AUB06dLFovz9998XAMTYsWOFUqkUrVq1EkOGDBF16tQRAES9evVEcnKyaNOmjahatap4+umnRdeuXYVEIhESiUT8/vvvFvvbt2+fACD69esnatSoIYKCgsSQIUNEz549hVwutxm3EEJERUWJ6tWrCwCiVq1aon///qJ79+5CrVYLAOKjjz6y+Xzatm0rmjdvLvz8/MSAAQPEoEGDxKxZs4p8nc+dOyeqVKkiAIj69euLoUOHikceeUQAEHK5XPz444/mdTdv3ixefPFFERQUJACIwYMHixdffFG8+OKL4uLFi0U+Vl4vvviizddg/fr1AoAYOnSo8Pb2FrVr1xbPPvus6NWrl9i2bZsQQojp06cLiUQimjZtKp588kkxePBg0bBhQwFAeHt7i7Nnz1o9Xs2aNUX+ryzTa9ehQwfRqVMn4e/vLwYNGiR69+4t3N3dBQAxY8aMMt3Xxo0bhUwmEwBE+/btxdChQ0Xjxo2FUqkUY8eOLfB9YsucOXMEAFG9enVx//59u7YRQohnn31WyOVy0bJlSzFgwAAxYMAAUatWLQFAhISEiDt37hT4XNu1a2dxPkilUiGTycSuXbvE4sWLhVwuFz169BADBw4UAQEBAoB47rnnCnxNx40bJxQKhejZs6cYMmSI+b3WrFkzkZKSYrGN6Zxdv369RblGoxHdunUTAIS/v7/o2bOn6Nevn6hataoAIJ566imh1+vN66emporatWubX7sBAwaIp59+WrRt21YoFAq7X3/TOZ//s8WkV69eAoBYs2aNRfn9+/dFs2bNBAARHBwsnnjiCfH4448LHx8fAUCMHz/eal8ARM2aNcXAgQOFj4+PGDBggOjVq5dQKpUCgFi3bp346aefhFwuF506dRJPP/20CAsLEwBEp06dhMFgsNjf7du3RXh4uAAgatSoIZ599lnRvXt383tz8eLFFuv36dNHABC7d++2+VyfffZZAUBs3LjRXKbX68WwYcMEAOHp6Sm6du0qBg4caP68a9u2rdBoNBb7uXbtmggODhYARHh4uBg6dKjo3LmzkEgkYvz48YW+3kQVFZMvIioTeZOve/fuCbVaLTw8PERsbKx5nbJIvgCI1atXm8uzs7PFY489JgCIRo0aie7du4v09HTz8jVr1ggA4tFHH7XYn+lCDIDo2bOnxTbHjh0Tnp6eQiqVipMnT5rLc3JyRNOmTQUAsWDBAosLxH///VfUrl1byGQy8ffff1s9H9PFcFJSUsEvaj4Gg8H8eNOmTbO4INuwYYOQSqXC09NT3L1712K7wl5fexWVfJkuOnNycqy2/euvv8S1a9esytetWycAiG7dulktKyxhMr0f8l7cHz9+XMhkMqFWq0VaWlqZ7CslJUX4+/sLAOKbb76x2N/MmTPN+7P34v/q1atCpVIJAMLLy0u8+OKL4rPPPhOnTp2y+Tqa7N27V9y7d8+iTK/Xi9mzZwsA4qWXXirwueY/H0zHr06dOsLPz08cP37cvOzOnTvmBOjq1asW+zS9pt7e3uLEiRPm8rS0NNG9e3cBQEyaNMlim4KSr9dff10AEM8++6xITk42l6empoonnnjC6hw3vW/yJ2VCCBEbG2txvhWmsOQrISFB+Pr6CplMJm7cuGGxzBTTpEmTRFZWlrn83r17onXr1gKA+O233yy2Mb3+TZo0sfhM3Lt3rzlpDggIMP9gIYTx/da4cWMBQOzdu9dif08++aQ5MdZqtebygwcPCrVaLWQymTh9+rS5/KuvvhIAxKhRo6yea2pqqlCpVMLHx8fi+SxYsEAAEF27dhUxMTHmcq1WK0aPHm3zBzZTkjdq1Cih0+nM5T///LM5MWTyRZUNky8iKhN5ky8hhJg6daoAIKZOnWouK4vkq1OnTlbbbN26VQAQUqlUREVFWSzLyckRVapUEQqFQmRnZ5vLTRdiUqlUXLp0yWqf06dPFwDE6NGjzWWbN2821yjZsmnTJgFATJw40er5ALC40LWH6UKtRo0aFrGbDBo0SAAQ8+bNsyh3RPIVGBgoMjIyir3fjh07ColEYnHRLUThCVNBx8h0Qbpv374y2ddnn30mAIgePXpYra/T6cyPY2/yJYQQu3fvNtck5L35+vqK1157zSqRLkpYWJgICAiwKMv7XPOfD3q93lyT+u6771rtz1SLnT9hMj3Xd955x2qb8+fPC4lEIjw9PUVmZqa53Fbydf/+faFQKET16tWtalGEECImJka4ubmJZs2amcs+/PBDAUAsWbKk0NeiKLaSr8zMTHH8+HHx6KOPmn9Uyev06dMCgGjTpo1V4ieEEKdOnTInhnmZjqutmqcWLVoIAGLEiBFWy5YuXWr1nrp69aq5NiohIcFqmylTpggA4uWXXzaXpaenCw8PD6sESwghvvjiC6vPNp1OJ6pUqSI8PDysEn0hjLWVwcHBws/Pz/w6mOLy9va2Op+FeFC7xuSLKhv2+SIih5g+fTo8PDywevVq3L9/v8wep1evXlZlpoElatWqhXr16lksk8lkqFmzJnQ6HeLj4622bd68OerXr29VPmzYMADAwYMHzWWmvhX5+0mYdO7cGQBs9kEJCQlB69atbW5XENNjDxkyBAqFwmr5888/bxWjozz22GNQq9UFLk9PT8d3332H6dOn45VXXsHIkSMxcuRIxMTEQAiBq1ev2v1YNWvWtHmMTMc6JiamTPZ16NAhAMAzzzxjtb5cLsfgwYPtflyTHj164MqVK9i0aRPGjh2Lli1bQi6XIzk5GatXr0bz5s0RFRVltV1CQgLWr1+PqVOnYvTo0ebXU6fTISEhAYmJiVbb2DofpFIpatasCaDwc6mg13To0KFWZY0aNUJkZCTS09Nx+vTpQp///v37odPp0KdPH6hUKqvlwcHBqFu3Lv7++29z37NWrVoBAD766CN8//33SEtLK/QxinLgwAFzX0iVSoU2bdrg4MGD+PLLL/Hmm29arGs65wcMGACp1PqSytQHzNY5r1AobM5xZXqN7X39//zzTwBAnz594O/vb7WNrc8BDw8P9O/fHykpKfj1118t1v/mm28AACNGjDCXnTp1CvHx8XjkkUcQFBRk9RgqlQqtWrVCUlIS/v33X6u4fHx8rLYxfYYSVTYcap6IHCIwMBDjxo3DggUL8H//93/4+OOPy+RxwsLCrMo8PT0LXJZ3uVartVpmuhDNzzTUdN4BDUwDhAwfPhzDhw8vMEZbSV6NGjUKXL8gpscuaNhrU/mdO3eKve+HVdjz2bt3L4YOHWoeYMKW4lxAV6tWzWa5l5cXANvHtTT2ZboArl69us1tSnJMAcDNzQ0DBw7EwIEDAQDJycn4/vvv8c477yA2Nhbjx4/Hrl27zOt/9913GDNmDNLT0wvcZ1pamtWFeVHnQ2HnUkGvaWHny5kzZ2wOAJKX6Rz67LPP8NlnnxW6bmJiIsLCwtCjRw/85z//wZIlSzBs2DDI5XK0bNkSPXv2xKhRo4o9qmfeeb5SUlLw119/4f79+xg/fjwiIyPRrFkzq3hnzJiBGTNmFLjPrKwsq7Lg4GDIZDKr8uK+/iX9HBg+fDi+/fZbfPPNN+YfjO7fv489e/agWrVqePTRR83rmp7nrl27ipygPD4+HvXr1zfHVdRnKFFlw+SLiBzmzTffxKpVq/C///0P06ZNK9E+Chqdy8TWr8/2LCsNptj69Olj89dhE1uTELu7u5d6PEVdJJWlgp5Peno6hgwZgsTERLz33nsYOnQoatasCZVKBYlEgueeew7fffddgSML2lKax7Ws3yMl4evri7FjxyI0NBT9+/fHvn37oNFooFarER0djZEjRwIAlixZgr59+yIsLMxca/TII4/gr7/+svl6FvVcnfFamM6h5s2bIzIystB1lUql+f7ixYvx6quvYuvWrdi9ezcOHTqEY8eOYcGCBfjuu++KVQuZf54vrVaLYcOGYfPmzXj++edx+vRp82tjirdTp06IiIiw+zEAx73+BX0O9OrVC4GBgfj111+RkpICHx8ffP/999Dr9Rg2bJjF45ueZ506ddCxY8dCHy8gIKBU4iaqqJh8EZHDVKlSBRMmTMD8+fMxf/58hIaG2lzPzc0NAGz+kn/r1q0yjTG/6OjoQsvzPgdTrcnLL79coiZnxWV67IJiNP1aXVANhzMcPHgQCQkJePrppzF79myr5deuXXNCVCUTEhICoOD3ZGm/V00TYuv1eiQnJ0OtVmP79u3Izs7GG2+8gUmTJllt4+jXMzo6Gk2bNrVZDqDAc97EdA516tQJy5cvL9Zj169fH9OmTcO0adOQlZWFFStW4M0338Rrr732UOejUqnE+vXrcejQIZw7dw5ff/01XnjhBYt4BwwYgKlTp5b4MR5GST8H5HI5nn32WaxYsQIbN27EqFGjbDY5BB48T1sTUBfEdH4U9RlKVNmUv5/4iKhCmzp1Kry8vPDpp58W2BzO9KVta86nvM2tHOHMmTPmPgx5ff/99wCMF4kmPXv2BABs3rzZIbGZ+pD99NNP0Ov1VstNc+mY1isPkpKSANhu3nflyhWcOnXK0SGVmKkGYOPGjVbL9Ho9Nm3aVKz9FVXbd+XKFQDGHydMtaeFvZ5//PFHmfavtOXHH3+0Krt06RLOnDkDT09PNG/evNDtu3XrBplMhm3btkGn05U4Dnd3d7zxxhsICQlBXFyczbm5isPHxwdvv/02AGD+/PnmmiBHn/O2mD6Dfv/9d5vzqRX2OWBqHv3tt9/i33//xfHjx9GkSROLppUA0KZNG/j4+ODAgQM2+w8WFVdqaqrVctNnKFFlw+SLiBwqICAAEydOhFarxdq1a22u06VLFwDAokWLLCbu3Lt3L5YsWeKIMM0MBgMmTJhgEcfJkyexYsUKSCQSvPbaa+bywYMHo1GjRvjmm28wd+5cq34xQggcOnTIPFDDw+ratSuaNm2KGzdu4L333rO4eN+8eTM2bdoET09PjBo1qlQerzSYBnjYtGmTRZ+v5ORkjB49+qEuuB3tmWeegb+/P3bt2mV1ITlv3jxcv369WPubOXMm3nzzTZuDjdy5cwevvvoqAOCpp54y1w6bXs+vv/4aGRkZFuuPHTu2WI9fGpYvX24xqIZGo8GECRMghMBLL71kcxCNvMLCwjBq1CjcuHEDw4YNs5k8XrlyxSLh3bJlC44cOWK13smTJ3H//n14enrC19e35E8q16uvvoqQkBBcunTJ/Pjt2rVDz549cejQIYwbN85mknH27Fn8/vvvD/34BQkPD0ffvn2RlpaGSZMmWZxDf/31F1avXg2ZTIZx48ZZbdu+fXtERERg3759WLhwIQDY7K+qVCoxbdo0pKWlYdCgQTZrVO/cuYOvvvrK/H9ERAR69eqF1NRUTJ061eIHou3bt+Onn356qOdN5KqYfBGRw02dOhXe3t7m0cryGzZsGOrXr4/Dhw+jYcOGePrpp9G+fXv07NnTItlxhCeffBIXLlxAREQEnn32WfTp0wcdOnRAWloaZsyYYTFCoVwux5YtW1C7dm289957qFGjBnr27Inhw4ejd+/eCA4ORqdOnXD8+PFSiU0ikeCbb75BQEAAPvjgAzRu3BjPPfccOnXqhEGDBkEqlWLt2rXmmsTyoHXr1ujZsydu3ryJevXqmQeWqF27Nu7evYv+/fs7O0S7+fj44LPPPoNMJsOwYcPwyCOP4LnnnkPTpk3xwQcfYMyYMQAeNKMtSnp6OhYuXIg6deqgfv36GDhwIIYNG4bOnTujdu3aOHbsGOrUqWPxA8RTTz2Fxo0b48SJE6hTpw6efvppPPnkk6hXrx78/PzwyCOPlMVTL9CIESPQrl079OnTB88++ywiIiKwe/duNG7cGHPnzrVrH0uXLkXPnj2xceNGREREoFOnTnjuuefQv39/1K1bF3Xr1rW4yN+/fz86dOiAatWqoV+/fhg+fDi6deuGdu3awWAwYPbs2XYfg8KoVCq89dZbAIy1XyZff/01WrRogVWrVqFmzZro1q0bhg8fjieffBI1atRA8+bNyzT5AoBPPvkEtWvXxpdffom6deti2LBheOyxx9C5c2dkZGRgwYIFBdY6Dh8+HAaDAZ9++qm536Utb731Fp5//nkcOHAADRs2RPv27TFs2DAMHjwYTZo0QfXq1bFo0SKLbVavXo2goCCsWbMG9evXx7Bhw9ClSxc8+eST5h8TiCobJl9E5HB+fn6YPHlygctVKhX27NmDYcOGIS0tDdu3b4der8cPP/xg89fbshQQEIAjR47gsccew759+7B//340atQI69evt3kxWbduXZw+fRrz5s1DtWrVcOTIEWzatAmXL19GixYtsHLlSqv+FA+jadOmOHXqFF555RWkp6djw4YNiIqKwoABA3Do0CEMGTKk1B6rtGzduhUzZsxAYGAgfvvtN5w8eRJDhw7FkSNHSqWGwpEGDRqE3bt3o2vXrjh37hx+/fVXhIaG4uDBg+bRDu0dgODdd9/FV199hREjRkCpVOLgwYPYsGEDLly4gLZt22LBggU4c+aMRd8dNzc3HDx4EK+99hrc3d2xbds2XLx4ERMmTMCuXbtsTkFQlpYtW4b58+cjOjoaW7duhUQiwbhx43Dw4EGbw43bolKp8Ntvv+GLL75Au3btcPHiRWzYsAEnTpxAYGAgZs+ejQULFpjXHzlyJKZOnYrQ0FAcO3YMGzduxPXr1/HEE09g9+7dmDJlSqk9vzFjxiAsLAynT5/G9u3bAQBVq1bF4cOHsWzZMjRq1AinT5/Ghg0bcO7cOYSHh+Ojjz7CG2+8UWox2BIWFobjx49j6tSpkMvl2LRpE06ePIkePXpgx44dhb4GeWu6OnfuXOAonVKpFF9++SW2bt2Knj174vr169i4cSP+/PNPuLu7480338S6desstgkPD8fRo0fx3HPPITk5GVu2bEFqairWr19f5q8JUXklEcUZUoqIiIjs0qdPH+zYsQNHjhxBu3btnB1OmapVqxaio6OLNUolEVFlxJovIiKiErpz545VvySDwYCPP/4YO3bsQL169dC2bVsnRUdEROUNh5onIiIqoYMHD2LEiBFo0aIFatasCa1Wi3/++Qc3btyAWq3GmjVrnDrfGhERlS+s+SIiIiqhVq1a4YUXXkBycjJ27tyJHTt2QK/X4/nnn8fx48fL1TD/RETkfOzzRURERERE5ACs+SIiIiIiInIAJl9EREREREQOwOSLiIiIiIjIAZh8EREREREROQCTLyIiIiIiIgdg8kVEREREROQATL6IiIiIiIgcgMkXERERERGRAzD5IiIiIiIicgAmX0RERERERA7A5IuIiIiIiMgBmHwRERERERE5AJMvIiIiIiIiB2DyRURERERE5AD/DwKXO1vYsCrAAAAAAElFTkSuQmCC", "text/plain": [ "
" ] diff --git a/examples/wikitext/README.md b/examples/wikitext/README.md index c08cd48..9acccd0 100644 --- a/examples/wikitext/README.md +++ b/examples/wikitext/README.md @@ -1,48 +1,135 @@ # WikiText & GPT-2 Example -This directory contains scripts for fine-tuning GPT-2 on WikiText2 dataset. The pipeline is motivated from -[HuggingFace Example](https://github.com/huggingface/transformers/tree/main/examples/pytorch/language-modeling). -Please begin by installing necessary packages. +This repository contains scripts for fine-tuning GPT-2 and computing influence scores on the WikiText2 dataset. The pipeline is inspired by [this HuggingFace example](https://github.com/huggingface/transformers/tree/main/examples/pytorch/language-modeling). +Install the necessary packages: + ```bash pip install -r requirements.txt ``` ## Training -To fine-tune GPT-2, run the following command: +To fine-tune GPT-2, run: + ```bash python train.py --checkpoint_dir ./checkpoints \ --train_batch_size 8 \ --eval_batch_size 16 \ --learning_rate 3e-05 \ --weight_decay 0.01 \ - --num_train_epochs 3 + --num_train_epochs 3 \ + --seed 1004 ``` +This will fine-tune the model using the specified hyperparameters and save the final checkpoint in the `./checkpoints` directory. + ## Computing Pairwise Influence Scores -To obtain a pairwise influence scores on 481 query data points using `ekfac`, run the following command: +To compute pairwise influence scores using the `ekfac` strategy, run the following command: + ```bash python analyze.py --query_batch_size 32 \ --train_batch_size 64 \ --checkpoint_dir ./checkpoints \ --factor_strategy ekfac ``` -You can also use `identity`, `diagonal`, and `kfac`. On A100 (80GB), it takes roughly 50 minutes to compute the -pairwise scores (including computing EKFAC factors). +You can also use `identity`, `diagonal`, and `kfac` for `factor_strategy`. On an A100 (80GB) GPU, this process takes approximately 40 minutes: + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 11 | 2357.4 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Compute Pairwise Score | 1888.2 | 1 | 1888.2 | 80.098 | +| Fit Lambda | 274.64 | 1 | 274.64 | 11.651 | +| Fit Covariance | 180.27 | 1 | 180.27 | 7.6471 | +| Perform Eigendecomposition | 7.7754 | 1 | 7.7754 | 0.32984 | +| Save Eigendecomposition | 3.0652 | 1 | 3.0652 | 0.13003 | +| Save Covariance | 2.6799 | 1 | 2.6799 | 0.11368 | +| Save Lambda | 0.66036 | 1 | 0.66036 | 0.028013 | +| Load Covariance | 0.033343 | 1 | 0.033343 | 0.0014144 | +| Save Pairwise Score | 0.016471 | 1 | 0.016471 | 0.0006987 | +| Load All Factors | 0.0086084 | 1 | 0.0086084 | 0.00036517 | +| Load Eigendecomposition | 0.0054964 | 1 | 0.0054964 | 0.00023316 | +---------------------------------------------------------------------------------------------------------------------------------- +``` + +For more efficient computation, use half precision: + +```bash +python analyze.py --query_batch_size 32 \ + --train_batch_size 64 \ + --checkpoint_dir ./checkpoints \ + --factor_strategy ekfac \ + --use_half_precision +``` + +This reduces computation time to about 15 minutes on an A100 (80GB) GPU: + +``` +---------------------------------------------------------------------------------------------------------------------------------- +| Action | Mean duration (s) | Num calls | Total time (s) | Percentage % | +---------------------------------------------------------------------------------------------------------------------------------- +| Total | - | 11 | 785.92 | 100 % | +---------------------------------------------------------------------------------------------------------------------------------- +| Compute Pairwise Score | 654.62 | 1 | 654.62 | 83.294 | +| Fit Lambda | 74.662 | 1 | 74.662 | 9.4999 | +| Fit Covariance | 45.784 | 1 | 45.784 | 5.8256 | +| Perform Eigendecomposition | 7.5987 | 1 | 7.5987 | 0.96685 | +| Save Eigendecomposition | 1.4441 | 1 | 1.4441 | 0.18375 | +| Save Covariance | 1.3445 | 1 | 1.3445 | 0.17107 | +| Save Lambda | 0.38279 | 1 | 0.38279 | 0.048705 | +| Load Covariance | 0.058189 | 1 | 0.058189 | 0.0074039 | +| Save Pairwise Score | 0.0094807 | 1 | 0.0094807 | 0.0012063 | +| Load All Factors | 0.0083676 | 1 | 0.0083676 | 0.0010647 | +| Load Eigendecomposition | 0.0053729 | 1 | 0.0053729 | 0.00068364 | +---------------------------------------------------------------------------------------------------------------------------------- +``` + +The `half_precision_analysis.py` script compares the correlations between `float32` and `bfloat16` scores. + +

+Half Precision +

+ +The average correlation for 481 data points is `0.96`. ## Counterfactual Experiment -We can conduct counterfactual experiment by observing the increase in validation perplexity when removing top influential sequences. -We show a simple demo in `run_counterfactual.py` (the code assumes that you have computed the pairwise influence scores with `ekfac` and `identity`). +`run_counterfactual.py` demonstrates a counterfactual experiment by observing the increase in validation perplexity when removing top influential sequences. +(Note: This requires pre-computed pairwise influence scores with `ekfac` and `identity` strategies.) +

Counterfactual

+## Evaluating Linear Datamodeling Score + +The `evaluate_lds.py` script computes the [linear datamodeling score (LDS)](https://arxiv.org/abs/2303.14186). It measures the LDS obtained by +retraining the network 500 times with different subsets of the dataset (5 repeats and 100 masks). We obtain `0.44` LDS +(`0.42` LDS with the half precision and `0.12` LDS with the `identity` strategy). +The script also includes functionality to print out top influential sequences for a given query. -## Computing Linear Datamodeling Score +``` +Query Example: + = Homarus gammarus = + Homarus gammarus, known as the European lobster or common lobster, is a species of clawed lobster from the eastern Atlantic Ocean, Mediterranean Sea and parts of the Black Sea. It is closely related to the American lobster, H. americanus. It may grow to a length of 60 cm ( 24 in ) and a mass of 6 kilograms ( 13 lb ), and bears a conspicuous pair of claws. In life, the lobsters are blue, only becoming " lobster red " on cooking. Mating occurs in the summer, producing eggs which are carried by the females for up to a year before hatching into planktonic larvae. Homarus gammarus is a highly esteemed food, and is widely caught using lobster pots, mostly around the British Isles. + = = Description = = + Homarus gammarus is a large crustacean, with a body length up to 60 centimetres ( 24 in ) and weighing up to 5 – 6 kilograms ( 11 – 13 lb ), although the lobsters caught in lobster pots are usually 23 – 38 cm ( 9 – 15 in ) long and weigh 0 @.@ 7 – 2 @.@ 2 kg ( 1 @.@ 5 – 4 @.@ 9 lb ). Like other crustaceans, lobsters have a hard exoskeleton which they must shed in order to grow, in a process called ecdysis ( moulting ). This may occur several times a year for young lobsters, but decreases to once every 1 – 2 years for larger animals. + The first pair of pereiopods is armed with a large, asymmetrical pair of claws. The larger one is the " crusher ", and has rounded nodules used for crushing prey ; the other is the " cutter ", which has sharp inner edges, and is used for holding or tearing the prey. Usually, the left claw is the crusher, and the right is the cutter. + The exoskeleton is generally blue above, with spots that coalesce, and yellow below. The red colour associated with lobsters only appears after cooking. This occurs because, in life, the red pigment astaxanthin is bound to a protein complex, but the complex is broken up by the heat of cooking, releasing the red pigment. + The closest relative of H. gammarus is the American lobster, Homarus americanus. The two species are very similar, and can be crossed artificially -We can also compute the [Linear Datamodeling Score (LDS)](https://arxiv.org/abs/2303.14186). The code in `evaluate_lds.py` measures the LDS obtained by -retraining the network 600 times with different subsets of the dataset (5 repeats and 120 masks). We can obtain `0.37` LDS. \ No newline at end of file +Top Influential Example: + Sector Headquarters, Port Moresby + = Cape lobster = + The Cape lobster, Homarinus capensis, is a species of small lobster that lives off the coast of South Africa, from Dassen Island to Haga Haga. Only a few dozen specimens are known, mostly regurgitated by reef @-@ dwelling fish. It lives in rocky reefs, and is thought to lay large eggs that have a short larval phase, or that hatch directly as a juvenile. The species grows to a total length of 10 cm ( 3 @.@ 9 in ), and resembles a small European or American lobster ; it was previously included in the same genus, Homarus, although it is not very closely related to those species, and is now considered to form a separate, monotypic genus – Homarinus. Its closest relatives are the genera Thymops and Thymopides. + = = Distribution and ecology = = + The Cape lobster is endemic to South Africa. It occurs from Dassen Island, Western Cape in the west to Haga Haga, Eastern Cape in the east, a range of 900 kilometres ( 560 mi ). Most of the known specimens were regurgitated by fish caught on reefs at depths of 20 – 40 metres ( 66 – 131 ft ). This suggests that the Cape lobster inhabits rocky substrates, and may explain its apparent rarity, since such areas are not amenable to dredging or trawling, and the species may be too small to be retained by lobster traps. + = = Description = = + Homarinus capensis is considerably smaller than the large northern lobsters of the Atlantic Ocean, Homarus gammarus ( the European lobster ) and Homarus americanus ( the American lobster ), at 8 – 10 centimetres ( 3 @.@ 1 – 3 @.@ 9 in ) total length, or 4 – 5 cm ( 1 @.@ 6 – 2 @.@ 0 in ) carapace length. Accounts of the colouration of H. capensis are very variable, from tawny, red or yellow to " a rather dark olive ", similar to Homarus gammarus. + Homarinus and Homarus are considered to be the most plesiomorphic genera in the family Nephropidae. Nonetheless, the Cape lobster differs from Homarus in a number of characters. The rostrum of the Cape lobster is flattened, while that of Homarus is rounded in section +``` \ No newline at end of file diff --git a/examples/wikitext/analyze.py b/examples/wikitext/analyze.py index d19b5e9..8afbc30 100644 --- a/examples/wikitext/analyze.py +++ b/examples/wikitext/analyze.py @@ -1,7 +1,7 @@ import argparse import logging import os -from typing import Dict, List, Optional +from typing import Dict, List import torch import torch.nn.functional as F @@ -12,6 +12,8 @@ from kronfluence.analyzer import Analyzer, prepare_model from kronfluence.arguments import FactorArguments, ScoreArguments from kronfluence.task import Task +from kronfluence.utils.common.factor_arguments import all_low_precision_factor_arguments +from kronfluence.utils.common.score_arguments import all_low_precision_score_arguments from kronfluence.utils.dataset import DataLoaderKwargs BATCH_TYPE = Dict[str, torch.Tensor] @@ -27,6 +29,12 @@ def parse_args(): help="A path that is storing the final checkpoint of the model.", ) + parser.add_argument( + "--factor_strategy", + type=str, + default="ekfac", + help="Strategy to compute influence factors.", + ) parser.add_argument( "--query_gradient_rank", type=int, @@ -35,10 +43,16 @@ def parse_args(): ) parser.add_argument( "--use_half_precision", - type=bool, + action="store_true", default=False, help="Whether to use half precision for computing factors and scores.", ) + parser.add_argument( + "--use_compile", + action="store_true", + default=False, + help="Whether to use torch compile for computing factors and scores.", + ) parser.add_argument( "--query_batch_size", type=int, @@ -52,16 +66,16 @@ def parse_args(): help="Batch size for computing query gradients.", ) parser.add_argument( - "--factor_strategy", - type=str, - default="ekfac", - help="Strategy to compute influence factors.", + "--profile", + action="store_true", + default=False, + help="Boolean flag to profile computations.", ) - args = parser.parse_args() if args.checkpoint_dir is not None: os.makedirs(args.checkpoint_dir, exist_ok=True) + return args @@ -76,23 +90,21 @@ def compute_train_loss( input_ids=batch["input_ids"], attention_mask=batch["attention_mask"], ).logits - - shift_logits = logits[..., :-1, :].contiguous() + logits = logits[..., :-1, :].contiguous() + logits = logits.view(-1, logits.size(-1)) if not sample: labels = batch["labels"] - shift_labels = labels[..., 1:].contiguous() - reshaped_shift_logits = shift_logits.view(-1, shift_logits.size(-1)) - summed_loss = F.cross_entropy(reshaped_shift_logits, shift_labels.view(-1), reduction="sum") + labels = labels[..., 1:].contiguous() + summed_loss = F.cross_entropy(logits, labels.view(-1), reduction="sum") else: - reshaped_shift_logits = shift_logits.view(-1, shift_logits.size(-1)) with torch.no_grad(): - probs = torch.nn.functional.softmax(reshaped_shift_logits, dim=-1) + probs = torch.nn.functional.softmax(logits.detach(), dim=-1) sampled_labels = torch.multinomial( probs, num_samples=1, ).flatten() - summed_loss = F.cross_entropy(reshaped_shift_logits, sampled_labels.detach(), reduction="sum") + summed_loss = F.cross_entropy(logits, sampled_labels, reduction="sum") return summed_loss def compute_measurement( @@ -103,7 +115,7 @@ def compute_measurement( # We could also compute the log-likelihood or averaged margin. return self.compute_train_loss(batch, model) - def tracked_modules(self) -> List[str]: + def get_influence_tracked_modules(self) -> List[str]: total_modules = [] for i in range(12): @@ -116,7 +128,7 @@ def tracked_modules(self) -> List[str]: return total_modules - def get_attention_mask(self, batch: BATCH_TYPE) -> Optional[torch.Tensor]: + def get_attention_mask(self, batch: BATCH_TYPE) -> torch.Tensor: return batch["attention_mask"] @@ -143,10 +155,14 @@ def main(): task = LanguageModelingTask() model = prepare_model(model, task) + if args.use_compile: + model = torch.compile(model) + analyzer = Analyzer( analysis_name="wikitext", model=model, task=task, + profile=args.profile, ) # Configure parameters for DataLoader. dataloader_kwargs = DataLoaderKwargs(collate_fn=default_data_collator) @@ -156,43 +172,41 @@ def main(): factors_name = args.factor_strategy factor_args = FactorArguments(strategy=args.factor_strategy) if args.use_half_precision: - factor_args.activation_covariance_dtype = torch.bfloat16 - factor_args.gradient_covariance_dtype = torch.bfloat16 - factor_args.lambda_dtype = torch.bfloat16 + factor_args = all_low_precision_factor_arguments(strategy=args.factor_strategy, dtype=torch.bfloat16) factors_name += "_half" - + if args.use_compile: + factors_name += "_compile" analyzer.fit_all_factors( factors_name=factors_name, dataset=train_dataset, per_device_batch_size=None, factor_args=factor_args, + initial_per_device_batch_size_attempt=64, overwrite_output_dir=False, - initial_per_device_batch_size_attempt=128, ) # Compute pairwise scores. + score_args = ScoreArguments() + scores_name = factor_args.strategy + if args.use_half_precision: + score_args = all_low_precision_score_arguments(dtype=torch.bfloat16) + scores_name += "_half" + if args.use_compile: + scores_name += "_compile" rank = args.query_gradient_rank if args.query_gradient_rank != -1 else None - score_args = ScoreArguments(query_gradient_rank=rank, query_gradient_svd_dtype=torch.float32) - scores_name = f"{factor_args.strategy}_pairwise" if rank is not None: + score_args.query_gradient_low_rank = rank + score_args.query_gradient_accumulation_steps = 10 scores_name += f"_qlr{rank}" - - if args.use_half_precision: - score_args.per_sample_gradient_dtype = torch.bfloat16 - score_args.score_dtype = torch.bfloat16 - score_args.cached_activation_cpu_offload = True - scores_name += "_half" - analyzer.compute_pairwise_scores( scores_name=scores_name, score_args=score_args, - factors_name=args.factor_strategy, + factors_name=factors_name, query_dataset=eval_dataset, - query_indices=list(range(min([len(eval_dataset), 2000]))), train_dataset=train_dataset, per_device_query_batch_size=args.query_batch_size, per_device_train_batch_size=args.train_batch_size, - overwrite_output_dir=True, + overwrite_output_dir=False, ) scores = analyzer.load_pairwise_scores(scores_name)["all_modules"] logging.info(f"Scores shape: {scores.shape}") diff --git a/examples/wikitext/evaluate_lds.py b/examples/wikitext/evaluate_lds.py index 407e510..654dbb8 100644 --- a/examples/wikitext/evaluate_lds.py +++ b/examples/wikitext/evaluate_lds.py @@ -2,27 +2,63 @@ import numpy as np import torch +import tqdm from scipy.stats import spearmanr +from transformers import AutoTokenizer +from examples.wikitext.pipeline import get_wikitext_dataset from kronfluence.analyzer import Analyzer +def evaluate_correlations(scores: torch.Tensor) -> float: + losses = torch.from_numpy(torch.load(open("files/losses.pt", "rb"))) + masks = torch.from_numpy(torch.load(open("files/masks.pt", "rb"))).float() + + val_indices = np.arange(481) + preds = -masks @ scores.T + + rs = [] + ps = [] + for j in tqdm.tqdm(val_indices): + r, p = spearmanr(preds[:, j], losses[:, j]) + rs.append(r) + ps.append(p) + rs, ps = np.array(rs), np.array(ps) + return rs.mean() + + def main(): logging.basicConfig(level=logging.INFO) - results = torch.load("files/lds_results.pt") - diff_loss = torch.from_numpy(results["diff_loss"]) - mask = torch.from_numpy(results["mask"]).float() - mask = ((mask + 1) % 2).to(dtype=torch.float64).t() - # You might need to change the path. - scores = Analyzer.load_file("scores_pairwise/ekfac_pairwise.safetensors")["all_modules"].to(dtype=torch.float64) - preds = (scores @ mask).t().numpy() + scores = Analyzer.load_file("influence_results/wikitext/scores_ekfac/pairwise_scores.safetensors")[ + "all_modules" + ].to(dtype=torch.float32) + # scores = Analyzer.load_file("influence_results/wikitext/scores_ekfac_half/pairwise_scores.safetensors")[ + # "all_modules" + # ].to(dtype=torch.float32) + # scores = Analyzer.load_file("influence_results/wikitext/scores_identity/pairwise_scores.safetensors")[ + # "all_modules" + # ].to(dtype=torch.float32) + + corr_mean = evaluate_correlations(scores) + logging.info(f"LDS: {np.mean(corr_mean)}") + + # We can also visualize the top influential sequences. + eval_idx = 0 + train_dataset = get_wikitext_dataset( + split="eval_train", + ) + eval_dataset = get_wikitext_dataset( + split="valid", + ) + tokenizer = AutoTokenizer.from_pretrained("gpt2", use_fast=True, trust_remote_code=True) + print("Query Data Example:") + print(tokenizer.decode(eval_dataset[eval_idx]["input_ids"])) - corr_lst = [] - for i in range(diff_loss.shape[1]): - corr_lst.append(spearmanr(diff_loss[:, i], preds[:, i])[0]) - logging.info(f"LDS: {np.mean(corr_lst)}") + top_idx = int(torch.argsort(scores[eval_idx], descending=True)[0]) + print("Top Influential Example:") + print(tokenizer.decode(train_dataset[top_idx]["input_ids"])) if __name__ == "__main__": diff --git a/examples/wikitext/figure/half_precision.png b/examples/wikitext/figure/half_precision.png new file mode 100644 index 0000000..4de9d5e Binary files /dev/null and b/examples/wikitext/figure/half_precision.png differ diff --git a/examples/wikitext/files/losses.pt b/examples/wikitext/files/losses.pt new file mode 100644 index 0000000..7bc4c8c Binary files /dev/null and b/examples/wikitext/files/losses.pt differ diff --git a/examples/wikitext/files/masks.pt b/examples/wikitext/files/masks.pt new file mode 100644 index 0000000..40d0b18 Binary files /dev/null and b/examples/wikitext/files/masks.pt differ diff --git a/examples/glue/query_batching_analysis.py b/examples/wikitext/half_precision_analysis.py similarity index 52% rename from examples/glue/query_batching_analysis.py rename to examples/wikitext/half_precision_analysis.py index ad00ff4..21cfeb1 100644 --- a/examples/glue/query_batching_analysis.py +++ b/examples/wikitext/half_precision_analysis.py @@ -12,22 +12,28 @@ def main(): logging.basicConfig(level=logging.INFO) # Load the scores. You might need to modify the path. - full_scores = Analyzer.load_file("scores_ekfac/pairwise_scores.safetensors")["all_modules"] - lr_scores = Analyzer.load_file("scores_ekfac_qlr32/pairwise_scores.safetensors")["all_modules"] + full_scores = Analyzer.load_file("influence_results/wikitext/scores_ekfac/pairwise_scores.safetensors")[ + "all_modules" + ] + half_scores = Analyzer.load_file("influence_results/wikitext/scores_ekfac_half/pairwise_scores.safetensors")[ + "all_modules" + ].float() # Only plot first 1000 points to avoid clutter. + index = 5 plt.rcParams.update({"figure.dpi": 150}) plt.rcParams.update(markers.with_edge()) plt.rcParams["axes.axisbelow"] = True - plt.scatter(lr_scores[0][:1000], full_scores[0][:1000], edgecolor="k") + plt.scatter(half_scores[index], full_scores[index], edgecolor="k") plt.grid() - plt.xlabel("Full Rank Score") - plt.ylabel("Low Rank (32) Score") + plt.xlabel("bfloat32 Score") + plt.ylabel("float32 Score") plt.show() + # Compute the averaged spearman correlation. all_corr = [] - for i in range(100): - all_corr.append(spearmanr(full_scores[i], lr_scores[i])[0]) + for i in range(481): + all_corr.append(spearmanr(full_scores[i], half_scores[i])[0]) logging.info(f"Averaged Spearman Correlation: {np.array(all_corr).mean()}") diff --git a/examples/wikitext/pipeline.py b/examples/wikitext/pipeline.py index b0f5986..2040789 100644 --- a/examples/wikitext/pipeline.py +++ b/examples/wikitext/pipeline.py @@ -1,6 +1,7 @@ from itertools import chain from typing import List +import torch from datasets import load_dataset from torch import nn from torch.utils import data @@ -8,6 +9,7 @@ from transformers.pytorch_utils import Conv1D +@torch.no_grad() def replace_conv1d_modules(model: nn.Module) -> None: # GPT-2 is defined in terms of Conv1D. However, this does not work for Kronfluence. # Here, we convert these Conv1D modules to linear modules recursively. diff --git a/examples/wikitext/run_counterfactual.py b/examples/wikitext/run_counterfactual.py index d682b7d..e91e189 100644 --- a/examples/wikitext/run_counterfactual.py +++ b/examples/wikitext/run_counterfactual.py @@ -17,10 +17,10 @@ def main(): train_dataset = get_wikitext_dataset(split="train") # You might need to change the path. - identity_scores = Analyzer.load_file("analyses/wikitext/scores_identity_pairwise/pairwise_scores.safetensors")[ + identity_scores = Analyzer.load_file("influence_results/wikitext/scores_identity/pairwise_scores.safetensors")[ "all_modules" ][:50].sum(dim=0) - ekfac_scores = Analyzer.load_file("analyses/wikitext/scores_ekfac_pairwise/pairwise_scores.safetensors")[ + ekfac_scores = Analyzer.load_file("influence_results/wikitext/scores_ekfac/pairwise_scores.safetensors")[ "all_modules" ][:50].sum(dim=0) @@ -32,19 +32,19 @@ def get_topk_keep_indices(current_score: torch.Tensor, topk: int = 1) -> List[in remove_indices = [tensor.item() for tensor in remove_indices] return list(set(list(range(len(train_dataset)))) - set(remove_indices)) - eval_train_dataset = get_wikitext_dataset(split="valid", indices=list(range(50))) + valid_dataset = get_wikitext_dataset(split="valid", indices=list(range(50))) - def train_and_evaluate(indices): - train_dataset = get_wikitext_dataset(split="train", indices=indices) + def train_and_evaluate(indices) -> float: model = train( - dataset=train_dataset, + dataset=get_wikitext_dataset(split="train", indices=indices), batch_size=8, num_train_epochs=3, learning_rate=3e-05, weight_decay=0.01, ) - return evaluate_model(model, eval_train_dataset, batch_size=16) + return evaluate_model(model, valid_dataset, batch_size=16) + # We do this to make the experiment faster; set `num_iter = 5` for proper experiment. num_iter = 1 topk_lst = [0, 50, 100, 150, 200] diff --git a/examples/wikitext/train.py b/examples/wikitext/train.py index ebb880c..fcad14e 100644 --- a/examples/wikitext/train.py +++ b/examples/wikitext/train.py @@ -5,10 +5,8 @@ import time import torch -import torch.nn.functional as F from accelerate.utils import set_seed from torch import nn -from torch.nn import CrossEntropyLoss from torch.utils import data from transformers import default_data_collator @@ -55,7 +53,7 @@ def parse_args(): parser.add_argument( "--seed", type=int, - default=0, + default=1004, help="A seed for reproducible training pipeline.", ) parser.add_argument( @@ -64,7 +62,6 @@ def parse_args(): default="./checkpoints", help="A path to store the final checkpoint.", ) - args = parser.parse_args() if args.checkpoint_dir is not None: @@ -90,25 +87,21 @@ def train( model = construct_gpt2().to(DEVICE) optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate, weight_decay=weight_decay) - loss_fn = CrossEntropyLoss(reduction="mean") start_time = time.time() - model.eval() + model.train() for epoch in range(num_train_epochs): total_loss = 0.0 for batch in train_dataloader: - model.zero_grad() - lm_logits = model( + optimizer.zero_grad(set_to_none=True) + loss = model( input_ids=batch["input_ids"].to(device=DEVICE), attention_mask=batch["attention_mask"].to(device=DEVICE), - ).logits - labels = batch["labels"].to(device=DEVICE) - shift_logits = lm_logits[..., :-1, :].contiguous() - shift_labels = labels[..., 1:].contiguous() - loss = loss_fn(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)) + labels=batch["labels"].to(device=DEVICE), + ).loss loss.backward() optimizer.step() - total_loss += loss.detach().float() + total_loss += loss.detach().float() * batch["input_ids"].shape[0] logging.info(f"Epoch {epoch + 1} - Averaged Loss: {total_loss / len(dataset)}") end_time = time.time() elapsed_time = end_time - start_time @@ -126,17 +119,13 @@ def evaluate_model(model: nn.Module, dataset: data.Dataset, batch_size: int) -> total_num = 0 for batch in dataloader: with torch.no_grad(): - lm_logits = model( + loss = model( input_ids=batch["input_ids"].to(device=DEVICE), attention_mask=batch["attention_mask"].to(device=DEVICE), - ).logits - labels = batch["labels"].to(device=DEVICE) - shift_logits = lm_logits[..., :-1, :].contiguous() - shift_labels = labels[..., 1:].contiguous() - reshaped_shift_logits = shift_logits.view(-1, shift_logits.size(-1)) - loss = F.cross_entropy(reshaped_shift_logits, shift_labels.view(-1), reduction="sum").detach().float() - total_loss += loss - total_num += reshaped_shift_logits.shape[0] + labels=batch["labels"].to(device=DEVICE), + ).loss + total_loss += loss.detach() * batch["input_ids"].shape[0] + total_num += batch["input_ids"].shape[0] return total_loss.item() / total_num @@ -164,7 +153,10 @@ def main(): eval_dataset = get_wikitext_dataset(split="valid") eval_loss = evaluate_model(model=model, dataset=eval_dataset, batch_size=args.eval_batch_size) - eval_perplexity = math.exp(eval_loss) + try: + eval_perplexity = math.exp(eval_loss) + except OverflowError: + eval_perplexity = float("inf") logger.info(f"Evaluation perplexity: {eval_perplexity}") if args.checkpoint_dir is not None: diff --git a/kronfluence/analyzer.py b/kronfluence/analyzer.py index 710eb26..615c1c0 100644 --- a/kronfluence/analyzer.py +++ b/kronfluence/analyzer.py @@ -1,3 +1,4 @@ +import copy from pathlib import Path from typing import Dict, Optional, Union @@ -20,34 +21,32 @@ def prepare_model( model: nn.Module, task: Task, ) -> nn.Module: - """Prepares the model before passing it to `Analyzer`. This function sets `param.requires_grad = False` - for all modules that does not require influence computations and installs `TrackedModule` to supported - modules. This `TrackedModule` keeps track of relevant statistics needed to compute influence scores. + """Prepares the model for analysis by setting all parameters and buffers to non-trainable + and installing `TrackedModule` wrappers on supported modules. Args: model (nn.Module): - The PyTorch model to be analyzed. + The PyTorch model to be prepared for analysis. task (Task): - The specific task associated with the model. + The specific task associated with the model, used for `TrackedModule` installation. Returns: nn.Module: - The same PyTorch model with `param.requires_grad = False` on all modules that does not require influence - computations and with `TrackedModule` installed. + The prepared model with non-trainable parameters and `TrackedModule` wrappers. """ model.eval() for params in model.parameters(): params.requires_grad = False - # Install `TrackedModule` to the model. + for buffers in model.buffers(): + buffers.requires_grad = False + + # Install `TrackedModule` wrappers on supported modules. model = wrap_tracked_modules(model=model, task=task) return model class Analyzer(FactorComputer, ScoreComputer): - """ - Handles the computation of all factors (e.g., covariance and Lambda matrices for EKFAC) - and influence scores for a given PyTorch model. - """ + """Handles the computation of factors (e.g., covariance matrices) and scores for a given PyTorch model.""" def __init__( self, @@ -58,34 +57,37 @@ def __init__( log_level: Optional[int] = None, log_main_process_only: bool = True, profile: bool = False, - output_dir: str = "./analyses", + disable_tqdm: bool = False, + output_dir: str = "./influence_results", disable_model_save: bool = True, ) -> None: - """Initializes an instance of the Analyzer class. + """Initializes an instance of the `Analyzer` class. Args: analysis_name (str): - The unique identifier for the analysis, used to organize and retrieve the results. + Unique identifier for the analysis, used for organizing results. model (nn.Module): The PyTorch model to be analyzed. task (Task): The specific task associated with the model. cpu (bool, optional): - Specifies whether the analysis should be explicitly performed using the CPU. - Defaults to False, utilizing GPU resources if available. + If `True`, forces analysis to be performed on CPU. Defaults to `False`. log_level (int, optional): - The logging level to use (e.g., logging.DEBUG, logging.INFO). Defaults to the root logging level. + Logging level (e.g., logging.DEBUG, logging.INFO). Defaults to root logging level. log_main_process_only (bool, optional): - If True, restricts logging to the main process. Defaults to True. + If `True`, restricts logging to the main process. Defaults to `True`. profile (bool, optional): - Enables the generation of performance profiling logs. This can be useful for - identifying bottlenecks or performance issues. Defaults to False. + If `True`, enables performance profiling logs. Defaults to `False`. + disable_tqdm (bool, optional): + If `True`, disables TQDM progress bars. Defaults to `False`. output_dir (str): - The file path to the directory, where analysis results will be stored. If the directory - does not exist, it will be created. Defaults to './analyses'. + Directory path for storing analysis results. Defaults to './influence_results'. disable_model_save (bool, optional): - If set to True, prevents the saving of the model's state_dict. When the provided model is different - from the previously saved model, it will raise an Exception. Defaults to True. + If `True`, prevents saving the model's `state_dict`. Defaults to `True`. + + Raises: + ValueError: + If the provided model differs from a previously saved model when `disable_model_save=False`. """ super().__init__( name=analysis_name, @@ -95,22 +97,23 @@ def __init__( log_level=log_level, log_main_process_only=log_main_process_only, profile=profile, + disable_tqdm=disable_tqdm, output_dir=output_dir, ) - self.logger.info(f"Initializing Computer with parameters: {locals()}") - self.logger.debug(f"Process state configuration:\n{repr(self.state)}") + self.logger.info(f"Initializing `Analyzer` with parameters: {locals()}") + self.logger.info(f"Process state configuration:\n{repr(self.state)}") - # Save model parameters. + # Save model parameters if necessary. if self.state.is_main_process and not disable_model_save: self._save_model() self.state.wait_for_everyone() def set_dataloader_kwargs(self, dataloader_kwargs: DataLoaderKwargs) -> None: - """Sets the default DataLoader parameters to use for all DataLoaders. + """Sets the default DataLoader arguments. Args: dataloader_kwargs (DataLoaderKwargs): - The object containing parameters for DataLoader. + The object containing arguments for PyTorch DataLoader. """ self._dataloader_params = dataloader_kwargs @@ -118,17 +121,16 @@ def set_dataloader_kwargs(self, dataloader_kwargs: DataLoaderKwargs) -> None: def _save_model(self) -> None: """Saves the model to the output directory.""" model_save_path = self.output_dir / "model.safetensors" - extracted_model = extract_model_from_parallel(model=self.model, keep_fp32_wrapper=True) + extracted_model = extract_model_from_parallel(model=copy.deepcopy(self.model), keep_fp32_wrapper=True) if model_save_path.exists(): self.logger.info(f"Found existing saved model at `{model_save_path}`.") - # Load the existing model's state_dict for comparison. + # Load existing model's `state_dict` for comparison. loaded_state_dict = load_file(model_save_path) if not verify_models_equivalence(loaded_state_dict, extracted_model.state_dict()): error_msg = ( "Detected a difference between the current model and the one saved at " - f"`{model_save_path}`. Consider using a different `analysis_name` to " - f"avoid conflicts." + f"`{model_save_path}`. Consider using a different `analysis_name` to avoid conflicts." ) self.logger.error(error_msg) raise ValueError(error_msg) @@ -149,27 +151,24 @@ def fit_all_factors( factor_args: Optional[FactorArguments] = None, overwrite_output_dir: bool = False, ) -> None: - """Computes all necessary factors for the given factor strategy. As an example, EK-FAC - requires (1) computing covariance matrices, (2) performing Eigendecomposition, and - (3) computing Lambda (corrected-eigenvalues) matrices. + """Computes all necessary factors for the given strategy. Args: factors_name (str): - The unique identifier for the factor, used to organize and retrieve the results. + Unique identifier for the factor, used for organizing results. dataset (data.Dataset): - The dataset that will be used to fit all the factors. + Dataset used to fit all influence factors. per_device_batch_size (int, optional): - The per-device batch size used to fit the factors. If not specified, executable - per-device batch size is automatically determined. + Per-device batch size for factor fitting. If not specified, executable per-device batch size + is automatically determined. initial_per_device_batch_size_attempt (int): - The initial attempted per-device batch size when the batch size is not provided. + Initial batch size attempt when `per_device_batch_size` is not explicitly provided. Defaults to `4096`. dataloader_kwargs (DataLoaderKwargs, optional): - Controls additional arguments for PyTorch's DataLoader. + Additional arguments for PyTorch's DataLoader. factor_args (FactorArguments, optional): - Arguments related to computing the factors. If not specified, - the default values of `FactorArguments` will be used. + Arguments for factor computation. Defaults to `FactorArguments` default values. overwrite_output_dir (bool, optional): - If True, the existing factors with the same name will be overwritten. + If `True`, overwrites existing factors with the same `factors_name`. Defaults to `False`. """ self.fit_covariance_matrices( factors_name=factors_name, @@ -197,36 +196,41 @@ def fit_all_factors( @staticmethod def load_file(path: Union[str, Path]) -> Dict[str, torch.Tensor]: - """Loads the `.safetensors` file at the given path from disk. - - See https://github.com/huggingface/safetensors. + """Loads a `safetensors` file from the given path. Args: path (Path): - The path to the `.safetensors` file. + The path to the `safetensors` file. Returns: Dict[str, torch.Tensor]: - The contents of the file, which is the dictionary mapping string to tensors. + Dictionary mapping strings to tensors from the loaded file. + + Raises: + FileNotFoundError: + If the specified file does not exist. + + Note: + For more information on `safetensors`, see https://github.com/huggingface/safetensors. """ if isinstance(path, str): path = Path(path).resolve() if not path.exists(): - raise FileNotFoundError(f"File does not exists at `{path}`.") + raise FileNotFoundError(f"File not found: {path}.") return load_file(path) @staticmethod def get_module_summary(model: nn.Module) -> str: - """Returns the formatted summary of the modules in model. Useful identifying which modules to - compute influence scores. + """Generates a formatted summary of the model's modules, excluding those without parameters. This summary is + useful for identifying which modules to compute influence scores for. Args: model (nn.Module): - The PyTorch model to be investigated. + The PyTorch model to be summarized. Returns: str: - The formatted string summary of the model. + A formatted string containing the model summary, including module names and representations. """ format_str = "==Model Summary==" for module_name, module in model.named_modules(): diff --git a/kronfluence/arguments.py b/kronfluence/arguments.py index cc0091c..a5ae0c3 100644 --- a/kronfluence/arguments.py +++ b/kronfluence/arguments.py @@ -10,7 +10,12 @@ class Arguments: """Base class for specifying arguments for computing factors and influence scores.""" def to_dict(self) -> Dict[str, Any]: - """Converts the arguments to a dictionary.""" + """Converts the arguments to a dictionary. + + Returns: + Dict[str, Any]: + A dictionary representation of the arguments, with `torch.dtype` values converted to strings. + """ config = copy.deepcopy(self.__dict__) for key, value in config.items(): if isinstance(value, torch.dtype): @@ -18,7 +23,12 @@ def to_dict(self) -> Dict[str, Any]: return config def to_str_dict(self) -> Dict[str, str]: - """Converts the arguments to a dictionary, where all values are converted to strings.""" + """Converts the arguments to a dictionary with all values as strings. + + Returns: + Dict[str, str]: + A dictionary representation of the arguments, with all values converted to strings. + """ config = copy.deepcopy(self.__dict__) for key, value in config.items(): config[key] = str(value) @@ -29,188 +39,236 @@ def to_str_dict(self) -> Dict[str, str]: class FactorArguments(Arguments): """Arguments for computing influence factors.""" - # General configuration. # + # General configuration # strategy: str = field( default="ekfac", - metadata={"help": "Strategy for computing preconditioning factors."}, + metadata={"help": "Specifies the algorithm for computing influence factors. Default is 'ekfac'."}, ) use_empirical_fisher: bool = field( default=False, metadata={ - "help": "Whether to use empirical Fisher (using labels from batch) instead of " - "true Fisher (using sampled labels)." - }, - ) - distributed_sync_steps: int = field( - default=1_000, - metadata={ - "help": "Specifies the total iteration step to synchronize the process when using distributed setting." + "help": "If `True`, approximates empirical Fisher (using true labels) instead of " + "true Fisher (using sampled labels from model's outputs)." }, ) amp_dtype: Optional[torch.dtype] = field( default=None, - metadata={"help": "Dtype for automatic mixed precision (AMP). Disables AMP if None."}, + metadata={"help": "Data type for automatic mixed precision (AMP). If `None`, AMP is disabled."}, + ) + amp_scale: float = field( + default=2.0**16, + metadata={"help": "Scale factor for AMP (only applicable when `amp_dtype=torch.float16`)."}, + ) + has_shared_parameters: bool = field( + default=False, + metadata={"help": "Indicates whether shared parameters are present in the model's forward pass."}, ) # Configuration for fitting covariance matrices. # covariance_max_examples: Optional[int] = field( default=100_000, - metadata={ - "help": "Maximum number of examples for fitting covariance matrices. " - "Uses all data examples for the given dataset if None." - }, + metadata={"help": "Maximum number of examples for fitting covariance matrices. Uses entire dataset if `None`."}, ) - covariance_data_partition_size: int = field( + covariance_data_partitions: int = field( default=1, - metadata={ - "help": "Number of data partitions for computing covariance matrices. " - "For example, when `covariance_data_partition_size = 2`, the dataset is split " - "into 2 chunks and covariance matrices are separately computed for each chunk." - }, + metadata={"help": "Number of partitions to divide the dataset into for covariance matrix computation."}, ) - covariance_module_partition_size: int = field( + covariance_module_partitions: int = field( default=1, metadata={ - "help": "Number of module partitions for computing covariance matrices. " - "For example, when `covariance_module_partition_size = 2`, the modules (layers) are split " - "into 2 chunks and covariance matrices are separately computed for each chunk." + "help": "Number of partitions to divide the model's modules (layers) into for " + "covariance matrix computation." }, ) activation_covariance_dtype: torch.dtype = field( default=torch.float32, - metadata={"help": "Dtype for computing activation covariance matrices."}, + metadata={"help": "Data type for activation covariance computation."}, ) gradient_covariance_dtype: torch.dtype = field( default=torch.float32, - metadata={"help": "Dtype for computing pseudo-gradient covariance matrices."}, + metadata={"help": "Data type for pseudo-gradient covariance computation."}, ) - # Configuration for performing eigendecomposition. # + # Configuration for performing eigendecomposition # eigendecomposition_dtype: torch.dtype = field( default=torch.float64, - metadata={"help": "Dtype for performing Eigendecomposition. Recommended to use `torch.float64`."}, + metadata={ + "help": "Data type for eigendecomposition. Double precision (`torch.float64`) is recommended " + "for numerical stability." + }, ) - # Configuration for fitting Lambda matrices. # + # Configuration for fitting Lambda matrices # lambda_max_examples: Optional[int] = field( default=100_000, - metadata={ - "help": "Maximum number of examples for fitting Lambda matrices. " - "Uses all data examples for the given dataset if None." - }, + metadata={"help": "Maximum number of examples for fitting Lambda matrices. Uses entire dataset if `None`."}, ) - lambda_data_partition_size: int = field( + lambda_data_partitions: int = field( default=1, - metadata={ - "help": "Number of data partitions for computing Lambda matrices. " - "For example, when `lambda_data_partition_size = 2`, the dataset is split " - "into 2 chunks and Lambda matrices are separately computed for each chunk." - }, + metadata={"help": "Number of partitions to divide the dataset into for Lambda matrix computation."}, ) - lambda_module_partition_size: int = field( + lambda_module_partitions: int = field( default=1, metadata={ - "help": "Number of module partitions for computing Lambda matrices. " - "For example, when `lambda_module_partition_size = 2`, the modules (layers) are split " - "into 2 chunks and Lambda matrices are separately computed for each chunk." + "help": "Number of partitions to divide the model's modules (layers) into for Lambda matrix computation." }, ) - lambda_iterative_aggregate: bool = field( + use_iterative_lambda_aggregation: bool = field( default=False, metadata={ - "help": "Whether to aggregate squared sum of projected per-sample-gradient with for-loop iterations." + "help": "If `True`, aggregates the squared sum of projected per-sample gradients " + "iteratively (with for-loop) to reduce GPU memory usage." }, ) - cached_activation_cpu_offload: bool = field( + offload_activations_to_cpu: bool = field( default=False, - metadata={"help": "Whether to offload cached activation to CPU for computing the per-sample-gradient."}, + metadata={"help": "If `True`, offloads cached activations to CPU memory when computing per-sample gradients."}, + ) + per_sample_gradient_dtype: torch.dtype = field( + default=torch.float32, + metadata={"help": "Data type for per-sample gradient computation."}, ) lambda_dtype: torch.dtype = field( default=torch.float32, - metadata={"help": "Dtype for computing Lambda (corrected eigenvalues) matrices."}, + metadata={"help": "Data type for Lambda matrix computation."}, ) + def __post_init__(self) -> None: + if self.covariance_max_examples is not None and self.covariance_max_examples <= 0: + raise ValueError("`covariance_max_examples` must be `None` or positive.") + + if self.lambda_max_examples is not None and self.lambda_max_examples <= 0: + raise ValueError("`lambda_max_examples` must be `None` or positive.") + + if any( + partition <= 0 + for partition in [ + self.covariance_data_partitions, + self.covariance_module_partitions, + self.lambda_data_partitions, + self.lambda_module_partitions, + ] + ): + raise ValueError("All data and module partitions must be positive.") + + # For backward compatibility: + if not hasattr(self, "amp_scale"): + self.amp_scale = 2.0**16 + @dataclass class ScoreArguments(Arguments): """Arguments for computing influence scores.""" - # General configuration. # - damping: Optional[float] = field( + # General configuration # + damping_factor: Optional[float] = field( default=1e-08, metadata={ - "help": "A damping factor for the damped inverse Hessian-vector product (iHVP). " - "Uses a heuristic based on mean eigenvalues (0.1 x mean eigenvalues) if set to None." + "help": "Damping factor for inverse Hessian-vector product (iHVP). " + "If `None`, uses a heuristic of 0.1 times the mean eigenvalue." }, ) - cached_activation_cpu_offload: bool = field( + amp_dtype: Optional[torch.dtype] = field( + default=None, + metadata={"help": "Data type for automatic mixed precision (AMP). If `None`, AMP is disabled."}, + ) + offload_activations_to_cpu: bool = field( default=False, - metadata={"help": "Whether to offload cached activation to CPU for computing the per-sample-gradient."}, + metadata={"help": "If `True`, offloads cached activations to CPU memory when computing per-sample gradients."}, + ) + + # Partition configuration # + data_partitions: int = field( + default=1, + metadata={"help": "Number of partitions to divide the dataset for influence score computation."}, ) - distributed_sync_steps: int = field( - default=1_000, + module_partitions: int = field( + default=1, metadata={ - "help": "Specifies the total iteration step to synchronize the process when using distributed setting." + "help": "Number of partitions to divide the model's modules (layers) into for influence score computation." }, ) - amp_dtype: Optional[torch.dtype] = field( - default=None, - metadata={"help": "Dtype for automatic mixed precision (AMP). Disables AMP if None."}, - ) - # Partition configuration. # - data_partition_size: int = field( - default=1, + # General score configuration # + compute_per_module_scores: bool = field( + default=False, + metadata={"help": "If `True`, computes separate scores for each module instead of summing across all."}, + ) + compute_per_token_scores: bool = field( + default=False, metadata={ - "help": "Number of data partitions for computing influence scores. For example, when " - "`data_partition_size = 2`, the dataset is split into 2 chunks and scores are separately " - "computed for each chunk." + "help": "If `True`, computes separate scores for each token instead of summing across all. " + "Only applicable to transformer-based models." }, ) - module_partition_size: int = field( + + # Pairwise influence score configuration # + query_gradient_accumulation_steps: int = field( default=1, + metadata={"help": "Number of query batches to accumulate before processing training examples."}, + ) + query_gradient_low_rank: Optional[int] = field( + default=None, metadata={ - "help": "Number of module partitions for computing influence scores. For example, when " - "`module_partition_size = 2`, the module (layers) are split into 2 modules and scores are separately " - "computed for each chunk." + "help": "Rank for the low-rank approximation of the query gradient (query batching). " + "If `None`, no low-rank approximation is applied." }, ) - - # Score configuration. # - per_module_score: bool = field( + use_full_svd: bool = field( default=False, metadata={ - "help": "Whether to obtain per-module scores instead of the summed scores across all modules. " - "This is useful when performing layer-wise influence analysis." + "help": "If `True`, uses `torch.linalg.svd` instead of `torch.svd_lowrank` for query batching. " + "Full SVD is more accurate but slower and more memory-intensive." }, ) - query_gradient_rank: Optional[int] = field( - default=None, - metadata={"help": "Rank for the query gradient. Does not apply low-rank approximation if None."}, + aggregate_query_gradients: bool = field( + default=False, + metadata={ + "help": "If `True`, uses the summed query gradient instead of per-sample query gradients " + "for pairwise influence computation." + }, ) - num_query_gradient_aggregations: int = field( - default=1, - metadata={"help": "Number of query batches to aggregate over."}, + aggregate_train_gradients: bool = field( + default=False, + metadata={ + "help": "If `True`, uses the summed training gradient instead of per-sample training gradients " + "for pairwise influence computation." + }, ) + + # Self-influence score configuration # use_measurement_for_self_influence: bool = field( default=False, - metadata={"help": "Whether to use the measurement (instead of the loss) for computing self-influence scores."}, + metadata={"help": "If `True`, uses the measurement (instead of the loss) for computing self-influence scores."}, ) - # Dtype configuration. # + # Precision configuration # query_gradient_svd_dtype: torch.dtype = field( default=torch.float32, - metadata={"help": "Dtype for performing singular value decomposition (SVD) on the query gradient."}, - ) - score_dtype: torch.dtype = field( - default=torch.float32, - metadata={"help": "Dtype for computing and storing influence scores."}, + metadata={"help": "Data type for singular value decomposition (SVD) of query gradient."}, ) per_sample_gradient_dtype: torch.dtype = field( default=torch.float32, - metadata={"help": "Dtype for computing per-sample-gradients."}, + metadata={"help": "Data type for query per-sample gradient computation."}, ) precondition_dtype: torch.dtype = field( default=torch.float32, - metadata={"help": "Dtype for computing the preconditioned gradient. Recommended to use `torch.float32`."}, + metadata={"help": "Data type for preconditioned gradient computation."}, ) + score_dtype: torch.dtype = field( + default=torch.float32, + metadata={"help": "Data type for influence score computation."}, + ) + + def __post_init__(self) -> None: + if self.damping_factor is not None and self.damping_factor < 0: + raise ValueError("`damping_factor` must be `None` or positive.") + + if any(partition <= 0 for partition in [self.data_partitions, self.module_partitions]): + raise ValueError("Both data and module partitions must be positive.") + + if self.query_gradient_accumulation_steps <= 0: + raise ValueError("`query_gradient_accumulation_steps` must be positive.") + + if self.query_gradient_low_rank is not None and self.query_gradient_low_rank <= 0: + raise ValueError("`query_gradient_low_rank` must be `None` or positive.") diff --git a/kronfluence/computer/computer.py b/kronfluence/computer/computer.py index 9e29630..876e6f2 100644 --- a/kronfluence/computer/computer.py +++ b/kronfluence/computer/computer.py @@ -1,5 +1,6 @@ import logging import os +import time from abc import ABC from pathlib import Path from typing import Any, Dict, List, Optional, Sequence, Tuple, Union @@ -22,11 +23,23 @@ load_eigendecomposition, load_lambda_matrices, ) -from kronfluence.module.utils import get_tracked_module_names, make_modules_partition +from kronfluence.module.tracked_module import ModuleMode +from kronfluence.module.utils import ( + get_tracked_module_names, + make_modules_partition, + set_mode, +) from kronfluence.score.pairwise import load_pairwise_scores, pairwise_scores_exist from kronfluence.score.self import load_self_scores, self_scores_exist from kronfluence.task import Task -from kronfluence.utils.constants import FACTOR_TYPE, SCORE_TYPE +from kronfluence.utils.constants import ( + FACTOR_ARGUMENTS_NAME, + FACTOR_SAVE_PREFIX, + FACTOR_TYPE, + SCORE_ARGUMENTS_NAME, + SCORE_SAVE_PREFIX, + SCORE_TYPE, +) from kronfluence.utils.dataset import ( DataLoaderKwargs, DistributedEvalSampler, @@ -38,15 +51,8 @@ TrackedModuleNotFoundError, ) from kronfluence.utils.logger import PassThroughProfiler, Profiler, get_logger -from kronfluence.utils.save import ( - FACTOR_ARGUMENTS_NAME, - FACTOR_SAVE_PREFIX, - SCORE_ARGUMENTS_NAME, - SCORE_SAVE_PREFIX, - load_json, - save_json, -) -from kronfluence.utils.state import State +from kronfluence.utils.save import load_json, save_json +from kronfluence.utils.state import State, release_memory class Computer(ABC): @@ -62,8 +68,9 @@ def __init__( log_level: Optional[int] = logging.INFO, log_main_process_only: bool = True, profile: bool = False, + disable_tqdm: bool = False, ) -> None: - """Initializes an instance of the Computer class.""" + """Initializes an instance of the `Computer` class. See `Analyzer` for more information.""" self.state = State(cpu=cpu) # Create and configure logger. @@ -77,19 +84,17 @@ def __init__( if len(tracked_module_names) == 0: error_msg = ( f"No tracked modules found in the provided model: {self.model}. " - f"Please make sure to run `prepare_model` before passing it in to the " - f"Analyzer." + f"Please ensure you've run `prepare_model` before passing it to the Analyzer." ) self.logger.error(error_msg) raise TrackedModuleNotFoundError(error_msg) self.logger.info(f"Tracking modules with names: {tracked_module_names}.") if self.state.use_distributed and not isinstance(model, (DDP, FSDP)): - warning_msg = ( - "Creating a DDP module. If specific configuration needs to be used " - "for DDP, please pass in the model after the manual DDP wrapping." + self.logger.warning( + "Creating a DDP module. For custom DDP configuration, " + "please manually wrap the model with DDP before passing it in." ) - self.logger.warning(warning_msg) self.model.to(self.state.device) self.model = DDP( self.model, @@ -98,7 +103,10 @@ def __init__( ) if cpu and isinstance(model, (DataParallel, DDP, FSDP)): - error_msg = "To enforce CPU, the model should not be wrapped with DP, DDP, or FSDP." + error_msg = ( + "CPU enforcement is incompatible with DP, DDP, or FSDP wrapped models. " + "Please provide an unwrapped model when using `cpu=True`." + ) self.logger.error(error_msg) raise ValueError(error_msg) @@ -111,11 +119,9 @@ def __init__( # Create and configure profiler. self.profiler = Profiler(state=self.state) if profile else PassThroughProfiler(state=self.state) - # Create directory to save profiler output. - self.profiler_dir = (self.output_dir / "profiler_output").resolve() - os.makedirs(name=self.profiler_dir, exist_ok=True) + self.disable_tqdm = disable_tqdm - # DataLoader parameters. + # Set PyTorch DataLoader arguments. self._dataloader_params = DataLoaderKwargs() def factors_output_dir(self, factors_name: str) -> Path: @@ -140,11 +146,11 @@ def _save_arguments( loaded_arguments = load_json(arguments_save_path) if loaded_arguments != arguments.to_dict(): error_msg = ( - "Attempting to use the arguments that differs from the one already saved. " - "Please set `overwrite_output_dir=True` to overwrite existing experiment." + f"New arguments differ from saved arguments at `{arguments_save_path}`. " + "Set `overwrite_output_dir=True` to overwrite existing experiment.