Skip to content

Latest commit

 

History

History
166 lines (159 loc) · 9.06 KB

CHANGELOG.md

File metadata and controls

166 lines (159 loc) · 9.06 KB
Raw

Change log

v0.4.0

  • Support param version 2
  • Added pydrobert.torch.argcheck and standardized module argument checking a bit.
  • Added environment.yml for local dev.
  • LookupLanguageModel has been refactored and reimplemented. It is no longer compatible with previous versions of the model.
  • parse_arpa has been enhanced: it can handle log-probs in scientific notation (e.g. 1e4), conversion from (implicitly) log-base-10 to log-base-e probabilities; and storing log probabilities as Numpy floating-point types.
  • LookupLanguageModel and parse_arpa now have an optional logger argument to log progress building the trie and parsing the file, respectively.
  • Added best_is_train flag to TrainingController.update_for_epoch
  • Refactored get-torch-spect-data-dir-info to be faster
  • subset-torch-spect-data-dir command has been added
  • print-torch-{ali,ref}-data-dir-length-moments commands have been added
  • LookupLanguageMode.prob_list has been renamed to prob_dicts
  • Added ShallowFusionLanguageModel, ExtractableShallowFusionLanguageModel, and MixableShallowFusionLanguageModel
  • Slicing and chunking modules SliceSpectData, ChunkBySlices, and ChunkTokenSequenceBySlices, as well as the command chunk-torch-spect-data-dir, which puts them together.
  • Code for handling TextGrid files, including the functions read_textgrid and write_textgrid, as well as the commands torch-token-data-dir-to-textgrids and textgrids-to-torch-token-data-dir.
  • Commands for switching between ref and ali format: torch-ali-data-dir-to-torch-token-data-dir and torch-token-data-dir-to-torch-ali-data-dir.
  • Added py 3.10 support; removed py 3.6 support.
  • Initial (undocumented) support for PyTorch-Lightning in pydrobert.torch.lightning submodule. Will document when I get some time.
  • Refactored much of pydrobert.torch.data. Best just to look at the API. ContextWindowEvaluationDataLoader, ContextWindowTrainingDataLoader, SpectEvaluationDataLoader, SpectTrainingDataLoader, DataSetParams, SpectDataSetParams, and ContextWindowDataSetParams are now deprecated. The data loaders have been simplified to ContextWindowDataLoader and SpectDataLoader. Keyword arguments (like shuffle) now control their behaviour. The *DataSetParams have been renamed *DataLoaderParams with some of the parameters moved around. Notably, LangDataParams now stores sos, eos, and subset_ids parameters, from which a number of parameter objects inherit. SpectDataLoaderParams inherits from LangDataLoaderParams, which in turn inherits from DynamicLengthDataLoaderParams. The latter allows the loader's batch elements to be bucketed by length using the new BucketBatchSampler. It and a number of other samplers inherit from AbstractEpochSampler to help facilitate the simplified loaders and better resemble the PyTorch API. Mean-variance normalization of features is possible through the loaders and the new MeanVarianceNormalization module. LangDataSet and LangDataLoader have been introduced to facilitate language mdoel training. Finally, loaders (and samplers) are compatible with DistributedDataParallel environments.
  • Mean-variance statistics for normalization may be estimated from a data partition using the command compute-mvn-stats-for-torch-feat-data-dir.
  • Added torch-spect-data-dir-to-wds to convert a data dir to a WebDataset.
  • Changed method of constructing random state in EpochRandomSampler. Rerunning training on this new version with the same seed will likely result in different results from the old version!
  • FeatureDeltas now a module, in case you want to compute them online rather than waste disk space.
  • Added PadMaskedSequence.
  • Added FillAfterEndOfSequence.
  • Added binomial_coefficient, enumerate_binary_sequences, enumerate_vocab_sequences, and enumerate_binary_sequences_with_cardinality.
  • Docstrings updated to hopefully be clearer. Use "Call Parameters" and "Returns" sections for pytorch modules.
  • readthedocs updated.
  • Fixed up formatting of CLI help documentation.
  • Data sets can now initialize some of their parameters with the values in their associated param containers. For example, sos and eos are now set in SpectDataSet by passing an optional SpectDataParam instance. The old method (by argument) is now deprecated.
