v2.10.1

What's new

Fixed ✅

• Fixed redundant TextField wrapping in TransformerSuperGlueRteReader.

Commits

e5d1f59 Update torch requirement from <1.12.0,>=1.7.0 to >=1.7.0,<1.13.0 (#342)
3e3b3ec Remove duplicate TextField (#344)

v2.10.0

What's new

Added 🎉

• Changed the token-based verbose metric in the CrfTagger model (when verbose_metrics is True and calculate_span_f1 is False) to be FBetaVerboseMeasure instead of FBetaMeasure.
• Added option weight_strategy to CrfTagger in order to support three sample weighting techniques.

Commits

97df196 Implementation of Weighted CRF Tagger (handling unbalanced datasets) (#341)
704a244 Disable dependabot, add notice of future sunsetting (#343)
92372a3 Update results for coref (#339)
deddfbe Bump mypy from 0.960 to 0.961 (#340)
900c447 Bump mypy from 0.950 to 0.960 (#337)
af43511 Update mkdocs-material requirement from <8.3.0,>=5.5.0 to >=5.5.0,<8.4.0 (#338)
487191f Bump conllu from 4.4.1 to 4.4.2 (#336)
f233052 Bump mypy from 0.942 to 0.950 (#335)

v2.9.3

What's new

Added 🎉

• Added jpeg extension to __init__ of VisionReader.

Commits

87ae1da Giving up on Docspec2 (#331)
ccf692b Bump mypy from 0.931 to 0.942 (#332)
3c63cda Bump mkdocs from 1.2.3 to 1.3.0 (#333)
54eec30 Add option to specify the image extensions to be loaded/discovered (#324)
7bcc749 fix platform for GPU tests
eca99b7 Update torch requirement from <1.11.0,>=1.7.0 to >=1.7.0,<1.12.0 (#325)
198efc0 Update mkdocs-material requirement from <8.2.0,>=5.5.0 to >=5.5.0,<8.3.0 (#321)
2514f0d Update CI to install allennlp with optional deps for tests (#323)

v2.9.0

What's new

Added 🎉

• Added Python 3.9 to the testing matrix

Changed ⚠️

• Following a breaking change in the NLTK API, we now depend on the most recent version only.
• Added Tensorboard callbacks to the RC models

Removed 👋

• Removed the dependency on the overrides package
• Removed Tango components, since they now live at https://github.com/allenai/tango

Commits

8c5f4d0 Gives more accurate LERC instructions. (#320)
9fb6c09 Bump mypy from 0.930 to 0.931 (#319)
a36aed5 Removes stuff that now lives in the tango repo (#313)
eb3994e Bump mypy from 0.910 to 0.930 (#318)
9f32599 Update mkdocs-material requirement from <8.1.0,>=5.5.0 to >=5.5.0,<8.2.0 (#314)
b6923c3 Model updates (#315)
4866862 Add option to use scheduled sampling in CopyNet (#309)
8644437 Update mkdocs-material requirement from <7.4.0,>=5.5.0 to >=5.5.0,<8.1.0 (#310)
1aff4cc Python 3.9 (#312)
1591777 Remove dependency on the overrides package (#311)
05dd4e6 Update nltk requirement from <3.6.5 to <3.6.6 (#307)
6c4daca Update mkdocs-material requirement from <7.3.0,>=5.5.0 to >=5.5.0,<7.4.0 (#303)

v2.8.0

What's new

Changed ⚠️

• Seperate start/end token check in Seq2SeqDatasetReader for source and target tokenizers.

Commits

ca7e79a Prepare for release v2.8.0
5ff0f79 Seperate start/end token check for source and target tokenizer (#308)
84ba7cf Fix the IndexError when CNNDailyMailDatasetReader reads test data. (#306)
d1b81c3 Update torch requirement from <1.10.0,>=1.7.0 to >=1.7.0,<1.11.0 (#305)
39557dd pin NLTK to avoid breaking change to meteor score (#304)
c814aa1 Finished record reader (#300)
c57b931 fix tests for new overrides functionality (#302)
25e2b1a require Python>=3.7 (#301)

v2.7.0

What's new

Added 🎉

• Added some additional __init__() parameters to the T5 model in allennlp_models.generation for customizing.
  beam search and other options.
• Added a configuration file for fine-tuning t5-11b on 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.

