Skip to content

Latest commit

 

History

History
928 lines (712 loc) · 57.1 KB

CHANGELOG.md

File metadata and controls

928 lines (712 loc) · 57.1 KB
Raw

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.

v2.10.1 - 2022-10-18

Fixed

  • Updated dependencies

v2.10.0 - 2022-07-14

Added

  • Added metric FBetaVerboseMeasure which extends FBetaMeasure to ensure compatibility with logging plugins and add some options.
  • Added three sample weighting techniques to ConditionalRandomField by supplying three new subclasses: ConditionalRandomFieldWeightEmission, ConditionalRandomFieldWeightTrans, and ConditionalRandomFieldWeightLannoy.

Fixed

  • Fix error from cached-path version update.

v2.9.3 - 2022-04-13

Added

  • Added verification_tokens argument to TestPretrainedTransformerTokenizer.

Fixed

  • Updated various dependencies

v2.9.2 - 2022-03-21

Fixed

  • Removed unnecessary dependencies
  • Restore functionality of CLI in absence of now-optional checklist-package

v2.9.1 - 2022-03-09

Fixed

  • Updated dependencies, especially around doc creation.
  • Running the test suite out-of-tree (e.g. after installation) is now possible by pointing the environment variable ALLENNLP_SRC_DIR to the sources.
  • Silenced a warning that happens when you inappropriately clone a tensor.
  • Adding more clarification to the Vocabulary documentation around min_pretrained_embeddings and only_include_pretrained_words.
  • Fixed bug with type mismatch caused by latest release of cached-path that now returns a Path instead of a str.

Added

  • We can now transparently read compressed input files during prediction.
  • LZMA compression is now supported.
  • Added a way to give JSON blobs as input to dataset readers in the evaluate command.
  • Added the argument sub_module in PretrainedTransformerMismatchedEmbedder
  • Updated the docs for PytorchSeq2VecWrapper to specify that mask is required rather than sequence lengths for clarity.

Changed

  • You can automatically include all words from a pretrained file when building a vocabulary by setting the value in min_pretrained_embeddings to -1 for that particular namespace.

v2.9.0 - 2022-01-27

Added

  • Added an Evaluator class to make comparing source, target, and predictions easier.
  • Added a way to resize the vocabulary in the T5 module
  • Added an argument reinit_modules to cached_transformers.get() that allows you to re-initialize the pretrained weights of a transformer model, using layer indices or regex strings.
  • Added attribute _should_validate_this_epoch to GradientDescentTrainer that controls whether validation is run at the end of each epoch.
  • Added ShouldValidateCallback that can be used to configure the frequency of validation during training.
  • Added a MaxPoolingSpanExtractor. This SpanExtractor represents each span by a component wise max-pooling-operation.
  • Added support for dist_metric kwarg in initializing fairness metrics, which allows optionally using wasserstein distance (previously only KL divergence was supported).

Fixed

  • Fixed the docstring information for the FBetaMultiLabelMeasure metric.
  • Various fixes for Python 3.9
  • Fixed the name that the push-to-hf command uses to store weights.
  • FBetaMultiLabelMeasure now works with multiple dimensions
  • Support for inferior operating systems when making hardlinks
  • Use , as a separator for filenames in the evaluate command, thus allowing for URLs (eg. gs://...) as input files.
  • Removed a spurious error message "'torch.cuda' has no attribute '_check_driver'" that would be appear in the logs when a ConfigurationError for missing GPU was raised.
  • Load model on CPU post training to save GPU memory.
  • Fixed a bug in ShouldValidateCallback that leads to validation occuring after the first epoch regardless of validation_start value.
  • Fixed a bug in ShouldValidateCallback that leads to validation occuring every validation_interval + 1 epochs, instead of every validation_interval epochs.
  • Fixed a bug in ShouldValidateCallback that leads to validation never occuring at the end of training.

Removed

Changed

  • Make checklist an optional dependency.

v2.8.0 - 2021-11-01

Added

  • Added support to push models directly to the Hugging Face Hub with the command allennlp push-to-hf.
  • More default tests for the TextualEntailmentSuite.

Changed

  • The behavior of --overrides has changed. Previously the final configuration params were simply taken as the union over the original params and the --overrides params. But now you can use --overrides to completely replace any part of the original config. For example, passing --overrides '{"model":{"type":"foo"}}' will completely replace the "model" part of the original config. However, when you just want to change a single field in the JSON structure without removing / replacing adjacent fields, you can still use the "dot" syntax. For example, --overrides '{"model.num_layers":3}' will only change the num_layers parameter to the "model" part of the config, leaving everything else unchanged.
  • Integrated cached_path library to replace existing functionality in common.file_utils. This introduces some improvements without any breaking changes.

Fixed

  • Fixed the implementation of PairedPCABiasDirection in allennlp.fairness.bias_direction, where the difference vectors should not be centered when performing the PCA.
  • Fixed the docstring of ExponentialMovingAverage, which was causing its argument descriptions to render inccorrectly in the docs.