\n" + f"New arguments: {arguments.to_dict()}\n" + f"Saved arguments: {loaded_arguments}" ) - error_msg += f"\nNew arguments: {arguments.to_dict()}." - error_msg += f"\nSaved arguments: {loaded_arguments}." self.logger.error(error_msg) raise ValueError(error_msg) else: @@ -169,7 +175,7 @@ def _save_dataset_metadata( if dataset_metadata_save_path.exists() and not overwrite_output_dir: self.logger.info(f"Found existing saved dataset metadata at `{dataset_metadata_save_path}`.") - # Load the existing dataset metadata for comparison. + # Loads the existing dataset metadata for comparison. loaded_metadata = load_json(dataset_metadata_save_path) if loaded_metadata != dataset_metadata: error_msg = ( @@ -193,13 +199,13 @@ def _get_dataloader( allow_duplicates: bool = False, stack: bool = False, ) -> data.DataLoader: - """Returns the DataLoader for the given dataset, per_device_batch_size, and additional parameters.""" + """Returns the DataLoader with the provided configuration.""" if indices is not None: dataset = data.Subset(dataset=dataset, indices=indices) if self.state.use_distributed and not allow_duplicates: if stack: - error_msg = "DistributedEvalSampler is not currently supported with `stack=True`." + error_msg = "DistributedEvalSampler is incompatible with `stack=True`." self.logger.error(error_msg) raise ValueError(error_msg) sampler = DistributedEvalSampler( @@ -244,35 +250,31 @@ def _configure_dataloader(self, dataloader_kwargs: DataLoaderKwargs) -> Dict[str def _get_data_partition( self, total_data_examples: int, - data_partition_size: int, + data_partitions: int, target_data_partitions: Optional[Union[int, List[int]]], ) -> Tuple[List[Tuple[int, int]], List[int]]: """Partitions the dataset into several chunks.""" - if total_data_examples < data_partition_size: + if total_data_examples < data_partitions: error_msg = ( - f"Data partition size ({data_partition_size}) cannot be greater than the " - f"total data points ({total_data_examples}). Please reduce the data partition " - f"size in the argument." + f"Data partition size ({data_partitions}) exceeds total data points ({total_data_examples}). " + "Please reduce the data partition size." ) self.logger.error(error_msg) raise ValueError(error_msg) indices_partitions = make_indices_partition( - total_data_examples=total_data_examples, partition_size=data_partition_size + total_data_examples=total_data_examples, partition_size=data_partitions ) if target_data_partitions is None: - target_data_partitions = list(range(data_partition_size)) + target_data_partitions = list(range(data_partitions)) if isinstance(target_data_partitions, int): target_data_partitions = [target_data_partitions] for data_partition in target_data_partitions: - if data_partition < 0 or data_partition > data_partition_size: - error_msg = ( - f"Invalid data partition {data_partition} encountered. " - f"The module partition needs to be in between [0, {data_partition_size})." - ) + if data_partition < 0 or data_partition > data_partitions: + error_msg = f"Invalid data partition {data_partition}. Must be in range [0, {data_partitions})." self.logger.error(error_msg) raise ValueError(error_msg) @@ -280,48 +282,53 @@ def _get_data_partition( def _get_module_partition( self, - module_partition_size: int, + module_partitions: int, target_module_partitions: Optional[Union[int, List[int]]], ) -> Tuple[List[List[str]], List[int]]: """Partitions the modules into several chunks.""" tracked_module_names = get_tracked_module_names(self.model) - if len(tracked_module_names) < module_partition_size: + if len(tracked_module_names) < module_partitions: error_msg = ( - f"Module partition size ({module_partition_size}) cannot be greater than the " - f"total tracked modules ({len(tracked_module_names)}). Please reduce the module partition " - f"size in the argument." + f"Module partition size ({module_partitions}) exceeds total tracked modules " + f"({len(tracked_module_names)}). Please reduce the module partition size." ) self.logger.error(error_msg) raise ValueError(error_msg) modules_partition_list = make_modules_partition( total_module_names=tracked_module_names, - partition_size=module_partition_size, + partition_size=module_partitions, ) if target_module_partitions is None: - target_module_partitions = list(range(module_partition_size)) + target_module_partitions = list(range(module_partitions)) if isinstance(target_module_partitions, int): target_module_partitions = [target_module_partitions] for module_partition in target_module_partitions: - if module_partition < 0 or module_partition > module_partition_size: - error_msg = ( - f"Invalid module partition {module_partition} encountered. " - f"The module partition needs to be in between [0, {module_partition_size})." - ) + if module_partition < 0 or module_partition > module_partitions: + error_msg = f"Invalid module partition {module_partition}. Must be in range [0, {module_partitions})." self.logger.error(error_msg) raise ValueError(error_msg) return modules_partition_list, target_module_partitions - def _log_profile_summary(self) -> None: - """Log the summary of the profiling results.""" + def _reset_memory(self) -> None: + """Clears all cached memory.""" + self.model.zero_grad(set_to_none=True) + set_mode(model=self.model, mode=ModuleMode.DEFAULT, release_memory=True) + release_memory() + + def _log_profile_summary(self, name: str) -> None: + """Saves the summary of the profiling results.""" profile_summary = self.profiler.summary() - profile_save_path = (self.profiler_dir / f"summary_rank_{self.state.process_index}.txt").resolve() + time_str = time.strftime("%Y%m%d_%H%M%S") + profiler_dir = (self.output_dir / "profiler_output").resolve() + profile_save_path = (profiler_dir / f"{name}_summary_rank_{self.state.process_index}_{time_str}.txt").resolve() if profile_summary != "": + os.makedirs(name=profiler_dir, exist_ok=True) self.logger.info(profile_summary) with open(profile_save_path, "a", encoding="utf-8") as f: f.write(profile_summary) @@ -378,7 +385,7 @@ def load_self_scores(self, scores_name: str) -> Optional[SCORE_TYPE]: return None def load_all_factors(self, factors_name: str) -> FACTOR_TYPE: - """Loads factors from disk.""" + """Loads all relevant factors from disk.""" from kronfluence.factor.config import ( # pylint: disable=import-outside-toplevel FactorConfig, ) @@ -388,7 +395,7 @@ def load_all_factors(self, factors_name: str) -> FACTOR_TYPE: if factor_args is None: error_msg = f"Factors with name `{factors_name}` was not found at `{factors_output_dir}`." self.logger.error(error_msg) - raise ValueError(error_msg) + raise FileNotFoundError(error_msg) loaded_factors: FACTOR_TYPE = {} factor_config = FactorConfig.CONFIGS[factor_args.strategy] @@ -396,7 +403,7 @@ def load_all_factors(self, factors_name: str) -> FACTOR_TYPE: covariance_factors = self.load_covariance_matrices(factors_name=factors_name) if covariance_factors is None: error_msg = ( - f"Strategy `{factor_args.strategy}` computing covariance matrices. " + f"Strategy `{factor_args.strategy}` requires covariance matrices. " f"However, the covariance matrices were not found." ) self.logger.error(error_msg) @@ -407,7 +414,7 @@ def load_all_factors(self, factors_name: str) -> FACTOR_TYPE: eigen_factors = self.load_eigendecomposition(factors_name=factors_name) if eigen_factors is None: error_msg = ( - f"Strategy `{factor_args.strategy}` computing Eigendecomposition. " + f"Strategy `{factor_args.strategy}` requires Eigendecomposition results. " f"However, the Eigendecomposition results were not found." ) self.logger.error(error_msg) @@ -418,7 +425,7 @@ def load_all_factors(self, factors_name: str) -> FACTOR_TYPE: lambda_factors = self.load_lambda_matrices(factors_name=factors_name) if lambda_factors is None: error_msg = ( - f"Strategy `{factor_args.strategy}` computing Lambda matrices. " + f"Strategy `{factor_args.strategy}` requires Lambda matrices. " f"However, the Lambda matrices were not found." ) self.logger.error(error_msg) diff --git a/kronfluence/computer/factor_computer.py b/kronfluence/computer/factor_computer.py index a25fff7..5b7b1d4 100644 --- a/kronfluence/computer/factor_computer.py +++ b/kronfluence/computer/factor_computer.py @@ -25,14 +25,10 @@ save_eigendecomposition, save_lambda_matrices, ) -from kronfluence.module.tracked_module import ModuleMode -from kronfluence.module.utils import set_mode -from kronfluence.utils.constants import FACTOR_TYPE +from kronfluence.utils.constants import FACTOR_ARGUMENTS_NAME, FACTOR_TYPE from kronfluence.utils.dataset import DataLoaderKwargs, find_executable_batch_size from kronfluence.utils.exceptions import FactorsNotFoundError from kronfluence.utils.logger import get_time -from kronfluence.utils.save import FACTOR_ARGUMENTS_NAME -from kronfluence.utils.state import release_memory class FactorComputer(Computer): @@ -41,7 +37,7 @@ class FactorComputer(Computer): def _configure_and_save_factor_args( self, factor_args: Optional[FactorArguments], factors_output_dir: Path, overwrite_output_dir: bool ) -> FactorArguments: - """Configure the provided factor arguments and save it in disk.""" + """Configures and saves factor arguments to disk.""" if factor_args is None: factor_args = FactorArguments() self.logger.info(f"Factor arguments not provided. Using the default configuration: {factor_args}.") @@ -62,31 +58,31 @@ def _configure_and_save_factor_args( def _aggregate_factors( self, factors_name: str, - data_partition_size: int, - module_partition_size: int, - exists_fnc: Callable, + data_partitions: int, + module_partitions: int, + exist_fnc: Callable, load_fnc: Callable, save_fnc: Callable, ) -> Optional[FACTOR_TYPE]: """Aggregates factors computed for all data and module partitions.""" factors_output_dir = self.factors_output_dir(factors_name=factors_name) if not factors_output_dir.exists(): - error_msg = f"Factors directory `{factors_output_dir}` is not found when trying to aggregate factors." + error_msg = f"Factors directory `{factors_output_dir}` not found when trying to aggregate factors." self.logger.error(error_msg) raise FileNotFoundError(error_msg) - all_required_partitions = [(i, j) for i in range(data_partition_size) for j in range(module_partition_size)] + all_required_partitions = [(i, j) for i in range(data_partitions) for j in range(module_partitions)] all_partition_exists = all( - exists_fnc(output_dir=factors_output_dir, partition=partition) for partition in all_required_partitions + exist_fnc(output_dir=factors_output_dir, partition=partition) for partition in all_required_partitions ) if not all_partition_exists: - self.logger.warning("Factors are not aggregated as factors for some partitions are not yet computed.") + self.logger.info("Factors are not aggregated as factors for some partitions are not yet computed.") return start_time = time.time() aggregated_factors: FACTOR_TYPE = {} - for data_partition in range(data_partition_size): - for module_partition in range(module_partition_size): + for data_partition in range(data_partitions): + for module_partition in range(module_partitions): loaded_factors = load_fnc( output_dir=factors_output_dir, partition=(data_partition, module_partition), @@ -97,9 +93,11 @@ def _aggregate_factors( for module_name in factors: if module_name not in aggregated_factors[factor_name]: - aggregated_factors[factor_name][module_name] = factors[module_name] - else: - aggregated_factors[factor_name][module_name].add_(factors[module_name]) + aggregated_factors[factor_name][module_name] = torch.zeros_like( + factors[module_name], + requires_grad=False, + ) + aggregated_factors[factor_name][module_name].add_(factors[module_name]) del loaded_factors save_fnc( output_dir=factors_output_dir, @@ -121,7 +119,7 @@ def _find_executable_factors_batch_size( """Automatically finds executable batch size for performing `func`.""" if self.state.use_distributed: error_msg = ( - "Automatic batch size search is currently not supported for multi-GPU training. " + "Automatic batch size search is not supported for multi-GPU setting. " "Please manually configure the batch size by passing in `per_device_batch_size`." ) self.logger.error(error_msg) @@ -139,10 +137,8 @@ def _find_executable_factors_batch_size( def executable_batch_size_func(batch_size: int) -> None: self.logger.info(f"Attempting to set per-device batch size to {batch_size}.") - # Release all memory that could be caused by the previous OOM. - set_mode(model=self.model, mode=ModuleMode.DEFAULT, keep_factors=False) - self.model.zero_grad(set_to_none=True) - release_memory() + # Releases all memory that could be caused by the previous OOM. + self._reset_memory() total_batch_size = batch_size * self.state.num_processes loader = self._get_dataloader( dataset=dataset, @@ -178,7 +174,7 @@ def fit_covariance_matrices( factors_name (str): The unique identifier for the factor, used to organize and retrieve the results. dataset (data.Dataset): - The dataset that will be used to fit covariance matrices. + The dataset that will be used for fitting covariance matrices. per_device_batch_size (int, optional): The per-device batch size used to fit the factors. If not specified, executable batch size is automatically determined. @@ -187,8 +183,7 @@ def fit_covariance_matrices( dataloader_kwargs (DataLoaderKwargs, optional): Controls additional arguments for PyTorch's DataLoader. factor_args (FactorArguments, optional): - Arguments related to computing the factors. If not specified, the default values of - `FactorArguments` will be used. + Arguments for factor computation. target_data_partitions(Sequence[int], int, optional): The list of data partition to fit covariance matrices. By default, covariance matrices will be computed for all partitions. @@ -196,7 +191,7 @@ def fit_covariance_matrices( The list of module partition to fit covariance matrices. By default, covariance matrices will be computed for all partitions. overwrite_output_dir (bool, optional): - If True, the existing factors with the same `factors_name` will be overwritten. + Whether to overwrite existing output. """ self.logger.debug(f"Fitting covariance matrices with parameters: {locals()}") @@ -231,9 +226,7 @@ def fit_covariance_matrices( total_data_examples = min([factor_args.covariance_max_examples, len(dataset)]) self.logger.info(f"Total data examples to fit covariance matrices: {total_data_examples}.") - no_partition = ( - factor_args.covariance_data_partition_size == 1 and factor_args.covariance_module_partition_size == 1 - ) + no_partition = factor_args.covariance_data_partitions == 1 and factor_args.covariance_module_partitions == 1 partition_provided = target_data_partitions is not None or target_module_partitions is not None if no_partition and partition_provided: error_msg = ( @@ -245,17 +238,17 @@ def fit_covariance_matrices( data_partition_indices, target_data_partitions = self._get_data_partition( total_data_examples=total_data_examples, - data_partition_size=factor_args.covariance_data_partition_size, + data_partitions=factor_args.covariance_data_partitions, target_data_partitions=target_data_partitions, ) - max_partition_examples = total_data_examples // factor_args.covariance_data_partition_size + max_partition_examples = total_data_examples // factor_args.covariance_data_partitions module_partition_names, target_module_partitions = self._get_module_partition( - module_partition_size=factor_args.covariance_module_partition_size, + module_partitions=factor_args.covariance_module_partitions, target_module_partitions=target_module_partitions, ) if max_partition_examples < self.state.num_processes: - error_msg = "The number of processes are more than the data examples. Try reducing the number of processes." + error_msg = "The number of processes are larger than total data examples. Try reducing number of processes." self.logger.error(error_msg) raise ValueError(error_msg) @@ -293,6 +286,7 @@ def fit_covariance_matrices( "task": self.task, "factor_args": factor_args, "tracked_module_names": module_partition_names[module_partition], + "disable_tqdm": True, } per_device_batch_size = self._find_executable_factors_batch_size( func=fit_covariance_matrices_with_loader, @@ -303,7 +297,7 @@ def fit_covariance_matrices( total_data_examples=max_partition_examples, ) - release_memory() + self._reset_memory() start_time = get_time(state=self.state) with self.profiler.profile("Fit Covariance"): loader = self._get_dataloader( @@ -320,6 +314,7 @@ def fit_covariance_matrices( loader=loader, factor_args=factor_args, tracked_module_names=module_partition_names[module_partition], + disable_tqdm=self.disable_tqdm, ) end_time = get_time(state=self.state) elapsed_time = end_time - start_time @@ -337,8 +332,9 @@ def fit_covariance_matrices( metadata=factor_args.to_str_dict(), ) self.state.wait_for_everyone() - del covariance_factors, loader self.logger.info(f"Saved covariance matrices at `{factors_output_dir}`.") + del num_data_processed, covariance_factors, loader + self._reset_memory() all_end_time = get_time(state=self.state) elapsed_time = all_end_time - all_start_time @@ -348,7 +344,7 @@ def fit_covariance_matrices( self.aggregate_covariance_matrices(factors_name=factors_name) self.logger.info(f"Saved aggregated covariance matrices at `{factors_output_dir}`.") self.state.wait_for_everyone() - self._log_profile_summary() + self._log_profile_summary(name=f"factors_{factors_name}_covariance") @torch.no_grad() def aggregate_covariance_matrices( @@ -366,7 +362,7 @@ def aggregate_covariance_matrices( if factor_args is None: error_msg = ( f"Arguments for factors with name `{factors_name}` was not found when trying to " - f"aggregated covariance matrices." + f"aggregate covariance matrices." ) self.logger.error(error_msg) raise ValueError(error_msg) @@ -374,9 +370,9 @@ def aggregate_covariance_matrices( with self.profiler.profile("Aggregate Covariance"): self._aggregate_factors( factors_name=factors_name, - data_partition_size=factor_args.covariance_data_partition_size, - module_partition_size=factor_args.covariance_module_partition_size, - exists_fnc=covariance_matrices_exist, + data_partitions=factor_args.covariance_data_partitions, + module_partitions=factor_args.covariance_module_partitions, + exist_fnc=covariance_matrices_exist, load_fnc=load_covariance_matrices, save_fnc=save_covariance_matrices, ) @@ -388,26 +384,25 @@ def perform_eigendecomposition( overwrite_output_dir: bool = False, load_from_factors_name: Optional[str] = None, ) -> None: - """Performs Eigendecomposition on all covariance matrices. + """Performs eigendecomposition on all covariance matrices. Args: factors_name (str): The unique identifier for the factor, used to organize and retrieve the results. factor_args (FactorArguments, optional): - Arguments related to computing the factors. If not specified, the default values of - `FactorArguments` will be used. + Arguments for factor computation. overwrite_output_dir (bool, optional): - If True, the existing factors with the same `factors_name` will be overwritten. + Whether to overwrite existing output. load_from_factors_name (str, optional): The `factor_name` to load covariance matrices from. By default, covariance matrices with the same `factor_name` will be used. """ - self.logger.debug(f"Performing Eigendecomposition with parameters: {locals()}") + self.logger.debug(f"Performing eigendecomposition with parameters: {locals()}") factors_output_dir = self.factors_output_dir(factors_name=factors_name) os.makedirs(factors_output_dir, exist_ok=True) if eigendecomposition_exist(output_dir=factors_output_dir) and not overwrite_output_dir: - self.logger.info(f"Found existing Eigendecomposition results at `{factors_output_dir}`. Skipping.") + self.logger.info(f"Found existing eigendecomposition results at `{factors_output_dir}`. Skipping.") return factor_args = self._configure_and_save_factor_args( @@ -416,7 +411,7 @@ def perform_eigendecomposition( if not FactorConfig.CONFIGS[factor_args.strategy].requires_eigendecomposition: self.logger.info( - f"Strategy `{factor_args.strategy}` does not require performing Eigendecomposition. Skipping." + f"Strategy `{factor_args.strategy}` does not require performing eigendecomposition. Skipping." ) return None @@ -428,7 +423,7 @@ def perform_eigendecomposition( if not covariance_matrices_exist(output_dir=load_factors_output_dir): error_msg = ( f"Covariance matrices not found at `{load_factors_output_dir}`. " - f"To perform Eigendecomposition, covariance matrices need to be first computed." + f"To perform eigendecomposition, covariance matrices need to be first computed." ) self.logger.error(error_msg) raise FactorsNotFoundError(error_msg) @@ -449,9 +444,9 @@ def perform_eigendecomposition( ) self.state.wait_for_everyone() + self._reset_memory() eigen_factors = None if self.state.is_main_process: - release_memory() start_time = time.time() with self.profiler.profile("Perform Eigendecomposition"): eigen_factors = perform_eigendecomposition( @@ -459,18 +454,21 @@ def perform_eigendecomposition( model=self.model, state=self.state, factor_args=factor_args, + disable_tqdm=self.disable_tqdm, ) end_time = time.time() elapsed_time = end_time - start_time - self.logger.info(f"Performed Eigendecomposition in {elapsed_time:.2f} seconds.") + self.logger.info(f"Performed eigendecomposition in {elapsed_time:.2f} seconds.") with self.profiler.profile("Save Eigendecomposition"): save_eigendecomposition( output_dir=factors_output_dir, factors=eigen_factors, metadata=factor_args.to_str_dict() ) - self.logger.info(f"Saved Eigendecomposition results at `{factors_output_dir}`.") + self.logger.info(f"Saved eigendecomposition results at `{factors_output_dir}`.") + del eigen_factors + self._reset_memory() self.state.wait_for_everyone() - self._log_profile_summary() + self._log_profile_summary(name=f"factors_{factors_name}_eigendecomposition") def fit_lambda_matrices( self, @@ -491,7 +489,7 @@ def fit_lambda_matrices( factors_name (str): The unique identifier for the factor, used to organize and retrieve the results. dataset (data.Dataset): - The dataset that will be used to fit Lambda matrices. + The dataset that will be used for fitting Lambda matrices. per_device_batch_size (int, optional): The per-device batch size used to fit the factors. If not specified, executable batch size is automatically determined. @@ -500,8 +498,7 @@ def fit_lambda_matrices( dataloader_kwargs (DataLoaderKwargs, optional): Controls additional arguments for PyTorch's DataLoader. factor_args (FactorArguments, optional): - Arguments related to computing the factors. If not specified, the default values of - `FactorArguments` will be used. + Arguments for factor computation. target_data_partitions(Sequence[int], int, optional): The list of data partition to fit Lambda matrices. By default, Lambda matrices will be computed for all partitions. @@ -509,9 +506,9 @@ def fit_lambda_matrices( The list of module partition to fit Lambda matrices. By default, Lambda matrices will be computed for all partitions. overwrite_output_dir (bool, optional): - If True, the existing factors with the same `factors_name` will be overwritten. + Whether to overwrite existing output. load_from_factors_name (str, optional): - The `factor_name` to load Eigendecomposition results from. By default, Eigendecomposition + The `factor_name` to load eigendecomposition results from. By default, eigendecomposition results with the same `factor_name` will be used. """ self.logger.debug(f"Fitting Lambda matrices with parameters: {locals()}") @@ -541,7 +538,7 @@ def fit_lambda_matrices( if load_from_factors_name is not None: self.logger.info( - f"Will be loading Eigendecomposition results from factors with name `{load_from_factors_name}`." + f"Will be loading eigendecomposition results from factors with name `{load_from_factors_name}`." ) load_factors_output_dir = self.factors_output_dir(factors_name=load_from_factors_name) else: @@ -553,7 +550,7 @@ def fit_lambda_matrices( ): error_msg = ( f"Eigendecomposition results not found at `{load_factors_output_dir}`. " - f"To fit Lambda matrices for `{factor_args.strategy}`, Eigendecomposition must be " + f"To fit Lambda matrices for `{factor_args.strategy}`, eigendecomposition must be " f"performed before computing Lambda matrices." ) self.logger.error(error_msg) @@ -581,7 +578,7 @@ def fit_lambda_matrices( total_data_examples = min([factor_args.lambda_max_examples, len(dataset)]) self.logger.info(f"Total data examples to fit Lambda matrices: {total_data_examples}.") - no_partition = factor_args.lambda_data_partition_size == 1 and factor_args.lambda_module_partition_size == 1 + no_partition = factor_args.lambda_data_partitions == 1 and factor_args.lambda_module_partitions == 1 partition_provided = target_data_partitions is not None or target_module_partitions is not None if no_partition and partition_provided: error_msg = ( @@ -593,17 +590,17 @@ def fit_lambda_matrices( data_partition_indices, target_data_partitions = self._get_data_partition( total_data_examples=total_data_examples, - data_partition_size=factor_args.lambda_data_partition_size, + data_partitions=factor_args.lambda_data_partitions, target_data_partitions=target_data_partitions, ) - max_partition_examples = total_data_examples // factor_args.lambda_data_partition_size + max_partition_examples = total_data_examples // factor_args.lambda_data_partitions module_partition_names, target_module_partitions = self._get_module_partition( - module_partition_size=factor_args.lambda_module_partition_size, + module_partitions=factor_args.lambda_module_partitions, target_module_partitions=target_module_partitions, ) if max_partition_examples < self.state.num_processes: - error_msg = "The number of processes are more than the data examples. Try reducing the number of processes." + error_msg = "The number of processes are larger than total data examples. Try reducing number of processes." self.logger.error(error_msg) raise ValueError(error_msg) @@ -642,6 +639,7 @@ def fit_lambda_matrices( "task": self.task, "factor_args": factor_args, "tracked_module_names": module_partition_names[module_partition], + "disable_tqdm": True, } per_device_batch_size = self._find_executable_factors_batch_size( func=fit_lambda_matrices_with_loader, @@ -652,7 +650,7 @@ def fit_lambda_matrices( total_data_examples=max_partition_examples, ) - release_memory() + self._reset_memory() start_time = get_time(state=self.state) with self.profiler.profile("Fit Lambda"): loader = self._get_dataloader( @@ -670,6 +668,7 @@ def fit_lambda_matrices( loader=loader, factor_args=factor_args, tracked_module_names=module_partition_names[module_partition], + disable_tqdm=self.disable_tqdm, ) end_time = get_time(state=self.state) elapsed_time = end_time - start_time @@ -687,8 +686,9 @@ def fit_lambda_matrices( metadata=factor_args.to_str_dict(), ) self.state.wait_for_everyone() - del lambda_factors, loader self.logger.info(f"Saved Lambda matrices at `{factors_output_dir}`.") + del num_data_processed, lambda_factors, loader + self._reset_memory() all_end_time = get_time(state=self.state) elapsed_time = all_end_time - all_start_time @@ -698,7 +698,7 @@ def fit_lambda_matrices( self.aggregate_lambda_matrices(factors_name=factors_name) self.logger.info(f"Saved aggregated Lambda matrices at `{factors_output_dir}`.") self.state.wait_for_everyone() - self._log_profile_summary() + self._log_profile_summary(name=f"factors_{factors_name}_lambda") @torch.no_grad() def aggregate_lambda_matrices( @@ -716,7 +716,7 @@ def aggregate_lambda_matrices( if factor_args is None: error_msg = ( f"Arguments for factors with name `{factors_name}` was not found when trying " - f"to aggregated Lambda matrices." + f"to aggregate Lambda matrices." ) self.logger.error(error_msg) raise ValueError(error_msg) @@ -724,9 +724,9 @@ def aggregate_lambda_matrices( with self.profiler.profile("Aggregate Lambda"): self._aggregate_factors( factors_name=factors_name, - data_partition_size=factor_args.lambda_data_partition_size, - module_partition_size=factor_args.lambda_module_partition_size, - exists_fnc=lambda_matrices_exist, + data_partitions=factor_args.lambda_data_partitions, + module_partitions=factor_args.lambda_module_partitions, + exist_fnc=lambda_matrices_exist, load_fnc=load_lambda_matrices, save_fnc=save_lambda_matrices, ) diff --git a/kronfluence/computer/score_computer.py b/kronfluence/computer/score_computer.py index fb24da9..d606124 100644 --- a/kronfluence/computer/score_computer.py +++ b/kronfluence/computer/score_computer.py @@ -8,9 +8,8 @@ from kronfluence.arguments import FactorArguments, ScoreArguments from kronfluence.computer.computer import Computer -from kronfluence.module.tracked_module import ModuleMode -from kronfluence.module.utils import set_mode from kronfluence.score.pairwise import ( + compute_pairwise_query_aggregated_scores_with_loaders, compute_pairwise_scores_with_loaders, load_pairwise_scores, pairwise_scores_exist, @@ -23,12 +22,15 @@ save_self_scores, self_scores_exist, ) -from kronfluence.utils.constants import FACTOR_TYPE, SCORE_TYPE +from kronfluence.utils.constants import ( + FACTOR_ARGUMENTS_NAME, + FACTOR_TYPE, + SCORE_ARGUMENTS_NAME, + SCORE_TYPE, +) from kronfluence.utils.dataset import DataLoaderKwargs, find_executable_batch_size from kronfluence.utils.exceptions import FactorsNotFoundError from kronfluence.utils.logger import get_time -from kronfluence.utils.save import FACTOR_ARGUMENTS_NAME, SCORE_ARGUMENTS_NAME -from kronfluence.utils.state import release_memory class ScoreComputer(Computer): @@ -41,7 +43,7 @@ def _configure_and_save_score_args( factors_name: str, overwrite_output_dir: bool, ) -> Tuple[FactorArguments, ScoreArguments]: - """Configure the provided factor arguments and save it in disk.""" + """Configures and saves score arguments to disk.""" if score_args is None: score_args = ScoreArguments() self.logger.info(f"Score arguments not provided. Using the default configuration: {score_args}.") @@ -51,7 +53,7 @@ def _configure_and_save_score_args( factor_args = self.load_factor_args(factors_name=factors_name) factors_output_dir = self.factors_output_dir(factors_name=factors_name) if factor_args is None: - error_msg = f"Factors with name `{factors_name}` was not found at `{factors_output_dir}`." + error_msg = f"Factors with name `{factors_name}` not found at `{factors_output_dir}`." self.logger.error(error_msg) raise FactorsNotFoundError(error_msg) self.logger.info(f"Loaded `FactorArguments` with configuration: {factor_args}.") @@ -77,7 +79,7 @@ def _aggregate_scores( self, scores_name: str, score_args: ScoreArguments, - exists_fnc: Callable, + exist_fnc: Callable, load_fnc: Callable, save_fnc: Callable, dim: int, @@ -85,47 +87,43 @@ def _aggregate_scores( """Aggregates influence scores computed for all data and module partitions.""" scores_output_dir = self.scores_output_dir(scores_name=scores_name) if not scores_output_dir.exists(): - error_msg = ( - f"Scores output directory `{scores_output_dir}` is not found " - f"when trying to aggregate partitioned scores." - ) + error_msg = f"Scores directory `{scores_output_dir}` not found when trying to aggregate scores." self.logger.error(error_msg) raise FileNotFoundError(error_msg) - data_partition_size = score_args.data_partition_size - module_partition_size = score_args.module_partition_size all_required_partitions = [ - (i, j) for i in range(score_args.data_partition_size) for j in range(score_args.module_partition_size) + (i, j) for i in range(score_args.data_partitions) for j in range(score_args.module_partitions) ] all_partition_exists = all( - exists_fnc(output_dir=scores_output_dir, partition=partition) for partition in all_required_partitions + exist_fnc(output_dir=scores_output_dir, partition=partition) for partition in all_required_partitions ) if not all_partition_exists: - self.logger.info("Influence scores are not aggregated as scores for some partitions are not yet computed.") + self.logger.info("Scores are not aggregated as scores for some partitions are not yet computed.") return start_time = time.time() aggregated_scores: SCORE_TYPE = {} - with self.profiler.profile("Aggregate Score"): - for data_partition in range(data_partition_size): - aggregated_module_scores = {} + for data_partition in range(score_args.data_partitions): + aggregated_module_scores = {} - for module_partition in range(module_partition_size): - loaded_scores = load_fnc( - output_dir=scores_output_dir, - partition=(data_partition, module_partition), - ) + for module_partition in range(score_args.module_partitions): + loaded_scores = load_fnc( + output_dir=scores_output_dir, + partition=(data_partition, module_partition), + ) - for module_name, scores in loaded_scores.items(): - if module_name not in aggregated_module_scores: - aggregated_module_scores[module_name] = scores - else: - aggregated_module_scores[module_name].add_(scores) - del loaded_scores + for module_name, scores in loaded_scores.items(): + if module_name not in aggregated_module_scores: + aggregated_module_scores[module_name] = torch.zeros_like(scores, requires_grad=False) + aggregated_module_scores[module_name].add_(scores) + del loaded_scores - for module_name, scores in aggregated_module_scores.items(): - if module_name not in aggregated_scores: - aggregated_scores[module_name] = scores + for module_name, scores in aggregated_module_scores.items(): + if module_name not in aggregated_scores: + aggregated_scores[module_name] = scores.clone() + else: + if score_args.aggregate_train_gradients: + aggregated_scores[module_name].add_(scores) else: aggregated_scores[module_name] = torch.cat( ( @@ -134,10 +132,10 @@ def _aggregate_scores( ), dim=dim, ) - save_fnc(output_dir=scores_output_dir, scores=aggregated_scores, metadata=score_args.to_str_dict()) + save_fnc(output_dir=scores_output_dir, scores=aggregated_scores, metadata=score_args.to_str_dict()) end_time = time.time() elapsed_time = end_time - start_time - self.logger.info(f"Aggregated all partitioned scores in {elapsed_time:.2f} seconds.") + self.logger.info(f"Aggregated all scores in {elapsed_time:.2f} seconds.") return aggregated_scores def _find_executable_pairwise_scores_batch_size( @@ -156,8 +154,8 @@ def _find_executable_pairwise_scores_batch_size( """Automatically finds executable training batch size for computing pairwise influence scores.""" if self.state.use_distributed: error_msg = ( - "Automatic batch size search is currently not supported for multi-GPU training. " - "Please manually configure the batch size by passing in `per_device_train_batch_size`." + "Automatic batch size search is not supported for multi-GPU setting. " + "Please manually configure the batch size by passing in `per_device_batch_size`." ) self.logger.error(error_msg) raise NotImplementedError(error_msg) @@ -173,9 +171,8 @@ def _find_executable_pairwise_scores_batch_size( def executable_batch_size_func(batch_size: int) -> None: self.logger.info(f"Attempting to set per-device batch size to {batch_size}.") - # Release all memory that could be caused by the previous OOM. - set_mode(model=self.model, mode=ModuleMode.DEFAULT, keep_factors=False) - release_memory() + # Releases all memory that could be caused by the previous OOM. + self._reset_memory() total_batch_size = batch_size * self.state.num_processes query_loader = self._get_dataloader( dataset=query_dataset, @@ -192,7 +189,12 @@ def executable_batch_size_func(batch_size: int) -> None: allow_duplicates=True, stack=True, ) - compute_pairwise_scores_with_loaders( + func = ( + compute_pairwise_scores_with_loaders + if not score_args.aggregate_query_gradients + else compute_pairwise_query_aggregated_scores_with_loaders + ) + func( model=self.model, state=self.state, task=self.task, @@ -203,6 +205,7 @@ def executable_batch_size_func(batch_size: int) -> None: train_loader=train_loader, per_device_query_batch_size=per_device_query_batch_size, tracked_module_names=tracked_modules_name, + disable_tqdm=True, ) per_device_batch_size = find_executable_batch_size( @@ -229,9 +232,7 @@ def compute_pairwise_scores( target_module_partitions: Optional[Sequence[int]] = None, overwrite_output_dir: bool = False, ) -> Optional[SCORE_TYPE]: - """Computes pairwise influence scores for the given score configuration. As an example, - for Q query dataset and T training dataset, the pairwise influence scores are - 2-dimensional matrix with dimension `Q x T`. + """Computes pairwise influence scores with the given score configuration. Args: scores_name (str): @@ -258,8 +259,7 @@ def compute_pairwise_scores( dataloader_kwargs (DataLoaderKwargs, optional): Controls additional arguments for PyTorch's DataLoader. score_args (ScoreArguments, optional): - Arguments related to computing the pairwise scores. If not specified, the default values - of `ScoreArguments` will be used. + Arguments for score computation. target_data_partitions (Sequence[int], optional): Specific data partitions to compute influence scores. If not specified, scores for all data partitions will be computed. @@ -267,7 +267,7 @@ def compute_pairwise_scores( Specific module partitions to compute influence scores. If not specified, scores for all module partitions will be computed. overwrite_output_dir (bool, optional): - If True, the existing factors with the same name will be overwritten. + Whether to overwrite existing output. """ self.logger.debug(f"Computing pairwise scores with parameters: {locals()}") @@ -284,6 +284,30 @@ def compute_pairwise_scores( overwrite_output_dir=overwrite_output_dir, ) + if score_args.compute_per_token_scores and score_args.aggregate_train_gradients: + warning_msg = ( + "Token-wise influence computation is not compatible with `aggregate_train_gradients=True`. " + "Disabling `compute_per_token_scores`." + ) + score_args.compute_per_token_scores = False + self.logger.warning(warning_msg) + + if score_args.compute_per_token_scores and factor_args.has_shared_parameters: + warning_msg = ( + "Token-wise influence computation is not compatible with `has_shared_parameters=True`. " + "Disabling `compute_per_token_scores`." + ) + score_args.compute_per_token_scores = False + self.logger.warning(warning_msg) + + if score_args.compute_per_token_scores and self.task.enable_post_process_per_sample_gradient: + warning_msg = ( + "Token-wise influence computation is not compatible with tasks that requires " + "`enable_post_process_per_sample_gradient`. Disabling `compute_per_token_scores`." + ) + score_args.compute_per_token_scores = False + self.logger.warning(warning_msg) + dataloader_params = self._configure_dataloader(dataloader_kwargs) if self.state.is_main_process: self._save_dataset_metadata( @@ -313,7 +337,7 @@ def compute_pairwise_scores( factors_name=factors_name, ) - no_partition = score_args.data_partition_size == 1 and score_args.module_partition_size == 1 + no_partition = score_args.data_partitions == 1 and score_args.module_partitions == 1 partition_provided = target_data_partitions is not None or target_module_partitions is not None if no_partition and partition_provided: error_msg = ( @@ -325,12 +349,12 @@ def compute_pairwise_scores( data_partition_indices, target_data_partitions = self._get_data_partition( total_data_examples=len(train_dataset), - data_partition_size=score_args.data_partition_size, + data_partitions=score_args.data_partitions, target_data_partitions=target_data_partitions, ) - max_partition_examples = len(train_dataset) // score_args.data_partition_size + max_partition_examples = len(train_dataset) // score_args.data_partitions module_partition_names, target_module_partitions = self._get_module_partition( - module_partition_size=score_args.module_partition_size, + module_partitions=score_args.module_partitions, target_module_partitions=target_module_partitions, ) @@ -357,14 +381,16 @@ def compute_pairwise_scores( start_index, end_index = data_partition_indices[data_partition] self.logger.info( - f"Fitting pairwise scores with data indices ({start_index}, {end_index}) and " + f"Computing pairwise scores with data indices ({start_index}, {end_index}) and " f"modules {module_partition_names[module_partition]}." ) if per_device_train_batch_size is None: per_device_train_batch_size = self._find_executable_pairwise_scores_batch_size( query_dataset=query_dataset, - per_device_query_batch_size=per_device_query_batch_size, + per_device_query_batch_size=per_device_query_batch_size + if not score_args.aggregate_query_gradients + else 1, train_dataset=train_dataset, initial_per_device_train_batch_size_attempt=initial_per_device_train_batch_size_attempt, loaded_factors=loaded_factors, @@ -375,24 +401,29 @@ def compute_pairwise_scores( tracked_modules_name=module_partition_names[module_partition], ) - release_memory() + self._reset_memory() start_time = get_time(state=self.state) with self.profiler.profile("Compute Pairwise Score"): query_loader = self._get_dataloader( dataset=query_dataset, per_device_batch_size=per_device_query_batch_size, dataloader_params=dataloader_params, - allow_duplicates=True, + allow_duplicates=not score_args.aggregate_query_gradients, ) train_loader = self._get_dataloader( dataset=train_dataset, per_device_batch_size=per_device_train_batch_size, indices=list(range(start_index, end_index)), dataloader_params=dataloader_params, - allow_duplicates=True, - stack=True, + allow_duplicates=not score_args.aggregate_train_gradients, + stack=not score_args.aggregate_train_gradients, ) - scores = compute_pairwise_scores_with_loaders( + func = ( + compute_pairwise_scores_with_loaders + if not score_args.aggregate_query_gradients + else compute_pairwise_query_aggregated_scores_with_loaders + ) + scores = func( model=self.model, state=self.state, task=self.task, @@ -403,6 +434,7 @@ def compute_pairwise_scores( score_args=score_args, factor_args=factor_args, tracked_module_names=module_partition_names[module_partition], + disable_tqdm=self.disable_tqdm, ) end_time = get_time(state=self.state) elapsed_time = end_time - start_time @@ -418,6 +450,7 @@ def compute_pairwise_scores( ) self.state.wait_for_everyone() del scores, query_loader, train_loader + self._reset_memory() self.logger.info(f"Saved pairwise scores at {scores_output_dir}.") all_end_time = get_time(state=self.state) @@ -428,7 +461,7 @@ def compute_pairwise_scores( self.aggregate_pairwise_scores(scores_name=scores_name) self.logger.info(f"Saved aggregated pairwise scores at `{scores_output_dir}`.") self.state.wait_for_everyone() - self._log_profile_summary() + self._log_profile_summary(name=f"scores_{scores_name}_pairwise") @torch.no_grad() def aggregate_pairwise_scores(self, scores_name: str) -> None: @@ -443,19 +476,20 @@ def aggregate_pairwise_scores(self, scores_name: str) -> None: if score_args is None: error_msg = ( f"Arguments for scores with name `{score_args}` was not found when trying " - f"to aggregated pairwise influence scores." + f"to aggregate pairwise influence scores." ) self.logger.error(error_msg) raise ValueError(error_msg) - self._aggregate_scores( - scores_name=scores_name, - score_args=score_args, - exists_fnc=pairwise_scores_exist, - load_fnc=load_pairwise_scores, - save_fnc=save_pairwise_scores, - dim=1, - ) + with self.profiler.profile("Aggregate Score"): + self._aggregate_scores( + scores_name=scores_name, + score_args=score_args, + exist_fnc=pairwise_scores_exist, + load_fnc=load_pairwise_scores, + save_fnc=save_pairwise_scores, + dim=1, + ) def _find_executable_self_scores_batch_size( self, @@ -471,8 +505,8 @@ def _find_executable_self_scores_batch_size( """Automatically finds executable training batch size for computing self-influence scores.""" if self.state.use_distributed: error_msg = ( - "Automatic batch size search is currently not supported for multi-GPU training. " - "Please manually configure the batch size by passing in `per_device_train_batch_size`." + "Automatic batch size search is not supported for multi-GPU setting. " + "Please manually configure the batch size by passing in `per_device_batch_size`." ) self.logger.error(error_msg) raise NotImplementedError(error_msg) @@ -487,9 +521,8 @@ def _find_executable_self_scores_batch_size( def executable_batch_size_func(batch_size: int) -> None: self.logger.info(f"Attempting to set per-device batch size to {batch_size}.") - # Release all memory that could be caused by the previous OOM. - set_mode(model=self.model, mode=ModuleMode.DEFAULT, keep_factors=False) - release_memory() + # Releases all memory that could be caused by the previous OOM. + self._reset_memory() total_batch_size = batch_size * self.state.num_processes train_loader = self._get_dataloader( dataset=train_dataset, @@ -512,6 +545,7 @@ def executable_batch_size_func(batch_size: int) -> None: score_args=score_args, factor_args=factor_args, tracked_module_names=tracked_modules_name, + disable_tqdm=True, ) per_device_batch_size = find_executable_batch_size( @@ -535,8 +569,7 @@ def compute_self_scores( target_module_partitions: Optional[Sequence[int]] = None, overwrite_output_dir: bool = False, ) -> Optional[SCORE_TYPE]: - """Computes self-influence scores for the given score configuration. As an example, - for T training dataset, the self-influence scores are represented as T-dimensional vector. + """Computes self-influence scores with the given score configuration. Args: scores_name (str): @@ -556,8 +589,7 @@ def compute_self_scores( dataloader_kwargs (DataLoaderKwargs, optional): Controls additional arguments for PyTorch's DataLoader. score_args (ScoreArguments, optional): - Arguments related to computing the self-influence scores. If not specified, the default values - of `ScoreArguments` will be used. + Arguments for score computation. target_data_partitions (Sequence[int], optional): Specific data partitions to compute influence scores. If not specified, scores for all data partitions will be computed. @@ -565,7 +597,7 @@ def compute_self_scores( Specific module partitions to compute influence scores. If not specified, scores for all module partitions will be computed. overwrite_output_dir (bool, optional): - If True, the existing factors with the same name will be overwritten. + Whether to overwrite existing output. """ self.logger.debug(f"Computing self-influence scores with parameters: {locals()}") @@ -582,11 +614,28 @@ def compute_self_scores( overwrite_output_dir=overwrite_output_dir, ) - if score_args.query_gradient_rank is not None: + if score_args.query_gradient_accumulation_steps != 1: + warning_msg = "Query gradient accumulation is not supported for self-influence computation." + score_args.query_gradient_accumulation_steps = 1 + self.logger.warning(warning_msg) + + if score_args.query_gradient_low_rank is not None: warning_msg = ( "Low rank query gradient approximation is not supported for self-influence computation. " "No low rank query approximation will be performed." ) + score_args.query_gradient_low_rank = None + self.logger.warning(warning_msg) + + if score_args.aggregate_query_gradients or score_args.aggregate_train_gradients: + warning_msg = "Query or train gradient aggregation is not supported for self-influence computation." + score_args.aggregate_train_gradients = False + score_args.aggregate_query_gradients = False + self.logger.warning(warning_msg) + + if score_args.compute_per_token_scores: + warning_msg = "Token-wise influence computation is not compatible with self-influence scores. " + score_args.compute_per_token_scores = False self.logger.warning(warning_msg) dataloader_params = self._configure_dataloader(dataloader_kwargs) @@ -607,7 +656,7 @@ def compute_self_scores( factors_name=factors_name, ) - no_partition = score_args.data_partition_size == 1 and score_args.module_partition_size == 1 + no_partition = score_args.data_partitions == 1 and score_args.module_partitions == 1 partition_provided = target_data_partitions is not None or target_module_partitions is not None if no_partition and partition_provided: error_msg = ( @@ -619,12 +668,12 @@ def compute_self_scores( data_partition_indices, target_data_partitions = self._get_data_partition( total_data_examples=len(train_dataset), - data_partition_size=score_args.data_partition_size, + data_partitions=score_args.data_partitions, target_data_partitions=target_data_partitions, ) - max_partition_examples = len(train_dataset) // score_args.data_partition_size + max_partition_examples = len(train_dataset) // score_args.data_partitions module_partition_names, target_module_partitions = self._get_module_partition( - module_partition_size=score_args.module_partition_size, + module_partitions=score_args.module_partitions, target_module_partitions=target_module_partitions, ) @@ -651,7 +700,7 @@ def compute_self_scores( start_index, end_index = data_partition_indices[data_partition] self.logger.info( - f"Fitting self-influence scores with data indices ({start_index}, {end_index}) and " + f"Computing self-influence scores with data indices ({start_index}, {end_index}) and " f"modules {module_partition_names[module_partition]}." ) @@ -667,7 +716,7 @@ def compute_self_scores( tracked_modules_name=module_partition_names[module_partition], ) - release_memory() + self._reset_memory() start_time = get_time(state=self.state) with self.profiler.profile("Compute Self-Influence Score"): train_loader = self._get_dataloader( @@ -691,6 +740,7 @@ def compute_self_scores( score_args=score_args, factor_args=factor_args, tracked_module_names=module_partition_names[module_partition], + disable_tqdm=self.disable_tqdm, ) end_time = get_time(state=self.state) elapsed_time = end_time - start_time @@ -706,6 +756,7 @@ def compute_self_scores( ) self.state.wait_for_everyone() del scores, train_loader + self._reset_memory() self.logger.info(f"Saved self-influence scores at `{scores_output_dir}`.") all_end_time = get_time(state=self.state) @@ -716,7 +767,7 @@ def compute_self_scores( self.aggregate_self_scores(scores_name=scores_name) self.logger.info(f"Saved aggregated self-influence scores at `{scores_output_dir}`.") self.state.wait_for_everyone() - self._log_profile_summary() + self._log_profile_summary(name=f"scores_{scores_name}_self") @torch.no_grad() def aggregate_self_scores(self, scores_name: str) -> None: @@ -731,15 +782,16 @@ def aggregate_self_scores(self, scores_name: str) -> None: if score_args is None: error_msg = ( f"Arguments for scores with name `{score_args}` was not found when trying " - f"to aggregated self-influence scores." + f"to aggregate self-influence scores." ) self.logger.error(error_msg) raise ValueError(error_msg) + score_args.aggregate_query_gradients = score_args.aggregate_train_gradients = False self._aggregate_scores( scores_name=scores_name, score_args=score_args, - exists_fnc=self_scores_exist, + exist_fnc=self_scores_exist, load_fnc=load_self_scores, save_fnc=save_self_scores, dim=0, diff --git a/kronfluence/factor/config.py b/kronfluence/factor/config.py index 3da75e1..3ba757f 100644 --- a/kronfluence/factor/config.py +++ b/kronfluence/factor/config.py @@ -9,6 +9,7 @@ ACTIVATION_EIGENVECTORS_NAME, GRADIENT_EIGENVALUES_NAME, GRADIENT_EIGENVECTORS_NAME, + HEURISTIC_DAMPING_SCALE, LAMBDA_MATRIX_NAME, NUM_LAMBDA_PROCESSED, ) @@ -17,7 +18,7 @@ class FactorStrategy(str, BaseEnum): - """Strategy for computing preconditioning factors.""" + """Strategies for computing preconditioning factors.""" IDENTITY = "identity" DIAGONAL = "diagonal" @@ -26,7 +27,7 @@ class FactorStrategy(str, BaseEnum): class FactorConfig(metaclass=ABCMeta): - """Configuration for each supported factor strategy.""" + """Configurations for each supported factor strategy.""" CONFIGS: Dict[FactorStrategy, Any] = {} @@ -40,32 +41,32 @@ def __init_subclass__(cls, factor_strategy: Optional[FactorStrategy] = None, **k @property @abstractmethod def requires_covariance_matrices(self) -> bool: - """Returns True if the strategy requires computing covariance matrices.""" + """Returns `True` if the strategy requires computing covariance matrices.""" raise NotImplementedError("Subclasses must implement the `requires_covariance_matrices` property.") @property @abstractmethod def requires_eigendecomposition(self) -> bool: - """Returns True if the strategy requires performing Eigendecomposition.""" + """Returns `True` if the strategy requires performing Eigendecomposition.""" raise NotImplementedError("Subclasses must implement the `requires_eigendecomposition` property.") @property @abstractmethod def requires_lambda_matrices(self) -> bool: - """Returns True if the strategy requires computing Lambda matrices.""" + """Returns `True` if the strategy requires computing Lambda matrices.""" raise NotImplementedError("Subclasses must implement the `requires_lambda_matrices` property.") @property @abstractmethod def requires_eigendecomposition_for_lambda(self) -> bool: - """Returns True if the strategy requires loading Eigendecomposition results, before computing + """Returns `True` if the strategy requires loading Eigendecomposition results, before computing Lambda matrices.""" raise NotImplementedError("Subclasses must implement the `requires_eigendecomposition_for_lambda` property.") @property @abstractmethod def requires_covariance_matrices_for_precondition(self) -> bool: - """Returns True if the strategy requires loading covariance matrices, before computing + """Returns `True` if the strategy requires loading covariance matrices, before computing preconditioned gradient.""" raise NotImplementedError( "Subclasses must implement the `requires_covariance_matrices_for_precondition` property." @@ -74,7 +75,7 @@ def requires_covariance_matrices_for_precondition(self) -> bool: @property @abstractmethod def requires_eigendecomposition_for_precondition(self) -> bool: - """Returns True if the strategy requires loading Eigendecomposition results, before computing + """Returns `True` if the strategy requires loading Eigendecomposition results, before computing preconditioned gradient.""" raise NotImplementedError( "Subclasses must implement the `requires_eigendecomposition_for_precondition` property." @@ -83,34 +84,42 @@ def requires_eigendecomposition_for_precondition(self) -> bool: @property @abstractmethod def requires_lambda_matrices_for_precondition(self) -> bool: - """Returns True if the strategy requires loading Lambda matrices, before computing + """Returns `True` if the strategy requires loading Lambda matrices, before computing the preconditioned gradient.""" raise NotImplementedError("Subclasses must implement the `requires_lambda_matrices_for_precondition` property.") + def prepare(self, storage: STORAGE_TYPE, score_args: Any, device: torch.device) -> None: + """Performs necessary operations before computing the preconditioned gradient. + + Args: + storage (STORAGE_TYPE): + A dictionary containing various factors required to compute the preconditioned gradient. + See `.storage` in `TrackedModule` for details. + score_args (ScoreArguments): + Arguments for computing the preconditioned gradient. + device (torch.device): + Device used for computing the preconditioned gradient. + """ + @abstractmethod def precondition_gradient( self, gradient: torch.Tensor, storage: STORAGE_TYPE, - damping: Optional[float] = None, ) -> torch.Tensor: - """Preconditions the per-sample-gradient. The per-sample-gradient is a 3-dimensional - tensor with the shape `batch_size x output_dim x input_dim`. + """Preconditions the per-sample gradient. The per-sample gradient is a 3-dimensional + tensor with shape `batch_size x output_dim x input_dim`. Args: gradient (torch.Tensor): - The per-sample-gradient tensor. + The per-sample gradient tensor. storage (STORAGE_TYPE): A dictionary containing various factors required to compute the preconditioned gradient. See `.storage` in `TrackedModule` for details. - damping (float, optional): - The damping factor when computing the preconditioned gradient. If not provided, sets - the damping term with some heuristic. Returns: torch.Tensor: - The preconditioned per-sample-gradient tensor. The dimension should be the same as the original - per-sample-gradient. + The preconditioned per-sample gradient tensor. """ raise NotImplementedError("Subclasses must implement the `precondition_gradient` method.") @@ -150,9 +159,8 @@ def precondition_gradient( self, gradient: torch.Tensor, storage: STORAGE_TYPE, - damping: Optional[float] = None, ) -> torch.Tensor: - del storage, damping + del storage return gradient @@ -187,18 +195,23 @@ def requires_eigendecomposition_for_precondition(self) -> bool: def requires_lambda_matrices_for_precondition(self) -> bool: return True + def prepare(self, storage: STORAGE_TYPE, score_args: Any, device: torch.device) -> None: + lambda_matrix = storage[LAMBDA_MATRIX_NAME].to(device=device) + lambda_matrix.div_(storage[NUM_LAMBDA_PROCESSED].to(device=device)) + damping_factor = score_args.damping_factor + if damping_factor is None: + damping_factor = HEURISTIC_DAMPING_SCALE * torch.mean(lambda_matrix) + lambda_matrix.add_(damping_factor) + storage[LAMBDA_MATRIX_NAME] = lambda_matrix.to(dtype=score_args.precondition_dtype, device="cpu").contiguous() + storage[NUM_LAMBDA_PROCESSED] = None + def precondition_gradient( self, gradient: torch.Tensor, storage: STORAGE_TYPE, - damping: Optional[float] = None, ) -> torch.Tensor: - lambda_matrix = storage[LAMBDA_MATRIX_NAME].to(dtype=gradient.dtype, device=gradient.device) - num_lambda_processed = storage[NUM_LAMBDA_PROCESSED].to(device=gradient.device) - lambda_matrix = lambda_matrix / num_lambda_processed - if damping is None: - damping = 0.1 * torch.mean(lambda_matrix) - return gradient / (lambda_matrix + damping) + lambda_matrix = storage[LAMBDA_MATRIX_NAME].to(device=gradient.device) + return gradient / lambda_matrix class Kfac(FactorConfig, factor_strategy=FactorStrategy.KFAC): @@ -235,26 +248,38 @@ def requires_eigendecomposition_for_precondition(self) -> bool: def requires_lambda_matrices_for_precondition(self) -> bool: return False + def prepare(self, storage: STORAGE_TYPE, score_args: Any, device: torch.device) -> None: + storage[ACTIVATION_EIGENVECTORS_NAME] = ( + storage[ACTIVATION_EIGENVECTORS_NAME].to(dtype=score_args.precondition_dtype).contiguous() + ) + storage[GRADIENT_EIGENVECTORS_NAME] = ( + storage[GRADIENT_EIGENVECTORS_NAME].to(dtype=score_args.precondition_dtype).contiguous() + ) + activation_eigenvalues = storage[ACTIVATION_EIGENVALUES_NAME].to(device=device) + gradient_eigenvalues = storage[GRADIENT_EIGENVALUES_NAME].to(device=device) + lambda_matrix = torch.kron(activation_eigenvalues.unsqueeze(0), gradient_eigenvalues.unsqueeze(-1)).unsqueeze(0) + damping_factor = score_args.damping_factor + if damping_factor is None: + damping_factor = HEURISTIC_DAMPING_SCALE * torch.mean(lambda_matrix) + lambda_matrix.add_(damping_factor) + storage[LAMBDA_MATRIX_NAME] = lambda_matrix.to(dtype=score_args.precondition_dtype, device="cpu").contiguous() + storage[NUM_LAMBDA_PROCESSED] = None + storage[ACTIVATION_EIGENVALUES_NAME] = None + storage[GRADIENT_EIGENVALUES_NAME] = None + @torch.no_grad() def precondition_gradient( self, gradient: torch.Tensor, storage: STORAGE_TYPE, - damping: Optional[float] = None, ) -> torch.Tensor: - activation_eigenvectors = storage[ACTIVATION_EIGENVECTORS_NAME].to(dtype=gradient.dtype, device=gradient.device) - gradient_eigenvectors = storage[GRADIENT_EIGENVECTORS_NAME].to(dtype=gradient.dtype, device=gradient.device) - activation_eigenvalues = storage[ACTIVATION_EIGENVALUES_NAME].to(dtype=gradient.dtype, device=gradient.device) - gradient_eigenvalues = storage[GRADIENT_EIGENVALUES_NAME].to(dtype=gradient.dtype, device=gradient.device) - lambda_matrix = torch.kron(activation_eigenvalues.unsqueeze(0), gradient_eigenvalues.unsqueeze(-1)).unsqueeze(0) - + activation_eigenvectors = storage[ACTIVATION_EIGENVECTORS_NAME].to(device=gradient.device) + gradient_eigenvectors = storage[GRADIENT_EIGENVECTORS_NAME].to(device=gradient.device) + lambda_matrix = storage[LAMBDA_MATRIX_NAME].to(device=gradient.device) gradient = torch.matmul(gradient_eigenvectors.t(), torch.matmul(gradient, activation_eigenvectors)) - - if damping is None: - damping = 0.1 * torch.mean(lambda_matrix) - - gradient.div_(lambda_matrix + damping) - return torch.matmul(gradient_eigenvectors, torch.matmul(gradient, activation_eigenvectors.t())) + gradient.div_(lambda_matrix) + gradient = torch.matmul(gradient_eigenvectors, torch.matmul(gradient, activation_eigenvectors.t())) + return gradient class Ekfac(FactorConfig, factor_strategy=FactorStrategy.EKFAC): @@ -291,23 +316,34 @@ def requires_eigendecomposition_for_precondition(self) -> bool: def requires_lambda_matrices_for_precondition(self) -> bool: return True + def prepare(self, storage: STORAGE_TYPE, score_args: Any, device: torch.device) -> None: + storage[ACTIVATION_EIGENVECTORS_NAME] = ( + storage[ACTIVATION_EIGENVECTORS_NAME].to(dtype=score_args.precondition_dtype).contiguous() + ) + storage[GRADIENT_EIGENVECTORS_NAME] = ( + storage[GRADIENT_EIGENVECTORS_NAME].to(dtype=score_args.precondition_dtype).contiguous() + ) + storage[ACTIVATION_EIGENVALUES_NAME] = None + storage[GRADIENT_EIGENVALUES_NAME] = None + lambda_matrix = storage[LAMBDA_MATRIX_NAME].to(device=device) + lambda_matrix.div_(storage[NUM_LAMBDA_PROCESSED].to(device=device)) + damping_factor = score_args.damping_factor + if damping_factor is None: + damping_factor = HEURISTIC_DAMPING_SCALE * torch.mean(lambda_matrix) + lambda_matrix.add_(damping_factor) + storage[LAMBDA_MATRIX_NAME] = lambda_matrix.to(dtype=score_args.precondition_dtype, device="cpu").contiguous() + storage[NUM_LAMBDA_PROCESSED] = None + @torch.no_grad() def precondition_gradient( self, gradient: torch.Tensor, storage: STORAGE_TYPE, - damping: Optional[float] = None, ) -> torch.Tensor: - activation_eigenvectors = storage[ACTIVATION_EIGENVECTORS_NAME].to(dtype=gradient.dtype, device=gradient.device) - gradient_eigenvectors = storage[GRADIENT_EIGENVECTORS_NAME].to(dtype=gradient.dtype, device=gradient.device) - lambda_matrix = storage[LAMBDA_MATRIX_NAME].to(dtype=gradient.dtype, device=gradient.device) - num_lambda_processed = storage[NUM_LAMBDA_PROCESSED].to(device=gradient.device) - lambda_matrix = lambda_matrix / num_lambda_processed + activation_eigenvectors = storage[ACTIVATION_EIGENVECTORS_NAME].to(device=gradient.device) + gradient_eigenvectors = storage[GRADIENT_EIGENVECTORS_NAME].to(device=gradient.device) + lambda_matrix = storage[LAMBDA_MATRIX_NAME].to(device=gradient.device) gradient = torch.matmul(gradient_eigenvectors.t(), torch.matmul(gradient, activation_eigenvectors)) - - if damping is None: - damping = 0.1 * torch.mean(lambda_matrix) - - gradient.div_(lambda_matrix + damping) + gradient.div_(lambda_matrix) gradient = torch.matmul(gradient_eigenvectors, torch.matmul(gradient, activation_eigenvectors.t())) return gradient diff --git a/kronfluence/factor/covariance.py b/kronfluence/factor/covariance.py index a1450e2..b32ee04 100644 --- a/kronfluence/factor/covariance.py +++ b/kronfluence/factor/covariance.py @@ -13,18 +13,18 @@ from kronfluence.arguments import FactorArguments from kronfluence.module.tracked_module import ModuleMode from kronfluence.module.utils import ( + get_tracked_module_names, load_factors, - remove_attention_mask, - remove_gradient_scale, set_attention_mask, set_gradient_scale, set_mode, - synchronize_covariance_matrices, + synchronize_modules, update_factor_args, ) from kronfluence.task import Task from kronfluence.utils.constants import ( COVARIANCE_FACTOR_NAMES, + DISTRIBUTED_SYNC_INTERVAL, FACTOR_TYPE, PARTITION_TYPE, ) @@ -37,7 +37,24 @@ def covariance_matrices_save_path( factor_name: str, partition: Optional[PARTITION_TYPE] = None, ) -> Path: - """Generates the path for saving/loading covariance matrices.""" + """Generates the path for saving or loading covariance matrices. + + Args: + output_dir (Path): + Directory to save or load the matrices. + factor_name (str): + Name of the factor (must be in `COVARIANCE_FACTOR_NAMES`). + partition (PARTITION_TYPE, optional): + Partition information, if any. + + Returns: + Path: + The full path for the covariance matrix file. + + Raises: + AssertionError: + If `factor_name` is not in `COVARIANCE_FACTOR_NAMES`. + """ assert factor_name in COVARIANCE_FACTOR_NAMES if partition is not None: data_partition, module_partition = partition @@ -53,7 +70,22 @@ def save_covariance_matrices( partition: Optional[PARTITION_TYPE] = None, metadata: Optional[Dict[str, str]] = None, ) -> None: - """Saves covariance matrices to disk.""" + """Saves covariance matrices to disk. + + Args: + output_dir (Path): + Directory to save the matrices. + factors (FACTOR_TYPE): + Dictionary of factors to save. + partition (PARTITION_TYPE, optional): + Partition information, if any. + metadata (Dict[str, str], optional): + Additional metadata to save with the factors. + + Raises: + AssertionError: + If factors keys don't match `COVARIANCE_FACTOR_NAMES`. + """ assert set(factors.keys()) == set(COVARIANCE_FACTOR_NAMES) for factor_name in factors: save_path = covariance_matrices_save_path( @@ -68,7 +100,18 @@ def load_covariance_matrices( output_dir: Path, partition: Optional[PARTITION_TYPE] = None, ) -> FACTOR_TYPE: - """Loads covariance matrices from disk.""" + """Loads covariance matrices from disk. + + Args: + output_dir (Path): + Directory to load the matrices from. + partition (PARTITION_TYPE, optional): + Partition information, if any. + + Returns: + FACTOR_TYPE: + Dictionary of loaded covariance factors. + """ covariance_factors = {} for factor_name in COVARIANCE_FACTOR_NAMES: save_path = covariance_matrices_save_path( @@ -84,7 +127,18 @@ def covariance_matrices_exist( output_dir: Path, partition: Optional[PARTITION_TYPE] = None, ) -> bool: - """Checks if covariance matrices exist at the specified directory.""" + """Checks if covariance matrices exist at the specified directory. + + Args: + output_dir (Path): + Directory to check for matrices. + partition (PARTITION_TYPE, optional): + Partition information, if any. + + Returns: + bool: + `True` if all covariance matrices exist, `False` otherwise. + """ for factor_name in COVARIANCE_FACTOR_NAMES: save_path = covariance_matrices_save_path( output_dir=output_dir, @@ -103,6 +157,7 @@ def fit_covariance_matrices_with_loader( loader: data.DataLoader, factor_args: FactorArguments, tracked_module_names: Optional[List[str]] = None, + disable_tqdm: bool = False, ) -> Tuple[torch.Tensor, FACTOR_TYPE]: """Computes activation and pseudo-gradient covariance matrices for a given model and task. @@ -120,38 +175,45 @@ def fit_covariance_matrices_with_loader( tracked_module_names (List[str], optional): A list of module names for which covariance matrices will be computed. If not specified, covariance matrices will be computed for all tracked modules. + disable_tqdm (bool, optional): + Whether to disable the progress bar. Defaults to `False`. Returns: Tuple[torch.Tensor, FACTOR_TYPE]: - A tuple containing the number of data points processed and computed covariance matrices in CPU. - The covariance matrices are organized in nested dictionaries, where the first key is the name of the - covariance matrix (e.g., activation covariance and pseudo-gradient covariance) and the second key is - the module name. + - Number of data points processed. + - Computed covariance matrices (nested dict: factor_name -> module_name -> tensor). """ - with torch.no_grad(): - update_factor_args(model=model, factor_args=factor_args) - set_mode( - model=model, - tracked_module_names=tracked_module_names, - mode=ModuleMode.COVARIANCE, - ) + update_factor_args(model=model, factor_args=factor_args) + if tracked_module_names is None: + tracked_module_names = get_tracked_module_names(model=model) + set_mode( + model=model, + tracked_module_names=tracked_module_names, + mode=ModuleMode.COVARIANCE, + release_memory=True, + ) + total_steps = 0 num_data_processed = torch.zeros((1,), dtype=torch.int64, requires_grad=False) enable_amp = factor_args.amp_dtype is not None - scaler = GradScaler(enabled=enable_amp) + enable_grad_scaler = enable_amp and factor_args.amp_dtype == torch.float16 + scaler = GradScaler(init_scale=factor_args.amp_scale, enabled=enable_grad_scaler) + if enable_grad_scaler: + gradient_scale = 1.0 / scaler.get_scale() + set_gradient_scale(model=model, gradient_scale=gradient_scale) with tqdm( total=len(loader), desc="Fitting covariance matrices", bar_format=TQDM_BAR_FORMAT, - disable=not state.is_main_process, + disable=not state.is_main_process or disable_tqdm, ) as pbar: for index, batch in enumerate(loader): batch = send_to_device(batch, device=state.device) - with torch.no_grad(): - attention_mask = task.get_attention_mask(batch=batch) - if attention_mask is not None: - set_attention_mask(model=model, attention_mask=attention_mask) + + attention_mask = task.get_attention_mask(batch=batch) + if attention_mask is not None: + set_attention_mask(model=model, attention_mask=attention_mask) with no_sync(model=model, state=state): model.zero_grad(set_to_none=True) @@ -161,40 +223,44 @@ def fit_covariance_matrices_with_loader( model=model, sample=not factor_args.use_empirical_fisher, ) - scaled_loss = scaler.scale(loss) - if enable_amp: - gradient_scale = 1.0 / scaler.get_scale() - set_gradient_scale(model=model, gradient_scale=gradient_scale) - scaled_loss.backward() - - num_data_processed += find_batch_size(data=batch) - total_steps += 1 + scaler.scale(loss).backward() if ( state.use_distributed - and total_steps % factor_args.distributed_sync_steps == 0 + and total_steps % DISTRIBUTED_SYNC_INTERVAL == 0 and index not in [len(loader) - 1, len(loader) - 2] ): - # Periodically synchronize all processes to avoid timeout at the final covariance synchronization. state.wait_for_everyone() + num_data_processed.add_(find_batch_size(data=batch)) + del loss + total_steps += 1 pbar.update(1) - with torch.no_grad(): - del loss - model.zero_grad(set_to_none=True) - remove_attention_mask(model=model) - remove_gradient_scale(model=model) + if state.use_distributed: + synchronize_modules(model=model, tracked_module_names=tracked_module_names) + num_data_processed = num_data_processed.to(device=state.device) + dist.all_reduce(tensor=num_data_processed, op=torch.distributed.ReduceOp.SUM) + num_data_processed = num_data_processed.cpu() - if state.use_distributed: - # Aggregate covariance matrices across multiple devices or nodes. - synchronize_covariance_matrices(model=model) - num_data_processed = num_data_processed.to(device=state.device) - dist.all_reduce(tensor=num_data_processed, op=torch.distributed.ReduceOp.SUM) - - saved_factors: FACTOR_TYPE = {} + saved_factors: FACTOR_TYPE = {} + if state.is_main_process: for factor_name in COVARIANCE_FACTOR_NAMES: - saved_factors[factor_name] = load_factors(model=model, factor_name=factor_name) - state.wait_for_everyone() - set_mode(model=model, mode=ModuleMode.DEFAULT, keep_factors=False) + factor = load_factors( + model=model, + factor_name=factor_name, + tracked_module_names=tracked_module_names, + cpu=True, + ) + if len(factor) == 0: + raise ValueError(f"Factor `{factor_name}` has not been computed.") + saved_factors[factor_name] = factor + + model.zero_grad(set_to_none=True) + set_attention_mask(model=model, attention_mask=None) + if enable_grad_scaler: + set_gradient_scale(model=model, gradient_scale=1.0) + set_mode(model=model, mode=ModuleMode.DEFAULT, release_memory=True) + state.wait_for_everyone() + return num_data_processed, saved_factors diff --git a/kronfluence/factor/eigen.py b/kronfluence/factor/eigen.py index 4559d7f..dfea4f6 100644 --- a/kronfluence/factor/eigen.py +++ b/kronfluence/factor/eigen.py @@ -4,6 +4,7 @@ import torch import torch.distributed as dist from accelerate.utils import find_batch_size, send_to_device +from accelerate.utils.memory import should_reduce_batch_size from safetensors.torch import load_file, save_file from torch import autocast, nn from torch.cuda.amp import GradScaler @@ -13,13 +14,13 @@ from kronfluence.arguments import FactorArguments from kronfluence.module.tracked_module import ModuleMode from kronfluence.module.utils import ( + finalize_iteration, get_tracked_module_names, load_factors, - remove_gradient_scale, set_factors, set_gradient_scale, set_mode, - synchronize_lambda_matrices, + synchronize_modules, update_factor_args, ) from kronfluence.task import Task @@ -27,30 +28,60 @@ ACTIVATION_COVARIANCE_MATRIX_NAME, ACTIVATION_EIGENVALUES_NAME, ACTIVATION_EIGENVECTORS_NAME, + DISTRIBUTED_SYNC_INTERVAL, EIGENDECOMPOSITION_FACTOR_NAMES, FACTOR_TYPE, GRADIENT_COVARIANCE_MATRIX_NAME, GRADIENT_EIGENVALUES_NAME, GRADIENT_EIGENVECTORS_NAME, LAMBDA_FACTOR_NAMES, - NUM_COVARIANCE_PROCESSED, + NUM_ACTIVATION_COVARIANCE_PROCESSED, + NUM_GRADIENT_COVARIANCE_PROCESSED, PARTITION_TYPE, ) from kronfluence.utils.logger import TQDM_BAR_FORMAT -from kronfluence.utils.state import State, no_sync +from kronfluence.utils.state import State, no_sync, release_memory def eigendecomposition_save_path( output_dir: Path, factor_name: str, ) -> Path: - """Generates the path for saving/loading Eigendecomposition results.""" + """Generates the path for saving or loading eigendecomposition results. + + Args: + output_dir (Path): + Directory to save or load eigenvectors and eigenvalues. + factor_name (str): + Name of the factor (must be in `EIGENDECOMPOSITION_FACTOR_NAMES`). + + Returns: + Path: + The full path for the eigendecomposition file. + + Raises: + AssertionError: + If `factor_name` is not in `EIGENDECOMPOSITION_FACTOR_NAMES`. + """ assert factor_name in EIGENDECOMPOSITION_FACTOR_NAMES return output_dir / f"{factor_name}.safetensors" def save_eigendecomposition(output_dir: Path, factors: FACTOR_TYPE, metadata: Optional[Dict[str, str]] = None) -> None: - """Saves Eigendecomposition results to disk.""" + """Saves eigendecomposition results to disk. + + Args: + output_dir (Path): + Directory to save the eigenvectors and eigenvalues. + factors (FACTOR_TYPE): + Dictionary of factors to save. + metadata (Dict[str, str], optional): + Additional metadata to save with the factors. + + Raises: + AssertionError: + If factors keys don't match `EIGENDECOMPOSITION_FACTOR_NAMES`. + """ assert set(factors.keys()) == set(EIGENDECOMPOSITION_FACTOR_NAMES) for factor_name in factors: save_path = eigendecomposition_save_path( @@ -63,7 +94,16 @@ def save_eigendecomposition(output_dir: Path, factors: FACTOR_TYPE, metadata: Op def load_eigendecomposition( output_dir: Path, ) -> FACTOR_TYPE: - """Loads Eigendecomposition results from disk.""" + """Loads eigendecomposition results from disk. + + Args: + output_dir (Path): + Directory to load the results from. + + Returns: + FACTOR_TYPE: + Dictionary of loaded eigendecomposition results. + """ eigen_factors = {} for factor_name in EIGENDECOMPOSITION_FACTOR_NAMES: save_path = eigendecomposition_save_path( @@ -77,7 +117,16 @@ def load_eigendecomposition( def eigendecomposition_exist( output_dir: Path, ) -> bool: - """Checks if Eigendecomposition results exist at the specified path.""" + """Checks if eigendecomposition results exist at the specified directory. + + Args: + output_dir (Path): + Directory to check for results. + + Returns: + bool: + `True` if all eigendecomposition results exist, `False` otherwise. + """ for factor_name in EIGENDECOMPOSITION_FACTOR_NAMES: save_path = eigendecomposition_save_path( output_dir=output_dir, @@ -94,24 +143,26 @@ def perform_eigendecomposition( model: nn.Module, state: State, factor_args: FactorArguments, + disable_tqdm: bool = False, ) -> FACTOR_TYPE: - """Performs Eigendecomposition on activation and pseudo-gradient covariance matrices. + """Performs eigendecomposition on activation and pseudo-gradient covariance matrices. Args: