Skip to content

Latest commit

 

History

History
353 lines (254 loc) · 22.7 KB

CHANGES.md

File metadata and controls

353 lines (254 loc) · 22.7 KB

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Added

Changed

Fixed

  • Fix an issue with using NeuralNetBinaryClassifier with torch.compile (#1058)

1.0.0 - 2024-05-27

The 1.0.0 release of skorch is here. We think that skorch is at a very stable point, which is why a 1.0.0 release is appropriate. There are no plans to add any breaking changes or major revisions in the future. Instead, our focus now is to keep skorch up-to-date with the latest versions of PyTorch and scikit-learn, and to fix any bugs that may arise.

0.15.0 - 2023-09-04

Added

  • Add the option to globally override the use of caching in scoring callbacks on the net by setting the use_caching argument on the net (this overrides the settings of individual callbacks) (#971)
  • Add support for saving and loading parameters with safetensors; use net.save_params(..., use_safetensors=True) and net.load_params(..., use_safetensors=True) (requires to install the safetensors library) (#970)

Changed

  • Nets pickled with skorch version 0.11 can no longer be loaded in version 0.15 (see #880); to transition these nets, pickle them in a skorch version between 0.12 and 0.14, then load them in 0.15

Fixed

  • Fixed a couple of issues when saving and loading parameters while using accelerate (via AccelerateMixin) in a multi-GPU setting, and some other minor accelerate issues (#1008, #1009)
  • Installing skorch with the [testing] option now installs all dev requirements (#1015)

0.14.0 - 2023-06-24

Added

  • Add version logging to NeptuneLogger callback (#964)
  • Add support for zero-shot and few-shot classification with the help of Large Language Models and the Hugging Face transformers library

Changed

  • Moved from pkg_resources to importlib and subsequently dropping support for Python 3.7 as PyTorch moved dropped support and the version itself hit EOL (#928 and #983)

  • NeuralNetRegressor can now be fitted with 1-dimensional y, which is necessary in some specific circumstances (e.g. in conjunction with sklearn's BaggingRegressor, see #972); for this to work correctly, the output of the of the PyTorch module should also be 1-dimensional; the existing default, i.e. having y and y_pred be 2-dimensional, remains the recommended way of using NeuralNetRegressor

Fixed

0.13.0 - 2023-05-17

Added

  • Add support for compiled PyTorch modules using the torch.compile function, introduced in PyTorch 2.0 release, which can greatly improve performance on new GPU architectures; to use it, initialize your net with the compile=True argument, further compilation arguments can be specified using the dunder notation, e.g. compile__dynamic=True
  • Add a class DistributedHistory which should be used when training in a multi GPU setting (#955)
  • SkorchDoctor: A helper class that assists in understanding and debugging the neural net training, see this notebook (#912)
  • When using AccelerateMixin, it is now possible to prevent unwrapping of the modules by setting unwrap_after_train=True (#963)

Changed

Fixed

  • Fixed install command to work with recent changes in Google Colab (#928)
  • Fixed a couple of bugs related to using non-default modules and criteria (#927)
  • Fixed a bug when using AccelerateMixin in a multi-GPU setup (#947)
  • _get_param_names returns a list instead of a generator so that subsequent error messages return useful information instead of a generator repr string (#925)
  • Fixed a bug that caused modules to not be sufficiently unwrapped at the end of training when using AccelerateMixin, which could prevent them from being pickleable (#963)

0.12.1 - 2022-11-18

Changed

  • NeptuneLogger was updated to work with recent versions of Neptune client (v0.14.3 or higher); it now logs some additional data, including the model summary, configuration, and learning rate (when available) (#906)

Fixed

  • Fixed an error that could occur with specific combinations of gpytorch and PyTorch versions (#913)

0.12.0 - 2022-10-07

Added

  • Added load_best attribute to EarlyStopping callback to automatically load module weights of the best result at the end of training
  • Added a method, trim_for_prediction, on the net classes, which trims the net from everything not required for using it for prediction; call this after fitting to reduce the size of the net
  • Added experimental support for huggingface accelerate; use the provided mixin class to add advanced training capabilities provided by the accelerate library to skorch
  • Add integration for Huggingface tokenizers; use skorch.hf.HuggingfaceTokenizer to train a Huggingface tokenizer on your custom data; use skorch.hf.HuggingfacePretrainedTokenizer to load a pre-trained Huggingface tokenizer
  • Added support for creating model checkpoints on Hugging Face Hub using HfHubStorage
  • Added a notebook that shows how to use skorch with PyTorch Geometric (#863)

