All notable changes to this project will be documented in this file.
- Changed the token-based verbose metric in the
False) to be
- Added option
CrfTaggerin order to support three sample weighting techniques.
- Added Python 3.9 to the testing matrix
- Following a breaking change in the NLTK API, we now depend on the most recent version only.
- Added Tensorboard callbacks to the RC models
- The error message you get when perl isn't installed is now more readable.
- Removed the dependency on the
- Removed Tango components, since they now live at https://github.com/allenai/tango
- Seperate start/end token check in
Seq2SeqDatasetReaderfor source and target tokenizers.
superglue_recordto the rc readers for SuperGLUE's Reading Comprehension with Commonsense Reasoning task
- Added some additional
__init__()parameters to the
allennlp_models.generationfor customizing. beam search and other options.
- Added a configuration file for fine-tuning
t5-11bon CCN-DM (requires at least 8 GPUs).
- Added a configuration to train on the PIQA dataset with AllenNLP Tango.
- Added a transformer classification model.
- Added a configuration to train on the IMDB dataset with AllenNLP Tango.
CopyNetSeq2Seqto use scheduled sampling during training.
- Fixed tests for Spacy versions greater than 3.1.
- Fixed the last step decoding when training CopyNet.
- Allow singleton clusters in
VisionReaderto yield all of
RegionDetectorOutput's keys in processing.
- Added support for NLVR2 visual entailment, including a data loader, two models, and training configs.
StanfordSentimentTreeBankDatasetReader.apply_token_indexers()to add token_indexers rather than in
- Added support for Flickr30k image retrieval, including a dataset reader, a model, and a training config.
CopyNetSeq2Relto smooth generation targets.
vocabas argument to
binary-gender-bias-mitigated-roberta-snlimodel card to indicate that model requires
- Fixed registered model name in the
- The multiple choice models now use the new
TransformerTextFieldand the transformer toolkit generally.
- Updated all instances of
num_serialized_models_to_keepparameter is now called
- Improvements to the vision models and other models that use
allennlp.modules.transformerunder the hood.
- Added tests for checklist suites for SQuAD-style reading comprehension models (
bidaf), and textual entailment models (
- Added an optional "weight" parameter to
CopyNetSeq2Seq.forward()for calculating a weighted loss instead of the simple average over the the negative log likelihoods for each instance in the batch.
- Added a way to initialize the
SrlBertmodel without caching/loading pretrained transformer weights. You need to set the
bert_modelparameter to the dictionary form of the corresponding
BertConfigfrom HuggingFace. See PR #257 for more details.
- Added a
beam_searchparameter to the
generationmodels so that a
BeamSearchobject can be specified in their configs.
- Added a binary gender bias-mitigated RoBERTa model for SNLI.
T5model for generation.
- Added a classmethod constructor on
- Added a parameter called
CNNDailyMailDatasetReader. This is useful with T5, for example, by setting
source_prefixto "summarization: ".
- Tests for
- Distributed tests for
- Added dataset reader for visual genome QA.
pretrained.load_predictor()now allows for loading model onto GPU.
VqaMeasurenow calculates correctly in the distributed case.
ConllCorefScoresnow calculates correctly in the distributed case.
SrlEvalScorerraises an appropriate error if run in the distributed setting.
nullin model cards for the models where it was the same as the default predictor.
- Fixed bug in
min_countto have key
answers. Resolves failure of model checks that involve calling
TransformerQAnow outputs span probabilities as well as scores.
predictions_to_labeled_instances, which is required for the interpret module.
- Added script that produces the coref training data.
- Added tests for using
allennlp predicton multitask models.
- Added reader and training config for RoBERTa on SuperGLUE's Recognizing Textual Entailment task
- Evaluating RC task card and associated LERC model card
- Compatibility with PyTorch 1.8
- Allows the order of examples in the task cards to be specified explicitly
- Dataset reader for SuperGLUE BoolQ
- Add option
SnliDatasetReaderto support only having "non-entailment" and "entailment" as output labels.
- Made all the models run on AllenNLP 2.1
- Add option
CrfTaggerto set the flag outside its forward function.
make_output_human_readablefor pair classification models (
- Fixed https://github.com/allenai/allennlp/issues/4745.
NumericallyAugmentedQaNetmodels to remove bias for layers that are followed by normalization layers.
- Updated the model cards for
- Predictors now work for the vilbert-multitask model.
- Support unlabeled instances in
coding_schemeparameter is now deprecated in
Conll2000DatasetReader, please use
- BART model now adds a
make_output_human_readablethat has the cleaned text corresponding to
TransformerMCReader.text_to_instanceoptional with default of
- Updated many of the models for version 2.1.0. Fixed and re-trained many of the models.
OpenIePredictor.predict_jsonso it treats auxiliary verbs as verbs when the language is English.
- Made the training configs compatible with the tensorboard logging changes in the main repo
- Dataset readers, models, metrics, and training configs for VQAv2, GQA, and Visual Entailment
training_configs/rc/dialog_qa.jsonnetto work with new data loading API.
- Fixed the potential for a dead-lock when training the
TransformerQAmodel on multiple GPUs when nodes receive different sized batches.
- Fixed BART. This implementation had some major bugs in it that caused poor performance during prediction.