covariance_factors (FACTOR_TYPE): - The covariance matrices to perform Eigendecomposition on. + Computed covariance factors. model (nn.Module): - The model which contains modules which Eigendecomposition will be performed. + The model used to compute covariance matrices. state (State): The current process's information (e.g., device being used). factor_args (FactorArguments): - Arguments related to performing Eigendecomposition. + Arguments for performing eigendecomposition. + disable_tqdm (bool, optional): + Whether to disable the progress bar. Defaults to `False`. Returns: FACTOR_TYPE: - The Eigendecomposition results in CPU. The Eigendecomposition results are organized in - nested dictionaries, where the first key is the name of the Eigendecomposition factor (e.g., - activation eigenvector), and the second key is the module name. + The results are organized in nested dictionaries, where the first key is the name of the factor + (e.g., activation eigenvector), and the second key is the module name. """ eigen_factors: FACTOR_TYPE = {} for factor_name in EIGENDECOMPOSITION_FACTOR_NAMES: @@ -122,17 +173,19 @@ def perform_eigendecomposition( total=len(tracked_module_names), desc="Performing Eigendecomposition", bar_format=TQDM_BAR_FORMAT, - disable=not state.is_main_process, + disable=not state.is_main_process or disable_tqdm, ) as pbar: for module_name in tracked_module_names: - for covariance_name, eigenvectors_name, eigenvalues_name in [ + for covariance_name, num_processed_name, eigenvectors_name, eigenvalues_name in [ ( ACTIVATION_COVARIANCE_MATRIX_NAME, + NUM_ACTIVATION_COVARIANCE_PROCESSED, ACTIVATION_EIGENVECTORS_NAME, ACTIVATION_EIGENVALUES_NAME, ), ( GRADIENT_COVARIANCE_MATRIX_NAME, + NUM_GRADIENT_COVARIANCE_PROCESSED, GRADIENT_EIGENVECTORS_NAME, GRADIENT_EIGENVALUES_NAME, ), @@ -143,16 +196,31 @@ def perform_eigendecomposition( dtype=factor_args.eigendecomposition_dtype, ) # Normalize covariance matrices. - covariance_matrix.div_( - covariance_factors[NUM_COVARIANCE_PROCESSED][module_name].to(device=state.device) - ) + covariance_matrix.div_(covariance_factors[num_processed_name][module_name].to(device=state.device)) # In cases where covariance matrices are not exactly symmetric due to numerical issues. - covariance_matrix = 0.5 * (covariance_matrix + covariance_matrix.t()) - eigenvalues, eigenvectors = torch.linalg.eigh(covariance_matrix) - eigen_factors[eigenvalues_name][module_name] = eigenvalues.to(dtype=original_dtype).contiguous().cpu() - eigen_factors[eigenvectors_name][module_name] = eigenvectors.to(dtype=original_dtype).contiguous().cpu() + covariance_matrix = covariance_matrix + covariance_matrix.t() + covariance_matrix.mul_(0.5) + + try: + eigenvalues, eigenvectors = torch.linalg.eigh(covariance_matrix) + except Exception as e: # pylint: disable=broad-exception-caught + # If we get OOM error, release the memory and try again. + if should_reduce_batch_size(exception=e): # pylint: disable=no-else-continue + release_memory() + eigenvalues, eigenvectors = torch.linalg.eigh(covariance_matrix) + else: + raise + del covariance_matrix + eigen_factors[eigenvalues_name][module_name] = eigenvalues.contiguous().to( + dtype=original_dtype, device="cpu" + ) + eigen_factors[eigenvectors_name][module_name] = eigenvectors.contiguous().to( + dtype=original_dtype, device="cpu" + ) del eigenvalues, eigenvectors + pbar.update(1) + return eigen_factors @@ -161,7 +229,24 @@ def lambda_matrices_save_path( factor_name: str, partition: Optional[PARTITION_TYPE] = None, ) -> Path: - """Generates the path for saving/loading Lambda matrices.""" + """Generates the path for saving or loading Lambda matrices. + + Args: + output_dir (Path): + Directory to save or load the matrices. + factor_name (str): + Name of the factor (must be in `LAMBDA_FACTOR_NAMES`). + partition (PARTITION_TYPE, optional): + Partition information, if any. + + Returns: + Path: + The full path for the Lambda matrix file. + + Raises: + AssertionError: + If `factor_name` is not in `LAMBDA_FACTOR_NAMES`. + """ assert factor_name in LAMBDA_FACTOR_NAMES if partition is not None: data_partition, module_partition = partition @@ -177,7 +262,22 @@ def save_lambda_matrices( partition: Optional[PARTITION_TYPE] = None, metadata: Optional[Dict[str, str]] = None, ) -> None: - """Saves Lambda matrices to disk.""" + """Saves Lambda matrices to disk. + + Args: + output_dir (Path): + Directory to save the matrices. + factors (FACTOR_TYPE): + Dictionary of factors to save. + partition (PARTITION_TYPE, optional): + Partition information, if any. + metadata (Dict[str, str], optional): + Additional metadata to save with the factors. + + Raises: + AssertionError: + If factors keys don't match `LAMBDA_FACTOR_NAMES`. + """ assert set(factors.keys()) == set(LAMBDA_FACTOR_NAMES) for factor_name in factors: save_path = lambda_matrices_save_path( @@ -192,7 +292,18 @@ def load_lambda_matrices( output_dir: Path, partition: Optional[PARTITION_TYPE] = None, ) -> FACTOR_TYPE: - """Loads Lambda matrices from disk.""" + """Loads Lambda matrices from disk. + + Args: + output_dir (Path): + Directory to load the matrices from. + partition (PARTITION_TYPE, optional): + Partition information, if any. + + Returns: + FACTOR_TYPE: + Dictionary of loaded Lambda factors. + """ lambda_factors = {} for factor_name in LAMBDA_FACTOR_NAMES: save_path = lambda_matrices_save_path( @@ -208,7 +319,18 @@ def lambda_matrices_exist( output_dir: Path, partition: Optional[PARTITION_TYPE] = None, ) -> bool: - """Check if Lambda matrices exist at the specified path.""" + """Checks if Lambda matrices exist at the specified directory. + + Args: + output_dir (Path): + Directory to check for matrices. + partition (PARTITION_TYPE, optional): + Partition information, if any. + + Returns: + bool: + `True` if all Lambda matrices exist, `False` otherwise. + """ for factor_name in LAMBDA_FACTOR_NAMES: save_path = lambda_matrices_save_path( output_dir=output_dir, @@ -228,8 +350,9 @@ def fit_lambda_matrices_with_loader( factor_args: FactorArguments, eigen_factors: Optional[FACTOR_TYPE] = None, tracked_module_names: Optional[List[str]] = None, + disable_tqdm: bool = False, ) -> Tuple[torch.Tensor, FACTOR_TYPE]: - """Computes Lambda (corrected eigenvalues) matrices for a given model and task. + """Computes Lambda matrices for a given model and task. Args: model (nn.Module): @@ -243,37 +366,45 @@ def fit_lambda_matrices_with_loader( factor_args (FactorArguments): Arguments for computing Lambda matrices. eigen_factors (FACTOR_TYPE, optional): - The eigendecomposition results to use for computing Lambda matrices. + Computed eigendecomposition results. tracked_module_names (List[str], optional): A list of module names for which Lambda matrices will be computed. If not specified, Lambda matrices will be computed for all tracked modules. + disable_tqdm (bool, optional): + Whether to disable the progress bar. Defaults to `False`. Returns: Tuple[torch.Tensor, FACTOR_TYPE]: - A tuple containing the number of data points processed, and computed Lambda matrices in CPU. - The Lambda matrices are organized in nested dictionaries, where the first key is the name of - the computed variable and the second key is the module name. + - Number of data points processed. + - Computed Lambda matrices (nested dict: factor_name -> module_name -> tensor). """ - with torch.no_grad(): - update_factor_args(model=model, factor_args=factor_args) - set_mode( - model=model, - tracked_module_names=tracked_module_names, - mode=ModuleMode.LAMBDA, - ) - if eigen_factors is not None: - for name in eigen_factors: - set_factors(model=model, factor_name=name, factors=eigen_factors[name]) + update_factor_args(model=model, factor_args=factor_args) + if tracked_module_names is None: + tracked_module_names = get_tracked_module_names(model=model) + set_mode( + model=model, + tracked_module_names=tracked_module_names, + mode=ModuleMode.LAMBDA, + release_memory=True, + ) + if eigen_factors is not None: + for name in eigen_factors: + set_factors(model=model, factor_name=name, factors=eigen_factors[name], clone=True) + total_steps = 0 num_data_processed = torch.zeros((1,), dtype=torch.int64, requires_grad=False) enable_amp = factor_args.amp_dtype is not None - scaler = GradScaler(enabled=enable_amp) + enable_grad_scaler = enable_amp and factor_args.amp_dtype == torch.float16 + scaler = GradScaler(init_scale=factor_args.amp_scale, enabled=enable_grad_scaler) + if enable_grad_scaler: + gradient_scale = 1.0 / scaler.get_scale() + set_gradient_scale(model=model, gradient_scale=gradient_scale) with tqdm( total=len(loader), desc="Fitting Lambda matrices", bar_format=TQDM_BAR_FORMAT, - disable=not state.is_main_process, + disable=not state.is_main_process or disable_tqdm, ) as pbar: for index, batch in enumerate(loader): batch = send_to_device(tensor=batch, device=state.device) @@ -286,40 +417,46 @@ def fit_lambda_matrices_with_loader( model=model, sample=not factor_args.use_empirical_fisher, ) - scaled_loss = scaler.scale(loss) - if enable_amp: - gradient_scale = 1.0 / scaler.get_scale() - set_gradient_scale(model=model, gradient_scale=gradient_scale) - scaled_loss.backward() + scaler.scale(loss).backward() - num_data_processed += find_batch_size(data=batch) - total_steps += 1 + if factor_args.has_shared_parameters: + finalize_iteration(model=model, tracked_module_names=tracked_module_names) if ( state.use_distributed - and total_steps % factor_args.distributed_sync_steps == 0 + and total_steps % DISTRIBUTED_SYNC_INTERVAL == 0 and index not in [len(loader) - 1, len(loader) - 2] ): - # Periodically synchronize all processes to avoid timeout at the final Lambda synchronization. state.wait_for_everyone() + num_data_processed.add_(find_batch_size(data=batch)) + del loss + total_steps += 1 pbar.update(1) - with torch.no_grad(): - del loss - remove_gradient_scale(model=model) - model.zero_grad(set_to_none=True) - - if state.use_distributed: - # Aggregate Lambda matrices across multiple devices or nodes. - synchronize_lambda_matrices(model=model) - num_data_processed = num_data_processed.to(device=state.device) - dist.all_reduce(tensor=num_data_processed, op=torch.distributed.ReduceOp.SUM) - - saved_factors: FACTOR_TYPE = {} - if state.is_main_process: - for factor_name in LAMBDA_FACTOR_NAMES: - saved_factors[factor_name] = load_factors(model=model, factor_name=factor_name) - state.wait_for_everyone() - set_mode(model=model, mode=ModuleMode.DEFAULT, keep_factors=False) + if state.use_distributed: + synchronize_modules(model=model, tracked_module_names=tracked_module_names) + num_data_processed = num_data_processed.to(device=state.device) + dist.all_reduce(tensor=num_data_processed, op=torch.distributed.ReduceOp.SUM) + num_data_processed = num_data_processed.cpu() + + saved_factors: FACTOR_TYPE = {} + if state.is_main_process: + for factor_name in LAMBDA_FACTOR_NAMES: + factor = load_factors( + model=model, + factor_name=factor_name, + tracked_module_names=tracked_module_names, + cpu=True, + ) + if len(factor) == 0: + raise ValueError(f"Factor `{factor_name}` has not been computed.") + saved_factors[factor_name] = factor + + model.zero_grad(set_to_none=True) + if enable_grad_scaler: + set_gradient_scale(model=model, gradient_scale=1.0) + set_mode(model=model, mode=ModuleMode.DEFAULT, release_memory=True) + state.wait_for_everyone() + return num_data_processed, saved_factors diff --git a/kronfluence/module/conv2d.py b/kronfluence/module/conv2d.py index 03953c6..f7b606a 100644 --- a/kronfluence/module/conv2d.py +++ b/kronfluence/module/conv2d.py @@ -1,11 +1,11 @@ -from typing import Tuple, Union +from typing import Optional, Tuple, Union import torch import torch.nn.functional as F from einconv.utils import get_conv_paddings from einops import rearrange, reduce -from opt_einsum import contract -from torch import nn +from opt_einsum import DynamicProgramming, contract_path +from torch import _VF, nn from torch.nn.modules.utils import _pair from kronfluence.module.tracked_module import TrackedModule @@ -65,22 +65,45 @@ def extract_patches( class TrackedConv2d(TrackedModule, module_type=nn.Conv2d): - """A tracking wrapper for `nn.Conv2D` modules.""" - - def _get_flattened_activation( - self, input_activation: torch.Tensor - ) -> Tuple[torch.Tensor, Union[torch.Tensor, int]]: - """Returns the flattened activation tensor and the number of stacked activations. - - Args: - input_activation (torch.Tensor): - The input tensor to the module, provided by the PyTorch's forward hook. - - Returns: - Tuple[torch.Tensor, torch.Tensor]: - The flattened activation tensor and the number of stacked activations. The flattened - activation is a 2-dimensional matrix with dimension `activation_num x activation_dim`. - """ + """A wrapper for `nn.Conv2d` modules.""" + + @property + def in_channels(self) -> int: # pylint: disable=missing-function-docstring + return self.original_module.in_channels + + @property + def out_channels(self) -> int: # pylint: disable=missing-function-docstring + return self.original_module.out_channels + + @property + def kernel_size(self) -> Tuple[int, int]: # pylint: disable=missing-function-docstring + return self.original_module.kernel_size + + @property + def padding(self) -> Tuple[int, int]: # pylint: disable=missing-function-docstring + return self.original_module.padding + + @property + def dilation(self) -> Tuple[int, int]: # pylint: disable=missing-function-docstring + return self.original_module.dilation + + @property + def groups(self) -> int: # pylint: disable=missing-function-docstring + return self.original_module.groups + + @property + def padding_mode(self) -> str: # pylint: disable=missing-function-docstring + return self.original_module.padding_mode + + @property + def weight(self) -> torch.Tensor: # pylint: disable=missing-function-docstring + return self.original_module.weight + + @property + def bias(self) -> Optional[torch.Tensor]: # pylint: disable=missing-function-docstring + return self.original_module.bias + + def get_flattened_activation(self, input_activation: torch.Tensor) -> Tuple[torch.Tensor, Union[torch.Tensor, int]]: input_activation = extract_patches( inputs=input_activation, kernel_size=self.original_module.kernel_size, @@ -93,7 +116,6 @@ def _get_flattened_activation( tensor=input_activation, pattern="b o1_o2 c_in_k1_k2 -> (b o1_o2) c_in_k1_k2", ) - if self.original_module.bias is not None: input_activation = torch.cat( [ @@ -105,41 +127,11 @@ def _get_flattened_activation( count = input_activation.size(0) return input_activation, count - def _get_flattened_gradient(self, output_gradient: torch.Tensor) -> torch.Tensor: - """Returns the flattened gradient tensor. - - Args: - output_gradient (torch.Tensor): - The gradient tensor with respect to the output of the module, provided by the - PyTorch's backward hook. + def get_flattened_gradient(self, output_gradient: torch.Tensor) -> Tuple[torch.Tensor, Union[torch.Tensor, int]]: + output_gradient = rearrange(output_gradient, "b c o1 o2 -> (b o1 o2) c") + return output_gradient, output_gradient.size(0) - Returns: - torch.Tensor: - The flattened output gradient tensor. The flattened gradient is a 2-dimensional matrix - with dimension `gradient_num x gradient_dim`. - """ - return rearrange(output_gradient, "b c o1 o2 -> (b o1 o2) c") - - def _compute_per_sample_gradient( - self, - input_activation: torch.Tensor, - output_gradient: torch.Tensor, - ) -> torch.Tensor: - """Returns the flattened per-sample-gradient tensor. - - Args: - input_activation (torch.Tensor): - The input tensor to the module, provided by the PyTorch's forward hook. - output_gradient (torch.Tensor): - The gradient tensor with respect to the output of the module, provided by the - PyTorch's backward hook. - - Returns: - torch.Tensor: - The per-sample-gradient tensor. The per-sample-gradient is a 3-dimensional matrix - with dimension `batch_size x gradient_dim x activation_dim`. An additional dimension is added - when the bias term is used. - """ + def _flatten_input_activation(self, input_activation: torch.Tensor) -> torch.Tensor: input_activation = extract_patches( inputs=input_activation, kernel_size=self.original_module.kernel_size, @@ -152,7 +144,6 @@ def _compute_per_sample_gradient( tensor=input_activation, pattern="b o1_o2 c_in_k1_k2 -> (b o1_o2) c_in_k1_k2", ) - if self.original_module.bias is not None: input_activation = torch.cat( [ @@ -161,6 +152,76 @@ def _compute_per_sample_gradient( ], dim=-1, ) + return input_activation + + def compute_summed_gradient(self, input_activation: torch.Tensor, output_gradient: torch.Tensor) -> torch.Tensor: + input_activation = self._flatten_input_activation(input_activation=input_activation) + input_activation = input_activation.view(output_gradient.size(0), -1, input_activation.size(-1)) + output_gradient = rearrange(tensor=output_gradient, pattern="b o i1 i2 -> b (i1 i2) o") + summed_gradient = torch.einsum("bci,bco->io", output_gradient, input_activation).unsqueeze_(dim=0) + return summed_gradient + + def compute_per_sample_gradient( + self, + input_activation: torch.Tensor, + output_gradient: torch.Tensor, + ) -> torch.Tensor: + input_activation = self._flatten_input_activation(input_activation=input_activation) + input_activation = input_activation.view(output_gradient.size(0), -1, input_activation.size(-1)) + output_gradient = rearrange(tensor=output_gradient, pattern="b o i1 i2 -> b (i1 i2) o") + per_sample_gradient = torch.einsum("bci,bco->bio", output_gradient, input_activation) + if self.per_sample_gradient_process_fnc is not None: + per_sample_gradient = self.per_sample_gradient_process_fnc( + module_name=self.name, gradient=per_sample_gradient + ) + return per_sample_gradient + + def compute_pairwise_score( + self, preconditioned_gradient: torch.Tensor, input_activation: torch.Tensor, output_gradient: torch.Tensor + ) -> torch.Tensor: + input_activation = self._flatten_input_activation(input_activation=input_activation) + input_activation = input_activation.view(output_gradient.size(0), -1, input_activation.size(-1)) + output_gradient = rearrange(tensor=output_gradient, pattern="b o i1 i2 -> b (i1 i2) o") + if isinstance(preconditioned_gradient, list): + left_mat, right_mat = preconditioned_gradient + expr = "qik,qko,b...i,b...o->qb" + if self.einsum_path is None: + path = contract_path( + expr, + left_mat, + right_mat, + output_gradient, + input_activation, + optimize=DynamicProgramming(search_outer=True, minimize="flops"), + )[0] + self.einsum_path = [item for pair in path for item in pair] + return _VF.einsum(expr, (left_mat, right_mat, output_gradient, input_activation), path=self.einsum_path) # pylint: disable=no-member + expr = "qio,bti,bto->qb" + if self.einsum_path is None: + path = contract_path( + expr, + preconditioned_gradient, + output_gradient, + input_activation, + optimize=DynamicProgramming(search_outer=True, minimize="flops"), + )[0] + self.einsum_path = [item for pair in path for item in pair] + return _VF.einsum(expr, (preconditioned_gradient, output_gradient, input_activation), path=self.einsum_path) # pylint: disable=no-member + + def compute_self_measurement_score( + self, preconditioned_gradient: torch.Tensor, input_activation: torch.Tensor, output_gradient: torch.Tensor + ) -> torch.Tensor: + input_activation = self._flatten_input_activation(input_activation=input_activation) input_activation = input_activation.view(output_gradient.size(0), -1, input_activation.size(-1)) output_gradient = rearrange(tensor=output_gradient, pattern="b o i1 i2 -> b (i1 i2) o") - return contract("abm,abn->amn", output_gradient, input_activation) + expr = "bio,bci,bco->b" + if self.einsum_path is None: + path = contract_path( + expr, + preconditioned_gradient, + output_gradient, + input_activation, + optimize=DynamicProgramming(search_outer=True, minimize="flops"), + )[0] + self.einsum_path = [item for pair in path for item in pair] + return _VF.einsum(expr, (preconditioned_gradient, output_gradient, input_activation), path=self.einsum_path) # pylint: disable=no-member diff --git a/kronfluence/module/linear.py b/kronfluence/module/linear.py index c1d74a2..2fcafec 100644 --- a/kronfluence/module/linear.py +++ b/kronfluence/module/linear.py @@ -1,38 +1,39 @@ -from typing import Tuple, Union +from typing import Optional, Tuple, Union import torch from einops import rearrange -from opt_einsum import contract -from torch import nn +from opt_einsum import DynamicProgramming, contract_path +from torch import _VF, nn from kronfluence.module.tracked_module import TrackedModule class TrackedLinear(TrackedModule, module_type=nn.Linear): - """A tracking wrapper for `nn.Linear` modules.""" - - def _get_flattened_activation( - self, input_activation: torch.Tensor - ) -> Tuple[torch.Tensor, Union[torch.Tensor, int]]: - """Returns the flattened activation tensor and the number of stacked activations. - - Args: - input_activation (torch.Tensor): - The input tensor to the module, provided by the PyTorch's forward hook. - - Returns: - Tuple[torch.Tensor, torch.Tensor]: - The flattened activation tensor and the number of stacked activations. The flattened - activation is a 2-dimensional matrix with dimension `activation_num x activation_dim`. - """ + """A wrapper for `nn.Linear` modules.""" + + @property + def in_features(self) -> int: # pylint: disable=missing-function-docstring + return self.original_module.in_features + + @property + def out_features(self) -> int: # pylint: disable=missing-function-docstring + return self.original_module.out_features + + @property + def weight(self) -> torch.Tensor: # pylint: disable=missing-function-docstring + return self.original_module.weight + + @property + def bias(self) -> Optional[torch.Tensor]: # pylint: disable=missing-function-docstring + return self.original_module.bias + + def get_flattened_activation(self, input_activation: torch.Tensor) -> Tuple[torch.Tensor, Union[torch.Tensor, int]]: flattened_activation = rearrange(tensor=input_activation, pattern="b ... d_in -> (b ...) d_in") flattened_attention_mask = None - if self._attention_mask is not None and flattened_activation.size(0) == self._attention_mask.numel(): + if self.attention_mask is not None and flattened_activation.size(0) == self.attention_mask.numel(): # If the binary attention mask is provided, zero-out appropriate activations. - flattened_attention_mask = rearrange(tensor=self._attention_mask, pattern="b ... -> (b ...) 1") - # Make sure in-place operation does not change the activation during the forward pass. - flattened_activation = flattened_activation.clone() + flattened_attention_mask = rearrange(tensor=self.attention_mask, pattern="b ... -> (b ...) 1") flattened_activation.mul_(flattened_attention_mask) if self.original_module.bias is not None: @@ -44,41 +45,94 @@ def _get_flattened_activation( count = flattened_activation.size(0) if flattened_attention_mask is None else flattened_attention_mask.sum() return flattened_activation, count - def _get_flattened_gradient(self, output_gradient: torch.Tensor) -> torch.Tensor: - """Returns the flattened gradient tensor. - - Args: - output_gradient (torch.Tensor): - The gradient tensor with respect to the output of the module, provided by the - PyTorch's backward hook. + def get_flattened_gradient(self, output_gradient: torch.Tensor) -> Tuple[torch.Tensor, Union[torch.Tensor, int]]: + flattened_gradient = rearrange(tensor=output_gradient, pattern="b ... d_out -> (b ...) d_out") + if self.attention_mask is not None and flattened_gradient.size(0) == self.attention_mask.numel(): + count = self.attention_mask.sum() + else: + count = flattened_gradient.size(0) + return flattened_gradient, count - Returns: - torch.Tensor: - The flattened output gradient tensor. The flattened gradient is a 2-dimensional matrix - with dimension `gradient_num x gradient_dim`. - """ - return rearrange(tensor=output_gradient, pattern="b ... d_out -> (b ...) d_out") - - def _compute_per_sample_gradient( - self, input_activation: torch.Tensor, output_gradient: torch.Tensor - ) -> torch.Tensor: - """Returns the flattened per-sample-gradient tensor. - - Args: - input_activation (torch.Tensor): - The input tensor to the module, provided by the PyTorch's forward hook. - output_gradient (torch.Tensor): - The gradient tensor with respect to the output of the module, provided by the - PyTorch's backward hook. - - Returns: - torch.Tensor: - The per-sample-gradient tensor. The per-sample-gradient is a 3-dimensional matrix - with dimension `batch_size x gradient_dim x activation_dim`. An additional dimension is added - when the bias term is used. - """ + def _flatten_input_activation(self, input_activation: torch.Tensor) -> torch.Tensor: if self.original_module.bias is not None: shape = list(input_activation.size()[:-1]) + [1] append_term = input_activation.new_ones(shape, requires_grad=False) input_activation = torch.cat([input_activation, append_term], dim=-1) - return contract("b...i,b...o->bio", output_gradient, input_activation) + return input_activation + + def compute_summed_gradient(self, input_activation: torch.Tensor, output_gradient: torch.Tensor) -> torch.Tensor: + input_activation = self._flatten_input_activation(input_activation=input_activation) + summed_gradient = torch.einsum("b...i,b...o->io", output_gradient, input_activation).unsqueeze_(dim=0) + return summed_gradient + + def compute_per_sample_gradient( + self, input_activation: torch.Tensor, output_gradient: torch.Tensor + ) -> torch.Tensor: + input_activation = self._flatten_input_activation(input_activation=input_activation) + per_sample_gradient = torch.einsum("b...i,b...o->bio", output_gradient, input_activation) + if self.per_sample_gradient_process_fnc is not None: + per_sample_gradient = self.per_sample_gradient_process_fnc( + module_name=self.name, gradient=per_sample_gradient + ) + return per_sample_gradient + + def compute_pairwise_score( + self, preconditioned_gradient: torch.Tensor, input_activation: torch.Tensor, output_gradient: torch.Tensor + ) -> torch.Tensor: + input_activation = self._flatten_input_activation(input_activation=input_activation) + if isinstance(preconditioned_gradient, list): + left_mat, right_mat = preconditioned_gradient + if self.score_args.compute_per_token_scores and len(input_activation.shape) == 3: + expr = "qik,qko,bti,bto->qbt" + else: + expr = "qik,qko,b...i,b...o->qb" + if self.einsum_path is None: + path = contract_path( + expr, + left_mat, + right_mat, + output_gradient, + input_activation, + optimize=DynamicProgramming(search_outer=True, minimize="flops"), + )[0] + self.einsum_path = [item for pair in path for item in pair] + return _VF.einsum(expr, (left_mat, right_mat, output_gradient, input_activation), path=self.einsum_path) # pylint: disable=no-member + if self.score_args.compute_per_token_scores and len(input_activation.shape) == 3: + expr = "qio,bti,bto->qbt" + if self.einsum_path is None: + path = contract_path( + expr, + preconditioned_gradient, + output_gradient, + input_activation, + optimize=DynamicProgramming(search_outer=True, minimize="flops"), + )[0] + self.einsum_path = [item for pair in path for item in pair] + return _VF.einsum(expr, (preconditioned_gradient, output_gradient, input_activation), path=self.einsum_path) # pylint: disable=no-member + expr = "qio,b...i,b...o->qb" + if self.einsum_path is None: + path = contract_path( + expr, + preconditioned_gradient, + output_gradient, + input_activation, + optimize=DynamicProgramming(search_outer=True, minimize="flops"), + )[0] + self.einsum_path = [item for pair in path for item in pair] + return _VF.einsum(expr, (preconditioned_gradient, output_gradient, input_activation), path=self.einsum_path) # pylint: disable=no-member + + def compute_self_measurement_score( + self, preconditioned_gradient: torch.Tensor, input_activation: torch.Tensor, output_gradient: torch.Tensor + ) -> torch.Tensor: + input_activation = self._flatten_input_activation(input_activation=input_activation) + expr = "bio,b...i,b...o->b" + if self.einsum_path is None: + path = contract_path( + expr, + preconditioned_gradient, + output_gradient, + input_activation, + optimize=DynamicProgramming(search_outer=True, minimize="flops"), + )[0] + self.einsum_path = [item for pair in path for item in pair] + return _VF.einsum(expr, (preconditioned_gradient, output_gradient, input_activation), path=self.einsum_path) # pylint: disable=no-member diff --git a/kronfluence/module/tracked_module.py b/kronfluence/module/tracked_module.py index f15be0d..22b90c7 100644 --- a/kronfluence/module/tracked_module.py +++ b/kronfluence/module/tracked_module.py @@ -1,37 +1,40 @@ from abc import abstractmethod -from typing import Any, Dict, List, Optional, Tuple, Type, Union +from typing import Any, Callable, Dict, List, Optional, Tuple, Type, Union import torch -import torch.distributed as dist from accelerate.utils.dataclasses import BaseEnum -from opt_einsum import contract from torch import nn -from torch.utils.hooks import RemovableHandle from kronfluence.arguments import FactorArguments, ScoreArguments from kronfluence.factor.config import FactorConfig +from kronfluence.module.tracker.base import BaseTracker +from kronfluence.module.tracker.factor import CovarianceTracker, LambdaTracker +from kronfluence.module.tracker.gradient import GradientTracker +from kronfluence.module.tracker.pairwise_score import PairwiseScoreTracker +from kronfluence.module.tracker.precondition import PreconditionTracker +from kronfluence.module.tracker.self_score import ( + SelfScoreTracker, + SelfScoreWithMeasurementTracker, +) from kronfluence.utils.constants import ( - ACTIVATION_COVARIANCE_MATRIX_NAME, - ACTIVATION_EIGENVECTORS_NAME, - AGGREGATED_PRECONDITIONED_GRADIENT_NAME, + ACCUMULATED_PRECONDITIONED_GRADIENT_NAME, + AGGREGATED_GRADIENT_NAME, COVARIANCE_FACTOR_NAMES, EIGENDECOMPOSITION_FACTOR_NAMES, - GRADIENT_COVARIANCE_MATRIX_NAME, - GRADIENT_EIGENVECTORS_NAME, LAMBDA_FACTOR_NAMES, - LAMBDA_MATRIX_NAME, - NUM_COVARIANCE_PROCESSED, - NUM_LAMBDA_PROCESSED, PAIRWISE_SCORE_MATRIX_NAME, PRECONDITIONED_GRADIENT_NAME, + PRECONDITIONED_GRADIENT_TYPE, SELF_SCORE_VECTOR_NAME, ) -from kronfluence.utils.exceptions import FactorsNotFoundError class ModuleMode(str, BaseEnum): - """Enum to represent a module's mode, indicating which factors need to be computed - during forward and backward passes.""" + """Enum representing a module's mode for factor and score computation. + + This enum indicates which factors and scores need to be computed during + forward and backward passes. + """ DEFAULT = "default" COVARIANCE = "covariance" @@ -40,15 +43,27 @@ class ModuleMode(str, BaseEnum): PAIRWISE_SCORE = "pairwise_score" SELF_SCORE = "self_score" SELF_MEASUREMENT_SCORE = "self_measurement_score" + GRADIENT_AGGREGATION = "gradient_aggregation" class TrackedModule(nn.Module): - """A wrapper class for PyTorch modules to compute preconditioning factors and influence scores.""" + """A wrapper class for PyTorch modules to compute influence factors and scores. + + This class extends `nn.Module` to add functionality for tracking and computing + various influence-related metrics. + """ SUPPORTED_MODULES: Dict[Type[nn.Module], Any] = {} - def __init_subclass__(cls, module_type: Type[nn.Module] = None, **kwargs) -> None: - """Automatically registers subclasses as supported modules.""" + def __init_subclass__(cls, module_type: Type[nn.Module] = None, **kwargs: Any) -> None: + """Automatically registers subclasses as supported modules. + + Args: + module_type (Type[nn.Module], optional): + The type of module this subclass supports. + **kwargs: + Additional keyword arguments. + """ super().__init_subclass__(**kwargs) if module_type is not None: cls.SUPPORTED_MODULES[module_type] = cls @@ -59,233 +74,267 @@ def __init__( original_module: nn.Module, factor_args: Optional[FactorArguments] = None, score_args: Optional[ScoreArguments] = None, + per_sample_gradient_process_fnc: Optional[Callable] = None, ) -> None: - """Initializes an instance of the TrackedModule class. + """Initializes an instance of the `TrackedModule` class. Args: name (str): The original name of the module. original_module (nn.Module): - The original module that will be wrapped with this module. + The original module to be wrapped. factor_args (FactorArguments, optional): - Arguments related to computing the influence factors. + Arguments for computing factors. score_args (ScoreArguments, optional): - Arguments related to computing the influence scores. + Arguments for computing influence scores. + per_sample_gradient_process_fnc (Callable, optional): + Optional function to post-process per-sample gradients. """ super().__init__() self.name = name self.original_module = original_module - - # A way to avoid Autograd computing the gradient with respect to the model parameters. self._constant: torch.Tensor = nn.Parameter( torch.zeros( 1, + dtype=self.original_module.weight.dtype, requires_grad=True, - device=self.original_module.weight.device, - dtype=torch.float16, ) ) - - if factor_args is None: - factor_args = FactorArguments() - self.factor_args = factor_args - if score_args is None: - score_args = ScoreArguments() - self.score_args = score_args - - self._mode: ModuleMode = ModuleMode.DEFAULT - self._cached_activations: List[torch.Tensor] = [] - self._cached_per_sample_gradient: Optional[torch.Tensor] = None - self._attention_mask: Optional[torch.Tensor] = None - self._gradient_scale: float = 1.0 - self._registered_hooks: List[RemovableHandle] = [] - self._cached_hooks: List[RemovableHandle] = [] - self._storage: Dict[str, Optional[Any]] = {} - self._storge_at_current_device: bool = False - - # Storage for activation and pseudo-gradient covariance matrices. # + self.current_mode = ModuleMode.DEFAULT + self.factor_args = FactorArguments() if factor_args is None else factor_args + self.score_args = ScoreArguments() if score_args is None else score_args + self.per_sample_gradient_process_fnc = per_sample_gradient_process_fnc + + self._trackers = { + ModuleMode.DEFAULT: BaseTracker(self), + ModuleMode.COVARIANCE: CovarianceTracker(self), + ModuleMode.LAMBDA: LambdaTracker(self), + ModuleMode.GRADIENT_AGGREGATION: GradientTracker(self), + ModuleMode.PRECONDITION_GRADIENT: PreconditionTracker(self), + ModuleMode.PAIRWISE_SCORE: PairwiseScoreTracker(self), + ModuleMode.SELF_SCORE: SelfScoreTracker(self), + ModuleMode.SELF_MEASUREMENT_SCORE: SelfScoreWithMeasurementTracker(self), + } + + self.attention_mask: Optional[torch.Tensor] = None + self.gradient_scale: float = 1.0 + self.storage: Dict[str, Optional[Union[torch.Tensor, PRECONDITIONED_GRADIENT_TYPE]]] = {} + self.einsum_path: Optional[List[int]] = None + self._initialize_storage() + + def _initialize_storage(self) -> None: + """Initializes storage for various factors and scores.""" + + # Storage for activation and pseudo-gradient covariance matrices # for covariance_factor_name in COVARIANCE_FACTOR_NAMES: - self._storage[covariance_factor_name] = None + self.storage[covariance_factor_name]: Optional[torch.Tensor] = None - # Storage for eigenvectors and eigenvalues. # + # Storage for eigenvectors and eigenvalues # for eigen_factor_name in EIGENDECOMPOSITION_FACTOR_NAMES: - self._storage[eigen_factor_name] = None + self.storage[eigen_factor_name]: Optional[torch.Tensor] = None - # Storage for lambda matrices. # + # Storage for lambda matrices # for lambda_factor_name in LAMBDA_FACTOR_NAMES: - self._storage[lambda_factor_name] = None + self.storage[lambda_factor_name]: Optional[torch.Tensor] = None + + # Storage for preconditioned gradients and influence scores # + self.storage[AGGREGATED_GRADIENT_NAME]: Optional[torch.Tensor] = None + self.storage[PRECONDITIONED_GRADIENT_NAME]: PRECONDITIONED_GRADIENT_TYPE = None + self.storage[ACCUMULATED_PRECONDITIONED_GRADIENT_NAME]: PRECONDITIONED_GRADIENT_TYPE = None + self.storage[PAIRWISE_SCORE_MATRIX_NAME]: Optional[torch.Tensor] = None + self.storage[SELF_SCORE_VECTOR_NAME]: Optional[torch.Tensor] = None + + def forward(self, inputs: torch.Tensor, *args: Any, **kwargs: Any) -> torch.Tensor: + """Performs a forward pass of the tracked module. + + This method should have identical behavior to that of the original module. + + Args: + inputs (torch.Tensor): + Input tensor to the module. + *args: + Variable length argument list. + **kwargs: + Arbitrary keyword arguments. - # Storage for preconditioned query gradients and influence scores. # - self._storage[PRECONDITIONED_GRADIENT_NAME] = None - self._storage[AGGREGATED_PRECONDITIONED_GRADIENT_NAME] = None - self._storage[PAIRWISE_SCORE_MATRIX_NAME] = None - self._storage[SELF_SCORE_VECTOR_NAME] = None + Returns: + torch.Tensor: + The output of the forward pass. + """ + outputs = self.original_module(inputs, *args, **kwargs) + if outputs.requires_grad: + return outputs + return outputs + self._constant + + def prepare_storage(self, device: torch.device) -> None: + """Prepares storage for computing influence scores. + + This method performs necessary operations on storage before computing influence scores. + + Args: + device (torch.device): + The device to prepare the storage for. + """ + FactorConfig.CONFIGS[self.factor_args.strategy].prepare( + storage=self.storage, + score_args=self.score_args, + device=device, + ) def update_factor_args(self, factor_args: FactorArguments) -> None: - """Updates the factor arguments.""" + """Updates the factor arguments. + + Args: + factor_args (FactorArguments): + New factor arguments to set. + """ self.factor_args = factor_args def update_score_args(self, score_args: ScoreArguments) -> None: - """Updates the score arguments.""" + """Updates the score arguments. + + Args: + score_args (ScoreArguments): + New score arguments to set. + """ self.score_args = score_args def get_factor(self, factor_name: str) -> Optional[torch.Tensor]: - """Returns the factor with the given name.""" - if factor_name not in self._storage: + """Retrieves a factor by name from storage. + + Args: + factor_name (str): + The name of the factor to retrieve. + + Returns: + Optional[torch.Tensor]: + The requested factor, or `None` if not found. + """ + if factor_name not in self.storage or self.storage[factor_name] is None: + return None + return self.storage[factor_name] + + def release_factor(self, factor_name: str) -> None: + """Releases a factor from memory. + + Args: + factor_name (str): + The name of the factor to release. + """ + if factor_name not in self.storage or self.storage[factor_name] is None: return None - return self._storage[factor_name] + del self.storage[factor_name] + self.storage[factor_name] = None def set_factor(self, factor_name: str, factor: Any) -> None: - """Sets the factor with the given name.""" - if factor_name in self._storage: - self._storage[factor_name] = factor - - def forward(self, inputs: torch.Tensor, *args, **kwargs) -> Any: - """A forward pass of the tracked module. This should have identical behavior to - the original module.""" - return self.original_module(inputs + self._constant, *args, **kwargs) - - def set_mode(self, mode: ModuleMode, keep_factors: bool = True) -> None: - """Sets the module mode of all `TrackedModule` instances within a model.""" - current_mode = self._mode - self.remove_registered_hooks() - - if current_mode == ModuleMode.COVARIANCE and not keep_factors: - self._release_covariance_matrices() - - if current_mode == ModuleMode.LAMBDA and not keep_factors: - self._release_eigendecomposition_results() - self._release_lambda_matrix() - - if ( - current_mode - in [ - ModuleMode.PRECONDITION_GRADIENT, - ModuleMode.PAIRWISE_SCORE, - ModuleMode.SELF_SCORE, - ModuleMode.SELF_MEASUREMENT_SCORE, - ] - and not keep_factors - ): - self._release_preconditioned_gradient() - self.release_scores() - - if mode == ModuleMode.DEFAULT and not keep_factors: - # Releases all factors when the mode is set to default. - self.remove_attention_mask() - self._release_covariance_matrices() - self._release_eigendecomposition_results() - self._release_lambda_matrix() - self._release_preconditioned_gradient() - self.release_scores() - - if mode == ModuleMode.COVARIANCE: - self._register_covariance_hooks() - - if mode == ModuleMode.LAMBDA: - self._register_lambda_hooks() - - if mode == ModuleMode.PRECONDITION_GRADIENT: - self._register_precondition_gradient_hooks() - - if mode == ModuleMode.PAIRWISE_SCORE: - self._register_pairwise_score_hooks() - - if mode == ModuleMode.SELF_SCORE: - self._register_self_score_hooks() - - if mode == ModuleMode.SELF_MEASUREMENT_SCORE: - self._register_self_measurement_score_hooks() - - self._mode = mode - - def remove_registered_hooks(self) -> None: - """Removes all registered hooks within the module.""" - while self._registered_hooks: - handle = self._registered_hooks.pop() - handle.remove() - self._registered_hooks = [] - while self._cached_hooks: - handle = self._cached_hooks.pop() - handle.remove() - self._cached_hooks = [] + """Sets a factor in storage. - def set_gradient_scale(self, scale: float = 1.0) -> None: - """Sets the scale of the gradient obtained from `GradScaler`.""" - self._gradient_scale = scale + Args: + factor_name (str): + The name of the factor to set. + factor (Any): + The factor value to store. + """ + if factor_name in self.storage: + self.storage[factor_name] = factor - def remove_gradient_scale(self) -> None: - """Resets the scale of the gradient.""" - self._gradient_scale = 1.0 + def set_mode(self, mode: ModuleMode, release_memory: bool = False) -> None: + """Sets the operating mode of the `TrackedModule`. + + This method changes the current mode and manages associated trackers and memory. + + Args: + mode (ModuleMode): + The new mode to set. + release_memory (bool): + Whether to release memory for all trackers. + """ + self._trackers[self.current_mode].release_hooks() + self.einsum_path = None + self.current_mode = mode + + if release_memory: + for _, tracker in self._trackers.items(): + tracker.release_memory() + + self._trackers[self.current_mode].register_hooks() - ############################################## - # Methods for computing covariance matrices. # - ############################################## def set_attention_mask(self, attention_mask: Optional[torch.Tensor] = None) -> None: - """Sets the attention mask for the activation covariance computation.""" - self._attention_mask = attention_mask + """Sets the attention mask for activation covariance computations. - def remove_attention_mask(self) -> None: - """Removes the currently set attention mask.""" - self._attention_mask = None + Args: + attention_mask (torch.Tensor, optional): + The attention mask to set. + """ + self.attention_mask = attention_mask - @abstractmethod - def _get_flattened_activation( - self, input_activation: torch.Tensor - ) -> Tuple[torch.Tensor, Union[torch.Tensor, int]]: - """Returns the flattened activation tensor and the number of stacked activations. + def set_gradient_scale(self, scale: float = 1.0) -> None: + """Sets the scale of the gradient obtained from `GradScaler`. Args: - input_activation (torch.Tensor): - The input tensor to the module, provided by the PyTorch's forward hook. + scale (float): + The scale factor to set. + """ + self.gradient_scale = scale + + def finalize_iteration(self) -> None: + """Finalizes statistics for the current iteration.""" + self._trackers[self.current_mode].finalize_iteration() + + def exist(self) -> bool: + """Checks if the desired statistics are available. Returns: - Tuple[torch.Tensor, torch.Tensor]: - The flattened activation tensor and the number of stacked activations. The flattened - activation is a 2-dimensional matrix with dimension `activation_num x activation_dim`. + bool: + `True` if statistics exist, `False` otherwise. + """ + return self._trackers[self.current_mode].exist() + + def synchronize(self, num_processes: int) -> None: + """Synchronizes statistics across multiple processes. + + Args: + num_processes (int): + The number of processes to synchronize across. + """ + self._trackers[self.current_mode].synchronize(num_processes=num_processes) + + def truncate(self, keep_size: int) -> None: + """Truncates stored statistics to a specified size. + + Args: + keep_size (int): + The number of dimension to keep. """ - raise NotImplementedError("Subclasses must implement the `_get_flattened_activation` method.") + self._trackers[self.current_mode].truncate(keep_size=keep_size) + + def accumulate_iterations(self) -> None: + """Accumulates (or prepares to accumulate) statistics across multiple iterations.""" + self._trackers[self.current_mode].accumulate_iterations() - @torch.no_grad() - def _update_activation_covariance_matrix(self, input_activation: torch.Tensor) -> None: - """Updates the activation covariance matrix. + def finalize_all_iterations(self) -> None: + """Finalizes statistics after all iterations.""" + self._trackers[self.current_mode].finalize_all_iterations() + + @abstractmethod + def get_flattened_activation(self, input_activation: torch.Tensor) -> Tuple[torch.Tensor, Union[torch.Tensor, int]]: + """Returns the flattened activation tensor and the number of stacked activations. Args: input_activation (torch.Tensor): The input tensor to the module, provided by the PyTorch's forward hook. + + Returns: + Tuple[torch.Tensor, Union[torch.Tensor, int]]: + The flattened activation tensor and the number of stacked activations. The flattened + activation is a 2-dimensional matrix with dimension `activation_num x activation_dim`. """ - input_activation = input_activation.to(dtype=self.factor_args.activation_covariance_dtype) - flattened_activation, count = self._get_flattened_activation(input_activation) - - if self._storage[ACTIVATION_COVARIANCE_MATRIX_NAME] is None: - dimension = flattened_activation.size(1) - self._storage[ACTIVATION_COVARIANCE_MATRIX_NAME] = torch.zeros( - size=(dimension, dimension), - dtype=flattened_activation.dtype, - device=flattened_activation.device, - requires_grad=False, - ) - # Add the current batch's activation covariance to the stored activation covariance matrix. - self._storage[ACTIVATION_COVARIANCE_MATRIX_NAME].addmm_(flattened_activation.t(), flattened_activation) - - if self._storage[NUM_COVARIANCE_PROCESSED] is None: - device = None - if isinstance(count, torch.Tensor): - # When using attention masks, `count` can be tensor. - device = count.device - self._storage[NUM_COVARIANCE_PROCESSED] = torch.zeros( - size=(1,), - dtype=torch.int64, - device=device, - requires_grad=False, - ) - # Keep track of total number of elements used to aggregate covariance matrices. - self._storage[NUM_COVARIANCE_PROCESSED].add_(count) + raise NotImplementedError("Subclasses must implement the `get_flattened_activation` method.") @abstractmethod - def _get_flattened_gradient(self, output_gradient: torch.Tensor) -> torch.Tensor: - """Returns the flattened gradient tensor. + def get_flattened_gradient(self, output_gradient: torch.Tensor) -> Tuple[torch.Tensor, Union[torch.Tensor, int]]: + """Returns the flattened output gradient tensor. Args: output_gradient (torch.Tensor): @@ -293,707 +342,75 @@ def _get_flattened_gradient(self, output_gradient: torch.Tensor) -> torch.Tensor PyTorch's backward hook. Returns: - torch.Tensor: - The flattened output gradient tensor. The flattened gradient is a 2-dimensional matrix - with dimension `gradient_num x gradient_dim`. + Tuple[torch.Tensor, Union[torch.Tensor, int]]: + The flattened output gradient tensor and the number of stacked gradients. The flattened + gradient is a 2-dimensional matrix with dimension `gradient_num x gradient_dim`. """ - raise NotImplementedError("Subclasses must implement the `_get_flattened_gradient` method.") + raise NotImplementedError("Subclasses must implement the `get_flattened_gradient` method.") - @torch.no_grad() - def _update_gradient_covariance_matrix(self, output_gradient: torch.Tensor) -> None: - """Updates the pseudo-gradient covariance matrix. + @abstractmethod + def compute_summed_gradient(self, input_activation: torch.Tensor, output_gradient: torch.Tensor) -> torch.Tensor: + """Returns the summed gradient tensor. Args: + input_activation (torch.Tensor): + The input tensor to the module, provided by the PyTorch's forward hook. output_gradient (torch.Tensor): - The gradient tensor with respect to the output of the module, provided by the - PyTorch's backward hook. - """ - output_gradient = output_gradient.to(dtype=self.factor_args.gradient_covariance_dtype) - flattened_gradient = self._get_flattened_gradient(output_gradient) - if self._gradient_scale != 1.0: - # Avoiding in-place operation here. - flattened_gradient = self._gradient_scale * flattened_gradient - - if self._storage[GRADIENT_COVARIANCE_MATRIX_NAME] is None: - # Initialize pseudo-gradient covariance matrix if it does not exist. - dimension = flattened_gradient.size(1) - self._storage[GRADIENT_COVARIANCE_MATRIX_NAME] = torch.zeros( - size=(dimension, dimension), - dtype=flattened_gradient.dtype, - device=flattened_gradient.device, - requires_grad=False, - ) - # Add the current batch's pseudo-gradient covariance to the stored pseudo-gradient covariance matrix. - self._storage[GRADIENT_COVARIANCE_MATRIX_NAME].addmm_(flattened_gradient.t(), flattened_gradient) - - def _register_covariance_hooks(self) -> None: - """Installs forward and backward hooks for computation of the covariance matrices.""" - - def forward_hook(module: nn.Module, inputs: Tuple[torch.Tensor], outputs: Tuple[torch.Tensor]) -> None: - del module - with torch.no_grad(): - # Compute and update activation covariance matrix in the forward pass. - self._update_activation_covariance_matrix(inputs[0].detach()) - # Register backward hook to obtain gradient with respect to the output. - self._cached_hooks.append(outputs.register_hook(backward_hook)) - - @torch.no_grad() - def backward_hook(output_gradient: torch.Tensor) -> None: - handle = self._cached_hooks.pop() - handle.remove() - # Compute and update pseudo-gradient covariance matrix in the backward pass. - self._update_gradient_covariance_matrix(output_gradient.detach()) - - self._registered_hooks.append(self.original_module.register_forward_hook(forward_hook)) - - def _release_covariance_matrices(self) -> None: - """Clears the stored activation and pseudo-gradient covariance matrices from memory.""" - for covariance_factor_name in COVARIANCE_FACTOR_NAMES: - del self._storage[covariance_factor_name] - self._storage[covariance_factor_name] = None - - def _covariance_matrices_available(self) -> bool: - """Checks if the covariance matrices are currently stored in the storage.""" - for covariance_factor_name in COVARIANCE_FACTOR_NAMES: - if self._storage[covariance_factor_name] is None: - return False - return True - - @torch.no_grad() - def synchronize_covariance_matrices(self) -> None: - """Aggregates covariance matrices across multiple devices or nodes in a distributed setting.""" - if dist.is_initialized() and torch.cuda.is_available() and self._covariance_matrices_available(): - # Note that only the main process holds the aggregated covariance matrix. - for covariance_factor_name in COVARIANCE_FACTOR_NAMES: - self._storage[covariance_factor_name] = self._storage[covariance_factor_name].cuda() - dist.reduce( - tensor=self._storage[covariance_factor_name], - op=dist.ReduceOp.SUM, - dst=0, - ) - - ########################################## - # Methods for computing Lambda matrices. # - ########################################## - def _release_eigendecomposition_results(self) -> None: - """Clears the stored eigenvectors and eigenvalues from memory.""" - for eigen_factor_name in EIGENDECOMPOSITION_FACTOR_NAMES: - del self._storage[eigen_factor_name] - self._storage[eigen_factor_name] = None + The gradient tensor with respect to the output of the module, provided by the PyTorch's backward hook. - def _eigendecomposition_results_available(self) -> bool: - """Checks if the eigendecomposition results are currently stored in storage.""" - for eigen_factor_name in EIGENDECOMPOSITION_FACTOR_NAMES: - if self._storage[eigen_factor_name] is None: - return False - return True + Returns: + torch.Tensor: + The aggregated gradient tensor. + """ + raise NotImplementedError("Subclasses must implement the `compute_summed_gradient` method.") @abstractmethod - def _compute_per_sample_gradient( + def compute_per_sample_gradient( self, input_activation: torch.Tensor, output_gradient: torch.Tensor ) -> torch.Tensor: - """Returns the flattened per-sample-gradient tensor. For the brief introduction to - per-sample-gradients, see https://pytorch.org/functorch/stable/notebooks/per_sample_grads.html. + """Returns the flattened per-sample gradient tensor. For a brief introduction to + per-sample gradient, see https://pytorch.org/functorch/stable/notebooks/per_sample_grads.html. Args: input_activation (torch.Tensor): The input tensor to the module, provided by the PyTorch's forward hook. output_gradient (torch.Tensor): - The gradient tensor with respect to the output of the module, provided by the - PyTorch's backward hook. + The gradient tensor with respect to the output of the module, provided by the PyTorch's backward hook. Returns: torch.Tensor: - The per-sample-gradient tensor. The per-sample-gradient is a 3-dimensional matrix - with dimension `batch_size x gradient_dim x activation_dim`. An additional dimension is added - when the bias term is used. + The per-sample gradient tensor. """ - raise NotImplementedError("Subclasses must implement the `_compute_per_sample_gradient` method.") + raise NotImplementedError("Subclasses must implement the `compute_per_sample_gradient` method.") - @torch.no_grad() - def _update_lambda_matrix(self, per_sample_gradient: torch.Tensor) -> None: - """Updates the Lambda matrix using the per-sample-gradient. + @abstractmethod + def compute_pairwise_score( + self, preconditioned_gradient: torch.Tensor, input_activation: torch.Tensor, output_gradient: torch.Tensor + ) -> torch.Tensor: + """Computes pairwise influence scores. Args: - per_sample_gradient (torch.Tensor): - The per-sample-gradient tensor for the given batch. + preconditioned_gradient (torch.Tensor): + The preconditioned gradient. + input_activation (torch.Tensor): + The input tensor to the module, provided by the PyTorch's forward hook. + output_gradient (torch.Tensor): + The gradient tensor with respect to the output of the module, provided by the PyTorch's backward hook. """ - batch_size = per_sample_gradient.size(0) - - if self._storage[LAMBDA_MATRIX_NAME] is None: - # Initialize Lambda matrix if it does not exist. - self._storage[LAMBDA_MATRIX_NAME] = torch.zeros( - size=(per_sample_gradient.size(1), per_sample_gradient.size(2)), - dtype=per_sample_gradient.dtype, - device=per_sample_gradient.device, - requires_grad=False, - ) - - if FactorConfig.CONFIGS[self.factor_args.strategy].requires_eigendecomposition_for_lambda: - if not self._eigendecomposition_results_available(): - error_msg = ( - f"The strategy {self.factor_args.strategy} requires Eigendecomposition " - f"results to be loaded for Lambda computations. However, Eigendecomposition " - f"results are not found." - ) - raise FactorsNotFoundError(error_msg) - - # Move activation and pseudo-gradient eigenvectors to appropriate devices. - self._storage[ACTIVATION_EIGENVECTORS_NAME] = self._storage[ACTIVATION_EIGENVECTORS_NAME].to( - dtype=self.factor_args.lambda_dtype, - device=per_sample_gradient.device, - ) - self._storage[GRADIENT_EIGENVECTORS_NAME] = self._storage[GRADIENT_EIGENVECTORS_NAME].to( - dtype=self.factor_args.lambda_dtype, - device=per_sample_gradient.device, - ) - - if self._storage[NUM_LAMBDA_PROCESSED] is None: - self._storage[NUM_LAMBDA_PROCESSED] = torch.zeros( - size=(1,), - dtype=torch.int64, - requires_grad=False, - ) - - if self._gradient_scale != 1.0: - per_sample_gradient.mul_(self._gradient_scale) - - if FactorConfig.CONFIGS[self.factor_args.strategy].requires_eigendecomposition_for_lambda: - if self.factor_args.lambda_iterative_aggregate: - # This batch-wise iterative update can be useful when the GPU memory is limited. - per_sample_gradient = torch.matmul( - per_sample_gradient, - self._storage[ACTIVATION_EIGENVECTORS_NAME], - ) - for i in range(batch_size): - sqrt_lambda = torch.matmul( - self._storage[GRADIENT_EIGENVECTORS_NAME].t(), - per_sample_gradient[i], - ) - self._storage[LAMBDA_MATRIX_NAME].add_(sqrt_lambda.square_()) - else: - per_sample_gradient = torch.matmul( - self._storage[GRADIENT_EIGENVECTORS_NAME].t(), - torch.matmul(per_sample_gradient, self._storage[ACTIVATION_EIGENVECTORS_NAME]), - ) - self._storage[LAMBDA_MATRIX_NAME].add_(per_sample_gradient.square_().sum(dim=0)) - else: - # Assume that the eigenbasis is identity. - self._storage[LAMBDA_MATRIX_NAME].add_(per_sample_gradient.square_().sum(dim=0)) - - self._storage[NUM_LAMBDA_PROCESSED].add_(batch_size) - - def _register_lambda_hooks(self) -> None: - """Installs forward and backward hooks for computation of the Lambda matrices.""" - - def forward_hook(module: nn.Module, inputs: Tuple[torch.Tensor], outputs: Tuple[torch.Tensor]) -> None: - del module - with torch.no_grad(): - cached_activation = inputs[0].detach().to(dtype=self.factor_args.lambda_dtype) - if self.factor_args.cached_activation_cpu_offload: - self._cached_activations.append(cached_activation.cpu()) - else: - self._cached_activations.append(cached_activation) - # Register backward hook to obtain gradient with respect to the output. - self._cached_hooks.append(outputs.register_hook(backward_hook)) - - @torch.no_grad() - def backward_hook(output_gradient: torch.Tensor) -> None: - handle = self._cached_hooks.pop() - handle.remove() - cached_activation = self._cached_activations.pop() - if self.factor_args.cached_activation_cpu_offload: - cached_activation = cached_activation.to(device=output_gradient.device) - - per_sample_gradient = self._compute_per_sample_gradient( - input_activation=cached_activation, - output_gradient=output_gradient.detach().to(dtype=self.factor_args.lambda_dtype), - ) - del cached_activation - - if self._cached_per_sample_gradient is None: - self._cached_per_sample_gradient = per_sample_gradient - else: - self._cached_per_sample_gradient.add_(per_sample_gradient) - del per_sample_gradient - - if len(self._cached_activations) == 0: - self._update_lambda_matrix(per_sample_gradient=self._cached_per_sample_gradient) - del self._cached_per_sample_gradient - self._cached_per_sample_gradient = None + raise NotImplementedError("Subclasses must implement the `compute_pairwise_score` method.") - self._registered_hooks.append(self.original_module.register_forward_hook(forward_hook)) - - def _release_lambda_matrix(self) -> None: - """Clears the stored Lambda matrix from memory.""" - for lambda_factor_name in LAMBDA_FACTOR_NAMES: - del self._storage[lambda_factor_name] - self._storage[lambda_factor_name] = None - self._cached_activations = [] - del self._cached_per_sample_gradient - self._cached_per_sample_gradient = None - - def _lambda_matrix_available(self) -> bool: - """Checks if the Lamda matrix is currently stored in storage.""" - for lambda_factor_name in LAMBDA_FACTOR_NAMES: - if self._storage[lambda_factor_name] is None: - return False - return True - - @torch.no_grad() - def synchronize_lambda_matrices(self) -> None: - """Aggregates Lambda matrices across multiple devices or nodes in a distributed setting.""" - if dist.is_initialized() and torch.cuda.is_available() and self._lambda_matrix_available(): - # Note that only the main process holds the aggregated Lambda matrix. - for lambda_factor_name in LAMBDA_FACTOR_NAMES: - self._storage[lambda_factor_name] = self._storage[lambda_factor_name].cuda() - torch.distributed.reduce( - tensor=self._storage[lambda_factor_name], - op=dist.ReduceOp.SUM, - dst=0, - ) - - ################################################## - # Methods for computing preconditioned gradient. # - ################################################## - @torch.no_grad() - def _compute_low_rank_preconditioned_gradient( - self, - preconditioned_gradient: torch.Tensor, - ) -> List[torch.Tensor]: - """Performs low-rank approximation of the preconditioned gradient with SVD. + @abstractmethod + def compute_self_measurement_score( + self, preconditioned_gradient: torch.Tensor, input_activation: torch.Tensor, output_gradient: torch.Tensor + ) -> torch.Tensor: + """Computes self-influence scores with measurement. Args: preconditioned_gradient (torch.Tensor): - The preconditioned per-sample-gradient matrix to be low-rank approximated. - - Returns: - List[torch.Tensor, torch.Tensor]: - Low-rank matrices that approximate the original preconditioned gradient. + The preconditioned gradient. + input_activation (torch.Tensor): + The input tensor to the module, provided by the PyTorch's forward hook. + output_gradient (torch.Tensor): + The gradient tensor with respect to the output of the module, provided by the PyTorch's backward hook. """ - U, S, V = torch.linalg.svd( # pylint: disable=not-callable - preconditioned_gradient.contiguous().to(dtype=self.score_args.query_gradient_svd_dtype), - full_matrices=False, - ) - rank = self.score_args.query_gradient_rank - # Avoid holding the full memory of the original tensor before indexing. - U_k = U[:, :, :rank].clone() - S_k = S[:, :rank].clone() - V_k = V[:, :rank, :].clone() - return [ - torch.matmul(U_k, torch.diag_embed(S_k)).to(dtype=self.score_args.score_dtype).contiguous(), - V_k.to(dtype=self.score_args.score_dtype).contiguous(), - ] - - def _register_precondition_gradient_hooks(self) -> None: - """Installs forward and backward hooks for computation of preconditioned per-sample-gradient.""" - - def forward_hook(module: nn.Module, inputs: Tuple[torch.Tensor], outputs: Tuple[torch.Tensor]) -> None: - del module - with torch.no_grad(): - cached_activation = inputs[0].detach().to(dtype=self.score_args.per_sample_gradient_dtype) - if self.score_args.cached_activation_cpu_offload: - self._cached_activations.append(cached_activation.cpu()) - else: - self._cached_activations.append(cached_activation) - # Register backward hook to obtain gradient with respect to the output. - self._cached_hooks.append(outputs.register_hook(backward_hook)) - - @torch.no_grad() - def backward_hook(output_gradient: torch.Tensor) -> None: - handle = self._cached_hooks.pop() - handle.remove() - cached_activation = self._cached_activations.pop() - if self.score_args.cached_activation_cpu_offload: - cached_activation = cached_activation.to(device=output_gradient.device) - - per_sample_gradient = self._compute_per_sample_gradient( - input_activation=cached_activation, - output_gradient=output_gradient.detach().to(dtype=self.score_args.per_sample_gradient_dtype), - ) - del cached_activation - - if self._cached_per_sample_gradient is None: - self._cached_per_sample_gradient = per_sample_gradient - else: - self._cached_per_sample_gradient.add_(per_sample_gradient) - del per_sample_gradient - - if len(self._cached_activations) == 0: - if self._gradient_scale != 1.0: - self._cached_per_sample_gradient.mul_(self._gradient_scale) - - preconditioned_gradient = FactorConfig.CONFIGS[self.factor_args.strategy].precondition_gradient( - gradient=self._cached_per_sample_gradient.to(dtype=self.score_args.precondition_dtype), - storage=self._storage, - damping=self.score_args.damping, - ) - del self._cached_per_sample_gradient - self._cached_per_sample_gradient = None - - if ( - self.score_args.query_gradient_rank is not None - and min(preconditioned_gradient.size()[1:]) > self.score_args.query_gradient_rank - ): - # Apply low-rank approximation to the preconditioned gradient. - preconditioned_gradient = self._compute_low_rank_preconditioned_gradient( - preconditioned_gradient=preconditioned_gradient - ) - self._storage[PRECONDITIONED_GRADIENT_NAME] = preconditioned_gradient - else: - self._storage[PRECONDITIONED_GRADIENT_NAME] = preconditioned_gradient.to( - dtype=self.score_args.score_dtype - ) - - self._registered_hooks.append(self.original_module.register_forward_hook(forward_hook)) - - def aggregate_preconditioned_gradient(self): - """Aggregates the preconditioned per-sample-gradients.""" - if self._storage[PRECONDITIONED_GRADIENT_NAME] is None: - return - - if isinstance(self._storage[PRECONDITIONED_GRADIENT_NAME], list): - if self._storage[AGGREGATED_PRECONDITIONED_GRADIENT_NAME] is not None: - self._storage[AGGREGATED_PRECONDITIONED_GRADIENT_NAME] = [ - torch.cat( - ( - self._storage[AGGREGATED_PRECONDITIONED_GRADIENT_NAME][0], - self._storage[PRECONDITIONED_GRADIENT_NAME][0], - ), - dim=0, - ), - torch.cat( - ( - self._storage[AGGREGATED_PRECONDITIONED_GRADIENT_NAME][1], - self._storage[PRECONDITIONED_GRADIENT_NAME][1], - ), - dim=0, - ), - ] - del self._storage[PRECONDITIONED_GRADIENT_NAME] - self._storage[PRECONDITIONED_GRADIENT_NAME] = None - else: - self._storage[AGGREGATED_PRECONDITIONED_GRADIENT_NAME] = self._storage[PRECONDITIONED_GRADIENT_NAME] - else: - if self._storage[AGGREGATED_PRECONDITIONED_GRADIENT_NAME] is not None: - self._storage[AGGREGATED_PRECONDITIONED_GRADIENT_NAME] = torch.cat( - ( - self._storage[AGGREGATED_PRECONDITIONED_GRADIENT_NAME], - self._storage[PRECONDITIONED_GRADIENT_NAME], - ), - dim=0, - ) - del self._storage[PRECONDITIONED_GRADIENT_NAME] - self._storage[PRECONDITIONED_GRADIENT_NAME] = None - else: - self._storage[AGGREGATED_PRECONDITIONED_GRADIENT_NAME] = self._storage[PRECONDITIONED_GRADIENT_NAME] - - def _release_preconditioned_gradient(self) -> None: - """Clears the preconditioned per-sample-gradient from memory.""" - del self._storage[AGGREGATED_PRECONDITIONED_GRADIENT_NAME] - self._storage[AGGREGATED_PRECONDITIONED_GRADIENT_NAME] = None - del self._storage[PRECONDITIONED_GRADIENT_NAME] - self._storage[PRECONDITIONED_GRADIENT_NAME] = None - self._cached_activations = [] - del self._cached_per_sample_gradient - self._cached_per_sample_gradient = None - - def get_preconditioned_gradient_batch_size(self) -> Optional[int]: - """Returns the saved batch dimension for the preconditioned gradient.""" - if self._storage[PRECONDITIONED_GRADIENT_NAME] is not None: - if isinstance(self._storage[PRECONDITIONED_GRADIENT_NAME], list): - return self._storage[PRECONDITIONED_GRADIENT_NAME][0].size(0) - return self._storage[PRECONDITIONED_GRADIENT_NAME].size(0) - return None - - @torch.no_grad() - def truncate_preconditioned_gradient(self, keep_size: int) -> None: - """Truncates and keeps only the first keep_size dimension for the preconditioned gradient.""" - if self._storage[PRECONDITIONED_GRADIENT_NAME] is not None: - if isinstance(self._storage[PRECONDITIONED_GRADIENT_NAME], list): - assert len(self._storage[PRECONDITIONED_GRADIENT_NAME]) == 2 - self._storage[PRECONDITIONED_GRADIENT_NAME] = [ - self._storage[PRECONDITIONED_GRADIENT_NAME][0][:keep_size].clone(), - self._storage[PRECONDITIONED_GRADIENT_NAME][1][:keep_size].clone(), - ] - else: - self._storage[PRECONDITIONED_GRADIENT_NAME] = self._storage[PRECONDITIONED_GRADIENT_NAME][ - :keep_size - ].clone() - - def _preconditioned_gradient_available(self) -> bool: - """Checks if the preconditioned matrices are currently stored in the storage.""" - return self._storage[PRECONDITIONED_GRADIENT_NAME] is not None - - @torch.no_grad() - def synchronize_preconditioned_gradient(self, num_processes: int) -> None: - """Stacks preconditioned gradient across multiple devices or nodes in a distributed setting.""" - if dist.is_initialized() and torch.cuda.is_available() and self._preconditioned_gradient_available(): - if isinstance(self._storage[PRECONDITIONED_GRADIENT_NAME], list): - assert len(self._storage[PRECONDITIONED_GRADIENT_NAME]) == 2 - for i in range(len(self._storage[PRECONDITIONED_GRADIENT_NAME])): - size = self._storage[PRECONDITIONED_GRADIENT_NAME][i].size() - stacked_matrix = torch.empty( - size=(num_processes, size[0], size[1], size[2]), - dtype=self._storage[PRECONDITIONED_GRADIENT_NAME][i].dtype, - device=self._storage[PRECONDITIONED_GRADIENT_NAME][i].device, - ) - torch.distributed.all_gather_into_tensor( - output_tensor=stacked_matrix, - input_tensor=self._storage[PRECONDITIONED_GRADIENT_NAME][i].contiguous(), - ) - self._storage[PRECONDITIONED_GRADIENT_NAME][i] = ( - stacked_matrix.transpose(0, 1) - .reshape(num_processes * size[0], size[1], size[2]) - .contiguous() - .clone() - ) - else: - size = self._storage[PRECONDITIONED_GRADIENT_NAME].size() - stacked_preconditioned_gradient = torch.empty( - size=(num_processes, size[0], size[1], size[2]), - dtype=self._storage[PRECONDITIONED_GRADIENT_NAME].dtype, - device=self._storage[PRECONDITIONED_GRADIENT_NAME].device, - ) - torch.distributed.all_gather_into_tensor( - output_tensor=stacked_preconditioned_gradient, - input_tensor=self._storage[PRECONDITIONED_GRADIENT_NAME].contiguous(), - ) - self._storage[PRECONDITIONED_GRADIENT_NAME] = ( - stacked_preconditioned_gradient.transpose(0, 1) - .reshape(num_processes * size[0], size[1], size[2]) - .contiguous() - .clone() - ) - - ########################################### - # Methods for computing influence scores. # - ########################################### - def _register_pairwise_score_hooks(self) -> None: - """Installs forward and backward hooks for computation of pairwise influence scores.""" - - def forward_hook(module: nn.Module, inputs: Tuple[torch.Tensor], outputs: Tuple[torch.Tensor]) -> None: - del