Changed

  • The minimum required scikit-learn version has been bumped to 0.22.0
  • Initialize data loaders for training and validation dataset once per fit call instead of once per epoch (migration guide)
  • It is now possible to call np.asarray with SliceDatasets (#858)

Fixed

  • Fixed a bug in SliceDataset that prevented it to be used with to_numpy (#858)
  • Fixed a bug that occurred when loading a net that has device set to None (#876)
  • Fixed a bug that in some cases could prevent loading a net that was trained with CUDA without CUDA
  • Enable skorch to work on M1/M2 Apple MacBooks (#884)

0.11.0 - 2021-10-11

Added

  • Added load_best attribute to Checkpoint callback to automatically load state of the best result at the end of training
  • Added a get_all_learnable_params method to retrieve the named parameters of all PyTorch modules defined on the net, including of criteria if applicable
  • Added MlflowLogger callback for logging to Mlflow (#769)
  • Added InputShapeSetter callback for automatically setting the input dimension of the PyTorch module
  • Added a new module to support Gaussian Processes through GPyTorch. To learn more about it, read the GP documentation or take a look at the GP notebook. This feature is experimental, i.e. the API could be changed in the future in a backwards incompatible way (#782)

Changed

  • Changed the signature of validation_step, train_step_single, train_step, evaluation_step, on_batch_begin, and on_batch_end such that instead of receiving X and y, they receive the whole batch; this makes it easier to deal with datasets that don't strictly return an (X, y) tuple, which is true for quite a few PyTorch datasets; please refer to the migration guide if you encounter problems (#699)
  • Checking of arguments to NeuralNet is now during .initialize(), not during __init__, to avoid raising false positives for yet unknown module or optimizer attributes
  • Modules, criteria, and optimizers that are added to a net by the user are now first class: skorch takes care of setting train/eval mode, moving to the indicated device, and updating all learnable parameters during training (check the docs for more details, #751)
  • CVSplit is renamed to ValidSplit to avoid confusion (#752)

Fixed

  • Fixed a few bugs in the net.history implementation (#776)
  • Fixed a bug in TrainEndCheckpoint that prevented it from being unpickled (#773)

0.10.0 - 2021-03-23

Added

  • Added SacredLogger callback for logging to Sacred (#725)
  • CLI helper function now also supports normal (i.e. non-skorch) sklearn estimators
  • Disabling all callbacks is now supported (which allows reducing overhead, which is especially relevant for small models).
  • LRScheduler now correctly passes the value being monitored to ReduceLROnPlateau. (#738)

Changed

  • We no longer pass the epoch parameter to LR schedulers, since that parameter has been deprecated. We now rely on the scheduler to keep track of the epoch itself.
  • Changed implementation of net.history access to make it faster; this should result in a nice speedup when dealing with very small model/data but otherwise not have any noticeable effects; if you encounter bugs, though, please create an issue

Fixed

0.9.0 - 2020-08-30

Added

  • Added the event_name argument for LRScheduler for optional recording of LR changes inside net.history. NOTE: Supported only in Pytorch>=1.4
  • Make it easier to add custom modules or optimizers to a neural net class by automatically registering them where necessary and by making them available to set_params
  • Added the step_every argument for LRScheduler to set whether the scheduler step should be taken on every epoch or on every batch.
  • Added the scoring module with loss_scoring function, which computes the net's loss (using get_loss) on provided input data.
  • Added a parameter predict_nonlinearity to NeuralNet which allows users to control the nonlinearity to be applied to the module output when calling predict and predict_proba (#637, #661)
  • Added the possibility to save the criterion with save_params and with checkpoint callbacks
  • Added the possibility to save custom modules with save_params and with checkpoint callbacks

Changed

  • Removed support for schedulers with a batch_step() method in LRScheduler.
  • Raise FutureWarning in CVSplit when random_state is not used. Will raise an exception in a future (#620)
  • The behavior of method net.get_params changed to make it more consistent with sklearn: it will no longer return "learned" attributes like module_; therefore, functions like sklearn.base.clone, when called with a fitted net, will no longer return a fitted net but instead an uninitialized net; if you want a copy of a fitted net, use copy.deepcopy instead;net.get_params is used under the hood by many sklearn functions and classes, such as GridSearchCV, whose behavior may thus be affected by the change. (#521, #527)
  • Raise FutureWarning when using CyclicLR scheduler, because the default behavior has changed from taking a step every batch to taking a step every epoch. (#626)
  • Set train/validation on criterion if it's a PyTorch module (#621)
  • Don't pass y=None to NeuralNet.train_split to enable the direct use of split functions without positional y in their signatures. This is useful when working with unsupervised data (#605).
  • to_numpy is now able to unpack dicts and lists/tuples (#657, #658)
  • When using CrossEntropyLoss, softmax is now automatically applied to the output when calling predict or predict_proba