v2.7.0 - 2021-09-01

Added

  • Added in a default behavior to the _to_params method of Registrable so that in the case it is not implemented by the child class, it will still produce a parameter dictionary.
  • Added in _to_params implementations to all tokenizers.
  • Added support to evaluate mutiple datasets and produce corresponding output files in the evaluate command.
  • Added more documentation to the learning rate schedulers to include a sample config object for how to use it.
  • Moved the pytorch learning rate schedulers wrappers to their own file called pytorch_lr_schedulers.py so that they will have their own documentation page.
  • Added a module allennlp.nn.parallel with a new base class, DdpAccelerator, which generalizes PyTorch's DistributedDataParallel wrapper to support other implementations. Two implementations of this class are provided. The default is TorchDdpAccelerator (registered at "torch"), which is just a thin wrapper around DistributedDataParallel. The other is FairScaleFsdpAccelerator, which wraps FairScale's FullyShardedDataParallel. You can specify the DdpAccelerator in the "distributed" section of a configuration file under the key "ddp_accelerator".
  • Added a module allennlp.nn.checkpoint with a new base class, CheckpointWrapper, for implementations of activation/gradient checkpointing. Two implentations are provided. The default implementation is TorchCheckpointWrapper (registered as "torch"), which exposes PyTorch's checkpoint functionality. The other is FairScaleCheckpointWrapper which exposes the more flexible checkpointing funtionality from FairScale.
  • The Model base class now takes a ddp_accelerator parameter (an instance of DdpAccelerator) which will be available as self.ddp_accelerator during distributed training. This is useful when, for example, instantiating submodules in your model's __init__() method by wrapping them with self.ddp_accelerator.wrap_module(). See the allennlp.modules.transformer.t5 for an example.
  • We now log batch metrics to tensorboard and wandb.
  • Added Tango components, to be explored in detail in a later post
  • Added ScaledDotProductMatrixAttention, and converted the transformer toolkit to use it
  • Added tests to ensure that all Attention and MatrixAttention implementations are interchangeable
  • Added a way for AllenNLP Tango to read and write datasets lazily.
  • Added a way to remix datasets flexibly
  • Added from_pretrained_transformer_and_instances constructor to Vocabulary
  • TransformerTextField now supports __len__.

Fixed

  • Fixed a bug in ConditionalRandomField: transitions and tag_sequence tensors were not initialized on the desired device causing high CPU usage (see allenai/allennlp#2884)
  • Fixed a mispelling: the parameter contructor_extras in Lazy() is now correctly called constructor_extras.
  • Fixed broken links in allennlp.nn.initializers docs.
  • Fixed bug in BeamSearch where last_backpointers was not being passed to any Constraints.
  • TransformerTextField can now take tensors of shape (1, n) like the tensors produced from a HuggingFace tokenizer.
  • tqdm lock is now set inside MultiProcessDataLoading when new workers are spawned to avoid contention when writing output.
  • ConfigurationError is now pickleable.
  • Checkpointer cleaning was fixed to work on Windows Paths
  • Multitask models now support TextFieldTensor in heads, not just in the backbone.
  • Fixed the signature of ScaledDotProductAttention to match the other Attention classes
  • allennlp commands will now catch SIGTERM signals and handle them similar to SIGINT (keyboard interrupt).
  • The MultiProcessDataLoader will properly shutdown its workers when a SIGTERM is received.
  • Fixed the way names are applied to Tango Step instances.
  • Fixed a bug in calculating loss in the distributed setting.
  • Fixed a bug when extending a sparse sequence by 0 items.

Changed

  • The type of the grad_norm parameter of GradientDescentTrainer is now Union[float, bool], with a default value of False. False means gradients are not rescaled and the gradient norm is never even calculated. True means the gradients are still not rescaled but the gradient norm is calculated and passed on to callbacks. A float value means gradients are rescaled.
  • TensorCache now supports more concurrent readers and writers.
  • We no longer log parameter statistics to tensorboard or wandb by default.

v2.6.0 - 2021-07-19

Added

  • Added on_backward training callback which allows for control over backpropagation and gradient manipulation.
  • Added AdversarialBiasMitigator, a Model wrapper to adversarially mitigate biases in predictions produced by a pretrained model for a downstream task.
  • Added which_loss parameter to ensure_model_can_train_save_and_load in ModelTestCase to specify which loss to test.
  • Added **kwargs to Predictor.from_path(). These key-word argument will be passed on to the Predictor's constructor.
  • The activation layer in the transformer toolkit now can be queried for its output dimension.
  • TransformerEmbeddings now takes, but ignores, a parameter for the attention mask. This is needed for compatibility with some other modules that get called the same way and use the mask.
  • TransformerPooler can now be instantiated from a pretrained transformer module, just like the other modules in the transformer toolkit.
  • TransformerTextField, for cases where you don't care about AllenNLP's advanced text handling capabilities.
  • Added TransformerModule._post_load_pretrained_state_dict_hook() method. Can be used to modify missing_keys and unexpected_keys after loading a pretrained state dictionary. This is useful when tying weights, for example.
  • Added an end-to-end test for the Transformer Toolkit.
  • Added vocab argument to BeamSearch, which is passed to each contraint in constraints (if provided).

Fixed

  • Fixed missing device mapping in the allennlp.modules.conditional_random_field.py file.
  • Fixed Broken link in allennlp.fairness.fairness_metrics.Separation docs
  • Ensured all allennlp submodules are imported with allennlp.common.plugins.import_plugins().
  • Fixed IndexOutOfBoundsException in MultiOptimizer when checking if optimizer received any parameters.
  • Removed confusing zero mask from VilBERT.
  • Ensured ensure_model_can_train_save_and_load is consistently random.
  • Fixed weight tying logic in T5 transformer module. Previously input/output embeddings were always tied. Now this is optional, and the default behavior is taken from the config.tie_word_embeddings value when instantiating from_pretrained_module().
  • Implemented slightly faster label smoothing.
  • Fixed the docs for PytorchTransformerWrapper
  • Fixed recovering training jobs with models that expect get_metrics() to not be called until they have seen at least one batch.
  • Made the Transformer Toolkit compatible with transformers that don't start their positional embeddings at 0.
  • Weights & Biases training callback ("wandb") now works when resuming training jobs.

Changed

  • Changed behavior of MultiOptimizer so that while a default optimizer is still required, an error is not thrown if the default optimizer receives no parameters.
  • Made the epsilon parameter for the layer normalization in token embeddings configurable.