module - with torch.no_grad(): - cached_activation = inputs[0].detach().to(dtype=self.score_args.per_sample_gradient_dtype) - if self.score_args.cached_activation_cpu_offload: - self._cached_activations.append(cached_activation.cpu()) - else: - self._cached_activations.append(cached_activation) - # Register backward hook to obtain gradient with respect to the output. - self._cached_hooks.append(outputs.register_hook(backward_hook)) - - @torch.no_grad() - def backward_hook(output_gradient: torch.Tensor) -> None: - handle = self._cached_hooks.pop() - handle.remove() - cached_activation = self._cached_activations.pop() - if self.score_args.cached_activation_cpu_offload: - cached_activation = cached_activation.to(device=output_gradient.device) - - per_sample_gradient = self._compute_per_sample_gradient( - input_activation=cached_activation, - output_gradient=output_gradient.detach().to(dtype=self.score_args.per_sample_gradient_dtype), - ).to(dtype=self.score_args.score_dtype) - del cached_activation - - if self._cached_per_sample_gradient is None: - self._cached_per_sample_gradient = per_sample_gradient - else: - self._cached_per_sample_gradient.add_(per_sample_gradient) - del per_sample_gradient - - # If the module was used multiple times throughout the forward pass, - # only compute scores after aggregating all per-sample-gradients. - if len(self._cached_activations) == 0: - if self._gradient_scale != 1.0: - self._cached_per_sample_gradient.mul_(self._gradient_scale) - - if isinstance(self._storage[AGGREGATED_PRECONDITIONED_GRADIENT_NAME], list): - # The preconditioned gradient is stored as a low-rank approximation. - left_mat, right_mat = self._storage[AGGREGATED_PRECONDITIONED_GRADIENT_NAME] - - input_dim = right_mat.size(2) - output_dim = left_mat.size(1) - query_batch_size = left_mat.size(0) - train_batch_size = self._cached_per_sample_gradient.size(0) - rank = self.score_args.query_gradient_rank - if ( - train_batch_size * query_batch_size * rank * min((input_dim, output_dim)) - > query_batch_size * input_dim * output_dim - ): - # If reconstructing the gradient is more memory efficient, reconstruct and compute the score. - self._storage[PAIRWISE_SCORE_MATRIX_NAME] = contract( - "qki,toi,qok->qt", - right_mat, - self._cached_per_sample_gradient, - left_mat, - ) - # Otherwise, try to avoid reconstructing the full per-sample-gradient. - elif output_dim >= input_dim: - intermediate = contract("toi,qok->qtik", self._cached_per_sample_gradient, left_mat) - self._storage[PAIRWISE_SCORE_MATRIX_NAME] = contract("qki,qtik->qt", right_mat, intermediate) - else: - intermediate = contract("qki,toi->qtko", right_mat, self._cached_per_sample_gradient) - self._storage[PAIRWISE_SCORE_MATRIX_NAME] = contract("qtko,qok->qt", intermediate, left_mat) - else: - self._storage[PAIRWISE_SCORE_MATRIX_NAME] = contract( - "qio,tio->qt", - self._storage[AGGREGATED_PRECONDITIONED_GRADIENT_NAME], - self._cached_per_sample_gradient, - ) - del self._cached_per_sample_gradient - self._cached_per_sample_gradient = None - - self._registered_hooks.append(self.original_module.register_forward_hook(forward_hook)) - - def _register_self_score_hooks(self) -> None: - """Installs forward and backward hooks for computation of self-influence scores.""" - - def forward_hook(module: nn.Module, inputs: Tuple[torch.Tensor], outputs: Tuple[torch.Tensor]) -> None: - del module - with torch.no_grad(): - cached_activation = inputs[0].detach().to(dtype=self.score_args.per_sample_gradient_dtype) - if self.score_args.cached_activation_cpu_offload: - self._cached_activations.append(cached_activation.cpu()) - else: - self._cached_activations.append(cached_activation) - # Register backward hook to obtain gradient with respect to the output. - self._cached_hooks.append(outputs.register_hook(backward_hook)) - - @torch.no_grad() - def backward_hook(output_gradient: torch.Tensor) -> None: - handle = self._cached_hooks.pop() - handle.remove() - cached_activation = self._cached_activations.pop() - if self.score_args.cached_activation_cpu_offload: - cached_activation = cached_activation.to(device=output_gradient.device) - - per_sample_gradient = self._compute_per_sample_gradient( - input_activation=cached_activation, - output_gradient=output_gradient.detach().to(dtype=self.score_args.per_sample_gradient_dtype), - ) - del cached_activation - - # The preconditioning factors need to be loaded to appropriate device as they will be - # used at each iteration. - if not self._storge_at_current_device: - self._move_storage_to_device(target_device=per_sample_gradient.device) - self._storge_at_current_device = True - - if self._cached_per_sample_gradient is None: - self._cached_per_sample_gradient = per_sample_gradient - else: - self._cached_per_sample_gradient.add_(per_sample_gradient) - del per_sample_gradient - - # If the module was used multiple times throughout the forward pass, - # only compute scores after aggregating all per-sample-gradients. - if len(self._cached_activations) == 0: - if self._gradient_scale != 1.0: - self._cached_per_sample_gradient.mul_(self._gradient_scale) - - preconditioned_gradient = ( - FactorConfig.CONFIGS[self.factor_args.strategy] - .precondition_gradient( - gradient=self._cached_per_sample_gradient.to(dtype=self.score_args.precondition_dtype), - storage=self._storage, - damping=self.score_args.damping, - ) - .to(dtype=self.score_args.score_dtype) - ) - self._cached_per_sample_gradient = self._cached_per_sample_gradient.to( - dtype=self.score_args.score_dtype - ) - preconditioned_gradient.mul_(self._cached_per_sample_gradient) - self._storage[SELF_SCORE_VECTOR_NAME] = preconditioned_gradient.sum(dim=(1, 2)) - self._cached_per_sample_gradient = None - - self._registered_hooks.append(self.original_module.register_forward_hook(forward_hook)) - - def _register_self_measurement_score_hooks(self) -> None: - """Installs forward and backward hooks for computation of self-influence scores.""" - - def forward_hook(module: nn.Module, inputs: Tuple[torch.Tensor], outputs: Tuple[torch.Tensor]) -> None: - del module - with torch.no_grad(): - cached_activation = inputs[0].detach().to(dtype=self.score_args.per_sample_gradient_dtype) - if self.score_args.cached_activation_cpu_offload: - self._cached_activations.append(cached_activation.cpu()) - else: - self._cached_activations.append(cached_activation) - # Register backward hook to obtain gradient with respect to the output. - self._cached_hooks.append(outputs.register_hook(backward_hook)) - - @torch.no_grad() - def backward_hook(output_gradient: torch.Tensor) -> None: - handle = self._cached_hooks.pop() - handle.remove() - cached_activation = self._cached_activations.pop() - if self.score_args.cached_activation_cpu_offload: - cached_activation = cached_activation.to(device=output_gradient.device) - - per_sample_gradient = self._compute_per_sample_gradient( - input_activation=cached_activation, - output_gradient=output_gradient.detach().to(dtype=self.score_args.per_sample_gradient_dtype), - ) - del cached_activation - - # The preconditioning factors need to be loaded to appropriate device as they will be - # used at each iteration. - if not self._storge_at_current_device: - self._move_storage_to_device(target_device=per_sample_gradient.device) - self._storge_at_current_device = True - - if self._cached_per_sample_gradient is None: - self._cached_per_sample_gradient = per_sample_gradient - else: - self._cached_per_sample_gradient.add_(per_sample_gradient) - del per_sample_gradient - - # If the module was used multiple times throughout the forward pass, - # only compute scores after aggregating all per-sample-gradients. - if len(self._cached_activations) == 0: - if self._gradient_scale != 1.0: - self._cached_per_sample_gradient.mul_(self._gradient_scale) - - self._cached_per_sample_gradient = self._cached_per_sample_gradient.to( - dtype=self.score_args.score_dtype - ) - self._storage[SELF_SCORE_VECTOR_NAME] = self._cached_per_sample_gradient.mul_( - self._storage[PRECONDITIONED_GRADIENT_NAME] - ).sum(dim=(1, 2)) - self._cached_per_sample_gradient = None - del self._storage[PRECONDITIONED_GRADIENT_NAME] - self._storage[PRECONDITIONED_GRADIENT_NAME] = None - - self._registered_hooks.append(self.original_module.register_forward_hook(forward_hook)) - - def _move_storage_to_device(self, target_device: torch.device) -> None: - """Moves stored factors into the target device.""" - for name, factor in self._storage.items(): - if factor is not None: - if isinstance(factor, torch.Tensor): - self._storage[name] = factor.to( - device=target_device, - dtype=self.score_args.precondition_dtype, - ) - elif isinstance(factor, list): - for i in range(len(self._storage[name])): - self._storage[name][i] = factor[i].to( - device=target_device, - dtype=self.score_args.precondition_dtype, - ) - else: - raise RuntimeError(f"`{name}` in `TrackedModule` storage does not have a valid type.") - - def release_scores(self) -> None: - """Clears the influence scores from memory.""" - del self._storage[PAIRWISE_SCORE_MATRIX_NAME] - self._storage[PAIRWISE_SCORE_MATRIX_NAME] = None - del self._storage[SELF_SCORE_VECTOR_NAME] - self._storage[SELF_SCORE_VECTOR_NAME] = None - self._cached_activations = [] - del self._cached_per_sample_gradient - self._cached_per_sample_gradient = None - self._storge_at_current_device = False + raise NotImplementedError("Subclasses must implement the `compute_self_measurement_score` method.") diff --git a/kronfluence/module/tracker/__init__.py b/kronfluence/module/tracker/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/kronfluence/module/tracker/base.py b/kronfluence/module/tracker/base.py new file mode 100644 index 0000000..0d7c091 --- /dev/null +++ b/kronfluence/module/tracker/base.py @@ -0,0 +1,88 @@ +from typing import List, Optional, Union + +import torch +from torch import nn +from torch.utils.hooks import RemovableHandle + + +class BaseTracker: + """Base class for tracking module activations, gradients, and scores.""" + + def __init__(self, module: nn.Module) -> None: + """Initializes an instance of the `BaseTracker` class. + + Args: + module (TrackedModule): + The `TrackedModule` that wraps the original module. + """ + self.module = module + self.registered_hooks: List[RemovableHandle] = [] + self.cached_hooks: List[RemovableHandle] = [] + self.cached_activations: Optional[Union[List[torch.Tensor]], torch.Tensor] = None + self.cached_per_sample_gradient: Optional[torch.Tensor] = None + + def release_hooks(self) -> None: + """Removes all registered hooks.""" + self.clear_all_cache() + while self.registered_hooks: + handle = self.registered_hooks.pop() + handle.remove() + self.registered_hooks = [] + + def clear_all_cache(self) -> None: + """Clears all cached data and removes cached hooks.""" + del self.cached_activations, self.cached_per_sample_gradient + self.cached_activations, self.cached_per_sample_gradient = None, None + while self.cached_hooks: + handle = self.cached_hooks.pop() + handle.remove() + self.cached_hooks = [] + + def _raise_cache_not_found_exception(self) -> None: + """Raises an exception when cached activations are not found.""" + raise RuntimeError( + f"Module '{self.module.name}' has no cached activations. This can occur if:\n" + f"1. The module was not used during the forward pass, or\n" + f"2. The module was encountered multiple times in the forward pass.\n" + f"For case 2, set 'has_shared_parameters=True' to enable parameter sharing." + ) + + def register_hooks(self) -> None: + """Registers hooks for the module.""" + + def finalize_iteration(self) -> None: + """Finalizes statistics for the current iteration.""" + + def exist(self) -> bool: + """Checks if the desired statistics are available. + + Returns: + bool: + `True` if statistics exist, `False` otherwise. + """ + return False + + def synchronize(self, num_processes: int) -> None: + """Synchronizes statistics across multiple processes. + + Args: + num_processes (int): + The number of processes to synchronize across. + """ + + def truncate(self, keep_size: int) -> None: + """Truncates stored statistics to a specified size. + + Args: + keep_size (int): + The number of dimensions to keep. + """ + + def accumulate_iterations(self) -> None: + """Accumulates (or prepares to accumulate) statistics across multiple iterations.""" + + def finalize_all_iterations(self) -> None: + """Finalizes statistics after all iterations.""" + + def release_memory(self) -> None: + """Releases any memory held by the tracker.""" diff --git a/kronfluence/module/tracker/factor.py b/kronfluence/module/tracker/factor.py new file mode 100644 index 0000000..0e4a9c2 --- /dev/null +++ b/kronfluence/module/tracker/factor.py @@ -0,0 +1,327 @@ +from typing import Tuple, Union + +import torch +import torch.distributed as dist +from torch import nn + +from kronfluence.factor.config import FactorConfig +from kronfluence.module.tracker.base import BaseTracker +from kronfluence.utils.constants import ( + ACTIVATION_COVARIANCE_MATRIX_NAME, + ACTIVATION_EIGENVECTORS_NAME, + COVARIANCE_FACTOR_NAMES, + EIGENDECOMPOSITION_FACTOR_NAMES, + GRADIENT_COVARIANCE_MATRIX_NAME, + GRADIENT_EIGENVECTORS_NAME, + LAMBDA_FACTOR_NAMES, + LAMBDA_MATRIX_NAME, + NUM_ACTIVATION_COVARIANCE_PROCESSED, + NUM_GRADIENT_COVARIANCE_PROCESSED, + NUM_LAMBDA_PROCESSED, +) +from kronfluence.utils.exceptions import FactorsNotFoundError + + +class CovarianceTracker(BaseTracker): + """Tracks and computes activation and gradient covariance matrices for a given module.""" + + _activation_covariance_initialized: bool = False + _gradient_covariance_initialized: bool = False + + def _update_activation_covariance_matrix( + self, input_activation: torch.Tensor, count: Union[torch.Tensor, int] + ) -> None: + """Computes and updates the activation covariance matrix. + + Args: + input_activation (torch.Tensor): + The flattened input tensor to the module, provided by PyTorch's forward hook. + count (int): + The number of activations. + """ + if not self._activation_covariance_initialized: + self.module.storage[NUM_ACTIVATION_COVARIANCE_PROCESSED] = torch.zeros( + size=(1,), + dtype=torch.int64, + device=count.device if isinstance(count, torch.Tensor) else None, + requires_grad=False, + ) + dimension = input_activation.size(1) + self.module.storage[ACTIVATION_COVARIANCE_MATRIX_NAME] = torch.zeros( + size=(dimension, dimension), + dtype=input_activation.dtype, + device=input_activation.device, + requires_grad=False, + ) + self._activation_covariance_initialized = True + self.module.storage[NUM_ACTIVATION_COVARIANCE_PROCESSED].add_(count) + self.module.storage[ACTIVATION_COVARIANCE_MATRIX_NAME].addmm_(input_activation.t(), input_activation) + + def _update_gradient_covariance_matrix( + self, output_gradient: torch.Tensor, count: Union[torch.Tensor, int] + ) -> None: + """Computes and updates the pseudo-gradient covariance matrix. + + Args: + output_gradient (torch.Tensor): + The flattened gradient tensor with respect to the output of the module, provided + by PyTorch's backward hook. + count (int): + The number of gradients. + """ + if not self._gradient_covariance_initialized: + # In most cases, `NUM_GRADIENT_COVARIANCE_PROCESSED` and `NUM_ACTIVATION_COVARIANCE_PROCESSED` are + # identical. However, they may differ when using gradient checkpointing or `torch.compile()`. + self.module.storage[NUM_GRADIENT_COVARIANCE_PROCESSED] = torch.zeros( + size=(1,), + dtype=torch.int64, + device=count.device if isinstance(count, torch.Tensor) else None, + requires_grad=False, + ) + dimension = output_gradient.size(1) + self.module.storage[GRADIENT_COVARIANCE_MATRIX_NAME] = torch.zeros( + size=(dimension, dimension), + dtype=output_gradient.dtype, + device=output_gradient.device, + requires_grad=False, + ) + self._gradient_covariance_initialized = True + self.module.storage[NUM_GRADIENT_COVARIANCE_PROCESSED].add_(count) + alpha = 1 + if self.module.gradient_scale != 1.0: + alpha = self.module.gradient_scale**2.0 + self.module.storage[GRADIENT_COVARIANCE_MATRIX_NAME].addmm_(output_gradient.t(), output_gradient, alpha=alpha) + + def register_hooks(self) -> None: + """Sets up hooks to compute activation and gradient covariance matrices.""" + + @torch.no_grad() + def forward_hook(module: nn.Module, inputs: Tuple[torch.Tensor], outputs: torch.Tensor) -> None: + del module + input_activation = ( + inputs[0] + .detach() + .to( + dtype=self.module.factor_args.activation_covariance_dtype, + copy=self.module.attention_mask is not None, + ) + ) + # Computes and updates activation covariance during forward pass. + input_activation, count = self.module.get_flattened_activation(input_activation=input_activation) + self._update_activation_covariance_matrix(input_activation=input_activation, count=count) + self.cached_hooks.append(outputs.register_hook(backward_hook)) + + @torch.no_grad() + def backward_hook(output_gradient: torch.Tensor) -> None: + handle = self.cached_hooks.pop() + handle.remove() + output_gradient = output_gradient.detach().to(dtype=self.module.factor_args.gradient_covariance_dtype) + # Computes and updates pseudo-gradient covariance during backward pass. + output_gradient, count = self.module.get_flattened_gradient(output_gradient=output_gradient) + self._update_gradient_covariance_matrix(output_gradient=output_gradient, count=count) + + self.registered_hooks.append(self.module.register_forward_hook(forward_hook)) + + def exist(self) -> bool: + """Checks if both activation and gradient covariance matrices are available.""" + for covariance_factor_name in COVARIANCE_FACTOR_NAMES: + if self.module.storage[covariance_factor_name] is None: + return False + return True + + def synchronize(self, num_processes: int) -> None: + """Aggregates covariance matrices across multiple devices or nodes in a distributed setting.""" + del num_processes + if dist.is_initialized() and torch.cuda.is_available() and self.exist(): + for covariance_factor_name in COVARIANCE_FACTOR_NAMES: + self.module.storage[covariance_factor_name] = self.module.storage[covariance_factor_name].cuda() + dist.reduce( + tensor=self.module.storage[covariance_factor_name], + op=dist.ReduceOp.SUM, + dst=0, + ) + + def release_memory(self) -> None: + """Clears all covariance matrices from memory.""" + self._activation_covariance_initialized = False + self._gradient_covariance_initialized = False + for covariance_factor_name in COVARIANCE_FACTOR_NAMES: + self.module.storage[covariance_factor_name] = None + + +class LambdaTracker(BaseTracker): + """Tracks and computes Lambda matrices for a given module.""" + + def _eigendecomposition_results_exist(self) -> bool: + """Checks if eigendecomposition results are available.""" + for eigen_factor_name in EIGENDECOMPOSITION_FACTOR_NAMES: + if self.module.storage[eigen_factor_name] is None: + return False + return True + + def _update_lambda_matrix(self, per_sample_gradient: torch.Tensor) -> None: + """Computes and updates the Lambda matrix using provided per-sample gradient. + + Args: + per_sample_gradient (torch.Tensor): + The per-sample gradient tensor for the given batch. + """ + batch_size = per_sample_gradient.size(0) + + if self.module.storage[NUM_LAMBDA_PROCESSED] is None: + self.module.storage[NUM_LAMBDA_PROCESSED] = torch.zeros( + size=(1,), + dtype=torch.int64, + requires_grad=False, + ) + self.module.storage[LAMBDA_MATRIX_NAME] = torch.zeros( + size=(per_sample_gradient.size(1), per_sample_gradient.size(2)), + dtype=per_sample_gradient.dtype, + device=per_sample_gradient.device, + requires_grad=False, + ) + + if FactorConfig.CONFIGS[self.module.factor_args.strategy].requires_eigendecomposition_for_lambda: + if not self._eigendecomposition_results_exist(): + raise FactorsNotFoundError( + f"The strategy {self.module.factor_args.strategy} requires eigendecomposition " + f"results for Lambda computations, but they are not found." + ) + + # Move activation and pseudo-gradient eigenvectors to appropriate devices. + self.module.storage[ACTIVATION_EIGENVECTORS_NAME] = self.module.storage[ + ACTIVATION_EIGENVECTORS_NAME + ].to( + dtype=per_sample_gradient.dtype, + device=per_sample_gradient.device, + ) + self.module.storage[GRADIENT_EIGENVECTORS_NAME] = self.module.storage[GRADIENT_EIGENVECTORS_NAME].to( + dtype=per_sample_gradient.dtype, + device=per_sample_gradient.device, + ) + + self.module.storage[NUM_LAMBDA_PROCESSED].add_(batch_size) + if FactorConfig.CONFIGS[self.module.factor_args.strategy].requires_eigendecomposition_for_lambda: + if self.module.factor_args.use_iterative_lambda_aggregation: + # This batch-wise iterative update can be useful when the GPU memory is limited. + per_sample_gradient = torch.matmul( + per_sample_gradient, + self.module.storage[ACTIVATION_EIGENVECTORS_NAME], + ) + for i in range(batch_size): + sqrt_lambda = torch.matmul( + self.module.storage[GRADIENT_EIGENVECTORS_NAME].t(), + per_sample_gradient[i], + ) + self.module.storage[LAMBDA_MATRIX_NAME].add_(sqrt_lambda.square_()) + else: + per_sample_gradient = ( + torch.matmul( + self.module.storage[GRADIENT_EIGENVECTORS_NAME].t(), + torch.matmul(per_sample_gradient, self.module.storage[ACTIVATION_EIGENVECTORS_NAME]), + ) + .square_() + .sum(dim=0) + ) + self.module.storage[LAMBDA_MATRIX_NAME].add_(per_sample_gradient) + else: + # Approximate the eigenbasis as identity. + per_sample_gradient = per_sample_gradient.square_().sum(dim=0) + self.module.storage[LAMBDA_MATRIX_NAME].add_(per_sample_gradient) + + def register_hooks(self) -> None: + """Sets up hooks to compute lambda matrices.""" + + @torch.no_grad() + def forward_hook(module: nn.Module, inputs: Tuple[torch.Tensor], outputs: torch.Tensor) -> None: + del module + cached_activation = inputs[0].detach() + device = "cpu" if self.module.factor_args.offload_activations_to_cpu else cached_activation.device + cached_activation = cached_activation.to( + device=device, + dtype=self.module.factor_args.per_sample_gradient_dtype, + copy=True, + ) + if self.module.factor_args.has_shared_parameters: + if self.cached_activations is None: + self.cached_activations = [] + self.cached_activations.append(cached_activation) + else: + self.cached_activations = cached_activation + self.cached_hooks.append( + outputs.register_hook( + shared_backward_hook if self.module.factor_args.has_shared_parameters else backward_hook + ) + ) + + @torch.no_grad() + def backward_hook(output_gradient: torch.Tensor) -> None: + if self.cached_activations is None: + self._raise_cache_not_found_exception() + handle = self.cached_hooks.pop() + handle.remove() + output_gradient = output_gradient.detach().to(dtype=self.module.factor_args.per_sample_gradient_dtype) + per_sample_gradient = self.module.compute_per_sample_gradient( + input_activation=self.cached_activations.to(device=output_gradient.device), + output_gradient=output_gradient, + ).to(dtype=self.module.factor_args.lambda_dtype) + self.clear_all_cache() + del output_gradient + if self.module.gradient_scale != 1.0: + per_sample_gradient.mul_(self.module.gradient_scale) + # Computes and updates lambda matrix during backward pass. + self._update_lambda_matrix(per_sample_gradient=per_sample_gradient) + + @torch.no_grad() + def shared_backward_hook(output_gradient: torch.Tensor) -> None: + handle = self.cached_hooks.pop() + handle.remove() + output_gradient = output_gradient.detach().to(dtype=self.module.factor_args.per_sample_gradient_dtype) + cached_activation = self.cached_activations.pop() + per_sample_gradient = self.module.compute_per_sample_gradient( + input_activation=cached_activation.to(device=output_gradient.device), + output_gradient=output_gradient, + ) + if self.cached_per_sample_gradient is None: + self.cached_per_sample_gradient = torch.zeros_like(per_sample_gradient, requires_grad=False) + # Aggregates per-sample gradients during backward pass. + self.cached_per_sample_gradient.add_(per_sample_gradient) + + self.registered_hooks.append(self.module.register_forward_hook(forward_hook)) + + @torch.no_grad() + def finalize_iteration(self) -> None: + """Updates Lambda matrix using cached per-sample gradients.""" + if self.module.factor_args.has_shared_parameters: + self.cached_per_sample_gradient = self.cached_per_sample_gradient.to( + dtype=self.module.factor_args.lambda_dtype + ) + if self.module.gradient_scale != 1.0: + self.cached_per_sample_gradient.mul_(self.module.gradient_scale) + self._update_lambda_matrix(per_sample_gradient=self.cached_per_sample_gradient) + self.clear_all_cache() + + def exist(self) -> bool: + """Checks if Lambda matrices are available.""" + for lambda_factor_name in LAMBDA_FACTOR_NAMES: + if self.module.storage[lambda_factor_name] is None: + return False + return True + + def synchronize(self, num_processes: int) -> None: + """Aggregates Lambda matrices across multiple devices or nodes in a distributed setting.""" + del num_processes + if dist.is_initialized() and torch.cuda.is_available() and self.exist(): + for lambda_factor_name in LAMBDA_FACTOR_NAMES: + self.module.storage[lambda_factor_name] = self.module.storage[lambda_factor_name].cuda() + dist.reduce( + tensor=self.module.storage[lambda_factor_name], + op=dist.ReduceOp.SUM, + dst=0, + ) + + def release_memory(self) -> None: + """Clears Lambda matrices from memory.""" + self.clear_all_cache() + for lambda_factor_name in LAMBDA_FACTOR_NAMES: + self.module.storage[lambda_factor_name] = None diff --git a/kronfluence/module/tracker/gradient.py b/kronfluence/module/tracker/gradient.py new file mode 100644 index 0000000..712cc7e --- /dev/null +++ b/kronfluence/module/tracker/gradient.py @@ -0,0 +1,93 @@ +from typing import Tuple + +import torch +import torch.distributed as dist +from torch import nn + +from kronfluence.module.tracker.base import BaseTracker +from kronfluence.utils.constants import AGGREGATED_GRADIENT_NAME + + +class GradientTracker(BaseTracker): + """Tracks and computes aggregated gradient for a given module.""" + + def register_hooks(self) -> None: + """Sets up hooks to compute and keep track of aggregated gradient.""" + + @torch.no_grad() + def forward_hook(module: nn.Module, inputs: Tuple[torch.Tensor], outputs: torch.Tensor) -> None: + del module + cached_activation = inputs[0].detach() + device = "cpu" if self.module.score_args.offload_activations_to_cpu else cached_activation.device + cached_activation = cached_activation.to( + device=device, + dtype=self.module.score_args.per_sample_gradient_dtype, + copy=True, + ) + if self.module.factor_args.has_shared_parameters: + if self.cached_activations is None: + self.cached_activations = [] + self.cached_activations.append(cached_activation) + else: + self.cached_activations = cached_activation + self.cached_hooks.append(outputs.register_hook(backward_hook)) + + @torch.no_grad() + def backward_hook(output_gradient: torch.Tensor) -> None: + if self.cached_activations is None: + self._raise_cache_not_found_exception() + handle = self.cached_hooks.pop() + handle.remove() + output_gradient = output_gradient.detach().to(dtype=self.module.score_args.per_sample_gradient_dtype) + if isinstance(self.cached_activations, list): + cached_activation = self.cached_activations.pop() + else: + cached_activation = self.cached_activations + if self.module.per_sample_gradient_process_fnc is None: + summed_gradient = self.module.compute_summed_gradient( + input_activation=cached_activation.to(device=output_gradient.device), + output_gradient=output_gradient, + ) + self.clear_all_cache() + else: + summed_gradient = self.module.compute_per_sample_gradient( + input_activation=cached_activation.to(device=output_gradient.device), + output_gradient=output_gradient, + ).sum(dim=0, keepdim=True) + if self.module.gradient_scale != 1.0: + summed_gradient.mul_(self.module.gradient_scale) + if self.module.storage[AGGREGATED_GRADIENT_NAME] is None: + self.module.storage[AGGREGATED_GRADIENT_NAME] = torch.zeros_like(summed_gradient, requires_grad=False) + self.module.storage[AGGREGATED_GRADIENT_NAME].add_(summed_gradient) + + self.registered_hooks.append(self.module.register_forward_hook(forward_hook)) + + def finalize_iteration(self): + """Clears all cached data from memory.""" + self.clear_all_cache() + + def exist(self) -> bool: + """Checks if aggregated gradient is available.""" + return self.module.storage[AGGREGATED_GRADIENT_NAME] is not None + + def synchronize(self, num_processes: int = 1) -> None: + """Aggregates summed gradient across multiple devices or nodes in a distributed setting.""" + del num_processes + if dist.is_initialized() and torch.cuda.is_available(): + if self.module.storage[AGGREGATED_GRADIENT_NAME] is None: + self.module.storage[AGGREGATED_GRADIENT_NAME] = torch.zeros( + size=(1,), + dtype=self.module.score_args.per_sample_gradient_dtype, + device="cuda", + requires_grad=False, + ) + self.module.storage[AGGREGATED_GRADIENT_NAME] = self.module.storage[AGGREGATED_GRADIENT_NAME].contiguous() + dist.all_reduce( + tensor=self.module.storage[AGGREGATED_GRADIENT_NAME], + op=dist.ReduceOp.SUM, + ) + + def release_memory(self) -> None: + """Clears aggregated gradients from memory.""" + self.clear_all_cache() + self.module.storage[AGGREGATED_GRADIENT_NAME] = None diff --git a/kronfluence/module/tracker/pairwise_score.py b/kronfluence/module/tracker/pairwise_score.py new file mode 100644 index 0000000..30ff5f3 --- /dev/null +++ b/kronfluence/module/tracker/pairwise_score.py @@ -0,0 +1,137 @@ +from typing import Tuple + +import torch +from opt_einsum import DynamicProgramming, contract_path +from torch import _VF, nn + +from kronfluence.module.tracker.base import BaseTracker +from kronfluence.utils.constants import ( + ACCUMULATED_PRECONDITIONED_GRADIENT_NAME, + AGGREGATED_GRADIENT_NAME, + PAIRWISE_SCORE_MATRIX_NAME, + PRECONDITIONED_GRADIENT_NAME, +) + + +class PairwiseScoreTracker(BaseTracker): + """Computes pairwise influence scores for a given module.""" + + def _compute_pairwise_score_with_gradient(self, per_sample_gradient: torch.Tensor) -> None: + """Computes pairwise influence scores using per-sample-gradient. + + Args: + per_sample_gradient (torch.Tensor): + The per-sample-gradient tensor for the given batch. + """ + precondition_name = ACCUMULATED_PRECONDITIONED_GRADIENT_NAME + if isinstance(self.module.storage[precondition_name], list): + left_mat, right_mat = self.module.storage[precondition_name] + expr = "qki,toi,qok->qt" + if self.module.einsum_path is None: + path = contract_path( + expr, + right_mat, + per_sample_gradient, + left_mat, + optimize=DynamicProgramming(search_outer=True, minimize="flops"), + )[0] + self.module.einsum_path = [item for pair in path for item in pair] + scores = _VF.einsum(expr, (right_mat, per_sample_gradient, left_mat), path=self.module.einsum_path) # pylint: disable=no-member + else: + scores = torch.einsum( + "qio,tio->qt", + self.module.storage[precondition_name], + per_sample_gradient, + ) + + if self.module.storage[PAIRWISE_SCORE_MATRIX_NAME] is not None: + self.module.storage[PAIRWISE_SCORE_MATRIX_NAME].add_(scores) + else: + self.module.storage[PAIRWISE_SCORE_MATRIX_NAME] = scores + + def register_hooks(self) -> None: + """Sets up hooks to compute pairwise influence scores.""" + + @torch.no_grad() + def forward_hook(module: nn.Module, inputs: Tuple[torch.Tensor], outputs: torch.Tensor) -> None: + del module + cached_activation = inputs[0].detach() + device = "cpu" if self.module.score_args.offload_activations_to_cpu else cached_activation.device + cached_activation = cached_activation.to( + device=device, + dtype=self.module.score_args.score_dtype, + copy=True, + ) + if self.module.factor_args.has_shared_parameters: + if self.cached_activations is None: + self.cached_activations = [] + self.cached_activations.append(cached_activation) + else: + self.cached_activations = cached_activation + self.cached_hooks.append(outputs.register_hook(backward_hook)) + + @torch.no_grad() + def backward_hook(output_gradient: torch.Tensor) -> None: + if self.cached_activations is None: + self._raise_cache_not_found_exception() + handle = self.cached_hooks.pop() + handle.remove() + output_gradient = output_gradient.detach().to(dtype=self.module.score_args.score_dtype) + if isinstance(self.cached_activations, list): + cached_activation = self.cached_activations.pop() + else: + cached_activation = self.cached_activations + # Computes pairwise influence scores during backward pass. + if self.module.per_sample_gradient_process_fnc is None: + self.module.storage[PAIRWISE_SCORE_MATRIX_NAME] = self.module.compute_pairwise_score( + preconditioned_gradient=self.module.storage[ACCUMULATED_PRECONDITIONED_GRADIENT_NAME], + input_activation=cached_activation.to(device=output_gradient.device), + output_gradient=output_gradient, + ) + if self.module.gradient_scale != 1.0: + self.module.storage[PAIRWISE_SCORE_MATRIX_NAME].mul_(self.module.gradient_scale) + del cached_activation, output_gradient + self.clear_all_cache() + else: + per_sample_gradient = self.module.compute_per_sample_gradient( + input_activation=cached_activation.to(device=output_gradient.device), + output_gradient=output_gradient, + ) + del cached_activation, output_gradient + if self.module.gradient_scale != 1.0: + per_sample_gradient.mul_(self.module.gradient_scale) + self._compute_pairwise_score_with_gradient(per_sample_gradient=per_sample_gradient) + + self.registered_hooks.append(self.module.register_forward_hook(forward_hook)) + + def finalize_iteration(self) -> None: + """Clears all cached data from memory.""" + self.clear_all_cache() + + def exist(self) -> bool: + """Checks if pairwise score is available.""" + return self.module.storage[PAIRWISE_SCORE_MATRIX_NAME] is not None + + def accumulate_iterations(self) -> None: + """Removes pairwise scores from memory after a single iteration.""" + self.release_memory() + + @torch.no_grad() + def finalize_all_iterations(self) -> None: + """Removes cached preconditioned gradient from memory. Additionally, if aggregated gradients are available, + computes the pairwise score using them.""" + if self.module.storage[AGGREGATED_GRADIENT_NAME] is not None: + self.module.storage[AGGREGATED_GRADIENT_NAME] = self.module.storage[AGGREGATED_GRADIENT_NAME].to( + dtype=self.module.score_args.score_dtype + ) + self._compute_pairwise_score_with_gradient( + per_sample_gradient=self.module.storage[AGGREGATED_GRADIENT_NAME] + ) + self.module.storage[ACCUMULATED_PRECONDITIONED_GRADIENT_NAME] = None + self.module.storage[PRECONDITIONED_GRADIENT_NAME] = None + self.clear_all_cache() + + def release_memory(self) -> None: + """Releases pairwise scores from memory.""" + self.clear_all_cache() + self.module.storage[PAIRWISE_SCORE_MATRIX_NAME] = None diff --git a/kronfluence/module/tracker/precondition.py b/kronfluence/module/tracker/precondition.py new file mode 100644 index 0000000..8d677b8 --- /dev/null +++ b/kronfluence/module/tracker/precondition.py @@ -0,0 +1,261 @@ +from typing import List, Tuple + +import torch +import torch.distributed as dist +from torch import nn + +from kronfluence.factor.config import FactorConfig +from kronfluence.module.tracker.base import BaseTracker +from kronfluence.utils.constants import ( + ACCUMULATED_PRECONDITIONED_GRADIENT_NAME, + AGGREGATED_GRADIENT_NAME, + PRECONDITIONED_GRADIENT_NAME, +) + + +class PreconditionTracker(BaseTracker): + """Computes preconditioned gradient for a given module.""" + + def _compute_low_rank_preconditioned_gradient( + self, + preconditioned_gradient: torch.Tensor, + target_dtype: torch.dtype, + ) -> List[torch.Tensor]: + """Performs low-rank approximation of the preconditioned gradient. + + Args: + preconditioned_gradient (torch.Tensor): + The preconditioned per-sample gradient tensor to be low-rank approximated. + target_dtype (torch.dtype): + The desired dtype for the output. + + Returns: + List[torch.Tensor, torch.Tensor]: + Low-rank matrices approximating the original preconditioned gradient. + """ + rank = self.module.score_args.query_gradient_low_rank + if self.module.score_args.use_full_svd: + U, S, V = torch.linalg.svd( # pylint: disable=not-callable + preconditioned_gradient, + full_matrices=False, + ) + U_k = U[:, :, :rank] + S_k = S[:, :rank] + # Avoid holding the full memory of the original tensor before indexing. + V_k = V[:, :rank, :].to(dtype=target_dtype, copy=True) + left_mat = torch.matmul(U_k, torch.diag_embed(S_k)).to(dtype=target_dtype) + return [left_mat, V_k] + + U, S, V = torch.svd_lowrank(preconditioned_gradient, q=rank) + left_mat = torch.matmul(U, torch.diag_embed(S)).to(dtype=target_dtype) + V = V.transpose(1, 2).to(dtype=target_dtype) + return [left_mat, V] + + def _process_preconditioned_gradient(self, preconditioned_gradient: torch.Tensor) -> None: + """Processes the preconditioned per-sample gradient. + + Args: + preconditioned_gradient (torch.Tensor): + The preconditioned per-sample gradient tensor for the given batch. + """ + if ( + self.module.score_args.query_gradient_low_rank is not None + and min(preconditioned_gradient.size()[1:]) > self.module.score_args.query_gradient_low_rank + ): + # Apply low-rank approximation to the preconditioned gradient. + preconditioned_gradient = preconditioned_gradient.to( + dtype=self.module.score_args.query_gradient_svd_dtype + ).contiguous() + preconditioned_gradient = self._compute_low_rank_preconditioned_gradient( + preconditioned_gradient=preconditioned_gradient, + target_dtype=self.module.score_args.score_dtype, + ) + else: + preconditioned_gradient = preconditioned_gradient.to(dtype=self.module.score_args.score_dtype) + self.module.storage[PRECONDITIONED_GRADIENT_NAME] = preconditioned_gradient + + def register_hooks(self) -> None: + """Sets up hooks to compute preconditioned per-sample gradient.""" + + @torch.no_grad() + def forward_hook(module: nn.Module, inputs: Tuple[torch.Tensor], outputs: torch.Tensor) -> None: + del module + cached_activation = inputs[0].detach() + device = "cpu" if self.module.score_args.offload_activations_to_cpu else cached_activation.device + cached_activation = cached_activation.to( + device=device, + dtype=self.module.score_args.per_sample_gradient_dtype, + copy=True, + ) + if self.module.factor_args.has_shared_parameters: + if self.cached_activations is None: + self.cached_activations = [] + self.cached_activations.append(cached_activation) + else: + self.cached_activations = cached_activation + self.cached_hooks.append( + outputs.register_hook( + shared_backward_hook if self.module.factor_args.has_shared_parameters else backward_hook + ) + ) + + @torch.no_grad() + def backward_hook(output_gradient: torch.Tensor) -> None: + if self.cached_activations is None: + self._raise_cache_not_found_exception() + handle = self.cached_hooks.pop() + handle.remove() + output_gradient = output_gradient.detach().to(dtype=self.module.score_args.per_sample_gradient_dtype) + per_sample_gradient = self.module.compute_per_sample_gradient( + input_activation=self.cached_activations.to(device=output_gradient.device), + output_gradient=output_gradient, + ).to(dtype=self.module.score_args.precondition_dtype) + self.clear_all_cache() + del output_gradient + # Computes preconditioned per-sample gradient during backward pass. + preconditioned_gradient = FactorConfig.CONFIGS[self.module.factor_args.strategy].precondition_gradient( + gradient=per_sample_gradient, + storage=self.module.storage, + ) + if self.module.gradient_scale != 1.0: + preconditioned_gradient.mul_(self.module.gradient_scale) + del per_sample_gradient + self._process_preconditioned_gradient(preconditioned_gradient=preconditioned_gradient) + + @torch.no_grad() + def shared_backward_hook(output_gradient: torch.Tensor) -> None: + handle = self.cached_hooks.pop() + handle.remove() + output_gradient = output_gradient.detach().to(dtype=self.module.score_args.per_sample_gradient_dtype) + cached_activation = self.cached_activations.pop() + per_sample_gradient = self.module.compute_per_sample_gradient( + input_activation=cached_activation.to(device=output_gradient.device), + output_gradient=output_gradient, + ) + if self.cached_per_sample_gradient is None: + self.cached_per_sample_gradient = torch.zeros_like(per_sample_gradient, requires_grad=False) + # Aggregates per-sample gradients during backward pass. + self.cached_per_sample_gradient.add_(per_sample_gradient) + + self.registered_hooks.append(self.module.register_forward_hook(forward_hook)) + + @torch.no_grad() + def finalize_iteration(self) -> None: + """Computes preconditioned gradient using cached per-sample gradients.""" + if self.module.factor_args.has_shared_parameters: + self.cached_per_sample_gradient = self.cached_per_sample_gradient.to( + dtype=self.module.score_args.precondition_dtype + ) + preconditioned_gradient = FactorConfig.CONFIGS[self.module.factor_args.strategy].precondition_gradient( + gradient=self.cached_per_sample_gradient, + storage=self.module.storage, + ) + self.cached_per_sample_gradient = None + if self.module.gradient_scale != 1.0: + preconditioned_gradient.mul_(self.module.gradient_scale) + self._process_preconditioned_gradient(preconditioned_gradient=preconditioned_gradient) + self.clear_all_cache() + + def exist(self) -> bool: + """Checks if preconditioned gradient is available.""" + return ( + self.module.storage[PRECONDITIONED_GRADIENT_NAME] is not None + or self.module.storage[ACCUMULATED_PRECONDITIONED_GRADIENT_NAME] is not None + ) + + def synchronize(self, num_processes: int = 1) -> None: + """Stacks preconditioned gradient across multiple devices or nodes in a distributed setting.""" + if ( + dist.is_initialized() + and torch.cuda.is_available() + and self.module.storage[PRECONDITIONED_GRADIENT_NAME] is not None + ): + if isinstance(self.module.storage[PRECONDITIONED_GRADIENT_NAME], list): + for i in range(len(self.module.storage[PRECONDITIONED_GRADIENT_NAME])): + size = self.module.storage[PRECONDITIONED_GRADIENT_NAME][i].size() + stacked_matrix = torch.empty( + size=(num_processes, size[0], size[1], size[2]), + dtype=self.module.storage[PRECONDITIONED_GRADIENT_NAME][i].dtype, + device=self.module.storage[PRECONDITIONED_GRADIENT_NAME][i].device, + ) + torch.distributed.all_gather_into_tensor( + output_tensor=stacked_matrix, + input_tensor=self.module.storage[PRECONDITIONED_GRADIENT_NAME][i].contiguous(), + ) + self.module.storage[PRECONDITIONED_GRADIENT_NAME][i] = stacked_matrix.transpose(0, 1).reshape( + num_processes * size[0], size[1], size[2] + ) + else: + size = self.module.storage[PRECONDITIONED_GRADIENT_NAME].size() + stacked_preconditioned_gradient = torch.empty( + size=(num_processes, size[0], size[1], size[2]), + dtype=self.module.storage[PRECONDITIONED_GRADIENT_NAME].dtype, + device=self.module.storage[PRECONDITIONED_GRADIENT_NAME].device, + ) + torch.distributed.all_gather_into_tensor( + output_tensor=stacked_preconditioned_gradient, + input_tensor=self.module.storage[PRECONDITIONED_GRADIENT_NAME].contiguous(), + ) + self.module.storage[PRECONDITIONED_GRADIENT_NAME] = stacked_preconditioned_gradient.transpose( + 0, 1 + ).reshape(num_processes * size[0], size[1], size[2]) + + def truncate(self, keep_size: int) -> None: + """Truncates preconditioned gradient to appropriate dimension.""" + if isinstance(self.module.storage[PRECONDITIONED_GRADIENT_NAME], list): + assert len(self.module.storage[PRECONDITIONED_GRADIENT_NAME]) == 2 + self.module.storage[PRECONDITIONED_GRADIENT_NAME] = [ + self.module.storage[PRECONDITIONED_GRADIENT_NAME][0][:keep_size].clone(), + self.module.storage[PRECONDITIONED_GRADIENT_NAME][1][:keep_size].clone(), + ] + else: + self.module.storage[PRECONDITIONED_GRADIENT_NAME] = self.module.storage[PRECONDITIONED_GRADIENT_NAME][ + :keep_size + ].clone() + + def accumulate_iterations(self) -> None: + """Accumulates preconditioned gradient across multiple iterations.""" + accumulated_gradient = self.module.storage[ACCUMULATED_PRECONDITIONED_GRADIENT_NAME] + gradient = self.module.storage[PRECONDITIONED_GRADIENT_NAME] + + if self.module.storage[ACCUMULATED_PRECONDITIONED_GRADIENT_NAME] is None: + if isinstance(self.module.storage[PRECONDITIONED_GRADIENT_NAME], list): + self.module.storage[ACCUMULATED_PRECONDITIONED_GRADIENT_NAME] = [ + tensor.contiguous() for tensor in gradient + ] + else: + self.module.storage[ACCUMULATED_PRECONDITIONED_GRADIENT_NAME] = gradient.contiguous() + + else: + if isinstance(self.module.storage[PRECONDITIONED_GRADIENT_NAME], list): + self.module.storage[ACCUMULATED_PRECONDITIONED_GRADIENT_NAME] = [ + torch.cat((accumulated_gradient[0], gradient[0]), dim=0).contiguous(), + torch.cat((accumulated_gradient[1], gradient[1]), dim=0).contiguous(), + ] + else: + self.module.storage[ACCUMULATED_PRECONDITIONED_GRADIENT_NAME] = torch.cat( + (accumulated_gradient, gradient), dim=0 + ).contiguous() + del self.module.storage[PRECONDITIONED_GRADIENT_NAME], gradient + self.module.storage[PRECONDITIONED_GRADIENT_NAME] = None + + @torch.no_grad() + def finalize_all_iterations(self) -> None: + """Preconditions aggregated gradient if it exists in storage.""" + if self.module.storage[AGGREGATED_GRADIENT_NAME] is not None: + self.module.storage[AGGREGATED_GRADIENT_NAME] = self.module.storage[AGGREGATED_GRADIENT_NAME].to( + dtype=self.module.score_args.precondition_dtype + ) + preconditioned_gradient = FactorConfig.CONFIGS[self.module.factor_args.strategy].precondition_gradient( + gradient=self.module.storage[AGGREGATED_GRADIENT_NAME], + storage=self.module.storage, + ) + self.module.storage[AGGREGATED_GRADIENT_NAME] = None + self._process_preconditioned_gradient(preconditioned_gradient=preconditioned_gradient) + self.accumulate_iterations() + + def release_memory(self) -> None: + """Clears preconditioned gradients from memory.""" + self.module.storage[ACCUMULATED_PRECONDITIONED_GRADIENT_NAME] = None + self.module.storage[PRECONDITIONED_GRADIENT_NAME] = None + self.clear_all_cache() diff --git a/kronfluence/module/tracker/self_score.py b/kronfluence/module/tracker/self_score.py new file mode 100644 index 0000000..679ecfa --- /dev/null +++ b/kronfluence/module/tracker/self_score.py @@ -0,0 +1,255 @@ +from typing import Tuple + +import torch +from torch import nn + +from kronfluence.factor.config import STORAGE_TYPE, FactorConfig +from kronfluence.module.tracker.base import BaseTracker +from kronfluence.utils.constants import ( + PRECONDITIONED_GRADIENT_NAME, + SELF_SCORE_VECTOR_NAME, +) + + +def move_storage_to_device(storage: STORAGE_TYPE, target_device: torch.device) -> None: + """Moves all stored factors in the storage dictionary to the specified target device. + + Args: + storage (STORAGE_TYPE): + A dictionary containing stored factors. + target_device (torch.device): + The target device to move the factors to. + """ + for name, factor in storage.items(): + if factor is not None: + if isinstance(factor, list): + for i in range(len(storage[name])): + storage[name][i] = factor[i].to(device=target_device) + if isinstance(factor, torch.Tensor): + storage[name] = factor.to(device=target_device) + + +class SelfScoreTracker(BaseTracker): + """Computes self-influence scores for a given module.""" + + storage_at_device: bool = False + + def _compute_self_score(self, per_sample_gradient: torch.Tensor) -> None: + """Computes self-influence scores using per-sample gradients. + + Args: + per_sample_gradient (torch.Tensor): + The per-sample gradient tensor for the given batch. + """ + if not self.storage_at_device: + move_storage_to_device( + storage=self.module.storage, + target_device=per_sample_gradient.device, + ) + self.storage_at_device = True + + preconditioned_gradient = ( + FactorConfig.CONFIGS[self.module.factor_args.strategy] + .precondition_gradient( + gradient=per_sample_gradient, + storage=self.module.storage, + ) + .to(dtype=self.module.score_args.score_dtype) + ) + per_sample_gradient = per_sample_gradient.to(dtype=self.module.score_args.score_dtype) + preconditioned_gradient.mul_(per_sample_gradient) + self.module.storage[SELF_SCORE_VECTOR_NAME] = preconditioned_gradient.sum(dim=(1, 2)) + + def register_hooks(self) -> None: + """Sets up hooks to compute self-influence scores.""" + + @torch.no_grad() + def forward_hook(module: nn.Module, inputs: Tuple[torch.Tensor], outputs: torch.Tensor) -> None: + del module + cached_activation = inputs[0].detach() + device = "cpu" if self.module.score_args.offload_activations_to_cpu else cached_activation.device + cached_activation = cached_activation.to( + device=device, + dtype=self.module.score_args.per_sample_gradient_dtype, + copy=True, + ) + if self.module.factor_args.has_shared_parameters: + if self.cached_activations is None: + self.cached_activations = [] + self.cached_activations.append(cached_activation) + else: + self.cached_activations = cached_activation + self.cached_hooks.append( + outputs.register_hook( + shared_backward_hook if self.module.factor_args.has_shared_parameters else backward_hook + ) + ) + + @torch.no_grad() + def backward_hook(output_gradient: torch.Tensor) -> None: + if self.cached_activations is None: + self._raise_cache_not_found_exception() + handle = self.cached_hooks.pop() + handle.remove() + output_gradient = output_gradient.detach().to(dtype=self.module.score_args.per_sample_gradient_dtype) + per_sample_gradient = self.module.compute_per_sample_gradient( + input_activation=self.cached_activations.to(device=output_gradient.device), + output_gradient=output_gradient, + ).to(dtype=self.module.score_args.precondition_dtype) + self.clear_all_cache() + del output_gradient + if self.module.gradient_scale != 1.0: + per_sample_gradient.mul_(self.module.gradient_scale) + self._compute_self_score(per_sample_gradient=per_sample_gradient) + + @torch.no_grad() + def shared_backward_hook(output_gradient: torch.Tensor) -> None: + handle = self.cached_hooks.pop() + handle.remove() + output_gradient = output_gradient.detach().to(dtype=self.module.score_args.per_sample_gradient_dtype) + cached_activation = self.cached_activations.pop() + per_sample_gradient = self.module.compute_per_sample_gradient( + input_activation=cached_activation.to(device=output_gradient.device), + output_gradient=output_gradient, + ) + if self.cached_per_sample_gradient is None: + self.cached_per_sample_gradient = torch.zeros_like(per_sample_gradient, requires_grad=False) + self.cached_per_sample_gradient.add_(per_sample_gradient) + + self.registered_hooks.append(self.module.register_forward_hook(forward_hook)) + + @torch.no_grad() + def finalize_iteration(self) -> None: + """Computes self-influence scores using cached per-sample gradients.""" + if self.module.factor_args.has_shared_parameters: + self.cached_per_sample_gradient = self.cached_per_sample_gradient.to( + dtype=self.module.score_args.precondition_dtype + ) + if self.module.gradient_scale != 1.0: + self.cached_per_sample_gradient.mul_(self.module.gradient_scale) + self._compute_self_score(per_sample_gradient=self.cached_per_sample_gradient) + self.clear_all_cache() + + def exist(self) -> bool: + """Checks if self-influence score is available.""" + return self.module.storage[SELF_SCORE_VECTOR_NAME] is not None + + def accumulate_iterations(self) -> None: + """Removes self-influence scores from memory after a single iteration.""" + self.release_memory() + + def release_memory(self) -> None: + """Releases self-influence scores from memory.""" + self.clear_all_cache() + if self.storage_at_device: + move_storage_to_device(storage=self.module.storage, target_device=torch.device("cpu")) + self.storage_at_device = False + del self.module.storage[SELF_SCORE_VECTOR_NAME] + self.module.storage[SELF_SCORE_VECTOR_NAME] = None + + +class SelfScoreWithMeasurementTracker(BaseTracker): + """Computes self-influence scores with measurement for a given module.""" + + storage_at_device: bool = False + + def _compute_self_measurement_score_with_gradient(self, per_sample_gradient: torch.Tensor) -> None: + """Computes self-influence scores with measurement using per-sample-gradients. + + Args: + per_sample_gradient (torch.Tensor): + The per-sample-gradient tensor for the given batch. + """ + scores = per_sample_gradient.mul_(self.module.storage[PRECONDITIONED_GRADIENT_NAME]).sum(dim=(1, 2)) + self.module.storage[PRECONDITIONED_GRADIENT_NAME] = None + if self.module.storage[SELF_SCORE_VECTOR_NAME] is None: + self.module.storage[SELF_SCORE_VECTOR_NAME] = scores + else: + self.module.storage[SELF_SCORE_VECTOR_NAME].add_(scores) + + def register_hooks(self) -> None: + """Sets up hooks to compute self-influence scores with measurement.""" + + @torch.no_grad() + def forward_hook(module: nn.Module, inputs: Tuple[torch.Tensor], outputs: torch.Tensor) -> None: + del module + cached_activation = inputs[0].detach() + device = "cpu" if self.module.score_args.offload_activations_to_cpu else cached_activation.device + cached_activation = cached_activation.to( + device=device, + dtype=self.module.score_args.score_dtype, + copy=True, + ) + if self.module.factor_args.has_shared_parameters: + if self.cached_activations is None: + self.cached_activations = [] + self.cached_activations.append(cached_activation) + else: + self.cached_activations = cached_activation + self.cached_hooks.append(outputs.register_hook(backward_hook)) + + @torch.no_grad() + def backward_hook(output_gradient: torch.Tensor) -> None: + if self.cached_activations is None: + self._raise_cache_not_found_exception() + + if not self.storage_at_device: + move_storage_to_device( + storage=self.module.storage, + target_device=output_gradient.device, + ) + self.storage_at_device = True + + handle = self.cached_hooks.pop() + handle.remove() + output_gradient = output_gradient.detach().to(dtype=self.module.score_args.score_dtype) + if isinstance(self.cached_activations, list): + cached_activation = self.cached_activations.pop() + else: + cached_activation = self.cached_activations + if self.module.per_sample_gradient_process_fnc is None: + scores = self.module.compute_self_measurement_score( + preconditioned_gradient=self.module.storage[PRECONDITIONED_GRADIENT_NAME], + input_activation=cached_activation.to(device=output_gradient.device), + output_gradient=output_gradient, + ) + self.module.storage[PRECONDITIONED_GRADIENT_NAME] = None + self.clear_all_cache() + if self.module.gradient_scale != 1.0: + scores.mul_(self.module.gradient_scale) + if self.module.storage[SELF_SCORE_VECTOR_NAME] is None: + self.module.storage[SELF_SCORE_VECTOR_NAME] = scores + else: + self.module.storage[SELF_SCORE_VECTOR_NAME].add_(scores) + else: + per_sample_gradient = self.module.compute_per_sample_gradient( + input_activation=cached_activation.to(device=output_gradient.device), + output_gradient=output_gradient, + ) + del cached_activation, output_gradient + if self.module.gradient_scale != 1.0: + per_sample_gradient.mul_(self.module.gradient_scale) + self._compute_self_measurement_score_with_gradient(per_sample_gradient=per_sample_gradient) + + self.registered_hooks.append(self.module.register_forward_hook(forward_hook)) + + def finalize_iteration(self) -> None: + """Clears all cached data from memory.""" + self.clear_all_cache() + + def exist(self) -> bool: + """Checks if self-influence score is available.""" + return self.module.storage[SELF_SCORE_VECTOR_NAME] is not None + + def accumulate_iterations(self) -> None: + """Removes self-influence scores from memory after a single iteration.""" + self.release_memory() + + def release_memory(self) -> None: + """Releases self-influence scores from memory.""" + self.clear_all_cache() + if self.storage_at_device: + move_storage_to_device(storage=self.module.storage, target_device=torch.device("cpu")) + self.storage_at_device = False + del self.module.storage[SELF_SCORE_VECTOR_NAME] + self.module.storage[SELF_SCORE_VECTOR_NAME] = None diff --git a/kronfluence/module/utils.py b/kronfluence/module/utils.py index 5266299..e858121 100644 --- a/kronfluence/module/utils.py +++ b/kronfluence/module/utils.py @@ -9,15 +9,22 @@ from kronfluence.arguments import FactorArguments, ScoreArguments from kronfluence.module.tracked_module import ModuleMode, TrackedModule from kronfluence.task import Task -from kronfluence.utils.exceptions import ( - IllegalTaskConfigurationError, - TrackedModuleNotFoundError, -) +from kronfluence.utils.exceptions import IllegalTaskConfigurationError def _get_submodules(model: nn.Module, key: str) -> Tuple[nn.Module, str]: - """Returns the parent module and its name given the name of the current module.""" - # The code is modified from: https://github.com/huggingface/peft/blob/main/src/peft/utils/other.py. + """Retrieves the parent module and its name given the name of the current module. + + Args: + model (nn.Module): + The PyTorch model to inspect. + key (str): + The full name of the current module. + + Returns: + Tuple[nn.Module, str]: + The parent module and the name of the target module. + """ parent = model.get_submodule(".".join(key.split(".")[:-1])) target_name = key.split(".")[-1] return parent, target_name @@ -37,46 +44,48 @@ def wrap_tracked_modules( task (Task): The specific task associated with the model. factor_args (FactorArguments, optional): - Arguments related to computing the preconditioning factors. + Arguments related to computing influence factors. score_args (ScoreArguments, optional): - Arguments related to computing the influence scores. + Arguments related to computing influence scores. Returns: nn.Module: - The wrapped model with `TrackedModule`. + The processed model with `TrackedModule` installed. """ if isinstance(model, (DP, DDP, FSDP)): raise ValueError( "The model is wrapped with DataParallel, DistributedDataParallel " - "or FullyShardedDataParallel. Call `wrap_tracked_modules` before wrapping the model." + "or FullyShardedDataParallel. Call `prepare_model` before wrapping the model." ) - tracked_module_count = 0 - tracked_module_names = task.tracked_modules() if task is not None else None + tracked_module_names = task.get_influence_tracked_modules() if task is not None else None tracked_module_exists_dict = None if tracked_module_names is not None: tracked_module_exists_dict = {name: False for name in tracked_module_names} + per_sample_gradient_process_fnc = None + if task is not None and task.enable_post_process_per_sample_gradient: + per_sample_gradient_process_fnc = task.post_process_per_sample_gradient named_modules = model.named_modules() for module_name, module in named_modules: if len(list(module.children())) > 0: continue - # Filter modules based on the task's `influence_modules` if specified. + # Filters modules based on the task's `get_influence_tracked_modules` if specified. if tracked_module_names is not None and module_name not in tracked_module_names: continue - # Wrap the module if it is currently supported (e.g., nn.Linear & nn.Conv2d). + # Wraps the module if it is currently supported (e.g., nn.Linear & nn.Conv2d). if isinstance(module, tuple(TrackedModule.SUPPORTED_MODULES)): tracked_module = TrackedModule.SUPPORTED_MODULES[type(module)]( name=module_name, original_module=module, + per_sample_gradient_process_fnc=per_sample_gradient_process_fnc, factor_args=factor_args, score_args=score_args, ) parent, target_name = _get_submodules(model=model, key=module_name) setattr(parent, target_name, tracked_module) - tracked_module_count += 1 if tracked_module_exists_dict is not None: tracked_module_exists_dict[module_name] = True @@ -87,22 +96,34 @@ def wrap_tracked_modules( ) raise IllegalTaskConfigurationError(error_msg) - if tracked_module_count == 0: - supported_modules_names = [module.__name__ for module in TrackedModule.SUPPORTED_MODULES] - error_msg = ( - f"Kronfluence currently supports modules in `{supported_modules_names}`. " - f"However, these modules were not found in the provided model. If you want to analyze " - "custom layers, consider rewriting your model to use the supported modules, " - "or define your own custom module by subclassing `TrackedModule`." + if not any(isinstance(module, TrackedModule) for module in model.modules()): + supported_modules = ", ".join(module.__name__ for module in TrackedModule.SUPPORTED_MODULES) + raise IllegalTaskConfigurationError( + f"No supported modules found. Kronfluence supports: {supported_modules}. " + "Consider rewriting your model or subclassing `TrackedModule` for custom layers.