Fixed

  • Fixed a bug where CyclicLR scheduler would update during both training and validation rather than just during training.
  • Fixed a bug introduced by moving the optimizer.zero_grad() call outside of the train step function, making it incompatible with LBFGS and other optimizers that call the train step several times per batch (#636)
  • Fixed pickling of the ProgressBar callback (#656)

0.8.0 - 2020-04-11

Added

  • Added NeptuneLogger callback for logging experiment metadata to neptune.ai (#586)
  • Add DataFrameTransformer, an sklearn compatible transformer that helps working with pandas DataFrames by transforming the DataFrame into a representation that works well with neural networks (#507)
  • Added WandbLogger callback for logging to Weights & Biases (#607)
  • Added None option to device which leaves the device(s) unmodified (#600)
  • Add PassthroughScoring, a scoring callback that just calculates the average score of a metric determined at batch level and then writes it to the epoch level (#595)

Changed

  • When using caching in scoring callbacks, no longer uselessly iterate over the data; this can save time if iteration is slow (#552, #557)
  • Cleaned up duplicate code in the fit_loop (#564)

Future Changes

  • WARNING: In release 0.10.0 of skorch, Python 3.5 support will be officially dropped (#634)

Fixed

  • Make skorch compatible with sklearn 0.22 (#571, #573, #575)
  • Fixed a bug that could occur when a new "settable" (via set_params) attribute was added to NeuralNet whose name starts the same as an existing attribute's name (#590)

0.7.0 - 2019-11-29

Added

  • More careful check for wrong parameter names being passed to NeuralNet (#500)
  • More helpful error messages when trying to predict using an uninitialized model
  • Add TensorBoard callback for automatic logging to tensorboard
  • Make NeuralNetBinaryClassifier work with sklearn.calibration.CalibratedClassifierCV
  • Improve NeuralNetBinaryClassifier compatibility with certain sklearn metrics (#515)
  • NeuralNetBinaryClassifier automatically squeezes module output if necessary (#515)
  • NeuralNetClassifier now has a classes_ attribute after fit is called, which is inferred from y by default (#465, #486)
  • NeuralNet.load_params with a checkpoint now initializes when needed (#497)

Changed

  • Improve numerical stability when using NLLLoss in NeuralNetClassifer (#491)
  • Refactor code to make gradient accumulation easier to implement (#506)
  • NeuralNetBinaryClassifier.predict_proba now returns a 2-dim array; to access the "old" y_proba, take y_proba[:, 1] (#515)
  • net.history is now a property that accesses net.history_, which stores the History object (#527)
  • Remove deprecated skorch.callbacks.CyclicLR, use torch.optim.lr_scheduler.CyclicLR instead

Future Changes

  • WARNING: In a future release, the behavior of method net.get_params will change to make it more consistent with sklearn: it will no longer return "learned" attributes like module_. Therefore, functions like sklearn.base.clone, when called with a fitted net, will no longer return a fitted net but instead an uninitialized net. If you want a copy of a fitted net, use copy.deepcopy instead. Note that net.get_params is used under the hood by many sklearn functions and classes, such as GridSearchCV, whose behavior may thus be affected by the change. (#521, #527)

Fixed

  • Fixed a bug that caused LoadInitState not to work with TrainEndCheckpoint (#528)
  • Fixed NeuralNetBinaryClassifier wrongly squeezing the batch dimension when using batch_size = 1 (#558)

0.6.0 - 2019-07-19

Added

  • Adds FAQ entry regarding the initialization behavior of NeuralNet when passed instantiated models. (#409)
  • Added CUDA pickle test including an artifact that supports testing on CUDA-less CI machines
  • Adds train_batch_count and valid_batch_count to history in training loop. (#445)
  • Adds score method for NeuralNetClassifier, NeuralNetBinaryClassifier, and NeuralNetRegressor (#469)
  • Wrapper class for torch Datasets to make them work with some sklearn features (e.g. grid search). (#443)

Changed

  • Repository moved to https://github.com/skorch-dev/skorch/, please change your git remotes
  • Treat cuda dependent attributes as prefix to cover values set using set_params since previously "criterion_" would not match net.criterion__weight as set by net.set_params(criterion__weight=w)
  • skorch pickle format changed in order to improve CUDA compatibility, if you have pickled models, please re-pickle them to be able to load them in the future
  • net.criterion_ and its parameters are now moved to target device when using criteria that inherit from torch.nn.Module. Previously the user had to make sure that parameters such as class weight are on the compute device
  • skorch now assumes PyTorch >= 1.1.0. This mainly affects learning rate schedulers, whose inner workings have been changed with version 1.1.0. This update will also invalidate pickled skorch models after a change introduced in PyTorch optimizers.

