v1.2.1

What's new

Added 🎉

• Added the TaskCard class and task cards for common tasks.
• Added a test for the interpret functionality

Changed ⚠️

• Added more information to model cards for pair classification models (pair-classification-decomposable-attention-elmo, pair-classification-roberta-snli, pair-classification-roberta-mnli, pair-classification-esim).

Fixed ✅

• 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

Commits

a6b5a2c Adds a test that checks the interpret functionality (#163)
6a81154 make naqanet test less flaky (#162)
1e4b67c Transformer ELMo config fixes (#131)
d383ea9 Task cards (#161)
b152d82 Updating pair classification model cards (#160)

v1.2.0

What's new

Changed ⚠️

• Updated docstring for Transformer MC.
• Added more information to model cards for multiple choice models (mc-roberta-commonsenseqa,
  mc-roberta-piqa, and mc-roberta-swag).

Fixed ✅

• Fixed many training configs to work out-of-the box. These include the configs for bart_cnn_dm, swag, bidaf, bidaf_elmo,
  naqanet, and qanet.
• Fixed minor bug in MaskedLanguageModel, where getting token ids used hard-coded assumptions (that
  could be wrong) instead of our standard utility function.

Commits

fcb6758 Prepare for release v1.2.0
43c9b51 remove autocast(False) around RNNs (#158)
bf42ee0 make coref test less flaky
84fb13f Adding info to Multiple Choice model cards (#156)
ab4ab67 make more configs work out-of-the-box (#153)
f253c57 Bump conllu from 4.2 to 4.2.1 (#155)

v1.2.0rc1

What's new

Added 🎉

• Added dataset reader support for SQuAD 2.0 with both the SquadReader and TransformerSquadReader.
• Updated the SQuAD v1.1 metric to work with SQuAD 2.0 as well.
• Updated the TransformerQA model to work for SQuAD 2.0.
• Added official support for Python 3.8.
• Added a json template for model cards.
• Added training_config as a field in model cards.
• Added a BeamSearchGenerator registrable class which can be provided to a NextTokenLM model
  to utilize beam search for predicting a sequence of tokens, instead of a single next token.
  BeamSearchGenerator is 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 NextTokenLM that uses a PretrainedTransformerEmbedder.
• Added an overrides parameter to pretrained.load_predictor().

Changed ⚠️

• rc-transformer-qa pretrained model is now an updated version trained on SQuAD v2.0.
• skip_invalid_examples parameter in SQuAD dataset readers has been deprecated. Please use
  skip_impossible_questions instead.

Fixed ✅

• Fixed lm-masked-language-model pretrained model.
• Fixed BART for latest transformers version.
• Fixed BiDAF predictor and BiDAF predictor tests.
• Fixed a bug with Seq2SeqDatasetReader that 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 start_symbol or end_symbol are not part of the tokenizer's vocabulary.

Commits

8b5c42c Don't use PretrainedModelInitializer when loading a model (#141)
f789c53 add beam search to NextTokenLm (#149)
414cb48 Bidaf predictor and test quick fix (#154)
2d03ab0 fix docstrings to render correctly in docs (#150)
cdd505e Update mkdocs-material requirement from <6.1.0,>=5.5.0 to >=5.5.0,<6.2.0 (#152)
4cee3bd Model cards update (#148)
223ef71 Bump mypy from 0.782 to 0.790 (#147)
18008d0 fix TransformerQA predictor test (#145)
e827130 update masked lm language model (#143)
83e668c update how env variables set (#140)
d90953a run pretrained tests on CPU-only runner (#142)
0b3853a SQuAD 2.0 support (#134)
bcaf3c8 Fix some typos
0b05944 tick version for nightly releases
a330876 fix seq2seq reader bug (#139)
dd36890 Adding more documentation to model cards spec (#137)
436e047 Update mkdocs-material requirement from <5.6.0,>=5.5.0 to >=5.5.0,<6.1.0 (#138)
0cb9e0f Model cards update (#136)
5fabe32 official support Python 3.8 🐍 (#135)
16513bd Bump conllu from 4.1 to 4.2 (#133)
a1f9a05 Create Dependabot config file (#132)
3425192 remove TODO
2d7e1f6 Fixes for new transformers release (#130)
57908bb allow multiple token embedders, but only use first with non-empty type (#129)

v1.1.0

What's new since version 1.0.0

Added

• 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.
• Added BART model
• Added ModelCard and related classes. Added model cards for all the pretrained models.
• Added a field registered_predictor_name to ModelCard.
• Added a method load_predictor to allennlp_models.pretrained.
• Added support to multi-layer decoder in simple seq2seq model.
• Added two models for fine-grained NER
• Added a category for multiple choice models, including a few reference implementations

Changed

• CopyNetDatasetReader no longer automatically adds START_TOKEN and END_TOKEN to the tokenized source. If you want these in the tokenized source, it's up to the source tokenizer.
• Implemented manual distributed sharding for the SNLI dataset reader.