Fixed ✅

• Fixed tests for Spacy versions greater than 3.1.
• Fixed the last step decoding when training CopyNet.
• Allow singleton clusters in ConllCorefScores.

Changed ⚠️

• Updated VisionReader to yield all of RegionDetectorOutput's keys in processing.

Commits

ae7942e Prepare for release v2.7.0
a8a3486 update conll (#298)
54de9d6 IMDB Model (#297)
8d5b21f Fix a bug in the last decoding step of training CopyNet (#296)
1156e49 Bump conllu from 4.4 to 4.4.1 (#295)
31649f5 PIQA in Tango (#294)
4eb7c27 Updating vision reader to also produce class probs and labels (#293)
7b7b9c1 Spacy new version (#290)
f991ae0 Update mkdocs-material requirement from <7.2.0,>=5.5.0 to >=5.5.0,<7.3.0 (#289)
db0e21a FairScale integration and T5-11B fine-tuning (#271)

v2.6.0

What's new

Added 🎉

• Added support for NLVR2 visual entailment, including a data loader, two models, and training configs.
• Added StanfordSentimentTreeBankDatasetReader.apply_token_indexers() to add token_indexers rather than in text_to_instance
• Added AdversarialBiasMitigator tests.
• Added adversarial-binary-gender-bias-mitigated-roberta-snli model.
• Added support for Flickr30k image retrieval, including a dataset reader, a model, and a training config.
• Added label_smoothing parameter to CopyNetSeq2Rel to smooth generation targets.
• Added vocab as argument to beam_search.construct in all generation models.

Fixed ✅

• Fixed binary-gender-bias-mitigated-roberta-snli model card to indicate that model requires [email protected].
• Fixed registered model name in the pair-classification-roberta-rte and vgqa-vilbert model cards.

Changed ⚠️

• The multiple choice models now use the new TransformerTextField and the transformer toolkit generally.

Commits

146ea76 Pass vocab to beam search (#288)
698b7ef TransformerTextField in the MC models (#286)
07fa124 Add label smoothing to CopyNet (#287)
e47da99 Flickr30k (#285)
fb35b2d Bump mypy from 0.812 to 0.910 (#284)
90f6259 Update Python environment setup in GitHub Actions (#283)
bdf82a1 added AdversarialBiasMitigator tests and model (#281)
8d2d84f Test fixes (#282)
ef004d3 Update torch requirement from <1.9.0,>=1.7.0 to >=1.7.0,<1.10.0 (#280)
8cb4b08 small CI fix
d721487 add StanfordSentimentTreeBankDatasetReader.apply_token_indexers() (#273)
996adff Update pair-classification-binary-gender-bias-mitigated-roberta-snli.json with correct versions of allennlp-models and allennlp (#272)
b17d114 some model card fixes (#274)
0a7901c Revert "CHANGELOG"
38064e1 CHANGELOG
664f38f Nlvr2 (#265)
e395e63 tick version for nightly releases

v2.5.0

allennlp-models release corresponding to allennlp v2.5.0.

What's new

Changed ⚠️

• Updated all instances of sanity_checks to confidence_checks.
• The num_serialized_models_to_keep parameter is now called keep_most_recent_by_count.
• Improvements to the vision models and other models that use allennlp.modules.transformer under the hood.