Removed

  • Removed TransformerModule._tied_weights. Weights should now just be tied directly in the __init__() method. You can also override TransformerModule._post_load_pretrained_state_dict_hook() to remove keys associated with tied weights from missing_keys after loading a pretrained state dictionary.

v2.5.0 - 2021-06-03

Added

  • Added TaskSuite base class and command line functionality for running checklist test suites, along with implementations for SentimentAnalysisSuite, QuestionAnsweringSuite, and TextualEntailmentSuite. These can be found in the allennlp.confidence_checks.task_checklists module.
  • Added BiasMitigatorApplicator, which wraps any Model and mitigates biases by finetuning on a downstream task.
  • Added allennlp diff command to compute a diff on model checkpoints, analogous to what git diff does on two files.
  • Meta data defined by the class allennlp.common.meta.Meta is now saved in the serialization directory and archive file when training models from the command line. This is also now part of the Archive named tuple that's returned from load_archive().
  • Added nn.util.distributed_device() helper function.
  • Added allennlp.nn.util.load_state_dict helper function.
  • Added a way to avoid downloading and loading pretrained weights in modules that wrap transformers such as the PretrainedTransformerEmbedder and PretrainedTransformerMismatchedEmbedder. You can do this by setting the parameter load_weights to False. See PR #5172 for more details.
  • Added SpanExtractorWithSpanWidthEmbedding, putting specific span embedding computations into the _embed_spans method and leaving the common code in SpanExtractorWithSpanWidthEmbedding to unify the arguments, and modified BidirectionalEndpointSpanExtractor, EndpointSpanExtractor and SelfAttentiveSpanExtractor accordingly. Now, SelfAttentiveSpanExtractor can also embed span widths.
  • Added a min_steps parameter to BeamSearch to set a minimum length for the predicted sequences.
  • Added the FinalSequenceScorer abstraction to calculate the final scores of the generated sequences in BeamSearch.
  • Added shuffle argument to BucketBatchSampler which allows for disabling shuffling.
  • Added allennlp.modules.transformer.attention_module which contains a generalized AttentionModule. SelfAttention and T5Attention both inherit from this.
  • Added a Constraint abstract class to BeamSearch, which allows for incorporating constraints on the predictions found by BeamSearch, along with a RepeatedNGramBlockingConstraint constraint implementation, which allows for preventing repeated n-grams in the output from BeamSearch.
  • Added DataCollator for dynamic operations for each batch.

Changed

  • Use dist_reduce_sum in distributed metrics.
  • Allow Google Cloud Storage paths in cached_path ("gs://...").
  • Renamed nn.util.load_state_dict() to read_state_dict to avoid confusion with torch.nn.Module.load_state_dict().
  • TransformerModule.from_pretrained_module now only accepts a pretrained model ID (e.g. "bert-base-case") instead of an actual torch.nn.Module. Other parameters to this method have changed as well.
  • Print the first batch to the console by default.
  • Renamed sanity_checks to confidence_checks (sanity_checks is deprecated and will be removed in AllenNLP 3.0).
  • Trainer callbacks can now store and restore state in case a training run gets interrupted.
  • VilBERT backbone now rolls and unrolls extra dimensions to handle input with > 3 dimensions.
  • BeamSearch is now a Registrable class.

Fixed

  • When PretrainedTransformerIndexer folds long sequences, it no longer loses the information from token type ids.
  • Fixed documentation for GradientDescentTrainer.cuda_device.
  • Re-starting a training run from a checkpoint in the middle of an epoch now works correctly.
  • When using the "moving average" weights smoothing feature of the trainer, training checkpoints would also get smoothed, with strange results for resuming a training job. This has been fixed.
  • When re-starting an interrupted training job, the trainer will now read out the data loader even for epochs and batches that can be skipped. We do this to try to get any random number generators used by the reader or data loader into the same state as they were the first time the training job ran.
  • Fixed the potential for a race condition with cached_path() when extracting archives. Although the race condition is still possible if used with force_extract=True.
  • Fixed wandb callback to work in distributed training.
  • Fixed tqdm logging into multiple files with allennlp-optuna.

v2.4.0 - 2021-04-22

Added

  • Added a T5 implementation to modules.transformers.

Changed

  • Weights & Biases callback can now work in anonymous mode (i.e. without the WANDB_API_KEY environment variable).

Fixed

  • The GradientDescentTrainer no longer leaves stray model checkpoints around when it runs out of patience.
  • Fixed cached_path() for "hf://" files.
  • Improved the error message for the PolynomialDecay LR scheduler when num_steps_per_epoch is missing.

v2.3.1 - 2021-04-20

Added

  • Added support for the HuggingFace Hub as an alternative way to handle loading files. Hub downloads should be made through the hf:// URL scheme.
  • Add new dimension to the interpret module: influence functions via the InfluenceInterpreter base class, along with a concrete implementation: SimpleInfluence.
  • Added a quiet parameter to the MultiProcessDataLoading that disables Tqdm progress bars.
  • The test for distributed metrics now takes a parameter specifying how often you want to run it.
  • Created the fairness module and added three fairness metrics: Independence, Separation, and Sufficiency.
  • Added four bias metrics to the fairness module: WordEmbeddingAssociationTest, EmbeddingCoherenceTest, NaturalLanguageInference, and AssociationWithoutGroundTruth.
  • Added four bias direction methods (PCABiasDirection, PairedPCABiasDirection, TwoMeansBiasDirection, ClassificationNormalBiasDirection) and four bias mitigation methods (LinearBiasMitigator, HardBiasMitigator, INLPBiasMitigator, OSCaRBiasMitigator).

Changed

  • Updated CONTRIBUTING.md to remind reader to upgrade pip setuptools to avoid spaCy installation issues.

Fixed

  • Fixed a bug with the ShardedDatasetReader when used with multi-process data loading (allenai/allennlp#5132).

v2.3.0 - 2021-04-14

Added

  • Ported the following Huggingface LambdaLR-based schedulers: ConstantLearningRateScheduler, ConstantWithWarmupLearningRateScheduler, CosineWithWarmupLearningRateScheduler, CosineHardRestartsWithWarmupLearningRateScheduler.
  • Added new sub_token_mode parameter to pretrained_transformer_mismatched_embedder class to support first sub-token embedding
  • Added a way to run a multi task model with a dataset reader as part of allennlp predict.
  • Added new eval_mode in PretrainedTransformerEmbedder. If it is set to True, the transformer is always run in evaluation mode, which, e.g., disables dropout and does not update batch normalization statistics.
  • Added additional parameters to the W&B callback: entity, group, name, notes, and wandb_kwargs.