Fixed

• 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.
• Fixed GraphParser.get_metrics so that it expects a dict from F1Measure.get_metric.
• CopyNet and SimpleSeq2Seq models now work with AMP.
• Made the SST reader a little more strict in the kinds of input it accepts.
• Updated the RoBERTa SST config to make proper use of the CLS token
• Updated RoBERTa SNLI and MNLI pretrained models for latest transformers version
• Updated the BERT SRL model to be compatible with the new huggingface tokenizers.
• CopyNetSeq2Seq model now works with pretrained transformers.
• A bug with NextTokenLM that caused simple gradient interpreters to fail.
• A bug in training_config of qanet and bimpm that used the old version of regularizer and initializer.
• 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/.

Commits

5ffc207 Prepare for release v1.1.0
36ad6b3 Bump conllu from 4.0 to 4.1 (#126)
2c2e4e4 Fixes the BERT SRL model (#124)
44a3ca5 Update BartEncoder docstring for embeddings guidance (#127)
2e54449 Combinable quora sequences (#119)
ce8ef9b formatting changes for new version of black (#125)
3742b4a Distributed metrics (#123)
31b00e7 Bump markdown-include from 0.5.1 to 0.6.0 (#121)
c4ef4c8 Prepare for release v1.1.0rc4
27c0ca5 always run configs CI job (#118)
7c3be82 upgrade to actions cache v2 (#116)
dbd46b4 Add test for sentiment analysis (#117)
a91f009 run config tests in subprocesses (#115)
c211baf Update simple_seq2seq.py (#90)
267b747 Bump conllu from 3.1.1 to 4.0 (#114)
87570ec validate pretrained configs in CI (#112)
4fa5fc1 Fix RoBERTa SST (#110)
0491690 Only pin mkdocs-material to minor version, ignore specific patch version (#113)
8d27e7b prepare for release v1.1.0rc3
959a5eb Update graph parser metrics (#109)
45f85ce Bump mkdocs-material from 5.5.3 to 5.5.5 (#111)
e69f4c4 fix docs CI
4b96dfa tick version for nightly releases
e5f5c62 Update some models for AMP training (#104)
4f0bca1 Bump mkdocs-material from 5.5.2 to 5.5.3 (#108)
cbd2b57 Adds the pretrained BART model (#107)
5d9098f Bump conllu from 3.0 to 3.1.1 (#105)
92f2a8f Bump mkdocs-material from 5.5.0 to 5.5.2 (#106)
a901f9f Prepare for release v1.1.0rc2
04561a8 updates for torch 1.6 (#103)
e7b8247 Update RoBERTa SNLI/MNLI models (#102)
008828b Adding ModelCard (#98)
eaa331e Roberta SST (#99)
75c8869 Bump mkdocs-material from 5.4.0 to 5.5.0 (#100)
a56a103 Implemented BART (#35)
a730fed Cuda devices (#97)
4d0e090 Make sure we ship the SRL eval Perl script (#96)
4f2e316 tick version for nightly releases
dd60f94 Prepare for release v1.1.0rc1
da83a4e build and publish models docs (#91)
4b2178b implement manual distributed sharding for SNLI reader (#89)
1a2a8f4 Updates the SRL model (#93)
8a93743 Updated the fine-grained NER transformer model (#92)
b913333 Multiple Choice (#75)
09395d2 updates for new transformers release (#88)
11c6814 fix the bug of bimpm and update CHANGELOG (#87)
a735ddd Fix the regularizer of QANet model (#86)
0ce14da fixes for next_token_lm (#85)
37136f8 skip docker build on nightly release
82aa9ac Fine grained NER (#84)
4b5b939 fix test fixture
947beb0 remove unused param in copynet reader
9ec65df fix nightly Docker workflow
3019a4e fix workflow skip conditions
cc60ab9 use small dummy transformer for copynet test (#83)
ac9f214 fix nightly workflow
596e6a7 Bump mypy from 0.781 to 0.782 (#82)
d210c2f add nightly releases (#81)
935a2a8 dont add START and END tokens to source in CopyNet (#79)
d6798ce update skip conditions on const-parser-test (#80)
2754f88 Bump mypy from 0.780 to 0.781 (#78)
d3588ad Make CopyNet work with pretrained transformer models (#76)

v1.1.0rc4

What's new since v1.1.0rc3

Added

• Added regression tests for training configs that run on a scheduled workflow.
• Added a test for the pretrained sentiment analysis model.

Commits

27c0ca5 always run configs CI job (#118)
7c3be82 upgrade to actions cache v2 (#116)
dbd46b4 Add test for sentiment analysis (#117)
a91f009 run config tests in subprocesses (#115)
c211baf Update simple_seq2seq.py (#90)
267b747 Bump conllu from 3.1.1 to 4.0 (#114)
87570ec validate pretrained configs in CI (#112)
4fa5fc1 Fix RoBERTa SST (#110)
0491690 Only pin mkdocs-material to minor version, ignore specific patch version (#113)

v1.1.0rc3

What's new since v1.1.0rc2

Fixed

• Fixed GraphParser.get_metrics so that it expects a dict from F1Measure.get_metric.
• CopyNet and SimpleSeq2Seq models now work with AMP.