\n" + f"Current Model:\n{model}" ) - error_msg += f"\n{model}" - raise IllegalTaskConfigurationError(error_msg) - return model def make_modules_partition(total_module_names: List[str], partition_size: int) -> List[List[str]]: - """Divides a list of module names into smaller partitions of a specified size.""" + """Divides a list of module names into smaller partitions of a specified size. + + Args: + total_module_names (List[str]): + The list of all module names. + partition_size (int): + The number of partitions to create. + + Returns: + List[List[str]]: + A list of partitioned module names. + + Raises: + ValueError: If `len(total_module_names)` is less than `partition_size`. + """ + if len(total_module_names) < partition_size: + raise ValueError("The total modules must be equal to or greater than the partition size.") # See https://stackoverflow.com/questions/2130016/splitting-a-list-into-n-parts-of-approximately-equal-length. div, mod = divmod(len(total_module_names), partition_size) return list( @@ -110,177 +131,146 @@ def make_modules_partition(total_module_names: List[str], partition_size: int) - ) -def update_factor_args(model: nn.Module, factor_args: FactorArguments) -> None: - """Updates the factor arguments for all `TrackedModule` instances within a model.""" - tracked_module_count = 0 - for module in model.modules(): - if isinstance(module, TrackedModule): - module.update_factor_args(factor_args=factor_args) - tracked_module_count += 1 - if tracked_module_count == 0: - raise TrackedModuleNotFoundError( - f"Tracked modules not found when trying to update factor arguments `{factor_args}`." - ) - - -def update_score_args(model: nn.Module, score_args: ScoreArguments) -> None: - """Updates the score arguments for all `TrackedModule` instances within a model.""" - tracked_module_count = 0 - for module in model.modules(): - if isinstance(module, TrackedModule): - module.update_score_args(score_args=score_args) - tracked_module_count += 1 - if tracked_module_count == 0: - raise TrackedModuleNotFoundError( - f"Tracked modules not found when trying to update score arguments `{score_args}`." - ) - +def set_mode( + model: nn.Module, + mode: ModuleMode, + tracked_module_names: List[str] = None, + release_memory: bool = False, +) -> None: + """Sets the module mode of specified `TrackedModule` instances within a model. -def get_tracked_module_names(model: nn.Module) -> List[str]: - """Returns the names of `TrackedModule` instances within a model.""" - tracked_modules = [] + Args: + model (nn.Module): + The PyTorch model containing `TrackedModule` instances. + mode (ModuleMode): + The new mode to set for `TrackedModule`. + tracked_module_names (List[str], optional): + Names of modules to update. If `None`, updates all. + release_memory (bool, optional): + If `True`, releases memory of existing factors. + """ for module in model.modules(): if isinstance(module, TrackedModule): - tracked_modules.append(module.name) - return tracked_modules - + if tracked_module_names is not None and module.name not in tracked_module_names: + continue + module.set_mode(mode=mode, release_memory=release_memory) -def synchronize_covariance_matrices(model: nn.Module) -> None: - """Synchronizes covariance matrices of all `TrackedModule` instances within a model.""" - tracked_module_count = 0 - for module in model.modules(): - if isinstance(module, TrackedModule): - module.synchronize_covariance_matrices() - tracked_module_count += 1 - if tracked_module_count == 0: - raise TrackedModuleNotFoundError("Tracked modules not found when trying to synchronize covariance matrices.") +def update_factor_args(model: nn.Module, factor_args: FactorArguments) -> None: + """Updates the factor arguments for all `TrackedModule` instances within a model. -def synchronize_lambda_matrices(model: nn.Module) -> None: - """Synchronizes Lambda matrices of all `TrackedModule` instances within a model.""" - tracked_module_count = 0 + Args: + model (nn.Module): + The PyTorch model containing `TrackedModule` instances. + factor_args (FactorArguments): + The new factor arguments to set. + """ for module in model.modules(): if isinstance(module, TrackedModule): - module.synchronize_lambda_matrices() - tracked_module_count += 1 - if tracked_module_count == 0: - raise TrackedModuleNotFoundError("Tracked modules not found when trying to synchronize lambda matrices.") - + module.update_factor_args(factor_args=factor_args) -def aggregate_preconditioned_gradient(model: nn.Module) -> None: - """Aggregates preconditioned gradient of all `TrackedModule` instances within a model.""" - tracked_module_count = 0 - for module in model.modules(): - if isinstance(module, TrackedModule): - module.aggregate_preconditioned_gradient() - tracked_module_count += 1 - if tracked_module_count == 0: - raise TrackedModuleNotFoundError("Tracked modules not found when trying to truncate preconditioned gradient.") +def update_score_args(model: nn.Module, score_args: ScoreArguments) -> None: + """Updates the score arguments for all `TrackedModule` instances within a model. -def truncate_preconditioned_gradient(model: nn.Module, keep_size: int) -> None: - """Truncates preconditioned gradient of all `TrackedModule` instances within a model.""" - tracked_module_count = 0 + Args: + model (nn.Module): + The PyTorch model containing `TrackedModule` instances. + score_args (ScoreArguments): + The new score arguments to set. + """ for module in model.modules(): if isinstance(module, TrackedModule): - module.truncate_preconditioned_gradient(keep_size=keep_size) - tracked_module_count += 1 - if tracked_module_count == 0: - raise TrackedModuleNotFoundError("Tracked modules not found when trying to truncate preconditioned gradient.") + module.update_score_args(score_args=score_args) -def synchronize_preconditioned_gradient(model: nn.Module, num_processes: int) -> None: - """Synchronizes preconditioned gradient of all `TrackedModule` instances within a model.""" - tracked_module_count = 0 - for module in model.modules(): - if isinstance(module, TrackedModule): - module.synchronize_preconditioned_gradient(num_processes=num_processes) - tracked_module_count += 1 - if tracked_module_count == 0: - raise TrackedModuleNotFoundError( - "Tracked modules not found when trying to synchronize preconditioned gradient." - ) +def get_tracked_module_names(model: nn.Module) -> List[str]: + """Returns the names of `TrackedModule` instances within a model. + Args: + model (nn.Module): + The PyTorch model to inspect. -def release_scores(model: nn.Module) -> None: - """Releases scores of all `TrackedModule` instances within a model.""" - tracked_module_count = 0 - for module in model.modules(): - if isinstance(module, TrackedModule): - module.release_scores() - tracked_module_count += 1 - if tracked_module_count == 0: - raise TrackedModuleNotFoundError("Tracked modules not found when trying to release scores.") + Returns: + List[str]: + A list of names of `TrackedModule` instances. + """ + return [module.name for module in model.modules() if isinstance(module, TrackedModule)] -def set_mode( +def load_factors( model: nn.Module, - mode: ModuleMode, + factor_name: str, tracked_module_names: List[str] = None, - keep_factors: bool = False, -) -> None: - """Sets the module mode of all `TrackedModule` instances within a model. For example, to compute - and aggregate the covariance matrices, the module mode needs to be set to `ModuleMode.COVARIANCE`. + cpu: bool = True, +) -> Dict[str, torch.Tensor]: + """Loads factors with the given name from specified `TrackedModule` instances. Args: model (nn.Module): - The PyTorch model which contains `TrackedModule`. - mode (ModuleMode): - The new mode to set for `TrackedModule`. - tracked_module_names (List[str], optional): - The list of names for `TrackedModule` to set the new mode. If not provided, the new mode is - set for all available `TrackedModule`. - keep_factors (bool, optional): - If True, existing factors are kept in memory. Defaults to False. + The PyTorch model containing `TrackedModule` instances. + factor_name (str): + The name of the factor to load. + tracked_module_names (Optional[List[str]]): + Names of modules to load from. If `None`, loads from all. + cpu (bool): + If `True`, moves factors to CPU and releases GPU memory. + + Returns: + Dict[str, torch.Tensor]: + A dictionary of loaded factors, keyed by module name. """ - tracked_module_count = 0 + loaded_factors = {} for module in model.modules(): if isinstance(module, TrackedModule): if tracked_module_names is not None and module.name not in tracked_module_names: continue - module.set_mode(mode=mode, keep_factors=keep_factors) - tracked_module_count += 1 - if tracked_module_count == 0: - raise TrackedModuleNotFoundError(f"Tracked modules not found when trying to set mode `{mode}`.") - - -def load_factors( - model: nn.Module, - factor_name: str, -) -> Dict[str, torch.Tensor]: - """Loads factors with the given name from all `TrackedModule` instances within a model.""" - loaded_factors = {} - for module in model.modules(): - if isinstance(module, TrackedModule): factor = module.get_factor(factor_name=factor_name) if factor is not None: - loaded_factors[module.name] = factor - if len(loaded_factors) == 0: - raise TrackedModuleNotFoundError( - f"Tracked modules not found when trying to load factors with name `{factor_name}`." - ) + if cpu: + loaded_factors[module.name] = factor.to(device="cpu", copy=True) + module.release_factor(factor_name=factor_name) + else: + loaded_factors[module.name] = factor return loaded_factors -def set_factors(model: nn.Module, factor_name: str, factors: Dict[str, torch.Tensor]) -> None: - """Sets new factor for all `TrackedModule` instances within a model.""" - tracked_module_count = 0 +def set_factors(model: nn.Module, factor_name: str, factors: Dict[str, torch.Tensor], clone: bool = False) -> None: + """Sets new factors for all `TrackedModule` instances within a model. + + Args: + model (nn.Module): + The PyTorch model containing `TrackedModule` instances. + factor_name (str): + The name of the factor to set. + factors (Dict[str, torch.Tensor]): + A dictionary of factors to set, keyed by module name. + clone (bool): + If `True`, clones the factors before setting. + """ for module in model.modules(): if isinstance(module, TrackedModule): - module.set_factor(factor_name=factor_name, factor=factors[module.name]) - tracked_module_count += 1 - if tracked_module_count == 0: - raise TrackedModuleNotFoundError( - f"Tracked modules not found when trying to set factor with name `{factor_name}`." - ) + module.set_factor( + factor_name=factor_name, factor=factors[module.name].clone() if clone else factors[module.name] + ) def set_attention_mask( model: nn.Module, attention_mask: Optional[Union[Dict[str, torch.Tensor], torch.Tensor]] = None, ) -> None: - """Sets the attention mask for all `TrackedModule` instances within a model.""" - tracked_module_count = 0 + """Sets the attention mask for all `TrackedModule` instances within a model. + + Args: + model (nn.Module): + The PyTorch model containing `TrackedModule` instances. + attention_mask (Optional[Union[Dict[str, torch.Tensor], torch.Tensor]]): + The attention mask to set. Can be a dictionary (keyed by module name) or a single tensor. + + Raises: + RuntimeError: + If an invalid attention mask is provided. + """ for module in model.modules(): if isinstance(module, TrackedModule): if isinstance(attention_mask, dict): @@ -290,44 +280,134 @@ def set_attention_mask( module.set_attention_mask(attention_mask=None) elif isinstance(attention_mask, torch.Tensor): module.set_attention_mask(attention_mask=attention_mask) + elif attention_mask is None: + module.set_attention_mask(attention_mask=None) else: raise RuntimeError(f"Invalid attention mask `{attention_mask}` provided.") - tracked_module_count += 1 - if tracked_module_count == 0: - raise TrackedModuleNotFoundError("Tracked modules not found when trying to set `attention_mask`.") - - -def remove_attention_mask(model: nn.Module) -> None: - """Removes the attention mask for all `TrackedModule` instances within a model.""" - tracked_module_count = 0 - for module in model.modules(): - if isinstance(module, TrackedModule): - module.remove_attention_mask() - tracked_module_count += 1 - if tracked_module_count == 0: - raise TrackedModuleNotFoundError("Tracked modules not found when trying to remove `attention_mask`.") def set_gradient_scale( model: nn.Module, gradient_scale: float = 1.0, ) -> None: - """Sets the gradient scale for all `TrackedModule` instances within a model.""" - tracked_module_count = 0 + """Sets the gradient scale for all `TrackedModule` instances within a model. + + Args: + model (nn.Module): + The PyTorch model containing `TrackedModule` instances. + gradient_scale (float): + The gradient scale to set. + """ for module in model.modules(): if isinstance(module, TrackedModule): module.set_gradient_scale(scale=gradient_scale) - tracked_module_count += 1 - if tracked_module_count == 0: - raise TrackedModuleNotFoundError("Tracked modules not found when trying to set `gradient_scale`.") -def remove_gradient_scale(model: nn.Module) -> None: - """Resets the gradient scale for all `TrackedModule` instances within a model.""" - tracked_module_count = 0 +def prepare_modules(model: nn.Module, tracked_module_names: List[str], device: torch.device) -> None: + """Prepares specified `TrackedModule` instances for score computation. + + Args: + model (nn.Module): + The PyTorch model containing `TrackedModule` instances. + tracked_module_names (List[str]): + Names of modules to prepare. + device (torch.device): + The device to prepare the modules for. + """ for module in model.modules(): - if isinstance(module, TrackedModule): - module.remove_gradient_scale() - tracked_module_count += 1 - if tracked_module_count == 0: - raise TrackedModuleNotFoundError("Tracked modules not found when trying to remove `gradient_scale`.") + if isinstance(module, TrackedModule) and module.name in tracked_module_names: + module.prepare_storage(device=device) + + +def synchronize_modules(model: nn.Module, tracked_module_names: List[str], num_processes: int = 1) -> None: + """Synchronizes specified `TrackedModule` instances across processes. + + Args: + model (nn.Module): + The PyTorch model containing `TrackedModule` instances. + tracked_module_names (List[str]): + Names of modules to synchronize. + num_processes (int): + The number of processes to synchronize across. + """ + for module in model.modules(): + if isinstance(module, TrackedModule) and module.name in tracked_module_names: + module.synchronize(num_processes=num_processes) + + +def truncate(model: nn.Module, tracked_module_names: List[str], keep_size: int) -> None: + """Truncates the data in specified `TrackedModule` instances. + + Args: + model (nn.Module): + The PyTorch model containing `TrackedModule` instances. + tracked_module_names (List[str]): + Names of modules to truncate. + keep_size (int): + The number of elements to keep after truncation. + """ + for module in model.modules(): + if isinstance(module, TrackedModule) and module.name in tracked_module_names: + module.truncate(keep_size=keep_size) + + +def exist_for_all_modules(model: nn.Module, tracked_module_names: List[str]) -> bool: + """Checks if all specified `TrackedModule` instances have existing factor. + + Args: + model (nn.Module): + The PyTorch model containing `TrackedModule` instances. + tracked_module_names (List[str]): + Names of modules to check. + + Returns: + bool: + `True` if all specified modules have existing factor, `False` otherwise. + """ + return all( + module.exist() + for module in model.modules() + if isinstance(module, TrackedModule) and module.name in tracked_module_names + ) + + +def accumulate_iterations(model: nn.Module, tracked_module_names: List[str]) -> None: + """Accumulates iterations for specified `TrackedModule` instances. + + Args: + model (nn.Module): + The PyTorch model containing `TrackedModule` instances. + tracked_module_names (List[str]): + Names of modules to accumulate iterations for. + """ + for module in model.modules(): + if isinstance(module, TrackedModule) and module.name in tracked_module_names: + module.accumulate_iterations() + + +def finalize_iteration(model: nn.Module, tracked_module_names: List[str]) -> None: + """Finalizes the current iteration for specified `TrackedModule` instances. + + Args: + model (nn.Module): + The PyTorch model containing `TrackedModule` instances. + tracked_module_names (List[str]): + Names of modules to finalize iteration for. + """ + for module in model.modules(): + if isinstance(module, TrackedModule) and module.name in tracked_module_names: + module.finalize_iteration() + + +def finalize_all_iterations(model: nn.Module, tracked_module_names: List[str]) -> None: + """Finalizes all iterations for specified `TrackedModule` instances. + + Args: + model (nn.Module): + The PyTorch model containing `TrackedModule` instances. + tracked_module_names (List[str]): + Names of modules to finalize all iterations for. + """ + for module in model.modules(): + if isinstance(module, TrackedModule) and module.name in tracked_module_names: + module.finalize_all_iterations() diff --git a/kronfluence/score/__init__.py b/kronfluence/score/__init__.py index 49a4267..41ac691 100644 --- a/kronfluence/score/__init__.py +++ b/kronfluence/score/__init__.py @@ -1,5 +1,5 @@ from .pairwise import ( - _compute_dot_products_with_loader, + compute_dot_products_with_loader, compute_pairwise_scores_with_loaders, load_pairwise_scores, pairwise_scores_exist, diff --git a/kronfluence/score/dot_product.py b/kronfluence/score/dot_product.py new file mode 100644 index 0000000..cd96f3e --- /dev/null +++ b/kronfluence/score/dot_product.py @@ -0,0 +1,257 @@ +from typing import Dict, List, Optional, Union + +import torch +import torch.distributed as dist +from accelerate.utils import send_to_device +from torch import autocast, nn +from torch.cuda.amp import GradScaler +from torch.utils import data +from tqdm import tqdm + +from kronfluence.arguments import FactorArguments, ScoreArguments +from kronfluence.module import TrackedModule +from kronfluence.module.tracked_module import ModuleMode +from kronfluence.module.utils import ( + accumulate_iterations, + exist_for_all_modules, + finalize_all_iterations, + finalize_iteration, + set_mode, + synchronize_modules, +) +from kronfluence.task import Task +from kronfluence.utils.constants import ( + ALL_MODULE_NAME, + DISTRIBUTED_SYNC_INTERVAL, + PAIRWISE_SCORE_MATRIX_NAME, + SCORE_TYPE, +) +from kronfluence.utils.logger import TQDM_BAR_FORMAT +from kronfluence.utils.state import State, no_sync, release_memory + +DIMENSION_NOT_MATCH_ERROR_MSG = ( + "The model does not support token-wise score computation. " + "Set `compute_per_module_scores=True` or `compute_per_token_scores=False` " + "to avoid this error." +) + + +def compute_dot_products_with_loader( + model: nn.Module, + task: Task, + state: State, + train_loader: data.DataLoader, + factor_args: FactorArguments, + score_args: ScoreArguments, + tracked_module_names: List[str], + scaler: GradScaler, + disable_tqdm: bool = False, +) -> Union[Dict[str, torch.Tensor], torch.Tensor]: + """After computing the preconditioned query gradient, compute dot products with individual training gradients.""" + model.zero_grad(set_to_none=True) + set_mode( + model=model, + mode=ModuleMode.PAIRWISE_SCORE, + tracked_module_names=tracked_module_names, + release_memory=False, + ) + release_memory() + + cached_module_lst = [] + for module in model.modules(): + if isinstance(module, TrackedModule) and module.name in tracked_module_names: + cached_module_lst.append(module) + + dataset_size = len(train_loader.dataset) + score_chunks: Dict[str, List[torch.Tensor]] = {} + if score_args.compute_per_module_scores: + for module in cached_module_lst: + score_chunks[module.name] = [] + else: + score_chunks[ALL_MODULE_NAME] = [] + + total_steps = 0 + enable_amp = score_args.amp_dtype is not None + + with tqdm( + total=len(train_loader), + desc="Computing pairwise scores (training gradient)", + bar_format=TQDM_BAR_FORMAT, + disable=not state.is_main_process or disable_tqdm, + ) as pbar: + for batch in train_loader: + batch = send_to_device(tensor=batch, device=state.device) + + with no_sync(model=model, state=state): + model.zero_grad(set_to_none=True) + with autocast(device_type=state.device.type, enabled=enable_amp, dtype=score_args.amp_dtype): + loss = task.compute_train_loss( + batch=batch, + model=model, + sample=False, + ) + scaler.scale(loss).backward() + + if factor_args.has_shared_parameters: + finalize_iteration(model=model, tracked_module_names=tracked_module_names) + + with torch.no_grad(): + if score_args.compute_per_module_scores: + for module in cached_module_lst: + score_chunks[module.name].append( + module.get_factor(factor_name=PAIRWISE_SCORE_MATRIX_NAME).to(device="cpu", copy=True) + ) + else: + pairwise_scores = None + for module in cached_module_lst: + if pairwise_scores is None: + pairwise_scores = torch.zeros_like( + module.get_factor(factor_name=PAIRWISE_SCORE_MATRIX_NAME), requires_grad=False + ) + try: + pairwise_scores.add_(module.get_factor(factor_name=PAIRWISE_SCORE_MATRIX_NAME)) + except RuntimeError as exc: + if score_args.compute_per_token_scores: + raise RuntimeError(DIMENSION_NOT_MATCH_ERROR_MSG) from exc + raise + pairwise_scores = pairwise_scores.cpu() + score_chunks[ALL_MODULE_NAME].append(pairwise_scores) + accumulate_iterations(model=model, tracked_module_names=tracked_module_names) + + if state.use_distributed and total_steps % DISTRIBUTED_SYNC_INTERVAL == 0: + state.wait_for_everyone() + + del loss + total_steps += 1 + pbar.update(1) + + model.zero_grad(set_to_none=True) + finalize_all_iterations(model=model, tracked_module_names=tracked_module_names) + set_mode( + model=model, + mode=ModuleMode.PRECONDITION_GRADIENT, + tracked_module_names=tracked_module_names, + release_memory=False, + ) + release_memory() + + total_scores: SCORE_TYPE = {} + for module_name, chunks in score_chunks.items(): + total_scores[module_name] = torch.cat(chunks, dim=1) + if state.use_distributed: + total_scores[module_name] = total_scores[module_name].to(device=state.device) + gather_list = None + if state.is_main_process: + gather_list = [torch.zeros_like(total_scores[module_name]) for _ in range(state.num_processes)] + dist.gather(total_scores[module_name], gather_list) + if state.is_main_process: + total_scores[module_name] = torch.cat(gather_list, dim=1)[:, :dataset_size].cpu() + else: + total_scores[module_name] = total_scores[module_name].cpu() + state.wait_for_everyone() + + return total_scores + + +def compute_aggregated_dot_products_with_loader( + model: nn.Module, + task: Task, + state: State, + train_loader: data.DataLoader, + factor_args: FactorArguments, + score_args: ScoreArguments, + tracked_module_names: List[str], + scaler: GradScaler, + disable_tqdm: bool = False, +) -> Union[Dict[str, torch.Tensor], torch.Tensor]: + """After computing the preconditioned query gradient, compute dot products with aggregated training gradients.""" + model.zero_grad(set_to_none=True) + set_mode( + model=model, + mode=ModuleMode.GRADIENT_AGGREGATION, + tracked_module_names=tracked_module_names, + release_memory=False, + ) + release_memory() + + scores: Dict[str, Optional[torch.Tensor]] = {} + if score_args.compute_per_module_scores: + for module in model.modules(): + if isinstance(module, TrackedModule) and module.name in tracked_module_names: + scores[module.name] = None + else: + scores[ALL_MODULE_NAME] = None + + enable_amp = score_args.amp_dtype is not None + + if not exist_for_all_modules(model=model, tracked_module_names=tracked_module_names): + with tqdm( + total=len(train_loader), + desc="Computing pairwise scores (training gradient)", + bar_format=TQDM_BAR_FORMAT, + disable=not state.is_main_process or disable_tqdm, + ) as pbar: + for batch in train_loader: + batch = send_to_device(tensor=batch, device=state.device) + + with no_sync(model=model, state=state): + model.zero_grad(set_to_none=True) + with autocast(device_type=state.device.type, enabled=enable_amp, dtype=score_args.amp_dtype): + loss = task.compute_train_loss( + batch=batch, + model=model, + sample=False, + ) + scaler.scale(loss).backward() + + if factor_args.has_shared_parameters: + finalize_iteration(model=model, tracked_module_names=tracked_module_names) + + del loss + pbar.update(1) + + if state.use_distributed: + synchronize_modules(model=model, tracked_module_names=tracked_module_names) + + set_mode( + model=model, + mode=ModuleMode.PAIRWISE_SCORE, + tracked_module_names=tracked_module_names, + release_memory=False, + ) + finalize_all_iterations(model=model, tracked_module_names=tracked_module_names) + + with torch.no_grad(): + if score_args.compute_per_module_scores: + for module in model.modules(): + if isinstance(module, TrackedModule) and module.name in tracked_module_names: + scores[module.name] = module.get_factor(factor_name=PAIRWISE_SCORE_MATRIX_NAME).to( + device="cpu", copy=True + ) + else: + pairwise_scores = None + for module in model.modules(): + if isinstance(module, TrackedModule) and module.name in tracked_module_names: + if pairwise_scores is None: + pairwise_scores = torch.zeros_like( + module.get_factor(factor_name=PAIRWISE_SCORE_MATRIX_NAME), requires_grad=False + ) + try: + pairwise_scores.add_(module.get_factor(factor_name=PAIRWISE_SCORE_MATRIX_NAME)) + except RuntimeError as exc: + if score_args.compute_per_token_scores: + raise RuntimeError(DIMENSION_NOT_MATCH_ERROR_MSG) from exc + raise + scores[ALL_MODULE_NAME] = pairwise_scores.cpu() + accumulate_iterations(model=model, tracked_module_names=tracked_module_names) + + model.zero_grad(set_to_none=True) + set_mode( + model=model, + mode=ModuleMode.PRECONDITION_GRADIENT, + tracked_module_names=tracked_module_names, + release_memory=False, + ) + release_memory() + + return scores diff --git a/kronfluence/score/pairwise.py b/kronfluence/score/pairwise.py index a20a4b3..ff7260d 100644 --- a/kronfluence/score/pairwise.py +++ b/kronfluence/score/pairwise.py @@ -10,38 +10,47 @@ from tqdm import tqdm from kronfluence.arguments import FactorArguments, ScoreArguments -from kronfluence.module import TrackedModule from kronfluence.module.tracked_module import ModuleMode from kronfluence.module.utils import ( - aggregate_preconditioned_gradient, + accumulate_iterations, + finalize_all_iterations, + finalize_iteration, get_tracked_module_names, - release_scores, - remove_gradient_scale, + prepare_modules, set_factors, set_gradient_scale, set_mode, - synchronize_preconditioned_gradient, - truncate_preconditioned_gradient, + synchronize_modules, + truncate, update_factor_args, update_score_args, ) -from kronfluence.task import Task -from kronfluence.utils.constants import ( - ALL_MODULE_NAME, - FACTOR_TYPE, - PAIRWISE_SCORE_MATRIX_NAME, - PARTITION_TYPE, - SCORE_TYPE, +from kronfluence.score.dot_product import ( + compute_aggregated_dot_products_with_loader, + compute_dot_products_with_loader, ) +from kronfluence.task import Task +from kronfluence.utils.constants import FACTOR_TYPE, PARTITION_TYPE, SCORE_TYPE from kronfluence.utils.logger import TQDM_BAR_FORMAT -from kronfluence.utils.state import State, no_sync, release_memory +from kronfluence.utils.state import State, no_sync def pairwise_scores_save_path( output_dir: Path, partition: Optional[PARTITION_TYPE] = None, ) -> Path: - """Generates the path for saving/loading pairwise scores.""" + """Generates the path for saving or loading pairwise influence scores. + + Args: + output_dir (Path): + Directory to save or load the matrices. + partition (PARTITION_TYPE, optional): + Partition information, if any. + + Returns: + Path: + The full path for the score file. + """ if partition is not None: data_partition, module_partition = partition return output_dir / ( @@ -50,25 +59,24 @@ def pairwise_scores_save_path( return output_dir / "pairwise_scores.safetensors" -def pairwise_scores_exist( - output_dir: Path, - partition: Optional[PARTITION_TYPE] = None, -) -> bool: - """Check if the pairwise scores exist at the specified path.""" - save_path = pairwise_scores_save_path( - output_dir=output_dir, - partition=partition, - ) - return save_path.exists() - - def save_pairwise_scores( output_dir: Path, scores: SCORE_TYPE, partition: Optional[PARTITION_TYPE] = None, metadata: Optional[Dict[str, str]] = None, ) -> None: - """Saves pairwise influence scores to disk.""" + """Saves pairwise scores to disk. + + Args: + output_dir (Path): + Directory to save the scores. + scores (SCORE_TYPE): + Dictionary of scores to save. + partition (PARTITION_TYPE, optional): + Partition information, if any. + metadata (Dict[str, str], optional): + Additional metadata to save with the scores. + """ save_path = pairwise_scores_save_path( output_dir=output_dir, partition=partition, @@ -79,8 +87,19 @@ def save_pairwise_scores( def load_pairwise_scores( output_dir: Path, partition: Optional[PARTITION_TYPE] = None, -) -> Dict[str, torch.Tensor]: - """Loads pairwise scores from disk.""" +) -> SCORE_TYPE: + """Loads pairwise scores from disk. + + Args: + output_dir (Path): + Directory to load the scores from. + partition (PARTITION_TYPE, optional): + Partition information, if any. + + Returns: + SCORE_TYPE: + Dictionary of loaded scores. + """ save_path = pairwise_scores_save_path( output_dir=output_dir, partition=partition, @@ -88,105 +107,27 @@ def load_pairwise_scores( return load_file(filename=save_path) -def _compute_dot_products_with_loader( - model: nn.Module, - task: Task, - state: State, - train_loader: data.DataLoader, - score_args: ScoreArguments, - tracked_module_names: List[str], -) -> Union[Dict[str, torch.Tensor], torch.Tensor]: - """After computing the preconditioned query gradient, compute dot products with training gradients.""" - with torch.no_grad(): - set_mode( - model=model, - mode=ModuleMode.PAIRWISE_SCORE, - tracked_module_names=tracked_module_names, - keep_factors=True, - ) - dataset_size = len(train_loader.dataset) - score_chunks: Dict[str, List[torch.Tensor]] = {} - if score_args.per_module_score: - for module in model.modules(): - if isinstance(module, TrackedModule) and module.name in tracked_module_names: - score_chunks[module.name] = [] - else: - score_chunks[ALL_MODULE_NAME] = [] - total_steps = 0 - enable_amp = score_args.amp_dtype is not None - scaler = GradScaler(enabled=enable_amp) - - with tqdm( - total=len(train_loader), - desc="Computing pairwise scores (training gradient)", - bar_format=TQDM_BAR_FORMAT, - disable=not state.is_main_process, - ) as pbar: - for batch in train_loader: - batch = send_to_device(tensor=batch, device=state.device) - - with no_sync(model=model, state=state): - model.zero_grad(set_to_none=True) - with autocast(device_type=state.device.type, enabled=enable_amp, dtype=score_args.amp_dtype): - loss = task.compute_train_loss( - batch=batch, - model=model, - sample=False, - ) - scaled_loss = scaler.scale(loss) - if enable_amp: - gradient_scale = 1.0 / scaler.get_scale() - set_gradient_scale(model=model, gradient_scale=gradient_scale) - scaled_loss.backward() - total_steps += 1 - - with torch.no_grad(): - if score_args.per_module_score: - for module in model.modules(): - if isinstance(module, TrackedModule) and module.name in tracked_module_names: - score_chunks[module.name].append( - module.get_factor(factor_name=PAIRWISE_SCORE_MATRIX_NAME).cpu() - ) - else: - # Aggregate the pairwise scores across all modules. - pairwise_scores = None - for module in model.modules(): - if isinstance(module, TrackedModule) and module.name in tracked_module_names: - if pairwise_scores is None: - pairwise_scores = module.get_factor(factor_name=PAIRWISE_SCORE_MATRIX_NAME) - else: - pairwise_scores.add_(module.get_factor(factor_name=PAIRWISE_SCORE_MATRIX_NAME)) - score_chunks[ALL_MODULE_NAME].append(pairwise_scores.cpu()) - release_scores(model=model) - - if state.use_distributed and total_steps % score_args.distributed_sync_steps == 0: - state.wait_for_everyone() +def pairwise_scores_exist( + output_dir: Path, + partition: Optional[PARTITION_TYPE] = None, +) -> bool: + """Checks if pairwise influence scores exist at the specified directory. - pbar.update(1) + Args: + output_dir (Path): + Directory to check for scores. + partition (PARTITION_TYPE, optional): + Partition information, if any. - with torch.no_grad(): - model.zero_grad(set_to_none=True) - remove_gradient_scale(model=model) - set_mode( - model=model, - mode=ModuleMode.PRECONDITION_GRADIENT, - tracked_module_names=tracked_module_names, - keep_factors=False, - ) - total_scores: SCORE_TYPE = {} - for module_name, chunks in score_chunks.items(): - total_scores[module_name] = torch.cat(chunks, dim=1) - if state.use_distributed: - release_memory() - total_scores[module_name] = total_scores[module_name].to(device=state.device) - gather_list = None - if state.is_main_process: - gather_list = [torch.zeros_like(total_scores[module_name]) for _ in range(state.num_processes)] - torch.distributed.gather(total_scores[module_name], gather_list) - if state.is_main_process: - total_scores[module_name] = torch.cat(gather_list, dim=1)[:, :dataset_size].cpu() - state.wait_for_everyone() - return total_scores + Returns: + bool: + `True` if scores exist, `False` otherwise. + """ + save_path = pairwise_scores_save_path( + output_dir=output_dir, + partition=partition, + ) + return save_path.exists() def compute_pairwise_scores_with_loaders( @@ -200,12 +141,13 @@ def compute_pairwise_scores_with_loaders( score_args: ScoreArguments, factor_args: FactorArguments, tracked_module_names: Optional[List[str]], + disable_tqdm: bool = False, ) -> Dict[str, torch.Tensor]: """Computes pairwise influence scores for a given model and task. Args: loaded_factors (FACTOR_TYPE): - The factor results to load from, before computing the pairwise scores. + Computed factors. model (nn.Module): The model for which pairwise influence scores will be computed. state (State): @@ -219,39 +161,38 @@ def compute_pairwise_scores_with_loaders( train_loader (data.DataLoader): The data loader that will be used to compute training gradients. score_args (ScoreArguments): - Arguments related to computing pairwise scores. + Arguments for computing pairwise scores. factor_args (FactorArguments): - Arguments related to computing preconditioning factors. + Arguments used to compute factors. tracked_module_names (List[str], optional): A list of module names that pairwise scores will be computed. If not specified, scores will be computed for all available tracked modules. + disable_tqdm (bool, optional): + Whether to disable the progress bar. Defaults to `False`. Returns: - Dict[str, torch.Tensor]: - A dictionary containing the module name and its pairwise influence scores. If - `score_args.per_module_score` is False, the pairwise scores will be aggregated across all modules. + SCORE_TYPE: + A dictionary containing the module name and its pairwise influence scores. """ - with torch.no_grad(): - update_factor_args(model=model, factor_args=factor_args) - update_score_args(model=model, score_args=score_args) - - # Loads necessary factors before computing pairwise influence scores. - if len(loaded_factors) > 0: - for name in loaded_factors: - set_factors( - model=model, - factor_name=name, - factors=loaded_factors[name], - ) - - if tracked_module_names is None: - tracked_module_names = get_tracked_module_names(model=model) - set_mode( - model=model, - mode=ModuleMode.PRECONDITION_GRADIENT, - tracked_module_names=tracked_module_names, - keep_factors=True, - ) + update_factor_args(model=model, factor_args=factor_args) + update_score_args(model=model, score_args=score_args) + if tracked_module_names is None: + tracked_module_names = get_tracked_module_names(model=model) + set_mode( + model=model, + mode=ModuleMode.PRECONDITION_GRADIENT, + tracked_module_names=tracked_module_names, + release_memory=True, + ) + if len(loaded_factors) > 0: + for name in loaded_factors: + set_factors( + model=model, + factor_name=name, + factors=loaded_factors[name], + clone=True, + ) + prepare_modules(model=model, tracked_module_names=tracked_module_names, device=state.device) total_scores_chunks: Dict[str, Union[List[torch.Tensor], torch.Tensor]] = {} total_query_batch_size = per_device_query_batch_size * state.num_processes @@ -259,15 +200,25 @@ def compute_pairwise_scores_with_loaders( num_batches = len(query_loader) query_iter = iter(query_loader) - num_aggregates = 0 + num_accumulations = 0 enable_amp = score_args.amp_dtype is not None - scaler = GradScaler(enabled=enable_amp) + enable_grad_scaler = enable_amp and factor_args.amp_dtype == torch.float16 + scaler = GradScaler(init_scale=factor_args.amp_scale, enabled=enable_grad_scaler) + if enable_grad_scaler: + gradient_scale = 1.0 / scaler.get_scale() + set_gradient_scale(model=model, gradient_scale=gradient_scale) + + dot_product_func = ( + compute_aggregated_dot_products_with_loader + if score_args.aggregate_train_gradients + else compute_dot_products_with_loader + ) with tqdm( total=num_batches, desc="Computing pairwise scores (query gradient)", bar_format=TQDM_BAR_FORMAT, - disable=not state.is_main_process, + disable=not state.is_main_process or disable_tqdm, ) as pbar: for query_index in range(num_batches): query_batch = next(query_iter) @@ -280,54 +231,163 @@ def compute_pairwise_scores_with_loaders( model.zero_grad(set_to_none=True) with autocast(device_type=state.device.type, enabled=enable_amp, dtype=score_args.amp_dtype): measurement = task.compute_measurement(batch=query_batch, model=model) - scaled_measurement = scaler.scale(measurement) - if enable_amp: - gradient_scale = 1.0 / scaler.get_scale() - set_gradient_scale(model=model, gradient_scale=gradient_scale) - scaled_measurement.backward() + scaler.scale(measurement).backward() + + if factor_args.has_shared_parameters: + finalize_iteration(model=model, tracked_module_names=tracked_module_names) if state.use_distributed: # Stack preconditioned query gradient across multiple devices or nodes. - synchronize_preconditioned_gradient(model=model, num_processes=state.num_processes) + synchronize_modules( + model=model, tracked_module_names=tracked_module_names, num_processes=state.num_processes + ) if query_index == len(query_loader) - 1 and query_remainder > 0: - # Remove duplicate data points if the dataset is not exactly divisible - # by the current batch size. - truncate_preconditioned_gradient(model=model, keep_size=query_remainder) - - aggregate_preconditioned_gradient(model=model) - num_aggregates += 1 + # Removes duplicate data points if the dataset is not evenly divisible by the current batch size. + truncate(model=model, tracked_module_names=tracked_module_names, keep_size=query_remainder) + accumulate_iterations(model=model, tracked_module_names=tracked_module_names) del query_batch, measurement - model.zero_grad(set_to_none=True) - remove_gradient_scale(model=model) - release_memory() - if num_aggregates < score_args.num_query_gradient_aggregations and query_index != len(query_loader) - 1: + num_accumulations += 1 + if ( + num_accumulations < score_args.query_gradient_accumulation_steps + and query_index != len(query_loader) - 1 + ): + pbar.update(1) continue - # Compute the dot product between preconditioning query gradient and all training gradients. - num_aggregates = 0 - scores = _compute_dot_products_with_loader( + # Computes the dot product between preconditioning query gradient and all training gradients. + scores = dot_product_func( model=model, state=state, task=task, train_loader=train_loader, + factor_args=factor_args, score_args=score_args, tracked_module_names=tracked_module_names, + scaler=scaler, + disable_tqdm=disable_tqdm, ) - with torch.no_grad(): - if state.is_main_process: - for module_name, current_scores in scores.items(): - if module_name not in total_scores_chunks: - total_scores_chunks[module_name] = [] - total_scores_chunks[module_name].append(current_scores) - state.wait_for_everyone() + if state.is_main_process: + for module_name, current_scores in scores.items(): + if module_name not in total_scores_chunks: + total_scores_chunks[module_name] = [] + total_scores_chunks[module_name].append(current_scores) + del scores + state.wait_for_everyone() + + num_accumulations = 0 pbar.update(1) - with torch.no_grad(): - set_mode(model=model, mode=ModuleMode.DEFAULT, keep_factors=False) - if state.is_main_process: - for module_name in total_scores_chunks: - total_scores_chunks[module_name] = torch.cat(total_scores_chunks[module_name], dim=0) - state.wait_for_everyone() + if state.is_main_process: + for module_name in total_scores_chunks: + total_scores_chunks[module_name] = torch.cat(total_scores_chunks[module_name], dim=0) + + model.zero_grad(set_to_none=True) + if enable_grad_scaler: + set_gradient_scale(model=model, gradient_scale=1.0) + finalize_all_iterations(model=model, tracked_module_names=tracked_module_names) + set_mode(model=model, mode=ModuleMode.DEFAULT, release_memory=True) + state.wait_for_everyone() + return total_scores_chunks + + +def compute_pairwise_query_aggregated_scores_with_loaders( + loaded_factors: FACTOR_TYPE, + model: nn.Module, + state: State, + task: Task, + query_loader: data.DataLoader, + per_device_query_batch_size: int, + train_loader: data.DataLoader, + score_args: ScoreArguments, + factor_args: FactorArguments, + tracked_module_names: Optional[List[str]], + disable_tqdm: bool = False, +) -> Dict[str, torch.Tensor]: + """Computes pairwise influence scores (with query gradients aggregated) for a given model and task. See + `compute_pairwise_scores_with_loaders` for detailed information.""" + del per_device_query_batch_size + update_factor_args(model=model, factor_args=factor_args) + update_score_args(model=model, score_args=score_args) + if tracked_module_names is None: + tracked_module_names = get_tracked_module_names(model=model) + set_mode( + model=model, + mode=ModuleMode.GRADIENT_AGGREGATION, + tracked_module_names=tracked_module_names, + release_memory=True, + ) + if len(loaded_factors) > 0: + for name in loaded_factors: + set_factors(model=model, factor_name=name, factors=loaded_factors[name], clone=True) + prepare_modules(model=model, tracked_module_names=tracked_module_names, device=state.device) + + enable_amp = score_args.amp_dtype is not None + enable_grad_scaler = enable_amp and factor_args.amp_dtype == torch.float16 + scaler = GradScaler(init_scale=factor_args.amp_scale, enabled=enable_grad_scaler) + if enable_grad_scaler: + gradient_scale = 1.0 / scaler.get_scale() + set_gradient_scale(model=model, gradient_scale=gradient_scale) + + dot_product_func = ( + compute_aggregated_dot_products_with_loader + if score_args.aggregate_train_gradients + else compute_dot_products_with_loader + ) + + with tqdm( + total=len(query_loader), + desc="Computing pairwise scores (query gradient)", + bar_format=TQDM_BAR_FORMAT, + disable=not state.is_main_process or disable_tqdm, + ) as pbar: + for query_batch in query_loader: + query_batch = send_to_device( + tensor=query_batch, + device=state.device, + ) + + with no_sync(model=model, state=state): + model.zero_grad(set_to_none=True) + with autocast(device_type=state.device.type, enabled=enable_amp, dtype=score_args.amp_dtype): + measurement = task.compute_measurement(batch=query_batch, model=model) + scaler.scale(measurement).backward() + + if factor_args.has_shared_parameters: + finalize_iteration(model=model, tracked_module_names=tracked_module_names) + + del measurement + pbar.update(1) + + if state.use_distributed: + synchronize_modules(model=model, tracked_module_names=tracked_module_names) + + set_mode( + model=model, + mode=ModuleMode.PRECONDITION_GRADIENT, + tracked_module_names=tracked_module_names, + release_memory=False, + ) + finalize_all_iterations(model=model, tracked_module_names=tracked_module_names) + + scores = dot_product_func( + model=model, + state=state, + task=task, + train_loader=train_loader, + factor_args=factor_args, + score_args=score_args, + tracked_module_names=tracked_module_names, + scaler=scaler, + disable_tqdm=disable_tqdm, + ) + + model.zero_grad(set_to_none=True) + if enable_grad_scaler: + set_gradient_scale(model=model, gradient_scale=1.0) + set_mode(model=model, mode=ModuleMode.DEFAULT, release_memory=True) + state.wait_for_everyone() + + return scores diff --git a/kronfluence/score/self.py b/kronfluence/score/self.py index 539fbe5..e2c2aa3 100644 --- a/kronfluence/score/self.py +++ b/kronfluence/score/self.py @@ -2,7 +2,8 @@ from typing import Dict, List, Optional import torch -from accelerate.utils import find_batch_size, send_to_device +import torch.distributed as dist +from accelerate.utils import send_to_device from safetensors.torch import load_file, save_file from torch import autocast, nn from torch.cuda.amp import GradScaler @@ -13,9 +14,10 @@ from kronfluence.module import TrackedModule from kronfluence.module.tracked_module import ModuleMode from kronfluence.module.utils import ( + accumulate_iterations, + finalize_iteration, get_tracked_module_names, - release_scores, - remove_gradient_scale, + prepare_modules, set_factors, set_gradient_scale, set_mode, @@ -25,6 +27,7 @@ from kronfluence.task import Task from kronfluence.utils.constants import ( ALL_MODULE_NAME, + DISTRIBUTED_SYNC_INTERVAL, FACTOR_TYPE, PARTITION_TYPE, SCORE_TYPE, @@ -38,7 +41,18 @@ def self_scores_save_path( output_dir: Path, partition: Optional[PARTITION_TYPE] = None, ) -> Path: - """Generates the path for saving/loading self-influence scores.""" + """Generates the path for saving or loading self-influence scores. + + Args: + output_dir (Path): + Directory to save or load the matrices. + partition (PARTITION_TYPE, optional): + Partition information, if any. + + Returns: + Path: + The full path for the score file. + """ if partition is not None: data_partition, module_partition = partition return output_dir / ( @@ -47,42 +61,75 @@ def self_scores_save_path( return output_dir / "self_scores.safetensors" -def self_scores_exist( +def save_self_scores( output_dir: Path, + scores: SCORE_TYPE, partition: Optional[PARTITION_TYPE] = None, -) -> bool: - """Check if the self-influence scores exist at the specified path.""" + metadata: Optional[Dict[str, str]] = None, +) -> None: + """Saves self-influence scores to disk. + + Args: + output_dir (Path): + Directory to save the scores. + scores (SCORE_TYPE): + Dictionary of scores to save. + partition (PARTITION_TYPE, optional): + Partition information, if any. + metadata (Dict[str, str], optional): + Additional metadata to save with the scores. + """ save_path = self_scores_save_path( output_dir=output_dir, partition=partition, ) - return save_path.exists() + save_file(tensors=scores, filename=save_path, metadata=metadata) -def save_self_scores( +def load_self_scores( output_dir: Path, - scores: SCORE_TYPE, partition: Optional[PARTITION_TYPE] = None, - metadata: Optional[Dict[str, str]] = None, -) -> None: - """Saves self-influence scores to disk.""" +) -> SCORE_TYPE: + """Loads self-influence scores from disk. + + Args: + output_dir (Path): + Directory to load the scores from. + partition (PARTITION_TYPE, optional): + Partition information, if any. + + Returns: + SCORE_TYPE: + Dictionary of loaded scores. + """ save_path = self_scores_save_path( output_dir=output_dir, partition=partition, ) - save_file(tensors=scores, filename=save_path, metadata=metadata) + return load_file(filename=save_path) -def load_self_scores( +def self_scores_exist( output_dir: Path, partition: Optional[PARTITION_TYPE] = None, -) -> Dict[str, torch.Tensor]: - """Loads self-influence scores from disk.""" +) -> bool: + """Checks if self-influence scores exist at the specified directory. + + Args: + output_dir (Path): + Directory to check for scores. + partition (PARTITION_TYPE, optional): + Partition information, if any. + + Returns: + bool: + `True` if scores exist, `False` otherwise. + """ save_path = self_scores_save_path( output_dir=output_dir, partition=partition, ) - return load_file(filename=save_path) + return save_path.exists() def compute_self_scores_with_loaders( @@ -94,12 +141,13 @@ def compute_self_scores_with_loaders( score_args: ScoreArguments, factor_args: FactorArguments, tracked_module_names: Optional[List[str]], + disable_tqdm: bool = False, ) -> Dict[str, torch.Tensor]: """Computes self-influence scores for a given model and task. Args: loaded_factors (FACTOR_TYPE): - The factor results to load from, before computing the self-influence scores. + Computed factors. model (nn.Module): The model for which self-influence scores will be computed. state (State): @@ -109,63 +157,63 @@ def compute_self_scores_with_loaders( train_loader (data.DataLoader): The data loader that will be used to compute training gradients. score_args (ScoreArguments): - Arguments related to computing self-influence scores. + Arguments for computing self-influence scores. factor_args (FactorArguments): - Arguments related to computing preconditioning factors. + Arguments used to compute factors. tracked_module_names (List[str], optional): A list of module names that self-influence scores will be computed. If not specified, scores will be computed for all available tracked modules. + disable_tqdm (bool, optional): + Whether to disable the progress bar. Defaults to `False`. Returns: - Dict[str, torch.Tensor]: - A dictionary containing the module name and its self-influence scores. If - `score_args.per_module_score` is False, the self-influence scores will be aggregated across all modules. + SCORE_TYPE: + A dictionary containing the module name and its self-influence scores. """ - with torch.no_grad(): - update_factor_args(model=model, factor_args=factor_args) - update_score_args(model=model, score_args=score_args) - set_mode( - model=model, - mode=ModuleMode.DEFAULT, - keep_factors=False, - ) - - # Loads necessary factors before computing self-influence scores. - if len(loaded_factors) > 0: - for name in loaded_factors: - set_factors( - model=model, - factor_name=name, - factors=loaded_factors[name], - ) - if tracked_module_names is None: - tracked_module_names = get_tracked_module_names(model=model) - set_mode( - model=model, - mode=ModuleMode.SELF_SCORE, - tracked_module_names=tracked_module_names, - keep_factors=True, - ) + update_factor_args(model=model, factor_args=factor_args) + update_score_args(model=model, score_args=score_args) + if tracked_module_names is None: + tracked_module_names = get_tracked_module_names(model=model) + set_mode( + model=model, + mode=ModuleMode.SELF_SCORE, + tracked_module_names=tracked_module_names, + release_memory=True, + ) + if len(loaded_factors) > 0: + for name in loaded_factors: + set_factors(model=model, factor_name=name, factors=loaded_factors[name], clone=True) + prepare_modules(model=model, tracked_module_names=tracked_module_names, device=state.device) dataset_size = len(train_loader.dataset) score_chunks: Dict[str, List[torch.Tensor]] = {} - if score_args.per_module_score: + if score_args.compute_per_module_scores: for module in model.modules(): if isinstance(module, TrackedModule) and module.name in tracked_module_names: score_chunks[module.name] = [] else: score_chunks[ALL_MODULE_NAME] = [] + + cached_module_lst = [] + for module in model.modules(): + if isinstance(module, TrackedModule) and module.name in tracked_module_names: + cached_module_lst.append(module) + total_steps = 0 enable_amp = score_args.amp_dtype is not None - scaler = GradScaler(enabled=enable_amp) + enable_grad_scaler = enable_amp and factor_args.amp_dtype == torch.float16 + scaler = GradScaler(init_scale=factor_args.amp_scale, enabled=enable_grad_scaler) + if enable_grad_scaler: + gradient_scale = 1.0 / scaler.get_scale() + set_gradient_scale(model=model, gradient_scale=gradient_scale) with tqdm( total=len(train_loader), desc="Computing self-influence scores", bar_format=TQDM_BAR_FORMAT, - disable=not state.is_main_process, + disable=not state.is_main_process or disable_tqdm, ) as pbar: - for batch in train_loader: + for index, batch in enumerate(train_loader): batch = send_to_device( tensor=batch, device=state.device, @@ -179,62 +227,66 @@ def compute_self_scores_with_loaders( model=model, sample=False, ) - scaled_loss = scaler.scale(loss) - if enable_amp: - gradient_scale = 1.0 / scaler.get_scale() - set_gradient_scale(model=model, gradient_scale=gradient_scale) - scaled_loss.backward() - total_steps += 1 + scaler.scale(loss).backward() + + if factor_args.has_shared_parameters: + finalize_iteration(model=model, tracked_module_names=tracked_module_names) with torch.no_grad(): - if score_args.per_module_score: - for module in model.modules(): - if isinstance(module, TrackedModule) and module.name in tracked_module_names: - score_chunks[module.name].append( - module.get_factor(factor_name=SELF_SCORE_VECTOR_NAME).cpu() - ) + if score_args.compute_per_module_scores: + for module in cached_module_lst: + score_chunks[module.name].append( + module.get_factor(factor_name=SELF_SCORE_VECTOR_NAME).to(device="cpu", copy=True) + ) else: - # Aggregate the self-influence scores across all modules. - batch_size = find_batch_size(batch) - self_scores = torch.zeros( - size=(batch_size,), - dtype=score_args.score_dtype, - device=state.device, - requires_grad=False, - ) - for module in model.modules(): - if isinstance(module, TrackedModule) and module.name in tracked_module_names: - self_scores.add_(module.get_factor(factor_name=SELF_SCORE_VECTOR_NAME)) - score_chunks[ALL_MODULE_NAME].append(self_scores.cpu()) - release_scores(model=model) - - if state.use_distributed and total_steps % score_args.distributed_sync_steps == 0: + self_scores = None + for module in cached_module_lst: + if self_scores is None: + self_scores = torch.zeros_like( + module.get_factor(factor_name=SELF_SCORE_VECTOR_NAME), requires_grad=False + ) + self_scores.add_(module.get_factor(factor_name=SELF_SCORE_VECTOR_NAME)) + self_scores = self_scores.cpu() + score_chunks[ALL_MODULE_NAME].append(self_scores) + accumulate_iterations(model=model, tracked_module_names=tracked_module_names) + + if ( + state.use_distributed + and total_steps % DISTRIBUTED_SYNC_INTERVAL == 0 + and index not in [len(train_loader) - 1, len(train_loader) - 2] + ): state.wait_for_everyone() + del loss + total_steps += 1 pbar.update(1) - with torch.no_grad(): - model.zero_grad(set_to_none=True) - remove_gradient_scale(model=model) - set_mode( - model=model, - mode=ModuleMode.DEFAULT, - tracked_module_names=tracked_module_names, - keep_factors=False, - ) - total_scores: SCORE_TYPE = {} - for module_name, chunks in score_chunks.items(): - total_scores[module_name] = torch.cat(chunks, dim=0) - if state.use_distributed: - release_memory() - total_scores[module_name] = total_scores[module_name].to(device=state.device) - gather_list = None - if state.is_main_process: - gather_list = [torch.zeros_like(total_scores[module_name]) for _ in range(state.num_processes)] - torch.distributed.gather(total_scores[module_name], gather_list) - if state.is_main_process: - total_scores[module_name] = torch.cat(gather_list, dim=0)[:dataset_size].cpu() + model.zero_grad(set_to_none=True) + if enable_grad_scaler: + set_gradient_scale(model=model, gradient_scale=1.0) + set_mode( + model=model, + mode=ModuleMode.DEFAULT, + tracked_module_names=tracked_module_names, + release_memory=True, + ) + release_memory() + + total_scores: SCORE_TYPE = {} + for module_name, chunks in score_chunks.items(): + total_scores[module_name] = torch.cat(chunks, dim=0) + if state.use_distributed: + total_scores[module_name] = total_scores[module_name].to(device=state.device) + gather_list = None + if state.is_main_process: + gather_list = [torch.zeros_like(total_scores[module_name]) for _ in range(state.num_processes)] + dist.gather(total_scores[module_name], gather_list) + if state.is_main_process: + total_scores[module_name] = torch.cat(gather_list, dim=0)[:dataset_size].cpu() + else: + total_scores[module_name] = total_scores[module_name].cpu() state.wait_for_everyone() + return total_scores @@ -247,72 +299,52 @@ def compute_self_measurement_scores_with_loaders( score_args: ScoreArguments, factor_args: FactorArguments, tracked_module_names: Optional[List[str]], + disable_tqdm: bool = False, ) -> Dict[str, torch.Tensor]: - """Computes self-influence scores with measurement (instead of the loss) for a given model and task. - - Args: - loaded_factors (FACTOR_TYPE): - The factor results to load from, before computing the self-influence scores. - model (nn.Module): - The model for which self-influence scores will be computed. - state (State): - The current process's information (e.g., device being used). - task (Task): - The specific task associated with the model. - train_loader (data.DataLoader): - The data loader that will be used to compute training gradients. - score_args (ScoreArguments): - Arguments related to computing self-influence scores. - factor_args (FactorArguments): - Arguments related to computing preconditioning factors. - tracked_module_names (List[str], optional): - A list of module names that self-influence scores will be computed. If not specified, scores - will be computed for all available tracked modules. - - Returns: - Dict[str, torch.Tensor]: - A dictionary containing the module name and its self-influence scores. If - `score_args.per_module_score` is False, the self-influence scores will be aggregated across all modules. - """ - with torch.no_grad(): - update_factor_args(model=model, factor_args=factor_args) - update_score_args(model=model, score_args=score_args) - set_mode( - model=model, - mode=ModuleMode.DEFAULT, - keep_factors=False, - ) + """Computes self-influence scores with measurement (instead of the loss) for a given model and task. See + `compute_self_scores_with_loaders` for the detailed docstring.""" + update_factor_args(model=model, factor_args=factor_args) + update_score_args(model=model, score_args=score_args) + if tracked_module_names is None: + tracked_module_names = get_tracked_module_names(model=model) + if len(loaded_factors) > 0: + for name in loaded_factors: + set_factors( + model=model, + factor_name=name, + factors=loaded_factors[name], + clone=True, + ) + prepare_modules(model=model, tracked_module_names=tracked_module_names, device=state.device) - # Loads necessary factors before computing self-influence scores. - if len(loaded_factors) > 0: - for name in loaded_factors: - set_factors( - model=model, - factor_name=name, - factors=loaded_factors[name], - ) - if tracked_module_names is None: - tracked_module_names = get_tracked_module_names(model=model) + cached_module_lst = [] + for module in model.modules(): + if isinstance(module, TrackedModule) and module.name in tracked_module_names: + cached_module_lst.append(module) dataset_size = len(train_loader.dataset) score_chunks: Dict[str, List[torch.Tensor]] = {} - if score_args.per_module_score: - for module in model.modules(): - if isinstance(module, TrackedModule) and module.name in tracked_module_names: - score_chunks[module.name] = [] + if score_args.compute_per_module_scores: + for module in cached_module_lst: + score_chunks[module.name] = [] else: score_chunks[ALL_MODULE_NAME] = [] + total_steps = 0 enable_amp = score_args.amp_dtype is not None - scaler = GradScaler(enabled=enable_amp) + enable_grad_scaler = enable_amp and factor_args.amp_dtype == torch.float16 + scaler = GradScaler(init_scale=factor_args.amp_scale, enabled=enable_grad_scaler) + if enable_grad_scaler: + gradient_scale = 1.0 / scaler.get_scale() + set_gradient_scale(model=model, gradient_scale=gradient_scale) with tqdm( total=len(train_loader), desc="Computing self-influence scores", bar_format=TQDM_BAR_FORMAT, - disable=not state.is_main_process, + disable=not state.is_main_process or disable_tqdm, ) as pbar: - for batch in train_loader: + for index, batch in enumerate(train_loader): batch = send_to_device( tensor=batch, device=state.device, @@ -322,23 +354,23 @@ def compute_self_measurement_scores_with_loaders( model=model, mode=ModuleMode.PRECONDITION_GRADIENT, tracked_module_names=tracked_module_names, - keep_factors=True, + release_memory=False, ) with no_sync(model=model, state=state): model.zero_grad(set_to_none=True) with autocast(device_type=state.device.type, enabled=enable_amp, dtype=score_args.amp_dtype): measurement = task.compute_measurement(batch=batch, model=model) - scaled_measurement = scaler.scale(measurement) - if enable_amp: - gradient_scale = 1.0 / scaler.get_scale() - set_gradient_scale(model=model, gradient_scale=gradient_scale) - scaled_measurement.backward() + scaler.scale(measurement).backward() + + if factor_args.has_shared_parameters: + finalize_iteration(model=model, tracked_module_names=tracked_module_names) + del measurement set_mode( model=model, mode=ModuleMode.SELF_MEASUREMENT_SCORE, tracked_module_names=tracked_module_names, - keep_factors=True, + release_memory=False, ) with no_sync(model=model, state=state): model.zero_grad(set_to_none=True) @@ -348,60 +380,64 @@ def compute_self_measurement_scores_with_loaders( model=model, sample=False, ) - scaled_loss = scaler.scale(loss) - if enable_amp: - gradient_scale = 1.0 / scaler.get_scale() - set_gradient_scale(model=model, gradient_scale=gradient_scale) - scaled_loss.backward() - total_steps += 1 + scaler.scale(loss).backward() + + if factor_args.has_shared_parameters: + finalize_iteration(model=model, tracked_module_names=tracked_module_names) + del loss with torch.no_grad(): - if score_args.per_module_score: - for module in model.modules(): - if isinstance(module, TrackedModule) and module.name in tracked_module_names: - score_chunks[module.name].append( - module.get_factor(factor_name=SELF_SCORE_VECTOR_NAME).cpu() - ) + if score_args.compute_per_module_scores: + for module in cached_module_lst: + score_chunks[module.name].append( + module.get_factor(factor_name=SELF_SCORE_VECTOR_NAME).to(device="cpu", copy=True) + ) else: - # Aggregate the self-influence scores across all modules. - batch_size = find_batch_size(batch) - self_scores = torch.zeros( - size=(batch_size,), - dtype=score_args.score_dtype, - device=state.device, - requires_grad=False, - ) - for module in model.modules(): - if isinstance(module, TrackedModule) and module.name in tracked_module_names: - self_scores.add_(module.get_factor(factor_name=SELF_SCORE_VECTOR_NAME)) - score_chunks[ALL_MODULE_NAME].append(self_scores.cpu()) - release_scores(model=model) - - if state.use_distributed and total_steps % score_args.distributed_sync_steps == 0: + self_scores = None + for module in cached_module_lst: + if self_scores is None: + self_scores = torch.zeros_like( + module.get_factor(factor_name=SELF_SCORE_VECTOR_NAME), requires_grad=False + ) + self_scores.add_(module.get_factor(factor_name=SELF_SCORE_VECTOR_NAME)) + self_scores = self_scores.cpu() + score_chunks[ALL_MODULE_NAME].append(self_scores) + accumulate_iterations(model=model, tracked_module_names=tracked_module_names) + + if ( + state.use_distributed + and total_steps % DISTRIBUTED_SYNC_INTERVAL == 0 + and index not in [len(train_loader) - 1, len(train_loader) - 2] + ): state.wait_for_everyone() + total_steps += 1 pbar.update(1) - with torch.no_grad(): - model.zero_grad(set_to_none=True) - remove_gradient_scale(model=model) - set_mode( - model=model, - mode=ModuleMode.DEFAULT, - tracked_module_names=tracked_module_names, - keep_factors=False, - ) - total_scores: SCORE_TYPE = {} - for module_name, chunks in score_chunks.items(): - total_scores[module_name] = torch.cat(chunks, dim=0) - if state.use_distributed: - release_memory() - total_scores[module_name] = total_scores[module_name].to(device=state.device) - gather_list = None - if state.is_main_process: - gather_list = [torch.zeros_like(total_scores[module_name]) for _ in range(state.num_processes)] - torch.distributed.gather(total_scores[module_name], gather_list) - if state.is_main_process: - total_scores[module_name] = torch.cat(gather_list, dim=0)[:dataset_size].cpu() + model.zero_grad(set_to_none=True) + if enable_grad_scaler: + set_gradient_scale(model=model, gradient_scale=1.0) + set_mode( + model=model, + mode=ModuleMode.DEFAULT, + tracked_module_names=tracked_module_names, + release_memory=True, + ) + release_memory() + + total_scores: SCORE_TYPE = {} + for module_name, chunks in score_chunks.items(): + total_scores[module_name] = torch.cat(chunks, dim=0) + if state.use_distributed: + total_scores[module_name] = total_scores[module_name].to(device=state.device) + gather_list = None + if state.is_main_process: + gather_list = [torch.zeros_like(total_scores[module_name]) for _ in range(state.num_processes)] + dist.gather(total_scores[module_name], gather_list) + if state.is_main_process: + total_scores[module_name] = torch.cat(gather_list, dim=0)[:dataset_size].cpu() + else: + total_scores[module_name] = total_scores[module_name].cpu() state.wait_for_everyone() + return total_scores diff --git a/kronfluence/task.py b/kronfluence/task.py index 1884e52..1c82dd5 100644 --- a/kronfluence/task.py +++ b/kronfluence/task.py @@ -9,9 +9,15 @@ class Task(ABC): """Abstract base class for task definitions. Extend this class to implement specific tasks (e.g., regression, classification, language modeling) - with custom pipelines (models, data loaders, training objectives). + with custom pipelines (e.g., models, data loaders, training objectives). + + Attributes: + enable_post_process_per_sample_gradient (bool): + Flag to enable post-processing of per-sample gradients. Defaults to `False`. """ + enable_post_process_per_sample_gradient: bool = False + @abstractmethod def compute_train_loss( self, @@ -19,23 +25,23 @@ def compute_train_loss( model: nn.Module, sample: bool = False, ) -> torch.Tensor: - """Computes training loss for a given batch and model. + """Computes the training loss for a given batch and model. Args: batch (Any): - Batch of data sourced from the DataLoader. + A batch of data from the DataLoader. model (nn.Module): - The PyTorch model for loss computation. + The PyTorch model used for loss computation. sample (bool): Indicates whether to sample from the model's outputs or to use the actual targets from the - batch. Defaults to False. The case where `sample` is set to True must be implemented to + batch. Defaults to `False`. The case where `sample=True` must be implemented to approximate the true Fisher. Returns: torch.Tensor: - The computed loss as a tensor. + The computed loss as a scalar tensor. """ - raise NotImplementedError("Subclasses must implement the `compute_train_loss` method.") + raise NotImplementedError(f"{self.__class__.__name__} must implement the `compute_train_loss` method.") @abstractmethod def compute_measurement( @@ -43,49 +49,68 @@ def compute_measurement( batch: Any, model: nn.Module, ) -> torch.Tensor: - """Computes a measurable quantity (e.g., loss, logit, log probability) for a given batch and model. - This is defined as f(θ) from https://arxiv.org/pdf/2308.03296.pdf. + """Computes a measurable quantity for a given batch and model. + + This method calculates f(θ) as defined in https://arxiv.org/pdf/2308.03296.pdf. The measurable quantity + can be a loss, logit, log probability, or any other relevant metric for the task. Args: batch (Any): - Batch of data sourced from the DataLoader. + A batch of data from the DataLoader. model (nn.Module): - The PyTorch model for measurement computation. + The PyTorch model used for measurement computation. Returns: torch.Tensor: - The measurable quantity as a tensor. + The computed measurable quantity as a tensor. """ - raise NotImplementedError("Subclasses must implement the `compute_measurement` method.") + raise NotImplementedError(f"{self.__class__.__name__} must implement the `compute_measurement` method.") - def tracked_modules(self) -> Optional[List[str]]: - """Specifies modules for influence score computations. + def get_influence_tracked_modules(self) -> Optional[List[str]]: + """Specifies which modules should be tracked for influence factor and score computations. - Returns None by default, applying computations to all supported modules (e.g., nn.Linear, nn.Conv2d). - Subclasses can override this method to return a list of specific module names if influence functions + Override this method in subclasses to return a list of specific module names if influence functions should only be computed for a subset of the model. Returns: Optional[List[str]]: - A list of module names for which to compute influence functions, or None to indicate that - influence functions should be computed for all applicable modules. + A list of module names to compute influence functions for, or `None` to compute for + all applicable modules (e.g., `nn.Linear` and `nn.Conv2d`). """ def get_attention_mask(self, batch: Any) -> Optional[Union[Dict[str, torch.Tensor], torch.Tensor]]: - """Returns masks for data points within a batch that have been padded extra tokens to ensure - consistent length across the batch. Typically, it returns None for models or datasets not requiring - masking. + """Gets attention masks for padded sequences in a batch. - See https://huggingface.co/docs/transformers/en/glossary#attention-mask. + This method is typically used for models or datasets that require masking, such as transformer-based + architectures. For more information, see: https://huggingface.co/docs/transformers/en/glossary#attention-mask. Args: batch (Any): - Batch of data sourced from the DataLoader. + A batch of data from the DataLoader. Returns: Optional[Union[Dict[str, torch.Tensor], torch.Tensor]]: - A binary tensor as the mask for the batch, or None if padding is not used. The mask dimensions should - match `batch_size x num_seq`. For models requiring different masks for various modules - (e.g., encoder-decoder architectures), returns a dictionary mapping module names to their - corresponding masks. + - `None` if padding is not used. + - A binary tensor with dimension `batch_size x num_seq` as the mask for the batch. + - A dictionary mapping module names to their corresponding masks for models requiring different + masks for different modules (e.g., encoder-decoder architectures). + """ + + def post_process_per_sample_gradient(self, module_name: str, gradient: torch.Tensor) -> torch.Tensor: + """Post-processes the per-sample gradient of a specific module. + + This method is called only if `do_post_process_per_sample_gradient` is set to `True`. + Override this method in subclasses to implement custom gradient post-processing. + + Args: + module_name (str): + The name of the module whose gradient is being processed. + gradient (torch.Tensor): + The per-sample gradient tensor with dimension `batch_size x gradient_dim x activation_dim`. + + Returns: + torch.Tensor: + The modified per-sample gradient tensor. """ + del module_name + return gradient diff --git a/kronfluence/utils/common/__init__.py b/kronfluence/utils/common/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/kronfluence/utils/common/factor_arguments.py b/kronfluence/utils/common/factor_arguments.py new file mode 100644 index 0000000..26861f5 --- /dev/null +++ b/kronfluence/utils/common/factor_arguments.py @@ -0,0 +1,62 @@ +import torch + +from kronfluence import FactorArguments + + +def default_factor_arguments(strategy: str = "ekfac") -> FactorArguments: + """Creates default factor arguments""" + factor_args = FactorArguments(strategy=strategy) + return factor_args + + +def pytest_factor_arguments(strategy: str = "ekfac") -> FactorArguments: + """Creates factor arguments for unit tests""" + factor_args = FactorArguments(strategy=strategy) + factor_args.use_empirical_fisher = True + factor_args.activation_covariance_dtype = torch.float64 + factor_args.gradient_covariance_dtype = torch.float64 + factor_args.per_sample_gradient_dtype = torch.float64 + factor_args.lambda_dtype = torch.float64 + return factor_args + + +def smart_low_precision_factor_arguments( + strategy: str = "ekfac", dtype: torch.dtype = torch.bfloat16 +) -> FactorArguments: + """Creates factor arguments with low precision, except for Lambda computations.""" + factor_args = FactorArguments(strategy=strategy) + factor_args.amp_dtype = dtype + factor_args.activation_covariance_dtype = dtype + factor_args.gradient_covariance_dtype = dtype + factor_args.per_sample_gradient_dtype = dtype + factor_args.lambda_dtype = torch.float32 + return factor_args + + +def all_low_precision_factor_arguments(strategy: str = "ekfac", dtype: torch.dtype = torch.bfloat16) -> FactorArguments: + """Creates factor arguments with low precision for all computations.""" + factor_args = FactorArguments(strategy=strategy) + factor_args.amp_dtype = dtype + factor_args.activation_covariance_dtype = dtype + factor_args.gradient_covariance_dtype = dtype + factor_args.per_sample_gradient_dtype = dtype + factor_args.lambda_dtype = dtype + return factor_args + + +def reduce_memory_factor_arguments(strategy: str = "ekfac", dtype: torch.dtype = torch.bfloat16) -> FactorArguments: + """Creates factor arguments with low precision and iterative lambda aggregations.""" + factor_args = all_low_precision_factor_arguments(strategy=strategy, dtype=dtype) + factor_args.use_iterative_lambda_aggregation = True + return factor_args + + +def extreme_reduce_memory_factor_arguments( + strategy: str = "ekfac", module_partitions: int = 1, dtype: torch.dtype = torch.bfloat16 +) -> FactorArguments: + """Creates factor arguments for models that are difficult to fit on a single GPU.""" + factor_args = reduce_memory_factor_arguments(strategy=strategy, dtype=dtype) + factor_args.offload_activations_to_cpu = True + factor_args.covariance_module_partitions = module_partitions + factor_args.lambda_module_partitions = module_partitions + return factor_args diff --git a/kronfluence/utils/common/score_arguments.py b/kronfluence/utils/common/score_arguments.py new file mode 100644 index 0000000..40675f6 --- /dev/null +++ b/kronfluence/utils/common/score_arguments.py @@ -0,0 +1,84 @@ +from typing import Optional + +import torch + +from kronfluence import ScoreArguments + + +def default_score_arguments( + damping_factor: Optional[float] = 1e-08, query_gradient_low_rank: Optional[int] = None +) -> ScoreArguments: + """Creates default score arguments""" + score_args = ScoreArguments(damping_factor=damping_factor, query_gradient_low_rank=query_gradient_low_rank) + if score_args.query_gradient_low_rank is not None: + score_args.query_gradient_accumulation_steps = 10 + return score_args + + +def pytest_score_arguments( + damping_factor: Optional[float] = 1e-08, query_gradient_low_rank: Optional[int] = None +) -> ScoreArguments: + """Creates score arguments for unit tests""" + score_args = ScoreArguments(damping_factor=damping_factor, query_gradient_low_rank=query_gradient_low_rank) + score_args.query_gradient_svd_dtype = torch.float64 + score_args.score_dtype = torch.float64 + score_args.per_sample_gradient_dtype = torch.float64 + score_args.precondition_dtype = torch.float64 + return score_args + + +def smart_low_precision_score_arguments( + damping_factor: Optional[float] = 1e-08, + query_gradient_low_rank: Optional[int] = None, + dtype: torch.dtype = torch.bfloat16, +) -> ScoreArguments: + """Creates score arguments with low precision, except for preconditioning computations.""" + score_args = default_score_arguments(damping_factor=damping_factor, query_gradient_low_rank=query_gradient_low_rank) + score_args.amp_dtype = dtype + score_args.score_dtype = dtype + score_args.per_sample_gradient_dtype = dtype + score_args.query_gradient_svd_dtype = torch.float32 + score_args.precondition_dtype = torch.float32 + return score_args + + +def all_low_precision_score_arguments( + damping_factor: Optional[float] = 1e-08, + query_gradient_low_rank: Optional[int] = None, + dtype: torch.dtype = torch.bfloat16, +) -> ScoreArguments: + """Creates score arguments with low precision for all computations.""" + score_args = default_score_arguments(damping_factor=damping_factor, query_gradient_low_rank=query_gradient_low_rank) + score_args.amp_dtype = dtype + score_args.score_dtype = dtype + score_args.per_sample_gradient_dtype = dtype + score_args.precondition_dtype = dtype + score_args.query_gradient_svd_dtype = torch.float32 + return score_args + + +def reduce_memory_score_arguments( + damping_factor: Optional[float] = 1e-08, + query_gradient_low_rank: Optional[int] = None, + dtype: torch.dtype = torch.bfloat16, +) -> ScoreArguments: + """Creates score arguments with low precision and CPU offloading.""" + score_args = all_low_precision_score_arguments( + damping_factor=damping_factor, query_gradient_low_rank=query_gradient_low_rank, dtype=dtype + ) + score_args.offload_activations_to_cpu = True + return score_args + + +def extreme_reduce_memory_score_arguments( + damping_factor: Optional[float] = 1e-08, + module_partitions: int = 4, + query_gradient_low_rank: Optional[int] = None, + dtype: torch.dtype = torch.bfloat16, +) -> ScoreArguments: + """Creates score arguments for models that are difficult to fit on a single GPU.""" + score_args = reduce_memory_score_arguments( + damping_factor=damping_factor, query_gradient_low_rank=query_gradient_low_rank, dtype=dtype + ) + score_args.module_partitions = module_partitions + return score_args diff --git a/kronfluence/utils/constants.py b/kronfluence/utils/constants.py index 74847dc..166a26d 100644 --- a/kronfluence/utils/constants.py +++ b/kronfluence/utils/constants.py @@ -1,25 +1,40 @@ """A collection of constants.""" -from typing import Dict, Tuple +from typing import Dict, List, Optional, Tuple, Union import torch FACTOR_TYPE = Dict[str, Dict[str, torch.Tensor]] PARTITION_TYPE = Tuple[int, int] SCORE_TYPE = Dict[str, torch.Tensor] +PRECONDITIONED_GRADIENT_TYPE = Optional[Union[torch.Tensor, List[torch.Tensor]]] + +# Constants for file naming conventions. +FACTOR_SAVE_PREFIX = "factors_" +SCORE_SAVE_PREFIX = "scores_" +FACTOR_ARGUMENTS_NAME = "factor" +SCORE_ARGUMENTS_NAME = "score" + +# The total iteration step to synchronize the process when using distributed setting. +DISTRIBUTED_SYNC_INTERVAL = 1_000 + +HEURISTIC_DAMPING_SCALE = 0.1 # Activation covariance matrix. ACTIVATION_COVARIANCE_MATRIX_NAME = "activation_covariance" # Pseudo-gradient covariance matrix. GRADIENT_COVARIANCE_MATRIX_NAME = "gradient_covariance" # Number of elements used to aggregate activation