Changed

  • Sanity checks in the GradientDescentTrainer can now be turned off by setting the run_sanity_checks parameter to False.
  • Allow the order of examples in the task cards to be specified explicitly
  • histogram_interval parameter is now deprecated in TensorboardWriter, please use distribution_interval instead.
  • Memory usage is not logged in tensorboard during training now. ConsoleLoggerCallback should be used instead.
  • If you use the min_count parameter of the Vocabulary, but you specify a namespace that does not exist, the vocabulary creation will raise a ConfigurationError.
  • Documentation updates made to SoftmaxLoss regarding padding and the expected shapes of the input and output tensors of forward.
  • Moved the data preparation script for coref into allennlp-models.
  • If a transformer is not in cache but has override weights, the transformer's pretrained weights are no longer downloaded, that is, only its config.json file is downloaded.
  • SanityChecksCallback now raises SanityCheckError instead of AssertionError when a check fails.
  • jsonpickle removed from dependencies.
  • Improved the error message from Registrable.by_name() when the name passed does not match any registered subclassess. The error message will include a suggestion if there is a close match between the name passed and a registered name.

Fixed

  • Fixed a bug where some Activation implementations could not be pickled due to involving a lambda function.
  • Fixed __str__() method on ModelCardInfo class.
  • Fixed a stall when using distributed training and gradient accumulation at the same time
  • Fixed an issue where using the from_pretrained_transformer Vocabulary constructor in distributed training via the allennlp train command would result in the data being iterated through unnecessarily.
  • Fixed a bug regarding token indexers with the InterleavingDatasetReader when used with multi-process data loading.
  • Fixed a warning from transformers when using max_length in the PretrainedTransformerTokenizer.

Removed

  • Removed the stride parameter to PretrainedTransformerTokenizer. This parameter had no effect.

v2.2.0 - 2021-03-26

Added

  • Add new method on Field class: .human_readable_repr() -> Any
  • Add new method on Instance class: .human_readable_dict() -> JsonDict.
  • Added WandBCallback class for Weights & Biases integration, registered as a callback under the name "wandb".
  • Added TensorBoardCallback to replace the TensorBoardWriter. Registered as a callback under the name "tensorboard".
  • Added NormalizationBiasVerification and SanityChecksCallback for model sanity checks.
  • SanityChecksCallback runs by default from the allennlp train command. It can be turned off by setting trainer.enable_default_callbacks to false in your config.

Changed

  • Use attributes of ModelOutputs object in PretrainedTransformerEmbedder instead of indexing.
  • Added support for PyTorch version 1.8 and torchvision version 0.9 .
  • Model.get_parameters_for_histogram_tensorboard_logging is deprecated in favor of Model.get_parameters_for_histogram_logging.

Fixed

  • Makes sure tensors that are stored in TensorCache always live on CPUs
  • Fixed a bug where FromParams objects wrapped in Lazy() couldn't be pickled.
  • Fixed a bug where the ROUGE metric couldn't be picked.
  • Fixed a bug reported by allenai/allennlp#5036. We keeps our spacy POS tagger on.

Removed

  • Removed TensorBoardWriter. Please use the TensorBoardCallback instead.

v2.1.0 - 2021-02-24

Changed

  • coding_scheme parameter is now deprecated in Conll2003DatasetReader, please use convert_to_coding_scheme instead.
  • Support spaCy v3

Added

  • Added ModelUsage to ModelCard class.
  • Added a way to specify extra parameters to the predictor in an allennlp predict call.
  • Added a way to initialize a Vocabulary from transformers models.
  • Added the ability to use Predictors with multitask models through the new MultiTaskPredictor.
  • Added an example for fields of type ListField[TextField] to apply_token_indexers API docs.
  • Added text_key and label_key parameters to TextClassificationJsonReader class.
  • Added MultiOptimizer, which allows you to use different optimizers for different parts of your model.
  • Added a clarification to predictions_to_labeled_instances API docs for attack from json

Fixed

  • @Registrable.register(...) decorator no longer masks the decorated class's annotations
  • Ensured that MeanAbsoluteError always returns a float metric value instead of a Tensor.
  • Learning rate schedulers that rely on metrics from the validation set were broken in v2.0.0. This brings that functionality back.
  • Fixed a bug where the MultiProcessDataLoading would crash when num_workers > 0, start_method = "spawn", max_instances_in_memory not None, and batches_per_epoch not None.
  • Fixed documentation and validation checks for FBetaMultiLabelMetric.
  • Fixed handling of HTTP errors when fetching remote resources with cached_path(). Previously the content would be cached even when certain errors - like 404s - occurred. Now an HTTPError will be raised whenever the HTTP response is not OK.
  • Fixed a bug where the MultiTaskDataLoader would crash when num_workers > 0
  • Fixed an import error that happens when PyTorch's distributed framework is unavailable on the system.

v2.0.1 - 2021-01-29

Added

  • Added tokenizer_kwargs and transformer_kwargs arguments to PretrainedTransformerBackbone
  • Resize transformers word embeddings layer for additional_special_tokens

Changed

  • GradientDescentTrainer makes serialization_dir when it's instantiated, if it doesn't exist.

Fixed

  • common.util.sanitize now handles sets.

v2.0.0 - 2021-01-27

Added

  • The TrainerCallback constructor accepts serialization_dir provided by Trainer. This can be useful for Logger callbacks those need to store files in the run directory.
  • The TrainerCallback.on_start() is fired at the start of the training.
  • The TrainerCallback event methods now accept **kwargs. This may be useful to maintain backwards-compability of callbacks easier in the future. E.g. we may decide to pass the exception/traceback object in case of failure to on_end() and this older callbacks may simply ignore the argument instead of raising a TypeError.
  • Added a TensorBoardCallback which wraps the TensorBoardWriter.

Changed

  • The TrainerCallack.on_epoch() does not fire with epoch=-1 at the start of the training. Instead, TrainerCallback.on_start() should be used for these cases.
  • TensorBoardBatchMemoryUsage is converted from BatchCallback into TrainerCallback.
  • TrackEpochCallback is converted from EpochCallback into TrainerCallback.
  • Trainer can accept callbacks simply with name callbacks instead of trainer_callbacks.
  • TensorboardWriter renamed to TensorBoardWriter, and removed as an argument to the GradientDescentTrainer. In order to enable TensorBoard logging during training, you should utilize the TensorBoardCallback instead.