Added 🎉

• Added tests for checklist suites for SQuAD-style reading comprehension models (bidaf), and textual entailment models (decomposable_attention and esim).
• 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 SrlBert model without caching/loading pretrained transformer weights.
  You need to set the bert_model parameter to the dictionary form of the corresponding BertConfig from HuggingFace.
  See PR #257 for more details.
• Added a beam_search parameter to the generation models so that a BeamSearch object can be specified in their configs.
• Added a binary gender bias-mitigated RoBERTa model for SNLI.

Commits

a98e13a Specify BeamSearch as a parameter (#267)
5dcf2b9 Added binary gender bias-mitigated RoBERTa model for SNLI (#268)
79d25e5 tick version for nightly release
50a0452 Checkpointing (#269)
07f1b56 Update nr-interface requirement from <0.0.4 to <0.0.6 (#266)
8bf4e1c cancel redundant GH Actions builds (#270)
2f1b779 Update roberta-sst.json (#264)
dea182c Avoid duplicate tokenization of context in training (#263)
dc633f1 Updates for transformer toolkit changes (#261)
53c61dd Renaming sanity_checks to confidence_checks (#262)
3ec87c7 set codecov to 'informational' mode
45068bb Vgqa dataset reader (#260)
77315fc Add weighting option to CopyNet (#258)
845fe4c Add way to initialize SrlBert without pretrained BERT weights (#257)
ab1e86a Checklist tests (#255)
659c71f Update pretrained.py: Quick fix to be able to load pertained_models directly to GPU. (#254)

v2.4.0

What's new

Added 🎉

• Added T5 model for generation.
• Added a classmethod constructor on Seq2SeqPredictor: .pretrained_t5_for_generation().
• Added a parameter called source_prefix to CNNDailyMailDatasetReader. This is useful with T5, for example, by setting source_prefix to "summarization: ".
• Tests for VqaMeasure.
• Distributed tests for ConllCorefScores and SrlEvalScorer metrics.

Fixed ✅

• VqaMeasure now calculates correctly in the distributed case.
• ConllCorefScores now calculates correctly in the distributed case.
• SrlEvalScorer raises an appropriate error if run in the distributed setting.

Changed ⚠️

• Updated registered_predictor_name to null in model cards for the models where it was the same as the default predictor.

Commits

7a6ee0c Add T5 for generation/summarization (#241)
5012f23 Fix name of variable in docstring (#252)
d18c837 Remove registered_predictor_name from all modelcards where the predictor is the same as the default (#253)
e5789cf Distributed metrics (#251)
c733f83 Roberta data reader (#247)
419bc90 Distributed VQA metric (#250)

v2.3.0

What's new

Fixed ✅

• Fixed bug in experiment_from_huggingface.jsonnet and experiment.jsonnet by changing min_count to have key labels instead of answers. Resolves failure of model checks that involve calling _extend in vocabulary.py
• TransformerQA now outputs span probabilities as well as scores.
• TransformerQAPredictor now implements predictions_to_labeled_instances, which is required for the interpret module.

Added 🎉

• Added script that produces the coref training data.
• Added tests for using allennlp predict on multitask models.

Commits

eaf76a7 Fix paths in training config
acc3424 Adds tests for multitask predictor (#248)
f4fb932 Making TransformerQAPredictor compatible with interpret modules (#249)
31a9ad3 Fix paths in training config
f1de60f update version in README
b358dad Add list of pretrained models to README (#246)
39f1a0f fix spacing for API docs
048355d big improvements to readme (#243)
2208341 Add best span probs (#244)
15ac4dc fix sanity check tests (#242)
4b13924 Add make_output_human_readable to pair classification models (#235)
e371cef Update mkdocs-material requirement from <7.1.0,>=5.5.0 to >=5.5.0,<7.2.0 (#237)
6b306b0 Adds the coref prep script (#239)
155fd76 Fix label namespace in experiment.jsonnet (#240)
abc7e19 Fixed bug in experiment_from_huggingface.jsonnet (#238)
a8e0b00 Make snli dataset reader work with unlabeled instances (#234)
eb72cdc Allow example categories to be ordered (#229)
d97ff8e tick version for nightly