Commits

8d27e7b prepare for release v1.1.0rc3
959a5eb Update graph parser metrics (#109)
45f85ce Bump mkdocs-material from 5.5.3 to 5.5.5 (#111)
e69f4c4 fix docs CI
4b96dfa tick version for nightly releases
e5f5c62 Update some models for AMP training (#104)
4f0bca1 Bump mkdocs-material from 5.5.2 to 5.5.3 (#108)
cbd2b57 Adds the pretrained BART model (#107)
5d9098f Bump conllu from 3.0 to 3.1.1 (#105)
92f2a8f Bump mkdocs-material from 5.5.0 to 5.5.2 (#106)

v1.1.0rc2

What's new since v1.1.0rc1

Changed

• Updated to PyTorch 1.6.

Fixed

• Updated the RoBERTa SST config to make proper use of the CLS token
• Updated RoBERTa SNLI and MNLI pretrained models for latest transformers version

Added

• Added BART model
• Added ModelCard and related classes. Added model cards for all the pretrained models.
• Added a method load_predictor to allennlp_models.pretrained.

Commits

a901f9f Prepare for release v1.1.0rc2
04561a8 updates for torch 1.6 (#103)
e7b8247 Update RoBERTa SNLI/MNLI models (#102)
008828b Adding ModelCard (#98)
eaa331e Roberta SST (#99)
75c8869 Bump mkdocs-material from 5.4.0 to 5.5.0 (#100)
a56a103 Implemented BART (#35)
a730fed Cuda devices (#97)
4d0e090 Make sure we ship the SRL eval Perl script (#96)
4f2e316 tick version for nightly releases

v1.1.0rc1

First v1.1 pre-release candidate. See the corresponding allennlp v1.1.0rc1 release candidate.

What's new since v1.0.0

Fixed

• Updated the BERT SRL model to be compatible with the new huggingface tokenizers.
• CopyNetSeq2Seq model now works with pretrained transformers.
• A bug with NextTokenLM that caused simple gradient interpreters to fail.
• A bug in training_config of qanet and bimpm that used the old version of regularizer and initializer.
• 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/.

Changed

• CopyNetDatasetReader no longer automatically adds START_TOKEN and END_TOKEN
  to the tokenized source. If you want these in the tokenized source, it's up to
  the source tokenizer.

Added

• 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.

Commits

dd60f94 Prepare for release v1.1.0rc1
da83a4e build and publish models docs (#91)
4b2178b implement manual distributed sharding for SNLI reader (#89)
1a2a8f4 Updates the SRL model (#93)
8a93743 Updated the fine-grained NER transformer model (#92)
b913333 Multiple Choice (#75)
09395d2 updates for new transformers release (#88)
11c6814 fix the bug of bimpm and update CHANGELOG (#87)
a735ddd Fix the regularizer of QANet model (#86)
0ce14da fixes for next_token_lm (#85)
37136f8 skip docker build on nightly release
82aa9ac Fine grained NER (#84)
4b5b939 fix test fixture
947beb0 remove unused param in copynet reader
9ec65df fix nightly Docker workflow
3019a4e fix workflow skip conditions
cc60ab9 use small dummy transformer for copynet test (#83)
ac9f214 fix nightly workflow
596e6a7 Bump mypy from 0.781 to 0.782 (#82)
d210c2f add nightly releases (#81)
935a2a8 dont add START and END tokens to source in CopyNet (#79)
d6798ce update skip conditions on const-parser-test (#80)
2754f88 Bump mypy from 0.780 to 0.781 (#78)
d3588ad Make CopyNet work with pretrained transformer models (#76)

v1.0.0

Release 1.0 corresponding to the AllenNLP 1.0 Release.

v1.0.0rc6

Changed

• Removed deprecated "simple_seq2seq" predictor

Fixed

• Replaced deepcopy of Instances with new Instance.duplicate() method.
• 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

• 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

Commits

1b2682b prepare for release v1.0.0rc6
c0821a0 Adds SST and NLI models to the list of pretrained models (#73)
4b010c0 Remove some files I merged by accident (#74)
2b9ca77 Sets a default predictor for several models (#71)
961da3e Fix casing in NLI model (#72)
efe66bc update squad reader
93273f0 Fix capitalization typo in SNLI RoBERTa config (#70)
646d254 fixes on some structured prediction dataset readers (#65)
5086811 Bump mypy from 0.770 to 0.780 (#69)
d20cc8f Fixes for mixed precision training (#68)
28a650c Fix training config (#66)
053f0c0 fix SentenceTaggerPredictor import (#67)
204be1a Update transformer_qa.jsonnet (#59)
0ed63fa transformer qa quick fix
ae1c0ec Bump conllu from 2.3.2 to 3.0 (#63)
4f17497 Use new Instance.duplicate() method instead of deepcopy (#64)
6fac5b7 Bump version number