Removed

  • Removed EpochCallback, BatchCallback in favour of TrainerCallback. The metaclass-wrapping implementation is removed as well.
  • Removed the tensorboard_writer parameter to GradientDescentTrainer. You should use the TensorBoardCallback now instead.

Fixed

  • Now Trainer always fires TrainerCallback.on_end() so all the resources can be cleaned up properly.
  • Fixed the misspelling, changed TensoboardBatchMemoryUsage to TensorBoardBatchMemoryUsage.
  • We set a value to epoch so in case of firing TrainerCallback.on_end() the variable is bound. This could have lead to an error in case of trying to recover a run after it was finished training.

v2.0.0rc1 - 2021-01-21

Added

  • Added TensorCache class for caching tensors on disk
  • Added abstraction and concrete implementation for image loading
  • Added abstraction and concrete implementation for GridEmbedder
  • Added abstraction and demo implementation for an image augmentation module.
  • Added abstraction and concrete implementation for region detectors.
  • A new high-performance default DataLoader: MultiProcessDataLoading.
  • A MultiTaskModel and abstractions to use with it, including Backbone and Head. The MultiTaskModel first runs its inputs through the Backbone, then passes the result (and whatever other relevant inputs it got) to each Head that's in use.
  • A MultiTaskDataLoader, with a corresponding MultiTaskDatasetReader, and a couple of new configuration objects: MultiTaskEpochSampler (for deciding what proportion to sample from each dataset at every epoch) and a MultiTaskScheduler (for ordering the instances within an epoch).
  • Transformer toolkit to plug and play with modular components of transformer architectures.
  • Added a command to count the number of instances we're going to be training with
  • Added a FileLock class to common.file_utils. This is just like the FileLock from the filelock library, except that it adds an optional flag read_only_ok: bool, which when set to True changes the behavior so that a warning will be emitted instead of an exception when lacking write permissions on an existing file lock. This makes it possible to use the FileLock class on a read-only file system.
  • Added a new learning rate scheduler: CombinedLearningRateScheduler. This can be used to combine different LR schedulers, using one after the other.
  • Added an official CUDA 10.1 Docker image.
  • Moving ModelCard and TaskCard abstractions into the main repository.
  • Added a util function allennlp.nn.util.dist_reduce(...) for handling distributed reductions. This is especially useful when implementing a distributed Metric.
  • Added a FileLock class to common.file_utils. This is just like the FileLock from the filelock library, except that it adds an optional flag read_only_ok: bool, which when set to True changes the behavior so that a warning will be emitted instead of an exception when lacking write permissions on an existing file lock. This makes it possible to use the FileLock class on a read-only file system.
  • Added a new learning rate scheduler: CombinedLearningRateScheduler. This can be used to combine different LR schedulers, using one after the other.
  • Moving ModelCard and TaskCard abstractions into the main repository.

Changed

  • DatasetReaders are now always lazy. This means there is no lazy parameter in the base class, and the _read() method should always be a generator.
  • The DataLoader now decides whether to load instances lazily or not. With the PyTorchDataLoader this is controlled with the lazy parameter, but with the MultiProcessDataLoading this is controlled by the max_instances_in_memory setting.
  • ArrayField is now called TensorField, and implemented in terms of torch tensors, not numpy.
  • Improved nn.util.move_to_device function by avoiding an unnecessary recursive check for tensors and adding a non_blocking optional argument, which is the same argument as in torch.Tensor.to().
  • If you are trying to create a heterogeneous batch, you now get a better error message.
  • Readers using the new vision features now explicitly log how they are featurizing images.
  • master_addr and master_port renamed to primary_addr and primary_port, respectively.
  • is_master parameter for training callbacks renamed to is_primary.
  • master branch renamed to main
  • Torch version bumped to 1.7.1 in Docker images.
  • 'master' branch renamed to 'main'
  • Torch version bumped to 1.7.1 in Docker images.

Removed

  • Removed nn.util.has_tensor.

Fixed

  • The build-vocab command no longer crashes when the resulting vocab file is in the current working directory.
  • VQA models now use the vqa_score metric for early stopping. This results in much better scores.
  • Fixed typo with LabelField string representation: removed trailing apostrophe.
  • Vocabulary.from_files and cached_path will issue a warning, instead of failing, when a lock on an existing resource can't be acquired because the file system is read-only.
  • TrackEpochCallback is now a EpochCallback.

v1.3.0 - 2020-12-15

Added

  • Added links to source code in docs.
  • Added get_embedding_layer and get_text_field_embedder to the Predictor class; to specify embedding layers for non-AllenNLP models.
  • Added Gaussian Error Linear Unit (GELU) as an Activation.

Changed

  • Renamed module allennlp.data.tokenizers.token to allennlp.data.tokenizers.token_class to avoid this bug.
  • transformers dependency updated to version 4.0.1.
  • BasicClassifier's forward method now takes a metadata field.

Fixed

  • Fixed a lot of instances where tensors were first created and then sent to a device with .to(device). Instead, these tensors are now created directly on the target device.
  • Fixed issue with GradientDescentTrainer when constructed with validation_data_loader=None and learning_rate_scheduler!=None.
  • Fixed a bug when removing all handlers in root logger.
  • ShardedDatasetReader now inherits parameters from base_reader when required.
  • Fixed an issue in FromParams where parameters in the params object used to a construct a class were not passed to the constructor if the value of the parameter was equal to the default value. This caused bugs in some edge cases where a subclass that takes **kwargs needs to inspect kwargs before passing them to its superclass.
  • Improved the band-aid solution for segmentation faults and the "ImportError: dlopen: cannot load any more object with static TLS" by adding a transformers import.
  • Added safety checks for extracting tar files
  • Turned superfluous warning to info when extending the vocab in the embedding matrix, if no pretrained file was provided

v1.2.2 - 2020-11-17

Added

  • Added Docker builds for other torch-supported versions of CUDA.
  • Adds allennlp-semparse as an official, default plugin.