TaskCardabstractions out of the models repository.
masterbranch renamed to
SquadEmAndF1metric can now also accept a batch of predictions and corresponding answers (instead of a single one) in the form of list (for each).
- Fix an index bug in BART prediction.
SemanticRoleLabelerPredictor.tokens_to_instancesso it treats auxiliary verbs as verbs when the language is English
- Added link to source code to API docs.
- Information updates for remaining model cards (also includes the ones in demo, but not in the repository).
Dockerfile.committo work with different CUDA versions.
- Changes required for the
transformersdependency update to version 4.0.1.
- Added missing folder for
- Changed AllenNLP dependency for releases to allow for a range of versions, instead of being pinned to an exact version.
- There will now be multiple Docker images pushed to Docker Hub for releases, each corresponding to a different supported CUDA version (currently just 10.2 and 11.0).
ValueErrorerror message in
- Better check for start and end symbols in
Seq2SeqDatasetReaderthat doesn't fail for BPE-based tokenizers.
- Information updates for all model cards.
- Added the
TaskCardclass and task cards for common tasks.
- Added a test for the interpret functionality
- Added more information to model cards for pair classification models (
- Fixed TransformerElmo config to work with the new AllenNLP
- Pinned the version of torch more tightly to make AMP work
- Fixed the somewhat fragile Bidaf test
- Updated docstring for Transformer MC.
- Added more information to model cards for multiple choice models (
- Fixed many training configs to work out-of-the box. These include the configs for
- Fixed minor bug in MaskedLanguageModel, where getting token ids used hard-coded assumptions (that could be wrong) instead of our standard utility function.
- Added dataset reader support for SQuAD 2.0 with both the
- Updated the SQuAD v1.1 metric to work with SQuAD 2.0 as well.
- Updated the
TransformerQAmodel to work for SQuAD 2.0.
- Added official support for Python 3.8.
- Added a json template for model cards.
training_configas a field in model cards.
- Added a
BeamSearchGeneratorregistrable class which can be provided to a
NextTokenLMmodel to utilize beam search for predicting a sequence of tokens, instead of a single next token.
BeamSearchGeneratoris an abstract class, so a concrete registered implementation needs to be used. One implementation is provided so far:
TransformerBeamSearchGenerator, registered as
transformer, which will work with any
NextTokenLMthat uses a
- Added an
rc-transformer-qapretrained model is now an updated version trained on SQuAD v2.0.
skip_invalid_examplesparameter in SQuAD dataset readers has been deprecated. Please use
- Fixed BART for latest
- Fixed BiDAF predictor and BiDAF predictor tests.
- Fixed a bug with
Seq2SeqDatasetReaderthat would cause an exception when the desired behavior is to not add start or end symbols to either the source or the target and the default
end_symbolare not part of the tokenizer's vocabulary.
LanguageModelTokenEmbedderto allow allow multiple token embedders, but only use first with non-empty type
- Fixed evaluation of metrics when using distributed setting.
- Fixed a bug introduced in 1.0 where the SRL model did not reproduce the original result.
- Added regression tests for training configs that run on a scheduled workflow.
- Added a test for the pretrained sentiment analysis model.
- Added way for questions from quora dataset to be concatenated like the sequences in the SNLI dataset.
GraphParser.get_metricsso that it expects a dict from
SimpleSeq2Seqmodels now work with AMP.
- Made the SST reader a little more strict in the kinds of input it accepts.
- Updated to PyTorch 1.6.
- Updated the RoBERTa SST config to make proper use of the CLS token
- Updated RoBERTa SNLI and MNLI pretrained models for latest
- Added BART model
ModelCardand related classes. Added model cards for all the pretrained models.
- Added a field
- Added a method
- Added support to multi-layer decoder in simple seq2seq model.
- Updated the BERT SRL model to be compatible with the new huggingface tokenizers.
CopyNetSeq2Seqmodel now works with pretrained transformers.
- A bug with
NextTokenLMthat caused simple gradient interpreters to fail.
- A bug in
bimpmthat used the old version of
- The fine-grained NER transformer model did not survive an upgrade of the transformers library, but it is now fixed.
- Fixed many minor formatting issues in docstrings. Docs are now published at https://docs.allennlp.org/models/.
CopyNetDatasetReaderno longer automatically adds
END_TOKENto the tokenized source. If you want these in the tokenized source, it's up to the source tokenizer.
- Added two models for fine-grained NER
- Added a category for multiple choice models, including a few reference implementations
- Implemented manual distributed sharding for SNLI dataset reader.
No additional note-worthy changes since rc6.
- Removed deprecated
Instances with new
- A bug where pretrained sentence taggers would fail to be initialized because some of the models were not imported.
- A bug in some RC models that would cause mixed precision training to crash when using NVIDIA apex.
- Predictor names were inconsistently switching between dashes and underscores. Now they all use underscores.
- Added option to SemanticDependenciesDatasetReader to not skip instances that have no arcs, for validation data
- Added a default predictors to several models
- Added sentiment analysis models to pretrained.py
- Added NLI models to pretrained.py
- Moved the models into categories based on their format
transformer_qapredictor accept JSON input with the keys "question" and "passage" to be consistent with the
conlludependency (previously part of
We first introduced this
CHANGELOG after release
v1.0.0rc4, so please refer to the GitHub release
notes for this and earlier releases.