  • Renamed DataSetParams to DataLoaderParams and deprecated former naming to better mesh with their use in data loaders.
  • Moved pydrobert.torch.util.parse_arpa_lm to pydrobert.torch.data
  • SimpleRandomSamplingWithoutReplacement has been added as a new distribution.
  • EnumerateEstimator, ImportanceSamplingEstimator, and IndependentMetropolisHastingsEstimator have been added as a new estimators.
  • pydrobert.torch.estimators has been rewritten from the ground-up, with old functionality deprecated. Distribution-related functions have been rewritten as torch.distributions.Distribution classes implementing a ConditionalStraightThrough interface and stored in pydrobert.torch.distributions. The REINFORCE and RELAX estimators now have an object-oriented interface subclassing MonteCarloEstimator as DirectEstimator and RelaxEstimator, respectively. The REBAR control variate is now distribution-specific and found in pydrobert.torch.modules.
  • Bug fixes to OptimalCompletion and HardOptimalCompletionDistillationLoss involving batch sizes.
  • Refactored code to move modules to pydrobert.torch.modules and functions to pydrobert.torch.functional.
  • Deprecated pydrobert.torch.layers and pydrobert.torch.util.
  • Added a number of modules to pydrobert.torch.modules to wrap functional API. Moved docstrings to modules.
  • Fixed a problem with warp_1d_grid/SpecAugment which made it sensitive to the length of other elements in the batch.
  • Added compatibility wrappers to avoid warnings across supported pytorch versions.
  • Refactored code and added tests to support JIT tracing and scripting for most functions/modules in pytorch >= 1.8.1. before the next release. I'll write up documentation shortly.
  • Added pydrobert.torch.config to store constants used in the module.
  • Removed setup.py.
  • Deleted conda recipe in prep for conda-forge.
  • Compatibility/determinism fixes for 1.5.1.
  • Bump minimum PyTorch version to 1.5.1. Actually testing this minimum!
  • version.py -> _version.py.
  • A number of modifications and additions related to decoding and language models, including:
    • beam_search_advance and random_walk_advance have been simplified, with much of the end-of-sequence logic punted to their associated modules.
    • Rejigged SequentialLanguageModel and LookupLanguageModel to be both simpler and compatible with decoder interfaces.
    • ctc_greedy_search and ctc_prefix_search_advance functions have been added.
    • ExtractableSequentialLanguageModel, MixableSequentialLanguageModel, BeamSearch, RandomWalk, and CTCPrefixSearch modules have been added.
    • A SequentialLanguageModelDistribution wrapping RandomWalk which implements PyTorch's Distribution interface. Language models now work with estimators!
    • A new documentation page on how to deal with all of that.
  • Fixed bug in controller that always compared thresholds against best, not the last point that reset the countdown (#55)
  • Added pad_variable and RandomShift (#54)
  • Modified error_rate, prefix_error_rates to actually compute error rates when non-default costs are used. Old functionality is now in edit_distance and prefix_edit_distances (#51)
  • Fixed bug in how padding is handled in string matching utilities.
  • Fixed logic errors in compute-torch-token-data-dir-error-rates (#50)
  • Modified frame end in pydrobert.torch.data.transcript_to_token and added some notes on the ambiguity of the conversion.
  • Added some more checks and a 'fix' flag to pydrobert.torch.data.validate_spect_data_set. Entry get-torch-spect-data-dir-info now has --fix flag, too.

v0.3.0

A considerable amount of refactoring occurred for this build, chiefly to get rid of Python 2.7 support. While the functionality did not change much for this version, we have switched from a pkgutil-style pydrobert namespace to PEP-420-style namespaces. As a result, this package is not backwards-compatible with previous pydrobert packages! Make sure that if any of the following are installed, they exceed the following version thresholds:

  • pydrobert-param >0.2.0
  • pydrobert-kaldi >0.5.3
  • pydrobert-speech >0.1.0

Miscellaneous other stuff:

  • Type hints everywhere
  • Shifted python source to src/
  • Black-formatted remaining source
  • Removed future dependency
  • Shifted most of the configuration to setup.cfg, leaving only a shell in setup.py to remain compatible with Conda builds
  • Added pyproject.toml for PEP 517.
  • tox.ini for TOX testing
  • Switched to AppVeyor for CI
  • Added changelog :D