Fixed

  • GumbelSampler now sorts the beams by their true log prob.

v1.2.1 - 2020-11-10

Added

  • Added an optional seed parameter to ModelTestCase.set_up_model which sets the random seed for random, numpy, and torch.
  • Added support for a global plugins file at ~/.allennlp/plugins.
  • Added more documentation about plugins.
  • Added sampler class and parameter in beam search for non-deterministic search, with several implementations, including MultinomialSampler, TopKSampler, TopPSampler, and GumbelSampler. Utilizing GumbelSampler will give Stochastic Beam Search.

Changed

  • Pass batch metrics to BatchCallback.

Fixed

  • Fixed a bug where forward hooks were not cleaned up with saliency interpreters if there was an exception.
  • Fixed the computation of saliency maps in the Interpret code when using mismatched indexing. Previously, we would compute gradients from the top of the transformer, after aggregation from wordpieces to tokens, which gives results that are not very informative. Now, we compute gradients with respect to the embedding layer, and aggregate wordpieces to tokens separately.
  • Fixed the heuristics for finding embedding layers in the case of RoBERTa. An update in the transformers library broke our old heuristic.
  • Fixed typo with registered name of ROUGE metric. Previously was rogue, fixed to rouge.
  • Fixed default masks that were erroneously created on the CPU even when a GPU is available.
  • Fixed pretrained embeddings for transformers that don't use end tokens.
  • Fixed the transformer tokenizer cache when the tokenizers are initialized with custom kwargs.

v1.2.0 - 2020-10-29

Changed

  • Enforced stricter typing requirements around the use of Optional[T] types.
  • Changed the behavior of Lazy types in from_params methods. Previously, if you defined a Lazy parameter like foo: Lazy[Foo] = None in a custom from_params classmethod, then foo would actually never be None. This behavior is now different. If no params were given for foo, it will be None. You can also now set default values for foo like foo: Lazy[Foo] = Lazy(Foo). Or, if you want you want a default value but also want to allow for None values, you can write it like this: foo: Optional[Lazy[Foo]] = Lazy(Foo).
  • Added support for PyTorch version 1.7.

Fixed

  • Made it possible to instantiate TrainerCallback from config files.
  • Fixed the remaining broken internal links in the API docs.
  • Fixed a bug where Hotflip would crash with a model that had multiple TokenIndexers and the input used rare vocabulary items.
  • Fixed a bug where BeamSearch would fail if max_steps was equal to 1.
  • Fixed BasicTextFieldEmbedder to not raise ConfigurationError if it has embedders that are empty and not in input

v1.2.0rc1 - 2020-10-22

Added

  • Added a warning when batches_per_epoch for the validation data loader is inherited from the train data loader.
  • Added a build-vocab subcommand that can be used to build a vocabulary from a training config file.
  • Added tokenizer_kwargs argument to PretrainedTransformerMismatchedIndexer.
  • Added tokenizer_kwargs and transformer_kwargs arguments to PretrainedTransformerMismatchedEmbedder.
  • Added official support for Python 3.8.
  • Added a script: scripts/release_notes.py, which automatically prepares markdown release notes from the CHANGELOG and commit history.
  • Added a flag --predictions-output-file to the evaluate command, which tells AllenNLP to write the predictions from the given dataset to the file as JSON lines.
  • Added the ability to ignore certain missing keys when loading a model from an archive. This is done by adding a class-level variable called authorized_missing_keys to any PyTorch module that a Model uses. If defined, authorized_missing_keys should be a list of regex string patterns.
  • Added FBetaMultiLabelMeasure, a multi-label Fbeta metric. This is a subclass of the existing FBetaMeasure.
  • Added ability to pass additional key word arguments to cached_transformers.get(), which will be passed on to AutoModel.from_pretrained().
  • Added an overrides argument to Predictor.from_path().
  • Added a cached-path command.
  • Added a function inspect_cache to common.file_utils that prints useful information about the cache. This can also be used from the cached-path command with allennlp cached-path --inspect.
  • Added a function remove_cache_entries to common.file_utils that removes any cache entries matching the given glob patterns. This can used from the cached-path command with allennlp cached-path --remove some-files-*.
  • Added logging for the main process when running in distributed mode.
  • Added a TrainerCallback object to support state sharing between batch and epoch-level training callbacks.
  • Added support for .tar.gz in PretrainedModelInitializer.
  • Made BeamSearch instantiable from_params.
  • Pass serialization_dir to Model and DatasetReader.
  • Added an optional include_in_archive parameter to the top-level of configuration files. When specified, include_in_archive should be a list of paths relative to the serialization directory which will be bundled up with the final archived model from a training run.

Changed

  • Subcommands that don't require plugins will no longer cause plugins to be loaded or have an --include-package flag.
  • Allow overrides to be JSON string or dict.
  • transformers dependency updated to version 3.1.0.
  • When cached_path is called on a local archive with extract_archive=True, the archive is now extracted into a unique subdirectory of the cache root instead of a subdirectory of the archive's directory. The extraction directory is also unique to the modification time of the archive, so if the file changes, subsequent calls to cached_path will know to re-extract the archive.
  • Removed the truncation_strategy parameter to PretrainedTransformerTokenizer. The way we're calling the tokenizer, the truncation strategy takes no effect anyways.
  • Don't use initializers when loading a model, as it is not needed.
  • Distributed training will now automatically search for a local open port if the master_port parameter is not provided.
  • In training, save model weights before evaluation.
  • allennlp.common.util.peak_memory_mb renamed to peak_cpu_memory, and allennlp.common.util.gpu_memory_mb renamed to peak_gpu_memory, and they both now return the results in bytes as integers. Also, the peak_gpu_memory function now utilizes PyTorch functions to find the memory usage instead of shelling out to the nvidia-smi command. This is more efficient and also more accurate because it only takes into account the tensor allocations of the current PyTorch process.
  • Make sure weights are first loaded to the cpu when using PretrainedModelInitializer, preventing wasted GPU memory.
  • Load dataset readers in load_archive.
  • Updated AllenNlpTestCase docstring to remove reference to unittest.TestCase

Removed

  • Removed common.util.is_master function.