and gradient covariance. -NUM_COVARIANCE_PROCESSED = "num_covariance_processed" +NUM_ACTIVATION_COVARIANCE_PROCESSED = "num_activation_covariance_processed" +NUM_GRADIENT_COVARIANCE_PROCESSED = "num_gradient_covariance_processed" + # A list of factors to keep track of when computing covariance matrices. COVARIANCE_FACTOR_NAMES = [ ACTIVATION_COVARIANCE_MATRIX_NAME, GRADIENT_COVARIANCE_MATRIX_NAME, - NUM_COVARIANCE_PROCESSED, + NUM_ACTIVATION_COVARIANCE_PROCESSED, + NUM_GRADIENT_COVARIANCE_PROCESSED, ] @@ -50,11 +65,14 @@ # Preconditioned per-sample gradient. PRECONDITIONED_GRADIENT_NAME = "preconditioned_gradient" -AGGREGATED_PRECONDITIONED_GRADIENT_NAME = "aggregated_preconditioned_gradient" +# Accumulated preconditioned per-sample gradient. +ACCUMULATED_PRECONDITIONED_GRADIENT_NAME = "accumulated_preconditioned_gradient" +# Aggregated gradient. +AGGREGATED_GRADIENT_NAME = "aggregated_gradient" # Pairwise influence scores. PAIRWISE_SCORE_MATRIX_NAME = "pairwise_score_matrix" # Self-influence scores. SELF_SCORE_VECTOR_NAME = "self_score_vector" -# The dictionary key for storing scores for all modules. +# The dictionary key for storing summed scores. ALL_MODULE_NAME = "all_modules" diff --git a/kronfluence/utils/dataset.py b/kronfluence/utils/dataset.py index 9b82744..9f33b22 100644 --- a/kronfluence/utils/dataset.py +++ b/kronfluence/utils/dataset.py @@ -16,8 +16,11 @@ @dataclass class DataLoaderKwargs(KwargsHandler): - """The object used to customize `DataLoader`. Please refer to https://pytorch.org/docs/stable/data.html for - detailed information of each argument. The default arguments are copied from PyTorch version 2.2. + """Customization options for DataLoader. + + This class encapsulates the arguments used to customize PyTorch's DataLoader. Default values are based on + PyTorch version 2.3. For detailed information on each argument, refer to: + https://pytorch.org/docs/stable/data.html. """ num_workers: int = 0 @@ -33,9 +36,23 @@ class DataLoaderKwargs(KwargsHandler): def make_indices_partition(total_data_examples: int, partition_size: int) -> List[Tuple[int, int]]: - """Returns partitioned indices from the total data examples.""" + """Partitions data indices into approximately equal-sized bins. + + Args: + total_data_examples (int): + Total number of data examples. + partition_size (int): + Number of partitions to create. + + Returns: + List[Tuple[int, int]]: + List of tuples, each containing start and end indices for a partition. + + Raises: + ValueError: If `total_data_examples` is less than `partition_size`. + """ if total_data_examples < partition_size: - raise ValueError("The total data examples must be equal or greater than the partition size.") + raise ValueError("The total data examples must be equal to or greater than the partition size.") # See https://stackoverflow.com/questions/2130016/splitting-a-list-into-n-parts-of-approximately-equal-length. bins = list(map(len, np.array_split(range(total_data_examples), partition_size))) start_idx = 0 @@ -47,8 +64,26 @@ def make_indices_partition(total_data_examples: int, partition_size: int) -> Lis def find_executable_batch_size(func: Callable, start_batch_size: int) -> int: - """Finds executable batch size for calling the function that does not encounter OOM error. The code is motivated - from https://github.com/huggingface/accelerate/blob/v0.27.2/src/accelerate/utils/memory.py#L83.""" + """Finds the largest batch size that can be executed without OOM errors. + + This function progressively reduces the batch size until it finds a size that can be executed + without running out of memory. The code is motivated from: + https://github.com/huggingface/accelerate/blob/v0.27.2/src/accelerate/utils/memory.py#L83 + + Args: + func (Callable): + Function to test with different batch sizes. + start_batch_size (int): + Initial batch size to try. + + Returns: + int: + The largest executable batch size. + + Raises: + RuntimeError: + If no executable batch size is found (reaches zero). + """ batch_size = start_batch_size while True: @@ -67,16 +102,15 @@ def find_executable_batch_size(func: Callable, start_batch_size: int) -> int: class DistributedEvalSampler(Sampler[T_co]): - """DistributedEvalSampler is different from `DistributedSampler`: it does not add extra samples to make - the dataset evenly divisible. DistributedEvalSampler should not be used for training; the distributed processes - could hang forever. See this issue for details: https://github.com/pytorch/pytorch/issues/22584. + """Sampler for distributed setting without adding extra samples. + Unlike `DistributedSampler`, it does not add extra samples to make the dataset evenly divisible across processes. The code is adapted from https://github.com/SeungjunNah/DeepDeblur-PyTorch/blob/master/src/data/sampler.py. """ def __init__( # pylint: disable=super-init-not-called self, - dataset: torch.utils.data.Dataset, + dataset: data.Dataset, num_replicas: Optional[int] = None, rank: Optional[int] = None, seed: int = 0, @@ -112,10 +146,10 @@ def __len__(self) -> int: class DistributedSamplerWithStack(Sampler[T_co]): - """DistributedSampleWithStack is different from `DistributedSampler`. Instead of subsampling, - it stacks the dataset. For example, when `num_replicas` is 3, and the dataset of [0, ..., 9] is given, - the first, second, and third rank should have [0, 1, 2], [3, 4, 5], and [6, 7, 8], respectively. However, - it still adds extra samples to make the dataset evenly divisible. + """Sampler that stacks the dataset for distributed setting. + + Instead of subsampling, this sampler stacks the dataset across processes. It ensures even distribution by + adding padding samples if necessary. """ def __init__( # pylint: disable=super-init-not-called diff --git a/kronfluence/utils/exceptions.py b/kronfluence/utils/exceptions.py index 597ca5b..e64f334 100644 --- a/kronfluence/utils/exceptions.py +++ b/kronfluence/utils/exceptions.py @@ -3,12 +3,12 @@ class FactorsNotFoundError(ValueError): class TrackedModuleNotFoundError(ValueError): - """Exception raised when the tracked module is not found.""" + """Exception raised when a tracked module is not found in the model.""" class IllegalTaskConfigurationError(ValueError): - """Exception raised when the provided task is determined to be invalid.""" + """Exception raised when the provided task configuration is determined to be invalid.""" class UnsupportableModuleError(NotImplementedError): - """Exception raised when the provided module is not supported.""" + """Exception raised when the provided module is not supported by the current implementation.""" diff --git a/kronfluence/utils/logger.py b/kronfluence/utils/logger.py index fba4b88..2ca1a0f 100644 --- a/kronfluence/utils/logger.py +++ b/kronfluence/utils/logger.py @@ -20,21 +20,33 @@ class MultiProcessAdapter(logging.LoggerAdapter): - """An adapter to assist with logging in multiprocess. + """An adapter for logging in multiprocess environments. - The code is copied from https://github.com/huggingface/accelerate/blob/main/src/accelerate/logging.py with - minor modifications. + The code is adapted from: https://github.com/huggingface/accelerate/blob/main/src/accelerate/logging.py. """ def log(self, level: int, msg: str, *args, **kwargs) -> None: - """Delegates logger call after checking if it should log.""" + """Log a message if logging is enabled for this process.""" if self.isEnabledFor(level) and not self.extra["disable_log"]: msg, kwargs = self.process(msg, kwargs) self.logger.log(level, msg, *args, **kwargs) def get_logger(name: str, disable_log: bool = False, log_level: int = None) -> MultiProcessAdapter: - """Returns the logger with an option to disable logging.""" + """Creates and returns a logger with optional disabling and log level setting. + + Args: + name (str): + Name of the logger. + disable_log (bool): + Whether to disable logging. Defaults to `False`. + log_level (int): + Logging level to set. Defaults to `None`. + + Returns: + MultiProcessAdapter: + Configured logger adapter. + """ logger = logging.getLogger(name) if log_level is not None: logger.setLevel(log_level) @@ -43,16 +55,15 @@ def get_logger(name: str, disable_log: bool = False, log_level: int = None) -> M class Profiler: - """Profiling object to measure the time taken to run a certain operation. The profiler is helpful - for checking any bottlenecks in the code. + """A profiling utility to measure execution time of operations. - The code is modified from: + The code is adapted from: - https://github.com/Lightning-AI/lightning/tree/master/src/pytorch_lightning/profilers. - https://github.com/mlcommons/algorithmic-efficiency/blob/main/algorithmic_efficiency/profiler.py. """ def __init__(self, state: State) -> None: - """Initializes an instance of the Profiler class. + """Initializes an instance of the `Profiler` class. Args: state (State): @@ -63,7 +74,7 @@ def __init__(self, state: State) -> None: self.recorded_durations = defaultdict(list) def start(self, action_name: str) -> None: - """Defines how to start recording an action.""" + """Start recording an action.""" if not self.state.is_main_process: return if action_name in self.current_actions: @@ -71,7 +82,7 @@ def start(self, action_name: str) -> None: self.current_actions[action_name] = _get_monotonic_time() def stop(self, action_name: str) -> None: - """Defines how to record the duration once an action is complete.""" + """Stop recording an action and log its duration.""" if not self.state.is_main_process: return end_time = _get_monotonic_time() @@ -83,7 +94,7 @@ def stop(self, action_name: str) -> None: @contextmanager def profile(self, action_name: str) -> Generator: - """Yields a context manager to encapsulate the scope of a profiled action.""" + """Context manager for profiling an action.""" try: self.start(action_name) yield action_name @@ -92,6 +103,7 @@ def profile(self, action_name: str) -> Generator: @torch.no_grad() def _make_report(self) -> Tuple[_TABLE_DATA, float, float]: + """Generate a report of profiled actions.""" total_duration = 0.0 for a, d in self.recorded_durations.items(): d_tensor = torch.tensor(d, dtype=torch.float64, requires_grad=False) @@ -110,7 +122,7 @@ def _make_report(self) -> Tuple[_TABLE_DATA, float, float]: return report, total_calls, total_duration def summary(self) -> str: - """Returns a formatted summary for the Profiler.""" + """Generate a formatted summary of the profiling results.""" sep = os.linesep output_string = "Profiler Report:" @@ -143,51 +155,48 @@ def log_row(action: str, mean: str, num_calls: str, total: str, per: str) -> str class PassThroughProfiler(Profiler): - """A pass through Profiler objective that does not record timing for the profiler.""" + """A no-op profiler that doesn't record any timing information.""" def start(self, action_name: str) -> None: - """Defines how to start recording an action.""" return def stop(self, action_name: str) -> None: - """Defines how to record the duration once an action is complete.""" return def summary(self) -> str: - """Returns a formatted summary for the Profiler.""" return "" class TorchProfiler(Profiler): - """A PyTorch Profiler objective that provides detailed profiling information: - https://pytorch.org/tutorials/recipes/recipes/profiler_recipe.html. + """A profiler that utilizes PyTorch's built-in profiling capabilities. + + This profiler provides detailed information about PyTorch operations, including CPU and CUDA events. + It's useful for low-level profiling in PyTorch. + + Note: This is not used by default and is intended for detailed performance analysis. """ def __init__(self, state: State) -> None: - """Initializes an instance of the PyTorch Profiler class.""" super().__init__(state=state) self.actions: list = [] self.trace_outputs: list = [] self._set_up_torch_profiler() def start(self, action_name: str) -> None: - """Defines how to start recording an action.""" if action_name in self.current_actions: raise ValueError(f"Attempted to start {action_name} which has already started.") - # Set dummy value, since only used to track duplicate actions + # Set dummy value, since only used to track duplicate actions. self.current_actions[action_name] = 0.0 self.actions.append(action_name) self._torch_prof.start() def stop(self, action_name: str) -> None: - """Defines how to stop recording an action.""" if action_name not in self.current_actions: raise ValueError(f"Attempting to stop recording an action " f"({action_name}) which was never started.") _ = self.current_actions.pop(action_name) self._torch_prof.stop() def _set_up_torch_profiler(self) -> None: - """Creates the PyTorch profiler object with the necessary arguments.""" self._torch_prof = t_prof.profile( activities=[t_prof.ProfilerActivity.CPU, t_prof.ProfilerActivity.CUDA], record_shapes=True, @@ -198,7 +207,6 @@ def _set_up_torch_profiler(self) -> None: ) def _trace_handler(self, p) -> None: - """Adds the PyTorch Profiler trace output to a list once it is ready.""" # Set metric to sort based on device. is_cpu = self.state.device == torch.device("cpu") sort_by_metric = "self_cpu_time_total" if is_cpu else "self_cuda_time_total" @@ -216,12 +224,10 @@ def _trace_handler(self, p) -> None: self.recorded_durations[self.actions[-1]].append(total_time) def _reset_output(self) -> None: - """Resets actions and outputs list.""" self.actions = [] self.trace_outputs = [] def _high_level_summary(self) -> str: - """Returns a formatted high level summary for the PyTorch Profiler.""" sep = os.linesep output_string = "Overall PyTorch Profiler Report:" @@ -253,7 +259,6 @@ def log_row(action: str, mean: str, num_calls: str, total: str, per: str) -> str return output_string def summary(self) -> str: - """Returns a formatted summary for the PyTorch Profiler.""" assert len(self.actions) == len(self.trace_outputs), ( "Mismatch in the number of actions and outputs collected: " + f"# Actions: {len(self.actions)}, # Ouptuts: {len(self.trace_outputs)}" @@ -270,10 +275,15 @@ def summary(self) -> str: return summary -# Timing utilities copied from +# Timing utilities copied from: # https://github.com/mlcommons/algorithmic-efficiency/blob/main/algorithmic_efficiency/pytorch_utils.py. def _get_monotonic_time() -> float: - """Gets the monotonic time after the CUDA synchronization if necessary.""" + """Gets the time after the CUDA synchronization. + + Returns: + float: + The current time. + """ if torch.cuda.is_available() and torch.cuda.is_initialized(): torch.cuda.synchronize() return time.monotonic() @@ -281,7 +291,16 @@ def _get_monotonic_time() -> float: @torch.no_grad() def get_time(state: State) -> float: - """Gets the current time after synchronizing with other devices.""" + """Gets the current time after synchronizing with other devices. + + Args: + state (State): + The current process's information (e.g., device being used). + + Returns: + float: + The current time. + """ if not state.use_distributed: if torch.cuda.is_available() and torch.cuda.is_initialized(): torch.cuda.synchronize() diff --git a/kronfluence/utils/model.py b/kronfluence/utils/model.py index 2731690..e58bd33 100644 --- a/kronfluence/utils/model.py +++ b/kronfluence/utils/model.py @@ -15,22 +15,30 @@ def apply_ddp( - model: torch.nn.Module, + model: nn.Module, local_rank: int, rank: int, world_size: int, ) -> DistributedDataParallel: - """ - Applies DistributedDataParallel (DDP) to the given model. + """Applies DistributedDataParallel (DDP) to the given PyTorch model. Args: - model (torch.nn.Module): The model to apply DDP to. - local_rank (int): The local rank of the current process. - rank (int): The rank of the current process. - world_size (int): The total number of processes. + model (nn.Module): + The PyTorch model to be parallelized. + local_rank (int): + The local rank of the current process within its node. + rank (int): + The global rank of the current process across all nodes. + world_size (int): + The total number of processes in the distributed setup. Returns: - DistributedDataParallel: The model wrapped with DDP. + DistributedDataParallel: + The input model wrapped with DDP. + + Raises: + RuntimeError: + If the distributed initialization fails. """ dist.init_process_group("nccl", rank=rank, world_size=world_size) device = torch.device(f"cuda:{local_rank}") @@ -48,7 +56,7 @@ def apply_ddp( def apply_fsdp( - model: torch.nn.Module, + model: nn.Module, local_rank: int, rank: int, world_size: int, @@ -57,27 +65,35 @@ def apply_fsdp( is_transformer: bool = False, layer_to_wrap: Optional[nn.Module] = None, ) -> FSDP: - """ - Applies FullyShardedDataParallel (FSDP) to the given model. + """Applies FullyShardedDataParallel (FSDP) to the given PyTorch model. Args: - model (torch.nn.Module): The model to apply FSDP to. - local_rank (int): The local rank of the current process. - rank (int): The rank of the current process. - world_size (int): The total number of processes. - sharding_strategy (str): The sharding strategy to use. - Defaults to "FULL_SHARD". - cpu_offload (bool): Whether to offload parameters to CPU. Check - https://pytorch.org/docs/2.2/fsdp.html#torch.distributed.fsdp.CPUOffload. - Defaults to True. - is_transformer (bool): Whether the model is a transformer model. - Defaults to False. - layer_to_wrap (nn.Module, optional): The specific layer to wrap - for transformer models. Required if `is_transformer` is True. - Defaults to None. + model (nn.Module): + The PyTorch model to be parallelized. + local_rank (int): + The local rank of the current process within its node. + rank (int): + The global rank of the current process across all nodes. + world_size (int): + The total number of processes in the distributed setup. + sharding_strategy (str): + The FSDP sharding strategy to use. Defaults to "FULL_SHARD". + cpu_offload (bool): + Whether to offload parameters to CPU. Defaults to `True`. + is_transformer (bool): + Whether the model is a transformer. Defaults to `False`. + layer_to_wrap (nn.Module, optional): + The specific layer to wrap for transformer models. Required if `is_transformer` is `True`. Returns: - FSDP: The model wrapped with FSDP. + FullyShardedDataParallel: + The input model wrapped with FSDP. + + Raises: + ValueError: + If an invalid sharding strategy is provided or if `layer_to_wrap` is not provided for transformer models. + RuntimeError: + If the distributed initialization fails. """ dist.init_process_group("nccl", rank=rank, world_size=world_size) device = torch.device(f"cuda:{local_rank}") diff --git a/kronfluence/utils/save.py b/kronfluence/utils/save.py index bb0566a..f129551 100644 --- a/kronfluence/utils/save.py +++ b/kronfluence/utils/save.py @@ -5,38 +5,87 @@ import torch from safetensors import safe_open -FACTOR_SAVE_PREFIX = "factors_" -SCORE_SAVE_PREFIX = "scores_" -FACTOR_ARGUMENTS_NAME = "factor" -SCORE_ARGUMENTS_NAME = "score" +def load_file(path: Path) -> Dict[str, torch.Tensor]: + """Loads a dictionary of tensors from a file using `safetensors`. + Args: + path (Path): + The path to the file containing tensor data. -def load_file(path: Path) -> Dict[str, torch.Tensor]: - """Loads a dictionary of tensors from the path.""" - load_dict = {} - with safe_open(path, framework="pt", device="cpu") as f: - for key in f.keys(): - load_dict[key] = f.get_tensor(name=key) - return load_dict + Returns: + Dict[str, torch.Tensor]: + A dictionary where keys are tensor names and values are the corresponding tensors. + """ + if not path.exists(): + raise FileNotFoundError(f"File not found: {path}.") + try: + with safe_open(path, framework="pt", device="cpu") as f: + return {key: f.get_tensor(key) for key in f.keys()} + except Exception as e: + raise RuntimeError(f"Error loading file {path}: {str(e)}") from e def save_json(obj: Any, path: Path) -> None: - """Saves the object to a JSON file.""" - with open(path, "w", encoding="utf-8") as f: - json.dump(obj, f, indent=4) + """Saves an object to a JSON file. + + This function serializes the given object to JSON format and writes it to a file. + + Args: + obj (Any): + The object to be saved. Must be JSON-serializable. + path (Path): + The path where the JSON file will be saved. + """ + path.parent.mkdir(parents=True, exist_ok=True) + try: + with open(path, "w", encoding="utf-8") as f: + json.dump(obj, f, indent=4, ensure_ascii=False) + except TypeError as e: + raise TypeError(f"Object is not JSON-serializable: {str(e)}") from e + except Exception as e: + raise IOError(f"Error saving JSON file {path}: {str(e)}") from e def load_json(path: Path) -> Dict[str, Any]: - """Loads an object from the JSON file.""" - with open(path, "rb") as f: - obj = json.load(f) - return obj + """Loads an object from a JSON file. + + Args: + path (Path): + The path to the JSON file to be loaded. + + Returns: + Dict[str, Any]: + The object loaded from the JSON file. + """ + if not path.exists(): + raise FileNotFoundError(f"File not found: {path}.") + with open(path, "r", encoding="utf-8") as f: + return json.load(f) -@torch.no_grad() def verify_models_equivalence(state_dict1: Dict[str, torch.Tensor], state_dict2: Dict[str, torch.Tensor]) -> bool: - """Checks if two models are equivalent given their `state_dict`.""" + """Check if two models are equivalent given their state dictionaries. + + This function compares two model state dictionaries to determine if they represent + equivalent models. It checks for equality in the number of parameters, parameter names, + and parameter values (within a small tolerance). + + Args: + state_dict1 (Dict[str, torch.Tensor]): + The state dictionary of the first model. + state_dict2 (Dict[str, torch.Tensor]): + The state dictionary of the second model. + + Returns: + bool: + `True` if the models are equivalent, `False` otherwise. + + Notes: + - The function uses a relative tolerance of 1.3e-6 and an absolute tolerance of 1e-5 + when comparing tensor values. + - Tensors are compared in float32 precision on the CPU to ensure consistency. + """ if len(state_dict1) != len(state_dict2): return False @@ -46,7 +95,7 @@ def verify_models_equivalence(state_dict1: Dict[str, torch.Tensor], state_dict2: for name in state_dict1: tensor1 = state_dict1[name].to(dtype=torch.float32).cpu() tensor2 = state_dict2[name].to(dtype=torch.float32).cpu() - if not torch.allclose(tensor1, tensor2, rtol=1e-3, atol=1e-5): + if not torch.allclose(tensor1, tensor2, rtol=1.3e-6, atol=1e-5): return False return True diff --git a/kronfluence/utils/state.py b/kronfluence/utils/state.py index 08705dd..73d5d1e 100644 --- a/kronfluence/utils/state.py +++ b/kronfluence/utils/state.py @@ -1,13 +1,12 @@ import contextlib import gc import os -from typing import Any, Callable, Dict +from typing import Any, Callable, Dict, List import torch import torch.distributed as dist from accelerate.state import SharedDict from torch import nn -from torch.distributed.fsdp import FullyShardedDataParallel class State: @@ -23,12 +22,11 @@ class State: _shared_state: Dict[str, Any] = SharedDict() def __init__(self, cpu: bool = False) -> None: - """Initializes an instance of the State class. + """Initializes an instance of the `State` class. Args: cpu (bool): - Specifies whether the analysis should be explicitly performed using the CPU. - Defaults to False, utilizing GPU resources if available. + If `True`, forces the use of CPU even if GPUs are available. Defaults to `False`. """ self.__dict__ = self._shared_state @@ -51,6 +49,12 @@ def __init__(self, cpu: bool = False) -> None: self.device = torch.device("cpu") if self.cpu else self.default_device def __repr__(self) -> str: + """Provides a string representation of the `State` instance. + + Returns: + str: + A formatted string containing process and device information. + """ return ( f"Num processes: {self.num_processes}\n" f"Process index: {self.process_index}\n" @@ -60,64 +64,101 @@ def __repr__(self) -> str: @staticmethod def _reset_state() -> None: - """Resets `_shared_state`, is used internally and should not be called.""" + """Resets the shared state. For internal use only.""" State._shared_state.clear() @property def initialized(self) -> bool: - """Returns whether the `PartialState` has been initialized.""" + """Checks if the `State` has been initialized.""" return self._shared_state != {} @property def use_distributed(self) -> bool: - """Whether the State is configured for distributed training.""" + """Checks if the setup is configured for distributed setting.""" return self.num_processes > 1 @property def is_main_process(self) -> bool: - """Returns whether the current process is the main process.""" + """Checks if the current process is the main process.""" return self.process_index == 0 @property def is_local_main_process(self) -> bool: - """Returns whether the current process is the main process on the local node.""" + """Checks if the current process is the main process on the local node.""" return self.local_process_index == 0 @property def is_last_process(self) -> bool: - """Returns whether the current process is the last one.""" + """Checks if the current process is the last one.""" return self.process_index == self.num_processes - 1 def wait_for_everyone(self) -> None: - """Will stop the execution of the current process until every other process has reached that point - (so this does nothing when the script is only run in one process).""" + """Synchronizes all processes. + + This method will pause the execution of the current process until all other processes + reach this point. It has no effect in single-process execution. + """ if self.use_distributed: dist.barrier() @property def default_device(self) -> torch.device: - """Finds the default device currently available.""" + """Determines the default device (CUDA if available, otherwise CPU). + + Returns: + torch.device: + The default device. + """ if torch.cuda.is_available(): return torch.device("cuda") return torch.device("cpu") def release_memory() -> None: - """Releases the memory by calling `gc.collect()` and `torch.cuda.empty_cache()`.""" + """Releases unused memory. + + This function calls Python's garbage collector and empties CUDA cache if CUDA is available. + """ gc.collect() + torch.compiler.reset() if torch.cuda.is_available(): torch.cuda.empty_cache() +def get_active_tensors() -> List[torch.Tensor]: + """Gets a list of active tensors in memory. + + Returns: + List[torch.Tensor]: + A list of tuples containing tensor type and size. + """ + tensor_lst = [] + for obj in gc.get_objects(): + if torch.is_tensor(obj) or (hasattr(obj, "data") and torch.is_tensor(obj.data)): + tensor_lst.append(type(obj), obj.size()) + return tensor_lst + + @contextlib.contextmanager def no_sync(model: nn.Module, state: State) -> Callable: - """A context manager to avoid DDP synchronization. The code is adapted from - https://github.com/huggingface/accelerate/blob/v0.27.2/src/accelerate/accelerator.py#L852.""" + """A context manager to temporarily disable gradient synchronization in distributed setting. + + Args: + model (nn.Module): + The PyTorch model. + state (State): + The current process state. + + Yields: + A context where gradient synchronization is disabled (if applicable). + + Note: + For FullyShardedDataParallel (FSDP) models, this may result in higher memory usage. + See: https://pytorch.org/docs/stable/fsdp.html. + """ context = contextlib.nullcontext - # `no_sync()` for FSDP instance can result in higher memory usage, detailed in: - # https://pytorch.org/docs/stable/fsdp.html. - if state.use_distributed and not isinstance(model, FullyShardedDataParallel): + if state.use_distributed: context = getattr(model, "no_sync", context) with context(): diff --git a/kronfluence/version.py b/kronfluence/version.py index 3dc1f76..5becc17 100644 --- a/kronfluence/version.py +++ b/kronfluence/version.py @@ -1 +1 @@ -__version__ = "0.1.0" +__version__ = "1.0.0" diff --git a/pyproject.toml b/pyproject.toml index a8ddb26..870cc40 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -33,4 +33,6 @@ disable = """ use-implicit-booleaness-not-comparison, implicit-str-concat, inconsistent-return-statements, + too-many-lines, + too-many-public-methods, """ \ No newline at end of file diff --git a/requirements.txt b/requirements.txt index 353244f..544e947 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,8 +1,8 @@ torch>=2.1.0 torchvision>=0.16.0 -accelerate>=0.27.2 -einops>=0.7.0 +accelerate>=0.31.0 +einops>=0.8.0 einconv>=0.1.0 opt_einsum>=3.3.0 safetensors>=0.4.2 -tqdm>=4.66.2 \ No newline at end of file +tqdm>=4.66.4 \ No newline at end of file diff --git a/setup.py b/setup.py index e7d8c73..903cc3d 100644 --- a/setup.py +++ b/setup.py @@ -37,14 +37,11 @@ }, packages=find_packages(), keywords=[ - "PyTorch", "Training Data Attribution", "Influence Functions", - "KFAC", - "EKFAC", ], classifiers=[ - "Development Status :: 3 - Alpha", + "Development Status :: 4 - Beta", "Intended Audience :: Developers", "Intended Audience :: Science/Research", "License :: OSI Approved :: Apache Software License", diff --git a/tests/factors/test_covariances.py b/tests/factors/test_covariances.py index 79f3ecd..8603f0e 100644 --- a/tests/factors/test_covariances.py +++ b/tests/factors/test_covariances.py @@ -3,18 +3,24 @@ import pytest import torch -from kronfluence.arguments import FactorArguments +from kronfluence.utils.common.factor_arguments import ( + default_factor_arguments, + pytest_factor_arguments, +) from kronfluence.utils.constants import ( ACTIVATION_COVARIANCE_MATRIX_NAME, COVARIANCE_FACTOR_NAMES, GRADIENT_COVARIANCE_MATRIX_NAME, - NUM_COVARIANCE_PROCESSED, + NUM_ACTIVATION_COVARIANCE_PROCESSED, + NUM_GRADIENT_COVARIANCE_PROCESSED, ) from kronfluence.utils.dataset import DataLoaderKwargs from tests.utils import ( ATOL, + DEFAULT_FACTORS_NAME, RTOL, check_tensor_dict_equivalence, + custom_factors_name, prepare_model_and_analyzer, prepare_test, ) @@ -26,13 +32,14 @@ "mlp", "repeated_mlp", "conv", - "conv_bn", "bert", + "roberta", "gpt", + "gpt_checkpoint", ], ) -@pytest.mark.parametrize("activation_covariance_dtype", [torch.float32, torch.bfloat16]) -@pytest.mark.parametrize("gradient_covariance_dtype", [torch.float32, torch.bfloat16]) +@pytest.mark.parametrize("activation_covariance_dtype", [torch.float32, torch.float16]) +@pytest.mark.parametrize("gradient_covariance_dtype", [torch.float32, torch.float16]) @pytest.mark.parametrize("train_size", [16]) @pytest.mark.parametrize("seed", [0]) def test_fit_covariance_matrices( @@ -42,7 +49,7 @@ def test_fit_covariance_matrices( train_size: int, seed: int, ) -> None: - # Make sure that the covariance computations are working properly. + # Makes sure that the covariance computations are working properly. model, train_dataset, _, data_collator, task = prepare_test( test_name=test_name, train_size=train_size, @@ -53,25 +60,23 @@ def test_fit_covariance_matrices( model=model, task=task, ) - - factor_args = FactorArguments( - activation_covariance_dtype=activation_covariance_dtype, - gradient_covariance_dtype=gradient_covariance_dtype, - ) - factors_name = f"pytest_{test_name}_{test_fit_covariance_matrices.__name__}" + factor_args = default_factor_arguments() + factor_args.activation_covariance_dtype = activation_covariance_dtype + factor_args.gradient_covariance_dtype = gradient_covariance_dtype analyzer.fit_covariance_matrices( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, + factor_args=factor_args, dataset=train_dataset, dataloader_kwargs=kwargs, - factor_args=factor_args, per_device_batch_size=train_size // 4, overwrite_output_dir=True, ) covariance_factors = analyzer.load_covariance_matrices( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, ) assert set(covariance_factors.keys()) == set(COVARIANCE_FACTOR_NAMES) assert len(covariance_factors[ACTIVATION_COVARIANCE_MATRIX_NAME]) > 0 + assert len(covariance_factors[GRADIENT_COVARIANCE_MATRIX_NAME]) > 0 for module_name in covariance_factors[ACTIVATION_COVARIANCE_MATRIX_NAME]: assert covariance_factors[ACTIVATION_COVARIANCE_MATRIX_NAME][module_name].dtype == activation_covariance_dtype assert covariance_factors[GRADIENT_COVARIANCE_MATRIX_NAME][module_name].dtype == gradient_covariance_dtype @@ -99,39 +104,32 @@ def test_covariance_matrices_batch_size_equivalence( seed=seed, ) kwargs = DataLoaderKwargs(collate_fn=data_collator) + model = model.to(dtype=torch.float64) model, analyzer = prepare_model_and_analyzer( model=model, task=task, ) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - ) + factor_args = pytest_factor_arguments() analyzer.fit_covariance_matrices( - factors_name=f"pytest_{test_name}_{test_covariance_matrices_batch_size_equivalence.__name__}_bs1", + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, factor_args=factor_args, per_device_batch_size=1, overwrite_output_dir=True, dataloader_kwargs=kwargs, ) - bs1_covariance_factors = analyzer.load_covariance_matrices( - factors_name=f"pytest_{test_name}_{test_covariance_matrices_batch_size_equivalence.__name__}_bs1" - ) + bs1_covariance_factors = analyzer.load_covariance_matrices(factors_name=DEFAULT_FACTORS_NAME) analyzer.fit_covariance_matrices( - factors_name=f"pytest_{test_name}_{test_covariance_matrices_batch_size_equivalence.__name__}_bs8", + factors_name=custom_factors_name(name="bs8"), dataset=train_dataset, factor_args=factor_args, per_device_batch_size=8, overwrite_output_dir=True, dataloader_kwargs=kwargs, ) - bs8_covariance_factors = analyzer.load_covariance_matrices( - factors_name=f"pytest_{test_name}_{test_covariance_matrices_batch_size_equivalence.__name__}_bs8" - ) + bs8_covariance_factors = analyzer.load_covariance_matrices(factors_name=custom_factors_name(name="bs8")) for name in COVARIANCE_FACTOR_NAMES: assert check_tensor_dict_equivalence( @@ -146,18 +144,18 @@ def test_covariance_matrices_batch_size_equivalence( "test_name", [ "mlp", - "conv", - "gpt", + "conv_bn", + "bert", ], ) -@pytest.mark.parametrize("data_partition_size", [1, 4]) -@pytest.mark.parametrize("module_partition_size", [1, 3]) +@pytest.mark.parametrize("data_partitions", [2, 4]) +@pytest.mark.parametrize("module_partitions", [2, 3]) @pytest.mark.parametrize("train_size", [62]) @pytest.mark.parametrize("seed", [2]) def test_covariance_matrices_partition_equivalence( test_name: str, - data_partition_size: int, - module_partition_size: int, + data_partitions: int, + module_partitions: int, train_size: int, seed: int, ) -> None: @@ -168,36 +166,27 @@ def test_covariance_matrices_partition_equivalence( seed=seed, ) kwargs = DataLoaderKwargs(collate_fn=data_collator) + model = model.to(dtype=torch.float64) model, analyzer = prepare_model_and_analyzer( model=model, task=task, ) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - ) - factors_name = f"pytest_{test_name}_{test_covariance_matrices_partition_equivalence.__name__}" + factor_args = pytest_factor_arguments() analyzer.fit_covariance_matrices( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, factor_args=factor_args, per_device_batch_size=8, overwrite_output_dir=True, dataloader_kwargs=kwargs, ) - covariance_factors = analyzer.load_covariance_matrices(factors_name=factors_name) + covariance_factors = analyzer.load_covariance_matrices(factors_name=DEFAULT_FACTORS_NAME) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - covariance_data_partition_size=data_partition_size, - covariance_module_partition_size=module_partition_size, - ) + factor_args.covariance_data_partitions = data_partitions + factor_args.covariance_module_partitions = module_partitions analyzer.fit_covariance_matrices( - factors_name=f"pytest_{test_name}_partitioned_{data_partition_size}_{module_partition_size}", + factors_name=custom_factors_name(f"{data_partitions}_{module_partitions}"), dataset=train_dataset, factor_args=factor_args, per_device_batch_size=7, @@ -205,7 +194,7 @@ def test_covariance_matrices_partition_equivalence( dataloader_kwargs=kwargs, ) partitioned_covariance_factors = analyzer.load_covariance_matrices( - factors_name=f"pytest_{test_name}_partitioned_{data_partition_size}_{module_partition_size}", + factors_name=custom_factors_name(f"{data_partitions}_{module_partitions}"), ) for name in COVARIANCE_FACTOR_NAMES: @@ -217,7 +206,7 @@ def test_covariance_matrices_partition_equivalence( ) -@pytest.mark.parametrize("test_name", ["bert", "wrong_bert"]) +@pytest.mark.parametrize("test_name", ["bert", "wrong_bert", "roberta"]) @pytest.mark.parametrize("train_size", [213]) @pytest.mark.parametrize("seed", [3]) def test_covariance_matrices_attention_mask( @@ -225,7 +214,7 @@ def test_covariance_matrices_attention_mask( train_size: int, seed: int, ) -> None: - # Make sure the attention mask is correctly implemented by comparing with the results + # Makes sure the attention mask is correctly implemented by comparing with the results # without any padding applied (and batch size of 1). model, train_dataset, _, data_collator, task = prepare_test( test_name=test_name, @@ -245,21 +234,15 @@ def test_covariance_matrices_attention_mask( seed=seed, ) model = model.to(dtype=torch.float64) - - kwargs = DataLoaderKwargs(collate_fn=data_collator) model, analyzer = prepare_model_and_analyzer( model=model, task=task, ) + kwargs = DataLoaderKwargs(collate_fn=data_collator) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - ) - factors_name = f"pytest_{test_name}_{test_covariance_matrices_attention_mask.__name__}" + factor_args = pytest_factor_arguments() analyzer.fit_covariance_matrices( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, factor_args=factor_args, per_device_batch_size=train_size // 4, @@ -267,11 +250,11 @@ def test_covariance_matrices_attention_mask( dataloader_kwargs=kwargs, ) covariance_factors = analyzer.load_covariance_matrices( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, ) analyzer.fit_covariance_matrices( - factors_name=factors_name + "_no_pad", + factors_name=custom_factors_name("no_pad"), dataset=no_padded_train_dataset, factor_args=factor_args, per_device_batch_size=1, @@ -279,11 +262,15 @@ def test_covariance_matrices_attention_mask( dataloader_kwargs=kwargs, ) no_padded_covariance_factors = analyzer.load_covariance_matrices( - factors_name=factors_name + "_no_pad", + factors_name=custom_factors_name("no_pad"), ) for name in COVARIANCE_FACTOR_NAMES: - if "wrong" in test_name and name in [ACTIVATION_COVARIANCE_MATRIX_NAME, NUM_COVARIANCE_PROCESSED]: + if "wrong" in test_name and name in [ + ACTIVATION_COVARIANCE_MATRIX_NAME, + NUM_ACTIVATION_COVARIANCE_PROCESSED, + NUM_GRADIENT_COVARIANCE_PROCESSED, + ]: assert not check_tensor_dict_equivalence( covariance_factors[name], no_padded_covariance_factors[name], @@ -307,36 +294,32 @@ def test_covariance_matrices_automatic_batch_size( train_size: int, seed: int, ) -> None: - # Make sure the automatic batch size search feature is working properly. + # Makes sure the automatic batch size search feature is working properly. model, train_dataset, _, data_collator, task = prepare_test( test_name=test_name, train_size=train_size, seed=seed, ) kwargs = DataLoaderKwargs(collate_fn=data_collator) + model = model.to(dtype=torch.float64) model, analyzer = prepare_model_and_analyzer( model=model, task=task, ) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - ) - factors_name = f"pytest_{test_name}_{test_covariance_matrices_automatic_batch_size.__name__}" + factor_args = pytest_factor_arguments() analyzer.fit_covariance_matrices( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, factor_args=factor_args, per_device_batch_size=8, overwrite_output_dir=True, dataloader_kwargs=kwargs, ) - covariance_factors = analyzer.load_covariance_matrices(factors_name=factors_name) + covariance_factors = analyzer.load_covariance_matrices(factors_name=DEFAULT_FACTORS_NAME) analyzer.fit_covariance_matrices( - factors_name=factors_name + "_auto", + factors_name=custom_factors_name("auto"), dataset=train_dataset, factor_args=factor_args, per_device_batch_size=None, @@ -344,7 +327,7 @@ def test_covariance_matrices_automatic_batch_size( dataloader_kwargs=kwargs, ) auto_covariance_factors = analyzer.load_covariance_matrices( - factors_name=factors_name + "_auto", + factors_name=custom_factors_name("auto"), ) for name in COVARIANCE_FACTOR_NAMES: @@ -357,16 +340,20 @@ def test_covariance_matrices_automatic_batch_size( @pytest.mark.parametrize("test_name", ["mlp"]) -@pytest.mark.parametrize("data_partition_size", [1, 4]) +@pytest.mark.parametrize("max_examples", [4, 26]) +@pytest.mark.parametrize("data_partitions", [1, 4]) +@pytest.mark.parametrize("module_partitions", [1, 2]) @pytest.mark.parametrize("train_size", [80]) @pytest.mark.parametrize("seed", [5]) def test_covariance_matrices_max_examples( test_name: str, - data_partition_size: int, + max_examples: int, + data_partitions: int, + module_partitions: int, train_size: int, seed: int, ) -> None: - # Make sure the max covariance data selection is working properly. + # Makes sure the max covariance data selection is working properly. model, train_dataset, _, data_collator, task = prepare_test( test_name=test_name, train_size=train_size, @@ -378,87 +365,147 @@ def test_covariance_matrices_max_examples( task=task, ) - MAX_EXAMPLES = 26 - factor_args = FactorArguments( - use_empirical_fisher=True, - covariance_max_examples=MAX_EXAMPLES, - covariance_data_partition_size=data_partition_size, - ) - factors_name = f"pytest_{test_name}_{test_covariance_matrices_max_examples.__name__}" + factor_args = pytest_factor_arguments() + factor_args.covariance_max_examples = max_examples + factor_args.covariance_data_partitions = data_partitions + factor_args.covariance_module_partitions = module_partitions + analyzer.fit_covariance_matrices( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, factor_args=factor_args, per_device_batch_size=32, overwrite_output_dir=True, dataloader_kwargs=kwargs, ) - covariance_factors = analyzer.load_covariance_matrices(factors_name=factors_name) + covariance_factors = analyzer.load_covariance_matrices(factors_name=DEFAULT_FACTORS_NAME) - for num_examples in covariance_factors[NUM_COVARIANCE_PROCESSED].values(): - assert num_examples == MAX_EXAMPLES + for num_examples in covariance_factors[NUM_ACTIVATION_COVARIANCE_PROCESSED].values(): + assert num_examples == max_examples + for num_examples in covariance_factors[NUM_GRADIENT_COVARIANCE_PROCESSED].values(): + assert num_examples == max_examples -@pytest.mark.parametrize( - "test_name", - [ - "mlp", - "conv", - ], -) + +@pytest.mark.parametrize("test_name", ["mlp", "gpt"]) @pytest.mark.parametrize("train_size", [100]) @pytest.mark.parametrize("seed", [6]) -def test_covariance_matrices_amp( +def test_covariance_matrices_gradient_checkpoint( test_name: str, train_size: int, seed: int, ) -> None: - # Covariance matrices should be similar when AMP is enabled. + # Covariance matrices should be the same even when gradient checkpointing is used. model, train_dataset, _, data_collator, task = prepare_test( test_name=test_name, train_size=train_size, seed=seed, ) kwargs = DataLoaderKwargs(collate_fn=data_collator) + + model = model.to(dtype=torch.float64) model, analyzer = prepare_model_and_analyzer( model=model, task=task, ) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - amp_dtype=torch.float16, - ) + factor_args = pytest_factor_arguments() analyzer.fit_covariance_matrices( - factors_name=f"pytest_{test_name}_{test_covariance_matrices_amp.__name__}", + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, - factor_args=factor_args, per_device_batch_size=8, overwrite_output_dir=True, dataloader_kwargs=kwargs, + factor_args=factor_args, ) covariance_factors = analyzer.load_covariance_matrices( - factors_name=f"pytest_{test_name}_{test_covariance_matrices_amp.__name__}" + factors_name=DEFAULT_FACTORS_NAME, ) + model, _, _, _, task = prepare_test( + test_name=test_name + "_checkpoint", + train_size=train_size, + seed=seed, + ) + model = model.to(dtype=torch.float64) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) analyzer.fit_covariance_matrices( - factors_name=f"pytest_{test_name}_{test_covariance_matrices_amp.__name__}_amp", + factors_name=custom_factors_name("cp"), dataset=train_dataset, + per_device_batch_size=4, + dataloader_kwargs=kwargs, + overwrite_output_dir=True, factor_args=factor_args, + ) + checkpoint_covariance_factors = analyzer.load_covariance_matrices( + factors_name=custom_factors_name("cp"), + ) + + assert check_tensor_dict_equivalence( + covariance_factors[GRADIENT_COVARIANCE_MATRIX_NAME], + checkpoint_covariance_factors[GRADIENT_COVARIANCE_MATRIX_NAME], + atol=ATOL, + rtol=RTOL, + ) + + +@pytest.mark.parametrize("train_size", [100]) +@pytest.mark.parametrize("seed", [7, 8]) +def test_covariance_matrices_inplace( + train_size: int, + seed: int, +) -> None: + # Covariance matrices should be the identical for with and without in-place ReLU. + model, train_dataset, _, data_collator, task = prepare_test( + test_name="conv", + train_size=train_size, + seed=seed, + ) + model = model.to(dtype=torch.float64) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + + factor_args = pytest_factor_arguments() + analyzer.fit_covariance_matrices( + factors_name=DEFAULT_FACTORS_NAME, + dataset=train_dataset, per_device_batch_size=8, overwrite_output_dir=True, - dataloader_kwargs=kwargs, + factor_args=factor_args, ) - amp_covariance_factors = analyzer.load_covariance_matrices( - factors_name=f"pytest_{test_name}_{test_covariance_matrices_amp.__name__}_amp" + covariance_factors = analyzer.load_covariance_matrices( + factors_name=DEFAULT_FACTORS_NAME, ) - for name in COVARIANCE_FACTOR_NAMES: - assert check_tensor_dict_equivalence( - covariance_factors[name], - amp_covariance_factors[name], - atol=ATOL, - rtol=RTOL, - ) + model, _, _, _, task = prepare_test( + test_name="conv_inplace", + train_size=train_size, + seed=seed, + ) + model = model.to(dtype=torch.float64) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + analyzer.fit_covariance_matrices( + factors_name=custom_factors_name("inplace"), + dataset=train_dataset, + per_device_batch_size=4, + overwrite_output_dir=True, + factor_args=factor_args, + ) + inplace_covariance_factors = analyzer.load_covariance_matrices( + factors_name=custom_factors_name("inplace"), + ) + + assert check_tensor_dict_equivalence( + covariance_factors[GRADIENT_COVARIANCE_MATRIX_NAME], + inplace_covariance_factors[GRADIENT_COVARIANCE_MATRIX_NAME], + atol=ATOL, + rtol=RTOL, + ) diff --git a/tests/factors/test_eigendecompositions.py b/tests/factors/test_eigendecompositions.py new file mode 100644 index 0000000..3198754 --- /dev/null +++ b/tests/factors/test_eigendecompositions.py @@ -0,0 +1,66 @@ +# pylint: skip-file + +import pytest +import torch + +from kronfluence.arguments import FactorArguments +from kronfluence.utils.constants import ( + ACTIVATION_EIGENVECTORS_NAME, + EIGENDECOMPOSITION_FACTOR_NAMES, + GRADIENT_EIGENVECTORS_NAME, +) +from kronfluence.utils.dataset import DataLoaderKwargs +from tests.utils import DEFAULT_FACTORS_NAME, prepare_model_and_analyzer, prepare_test + + +@pytest.mark.parametrize( + "test_name", + [ + "mlp", + "conv", + "bert", + ], +) +@pytest.mark.parametrize("eigendecomposition_dtype", [torch.float32, torch.float64]) +@pytest.mark.parametrize("train_size", [1, 30]) +@pytest.mark.parametrize("seed", [0]) +def test_perform_eigendecomposition( + test_name: str, + eigendecomposition_dtype: torch.dtype, + train_size: int, + seed: int, +) -> None: + # Makes sure that the Eigendecomposition computations are working properly. + model, train_dataset, _, data_collator, task = prepare_test( + test_name=test_name, + train_size=train_size, + seed=seed, + ) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + factor_args = FactorArguments( + eigendecomposition_dtype=eigendecomposition_dtype, + ) + analyzer.fit_covariance_matrices( + factors_name=DEFAULT_FACTORS_NAME, + factor_args=factor_args, + dataset=train_dataset, + per_device_batch_size=None, + overwrite_output_dir=True, + dataloader_kwargs=kwargs, + ) + analyzer.perform_eigendecomposition( + factors_name=DEFAULT_FACTORS_NAME, + factor_args=factor_args, + overwrite_output_dir=True, + ) + eigen_factors = analyzer.load_eigendecomposition(factors_name=DEFAULT_FACTORS_NAME) + assert set(eigen_factors.keys()) == set(EIGENDECOMPOSITION_FACTOR_NAMES) + assert len(eigen_factors[ACTIVATION_EIGENVECTORS_NAME]) > 0 + assert len(eigen_factors[GRADIENT_EIGENVECTORS_NAME]) > 0 + for module_name in eigen_factors[ACTIVATION_EIGENVECTORS_NAME]: + assert eigen_factors[ACTIVATION_EIGENVECTORS_NAME][module_name] is not None + assert eigen_factors[GRADIENT_EIGENVECTORS_NAME][module_name] is not None diff --git a/tests/factors/test_eigens.py b/tests/factors/test_eigens.py deleted file mode 100644 index 8ab7cbb..0000000 --- a/tests/factors/test_eigens.py +++ /dev/null @@ -1,389 +0,0 @@ -# pylint: skip-file - -import pytest -import torch - -from kronfluence.arguments import FactorArguments -from kronfluence.utils.constants import ( - ACTIVATION_EIGENVECTORS_NAME, - EIGENDECOMPOSITION_FACTOR_NAMES, - GRADIENT_EIGENVECTORS_NAME, - LAMBDA_FACTOR_NAMES, - LAMBDA_MATRIX_NAME, - NUM_LAMBDA_PROCESSED, -) -from kronfluence.utils.dataset import DataLoaderKwargs -from tests.utils import ( - ATOL, - RTOL, - check_tensor_dict_equivalence, - prepare_model_and_analyzer, - prepare_test, -) - - -@pytest.mark.parametrize( - "test_name", - [ - "mlp", - "repeated_mlp", - "conv", - "conv_bn", - "bert", - "gpt", - ], -) -@pytest.mark.parametrize("eigendecomposition_dtype", [torch.float32, torch.float64]) -@pytest.mark.parametrize("train_size", [16]) -@pytest.mark.parametrize("seed", [0]) -def test_perform_eigendecomposition( - test_name: str, - eigendecomposition_dtype: torch.dtype, - train_size: int, - seed: int, -) -> None: - # Make sure that the Eigendecomposition computations are working properly. - model, train_dataset, _, data_collator, task = prepare_test( - test_name=test_name, - train_size=train_size, - seed=seed, - ) - kwargs = DataLoaderKwargs(collate_fn=data_collator) - model, analyzer = prepare_model_and_analyzer( - model=model, - task=task, - ) - factors_name = f"pytest_{test_name}_{test_perform_eigendecomposition.__name__}" - analyzer.fit_covariance_matrices( - factors_name=factors_name, - dataset=train_dataset, - per_device_batch_size=4, - overwrite_output_dir=True, - dataloader_kwargs=kwargs, - ) - factor_args = FactorArguments( - eigendecomposition_dtype=eigendecomposition_dtype, - ) - analyzer.perform_eigendecomposition( - factors_name=factors_name, - factor_args=factor_args, - overwrite_output_dir=True, - ) - eigen_factors = analyzer.load_eigendecomposition(factors_name=factors_name) - assert set(eigen_factors.keys()) == set(EIGENDECOMPOSITION_FACTOR_NAMES) - assert len(eigen_factors[ACTIVATION_EIGENVECTORS_NAME]) > 0 - for module_name in eigen_factors[ACTIVATION_EIGENVECTORS_NAME]: - assert eigen_factors[ACTIVATION_EIGENVECTORS_NAME][module_name] is not None - assert eigen_factors[GRADIENT_EIGENVECTORS_NAME][module_name] is not None - - -@pytest.mark.parametrize( - "test_name", - [ - "mlp", - "repeated_mlp", - "conv", - "conv_bn", - "bert", - "gpt", - ], -) -@pytest.mark.parametrize("lambda_dtype", [torch.float32, torch.bfloat16]) -@pytest.mark.parametrize("train_size", [16]) -@pytest.mark.parametrize("seed", [0]) -def test_fit_lambda_matrices( - test_name: str, - lambda_dtype: torch.dtype, - train_size: int, - seed: int, -) -> None: - # Make sure that the Lambda computations are working properly. - model, train_dataset, _, data_collator, task = prepare_test( - test_name=test_name, - train_size=train_size, - seed=seed, - ) - kwargs = DataLoaderKwargs(collate_fn=data_collator) - model, analyzer = prepare_model_and_analyzer( - model=model, - task=task, - ) - - factor_args = FactorArguments( - lambda_dtype=lambda_dtype, - ) - factors_name = f"pytest_{test_name}_{test_fit_lambda_matrices.__name__}" - analyzer.fit_all_factors( - factors_name=factors_name, - dataset=train_dataset, - per_device_batch_size=train_size // 4, - factor_args=factor_args, - dataloader_kwargs=kwargs, - overwrite_output_dir=True, - ) - - lambda_factors = analyzer.load_lambda_matrices(factors_name=factors_name) - assert set(lambda_factors.keys()) == set(LAMBDA_FACTOR_NAMES) - assert len(lambda_factors[LAMBDA_MATRIX_NAME]) > 0 - for module_name in lambda_factors[LAMBDA_MATRIX_NAME]: - assert lambda_factors[LAMBDA_MATRIX_NAME][module_name].dtype == lambda_dtype - - -@pytest.mark.parametrize( - "test_name", - [ - "mlp", - "conv", - "gpt", - ], -) -@pytest.mark.parametrize("strategy", ["diagonal", "ekfac"]) -@pytest.mark.parametrize("train_size", [50]) -@pytest.mark.parametrize("seed", [1]) -def test_lambda_matrices_batch_size_equivalence( - test_name: str, - strategy: str, - train_size: int, - seed: int, -) -> None: - # Lambda matrices should be identical regardless of what batch size used. - model, train_dataset, _, data_collator, task = prepare_test( - test_name=test_name, - train_size=train_size, - seed=seed, - ) - kwargs = DataLoaderKwargs(collate_fn=data_collator) - model, analyzer = prepare_model_and_analyzer( - model=model, - task=task, - ) - - factor_args = FactorArguments( - strategy=strategy, - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, - ) - analyzer.fit_all_factors( - factors_name=f"pytest_{test_name}_{test_lambda_matrices_batch_size_equivalence.__name__}_{strategy}_bs1", - dataset=train_dataset, - per_device_batch_size=1, - factor_args=factor_args, - dataloader_kwargs=kwargs, - overwrite_output_dir=True, - ) - bs1_lambda_factors = analyzer.load_lambda_matrices( - factors_name=f"pytest_{test_name}_{test_lambda_matrices_batch_size_equivalence.__name__}_{strategy}_bs1", - ) - - analyzer.fit_all_factors( - factors_name=f"pytest_{test_name}_{test_lambda_matrices_batch_size_equivalence.__name__}_{strategy}_bs8", - dataset=train_dataset, - per_device_batch_size=8, - factor_args=factor_args, - dataloader_kwargs=kwargs, - overwrite_output_dir=True, - ) - bs8_lambda_factors = analyzer.load_lambda_matrices( - factors_name=f"pytest_{test_name}_{test_lambda_matrices_batch_size_equivalence.__name__}_{strategy}_bs8", - ) - - for name in LAMBDA_FACTOR_NAMES: - assert check_tensor_dict_equivalence(bs1_lambda_factors[name], bs8_lambda_factors[name], atol=ATOL, rtol=RTOL) - - -@pytest.mark.parametrize( - "test_name", - [ - "mlp", - "conv", - "gpt", - ], -) -@pytest.mark.parametrize("strategy", ["diagonal", "ekfac"]) -@pytest.mark.parametrize("data_partition_size", [1, 4]) -@pytest.mark.parametrize("module_partition_size", [1, 3]) -@pytest.mark.parametrize("train_size", [81]) -@pytest.mark.parametrize("seed", [2]) -def test_lambda_matrices_partition_equivalence( - test_name: str, - strategy: str, - data_partition_size: int, - module_partition_size: int, - train_size: int, - seed: int, -) -> None: - # Lambda matrices should be identical regardless of what the partition used. - model, train_dataset, _, data_collator, task = prepare_test( - test_name=test_name, - train_size=train_size, - seed=seed, - ) - kwargs = DataLoaderKwargs(collate_fn=data_collator) - model, analyzer = prepare_model_and_analyzer( - model=model, - task=task, - ) - - factor_args = FactorArguments( - strategy=strategy, - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, - ) - factors_name = f"pytest_{test_name}_{strategy}_{test_lambda_matrices_partition_equivalence.__name__}" - analyzer.fit_all_factors( - factors_name=factors_name, - dataset=train_dataset, - factor_args=factor_args, - per_device_batch_size=8, - overwrite_output_dir=True, - dataloader_kwargs=kwargs, - ) - lambda_factors = analyzer.load_lambda_matrices( - factors_name=factors_name, - ) - - factor_args = FactorArguments( - strategy=strategy, - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, - lambda_data_partition_size=data_partition_size, - lambda_module_partition_size=module_partition_size, - ) - analyzer.fit_all_factors( - factors_name=f"pytest_{test_name}_{strategy}_{data_partition_size}_{module_partition_size}", - dataset=train_dataset, - factor_args=factor_args, - per_device_batch_size=6, - overwrite_output_dir=True, - dataloader_kwargs=kwargs, - ) - partitioned_lambda_factors = analyzer.load_lambda_matrices( - factors_name=f"pytest_{test_name}_{strategy}_{data_partition_size}_{module_partition_size}", - ) - for name in LAMBDA_FACTOR_NAMES: - assert check_tensor_dict_equivalence( - lambda_factors[name], partitioned_lambda_factors[name], atol=ATOL, rtol=RTOL - ) - - -@pytest.mark.parametrize( - "test_name", - [ - "mlp", - "conv", - "gpt", - ], -) -@pytest.mark.parametrize("train_size", [63]) -@pytest.mark.parametrize("seed", [3]) -def test_lambda_matrices_iterative_aggregate( - test_name: str, - train_size: int, - seed: int, -) -> None: - # Make sure aggregated lambda computation is working properly. - model, train_dataset, _, data_collator, task = prepare_test( - test_name=test_name, - train_size=train_size, - seed=seed, - ) - kwargs = DataLoaderKwargs(collate_fn=data_collator) - model, analyzer = prepare_model_and_analyzer( - model=model, - task=task, - ) - - factors_name = f"pytest_{test_name}_{test_lambda_matrices_iterative_aggregate.__name__}" - factor_args = FactorArguments( - use_empirical_fisher=True, - lambda_iterative_aggregate=False, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, - ) - analyzer.fit_all_factors( - factors_name=factors_name, - dataset=train_dataset, - factor_args=factor_args, - per_device_batch_size=8, - overwrite_output_dir=True, - dataloader_kwargs=kwargs, - ) - lambda_factors = analyzer.load_lambda_matrices( - factors_name=factors_name, - ) - - factor_args = FactorArguments( - use_empirical_fisher=True, - lambda_iterative_aggregate=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, - ) - analyzer.fit_all_factors( - factors_name=factors_name + "_iterative", - dataset=train_dataset, - factor_args=factor_args, - per_device_batch_size=4, - overwrite_output_dir=True, - dataloader_kwargs=kwargs, - ) - iterative_lambda_factors = analyzer.load_lambda_matrices( - factors_name=factors_name + "_iterative", - ) - - for name in LAMBDA_FACTOR_NAMES: - assert check_tensor_dict_equivalence(lambda_factors[name], iterative_lambda_factors[name], atol=ATOL, rtol=RTOL) - - -@pytest.mark.parametrize( - "test_name", - [ - "mlp", - ], -) -@pytest.mark.parametrize("data_partition_size", [1, 4]) -@pytest.mark.parametrize("train_size", [82]) -@pytest.mark.parametrize("seed", [4]) -def test_lambda_matrices_max_examples( - test_name: str, - data_partition_size: int, - train_size: int, - seed: int, -) -> None: - # Make sure the max Lambda data selection is working properly. - model, train_dataset, _, data_collator, task = prepare_test( - test_name=test_name, - train_size=train_size, - seed=seed, - ) - kwargs = DataLoaderKwargs(collate_fn=data_collator) - model, analyzer = prepare_model_and_analyzer( - model=model, - task=task, - ) - - MAX_EXAMPLES = 28 - factor_args = FactorArguments( - use_empirical_fisher=True, lambda_max_examples=MAX_EXAMPLES, lambda_data_partition_size=data_partition_size - ) - factors_name = f"pytest_{test_name}_{test_lambda_matrices_max_examples.