Fixed

  • Include requirements in MANIFEST.in
  • Add criterion_ to NeuralNet.cuda_dependent_attributes_ to avoid issues with criterion weight tensors from, e.g., NLLLoss (#426)
  • TrainEndCheckpoint can be cloned by sklearn.base.clone. (#459)

0.5.0 - 2018-12-13

Added

  • Basic usage notebook now runs on Google Colab
  • Advanced usage notebook now runs on Google Colab
  • MNIST with scikit-learn and skorch now runs on Google Colab
  • Better user-facing messages when module or optimizer are re-initialized
  • Added an experimental API (net._register_virtual_param) to register "virtual" parameters on the network with custom setter functions. (#369)
  • Setting parameters lr, momentum, optimizer__lr, etc. no longer resets the optmizer. As of now you can do net.set_params(lr=0.03) or net.set_params(optimizer__param_group__0__momentum=0.86) without triggering a re-initialization of the optimizer (#369)
  • Support for scipy sparse CSR matrices as input (as, e.g., returned by sklearn's CountVectorizer); note that they are cast to dense matrices during batching
  • Helper functions to build command line interfaces with almost no boilerplate, example that shows usage

Changed

  • Reduce overhead of BatchScoring when using train_loss_score or valid_loss_score by skipping superfluous inference step (#381)
  • The on_grad_computed callback function will yield an iterable for named_parameters only when it is used to reduce the run-time overhead of the call (#379)
  • Default fn_prefix in TrainEndCheckpoint is now train_end_ (#391)
  • Issues a warning when Checkpoints's monitor parameter is set to monitor and the history contains <monitor>_best. (#399)

Fixed

  • Re-initialize optimizer when set_params is called with lr argument (#372)
  • Copying a SliceDict now returns a SliceDict instead of a dict (#388)
  • Calling == on SliceDicts now works as expected when values are numpy arrays and torch tensors

0.4.0 - 2018-10-24

Added

  • Support for PyTorch 0.4.1
  • There is no need to explicitly name callbacks anymore (names are assigned automatically, name conflicts are resolved).
  • You can now access the training data in the on_grad_computed event
  • There is a new image segmentation example
  • Easily create toy network instances for quick experiments using skorch.toy.make_classifier and friends
  • New ParamMapper callback to modify/freeze/unfreeze parameters at certain point in time during training:
>>> from sklearn.callbacks import Freezer, Unfreezer
>>> net = Net(module, callbacks=[Freezer('layer*.weight'), Unfreezer('layer*.weight', at=10)])
  • Refactored EpochScoring for easier sub-classing
  • Checkpoint callback now supports saving the optimizer, this avoids problems with stateful optimizers such as Adam or RMSprop (#360)
  • Added LoadInitState callback for easy continued training from checkpoints (#360)
  • NeuralNetwork.load_params now supports loading from Checkpoint instances
  • Added documentation for saving and loading

Changed

  • The ProgressBar callback now determines the batches per epoch automatically by default (batches_per_epoch=auto)
  • The on_grad_computed event now has access to the current training data batch

Deprecated

  • Deprecated filtered_optimizer in favor of Freezer callback (#346)
  • NeuralNet.load_params and NeuralNet.save_params deprecate f parameter for the sake of f_optimizer, f_params and f_history (#360)

Fixed

  • uses_placeholder_y should not require existence of y field (#311)
  • LR scheduler creates batch_idx on first run (#314)
  • Use OrderedDict for callbacks to fix python 3.5 compatibility issues (#331)
  • Make to_tensor work correctly with PackedSequence (#335)
  • Rewrite History to not use any recursion to avoid memory leaks during exceptions (#312)
  • Use flaky in some neural network tests to hide platform differences
  • Fixes ReduceLROnPlateau when mode == max (#363)
  • Fix disconnected weights between net and optimizer after copying the net with copy.deepcopy (#318)
  • Fix a bug that intefered with loading CUDA models when the model was a CUDA tensor but the net was configured to use the CPU (#354, #358)