Fixed

  • Fix CUDA/CPU device mismatch bug during distributed training for categorical accuracy metric.
  • Fixed a bug where the reported batch_loss metric was incorrect when training with gradient accumulation.
  • Class decorators now displayed in API docs.
  • Fixed up the documentation for the allennlp.nn.beam_search module.
  • Ignore *args when constructing classes with FromParams.
  • Ensured some consistency in the types of the values that metrics return.
  • Fix a PyTorch warning by explicitly providing the as_tuple argument (leaving it as its default value of False) to Tensor.nonzero().
  • Remove temporary directory when extracting model archive in load_archive at end of function rather than via atexit.
  • Fixed a bug where using cached_path() offline could return a cached resource's lock file instead of the cache file.
  • Fixed a bug where cached_path() would fail if passed a cache_dir with the user home shortcut ~/.
  • Fixed a bug in our doc building script where markdown links did not render properly if the "href" part of the link (the part inside the ()) was on a new line.
  • Changed how gradients are zeroed out with an optimization. See this video from NVIDIA at around the 9 minute mark.
  • Fixed a bug where parameters to a FromParams class that are dictionaries wouldn't get logged when an instance is instantiated from_params.
  • Fixed a bug in distributed training where the vocab would be saved from every worker, when it should have been saved by only the local master process.
  • Fixed a bug in the calculation of rouge metrics during distributed training where the total sequence count was not being aggregated across GPUs.
  • Fixed allennlp.nn.util.add_sentence_boundary_token_ids() to use device parameter of input tensor.
  • Be sure to close the TensorBoard writer even when training doesn't finish.
  • Fixed the docstring for PyTorchSeq2VecWrapper.
  • Fixed a bug in the cnn_encoder where activations involving masked tokens could be picked up by the max
  • Fix intra word tokenization for PretrainedTransformerTokenizer when disabling fast tokenizer.

v1.1.0 - 2020-09-08

Fixed

  • Fixed handling of some edge cases when constructing classes with FromParams where the class accepts **kwargs.
  • Fixed division by zero error when there are zero-length spans in the input to a PretrainedTransformerMismatchedIndexer.
  • Improved robustness of cached_path when extracting archives so that the cache won't be corrupted if a failure occurs during extraction.
  • Fixed a bug with the average and evalb_bracketing_score metrics in distributed training.

Added

  • Predictor.capture_model_internals() now accepts a regex specifying which modules to capture.

v1.1.0rc4 - 2020-08-20

Added

  • Added a workflow to GitHub Actions that will automatically close unassigned stale issues and ping the assignees of assigned stale issues.

Fixed

  • Fixed a bug in distributed metrics that caused nan values due to repeated addition of an accumulated variable.

v1.1.0rc3 - 2020-08-12

Fixed

  • Fixed how truncation was handled with PretrainedTransformerTokenizer. Previously, if max_length was set to None, the tokenizer would still do truncation if the transformer model had a default max length in its config. Also, when max_length was set to a non-None value, several warnings would appear for certain transformer models around the use of the truncation parameter.
  • Fixed evaluation of all metrics when using distributed training.
  • Added a py.typed marker. Fixed type annotations in allennlp.training.util.
  • Fixed problem with automatically detecting whether tokenization is necessary. This affected primarily the Roberta SST model.
  • Improved help text for using the --overrides command line flag.

v1.1.0rc2 - 2020-07-31

Changed

  • Upgraded PyTorch requirement to 1.6.
  • Replaced the NVIDIA Apex AMP module with torch's native AMP module. The default trainer (GradientDescentTrainer) now takes a use_amp: bool parameter instead of the old opt_level: str parameter.

Fixed

  • Removed unnecessary warning about deadlocks in DataLoader.
  • Fixed testing models that only return a loss when they are in training mode.
  • Fixed a bug in FromParams that caused silent failure in case of the parameter type being Optional[Union[...]].
  • Fixed a bug where the program crashes if evaluation_data_loader is a AllennlpLazyDataset.

Added

  • Added the option to specify requires_grad: false within an optimizer's parameter groups.
  • Added the file-friendly-logging flag back to the train command. Also added this flag to the predict, evaluate, and find-learning-rate commands.
  • Added an EpochCallback to track current epoch as a model class member.
  • Added the option to enable or disable gradient checkpointing for transformer token embedders via boolean parameter gradient_checkpointing.

Removed

  • Removed the opt_level parameter to Model.load and load_archive. In order to use AMP with a loaded model now, just run the model's forward pass within torch's autocast context.

v1.1.0rc1 - 2020-07-14

Fixed

  • Reduced the amount of log messages produced by allennlp.common.file_utils.
  • Fixed a bug where PretrainedTransformerEmbedder parameters appeared to be trainable in the log output even when train_parameters was set to False.
  • Fixed a bug with the sharded dataset reader where it would only read a fraction of the instances in distributed training.
  • Fixed checking equality of TensorFields.
  • Fixed a bug where NamespaceSwappingField did not work correctly with .empty_field().
  • Put more sensible defaults on the huggingface_adamw optimizer.
  • Simplified logging so that all logging output always goes to one file.
  • Fixed interaction with the python command line debugger.
  • Log the grad norm properly even when we're not clipping it.
  • Fixed a bug where PretrainedModelInitializer fails to initialize a model with a 0-dim tensor
  • Fixed a bug with the layer unfreezing schedule of the SlantedTriangular learning rate scheduler.
  • Fixed a regression with logging in the distributed setting. Only the main worker should write log output to the terminal.
  • Pinned the version of boto3 for package managers (e.g. poetry).
  • Fixed issue #4330 by updating the tokenizers dependency.
  • Fixed a bug in TextClassificationPredictor so that it passes tokenized inputs to the DatasetReader in case it does not have a tokenizer.
  • reg_loss is only now returned for models that have some regularization penalty configured.
  • Fixed a bug that prevented cached_path from downloading assets from GitHub releases.
  • Fixed a bug that erroneously increased last label's false positive count in calculating fbeta metrics.
  • Tqdm output now looks much better when the output is being piped or redirected.
  • Small improvements to how the API documentation is rendered.
  • Only show validation progress bar from main process in distributed training.