__name__}" - analyzer.fit_all_factors( - factors_name=factors_name, - dataset=train_dataset, - factor_args=factor_args, - per_device_batch_size=8, - overwrite_output_dir=True, - dataloader_kwargs=kwargs, - ) - lambda_factors = analyzer.load_lambda_matrices( - factors_name=factors_name, - ) - for num_examples in lambda_factors[NUM_LAMBDA_PROCESSED].values(): - assert num_examples == MAX_EXAMPLES diff --git a/tests/factors/test_lambdas.py b/tests/factors/test_lambdas.py new file mode 100644 index 0000000..d2f15e8 --- /dev/null +++ b/tests/factors/test_lambdas.py @@ -0,0 +1,502 @@ +# pylint: skip-file + +import pytest +import torch + +from kronfluence.arguments import FactorArguments +from kronfluence.utils.common.factor_arguments import pytest_factor_arguments +from kronfluence.utils.constants import ( + LAMBDA_FACTOR_NAMES, + LAMBDA_MATRIX_NAME, + NUM_LAMBDA_PROCESSED, +) +from kronfluence.utils.dataset import DataLoaderKwargs +from tests.utils import ( + ATOL, + DEFAULT_FACTORS_NAME, + RTOL, + check_tensor_dict_equivalence, + custom_factors_name, + prepare_model_and_analyzer, + prepare_test, +) + + +@pytest.mark.parametrize( + "test_name", + [ + "mlp", + "repeated_mlp", + "conv", + "bert", + "roberta", + "gpt", + "gpt_checkpoint", + ], +) +@pytest.mark.parametrize("per_sample_gradient_dtype", [torch.float32, torch.bfloat16]) +@pytest.mark.parametrize("lambda_dtype", [torch.float32, torch.bfloat16]) +@pytest.mark.parametrize("train_size", [16]) +@pytest.mark.parametrize("seed", [0]) +def test_fit_lambda_matrices( + test_name: str, + per_sample_gradient_dtype: torch.dtype, + lambda_dtype: torch.dtype, + train_size: int, + seed: int, +) -> None: + # Makes sure that the Lambda computations are working properly. + model, train_dataset, _, data_collator, task = prepare_test( + test_name=test_name, + train_size=train_size, + seed=seed, + ) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + + factor_args = FactorArguments( + lambda_dtype=lambda_dtype, + per_sample_gradient_dtype=per_sample_gradient_dtype, + ) + if test_name == "repeated_mlp": + factor_args.has_shared_parameters = True + + analyzer.fit_all_factors( + factors_name=DEFAULT_FACTORS_NAME, + dataset=train_dataset, + per_device_batch_size=train_size // 4, + factor_args=factor_args, + dataloader_kwargs=kwargs, + overwrite_output_dir=True, + ) + + lambda_factors = analyzer.load_lambda_matrices(factors_name=DEFAULT_FACTORS_NAME) + assert set(lambda_factors.keys()) == set(LAMBDA_FACTOR_NAMES) + assert len(lambda_factors[LAMBDA_MATRIX_NAME]) > 0 + for module_name in lambda_factors[LAMBDA_MATRIX_NAME]: + assert lambda_factors[LAMBDA_MATRIX_NAME][module_name].dtype == lambda_dtype + + +@pytest.mark.parametrize( + "test_name", + [ + "mlp", + "conv", + "roberta", + ], +) +@pytest.mark.parametrize("strategy", ["diagonal", "ekfac"]) +@pytest.mark.parametrize("train_size", [50]) +@pytest.mark.parametrize("seed", [1]) +def test_lambda_matrices_batch_size_equivalence( + test_name: str, + strategy: str, + train_size: int, + seed: int, +) -> None: + # Lambda matrices should be identical regardless of what batch size used. + model, train_dataset, _, data_collator, task = prepare_test( + test_name=test_name, + train_size=train_size, + seed=seed, + ) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + model = model.to(dtype=torch.float64) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + + factor_args = pytest_factor_arguments(strategy=strategy) + analyzer.fit_all_factors( + factors_name=DEFAULT_FACTORS_NAME, + dataset=train_dataset, + per_device_batch_size=1, + factor_args=factor_args, + dataloader_kwargs=kwargs, + overwrite_output_dir=True, + ) + bs1_lambda_factors = analyzer.load_lambda_matrices( + factors_name=DEFAULT_FACTORS_NAME, + ) + + analyzer.fit_all_factors( + factors_name=custom_factors_name("bs8"), + dataset=train_dataset, + per_device_batch_size=8, + factor_args=factor_args, + dataloader_kwargs=kwargs, + overwrite_output_dir=True, + ) + bs8_lambda_factors = analyzer.load_lambda_matrices( + factors_name=custom_factors_name("bs8"), + ) + + for name in LAMBDA_FACTOR_NAMES: + assert check_tensor_dict_equivalence(bs1_lambda_factors[name], bs8_lambda_factors[name], atol=ATOL, rtol=RTOL) + + analyzer.fit_all_factors( + factors_name=custom_factors_name("auto"), + dataset=train_dataset, + per_device_batch_size=None, + factor_args=factor_args, + dataloader_kwargs=kwargs, + overwrite_output_dir=True, + ) + auto_lambda_factors = analyzer.load_lambda_matrices( + factors_name=custom_factors_name("auto"), + ) + + for name in LAMBDA_FACTOR_NAMES: + assert check_tensor_dict_equivalence(bs1_lambda_factors[name], auto_lambda_factors[name], atol=ATOL, rtol=RTOL) + + +@pytest.mark.parametrize("test_name", ["mlp", "conv_bn"]) +@pytest.mark.parametrize("strategy", ["diagonal", "ekfac"]) +@pytest.mark.parametrize("data_partitions", [2, 4]) +@pytest.mark.parametrize("module_partitions", [2, 3]) +@pytest.mark.parametrize("train_size", [81]) +@pytest.mark.parametrize("seed", [2]) +def test_lambda_matrices_partition_equivalence( + test_name: str, + strategy: str, + data_partitions: int, + module_partitions: int, + train_size: int, + seed: int, +) -> None: + # Lambda matrices should be identical regardless of what the partition used. + model, train_dataset, _, data_collator, task = prepare_test( + test_name=test_name, + train_size=train_size, + seed=seed, + ) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + + factor_args = pytest_factor_arguments(strategy=strategy) + analyzer.fit_all_factors( + factors_name=DEFAULT_FACTORS_NAME, + dataset=train_dataset, + factor_args=factor_args, + per_device_batch_size=8, + overwrite_output_dir=True, + dataloader_kwargs=kwargs, + ) + lambda_factors = analyzer.load_lambda_matrices( + factors_name=DEFAULT_FACTORS_NAME, + ) + + factor_args.lambda_data_partitions = data_partitions + factor_args.lambda_module_partitions = module_partitions + analyzer.fit_all_factors( + factors_name=custom_factors_name(f"{data_partitions}_{module_partitions}"), + dataset=train_dataset, + factor_args=factor_args, + per_device_batch_size=6, + overwrite_output_dir=True, + dataloader_kwargs=kwargs, + ) + partitioned_lambda_factors = analyzer.load_lambda_matrices( + factors_name=custom_factors_name(f"{data_partitions}_{module_partitions}"), + ) + for name in LAMBDA_FACTOR_NAMES: + assert check_tensor_dict_equivalence( + lambda_factors[name], partitioned_lambda_factors[name], atol=ATOL, rtol=RTOL + ) + + +@pytest.mark.parametrize( + "test_name", + [ + "mlp", + "conv_bn", + "bert", + "gpt", + ], +) +@pytest.mark.parametrize("train_size", [63, 121]) +@pytest.mark.parametrize("seed", [3]) +def test_lambda_matrices_iterative_lambda_aggregation( + test_name: str, + train_size: int, + seed: int, +) -> None: + # Makes sure iterative lambda computation is working properly. + model, train_dataset, _, data_collator, task = prepare_test( + test_name=test_name, + train_size=train_size, + seed=seed, + ) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + model = model.to(dtype=torch.float64) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + + factor_args = pytest_factor_arguments() + factor_args.use_iterative_lambda_aggregation = False + analyzer.fit_all_factors( + factors_name=DEFAULT_FACTORS_NAME, + dataset=train_dataset, + factor_args=factor_args, + per_device_batch_size=8, + overwrite_output_dir=True, + dataloader_kwargs=kwargs, + ) + lambda_factors = analyzer.load_lambda_matrices( + factors_name=DEFAULT_FACTORS_NAME, + ) + + factor_args.use_iterative_lambda_aggregation = True + analyzer.fit_all_factors( + factors_name=custom_factors_name("iterative"), + dataset=train_dataset, + factor_args=factor_args, + per_device_batch_size=16, + overwrite_output_dir=True, + dataloader_kwargs=kwargs, + ) + iterative_lambda_factors = analyzer.load_lambda_matrices( + factors_name=custom_factors_name("iterative"), + ) + + for name in LAMBDA_FACTOR_NAMES: + assert check_tensor_dict_equivalence(lambda_factors[name], iterative_lambda_factors[name], atol=ATOL, rtol=RTOL) + + +@pytest.mark.parametrize( + "test_name", + ["conv_bn", "gpt"], +) +@pytest.mark.parametrize("max_examples", [4, 31]) +@pytest.mark.parametrize("data_partitions", [1, 3]) +@pytest.mark.parametrize("module_partitions", [1, 2]) +@pytest.mark.parametrize("train_size", [82]) +@pytest.mark.parametrize("seed", [4]) +def test_lambda_matrices_max_examples( + test_name: str, + max_examples: int, + data_partitions: int, + module_partitions: int, + train_size: int, + seed: int, +) -> None: + # Makes sure the max Lambda data selection is working properly. + model, train_dataset, _, data_collator, task = prepare_test( + test_name=test_name, + train_size=train_size, + seed=seed, + ) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + + factor_args = FactorArguments( + lambda_max_examples=max_examples, + lambda_data_partitions=data_partitions, + lambda_module_partitions=module_partitions, + ) + analyzer.fit_all_factors( + factors_name=DEFAULT_FACTORS_NAME, + dataset=train_dataset, + factor_args=factor_args, + per_device_batch_size=8, + overwrite_output_dir=True, + dataloader_kwargs=kwargs, + ) + lambda_factors = analyzer.load_lambda_matrices( + factors_name=DEFAULT_FACTORS_NAME, + ) + for num_examples in lambda_factors[NUM_LAMBDA_PROCESSED].values(): + assert num_examples == max_examples + + +@pytest.mark.parametrize( + "test_name", + [ + "mlp", + "gpt", + ], +) +@pytest.mark.parametrize("train_size", [105]) +@pytest.mark.parametrize("seed", [6]) +def test_lambda_matrices_gradient_checkpoint( + test_name: str, + train_size: int, + seed: int, +) -> None: + # Lambda matrices should be the same even when gradient checkpointing is used. + model, train_dataset, _, data_collator, task = prepare_test( + test_name=test_name, + train_size=train_size, + seed=seed, + ) + model = model.to(dtype=torch.float64) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + + factor_args = pytest_factor_arguments() + analyzer.fit_all_factors( + factors_name=DEFAULT_FACTORS_NAME, + dataset=train_dataset, + per_device_batch_size=5, + overwrite_output_dir=True, + factor_args=factor_args, + dataloader_kwargs=kwargs, + ) + lambda_factors = analyzer.load_lambda_matrices( + factors_name=DEFAULT_FACTORS_NAME, + ) + + model, _, _, _, task = prepare_test( + test_name=test_name + "_checkpoint", + train_size=train_size, + seed=seed, + ) + model = model.to(dtype=torch.float64) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + analyzer.fit_all_factors( + factors_name=custom_factors_name("cp"), + dataset=train_dataset, + per_device_batch_size=6, + overwrite_output_dir=True, + factor_args=factor_args, + dataloader_kwargs=kwargs, + ) + checkpoint_lambda_factors = analyzer.load_lambda_matrices( + factors_name=custom_factors_name("cp"), + ) + + for name in LAMBDA_FACTOR_NAMES: + assert check_tensor_dict_equivalence( + lambda_factors[name], checkpoint_lambda_factors[name], atol=ATOL, rtol=RTOL + ) + + +@pytest.mark.parametrize( + "test_name", + ["mlp", "conv", "gpt"], +) +@pytest.mark.parametrize("train_size", [105]) +@pytest.mark.parametrize("seed", [7]) +def test_lambda_matrices_shared_parameters( + test_name: str, + train_size: int, + seed: int, +) -> None: + # When there are no shared parameters, results with and without `has_shared_parameters` should + # produce the same results. + model, train_dataset, _, data_collator, task = prepare_test( + test_name=test_name, + train_size=train_size, + seed=seed, + ) + model = model.to(dtype=torch.float64) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + + factor_args = pytest_factor_arguments() + analyzer.fit_all_factors( + factors_name=DEFAULT_FACTORS_NAME, + dataset=train_dataset, + per_device_batch_size=5, + overwrite_output_dir=True, + factor_args=factor_args, + dataloader_kwargs=kwargs, + ) + lambda_factors = analyzer.load_lambda_matrices( + factors_name=DEFAULT_FACTORS_NAME, + ) + + factor_args.has_shared_parameters = True + analyzer.fit_all_factors( + factors_name=custom_factors_name("shared"), + dataset=train_dataset, + per_device_batch_size=6, + overwrite_output_dir=True, + factor_args=factor_args, + dataloader_kwargs=kwargs, + ) + shared_lambda_factors = analyzer.load_lambda_matrices( + factors_name=custom_factors_name("shared"), + ) + + for name in LAMBDA_FACTOR_NAMES: + assert check_tensor_dict_equivalence(lambda_factors[name], shared_lambda_factors[name], atol=ATOL, rtol=RTOL) + + +@pytest.mark.parametrize("train_size", [121]) +@pytest.mark.parametrize("seed", [8]) +def test_lambda_matrices_inplace( + train_size: int, + seed: int, +) -> None: + # Lambda matrices should be the identical for with and without in-place ReLU. + model, train_dataset, _, data_collator, task = prepare_test( + test_name="conv", + train_size=train_size, + seed=seed, + ) + model = model.to(dtype=torch.float64) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + + factor_args = pytest_factor_arguments() + analyzer.fit_all_factors( + factors_name=DEFAULT_FACTORS_NAME, + dataset=train_dataset, + per_device_batch_size=5, + overwrite_output_dir=True, + factor_args=factor_args, + dataloader_kwargs=kwargs, + ) + lambda_factors = analyzer.load_lambda_matrices( + factors_name=DEFAULT_FACTORS_NAME, + ) + + model, _, _, _, task = prepare_test( + test_name="conv_inplace", + train_size=train_size, + seed=seed, + ) + model = model.to(dtype=torch.float64) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + analyzer.fit_all_factors( + factors_name=custom_factors_name("inplace"), + dataset=train_dataset, + per_device_batch_size=6, + overwrite_output_dir=True, + factor_args=factor_args, + dataloader_kwargs=kwargs, + ) + inplace_lambda_factors = analyzer.load_lambda_matrices( + factors_name=custom_factors_name("inplace"), + ) + + for name in LAMBDA_FACTOR_NAMES: + assert check_tensor_dict_equivalence(lambda_factors[name], inplace_lambda_factors[name], atol=ATOL, rtol=RTOL) diff --git a/tests/gpu_tests/README.md b/tests/gpu_tests/README.md index 4d1dd31..e83f15d 100644 --- a/tests/gpu_tests/README.md +++ b/tests/gpu_tests/README.md @@ -2,42 +2,55 @@ This folder contains various GPU tests for Kronfluence. Before running the tests, you need to prepare the baseline results by training an MNIST model and saving the results obtained with a single GPU: + ```bash python prepare_tests.py ``` ### CPU Tests + To test if running on CPU yields the same result as the GPU, run: + ```bash python cpu_test.py ``` ### DDP Tests -To test if running with Distributed Data Parallel (DDP) with 3 GPUs obtains the same result, run: + +To test if running with Distributed Data Parallel (DDP) with 4 GPUs obtains the same result, run: + ```bash -torchrun --nnodes=1 --nproc_per_node=3 ddp_test.py +torchrun --nnodes=1 --nproc_per_node=4 ddp_test.py ``` ### FSDP Tests -To test if running with Fully Sharded Data Parallel (FSDP) with 3 GPUs obtains the same result, run: + +To test if running with Fully Sharded Data Parallel (FSDP) with 4 GPUs obtains the same result, run: + ```bash -torchrun --nnodes=1 --nproc_per_node=3 fsdp_test.py +torchrun --nnodes=1 --nproc_per_node=4 fsdp_test.py ``` ### torch.compile Tests + To test if running with `torch.compile` obtains the same result, run: + ```bash python compile_test.py ``` ### AMP Tests + To test if running with automatic mixed precision (AMP) obtains the similar result, run: + ```bash python amp_test.py ``` ### CPU Offload Test -To test if `cached_activation_cpu_offload` option is properly implemented, run: + +To test if `offload_activations_to_cpu` option is properly implemented, run: + ```bash pytest test_offload_cpu.py ``` \ No newline at end of file diff --git a/tests/gpu_tests/amp_test.py b/tests/gpu_tests/amp_test.py index 7c27df8..3a81d5f 100644 --- a/tests/gpu_tests/amp_test.py +++ b/tests/gpu_tests/amp_test.py @@ -7,7 +7,8 @@ from torch.utils import data from kronfluence.analyzer import Analyzer, prepare_model -from kronfluence.arguments import FactorArguments, ScoreArguments +from kronfluence.utils.common.factor_arguments import pytest_factor_arguments +from kronfluence.utils.common.score_arguments import pytest_score_arguments from kronfluence.utils.constants import ( ALL_MODULE_NAME, COVARIANCE_FACTOR_NAMES, @@ -45,13 +46,8 @@ def setUpClass(cls) -> None: def test_covariance_matrices(self) -> None: covariance_factors = self.analyzer.load_covariance_matrices(factors_name=OLD_FACTOR_NAME) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, - amp_dtype=torch.float16, - ) + factor_args = pytest_factor_arguments() + factor_args.amp_dtype = torch.float16 self.analyzer.fit_covariance_matrices( factors_name=NEW_FACTOR_NAME, dataset=self.train_dataset, @@ -69,19 +65,14 @@ def test_covariance_matrices(self) -> None: assert check_tensor_dict_equivalence( covariance_factors[name], new_covariance_factors[name], - atol=1e-5, - rtol=1e-3, + atol=1e-3, + rtol=1e-1, ) - def test_lambda_matrices(self): + def test_lambda_matrices(self) -> None: lambda_factors = self.analyzer.load_lambda_matrices(factors_name=OLD_FACTOR_NAME) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, - amp_dtype=torch.float16, - ) + factor_args = pytest_factor_arguments() + factor_args.amp_dtype = torch.float16 self.analyzer.fit_lambda_matrices( factors_name=NEW_FACTOR_NAME, dataset=self.train_dataset, @@ -107,12 +98,8 @@ def test_lambda_matrices(self): def test_pairwise_scores(self) -> None: pairwise_scores = self.analyzer.load_pairwise_scores(scores_name=OLD_SCORE_NAME) - score_args = ScoreArguments( - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - amp_dtype=torch.float16, - ) + score_args = pytest_score_arguments() + score_args.amp_dtype = torch.float16 self.analyzer.compute_pairwise_scores( scores_name=NEW_SCORE_NAME, factors_name=OLD_FACTOR_NAME, @@ -134,17 +121,13 @@ def test_pairwise_scores(self) -> None: assert check_tensor_dict_equivalence( pairwise_scores, new_pairwise_scores, - atol=1e-5, - rtol=1e-3, + atol=1e-3, + rtol=1e-1, ) def test_self_scores(self) -> None: - score_args = ScoreArguments( - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - amp_dtype=torch.float16, - ) + score_args = pytest_score_arguments() + score_args.amp_dtype = torch.float16 self.analyzer.compute_self_scores( scores_name=NEW_SCORE_NAME, factors_name=OLD_FACTOR_NAME, @@ -155,7 +138,6 @@ def test_self_scores(self) -> None: overwrite_output_dir=True, ) new_self_scores = self.analyzer.load_self_scores(scores_name=NEW_SCORE_NAME) - self_scores = self.analyzer.load_self_scores(scores_name=OLD_SCORE_NAME) print(f"Previous score: {self_scores[ALL_MODULE_NAME]}") print(f"Previous shape: {self_scores[ALL_MODULE_NAME].shape}") @@ -164,8 +146,8 @@ def test_self_scores(self) -> None: assert check_tensor_dict_equivalence( self_scores, new_self_scores, - atol=1e-5, - rtol=1e-3, + atol=1e-3, + rtol=1e-1, ) diff --git a/tests/gpu_tests/compile_test.py b/tests/gpu_tests/compile_test.py index 61e9db5..10b3a6e 100644 --- a/tests/gpu_tests/compile_test.py +++ b/tests/gpu_tests/compile_test.py @@ -7,7 +7,8 @@ from torch.utils import data from kronfluence.analyzer import Analyzer, prepare_model -from kronfluence.arguments import FactorArguments, ScoreArguments +from kronfluence.utils.common.factor_arguments import pytest_factor_arguments +from kronfluence.utils.common.score_arguments import pytest_score_arguments from kronfluence.utils.constants import ( ALL_MODULE_NAME, COVARIANCE_FACTOR_NAMES, @@ -15,7 +16,7 @@ ) from tests.gpu_tests.pipeline import GpuTestTask, construct_test_mlp, get_mnist_dataset from tests.gpu_tests.prepare_tests import QUERY_INDICES, TRAIN_INDICES -from tests.utils import check_tensor_dict_equivalence +from tests.utils import ATOL, RTOL, check_tensor_dict_equivalence logging.basicConfig(level=logging.DEBUG) OLD_FACTOR_NAME = "single_gpu" @@ -41,6 +42,9 @@ def setUpClass(cls) -> None: cls.model = cls.model.cuda() cls.model = torch.compile(cls.model) + print(cls.model) + print(list(cls.model.named_modules())) + cls.analyzer = Analyzer( analysis_name="gpu_test", model=cls.model, @@ -49,12 +53,7 @@ def setUpClass(cls) -> None: def test_covariance_matrices(self) -> None: covariance_factors = self.analyzer.load_covariance_matrices(factors_name=OLD_FACTOR_NAME) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, - ) + factor_args = pytest_factor_arguments() self.analyzer.fit_covariance_matrices( factors_name=NEW_FACTOR_NAME, dataset=self.train_dataset, @@ -69,15 +68,42 @@ def test_covariance_matrices(self) -> None: print(f"Name: {name, module_name}") print(f"Previous factor: {covariance_factors[name][module_name]}") print(f"New factor: {new_covariance_factors[name][module_name]}") + assert check_tensor_dict_equivalence( + covariance_factors[name], + new_covariance_factors[name], + atol=ATOL, + rtol=RTOL, + ) - def test_lambda_matrices(self): + def test_lambda_matrices(self) -> None: lambda_factors = self.analyzer.load_lambda_matrices(factors_name=OLD_FACTOR_NAME) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, + factor_args = pytest_factor_arguments() + self.analyzer.fit_lambda_matrices( + factors_name=NEW_FACTOR_NAME, + dataset=self.train_dataset, + factor_args=factor_args, + per_device_batch_size=512, + overwrite_output_dir=True, + load_from_factors_name=OLD_FACTOR_NAME, ) + new_lambda_factors = self.analyzer.load_lambda_matrices(factors_name=NEW_FACTOR_NAME) + + for name in LAMBDA_FACTOR_NAMES: + for module_name in lambda_factors[name]: + print(f"Name: {name, module_name}") + print(f"Previous factor: {lambda_factors[name][module_name]}") + print(f"New factor: {new_lambda_factors[name][module_name]}") + assert check_tensor_dict_equivalence( + lambda_factors[name], + new_lambda_factors[name], + atol=ATOL, + rtol=RTOL, + ) + + def test_lambda_shared_matrices(self) -> None: + lambda_factors = self.analyzer.load_lambda_matrices(factors_name=OLD_FACTOR_NAME) + factor_args = pytest_factor_arguments() + factor_args.has_shared_parameters = True self.analyzer.fit_lambda_matrices( factors_name=NEW_FACTOR_NAME, dataset=self.train_dataset, @@ -96,18 +122,12 @@ def test_lambda_matrices(self): assert check_tensor_dict_equivalence( lambda_factors[name], new_lambda_factors[name], - atol=1e-3, - rtol=1e-1, + atol=ATOL, + rtol=RTOL, ) def test_pairwise_scores(self) -> None: - pairwise_scores = self.analyzer.load_pairwise_scores(scores_name=OLD_SCORE_NAME) - - score_args = ScoreArguments( - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - ) + score_args = pytest_score_arguments() self.analyzer.compute_pairwise_scores( scores_name=NEW_SCORE_NAME, factors_name=OLD_FACTOR_NAME, @@ -121,8 +141,8 @@ def test_pairwise_scores(self) -> None: overwrite_output_dir=True, ) new_pairwise_scores = self.analyzer.load_pairwise_scores(scores_name=NEW_SCORE_NAME) + pairwise_scores = self.analyzer.load_pairwise_scores(scores_name=OLD_SCORE_NAME) - torch.set_printoptions(threshold=30_000) print(f"Previous score: {pairwise_scores[ALL_MODULE_NAME][10]}") print(f"Previous shape: {pairwise_scores[ALL_MODULE_NAME].shape}") print(f"New score: {new_pairwise_scores[ALL_MODULE_NAME][10]}") @@ -130,16 +150,12 @@ def test_pairwise_scores(self) -> None: assert check_tensor_dict_equivalence( pairwise_scores, new_pairwise_scores, - atol=1e-5, - rtol=1e-3, + atol=ATOL, + rtol=RTOL, ) def test_self_scores(self) -> None: - score_args = ScoreArguments( - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - ) + score_args = pytest_score_arguments() self.analyzer.compute_self_scores( scores_name=NEW_SCORE_NAME, factors_name=OLD_FACTOR_NAME, @@ -150,9 +166,8 @@ def test_self_scores(self) -> None: overwrite_output_dir=True, ) new_self_scores = self.analyzer.load_self_scores(scores_name=NEW_SCORE_NAME) - self_scores = self.analyzer.load_self_scores(scores_name=OLD_SCORE_NAME) - torch.set_printoptions(threshold=30_000) + print(f"Previous score: {self_scores[ALL_MODULE_NAME]}") print(f"Previous shape: {self_scores[ALL_MODULE_NAME].shape}") print(f"New score: {new_self_scores[ALL_MODULE_NAME]}") @@ -160,8 +175,8 @@ def test_self_scores(self) -> None: assert check_tensor_dict_equivalence( self_scores, new_self_scores, - atol=1e-5, - rtol=1e-3, + atol=ATOL, + rtol=RTOL, ) diff --git a/tests/gpu_tests/cpu_test.py b/tests/gpu_tests/cpu_test.py index 455327d..29c3f13 100644 --- a/tests/gpu_tests/cpu_test.py +++ b/tests/gpu_tests/cpu_test.py @@ -7,7 +7,8 @@ from torch.utils import data from kronfluence.analyzer import Analyzer, prepare_model -from kronfluence.arguments import FactorArguments, ScoreArguments +from kronfluence.utils.common.factor_arguments import pytest_factor_arguments +from kronfluence.utils.common.score_arguments import pytest_score_arguments from kronfluence.utils.constants import ( ALL_MODULE_NAME, COVARIANCE_FACTOR_NAMES, @@ -15,7 +16,7 @@ ) from tests.gpu_tests.pipeline import GpuTestTask, construct_test_mlp, get_mnist_dataset from tests.gpu_tests.prepare_tests import QUERY_INDICES, TRAIN_INDICES -from tests.utils import check_tensor_dict_equivalence +from tests.utils import ATOL, RTOL, check_tensor_dict_equivalence logging.basicConfig(level=logging.DEBUG) OLD_FACTOR_NAME = "single_gpu" @@ -43,12 +44,7 @@ def setUpClass(cls) -> None: def test_covariance_matrices(self) -> None: covariance_factors = self.analyzer.load_covariance_matrices(factors_name=OLD_FACTOR_NAME) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, - ) + factor_args = pytest_factor_arguments() self.analyzer.fit_covariance_matrices( factors_name=NEW_FACTOR_NAME, dataset=self.train_dataset, @@ -66,18 +62,13 @@ def test_covariance_matrices(self) -> None: assert check_tensor_dict_equivalence( covariance_factors[name], new_covariance_factors[name], - atol=1e-5, - rtol=1e-3, + atol=ATOL, + rtol=RTOL, ) - def test_lambda_matrices(self): + def test_lambda_matrices(self) -> None: lambda_factors = self.analyzer.load_lambda_matrices(factors_name=OLD_FACTOR_NAME) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, - ) + factor_args = pytest_factor_arguments() self.analyzer.fit_lambda_matrices( factors_name=NEW_FACTOR_NAME, dataset=self.train_dataset, @@ -96,18 +87,14 @@ def test_lambda_matrices(self): assert check_tensor_dict_equivalence( lambda_factors[name], new_lambda_factors[name], - atol=1e-3, - rtol=1e-1, + atol=ATOL, + rtol=RTOL, ) def test_pairwise_scores(self) -> None: pairwise_scores = self.analyzer.load_pairwise_scores(scores_name=OLD_SCORE_NAME) - score_args = ScoreArguments( - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - ) + score_args = pytest_score_arguments() self.analyzer.compute_pairwise_scores( scores_name=NEW_SCORE_NAME, factors_name=OLD_FACTOR_NAME, @@ -129,16 +116,12 @@ def test_pairwise_scores(self) -> None: assert check_tensor_dict_equivalence( pairwise_scores, new_pairwise_scores, - atol=1e-5, - rtol=1e-3, + atol=ATOL, + rtol=RTOL, ) def test_self_scores(self) -> None: - score_args = ScoreArguments( - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - ) + score_args = pytest_score_arguments() self.analyzer.compute_self_scores( scores_name=NEW_SCORE_NAME, factors_name=OLD_FACTOR_NAME, @@ -158,8 +141,34 @@ def test_self_scores(self) -> None: assert check_tensor_dict_equivalence( self_scores, new_self_scores, - atol=1e-5, - rtol=1e-3, + atol=ATOL, + rtol=RTOL, + ) + + def test_self_scores_with_measurement(self) -> None: + score_args = pytest_score_arguments() + score_args.use_measurement_for_self_influence = True + self.analyzer.compute_self_scores( + scores_name=NEW_SCORE_NAME + "_measurement", + factors_name=OLD_FACTOR_NAME, + train_dataset=self.train_dataset, + train_indices=list(range(TRAIN_INDICES)), + per_device_train_batch_size=512, + score_args=score_args, + overwrite_output_dir=True, + ) + new_self_scores = self.analyzer.load_self_scores(scores_name=NEW_SCORE_NAME + "_measurement") + + self_scores = self.analyzer.load_self_scores(scores_name=OLD_SCORE_NAME + "_measurement") + print(f"Previous score: {self_scores[ALL_MODULE_NAME]}") + print(f"Previous shape: {self_scores[ALL_MODULE_NAME].shape}") + print(f"New score: {new_self_scores[ALL_MODULE_NAME]}") + print(f"New shape: {new_self_scores[ALL_MODULE_NAME].shape}") + assert check_tensor_dict_equivalence( + self_scores, + new_self_scores, + atol=ATOL, + rtol=RTOL, ) diff --git a/tests/gpu_tests/ddp_test.py b/tests/gpu_tests/ddp_test.py index d310f53..7ad3e63 100644 --- a/tests/gpu_tests/ddp_test.py +++ b/tests/gpu_tests/ddp_test.py @@ -9,7 +9,8 @@ from torch.utils import data from kronfluence.analyzer import Analyzer, prepare_model -from kronfluence.arguments import FactorArguments, ScoreArguments +from kronfluence.utils.common.factor_arguments import pytest_factor_arguments +from kronfluence.utils.common.score_arguments import pytest_score_arguments from kronfluence.utils.constants import ( ALL_MODULE_NAME, COVARIANCE_FACTOR_NAMES, @@ -18,12 +19,12 @@ from kronfluence.utils.model import apply_ddp from tests.gpu_tests.pipeline import GpuTestTask, construct_test_mlp, get_mnist_dataset from tests.gpu_tests.prepare_tests import QUERY_INDICES, TRAIN_INDICES -from tests.utils import check_tensor_dict_equivalence +from tests.utils import ATOL, RTOL, check_tensor_dict_equivalence LOCAL_RANK = int(os.environ["LOCAL_RANK"]) WORLD_RANK = int(os.environ["RANK"]) WORLD_SIZE = int(os.environ["WORLD_SIZE"]) -logging.basicConfig(level=logging.DEBUG) +logging.basicConfig(level=logging.INFO) OLD_FACTOR_NAME = "single_gpu" NEW_FACTOR_NAME = "ddp" OLD_SCORE_NAME = "single_gpu" @@ -60,12 +61,7 @@ def setUpClass(cls) -> None: def test_covariance_matrices(self) -> None: covariance_factors = self.analyzer.load_covariance_matrices(factors_name=OLD_FACTOR_NAME) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, - ) + factor_args = pytest_factor_arguments() self.analyzer.fit_covariance_matrices( factors_name=NEW_FACTOR_NAME, dataset=self.train_dataset, @@ -85,18 +81,42 @@ def test_covariance_matrices(self) -> None: assert check_tensor_dict_equivalence( covariance_factors[name], new_covariance_factors[name], - atol=1e-5, - rtol=1e-3, + atol=ATOL, + rtol=RTOL, ) - def test_lambda_matrices(self): + def test_lambda_matrices(self) -> None: lambda_factors = self.analyzer.load_lambda_matrices(factors_name=OLD_FACTOR_NAME) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, + factor_args = pytest_factor_arguments() + self.analyzer.fit_lambda_matrices( + factors_name=NEW_FACTOR_NAME, + dataset=self.train_dataset, + factor_args=factor_args, + per_device_batch_size=512, + overwrite_output_dir=True, + load_from_factors_name=OLD_FACTOR_NAME, ) + new_lambda_factors = self.analyzer.load_lambda_matrices(factors_name=NEW_FACTOR_NAME) + + for name in LAMBDA_FACTOR_NAMES: + if LOCAL_RANK == 0: + for module_name in lambda_factors[name]: + print(f"Name: {name, module_name}") + print(f"Previous factor: {lambda_factors[name][module_name]}") + print(f"New factor: {new_lambda_factors[name][module_name]}") + if LOCAL_RANK == 0: + assert check_tensor_dict_equivalence( + lambda_factors[name], + new_lambda_factors[name], + atol=ATOL, + rtol=RTOL, + ) + + def test_lambda_partition_matrices(self) -> None: + lambda_factors = self.analyzer.load_lambda_matrices(factors_name=OLD_FACTOR_NAME) + factor_args = pytest_factor_arguments() + factor_args.lambda_module_partitions = 2 + factor_args.lambda_data_partitions = 2 self.analyzer.fit_lambda_matrices( factors_name=NEW_FACTOR_NAME, dataset=self.train_dataset, @@ -117,18 +137,14 @@ def test_lambda_matrices(self): assert check_tensor_dict_equivalence( lambda_factors[name], new_lambda_factors[name], - atol=1e-3, - rtol=1e-1, + atol=ATOL, + rtol=RTOL, ) def test_pairwise_scores(self) -> None: pairwise_scores = self.analyzer.load_pairwise_scores(scores_name=OLD_SCORE_NAME) - score_args = ScoreArguments( - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - ) + score_args = pytest_score_arguments() self.analyzer.compute_pairwise_scores( scores_name=NEW_SCORE_NAME, factors_name=OLD_FACTOR_NAME, @@ -155,18 +171,48 @@ def test_pairwise_scores(self) -> None: assert check_tensor_dict_equivalence( pairwise_scores, new_pairwise_scores, - atol=1e-5, - rtol=1e-3, + atol=ATOL, + rtol=RTOL, ) - def test_self_scores(self) -> None: - self_scores = self.analyzer.load_self_scores(scores_name=OLD_SCORE_NAME) + def test_pairwise_partition_scores(self) -> None: + pairwise_scores = self.analyzer.load_pairwise_scores(scores_name=OLD_SCORE_NAME) - score_args = ScoreArguments( - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, + score_args = pytest_score_arguments() + score_args.module_partitions = 2 + score_args.data_partitions = 2 + self.analyzer.compute_pairwise_scores( + scores_name=NEW_SCORE_NAME, + factors_name=OLD_FACTOR_NAME, + query_dataset=self.eval_dataset, + train_dataset=self.train_dataset, + train_indices=list(range(TRAIN_INDICES)), + query_indices=list(range(QUERY_INDICES)), + per_device_query_batch_size=12, + per_device_train_batch_size=512, + score_args=score_args, + overwrite_output_dir=True, ) + new_pairwise_scores = self.analyzer.load_pairwise_scores(scores_name=NEW_SCORE_NAME) + + if LOCAL_RANK == 0: + print(f"Previous score: {pairwise_scores[ALL_MODULE_NAME][0]}") + print(f"Previous shape: {pairwise_scores[ALL_MODULE_NAME].shape}") + print(f"New score: {new_pairwise_scores[ALL_MODULE_NAME][0]}") + print(f"New shape: {new_pairwise_scores[ALL_MODULE_NAME].shape}") + print(f"Previous score: {pairwise_scores[ALL_MODULE_NAME][50]}") + print(f"Previous shape: {pairwise_scores[ALL_MODULE_NAME].shape}") + print(f"New score: {new_pairwise_scores[ALL_MODULE_NAME][50]}") + print(f"New shape: {new_pairwise_scores[ALL_MODULE_NAME].shape}") + assert check_tensor_dict_equivalence( + pairwise_scores, + new_pairwise_scores, + atol=ATOL, + rtol=RTOL, + ) + + def test_self_scores(self) -> None: + score_args = pytest_score_arguments() self.analyzer.compute_self_scores( scores_name=NEW_SCORE_NAME, factors_name=OLD_FACTOR_NAME, @@ -177,6 +223,7 @@ def test_self_scores(self) -> None: overwrite_output_dir=True, ) new_self_scores = self.analyzer.load_self_scores(scores_name=NEW_SCORE_NAME) + self_scores = self.analyzer.load_self_scores(scores_name=OLD_SCORE_NAME) if LOCAL_RANK == 0: print(f"Previous score: {self_scores[ALL_MODULE_NAME]}") @@ -186,20 +233,38 @@ def test_self_scores(self) -> None: assert check_tensor_dict_equivalence( self_scores, new_self_scores, - atol=1e-5, - rtol=1e-3, + atol=ATOL, + rtol=RTOL, ) - def test_lr_pairwise_scores(self) -> None: - pairwise_scores = self.analyzer.load_pairwise_scores(scores_name="single_gpu_qb") - - score_args = ScoreArguments( - query_gradient_rank=32, - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - query_gradient_svd_dtype=torch.float64, + score_args.use_measurement_for_self_influence = True + self.analyzer.compute_self_scores( + scores_name=NEW_SCORE_NAME + "_measurement", + factors_name=OLD_FACTOR_NAME, + train_dataset=self.train_dataset, + train_indices=list(range(TRAIN_INDICES)), + per_device_train_batch_size=512, + score_args=score_args, + overwrite_output_dir=True, ) + new_self_scores = self.analyzer.load_self_scores(scores_name=NEW_SCORE_NAME + "_measurement") + self_scores = self.analyzer.load_self_scores(scores_name=OLD_SCORE_NAME + "_measurement") + + if LOCAL_RANK == 0: + print(f"Previous score: {self_scores[ALL_MODULE_NAME]}") + print(f"Previous shape: {self_scores[ALL_MODULE_NAME].shape}") + print(f"New score: {new_self_scores[ALL_MODULE_NAME]}") + print(f"New shape: {new_self_scores[ALL_MODULE_NAME].shape}") + assert check_tensor_dict_equivalence( + self_scores, + new_self_scores, + atol=ATOL, + rtol=RTOL, + ) + + def test_lr_pairwise_scores(self) -> None: + score_args = pytest_score_arguments() + score_args.query_gradient_low_rank = 32 self.analyzer.compute_pairwise_scores( scores_name="ddp_qb", factors_name=OLD_FACTOR_NAME, @@ -212,37 +277,85 @@ def test_lr_pairwise_scores(self) -> None: score_args=score_args, overwrite_output_dir=True, ) - new_pairwise_scores = self.analyzer.load_pairwise_scores(scores_name="ddp_qb") + + def test_per_module_pairwise_scores(self) -> None: + pairwise_scores = self.analyzer.load_pairwise_scores(scores_name=OLD_SCORE_NAME) + score_args = pytest_score_arguments() + score_args.compute_per_module_scores = True + self.analyzer.compute_pairwise_scores( + scores_name=NEW_SCORE_NAME + "_per_module", + factors_name=OLD_FACTOR_NAME, + query_dataset=self.eval_dataset, + train_dataset=self.train_dataset, + train_indices=list(range(TRAIN_INDICES)), + query_indices=list(range(QUERY_INDICES)), + per_device_query_batch_size=12, + per_device_train_batch_size=512, + score_args=score_args, + overwrite_output_dir=True, + ) + new_pairwise_scores = self.analyzer.load_pairwise_scores(scores_name=NEW_SCORE_NAME + "_per_module") if LOCAL_RANK == 0: - print(f"Previous score: {pairwise_scores[ALL_MODULE_NAME][0]}") - print(f"Previous shape: {pairwise_scores[ALL_MODULE_NAME].shape}") - print(f"New score: {new_pairwise_scores[ALL_MODULE_NAME][0]}") - print(f"New shape: {new_pairwise_scores[ALL_MODULE_NAME].shape}") - print(f"Previous score: {pairwise_scores[ALL_MODULE_NAME][50]}") - print(f"Previous shape: {pairwise_scores[ALL_MODULE_NAME].shape}") - print(f"New score: {new_pairwise_scores[ALL_MODULE_NAME][50]}") - print(f"New shape: {new_pairwise_scores[ALL_MODULE_NAME].shape}") + total_scores = None + for module_name in new_pairwise_scores: + if total_scores is None: + total_scores = new_pairwise_scores[module_name] + else: + total_scores.add_(new_pairwise_scores[module_name]) assert check_tensor_dict_equivalence( pairwise_scores, - new_pairwise_scores, - atol=1e-3, - rtol=1e-1, + {ALL_MODULE_NAME: total_scores}, + atol=ATOL, + rtol=RTOL, ) - def test_lr_aggregate_pairwise_scores(self) -> None: - pairwise_scores = self.analyzer.load_pairwise_scores(scores_name="single_gpu_qb") + def test_lr_accumulate_pairwise_scores(self) -> None: + score_args = pytest_score_arguments() + score_args.query_gradient_low_rank = 32 + score_args.query_gradient_accumulation_steps = 3 + self.analyzer.compute_pairwise_scores( + scores_name="ddp_qb_agg", + factors_name=OLD_FACTOR_NAME, + query_dataset=self.eval_dataset, + train_dataset=self.train_dataset, + train_indices=list(range(TRAIN_INDICES)), + query_indices=list(range(QUERY_INDICES)), + per_device_query_batch_size=2, + per_device_train_batch_size=512, + score_args=score_args, + overwrite_output_dir=True, + ) - score_args = ScoreArguments( - query_gradient_rank=32, - num_query_gradient_aggregations=3, - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - query_gradient_svd_dtype=torch.float64, + def test_aggregate_scores(self) -> None: + score_args = pytest_score_arguments() + score_args.aggregate_train_gradients = True + self.analyzer.compute_pairwise_scores( + scores_name="ddp", + factors_name=OLD_FACTOR_NAME, + query_dataset=self.eval_dataset, + train_dataset=self.train_dataset, + train_indices=list(range(TRAIN_INDICES)), + query_indices=list(range(QUERY_INDICES)), + per_device_query_batch_size=2, + per_device_train_batch_size=512, + score_args=score_args, + overwrite_output_dir=True, ) + new_pairwise_scores = self.analyzer.load_pairwise_scores(scores_name="ddp") + pairwise_scores = self.analyzer.load_pairwise_scores(scores_name="single_gpu_train_agg") + + if LOCAL_RANK == 0: + assert check_tensor_dict_equivalence( + pairwise_scores, + new_pairwise_scores, + atol=ATOL, + rtol=RTOL, + ) + + score_args.aggregate_query_gradients = True self.analyzer.compute_pairwise_scores( - scores_name="ddp_qb_agg", + scores_name="ddp", factors_name=OLD_FACTOR_NAME, query_dataset=self.eval_dataset, train_dataset=self.train_dataset, @@ -253,20 +366,15 @@ def test_lr_aggregate_pairwise_scores(self) -> None: score_args=score_args, overwrite_output_dir=True, ) - new_pairwise_scores = self.analyzer.load_pairwise_scores(scores_name="ddp_qb_agg") + new_pairwise_scores = self.analyzer.load_pairwise_scores(scores_name="ddp") + pairwise_scores = self.analyzer.load_pairwise_scores(scores_name="single_gpu_all_agg") if LOCAL_RANK == 0: - print(f"Previous score: {pairwise_scores[ALL_MODULE_NAME][0]}") - print(f"Previous shape: {pairwise_scores[ALL_MODULE_NAME].shape}") - print(f"New score: {new_pairwise_scores[ALL_MODULE_NAME][0]}") - print(f"New shape: {new_pairwise_scores[ALL_MODULE_NAME].shape}") - print(f"Previous score: {pairwise_scores[ALL_MODULE_NAME][44]}") - print(f"New score: {new_pairwise_scores[ALL_MODULE_NAME][44]}") assert check_tensor_dict_equivalence( pairwise_scores, new_pairwise_scores, - atol=1e-1, - rtol=1e-1, + atol=ATOL, + rtol=RTOL, ) @classmethod diff --git a/tests/gpu_tests/fsdp_test.py b/tests/gpu_tests/fsdp_test.py index 8c08975..43e627b 100644 --- a/tests/gpu_tests/fsdp_test.py +++ b/tests/gpu_tests/fsdp_test.py @@ -9,7 +9,8 @@ from torch.utils import data from kronfluence.analyzer import Analyzer, prepare_model -from kronfluence.arguments import FactorArguments, ScoreArguments +from kronfluence.utils.common.factor_arguments import pytest_factor_arguments +from kronfluence.utils.common.score_arguments import pytest_score_arguments from kronfluence.utils.constants import ( ALL_MODULE_NAME, COVARIANCE_FACTOR_NAMES, @@ -18,7 +19,7 @@ from kronfluence.utils.model import apply_fsdp from tests.gpu_tests.pipeline import GpuTestTask, construct_test_mlp, get_mnist_dataset from tests.gpu_tests.prepare_tests import QUERY_INDICES, TRAIN_INDICES -from tests.utils import check_tensor_dict_equivalence +from tests.utils import ATOL, RTOL, check_tensor_dict_equivalence LOCAL_RANK = int(os.environ["LOCAL_RANK"]) WORLD_RANK = int(os.environ["RANK"]) @@ -62,12 +63,7 @@ def setUpClass(cls) -> None: def test_covariance_matrices(self) -> None: covariance_factors = self.analyzer.load_covariance_matrices(factors_name=OLD_FACTOR_NAME) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, - ) + factor_args = pytest_factor_arguments() self.analyzer.fit_covariance_matrices( factors_name=NEW_FACTOR_NAME, dataset=self.train_dataset, @@ -87,18 +83,13 @@ def test_covariance_matrices(self) -> None: assert check_tensor_dict_equivalence( covariance_factors[name], new_covariance_factors[name], - atol=1e-5, - rtol=1e-3, + atol=ATOL, + rtol=RTOL, ) - def test_lambda_matrices(self): + def test_lambda_matrices(self) -> None: lambda_factors = self.analyzer.load_lambda_matrices(factors_name=OLD_FACTOR_NAME) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, - ) + factor_args = pytest_factor_arguments() self.analyzer.fit_lambda_matrices( factors_name=NEW_FACTOR_NAME, dataset=self.train_dataset, @@ -119,18 +110,14 @@ def test_lambda_matrices(self): assert check_tensor_dict_equivalence( lambda_factors[name], new_lambda_factors[name], - atol=1e-3, - rtol=1e-1, + atol=ATOL, + rtol=RTOL, ) def test_pairwise_scores(self) -> None: pairwise_scores = self.analyzer.load_pairwise_scores(scores_name=OLD_SCORE_NAME) - score_args = ScoreArguments( - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - ) + score_args = pytest_score_arguments() self.analyzer.compute_pairwise_scores( scores_name=NEW_SCORE_NAME, factors_name=OLD_FACTOR_NAME, @@ -153,18 +140,14 @@ def test_pairwise_scores(self) -> None: assert check_tensor_dict_equivalence( pairwise_scores, new_pairwise_scores, - atol=1e-5, - rtol=1e-3, + atol=ATOL, + rtol=RTOL, ) def test_self_scores(self) -> None: self_scores = self.analyzer.load_self_scores(scores_name=OLD_SCORE_NAME) - score_args = ScoreArguments( - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - ) + score_args = pytest_score_arguments() self.analyzer.compute_self_scores( scores_name=NEW_SCORE_NAME, factors_name=OLD_FACTOR_NAME, @@ -184,8 +167,8 @@ def test_self_scores(self) -> None: assert check_tensor_dict_equivalence( self_scores, new_self_scores, - atol=1e-5, - rtol=1e-3, + atol=ATOL, + rtol=RTOL, ) @classmethod diff --git a/tests/gpu_tests/pipeline.py b/tests/gpu_tests/pipeline.py index 117ec52..76f542d 100644 --- a/tests/gpu_tests/pipeline.py +++ b/tests/gpu_tests/pipeline.py @@ -25,12 +25,12 @@ def compute_train_loss( if not sample: return F.cross_entropy(logits, labels, reduction="sum") with torch.no_grad(): - probs = torch.nn.functional.softmax(logits, dim=-1) + probs = torch.nn.functional.softmax(logits.detach(), dim=-1) sampled_labels = torch.multinomial( probs, num_samples=1, ).flatten() - return F.cross_entropy(logits, sampled_labels.detach(), reduction="sum") + return F.cross_entropy(logits, sampled_labels, reduction="sum") def compute_measurement( self, diff --git a/tests/gpu_tests/prepare_tests.py b/tests/gpu_tests/prepare_tests.py index 1392edf..a9a140b 100644 --- a/tests/gpu_tests/prepare_tests.py +++ b/tests/gpu_tests/prepare_tests.py @@ -6,7 +6,8 @@ from tqdm import tqdm from kronfluence.analyzer import Analyzer, prepare_model -from kronfluence.arguments import FactorArguments, ScoreArguments +from kronfluence.utils.common.factor_arguments import pytest_factor_arguments +from kronfluence.utils.common.score_arguments import pytest_score_arguments from tests.gpu_tests.pipeline import GpuTestTask, construct_test_mlp, get_mnist_dataset # Pick difficult cases where the dataset is not perfectly divisible by batch size. @@ -92,12 +93,7 @@ def run_analysis() -> None: task=task, ) - factor_args = FactorArguments( - use_empirical_fisher=True, - activation_covariance_dtype=torch.float64, - gradient_covariance_dtype=torch.float64, - lambda_dtype=torch.float64, - ) + factor_args = pytest_factor_arguments() analyzer.fit_all_factors( factors_name="single_gpu", dataset=train_dataset, @@ -106,11 +102,7 @@ def run_analysis() -> None: overwrite_output_dir=True, ) - score_args = ScoreArguments( - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - ) + score_args = pytest_score_arguments() analyzer.compute_pairwise_scores( scores_name="single_gpu", factors_name="single_gpu", @@ -130,13 +122,19 @@ def run_analysis() -> None: overwrite_output_dir=True, ) - score_args = ScoreArguments( - query_gradient_rank=32, - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - query_gradient_svd_dtype=torch.float64, + score_args = pytest_score_arguments() + score_args.use_measurement_for_self_influence = True + analyzer.compute_self_scores( + scores_name="single_gpu_measurement", + factors_name="single_gpu", + train_dataset=train_dataset, + per_device_train_batch_size=512, + score_args=score_args, + overwrite_output_dir=True, ) + + score_args = pytest_score_arguments() + score_args.query_gradient_low_rank = 32 analyzer.compute_pairwise_scores( scores_name="single_gpu_qb", factors_name="single_gpu", @@ -148,6 +146,46 @@ def run_analysis() -> None: overwrite_output_dir=True, ) + score_args = pytest_score_arguments() + score_args.aggregate_train_gradients = True + analyzer.compute_pairwise_scores( + scores_name="single_gpu_train_agg", + factors_name="single_gpu", + query_dataset=eval_dataset, + train_dataset=train_dataset, + per_device_query_batch_size=12, + per_device_train_batch_size=512, + score_args=score_args, + overwrite_output_dir=True, + ) + + score_args = pytest_score_arguments() + score_args.aggregate_query_gradients = True + analyzer.compute_pairwise_scores( + scores_name="single_gpu_query_agg", + factors_name="single_gpu", + query_dataset=eval_dataset, + train_dataset=train_dataset, + per_device_query_batch_size=12, + per_device_train_batch_size=512, + score_args=score_args, + overwrite_output_dir=True, + ) + + score_args = pytest_score_arguments() + score_args.aggregate_train_gradients = True + score_args.aggregate_query_gradients = True + analyzer.compute_pairwise_scores( + scores_name="single_gpu_all_agg", + factors_name="single_gpu", + query_dataset=eval_dataset, + train_dataset=train_dataset, + per_device_query_batch_size=12, + per_device_train_batch_size=512, + score_args=score_args, + overwrite_output_dir=True, + ) + if __name__ == "__main__": train() diff --git a/tests/gpu_tests/test_offload_cpu.py b/tests/gpu_tests/test_offload_cpu.py index 655ef68..fca8ef7 100644 --- a/tests/gpu_tests/test_offload_cpu.py +++ b/tests/gpu_tests/test_offload_cpu.py @@ -1,27 +1,16 @@ # pylint: skip-file -from typing import Tuple import pytest -from torch import nn +import torch from kronfluence.analyzer import Analyzer, prepare_model from kronfluence.arguments import FactorArguments, ScoreArguments -from kronfluence.task import Task +from kronfluence.utils.common.factor_arguments import pytest_factor_arguments +from kronfluence.utils.common.score_arguments import pytest_score_arguments from kronfluence.utils.constants import ALL_MODULE_NAME from kronfluence.utils.dataset import DataLoaderKwargs -from tests.utils import prepare_test - - -def prepare_model_and_analyzer(model: nn.Module, task: Task) -> Tuple[nn.Module, Analyzer]: - model = prepare_model(model=model, task=task) - analyzer = Analyzer( - analysis_name=f"pytest_{__name__}", - model=model, - task=task, - disable_model_save=True, - ) - return model, analyzer +from tests.utils import ATOL, RTOL, check_tensor_dict_equivalence, prepare_test @pytest.mark.parametrize( @@ -29,19 +18,20 @@ def prepare_model_and_analyzer(model: nn.Module, task: Task) -> Tuple[nn.Module, [ "mlp", "repeated_mlp", + "mlp_checkpoint", "conv", "conv_bn", "bert", "gpt", ], ) -@pytest.mark.parametrize("cached_activation_cpu_offload", [True, False]) +@pytest.mark.parametrize("offload_activations_to_cpu", [True, False]) @pytest.mark.parametrize("query_size", [16]) @pytest.mark.parametrize("train_size", [32]) -@pytest.mark.parametrize("seed", [0]) +@pytest.mark.parametrize("seed", [1]) def test_cpu_offloads( test_name: str, - cached_activation_cpu_offload: bool, + offload_activations_to_cpu: bool, query_size: int, train_size: int, seed: int, @@ -53,13 +43,19 @@ def test_cpu_offloads( seed=seed, ) kwargs = DataLoaderKwargs(collate_fn=data_collator) - model, analyzer = prepare_model_and_analyzer( + model = prepare_model(model=model, task=task) + analyzer = Analyzer( + analysis_name=f"pytest_{__name__}", model=model, task=task, + disable_model_save=True, + disable_tqdm=True, ) factor_args = FactorArguments( - cached_activation_cpu_offload=cached_activation_cpu_offload, + offload_activations_to_cpu=offload_activations_to_cpu, ) + if test_name == "repeated_mlp": + factor_args.has_shared_parameters = True factors_name = f"pytest_{test_name}_{test_cpu_offloads.__name__}" analyzer.fit_all_factors( factors_name=factors_name, @@ -71,7 +67,7 @@ def test_cpu_offloads( ) score_args = ScoreArguments( - cached_activation_cpu_offload=cached_activation_cpu_offload, + offload_activations_to_cpu=offload_activations_to_cpu, ) scores_name = f"pytest_{test_name}_{test_cpu_offloads.__name__}_scores" analyzer.compute_pairwise_scores( @@ -89,3 +85,95 @@ def test_cpu_offloads( pairwise_scores = analyzer.load_pairwise_scores(scores_name=scores_name) assert pairwise_scores[ALL_MODULE_NAME].size(0) == query_size assert pairwise_scores[ALL_MODULE_NAME].size(1) == train_size + + +@pytest.mark.parametrize( + "test_name", + [ + "mlp", + "repeated_mlp", + "mlp_checkpoint", + "conv", + ], +) +@pytest.mark.parametrize("per_module_score", [False, True]) +@pytest.mark.parametrize("query_size", [50]) +@pytest.mark.parametrize("train_size", [102]) +@pytest.mark.parametrize("seed", [1]) +def test_cpu_offloads_identical( + test_name: str, + per_module_score: bool, + query_size: int, + train_size: int, + seed: int, +) -> None: + model, train_dataset, test_dataset, data_collator, task = prepare_test( + test_name=test_name, + query_size=query_size, + train_size=train_size, + seed=seed, + ) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + model = model.to(dtype=torch.float64) + model = prepare_model(model=model, task=task) + + analyzer = Analyzer( + analysis_name=f"pytest_{test_cpu_offloads_identical}_{__name__}", + model=model, + task=task, + disable_model_save=True, + disable_tqdm=True, + ) + factor_args = pytest_factor_arguments() + if test_name == "repeated_mlp": + factor_args.has_shared_parameters = True + factors_name = f"pytest_{test_name}_{test_cpu_offloads_identical.__name__}" + analyzer.fit_all_factors( + factors_name=factors_name, + dataset=train_dataset, + dataloader_kwargs=kwargs, + per_device_batch_size=32, + factor_args=factor_args, + overwrite_output_dir=True, + ) + score_args = pytest_score_arguments() + scores_name = f"pytest_{test_name}_{test_cpu_offloads_identical.__name__}_scores" + analyzer.compute_pairwise_scores( + scores_name=scores_name, + factors_name=factors_name, + query_dataset=test_dataset, + per_device_query_batch_size=4, + train_dataset=train_dataset, + per_device_train_batch_size=8, + dataloader_kwargs=kwargs, + score_args=score_args, + overwrite_output_dir=True, + ) + pairwise_scores = analyzer.load_pairwise_scores(scores_name=scores_name) + + factors_name = f"pytest_{test_name}_{test_cpu_offloads_identical.__name__}_cached" + factor_args.offload_activations_to_cpu = True + analyzer.fit_all_factors( + factors_name=factors_name, + dataset=train_dataset, + dataloader_kwargs=kwargs, + per_device_batch_size=16, + factor_args=factor_args, + overwrite_output_dir=True, + ) + score_args.offload_activations_to_cpu = True + scores_name = f"pytest_{test_name}_{test_cpu_offloads_identical.__name__}_cached_scores" + analyzer.compute_pairwise_scores( + scores_name=scores_name, + factors_name=factors_name, + query_dataset=test_dataset, + per_device_query_batch_size=6, + train_dataset=train_dataset, + per_device_train_batch_size=8, + dataloader_kwargs=kwargs, + score_args=score_args, + overwrite_output_dir=True, + ) + cached_pairwise_scores = analyzer.load_pairwise_scores(scores_name=scores_name) + + assert check_tensor_dict_equivalence(pairwise_scores, cached_pairwise_scores, atol=ATOL, rtol=RTOL) diff --git a/tests/modules/test_matmul.py b/tests/modules/test_matmul.py new file mode 100644 index 0000000..aceb943 --- /dev/null +++ b/tests/modules/test_matmul.py @@ -0,0 +1,237 @@ +import time + +import opt_einsum +import pytest +import torch +from accelerate.utils import set_seed +from opt_einsum import DynamicProgramming + + +def test_query_gradient_svd( + seed: int = 0, +) -> None: + input_dim = 2048 + output_dim = 1024 + batch_dim = 16 + set_seed(seed) + + gradient = torch.rand(size=(batch_dim, output_dim, input_dim), dtype=torch.float64) + + U, S, V = torch.linalg.svd( + gradient.contiguous(), + full_matrices=False, + ) + assert torch.allclose(gradient, U @ torch.diag_embed(S) @ V, atol=1e-5, rtol=1e-3) + + rank = 32 + U_k = U[:, :, :rank] + S_k = S[:, :rank] + V_k = V[:, :rank, :].clone() + left, right = torch.matmul(U_k, torch.diag_embed(S_k)).contiguous(), V_k.contiguous() + assert torch.bmm(left, right).shape == gradient.shape + + rank = input_dim + U, S, V = torch.linalg.svd( + gradient.contiguous(), + full_matrices=False, + ) + U_k = U[:, :, :rank] + S_k = S[:, :rank] + V_k = V[:, :rank, :].clone() + left, right = torch.matmul(U_k, torch.diag_embed(S_k)).contiguous(), V_k.contiguous() + assert torch.allclose(torch.bmm(left, right), gradient, atol=1e-5, rtol=1e-3) + + rank = 32 + lr_gradient1 = torch.rand(size=(batch_dim, output_dim, rank), dtype=torch.float64) + lr_gradient2 = torch.rand(size=(batch_dim, rank, input_dim), dtype=torch.float64) + gradient = torch.bmm(lr_gradient1, lr_gradient2) + U, S, V = torch.linalg.svd( + gradient.contiguous(), + full_matrices=False, + ) + U_k = U[:, :, :rank] + S_k = S[:, :rank] + V_k = V[:, :rank, :].clone() + left_mat, right_mat = torch.matmul(U_k, torch.diag_embed(S_k)).contiguous(), V_k.contiguous() + assert torch.allclose(torch.bmm(left_mat, right_mat), gradient, atol=1e-5, rtol=1e-3) + + query_batch_dim = 32 + new_gradient = torch.rand(size=(query_batch_dim, output_dim, input_dim), dtype=torch.float64) + score = opt_einsum.contract("toi,qoi->tq", gradient, new_gradient) + + lr_score = opt_einsum.contract("qki,toi,qok->qt", right_mat, new_gradient, left_mat) + assert torch.allclose(score, lr_score) + + lr_score_reconst_matmul = torch.matmul( + torch.matmul(left_mat, right_mat).view(left_mat.size(0), -1), new_gradient.view(new_gradient.shape[0], -1).t() + ) + assert torch.allclose(score, lr_score_reconst_matmul) + + # These should be able to avoid explicit reconstruction. This should be used when input_dim > output_dim. + intermediate = opt_einsum.contract("qki,toi->qtko", right_mat, new_gradient) + final = opt_einsum.contract("qtko,qok->qt", intermediate, left_mat) + assert torch.allclose(score, final) + print("Option 1") + print(intermediate.numel()) + + # This should be used when output_dim > input_dim. + intermediate2 = torch.einsum("toi,qok->qtik", new_gradient, left_mat) + final2 = opt_einsum.contract("qki,qtik->qt", right_mat, intermediate2) + assert torch.allclose(score, final2) + print("Option 2") + print(intermediate2.numel()) + + print("Reconstruction") + print((torch.matmul(left_mat, right_mat).view(left_mat.size(0), -1)).numel()) + path = opt_einsum.contract_path("qki,toi,qok->qt", right_mat, new_gradient, left_mat) + print(path) + + +@pytest.mark.parametrize("input_dim", [256, 512]) +@pytest.mark.parametrize("output_dim", [512, 1024]) +@pytest.mark.parametrize("batch_dim", [8, 16]) +@pytest.mark.parametrize("qbatch_dim", [8, 16]) +@pytest.mark.parametrize("rank", [32]) +@pytest.mark.parametrize("seed", [0]) +def test_query_gradient_svd_reconst( + input_dim: int, + output_dim: int, + batch_dim: int, + qbatch_dim: int, + rank: int, + seed: int, +) -> None: + set_seed(seed) + + lr_gradient1 = torch.rand(size=(batch_dim, output_dim, rank + 50), dtype=torch.float64) + lr_gradient2 = torch.rand(size=(batch_dim, rank + 50, input_dim), dtype=torch.float64) + gradient = torch.bmm(lr_gradient1, lr_gradient2) + U, S, V = torch.linalg.svd( + gradient.contiguous(), + full_matrices=False, + ) + U_k = U[:, :, :rank] + S_k = S[:, :rank] + V_k = V[:, :rank, :].clone() + left_mat, right_mat = torch.matmul(U_k, torch.diag_embed(S_k)).contiguous(), V_k.contiguous() + new_gradient = torch.rand(size=(qbatch_dim, output_dim, input_dim), dtype=torch.float64) + + lr_score = opt_einsum.contract("qki,toi,qok->qt", right_mat, new_gradient, left_mat) + lr_score_reconst_matmul = torch.matmul( + torch.matmul(left_mat, right_mat).view(left_mat.size(0), -1), new_gradient.view(new_gradient.shape[0], -1).t() + ) + assert torch.allclose(lr_score, lr_score_reconst_matmul) + + # This should be used when input_dim > output_dim. + intermediate = opt_einsum.contract("qki,toi->qtko", right_mat, new_gradient) + final = opt_einsum.contract("qtko,qok->qt", intermediate, left_mat) + assert torch.allclose(lr_score, final) + print("Option 1") + print(intermediate.numel()) + + # This should be used when output_dim > input_dim. + intermediate2 = torch.einsum("toi,qok->qtik", new_gradient, left_mat) + final2 = opt_einsum.contract("qki,qtik->qt", right_mat, intermediate2) + assert torch.allclose(lr_score, final2) + print("Option 2") + print(intermediate2.numel()) + + print("Reconstruction") + reconst_numel = (torch.matmul(left_mat, right_mat).view(left_mat.size(0), -1)).numel() + print(reconst_numel) + path = opt_einsum.contract_path("qki,toi,qok->qt", right_mat, new_gradient, left_mat) + print(path) + + if new_gradient.size(0) * right_mat.size(0) * rank * min((right_mat.size(2), left_mat.size(1))) > right_mat.size( + 0 + ) * right_mat.size(2) * left_mat.size(1): + assert intermediate2.numel() > reconst_numel and intermediate.numel() > reconst_numel + elif output_dim >= input_dim: + assert intermediate2.numel() <= reconst_numel + else: + assert intermediate.numel() <= reconst_numel + + +def test_compute_score_matmul( + seed: int = 0, +) -> None: + input_dim = 4096 + output_dim = 100 + token_dim = 1 + batch_dim = 1024 + query_batch_dim = 2 + set_seed(seed) + + input_activation = torch.rand(size=(batch_dim, token_dim, input_dim), dtype=torch.float64) + output_gradient = torch.rand(size=(batch_dim, token_dim, output_dim), dtype=torch.float64) + gradient = opt_einsum.contract("b...i,b...o->bio", output_gradient, input_activation) + new_gradient = torch.rand(size=(query_batch_dim, output_dim, input_dim), dtype=torch.float64) + + score = opt_einsum.contract("toi,qoi->tq", gradient, new_gradient) + path = opt_einsum.contract_path("toi,qoi->tq", gradient, new_gradient) + print(path) + + unsqueeze_score = opt_einsum.contract("t...,q...->tq", gradient, new_gradient) + assert torch.allclose(score, unsqueeze_score) + + path = opt_einsum.contract_path( + "bti,bto,qio->qb", + output_gradient, + input_activation, + new_gradient, + optimize=DynamicProgramming(search_outer=True, minimize="flops"), + ) + print(path) + + +def test_precondition_gradient( + seed: int = 0, +) -> None: + input_dim = 128 + output_dim = 256 + batch_dim = 8 + lambda_scale = 1000 + damping = 1e-08 + + set_seed(seed) + A = torch.rand(size=(input_dim, input_dim), dtype=torch.float64) + B = torch.rand(size=(output_dim, output_dim), dtype=torch.float64) + Lambda = torch.rand(size=(output_dim, input_dim), dtype=torch.float64) + gradient = torch.rand(size=(batch_dim, output_dim, input_dim), dtype=torch.float64) + + start_time = time.time() + rotated_gradient = torch.einsum( + "ij,bjl,lk->bik", + ( + B.t(), + gradient, + A, + ), + ) + rotated_gradient.div_(Lambda + damping) + results = lambda_scale * torch.einsum( + "ij,bjl,lk->bik", + (B, rotated_gradient, A.t()), + ) + print(f"Took {time.time() - start_time} seconds.") + + start_time = time.time() + grads_rot = torch.matmul( + B.t(), + torch.matmul( + gradient, + A, + ), + ) + scaled_lambda = Lambda / lambda_scale + grads_rot.div_(scaled_lambda) + raw_results = torch.matmul( + B, + torch.matmul( + grads_rot, + A.t(), + ), + ) + print(f"Took {time.time() - start_time} seconds.") + + assert torch.allclose(raw_results, results, atol=1e-5, rtol=1e-3) diff --git a/tests/modules/test_modules.py b/tests/modules/test_modules.py index 25dde9a..7e60489 100644 --- a/tests/modules/test_modules.py +++ b/tests/modules/test_modules.py @@ -9,13 +9,12 @@ from kronfluence.arguments import FactorArguments from kronfluence.module.tracked_module import ModuleMode from kronfluence.module.utils import set_mode, wrap_tracked_modules -from kronfluence.utils.save import verify_models_equivalence from tests.utils import prepare_test @pytest.mark.parametrize( "test_name", - ["mlp", "conv", "conv_bn", "bert", "gpt"], + ["mlp", "conv_bn", "gpt"], ) @pytest.mark.parametrize( "mode", @@ -24,6 +23,7 @@ ModuleMode.COVARIANCE, ModuleMode.LAMBDA, ModuleMode.PRECONDITION_GRADIENT, + ModuleMode.GRADIENT_AGGREGATION, ], ) @pytest.mark.parametrize("train_size", [32]) @@ -34,6 +34,7 @@ def test_tracked_modules_forward_equivalence( train_size: int, seed: int, ) -> None: + # The forward pass should produce the same results with and without wrapped modules. model, train_dataset, _, data_collator, task = prepare_test( test_name=test_name, train_size=train_size, @@ -82,16 +83,18 @@ def test_tracked_modules_forward_equivalence( ModuleMode.COVARIANCE, ModuleMode.LAMBDA, ModuleMode.PRECONDITION_GRADIENT, + ModuleMode.GRADIENT_AGGREGATION, ], ) @pytest.mark.parametrize("train_size", [32]) -@pytest.mark.parametrize("seed", [0]) +@pytest.mark.parametrize("seed", [1]) def test_tracked_modules_backward_equivalence( test_name: str, mode: ModuleMode, train_size: int, seed: int, ) -> None: + # The backward pass should produce the same results with and without wrapped modules. model, train_dataset, _, data_collator, task = prepare_test( test_name=test_name, train_size=train_size, @@ -125,30 +128,10 @@ def test_tracked_modules_backward_equivalence( wrapped_loss = task.compute_train_loss(batch, wrapped_model, sample=False) wrapped_loss.backward() for name, param in wrapped_model.named_parameters(): - wrapped_grads[name] = param.grad.detach() + if param.grad is not None: + wrapped_grads[name] = param.grad.detach() for name, grad in wrapped_grads.items(): original_name = name.replace(".original_module", "") if original_name in original_grads: assert torch.allclose(grad, original_grads[original_name]) - - -def test_verify_models_equivalence() -> None: - model1, _, _, _, _ = prepare_test( - test_name="mlp", - train_size=10, - seed=0, - ) - model2, _, _, _, _ = prepare_test( - test_name="mlp", - train_size=10, - seed=1, - ) - model3, _, _, _, _ = prepare_test( - test_name="conv", - train_size=10, - seed=1, - ) - assert verify_models_equivalence(model1.state_dict(), model1.state_dict()) - assert not verify_models_equivalence(model1.state_dict(), model2.state_dict()) - assert not verify_models_equivalence(model1.state_dict(), model3.state_dict()) diff --git a/tests/modules/test_per_sample_gradients.py b/tests/modules/test_per_sample_gradients.py index dde6b27..84a9481 100644 --- a/tests/modules/test_per_sample_gradients.py +++ b/tests/modules/test_per_sample_gradients.py @@ -4,7 +4,6 @@ import time from typing import Any, Dict, List -import opt_einsum import pytest import torch from accelerate.utils import find_batch_size, set_seed @@ -14,7 +13,12 @@ from kronfluence.analyzer import Analyzer, prepare_model from kronfluence.arguments import FactorArguments from kronfluence.module.tracked_module import ModuleMode, TrackedModule -from kronfluence.module.utils import set_mode, update_factor_args +from kronfluence.module.utils import ( + finalize_iteration, + get_tracked_module_names, + set_mode, + update_factor_args, +) from kronfluence.task import Task from kronfluence.utils.constants import LAMBDA_MATRIX_NAME, PRECONDITIONED_GRADIENT_NAME from kronfluence.utils.dataset import DataLoaderKwargs @@ -95,6 +99,7 @@ def for_loop_per_sample_gradient( "conv", "conv_bn", "bert", + "roberta", "gpt", ], ) @@ -114,7 +119,7 @@ def test_for_loop_per_sample_gradient_equivalence( ) original_model = copy.deepcopy(model) - batch_size = 4 + batch_size = 3 num_batches = train_size // batch_size train_loader = DataLoader( train_dataset, @@ -133,7 +138,10 @@ def test_for_loop_per_sample_gradient_equivalence( factor_args = FactorArguments( strategy="identity", ) + if test_name == "repeated_mlp": + factor_args.has_shared_parameters = True update_factor_args(model=model, factor_args=factor_args) + tracked_modules_names = get_tracked_module_names(model=model) per_sample_gradients = [] set_mode(model, ModuleMode.PRECONDITION_GRADIENT) @@ -148,6 +156,9 @@ def test_for_loop_per_sample_gradient_equivalence( loss = task.compute_train_loss(batch=batch_lst[i], model=model, sample=False) loss.backward() + if test_name == "repeated_mlp": + finalize_iteration(model=model, tracked_module_names=tracked_modules_names) + module_gradients = {} for module in model.modules(): if isinstance(module, TrackedModule): @@ -182,6 +193,7 @@ def test_for_loop_per_sample_gradient_equivalence( "conv", "conv_bn", "bert", + "roberta", "gpt", ], ) @@ -220,7 +232,10 @@ def test_mean_gradient_equivalence( factor_args = FactorArguments( strategy="identity", ) + if test_name == "repeated_mlp": + factor_args.has_shared_parameters = True update_factor_args(model=model, factor_args=factor_args) + tracked_modules_names = get_tracked_module_names(model=model) per_sample_gradients = [] set_mode(model, ModuleMode.PRECONDITION_GRADIENT) @@ -235,6 +250,9 @@ def test_mean_gradient_equivalence( loss = task.compute_train_loss(batch=batch_lst[i], model=model, sample=False) loss.backward() + if test_name == "repeated_mlp": + finalize_iteration(model=model, tracked_module_names=tracked_modules_names) + module_gradients = {} for module in model.modules(): if isinstance(module, TrackedModule): @@ -292,7 +310,7 @@ def test_mean_gradient_equivalence( "mlp", "repeated_mlp", "conv", - "gpt", + "roberta", ], ) @pytest.mark.parametrize("train_size", [32]) @@ -331,12 +349,15 @@ def test_lambda_equivalence( task=task, disable_model_save=True, cpu=True, + disable_tqdm=True, ) kwargs = DataLoaderKwargs(collate_fn=data_collator) factor_args = FactorArguments( strategy="diagonal", use_empirical_fisher=True, ) + if test_name == "repeated_mlp": + factor_args.has_shared_parameters = True analyzer.fit_lambda_matrices( factors_name=f"pytest_{test_name}_lambda_diag", dataset=train_dataset, @@ -372,224 +393,3 @@ def test_lambda_equivalence( atol=ATOL, rtol=RTOL, ) - - -def test_precondition_gradient( - seed: int = 0, -) -> None: - input_dim = 128 - output_dim = 256 - batch_dim = 8 - lambda_scale = 1000 - damping = 1e-08 - - set_seed(seed) - A = torch.rand(size=(input_dim, input_dim), dtype=torch.float64) - B = torch.rand(size=(output_dim, output_dim), dtype=torch.float64) - Lambda = torch.rand(size=(output_dim, input_dim), dtype=torch.float64) - gradient = torch.rand(size=(batch_dim, output_dim, input_dim), dtype=torch.float64) - - start_time = time.time() - rotated_gradient = torch.einsum( - "ij,bjl,lk->bik", - ( - B.t(), - gradient, - A, - ), - ) - rotated_gradient.div_(Lambda + damping) - results = lambda_scale * torch.einsum( - "ij,bjl,lk->bik", - (B, rotated_gradient, A.t()), - ) - print(f"Took {time.time() - start_time} seconds.") - - start_time = time.time() - grads_rot = torch.matmul( - B.t(), - torch.matmul( - gradient, - A, - ), - ) - scaled_lambda = Lambda / lambda_scale - grads_rot.div_(scaled_lambda) - raw_results = torch.matmul( - B, - torch.matmul( - grads_rot, - A.t(), - ), - ) - print(f"Took {time.time() - start_time} seconds.") - - assert torch.allclose(raw_results, results, atol=1e-5, rtol=1e-3) - - -def test_query_gradient_svd( - seed: int = 0, -) -> None: - input_dim = 2048 - output_dim = 1024 - batch_dim = 16 - set_seed(seed) - - gradient = torch.rand(size=(batch_dim, output_dim, input_dim), dtype=torch.float64) - - U, S, V = torch.linalg.svd( - gradient.contiguous(), - full_matrices=False, - ) - assert torch.allclose(gradient, U @ torch.diag_embed(S) @ V, atol=1e-5, rtol=1e-3) - - rank = 32 - U_k = U[:, :, :rank] - S_k = S[:, :rank] - V_k = V[:, :rank, :].clone() - left, right = torch.matmul(U_k, torch.diag_embed(S_k)).contiguous(), V_k.contiguous() - assert torch.bmm(left, right).shape == gradient.shape - - rank = input_dim - U, S, V = torch.linalg.svd( - gradient.contiguous(), - full_matrices=False, - ) - U_k = U[:, :, :rank] - S_k = S[:, :rank] - V_k = V[:, :rank, :].clone() - left, right = torch.matmul(U_k, torch.diag_embed(S_k)).contiguous(), V_k.contiguous() - assert torch.allclose(torch.bmm(left, right), gradient, atol=1e-5, rtol=1e-3) - - rank = 32 - lr_gradient1 = torch.rand(size=(batch_dim, output_dim, rank), dtype=torch.float64) - lr_gradient2 = torch.rand(size=(batch_dim, rank, input_dim), dtype=torch.float64) - gradient = torch.bmm(lr_gradient1, lr_gradient2) - U, S, V = torch.linalg.svd( - gradient.contiguous(), - full_matrices=False, - ) - U_k = U[:, :, :rank] - S_k = S[:, :rank] - V_k = V[:, :rank, :].clone() - left_mat, right_mat = torch.matmul(U_k, torch.diag_embed(S_k)).contiguous(), V_k.contiguous() - assert torch.allclose(torch.bmm(left_mat, right_mat), gradient, atol=1e-5, rtol=1e-3) - - query_batch_dim = 32 - new_gradient = torch.rand(size=(query_batch_dim, output_dim, input_dim), dtype=torch.float64) - score = opt_einsum.contract("toi,qoi->tq", gradient, new_gradient) - - lr_score = opt_einsum.contract("qki,toi,qok->qt", right_mat, new_gradient, left_mat) - assert torch.allclose(score, lr_score) - - lr_score_reconst_matmul = torch.matmul( - torch.matmul(left_mat, right_mat).view(left_mat.size(0), -1), new_gradient.view(new_gradient.shape[0], -1).t() - ) - assert torch.allclose(score, lr_score_reconst_matmul) - - # These should be able to avoid explicit reconstruction. - # This should be used when input_dim > output_dim. - intermediate = opt_einsum.contract("qki,toi->qtko", right_mat, new_gradient) - final = opt_einsum.contract("qtko,qok->qt", intermediate, left_mat) - assert torch.allclose(score, final) - print("Option 1") - print(intermediate.numel()) - - # This should be used when output_dim > input_dim. - intermediate2 = torch.einsum("toi,qok->qtik", new_gradient, left_mat) - final2 = opt_einsum.contract("qki,qtik->qt", right_mat, intermediate2) - assert torch.allclose(score, final2) - print("Option 2") - print(intermediate2.numel()) - - print("Reconstruction") - print((torch.matmul(left_mat, right_mat).view(left_mat.size(0), -1)).numel()) - path = opt_einsum.contract_path("qki,toi,qok->qt", right_mat, new_gradient, left_mat) - print(path) - - -@pytest.mark.parametrize("input_dim", [256, 512]) -@pytest.mark.parametrize("output_dim", [512, 1024]) -@pytest.mark.parametrize("batch_dim", [8, 16]) -@pytest.mark.parametrize("qbatch_dim", [8, 16]) -@pytest.mark.parametrize("rank", [32]) -@pytest.mark.parametrize("seed", [0]) -def test_query_gradient_svd_reconst( - input_dim: int, - output_dim: int, - batch_dim: int, - qbatch_dim: int, - rank: int, - seed: int, -) -> None: - set_seed(seed) - - lr_gradient1 = torch.rand(size=(batch_dim, output_dim, rank + 50), dtype=torch.float64) - lr_gradient2 = torch.rand(size=(batch_dim, rank + 50, input_dim), dtype=torch.float64) - gradient = torch.bmm(lr_gradient1, lr_gradient2) - U, S, V = torch.linalg.svd( - gradient.contiguous(), - full_matrices=False, - ) - U_k = U[:, :, :rank] - S_k = S[:, :rank] - V_k = V[:, :rank, :].clone() - left_mat, right_mat = torch.matmul(U_k, torch.diag_embed(S_k)).contiguous(), V_k.contiguous() - new_gradient = torch.rand(size=(qbatch_dim, output_dim, input_dim), dtype=torch.float64) - - lr_score = opt_einsum.contract("qki,toi,qok->qt", right_mat, new_gradient, left_mat) - lr_score_reconst_matmul = torch.matmul( - torch.matmul(left_mat, right_mat).view(left_mat.size(0), -1), new_gradient.view(new_gradient.shape[0], -1).t() - ) - assert torch.allclose(lr_score, lr_score_reconst_matmul) - - # This should be used when input_dim > output_dim. - intermediate = opt_einsum.contract("qki,toi->qtko", right_mat, new_gradient) - final = opt_einsum.contract("qtko,qok->qt", intermediate, left_mat) - assert torch.allclose(lr_score, final) - print("Option 1") - print(intermediate.numel()) - - # This should be used when output_dim > input_dim. - intermediate2 = torch.einsum("toi,qok->qtik", new_gradient, left_mat) - final2 = opt_einsum.contract("qki,qtik->qt", right_mat, intermediate2) - assert torch.allclose(lr_score, final2) - print("Option 2") - print(intermediate2.numel()) - - print("Reconstruction") - reconst_numel = (torch.matmul(left_mat, right_mat).view(left_mat.size(0), -1)).numel() - print(reconst_numel) - path = opt_einsum.contract_path("qki,toi,qok->qt", right_mat, new_gradient, left_mat) - print(path) - - if new_gradient.size(0) * right_mat.size(0) * rank * min((right_mat.size(2), left_mat.size(1))) > right_mat.size( - 0 - ) * right_mat.size(2) * left_mat.size(1): - assert intermediate2.numel() > reconst_numel and intermediate.numel() > reconst_numel - elif output_dim >= input_dim: - assert intermediate2.numel() <= reconst_numel - else: - assert intermediate.numel() <= reconst_numel - - -def test_compute_score_matmul( - seed: int = 0, -) -> None: - input_dim = 1024 - output_dim = 2048 - batch_dim = 16 - query_batch_dim = 64 - set_seed(seed) - - gradient = torch.rand(size=(batch_dim, output_dim, input_dim), dtype=torch.float64) - new_gradient = torch.rand(size=(query_batch_dim, output_dim, input_dim), dtype=torch.float64) - - score = opt_einsum.contract("toi,qoi->tq", gradient, new_gradient) - path = opt_einsum.contract_path("toi,qoi->tq", gradient, new_gradient) - print(path) - - unsqueeze_score = opt_einsum.contract("t...,q...->tq", gradient, new_gradient) - assert torch.allclose(score, unsqueeze_score) - path = opt_einsum.contract_path("t...,q...->tq", gradient, new_gradient) - print(path) diff --git a/tests/scores/test_pairwise_scores.py b/tests/scores/test_pairwise_scores.py index 6b3b454..087627f 100644 --- a/tests/scores/test_pairwise_scores.py +++ b/tests/scores/test_pairwise_scores.py @@ -4,14 +4,20 @@ import pytest import torch +from scipy.stats import spearmanr -from kronfluence.arguments import FactorArguments, ScoreArguments +from kronfluence.arguments import ScoreArguments +from kronfluence.utils.common.factor_arguments import pytest_factor_arguments +from kronfluence.utils.common.score_arguments import pytest_score_arguments from kronfluence.utils.constants import ALL_MODULE_NAME from kronfluence.utils.dataset import DataLoaderKwargs from tests.utils import ( ATOL, + DEFAULT_FACTORS_NAME, + DEFAULT_SCORES_NAME, RTOL, check_tensor_dict_equivalence, + custom_scores_name, prepare_model_and_analyzer, prepare_test, ) @@ -23,25 +29,26 @@ "mlp", "repeated_mlp", "conv", - "conv_bn", "bert", + "roberta", "gpt", + "gpt_checkpoint", ], ) -@pytest.mark.parametrize("score_dtype", [torch.float32, torch.bfloat16]) -@pytest.mark.parametrize("query_gradient_rank", [None, 16]) +@pytest.mark.parametrize("score_dtype", [torch.float32]) +@pytest.mark.parametrize("query_gradient_low_rank", [None, 16]) @pytest.mark.parametrize("query_size", [16]) @pytest.mark.parametrize("train_size", [32]) @pytest.mark.parametrize("seed", [0]) def test_compute_pairwise_scores( test_name: str, score_dtype: torch.dtype, - query_gradient_rank: Optional[int], + query_gradient_low_rank: Optional[int], query_size: int, train_size: int, seed: int, ) -> None: - # Make sure that the pairwise influence computations are working properly. + # Makes sure that pairwise influence computations are working properly. model, train_dataset, test_dataset, data_collator, task = prepare_test( test_name=test_name, query_size=query_size, @@ -53,9 +60,13 @@ def test_compute_pairwise_scores( model=model, task=task, ) - factors_name = f"pytest_{test_name}_{test_compute_pairwise_scores.__name__}" + factor_args = pytest_factor_arguments() + if test_name == "repeated_mlp": + factor_args.has_shared_parameters = True + analyzer.fit_all_factors( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, + factor_args=factor_args, dataset=train_dataset, dataloader_kwargs=kwargs, per_device_batch_size=32, @@ -64,12 +75,11 @@ def test_compute_pairwise_scores( score_args = ScoreArguments( score_dtype=score_dtype, - query_gradient_rank=query_gradient_rank, + query_gradient_low_rank=query_gradient_low_rank, ) - scores_name = f"pytest_{test_name}_{test_compute_pairwise_scores.__name__}_{query_gradient_rank}_scores" analyzer.compute_pairwise_scores( - scores_name=scores_name, - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, query_dataset=test_dataset, per_device_query_batch_size=4, train_dataset=train_dataset, @@ -79,36 +89,35 @@ def test_compute_pairwise_scores( overwrite_output_dir=True, ) - pairwise_scores = analyzer.load_pairwise_scores(scores_name=scores_name) + pairwise_scores = analyzer.load_pairwise_scores(scores_name=DEFAULT_SCORES_NAME) assert pairwise_scores[ALL_MODULE_NAME].size(0) == query_size assert pairwise_scores[ALL_MODULE_NAME].size(1) == train_size assert pairwise_scores[ALL_MODULE_NAME].dtype == score_dtype -@pytest.mark.parametrize( - "test_name", - [ - "mlp", - ], -) -@pytest.mark.parametrize("per_sample_gradient_dtype", [torch.float32, torch.bfloat16]) -@pytest.mark.parametrize("precondition_dtype", [torch.float32, torch.bfloat16]) -@pytest.mark.parametrize("score_dtype", [torch.float32, torch.bfloat16]) -@pytest.mark.parametrize("query_gradient_rank", [None, 16]) +@pytest.mark.parametrize("test_name", ["mlp"]) +@pytest.mark.parametrize("has_shared_parameters", [True, False]) +@pytest.mark.parametrize("per_sample_gradient_dtype", [torch.float32, torch.float16]) +@pytest.mark.parametrize("precondition_dtype", [torch.float32, torch.float16]) +@pytest.mark.parametrize("score_dtype", [torch.float32, torch.float16]) +@pytest.mark.parametrize("query_gradient_low_rank", [None, 16]) +@pytest.mark.parametrize("damping_factor", [None, 1e-08]) @pytest.mark.parametrize("query_size", [16]) @pytest.mark.parametrize("train_size", [32]) -@pytest.mark.parametrize("seed", [6]) +@pytest.mark.parametrize("seed", [1]) def test_compute_pairwise_scores_dtype( test_name: str, + has_shared_parameters: bool, per_sample_gradient_dtype: torch.dtype, precondition_dtype: torch.dtype, score_dtype: torch.dtype, - query_gradient_rank: Optional[int], + query_gradient_low_rank: Optional[int], + damping_factor: Optional[float], query_size: int, train_size: int, seed: int, ) -> None: - # Make sure that the pairwise influence computations are working properly with different dtypes. + # Makes sure that pairwise influence computations are working properly with different data types. model, train_dataset, test_dataset, data_collator, task = prepare_test( test_name=test_name, query_size=query_size, @@ -120,25 +129,28 @@ def test_compute_pairwise_scores_dtype( model=model, task=task, ) - factors_name = f"pytest_{test_name}_{test_compute_pairwise_scores_dtype.__name__}" + + factor_args = pytest_factor_arguments() + factor_args.has_shared_parameters = has_shared_parameters analyzer.fit_all_factors( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, + factor_args=factor_args, dataloader_kwargs=kwargs, per_device_batch_size=32, overwrite_output_dir=True, ) score_args = ScoreArguments( + damping_factor=damping_factor, score_dtype=score_dtype, - query_gradient_rank=query_gradient_rank, + query_gradient_low_rank=query_gradient_low_rank, per_sample_gradient_dtype=per_sample_gradient_dtype, precondition_dtype=precondition_dtype, ) - scores_name = f"pytest_{test_name}_{test_compute_pairwise_scores_dtype.__name__}_{query_gradient_rank}_scores" analyzer.compute_pairwise_scores( - scores_name=scores_name, - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, query_dataset=test_dataset, per_device_query_batch_size=4, train_dataset=train_dataset, @@ -148,7 +160,7 @@ def test_compute_pairwise_scores_dtype( overwrite_output_dir=True, ) - pairwise_scores = analyzer.load_pairwise_scores(scores_name=scores_name) + pairwise_scores = analyzer.load_pairwise_scores(scores_name=DEFAULT_SCORES_NAME) assert pairwise_scores[ALL_MODULE_NAME].size(0) == query_size assert pairwise_scores[ALL_MODULE_NAME].size(1) == train_size assert pairwise_scores[ALL_MODULE_NAME].dtype == score_dtype @@ -158,14 +170,13 @@ def test_compute_pairwise_scores_dtype( "test_name", [ "mlp", - "conv", - "gpt", + "conv_bn", ], ) @pytest.mark.parametrize("strategy", ["identity", "diagonal", "kfac", "ekfac"]) @pytest.mark.parametrize("query_size", [20]) @pytest.mark.parametrize("train_size", [50]) -@pytest.mark.parametrize("seed", [1]) +@pytest.mark.parametrize("seed", [2]) def test_pairwise_scores_batch_size_equivalence( test_name: str, strategy: str, @@ -187,12 +198,9 @@ def test_pairwise_scores_batch_size_equivalence( task=task, ) - factor_args = FactorArguments( - strategy=strategy, - ) - factors_name = f"pytest_{test_name}_{test_pairwise_scores_batch_size_equivalence.__name__}_{strategy}" + factor_args = pytest_factor_arguments(strategy=strategy) analyzer.fit_all_factors( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, dataloader_kwargs=kwargs, per_device_batch_size=4, @@ -200,15 +208,10 @@ def test_pairwise_scores_batch_size_equivalence( overwrite_output_dir=True, ) - score_args = ScoreArguments( - per_module_score=False, - score_dtype=torch.float64, - precondition_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - ) + score_args = pytest_score_arguments() analyzer.compute_pairwise_scores( - scores_name=f"pytest_{test_name}_{test_pairwise_scores_batch_size_equivalence.__name__}_{strategy}_score_bs1", - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, query_dataset=test_dataset, per_device_query_batch_size=4, train_dataset=train_dataset, @@ -218,12 +221,12 @@ def test_pairwise_scores_batch_size_equivalence( overwrite_output_dir=True, ) bs1_scores = analyzer.load_pairwise_scores( - scores_name=f"pytest_{test_name}_{test_pairwise_scores_batch_size_equivalence.__name__}_{strategy}_score_bs1", + scores_name=DEFAULT_SCORES_NAME, ) analyzer.compute_pairwise_scores( - scores_name=f"pytest_{test_name}_{test_pairwise_scores_batch_size_equivalence.__name__}_{strategy}_score_bs8", - factors_name=factors_name, + scores_name=custom_scores_name("bs8"), + factors_name=DEFAULT_FACTORS_NAME, query_dataset=test_dataset, per_device_query_batch_size=3, train_dataset=train_dataset, @@ -233,7 +236,7 @@ def test_pairwise_scores_batch_size_equivalence( overwrite_output_dir=True, ) bs8_scores = analyzer.load_pairwise_scores( - scores_name=f"pytest_{test_name}_{test_pairwise_scores_batch_size_equivalence.__name__}_{strategy}_score_bs8", + scores_name=custom_scores_name("bs8"), ) assert check_tensor_dict_equivalence( @@ -244,8 +247,8 @@ def test_pairwise_scores_batch_size_equivalence( ) analyzer.compute_pairwise_scores( - scores_name=f"pytest_{test_name}_{test_pairwise_scores_batch_size_equivalence.__name__}_{strategy}_score_auto", - factors_name=factors_name, + scores_name=custom_scores_name("auto"), + factors_name=DEFAULT_FACTORS_NAME, query_dataset=test_dataset, per_device_query_batch_size=10, train_dataset=train_dataset, @@ -255,7 +258,7 @@ def test_pairwise_scores_batch_size_equivalence( overwrite_output_dir=True, ) bs_auto_scores = analyzer.load_pairwise_scores( - scores_name=f"pytest_{test_name}_{test_pairwise_scores_batch_size_equivalence.__name__}_{strategy}_score_auto", + scores_name=custom_scores_name("auto"), ) assert check_tensor_dict_equivalence( @@ -271,19 +274,22 @@ def test_pairwise_scores_batch_size_equivalence( [ "mlp", "conv", + "gpt", ], ) -@pytest.mark.parametrize("data_partition_size", [1, 4]) -@pytest.mark.parametrize("module_partition_size", [1, 3]) -@pytest.mark.parametrize("per_module_score", [True, False]) +@pytest.mark.parametrize("data_partitions", [2, 4]) +@pytest.mark.parametrize("module_partitions", [2, 3]) +@pytest.mark.parametrize("compute_per_module_scores", [True, False]) +@pytest.mark.parametrize("compute_per_token_scores", [True, False]) @pytest.mark.parametrize("query_size", [32]) @pytest.mark.parametrize("train_size", [64]) -@pytest.mark.parametrize("seed", [2]) +@pytest.mark.parametrize("seed", [3]) def test_pairwise_scores_partition_equivalence( test_name: str, - data_partition_size: int, - module_partition_size: int, - per_module_score: bool, + data_partitions: int, + module_partitions: int, + compute_per_module_scores: bool, + compute_per_token_scores: bool, query_size: int, train_size: int, seed: int, @@ -302,25 +308,20 @@ def test_pairwise_scores_partition_equivalence( task=task, ) - factors_name = f"pytest_{test_name}_{test_pairwise_scores_partition_equivalence.__name__}" analyzer.fit_all_factors( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, dataloader_kwargs=kwargs, per_device_batch_size=8, overwrite_output_dir=True, ) - scores_name = f"pytest_{test_name}_{test_pairwise_scores_partition_equivalence.__name__}_scores" - score_args = ScoreArguments( - per_module_score=per_module_score, - score_dtype=torch.float64, - precondition_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - ) + score_args = pytest_score_arguments() + score_args.compute_per_module_scores = compute_per_module_scores + score_args.compute_per_token_scores = compute_per_token_scores analyzer.compute_pairwise_scores( - scores_name=scores_name, - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, query_dataset=test_dataset, per_device_query_batch_size=4, train_dataset=train_dataset, @@ -329,19 +330,13 @@ def test_pairwise_scores_partition_equivalence( score_args=score_args, overwrite_output_dir=True, ) - scores = analyzer.load_pairwise_scores(scores_name=scores_name) + scores = analyzer.load_pairwise_scores(scores_name=DEFAULT_SCORES_NAME) - score_args = ScoreArguments( - data_partition_size=data_partition_size, - module_partition_size=module_partition_size, - per_module_score=per_module_score, - score_dtype=torch.float64, - precondition_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - ) + score_args.data_partitions = data_partitions + score_args.module_partitions = module_partitions analyzer.compute_pairwise_scores( - scores_name=f"pytest_{test_name}_partition_{data_partition_size}_{module_partition_size}", - factors_name=factors_name, + scores_name=custom_scores_name(f"{data_partitions}_{module_partitions}"), + factors_name=DEFAULT_FACTORS_NAME, query_dataset=test_dataset, per_device_query_batch_size=10, train_dataset=train_dataset, @@ -351,7 +346,7 @@ def test_pairwise_scores_partition_equivalence( overwrite_output_dir=True, ) partitioned_scores = analyzer.load_pairwise_scores( - scores_name=f"pytest_{test_name}_partition_{data_partition_size}_{module_partition_size}", + scores_name=custom_scores_name(f"{data_partitions}_{module_partitions}"), ) assert check_tensor_dict_equivalence( @@ -372,7 +367,7 @@ def test_pairwise_scores_partition_equivalence( ) @pytest.mark.parametrize("query_size", [32]) @pytest.mark.parametrize("train_size", [64]) -@pytest.mark.parametrize("seed", [0]) +@pytest.mark.parametrize("seed", [4]) def test_per_module_scores_equivalence( test_name: str, query_size: int, @@ -393,25 +388,18 @@ def test_per_module_scores_equivalence( task=task, ) - factors_name = f"pytest_{test_name}_{test_per_module_scores_equivalence.__name__}" analyzer.fit_all_factors( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, dataloader_kwargs=kwargs, per_device_batch_size=8, overwrite_output_dir=True, ) - scores_name = f"pytest_{test_name}_{test_per_module_scores_equivalence.__name__}_scores" - score_args = ScoreArguments( - per_module_score=False, - score_dtype=torch.float64, - precondition_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - ) + score_args = pytest_score_arguments() analyzer.compute_pairwise_scores( - scores_name=scores_name, - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, query_dataset=test_dataset, per_device_query_batch_size=4, train_dataset=train_dataset, @@ -420,17 +408,12 @@ def test_per_module_scores_equivalence( score_args=score_args, overwrite_output_dir=True, ) - scores = analyzer.load_pairwise_scores(scores_name=scores_name) + scores = analyzer.load_pairwise_scores(scores_name=DEFAULT_SCORES_NAME) - score_args = ScoreArguments( - per_module_score=True, - score_dtype=torch.float64, - precondition_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - ) + score_args.compute_per_module_scores = True analyzer.compute_pairwise_scores( - scores_name=scores_name + "_per_module", - factors_name=factors_name, + scores_name=custom_scores_name("per_module"), + factors_name=DEFAULT_FACTORS_NAME, query_dataset=test_dataset, per_device_query_batch_size=4, train_dataset=train_dataset, @@ -439,7 +422,7 @@ def test_per_module_scores_equivalence( score_args=score_args, overwrite_output_dir=True, ) - per_module_scores = analyzer.load_pairwise_scores(scores_name=scores_name + "_per_module") + per_module_scores = analyzer.load_pairwise_scores(scores_name=custom_scores_name("per_module")) total_scores = None for module_name in per_module_scores: @@ -451,24 +434,99 @@ def test_per_module_scores_equivalence( assert torch.allclose(total_scores, scores[ALL_MODULE_NAME], atol=ATOL, rtol=RTOL) +@pytest.mark.parametrize("test_name", ["mlp", "conv", "gpt"]) +@pytest.mark.parametrize("compute_per_module_scores", [True, False]) +@pytest.mark.parametrize("query_size", [12]) +@pytest.mark.parametrize("train_size", [64]) +@pytest.mark.parametrize("seed", [5]) +def test_per_token_scores_equivalence( + test_name: str, + compute_per_module_scores: bool, + query_size: int, + train_size: int, + seed: int, +) -> None: + # Influence scores should be identical with and without per token score computations. + model, train_dataset, test_dataset, data_collator, task = prepare_test( + test_name=test_name, + query_size=query_size, + train_size=train_size, + seed=seed, + ) + model = model.to(dtype=torch.float64) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + + analyzer.fit_all_factors( + factors_name=DEFAULT_FACTORS_NAME, + dataset=train_dataset, + dataloader_kwargs=kwargs, + per_device_batch_size=8, + overwrite_output_dir=True, + ) + + score_args = pytest_score_arguments() + score_args.compute_per_module_scores = compute_per_module_scores + analyzer.compute_pairwise_scores( + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, + query_dataset=test_dataset, + per_device_query_batch_size=4, + train_dataset=train_dataset, + per_device_train_batch_size=8, + dataloader_kwargs=kwargs, + score_args=score_args, + overwrite_output_dir=True, + ) + scores = analyzer.load_pairwise_scores(scores_name=DEFAULT_SCORES_NAME) + + score_args.compute_per_token_scores = True + analyzer.compute_pairwise_scores( + scores_name=custom_scores_name("per_token"), + factors_name=DEFAULT_FACTORS_NAME, + query_dataset=test_dataset, + per_device_query_batch_size=4, + train_dataset=train_dataset, + per_device_train_batch_size=8, + dataloader_kwargs=kwargs, + score_args=score_args, + overwrite_output_dir=True, + ) + per_token_scores = analyzer.load_pairwise_scores(scores_name=custom_scores_name("per_token")) + + for module_name in per_token_scores: + if test_name == "gpt": + assert torch.allclose(per_token_scores[module_name].sum(dim=-1), scores[module_name], atol=ATOL, rtol=RTOL) + else: + assert torch.allclose(per_token_scores[module_name], scores[module_name], atol=ATOL, rtol=RTOL) + + @pytest.mark.parametrize( "test_name", [ "mlp", "conv_bn", - "gpt", ], ) +@pytest.mark.parametrize("data_partitions", [1, 2]) +@pytest.mark.parametrize("query_gradient_low_rank", [None, 4]) +@pytest.mark.parametrize("query_gradient_accumulation_steps", [1, 4]) @pytest.mark.parametrize("query_size", [60]) @pytest.mark.parametrize("train_size", [60]) -@pytest.mark.parametrize("seed", [0]) +@pytest.mark.parametrize("seed", [6]) def test_compute_pairwise_scores_with_indices( test_name: str, + data_partitions: int, + query_gradient_low_rank: Optional[int], + query_gradient_accumulation_steps: int, query_size: int, train_size: int, seed: int, ) -> None: - # Make sure the indices selection is correctly implemented. + # Makes sure the indices selection is correctly implemented. model, train_dataset, test_dataset, data_collator, task = prepare_test( test_name=test_name, train_size=train_size, @@ -480,20 +538,21 @@ def test_compute_pairwise_scores_with_indices( model=model, task=task, ) - factors_name = f"pytest_{test_name}_{test_compute_pairwise_scores_with_indices.__name__}" analyzer.fit_all_factors( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, dataloader_kwargs=kwargs, per_device_batch_size=32, overwrite_output_dir=True, ) - score_args = ScoreArguments(data_partition_size=2) - scores_name = f"pytest_{test_name}_{test_compute_pairwise_scores_with_indices.__name__}_scores" + score_args = pytest_score_arguments() + score_args.data_partitions = data_partitions + score_args.query_gradient_low_rank = query_gradient_low_rank + score_args.query_gradient_accumulation_steps = query_gradient_accumulation_steps analyzer.compute_pairwise_scores( - scores_name=scores_name, - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, query_dataset=test_dataset, query_indices=list(range(30)), per_device_query_batch_size=4, @@ -505,7 +564,7 @@ def test_compute_pairwise_scores_with_indices( overwrite_output_dir=True, ) - pairwise_scores = analyzer.load_pairwise_scores(scores_name=scores_name) + pairwise_scores = analyzer.load_pairwise_scores(scores_name=DEFAULT_SCORES_NAME) assert pairwise_scores[ALL_MODULE_NAME].size(0) == 30 assert pairwise_scores[ALL_MODULE_NAME].size(1) == 50 @@ -514,22 +573,23 @@ def test_compute_pairwise_scores_with_indices( "test_name", [ "mlp", - "repeated_mlp", - "conv", + "conv_bn", ], ) @pytest.mark.parametrize("query_size", [64]) @pytest.mark.parametrize("train_size", [32]) -@pytest.mark.parametrize("num_query_gradient_aggregations", [2, 5]) -@pytest.mark.parametrize("seed", [5]) -def test_query_aggregation( +@pytest.mark.parametrize("query_gradient_low_rank", [None]) +@pytest.mark.parametrize("query_gradient_accumulation_steps", [2, 5]) +@pytest.mark.parametrize("seed", [7]) +def test_query_accumulation_steps( test_name: str, query_size: int, train_size: int, - num_query_gradient_aggregations: int, + query_gradient_low_rank: Optional[int], + query_gradient_accumulation_steps: int, seed: int, ) -> None: - # Make sure the query aggregation is correctly implemented. + # Makes sure the query accumulation is correctly implemented. model, train_dataset, test_dataset, data_collator, task = prepare_test( test_name=test_name, query_size=query_size, @@ -543,26 +603,101 @@ def test_query_aggregation( task=task, ) - factors_name = f"pytest_{test_name}_{test_query_aggregation.__name__}" analyzer.fit_all_factors( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, dataloader_kwargs=kwargs, per_device_batch_size=8, overwrite_output_dir=True, ) - scores_name = f"pytest_{test_name}_{test_query_aggregation.__name__}_scores" - score_args = ScoreArguments( - query_gradient_rank=8, - num_query_gradient_aggregations=1, - score_dtype=torch.float64, - precondition_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, + score_args = pytest_score_arguments(query_gradient_low_rank=query_gradient_low_rank) + analyzer.compute_pairwise_scores( + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, + query_dataset=test_dataset, + per_device_query_batch_size=4, + train_dataset=train_dataset, + per_device_train_batch_size=8, + dataloader_kwargs=kwargs, + score_args=score_args, + overwrite_output_dir=True, ) + scores = analyzer.load_pairwise_scores(scores_name=DEFAULT_SCORES_NAME) + + score_args.query_gradient_accumulation_steps = query_gradient_accumulation_steps analyzer.compute_pairwise_scores( - scores_name=scores_name, - factors_name=factors_name, + scores_name=custom_scores_name("accumulation"), + factors_name=DEFAULT_FACTORS_NAME, + query_dataset=test_dataset, + per_device_query_batch_size=10, + train_dataset=train_dataset, + per_device_train_batch_size=5, + dataloader_kwargs=kwargs, + score_args=score_args, + overwrite_output_dir=True, + ) + accumulated_scores = analyzer.load_pairwise_scores( + scores_name=custom_scores_name("accumulation"), + ) + + assert check_tensor_dict_equivalence( + scores, + accumulated_scores, + atol=ATOL, + rtol=RTOL, + ) + + +@pytest.mark.parametrize( + "test_name", + ["mlp", "conv"], +) +@pytest.mark.parametrize("query_size", [50]) +@pytest.mark.parametrize("train_size", [32]) +@pytest.mark.parametrize("data_partitions", [3]) +@pytest.mark.parametrize("module_partitions", [3]) +@pytest.mark.parametrize("query_gradient_low_rank", [None]) +@pytest.mark.parametrize("seed", [8]) +def test_query_gradient_aggregation( + test_name: str, + query_size: int, + train_size: int, + data_partitions: int, + module_partitions: int, + query_gradient_low_rank: Optional[int], + seed: int, +) -> None: + # Makes sure the query gradient aggregation is correctly implemented. + model, train_dataset, test_dataset, data_collator, task = prepare_test( + test_name=test_name, + query_size=query_size, + train_size=train_size, + seed=seed, + ) + model = model.to(dtype=torch.float64) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + + factor_args = pytest_factor_arguments() + if test_name == "repeated_mlp": + factor_args.has_shared_parameters = True + analyzer.fit_all_factors( + factors_name=DEFAULT_FACTORS_NAME, + dataset=train_dataset, + factor_args=factor_args, + dataloader_kwargs=kwargs, + per_device_batch_size=8, + overwrite_output_dir=True, + ) + + score_args = pytest_score_arguments(query_gradient_low_rank=query_gradient_low_rank) + analyzer.compute_pairwise_scores( + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, query_dataset=test_dataset, per_device_query_batch_size=4, train_dataset=train_dataset, @@ -571,18 +706,99 @@ def test_query_aggregation( score_args=score_args, overwrite_output_dir=True, ) - scores = analyzer.load_pairwise_scores(scores_name=scores_name) + scores = analyzer.load_pairwise_scores(scores_name=DEFAULT_SCORES_NAME) - score_args = ScoreArguments( - query_gradient_rank=8, - num_query_gradient_aggregations=num_query_gradient_aggregations, - score_dtype=torch.float64, - precondition_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, + score_args.aggregate_query_gradients = True + score_args.data_partitions = data_partitions + score_args.module_partitions = data_partitions + analyzer.compute_pairwise_scores( + scores_name=custom_scores_name("aggregation"), + factors_name=DEFAULT_FACTORS_NAME, + query_dataset=test_dataset, + per_device_query_batch_size=10, + train_dataset=train_dataset, + per_device_train_batch_size=5, + dataloader_kwargs=kwargs, + score_args=score_args, + overwrite_output_dir=True, + ) + aggregated_scores = analyzer.load_pairwise_scores( + scores_name=custom_scores_name("aggregation"), + ) + + assert aggregated_scores[ALL_MODULE_NAME].shape[0] == 1 + assert torch.allclose( + scores[ALL_MODULE_NAME].sum(dim=0, keepdim=True), + aggregated_scores[ALL_MODULE_NAME], + atol=ATOL, + rtol=RTOL, + ) + + +@pytest.mark.parametrize( + "test_name", + ["mlp", "conv"], +) +@pytest.mark.parametrize("query_size", [64]) +@pytest.mark.parametrize("train_size", [32]) +@pytest.mark.parametrize("data_partitions", [3]) +@pytest.mark.parametrize("module_partitions", [2]) +@pytest.mark.parametrize("aggregate_query_gradients", [True, False]) +@pytest.mark.parametrize("query_gradient_low_rank", [None]) +@pytest.mark.parametrize("seed", [9]) +def test_train_gradient_aggregation( + test_name: str, + query_size: int, + train_size: int, + data_partitions: int, + module_partitions: int, + aggregate_query_gradients: bool, + query_gradient_low_rank: Optional[int], + seed: int, +) -> None: + # Makes sure the train gradient aggregation is correctly implemented. + model, train_dataset, test_dataset, data_collator, task = prepare_test( + test_name=test_name, + query_size=query_size, + train_size=train_size, + seed=seed, + ) + model = model.to(dtype=torch.float64) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + + analyzer.fit_all_factors( + factors_name=DEFAULT_FACTORS_NAME, + dataset=train_dataset, + dataloader_kwargs=kwargs, + per_device_batch_size=8, + overwrite_output_dir=True, + ) + + score_args = pytest_score_arguments(query_gradient_low_rank=query_gradient_low_rank) + score_args.aggregate_query_gradients = aggregate_query_gradients + analyzer.compute_pairwise_scores( + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, + query_dataset=test_dataset, + per_device_query_batch_size=4, + train_dataset=train_dataset, + per_device_train_batch_size=8, + dataloader_kwargs=kwargs, + score_args=score_args, + overwrite_output_dir=True, ) + scores = analyzer.load_pairwise_scores(scores_name=DEFAULT_SCORES_NAME) + + score_args.aggregate_train_gradients = True + score_args.data_partitions = data_partitions + score_args.module_partitions = module_partitions analyzer.compute_pairwise_scores( - scores_name=f"pytest_{test_name}_{test_query_aggregation.__name__}_aggregate_scores", - factors_name=factors_name, + scores_name=custom_scores_name("aggregation"), + factors_name=DEFAULT_FACTORS_NAME, query_dataset=test_dataset, per_device_query_batch_size=10, train_dataset=train_dataset, @@ -591,13 +807,172 @@ def test_query_aggregation( score_args=score_args, overwrite_output_dir=True, ) - partitioned_scores = analyzer.load_pairwise_scores( - scores_name=f"pytest_{test_name}_{test_query_aggregation.__name__}_aggregate_scores", + aggregated_scores = analyzer.load_pairwise_scores( + scores_name=custom_scores_name("aggregation"), ) + assert aggregated_scores[ALL_MODULE_NAME].shape[1] == 1 + assert torch.allclose( + scores[ALL_MODULE_NAME].sum(dim=1, keepdim=True), + aggregated_scores[ALL_MODULE_NAME], + atol=ATOL, + rtol=RTOL, + ) + + +@pytest.mark.parametrize( + "test_name", + [ + "mlp", + "conv", + "roberta", + ], +) +@pytest.mark.parametrize("query_size", [50]) +@pytest.mark.parametrize("train_size", [32]) +@pytest.mark.parametrize("seed", [10]) +def test_pairwise_shared_parameters( + test_name: str, + query_size: int, + train_size: int, + seed: int, +) -> None: + # Makes sure the scores are identical with and without `has_shared_parameters` flag. + model, train_dataset, test_dataset, data_collator, task = prepare_test( + test_name=test_name, + query_size=query_size, + train_size=train_size, + seed=seed, + ) + model = model.to(dtype=torch.float64) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + factor_args = pytest_factor_arguments() + score_args = pytest_score_arguments() + analyzer.fit_all_factors( + factors_name=DEFAULT_FACTORS_NAME, + factor_args=factor_args, + dataset=train_dataset, + dataloader_kwargs=kwargs, + per_device_batch_size=8, + overwrite_output_dir=True, + ) + analyzer.compute_pairwise_scores( + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, + query_dataset=test_dataset, + per_device_query_batch_size=4, + train_dataset=train_dataset, + score_args=score_args, + per_device_train_batch_size=8, + dataloader_kwargs=kwargs, + overwrite_output_dir=True, + ) + scores = analyzer.load_pairwise_scores(scores_name=DEFAULT_SCORES_NAME) + + factor_args.has_shared_parameters = True + analyzer.fit_all_factors( + factors_name=DEFAULT_FACTORS_NAME, + factor_args=factor_args, + dataset=train_dataset, + dataloader_kwargs=kwargs, + per_device_batch_size=8, + overwrite_output_dir=True, + ) + analyzer.compute_pairwise_scores( + scores_name=custom_scores_name("shared"), + factors_name=DEFAULT_FACTORS_NAME, + query_dataset=test_dataset, + per_device_query_batch_size=4, + train_dataset=train_dataset, + score_args=score_args, + per_device_train_batch_size=8, + dataloader_kwargs=kwargs, + overwrite_output_dir=True, + ) + shared_scores = analyzer.load_pairwise_scores(scores_name=custom_scores_name("shared")) + assert check_tensor_dict_equivalence( scores, - partitioned_scores, + shared_scores, atol=ATOL, rtol=RTOL, ) + + +@pytest.mark.parametrize( + "test_name", + ["mlp", "conv_bn", "gpt"], +) +@pytest.mark.parametrize("query_gradient_low_rank", [16, 32]) +@pytest.mark.parametrize("use_full_svd", [False, True]) +@pytest.mark.parametrize("query_gradient_accumulation_steps", [1, 3]) +@pytest.mark.parametrize("query_size", [64]) +@pytest.mark.parametrize("train_size", [160]) +@pytest.mark.parametrize("seed", [11]) +def test_pairwise_query_batching( + test_name: str, + query_gradient_low_rank: int, + use_full_svd: bool, + query_gradient_accumulation_steps: int, + query_size: int, + train_size: int, + seed: int, +) -> None: + # Makes sure similar results are obtained with and without query batching. + model, train_dataset, test_dataset, data_collator, task = prepare_test( + test_name=test_name, + query_size=query_size, + train_size=train_size, + seed=seed, + ) + model = model.to(dtype=torch.float64) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + factor_args = pytest_factor_arguments() + analyzer.fit_all_factors( + factors_name=DEFAULT_FACTORS_NAME, + factor_args=factor_args, + dataset=train_dataset, + dataloader_kwargs=kwargs, + per_device_batch_size=8, + overwrite_output_dir=True, + ) + score_args = pytest_score_arguments() + analyzer.compute_pairwise_scores( + scores_name=DEFAULT_SCORES_NAME, + score_args=score_args, + factors_name=DEFAULT_FACTORS_NAME, + query_dataset=test_dataset, + per_device_query_batch_size=4, + train_dataset=train_dataset, + per_device_train_batch_size=8, + dataloader_kwargs=kwargs, + overwrite_output_dir=True, + ) + scores = analyzer.load_pairwise_scores(scores_name=DEFAULT_SCORES_NAME) + + score_args.query_gradient_low_rank = query_gradient_low_rank + score_args.use_full_svd = use_full_svd + score_args.query_gradient_accumulation_steps = query_gradient_accumulation_steps + analyzer.compute_pairwise_scores( + scores_name=custom_scores_name("qb"), + score_args=score_args, + factors_name=DEFAULT_FACTORS_NAME, + query_dataset=test_dataset, + per_device_query_batch_size=3, + train_dataset=train_dataset, + per_device_train_batch_size=9, + dataloader_kwargs=kwargs, + overwrite_output_dir=True, + ) + qb_scores = analyzer.load_pairwise_scores(scores_name=custom_scores_name("qb")) + + for i in range(query_size): + assert spearmanr(scores[ALL_MODULE_NAME][i], qb_scores[ALL_MODULE_NAME][i])[0] > 0.9 diff --git a/tests/scores/test_self_scores.py b/tests/scores/test_self_scores.py index 1464140..c0ee190 100644 --- a/tests/scores/test_self_scores.py +++ b/tests/scores/test_self_scores.py @@ -1,15 +1,26 @@ # pylint: skip-file +from typing import Optional + import pytest import torch -from kronfluence.arguments import FactorArguments, ScoreArguments +from kronfluence.arguments import ScoreArguments +from kronfluence.utils.common.factor_arguments import ( + default_factor_arguments, + pytest_factor_arguments, +) +from kronfluence.utils.common.score_arguments import pytest_score_arguments from kronfluence.utils.constants import ALL_MODULE_NAME from kronfluence.utils.dataset import DataLoaderKwargs from tests.utils import ( ATOL, + DEFAULT_FACTORS_NAME, + DEFAULT_SCORES_NAME, RTOL, check_tensor_dict_equivalence, + custom_factors_name, + custom_scores_name, prepare_model_and_analyzer, prepare_test, ) @@ -21,21 +32,24 @@ "mlp", "repeated_mlp", "conv", - "conv_bn", "bert", + "roberta", "gpt", + "gpt_checkpoint", ], ) -@pytest.mark.parametrize("score_dtype", [torch.float32, torch.bfloat16]) +@pytest.mark.parametrize("use_measurement_for_self_influence", [False, True]) +@pytest.mark.parametrize("score_dtype", [torch.float32]) @pytest.mark.parametrize("train_size", [22]) @pytest.mark.parametrize("seed", [0]) def test_compute_self_scores( test_name: str, + use_measurement_for_self_influence: bool, score_dtype: torch.dtype, train_size: int, seed: int, ) -> None: - # Make sure that the self-influence computations are working properly. + # Makes sure that self-influence computations are working properly. model, train_dataset, _, data_collator, task = prepare_test( test_name=test_name, train_size=train_size, @@ -46,9 +60,13 @@ def test_compute_self_scores( model=model, task=task, ) - factors_name = f"pytest_{test_name}_{test_compute_self_scores.__name__}" + factor_args = default_factor_arguments() + if test_name == "repeated_mlp": + factor_args.has_shared_parameters = True + analyzer.fit_all_factors( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, + factor_args=factor_args, dataset=train_dataset, dataloader_kwargs=kwargs, per_device_batch_size=32, @@ -56,12 +74,12 @@ def test_compute_self_scores( ) score_args = ScoreArguments( + use_measurement_for_self_influence=use_measurement_for_self_influence, score_dtype=score_dtype, ) - scores_name = f"pytest_{test_name}_{test_compute_self_scores.__name__}_scores" analyzer.compute_self_scores( - scores_name=scores_name, - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, train_dataset=train_dataset, per_device_train_batch_size=4, dataloader_kwargs=kwargs, @@ -69,32 +87,31 @@ def test_compute_self_scores( overwrite_output_dir=True, ) - self_scores = analyzer.load_self_scores(scores_name=scores_name) + self_scores = analyzer.load_self_scores(scores_name=DEFAULT_SCORES_NAME) assert self_scores[ALL_MODULE_NAME].size(0) == train_size assert len(self_scores[ALL_MODULE_NAME].shape) == 1 assert self_scores[ALL_MODULE_NAME].dtype == score_dtype -@pytest.mark.parametrize( - "test_name", - [ - "mlp", - ], -) -@pytest.mark.parametrize("per_sample_gradient_dtype", [torch.float32, torch.bfloat16]) -@pytest.mark.parametrize("precondition_dtype", [torch.float32, torch.bfloat16]) -@pytest.mark.parametrize("score_dtype", [torch.float32, torch.bfloat16]) +@pytest.mark.parametrize("test_name", ["mlp"]) +@pytest.mark.parametrize("has_shared_parameters", [True, False]) +@pytest.mark.parametrize("per_sample_gradient_dtype", [torch.float32, torch.float16]) +@pytest.mark.parametrize("precondition_dtype", [torch.float32, torch.float16]) +@pytest.mark.parametrize("score_dtype", [torch.float32, torch.float16]) +@pytest.mark.parametrize("damping_factor", [None, 1e-08]) @pytest.mark.parametrize("train_size", [32]) -@pytest.mark.parametrize("seed", [6]) +@pytest.mark.parametrize("seed", [1]) def test_compute_self_scores_dtype( test_name: str, + has_shared_parameters: bool, per_sample_gradient_dtype: torch.dtype, precondition_dtype: torch.dtype, score_dtype: torch.dtype, + damping_factor: Optional[float], train_size: int, seed: int, ) -> None: - # Make sure that the self-influence computations are working properly with different dtypes. + # Makes sure that self-influence computations are working properly with different data types. model, train_dataset, _, data_collator, task = prepare_test( test_name=test_name, query_size=10, @@ -106,10 +123,13 @@ def test_compute_self_scores_dtype( model=model, task=task, ) - factors_name = f"pytest_{test_name}_{test_compute_self_scores_dtype.__name__}" + + factor_args = pytest_factor_arguments() + factor_args.has_shared_parameters = has_shared_parameters analyzer.fit_all_factors( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, + factor_args=factor_args, dataloader_kwargs=kwargs, per_device_batch_size=32, overwrite_output_dir=True, @@ -119,18 +139,18 @@ def test_compute_self_scores_dtype( score_dtype=score_dtype, per_sample_gradient_dtype=per_sample_gradient_dtype, precondition_dtype=precondition_dtype, + damping_factor=damping_factor, ) - scores_name = f"pytest_{test_name}_{test_compute_self_scores_dtype.__name__}_scores" analyzer.compute_self_scores( - scores_name=scores_name, - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, train_dataset=train_dataset, per_device_train_batch_size=8, dataloader_kwargs=kwargs, score_args=score_args, overwrite_output_dir=True, ) - self_scores = analyzer.load_self_scores(scores_name=scores_name) + self_scores = analyzer.load_self_scores(scores_name=DEFAULT_SCORES_NAME) assert self_scores[ALL_MODULE_NAME].size(0) == train_size assert len(self_scores[ALL_MODULE_NAME].shape) == 1 assert self_scores[ALL_MODULE_NAME].dtype == score_dtype @@ -140,13 +160,12 @@ def test_compute_self_scores_dtype( "test_name", [ "mlp", - "conv", - "gpt", + "conv_bn", ], ) @pytest.mark.parametrize("strategy", ["identity", "diagonal", "kfac", "ekfac"]) -@pytest.mark.parametrize("train_size", [50]) -@pytest.mark.parametrize("seed", [1]) +@pytest.mark.parametrize("train_size", [49]) +@pytest.mark.parametrize("seed", [2]) def test_self_scores_batch_size_equivalence( test_name: str, strategy: str, @@ -166,12 +185,9 @@ def test_self_scores_batch_size_equivalence( task=task, ) - factor_args = FactorArguments( - strategy=strategy, - ) - factors_name = f"pytest_{test_name}_{test_self_scores_batch_size_equivalence.__name__}" + factor_args = pytest_factor_arguments(strategy=strategy) analyzer.fit_all_factors( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, dataloader_kwargs=kwargs, per_device_batch_size=4, @@ -179,15 +195,10 @@ def test_self_scores_batch_size_equivalence( overwrite_output_dir=True, ) - score_args = ScoreArguments( - per_module_score=False, - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - ) + score_args = pytest_score_arguments() analyzer.compute_self_scores( - scores_name=f"pytest_{test_name}_{test_self_scores_batch_size_equivalence.__name__}_{strategy}_score_bs1", - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, train_dataset=train_dataset, per_device_train_batch_size=1, dataloader_kwargs=kwargs, @@ -195,12 +206,12 @@ def test_self_scores_batch_size_equivalence( overwrite_output_dir=True, ) bs1_scores = analyzer.load_self_scores( - scores_name=f"pytest_{test_name}_{test_self_scores_batch_size_equivalence.__name__}_{strategy}_score_bs1", + scores_name=DEFAULT_SCORES_NAME, ) analyzer.compute_self_scores( - scores_name=f"pytest_{test_name}_{test_self_scores_batch_size_equivalence.__name__}_{strategy}_score_bs8", - factors_name=factors_name, + scores_name=custom_scores_name("bs8"), + factors_name=DEFAULT_FACTORS_NAME, train_dataset=train_dataset, per_device_train_batch_size=8, dataloader_kwargs=kwargs, @@ -208,7 +219,7 @@ def test_self_scores_batch_size_equivalence( overwrite_output_dir=True, ) bs8_scores = analyzer.load_self_scores( - scores_name=f"pytest_{test_name}_{test_self_scores_batch_size_equivalence.__name__}_{strategy}_score_bs8", + scores_name=custom_scores_name("bs8"), ) assert check_tensor_dict_equivalence( @@ -219,8 +230,8 @@ def test_self_scores_batch_size_equivalence( ) analyzer.compute_self_scores( - scores_name=f"pytest_{test_name}_{test_self_scores_batch_size_equivalence.__name__}_{strategy}_score_auto", - factors_name=factors_name, + scores_name=custom_scores_name("auto"), + factors_name=DEFAULT_FACTORS_NAME, train_dataset=train_dataset, per_device_train_batch_size=None, dataloader_kwargs=kwargs, @@ -228,7 +239,7 @@ def test_self_scores_batch_size_equivalence( overwrite_output_dir=True, ) bs_auto_scores = analyzer.load_self_scores( - scores_name=f"pytest_{test_name}_{test_self_scores_batch_size_equivalence.__name__}_{strategy}_score_auto", + scores_name=custom_scores_name("auto"), ) assert check_tensor_dict_equivalence( @@ -247,14 +258,16 @@ def test_self_scores_batch_size_equivalence( "gpt", ], ) -@pytest.mark.parametrize("data_partition_size", [1, 4]) -@pytest.mark.parametrize("module_partition_size", [1, 3]) +@pytest.mark.parametrize("data_partitions", [2, 4]) +@pytest.mark.parametrize("module_partitions", [2, 3]) +@pytest.mark.parametrize("compute_per_module_scores", [True, False]) @pytest.mark.parametrize("train_size", [64]) -@pytest.mark.parametrize("seed", [2]) +@pytest.mark.parametrize("seed", [3]) def test_self_scores_partition_equivalence( test_name: str, - data_partition_size: int, - module_partition_size: int, + data_partitions: int, + module_partitions: int, + compute_per_module_scores: bool, train_size: int, seed: int, ) -> None: @@ -271,42 +284,32 @@ def test_self_scores_partition_equivalence( task=task, ) - factors_name = f"pytest_{test_name}_{test_self_scores_partition_equivalence.__name__}" analyzer.fit_all_factors( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, dataloader_kwargs=kwargs, per_device_batch_size=8, overwrite_output_dir=True, ) - scores_name = f"pytest_{test_name}_{test_self_scores_partition_equivalence.__name__}_scores" - score_args = ScoreArguments( - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - ) + score_args = pytest_score_arguments() + score_args.compute_per_module_scores = compute_per_module_scores analyzer.compute_self_scores( - scores_name=scores_name, - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, train_dataset=train_dataset, per_device_train_batch_size=8, dataloader_kwargs=kwargs, score_args=score_args, overwrite_output_dir=True, ) - scores = analyzer.load_self_scores(scores_name=scores_name) + scores = analyzer.load_self_scores(scores_name=DEFAULT_SCORES_NAME) - score_args = ScoreArguments( - data_partition_size=data_partition_size, - module_partition_size=module_partition_size, - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - ) + score_args.data_partitions = data_partitions + score_args.module_partitions = module_partitions analyzer.compute_self_scores( - scores_name=f"pytest_{test_name}_partition_{data_partition_size}_{module_partition_size}", - factors_name=factors_name, + scores_name=custom_scores_name(f"{data_partitions}_{module_partitions}"), + factors_name=DEFAULT_FACTORS_NAME, train_dataset=train_dataset, per_device_train_batch_size=5, dataloader_kwargs=kwargs, @@ -314,7 +317,7 @@ def test_self_scores_partition_equivalence( overwrite_output_dir=True, ) partitioned_scores = analyzer.load_self_scores( - scores_name=f"pytest_{test_name}_partition_{data_partition_size}_{module_partition_size}", + scores_name=custom_scores_name(f"{data_partitions}_{module_partitions}"), ) assert check_tensor_dict_equivalence( @@ -329,7 +332,7 @@ def test_self_scores_partition_equivalence( "test_name", [ "mlp", - "conv", + "conv_bn", "gpt", ], ) @@ -353,37 +356,37 @@ def test_per_module_scores_equivalence( task=task, ) - factors_name = f"pytest_{test_name}_{test_per_module_scores_equivalence.__name__}" analyzer.fit_all_factors( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, dataloader_kwargs=kwargs, per_device_batch_size=8, overwrite_output_dir=True, ) - scores_name = f"pytest_{test_name}_{test_per_module_scores_equivalence.__name__}_scores" + score_args = pytest_score_arguments() analyzer.compute_self_scores( - scores_name=scores_name, - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + score_args=score_args, + factors_name=DEFAULT_FACTORS_NAME, train_dataset=train_dataset, per_device_train_batch_size=8, dataloader_kwargs=kwargs, overwrite_output_dir=True, ) - scores = analyzer.load_self_scores(scores_name=scores_name) + scores = analyzer.load_self_scores(scores_name=DEFAULT_SCORES_NAME) - score_args = ScoreArguments(per_module_score=True) + score_args.per_module_score = True analyzer.compute_self_scores( - scores_name=scores_name + "_per_module", - factors_name=factors_name, + scores_name=custom_scores_name("per_module"), + factors_name=DEFAULT_FACTORS_NAME, train_dataset=train_dataset, per_device_train_batch_size=8, dataloader_kwargs=kwargs, score_args=score_args, overwrite_output_dir=True, ) - per_module_scores = analyzer.load_self_scores(scores_name=scores_name + "_per_module") + per_module_scores = analyzer.load_self_scores(custom_scores_name("per_module")) total_scores = None for module_name in per_module_scores: @@ -400,17 +403,18 @@ def test_per_module_scores_equivalence( [ "mlp", "conv_bn", - "gpt", ], ) +@pytest.mark.parametrize("data_partitions", [1, 2]) @pytest.mark.parametrize("train_size", [60]) -@pytest.mark.parametrize("seed", [7]) +@pytest.mark.parametrize("seed", [6]) def test_compute_self_scores_with_indices( test_name: str, + data_partitions: int, train_size: int, seed: int, ) -> None: - # Make sure the indices selection is correctly implemented. + # Makes sure the indices selection is correctly implemented. model, train_dataset, _, data_collator, task = prepare_test( test_name=test_name, train_size=train_size, @@ -421,20 +425,19 @@ def test_compute_self_scores_with_indices( model=model, task=task, ) - factors_name = f"pytest_{test_name}_{test_compute_self_scores_with_indices.__name__}" analyzer.fit_all_factors( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, dataloader_kwargs=kwargs, per_device_batch_size=32, overwrite_output_dir=True, ) - score_args = ScoreArguments(data_partition_size=2) - scores_name = f"pytest_{test_name}_{test_compute_self_scores_with_indices.__name__}_scores" + score_args = pytest_score_arguments() + score_args.data_partitions = data_partitions analyzer.compute_self_scores( - scores_name=scores_name, - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, train_dataset=train_dataset, train_indices=list(range(48)), per_device_train_batch_size=8, @@ -443,18 +446,13 @@ def test_compute_self_scores_with_indices( overwrite_output_dir=True, ) - self_scores = analyzer.load_self_scores(scores_name=scores_name) + self_scores = analyzer.load_self_scores(scores_name=DEFAULT_SCORES_NAME) assert self_scores[ALL_MODULE_NAME].size(0) == 48 -@pytest.mark.parametrize( - "test_name", - [ - "mlp", - ], -) +@pytest.mark.parametrize("test_name", ["mlp"]) @pytest.mark.parametrize("train_size", [60]) -@pytest.mark.parametrize("seed", [1]) +@pytest.mark.parametrize("seed", [7]) def test_compute_self_scores_with_diagonal_pairwise_equivalence( test_name: str, train_size: int, @@ -472,36 +470,29 @@ def test_compute_self_scores_with_diagonal_pairwise_equivalence( model=model, task=task, ) - factors_name = f"pytest_{test_name}_{test_compute_self_scores_with_diagonal_pairwise_equivalence.__name__}" analyzer.fit_all_factors( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, dataloader_kwargs=kwargs, per_device_batch_size=32, overwrite_output_dir=True, ) - scores_name = f"pytest_{test_name}_{test_compute_self_scores_with_diagonal_pairwise_equivalence.__name__}_scores" - score_args = ScoreArguments( - use_measurement_for_self_influence=False, - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - ) + score_args = pytest_score_arguments() analyzer.compute_self_scores( - scores_name=scores_name, - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, train_dataset=train_dataset, per_device_train_batch_size=8, dataloader_kwargs=kwargs, score_args=score_args, overwrite_output_dir=True, ) - self_scores = analyzer.load_self_scores(scores_name=scores_name) + self_scores = analyzer.load_self_scores(scores_name=DEFAULT_SCORES_NAME) analyzer.compute_pairwise_scores( - scores_name=scores_name, - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, train_dataset=train_dataset, per_device_train_batch_size=8, query_dataset=train_dataset, @@ -510,7 +501,7 @@ def test_compute_self_scores_with_diagonal_pairwise_equivalence( score_args=score_args, overwrite_output_dir=True, ) - pairwise_scores = analyzer.load_pairwise_scores(scores_name=scores_name) + pairwise_scores = analyzer.load_pairwise_scores(scores_name=DEFAULT_SCORES_NAME) assert torch.allclose( torch.diag(pairwise_scores[ALL_MODULE_NAME]), @@ -525,7 +516,7 @@ def test_compute_self_scores_with_diagonal_pairwise_equivalence( ["mlp", "conv", "conv_bn", "wrong_conv"], ) @pytest.mark.parametrize("train_size", [24]) -@pytest.mark.parametrize("seed", [7]) +@pytest.mark.parametrize("seed", [8]) def test_compute_self_measurement_scores_with_diagonal_pairwise_equivalence( test_name: str, train_size: int, @@ -543,40 +534,30 @@ def test_compute_self_measurement_scores_with_diagonal_pairwise_equivalence( model=model, task=task, ) - factors_name = ( - f"pytest_{test_name}_{test_compute_self_measurement_scores_with_diagonal_pairwise_equivalence.__name__}" - ) analyzer.fit_all_factors( - factors_name=factors_name, + factors_name=DEFAULT_FACTORS_NAME, dataset=train_dataset, dataloader_kwargs=kwargs, per_device_batch_size=32, overwrite_output_dir=True, ) - scores_name = ( - f"pytest_{test_name}_{test_compute_self_measurement_scores_with_diagonal_pairwise_equivalence.__name__}_scores" - ) - score_args = ScoreArguments( - use_measurement_for_self_influence=True, - score_dtype=torch.float64, - per_sample_gradient_dtype=torch.float64, - precondition_dtype=torch.float64, - ) + score_args = pytest_score_arguments() + score_args.use_measurement_for_self_influence = True analyzer.compute_self_scores( - scores_name=scores_name, - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, train_dataset=train_dataset, per_device_train_batch_size=8, dataloader_kwargs=kwargs, score_args=score_args, overwrite_output_dir=True, ) - self_scores = analyzer.load_self_scores(scores_name=scores_name) + self_scores = analyzer.load_self_scores(scores_name=DEFAULT_SCORES_NAME) analyzer.compute_pairwise_scores( - scores_name=scores_name, - factors_name=factors_name, + scores_name=DEFAULT_SCORES_NAME, + factors_name=DEFAULT_FACTORS_NAME, train_dataset=train_dataset, per_device_train_batch_size=8, query_dataset=train_dataset, @@ -585,7 +566,7 @@ def test_compute_self_measurement_scores_with_diagonal_pairwise_equivalence( score_args=score_args, overwrite_output_dir=True, ) - pairwise_scores = analyzer.load_pairwise_scores(scores_name=scores_name) + pairwise_scores = analyzer.load_pairwise_scores(scores_name=DEFAULT_SCORES_NAME) assert torch.allclose( torch.diag(pairwise_scores[ALL_MODULE_NAME]), @@ -593,3 +574,86 @@ def test_compute_self_measurement_scores_with_diagonal_pairwise_equivalence( atol=ATOL, rtol=RTOL, ) + + +@pytest.mark.parametrize( + "test_name", + [ + "mlp", + "conv", + "roberta", + ], +) +@pytest.mark.parametrize("use_measurement_for_self_influence", [False, True]) +@pytest.mark.parametrize("query_size", [50]) +@pytest.mark.parametrize("train_size", [32]) +@pytest.mark.parametrize("seed", [9]) +def test_self_shared_parameters( + test_name: str, + use_measurement_for_self_influence: bool, + query_size: int, + train_size: int, + seed: int, +) -> None: + # Makes sure the scores are identical with and without `has_shared_parameters` flag. + model, train_dataset, _, data_collator, task = prepare_test( + test_name=test_name, + query_size=query_size, + train_size=train_size, + seed=seed, + ) + model = model.to(dtype=torch.float64) + kwargs = DataLoaderKwargs(collate_fn=data_collator) + model, analyzer = prepare_model_and_analyzer( + model=model, + task=task, + ) + factor_args = pytest_factor_arguments() + factor_args.has_shared_parameters = False + score_args = pytest_score_arguments() + score_args.use_measurement_for_self_influence = use_measurement_for_self_influence + analyzer.fit_all_factors( + factors_name=DEFAULT_FACTORS_NAME, + factor_args=factor_args, + dataset=train_dataset, + dataloader_kwargs=kwargs, + per_device_batch_size=8, + overwrite_output_dir=True, + ) + analyzer.compute_self_scores( + scores_name=DEFAULT_SCORES_NAME, + score_args=score_args, + factors_name=DEFAULT_FACTORS_NAME, + train_dataset=train_dataset, + per_device_train_batch_size=8, + dataloader_kwargs=kwargs, + overwrite_output_dir=True, + ) + scores = analyzer.load_self_scores(scores_name=DEFAULT_SCORES_NAME) + + factor_args.has_shared_parameters = True + analyzer.fit_all_factors( + factors_name=custom_factors_name("shared"), + factor_args=factor_args, + dataset=train_dataset, + dataloader_kwargs=kwargs, + per_device_batch_size=8, + overwrite_output_dir=True, + ) + analyzer.compute_self_scores( + scores_name=custom_scores_name("shared"), + score_args=score_args, + factors_name=custom_factors_name("shared"), + train_dataset=train_dataset, + per_device_train_batch_size=8, + dataloader_kwargs=kwargs, + overwrite_output_dir=True, + ) + shared_scores = analyzer.load_self_scores(scores_name=custom_scores_name("shared")) + + assert check_tensor_dict_equivalence( + scores, + shared_scores, + atol=ATOL, + rtol=RTOL, + ) diff --git a/tests/test_analyzer.py b/tests/test_analyzer.py index cb5f8c1..89b4e77 100644 --- a/tests/test_analyzer.py +++ b/tests/test_analyzer.py @@ -13,9 +13,11 @@ "test_name", [ "mlp", - "conv", + "mlp_checkpoint", + "repeated_mlp", "conv_bn", "bert", + "roberta", "gpt", ], ) @@ -47,11 +49,15 @@ def test_analyzer( analysis_name=f"pytest_{test_name}", model=model, task=task, + disable_tqdm=True, disable_model_save=True, cpu=True, ) kwargs = DataLoaderKwargs(collate_fn=data_collator) + factor_args = FactorArguments(strategy=strategy) + if test_name == "repeated_mlp": + factor_args.has_shared_parameters = True analyzer.fit_all_factors( factors_name=f"pytest_{test_analyzer.__name__}_{test_name}", dataset=train_dataset, @@ -70,12 +76,24 @@ def test_analyzer( dataloader_kwargs=kwargs, overwrite_output_dir=True, ) + score_args = ScoreArguments() analyzer.compute_self_scores( scores_name="self", factors_name=f"pytest_{test_analyzer.__name__}_{test_name}", train_dataset=train_dataset, per_device_train_batch_size=8, dataloader_kwargs=kwargs, + score_args=score_args, + overwrite_output_dir=True, + ) + score_args.use_measurement_for_self_influence = True + analyzer.compute_self_scores( + scores_name="self", + factors_name=f"pytest_{test_analyzer.__name__}_{test_name}", + train_dataset=train_dataset, + per_device_train_batch_size=6, + dataloader_kwargs=kwargs, + score_args=score_args, overwrite_output_dir=True, ) @@ -85,41 +103,49 @@ def test_default_factor_arguments() -> None: assert factor_args.strategy == "ekfac" assert factor_args.use_empirical_fisher is False - assert factor_args.distributed_sync_steps == 1000 assert factor_args.amp_dtype is None + assert factor_args.amp_scale == 2.0**16 + assert factor_args.has_shared_parameters is False assert factor_args.covariance_max_examples == 100_000 - assert factor_args.covariance_data_partition_size == 1 - assert factor_args.covariance_module_partition_size == 1 + assert factor_args.covariance_data_partitions == 1 + assert factor_args.covariance_module_partitions == 1 assert factor_args.activation_covariance_dtype == torch.float32 assert factor_args.gradient_covariance_dtype == torch.float32 + assert factor_args.eigendecomposition_dtype == torch.float64 assert factor_args.lambda_max_examples == 100_000 - assert factor_args.lambda_data_partition_size == 1 - assert factor_args.lambda_module_partition_size == 1 - assert factor_args.lambda_iterative_aggregate is False - assert factor_args.cached_activation_cpu_offload is False + assert factor_args.lambda_data_partitions == 1 + assert factor_args.lambda_module_partitions == 1 + assert factor_args.use_iterative_lambda_aggregation is False + assert factor_args.offload_activations_to_cpu is False + assert factor_args.per_sample_gradient_dtype == torch.float32 assert factor_args.lambda_dtype == torch.float32 def test_default_score_arguments() -> None: score_args = ScoreArguments() - assert score_args.damping == 1e-08 - assert score_args.cached_activation_cpu_offload is False - assert score_args.distributed_sync_steps == 1000 + assert score_args.damping_factor == 1e-08 assert score_args.amp_dtype is None + assert score_args.offload_activations_to_cpu is False - assert score_args.data_partition_size == 1 - assert score_args.module_partition_size == 1 - assert score_args.per_module_score is False + assert score_args.data_partitions == 1 + assert score_args.module_partitions == 1 + + assert score_args.compute_per_module_scores is False + assert score_args.compute_per_token_scores is False + + assert score_args.query_gradient_accumulation_steps == 1 + assert score_args.query_gradient_low_rank is None + assert score_args.use_full_svd is False + assert score_args.aggregate_query_gradients is False + assert score_args.aggregate_train_gradients is False - assert score_args.query_gradient_rank is None - assert score_args.num_query_gradient_aggregations == 1 - assert score_args.query_gradient_svd_dtype == torch.float32 assert score_args.use_measurement_for_self_influence is False - assert score_args.score_dtype == torch.float32 + assert score_args.query_gradient_svd_dtype == torch.float32 assert score_args.per_sample_gradient_dtype == torch.float32 assert score_args.precondition_dtype == torch.float32 + assert score_args.score_dtype == torch.float32 diff --git a/tests/test_testable_tasks.py b/tests/test_testable_tasks.py index 722a46c..6d84d75 100644 --- a/tests/test_testable_tasks.py +++ b/tests/test_testable_tasks.py @@ -8,6 +8,7 @@ make_conv_model, ) from tests.testable_tasks.language_modeling import make_gpt_dataset, make_tiny_gpt +from tests.testable_tasks.multiple_choice import make_roberta_dataset, make_tiny_roberta from tests.testable_tasks.regression import make_mlp_model, make_regression_dataset from tests.testable_tasks.text_classification import make_bert_dataset, make_tiny_bert @@ -69,6 +70,27 @@ def test_bert(): logits.sum().backward() +def test_roberta(): + model = make_tiny_roberta(seed=0) + dataset = make_roberta_dataset(num_data=2, seed=0) + batch_size = 2 + loader = DataLoader( + dataset, + collate_fn=default_data_collator, + batch_size=batch_size, + drop_last=False, + shuffle=False, + ) + + batch = next(iter(loader)) + inputs = ( + batch["input_ids"], + batch["attention_mask"], + ) + logits = model(*inputs).logits + logits.sum().backward() + + def test_gpt(): model = make_tiny_gpt(seed=0) dataset = make_gpt_dataset(num_data=8, seed=0) diff --git a/tests/testable_tasks/classification.py b/tests/testable_tasks/classification.py index db89a42..1282f2a 100644 --- a/tests/testable_tasks/classification.py +++ b/tests/testable_tasks/classification.py @@ -26,6 +26,18 @@ def make_conv_model(bias: bool = True, seed: int = 0) -> nn.Module: ) +def make_conv_inplace_model(bias: bool = True, seed: int = 0) -> nn.Module: + set_seed(seed) + return nn.Sequential( + nn.Conv2d(3, 4, 3, 1, bias=bias), + nn.ReLU(inplace=True), + nn.Conv2d(4, 8, 3, 1, bias=bias), + nn.ReLU(inplace=True), + nn.Flatten(), + nn.Linear(1152, 5, bias=bias), + ) + + def make_conv_bn_model(bias: bool = True, seed: int = 0) -> nn.Module: set_seed(seed) return nn.Sequential( diff --git a/tests/testable_tasks/language_modeling.py b/tests/testable_tasks/language_modeling.py index aa2fd11..4c7f1e3 100644 --- a/tests/testable_tasks/language_modeling.py +++ b/tests/testable_tasks/language_modeling.py @@ -1,20 +1,28 @@ # pylint: skip-file from itertools import chain -from typing import Any, Dict, List, Optional +from typing import Any, Dict, List import torch import torch.nn.functional as F from datasets import load_dataset from torch import nn from torch.utils import data -from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, Conv1D +from transformers import ( + AutoConfig, + AutoModelForCausalLM, + AutoTokenizer, + Conv1D, + logging, +) from kronfluence.task import Task +logging.set_verbosity_error() BATCH_TYPE = Dict[str, torch.Tensor] +@torch.no_grad() def _replace_conv1d_modules(model: nn.Module) -> None: for name, module in model.named_children(): if len(list(module.children())) > 0: @@ -70,7 +78,8 @@ def tokenize_function(examples): def group_texts(examples): concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()} total_length = len(concatenated_examples[list(examples.keys())[0]]) - total_length = (total_length // block_size) * block_size + if total_length >= block_size: + total_length = (total_length // block_size) * block_size result = { k: [t[i : i + block_size] for i in range(0, total_length, block_size)] for k, t in concatenated_examples.items() @@ -128,7 +137,7 @@ def compute_measurement( ) -> torch.Tensor: return self.compute_train_loss(batch, model) - def tracked_modules(self) -> List[str]: + def get_influence_tracked_modules(self) -> List[str]: total_modules = [] for i in range(5): @@ -141,5 +150,5 @@ def tracked_modules(self) -> List[str]: return total_modules - def get_attention_mask(self, batch: Any) -> Optional[torch.Tensor]: + def get_attention_mask(self, batch: Any) -> torch.Tensor: return batch["attention_mask"] diff --git a/tests/testable_tasks/multiple_choice.py b/tests/testable_tasks/multiple_choice.py new file mode 100644 index 0000000..971b6fb --- /dev/null +++ b/tests/testable_tasks/multiple_choice.py @@ -0,0 +1,133 @@ +# pylint: skip-file + +from itertools import chain +from typing import Any, Dict + +import torch +import torch.nn.functional as F +from datasets import load_dataset +from torch import nn +from torch.utils import data +from transformers import AutoConfig, AutoModelForMultipleChoice, AutoTokenizer, logging + +from kronfluence.task import Task + +logging.set_verbosity_error() +BATCH_TYPE = Dict[str, torch.Tensor] + + +def make_tiny_roberta(seed: int = 0) -> nn.Module: + torch.manual_seed(seed) + config = AutoConfig.from_pretrained( + "hf-internal-testing/tiny-random-RobertaModel", + trust_remote_code=True, + ) + model = AutoModelForMultipleChoice.from_pretrained( + "hf-internal-testing/tiny-random-RobertaModel", + from_tf=False, + config=config, + ignore_mismatched_sizes=False, + trust_remote_code=True, + ) + return model + + +def make_roberta_dataset(num_data: int, seed: int = 0) -> data.Dataset: + torch.manual_seed(seed) + raw_datasets = load_dataset("swag", "regular") + tokenizer = AutoTokenizer.from_pretrained( + "hf-internal-testing/tiny-random-RobertaModel", use_fast=True, trust_remote_code=True + ) + + column_names = raw_datasets["train"].column_names + ending_names = [f"ending{i}" for i in range(4)] + context_name = "sent1" + question_header_name = "sent2" + label_column_name = "label" if "label" in column_names else "labels" + padding = "max_length" + + def preprocess_function(examples: Any): + first_sentences = [[context] * 4 for context in examples[context_name]] + question_headers = examples[question_header_name] + second_sentences = [ + [f"{header} {examples[end][i]}" for end in ending_names] for i, header in enumerate(question_headers) + ] + labels = examples[label_column_name] + + first_sentences = list(chain(*first_sentences)) + second_sentences = list(chain(*second_sentences)) + + tokenized_examples = tokenizer( + first_sentences, + second_sentences, + max_length=128, + padding=padding, + truncation=True, + ) + tokenized_inputs = {k: [v[i : i + 4] for i in range(0, len(v), 4)] for k, v in tokenized_examples.items()} + tokenized_inputs["labels"] = labels + return tokenized_inputs + + processed_datasets = raw_datasets.map( + preprocess_function, + batched=True, + remove_columns=raw_datasets["train"].column_names, + ) + + dataset = processed_datasets["train"].select(range(num_data)) + + return dataset + + +class MultipleChoiceTask(Task): + enable_post_process_per_sample_gradient = True + + def compute_train_loss( + self, + batch: BATCH_TYPE, + model: nn.Module, + sample: bool = False, + ) -> torch.Tensor: + logits = model( + input_ids=batch["input_ids"], + attention_mask=batch["attention_mask"], + ).logits + + if not sample: + return F.cross_entropy(logits, batch["labels"], reduction="sum") + with torch.no_grad(): + probs = torch.nn.functional.softmax(logits.detach(), dim=-1) + sampled_labels = torch.multinomial( + probs, + num_samples=1, + ).flatten() + return F.cross_entropy(logits, sampled_labels, reduction="sum") + + def compute_measurement( + self, + batch: BATCH_TYPE, + model: nn.Module, + ) -> torch.Tensor: + logits = model( + input_ids=batch["input_ids"], + attention_mask=batch["attention_mask"], + ).logits + + labels = batch["labels"] + bindex = torch.arange(logits.shape[0]).to(device=logits.device, non_blocking=False) + logits_correct = logits[bindex, labels] + + cloned_logits = logits.clone() + cloned_logits[bindex, labels] = torch.tensor(-torch.inf, device=logits.device, dtype=logits.dtype) + + margins = logits_correct - cloned_logits.logsumexp(dim=-1) + return -margins.sum() + + def get_attention_mask(self, batch: Any) -> torch.Tensor: + return batch["attention_mask"] + + def post_process_per_sample_gradient(self, module_name: str, gradient: torch.Tensor) -> torch.Tensor: + del module_name + total_batch_size = gradient.size(0) + true_batch_size = int(total_batch_size / 4) + return gradient.reshape(true_batch_size, 4, *gradient.size()[1:]).sum(dim=1) diff --git a/tests/testable_tasks/regression.py b/tests/testable_tasks/regression.py index e706efa..04e714e 100644 --- a/tests/testable_tasks/regression.py +++ b/tests/testable_tasks/regression.py @@ -8,6 +8,7 @@ from accelerate.utils import set_seed from torch import nn from torch.utils import data +from torch.utils.checkpoint import checkpoint_sequential from kronfluence.task import Task @@ -25,19 +26,32 @@ def make_mlp_model(bias: bool = True, seed: int = 0) -> nn.Module: ) +class RepeatedMLP(nn.Module): + def __init__(self, bias: bool) -> None: + super().__init__() + self.linear1 = nn.Linear(10, 16, bias=bias) + self.shared_linear = nn.Linear(16, 16, bias=bias) + self.linear2 = nn.Linear(16, 16, bias=bias) + self.linear3 = nn.Linear(16, 1, bias=bias) + + def forward(self, inputs: torch.Tensor) -> torch.Tensor: + x = self.linear1(inputs) + x = torch.relu(x) + x = self.shared_linear(x) + x = torch.relu(x) + x = self.shared_linear(x) + x = torch.relu(x) + x = self.linear2(x) + x = torch.relu(x) + x = self.shared_linear(x) + x = torch.relu(x) + return self.linear3(x) + + def make_repeated_mlp_model(bias: bool = True, seed: int = 0) -> nn.Module: set_seed(seed) - - shared_linear = nn.Linear(16, 16, bias=bias) - return nn.Sequential( - nn.Linear(10, 16, bias=bias), - nn.ReLU(), - shared_linear, - nn.ReLU(), - shared_linear, - nn.ReLU(), - nn.Linear(16, 1, bias=bias), - ) + model = RepeatedMLP(bias=bias) + return model def make_regression_dataset(num_data: int, seed: int = 0) -> data.Dataset: @@ -75,3 +89,32 @@ def compute_measurement( model: nn.Module, ) -> torch.Tensor: return self.compute_train_loss(batch, model, sample=False) + + +class GradientCheckpointRegressionTask(RegressionTask): + def compute_train_loss( + self, + batch: BATCH_TYPE, + model: nn.Module, + sample: bool = False, + ) -> torch.Tensor: + inputs, targets = batch + + if list(model.parameters())[1].dtype == torch.float64: + inputs = inputs.to(dtype=torch.float64) + targets = targets.to(dtype=torch.float64) + + outputs = checkpoint_sequential(functions=model, segments=2, input=inputs, use_reentrant=False) + + if not sample: + return F.mse_loss(outputs, targets, reduction="sum") + with torch.no_grad(): + sampled_targets = torch.normal(outputs.detach(), std=math.sqrt(0.5)) + return F.mse_loss(outputs, sampled_targets, reduction="sum") + + def compute_measurement( + self, + batch: BATCH_TYPE, + model: nn.Module, + ) -> torch.Tensor: + return self.compute_train_loss(batch, model, sample=False) diff --git a/tests/testable_tasks/text_classification.py b/tests/testable_tasks/text_classification.py index 0a2ac5f..6e9c005 100644 --- a/tests/testable_tasks/text_classification.py +++ b/tests/testable_tasks/text_classification.py @@ -7,10 +7,16 @@ from datasets import load_dataset from torch import nn from torch.utils import data -from transformers import AutoConfig, AutoModelForSequenceClassification, AutoTokenizer +from transformers import ( + AutoConfig, + AutoModelForSequenceClassification, + AutoTokenizer, + logging, +) from kronfluence.task import Task +logging.set_verbosity_error() BATCH_TYPE = Dict[str, torch.Tensor] @@ -113,7 +119,7 @@ def compute_measurement( margins = logits_correct - cloned_logits.logsumexp(dim=-1) return -margins.sum() - def get_attention_mask(self, batch: Any) -> Optional[torch.Tensor]: + def get_attention_mask(self, batch: Any) -> torch.Tensor: return batch["attention_mask"] diff --git a/tests/utils.py b/tests/utils.py index db74c44..03ca290 100644 --- a/tests/utils.py +++ b/tests/utils.py @@ -12,17 +12,24 @@ from kronfluence.utils.exceptions import UnsupportableModuleError from tests.testable_tasks.classification import ( ClassificationTask, + WrongClassificationTask, make_classification_dataset, make_conv_bn_model, + make_conv_inplace_model, make_conv_model, - WrongClassificationTask, ) from tests.testable_tasks.language_modeling import ( LanguageModelingTask, make_gpt_dataset, make_tiny_gpt, ) +from tests.testable_tasks.multiple_choice import ( + MultipleChoiceTask, + make_roberta_dataset, + make_tiny_roberta, +) from tests.testable_tasks.regression import ( + GradientCheckpointRegressionTask, RegressionTask, make_mlp_model, make_regression_dataset, @@ -38,15 +45,26 @@ RTOL = 1.3e-6 ATOL = 1e-5 +DEFAULT_FACTORS_NAME = "pytest" +DEFAULT_SCORES_NAME = "pytest" + + +def custom_factors_name(name: str) -> str: + return f"{DEFAULT_FACTORS_NAME}_{name}" + + +def custom_scores_name(name: str) -> str: + return f"{DEFAULT_FACTORS_NAME}_{name}" + def prepare_model_and_analyzer(model: nn.Module, task: Task) -> Tuple[nn.Module, Analyzer]: model = prepare_model(model=model, task=task) analyzer = Analyzer( - analysis_name=f"pytest_{__name__}", + analysis_name="pytest", model=model, task=task, disable_model_save=True, - cpu=True, + disable_tqdm=True, ) return model, analyzer @@ -70,12 +88,24 @@ def prepare_test( query_dataset = make_regression_dataset(num_data=query_size, seed=seed + 1) task = RegressionTask() data_collator = None + elif test_name == "mlp_checkpoint": + model = make_mlp_model(seed=seed) + train_dataset = make_regression_dataset(num_data=train_size, seed=seed) + query_dataset = make_regression_dataset(num_data=query_size, seed=seed + 1) + task = GradientCheckpointRegressionTask() + data_collator = None elif test_name == "conv": model = make_conv_model(seed=seed) train_dataset = make_classification_dataset(num_data=train_size, seed=seed) query_dataset = make_classification_dataset(num_data=query_size, seed=seed + 1) task = ClassificationTask() data_collator = None + elif test_name == "conv_inplace": + model = make_conv_inplace_model(seed=seed) + train_dataset = make_classification_dataset(num_data=train_size, seed=seed) + query_dataset = make_classification_dataset(num_data=query_size, seed=seed + 1) + task = ClassificationTask() + data_collator = None elif test_name == "wrong_conv": model = make_conv_model(seed=seed) train_dataset = make_classification_dataset(num_data=train_size, seed=seed) @@ -100,12 +130,25 @@ def prepare_test( query_dataset = make_bert_dataset(num_data=query_size, seed=seed + 1, do_not_pad=do_not_pad) task = WrongTextClassificationTask() data_collator = default_data_collator + elif test_name == "roberta": + model = make_tiny_roberta(seed=seed) + train_dataset = make_roberta_dataset(num_data=train_size, seed=seed) + query_dataset = make_roberta_dataset(num_data=query_size, seed=seed + 1) + task = MultipleChoiceTask() + data_collator = default_data_collator elif test_name == "gpt": model = make_tiny_gpt(seed=seed) train_dataset = make_gpt_dataset(num_data=train_size, seed=seed) query_dataset = make_gpt_dataset(num_data=query_size, seed=seed + 1) task = LanguageModelingTask() data_collator = default_data_collator + elif test_name == "gpt_checkpoint": + model = make_tiny_gpt(seed=seed) + model.gradient_checkpointing_enable() + train_dataset = make_gpt_dataset(num_data=train_size, seed=seed) + query_dataset = make_gpt_dataset(num_data=query_size, seed=seed + 1) + task = LanguageModelingTask() + data_collator = default_data_collator else: raise NotImplementedError(f"{test_name} is not a valid test configuration name.") model.eval() @@ -133,7 +176,7 @@ def reshape_parameter_gradient_to_module_matrix( module_name: str, gradient_dict: Dict[str, torch.Tensor], remove_gradient: bool = True, -) -> torch.Tensor: +) -> Optional[torch.Tensor]: if isinstance(module, nn.Linear): if module_name == "lm_head": # Edge case for small GPT model. @@ -164,6 +207,6 @@ def reshape_parameter_gradient_to_module_matrix( if remove_gradient: del gradient_dict[module_name + ".bias"] else: - error_msg = f"Unsupported module type: {type(module)}. Only nn.Linear or nn.Conv2d are supported." + error_msg = f"Unsupported module type: {type(module)}. Only `nn.Linear` or `nn.Conv2d` are supported." raise UnsupportableModuleError(error_msg) return gradient_matrix