Added

  • Adjust beam search to support multi-layer decoder.
  • A method to ModelTestCase for running basic model tests when you aren't using config files.
  • Added some convenience methods for reading files.
  • Added an option to file_utils.cached_path to automatically extract archives.
  • Added the ability to pass an archive file instead of a local directory to Vocab.from_files.
  • Added the ability to pass an archive file instead of a glob to ShardedDatasetReader.
  • Added a new "linear_with_warmup" learning rate scheduler.
  • Added a check in ShardedDatasetReader that ensures the base reader doesn't implement manual distributed sharding itself.
  • Added an option to PretrainedTransformerEmbedder and PretrainedTransformerMismatchedEmbedder to use a scalar mix of all hidden layers from the transformer model instead of just the last layer. To utilize this, just set last_layer_only to False.
  • cached_path() can now read files inside of archives.
  • Training metrics now include batch_loss and batch_reg_loss in addition to aggregate loss across number of batches.

Changed

  • Not specifying a cuda_device now automatically determines whether to use a GPU or not.
  • Discovered plugins are logged so you can see what was loaded.
  • allennlp.data.DataLoader is now an abstract registrable class. The default implementation remains the same, but was renamed to allennlp.data.PyTorchDataLoader.
  • BertPooler can now unwrap and re-wrap extra dimensions if necessary.
  • New transformers dependency. Only version >=3.0 now supported.

v1.0.0 - 2020-06-16

Fixed

  • Lazy dataset readers now work correctly with multi-process data loading.
  • Fixed race conditions that could occur when using a dataset cache.

Added

  • A bug where where all datasets would be loaded for vocab creation even if not needed.
  • A parameter to the DatasetReader class: manual_multi_process_sharding. This is similar to the manual_distributed_sharding parameter, but applies when using a multi-process DataLoader.

v1.0.0rc6 - 2020-06-11

Fixed

  • A bug where TextFields could not be duplicated since some tokenizers cannot be deep-copied. See allenai/allennlp#4270.
  • Our caching mechanism had the potential to introduce race conditions if multiple processes were attempting to cache the same file at once. This was fixed by using a lock file tied to each cached file.
  • get_text_field_mask() now supports padding indices that are not 0.
  • A bug where predictor.get_gradients() would return an empty dictionary if an embedding layer had trainable set to false
  • Fixes PretrainedTransformerMismatchedIndexer in the case where a token consists of zero word pieces.
  • Fixes a bug when using a lazy dataset reader that results in a UserWarning from PyTorch being printed at every iteration during training.
  • Predictor names were inconsistently switching between dashes and underscores. Now they all use underscores.
  • Predictor.from_path now automatically loads plugins (unless you specify load_plugins=False) so that you don't have to manually import a bunch of modules when instantiating predictors from an archive path.
  • allennlp-server automatically found as a plugin once again.

Added

  • A duplicate() method on Instances and Fields, to be used instead of copy.deepcopy()
  • A batch sampler that makes sure each batch contains approximately the same number of tokens (MaxTokensBatchSampler)
  • Functions to turn a sequence of token indices back into tokens
  • The ability to use Huggingface encoder/decoder models as token embedders
  • Improvements to beam search
  • ROUGE metric
  • Polynomial decay learning rate scheduler
  • A BatchCallback for logging CPU and GPU memory usage to tensorboard. This is mainly for debugging because using it can cause a significant slowdown in training.
  • Ability to run pretrained transformers as an embedder without training the weights
  • Add Optuna Integrated badge to README.md

Changed

  • Similar to our caching mechanism, we introduced a lock file to the vocab to avoid race conditions when saving/loading the vocab from/to the same serialization directory in different processes.
  • Changed the Token, Instance, and Batch classes along with all Field classes to "slots" classes. This dramatically reduces the size in memory of instances.
  • SimpleTagger will no longer calculate span-based F1 metric when calculate_span_f1 is False.
  • CPU memory for every worker is now reported in the logs and the metrics. Previously this was only reporting the CPU memory of the master process, and so it was only correct in the non-distributed setting.
  • To be consistent with PyTorch IterableDataset, AllennlpLazyDataset no longer implements __len__(). Previously it would always return 1.
  • Removed old tutorials, in favor of the new AllenNLP Guide
  • Changed the vocabulary loading to consider new lines for Windows/Linux and Mac.

v1.0.0rc5 - 2020-05-26

Fixed

  • Fix bug where PretrainedTransformerTokenizer crashed with some transformers (#4267)
  • Make cached_path work offline.
  • Tons of docstring inconsistencies resolved.
  • Nightly builds no longer run on forks.
  • Distributed training now automatically figures out which worker should see which instances
  • A race condition bug in distributed training caused from saving the vocab to file from the master process while other processing might be reading those files.
  • Unused dependencies in setup.py removed.

Added

  • Additional CI checks to ensure docstrings are consistently formatted.
  • Ability to train on CPU with multiple processes by setting cuda_devices to a list of negative integers in your training config. For example: "distributed": {"cuda_devices": [-1, -1]}. This is mainly to make it easier to test and debug distributed training code..
  • Documentation for when parameters don't need config file entries.

Changed

  • The allennlp test-install command now just ensures the core submodules can be imported successfully, and prints out some other useful information such as the version, PyTorch version, and the number of GPU devices available.
  • All of the tests moved from allennlp/tests to tests at the root level, and allennlp/tests/fixtures moved to test_fixtures at the root level. The PyPI source and wheel distributions will no longer include tests and fixtures.

v1.0.0rc4 - 2020-05-14

We first introduced this CHANGELOG after release v1.0.0rc4, so please refer to the GitHub release notes for this and earlier releases.