flairNLP/flair v0.15.0

Release 0.15.0

on GitHub

This release adds multi-GPU support, an improved documentation page with API docs and finally deprecates Python 3.8!

Improved Documentation and API Docs

Thanks to @konstantin-lukas we have a completely new design of our documentation page, which now includes API docs.

You can check it out here!

• Check out our tutorials
• Check out the new Python API docs

Future releases will improve docstring coverage and further improve upon the documentation!

PRs:

• Fix doc build by @helpmefindaname in #3528
• Rework Doc page by @konstantin-lukas in #3563
• Test new docstrings and apidocs deployment by @alanakbik in #3573

Multi-GPU Support

Flair now offers support for training models on multiple GPUs! Big thanks to @jeffpicard!

PRs:

• Add multi-GPU support by @jeffpicard in #3548
• Fix gradient accumulation and learning rate aggregation by @jeffpicard in #3583

Deprecations

Since Python3.8 is no longer supported, we are also dropping support for it, in favor of features added in python 3.9.
To acknowledge CVE-2024-10073, we decided to drop support for the flair.models.clustering module, since we aren't aware of any usage of it, we decided to do a hard drop instead of a deprecation.

• Drop python 3.8 by @helpmefindaname in #3560
• Remove clustering support by @helpmefindaname in #3567

Other Improvements

New Datasets

• Add CleanCoNLL object by @susannaruecker in #3557
• Add NoiseBench object by @elenamer in #3512

Performance Improvements

• perf: optimize dictionary items check by @MattGPT-ai in #3569
• Refactor fill_mean_token_embeddings for performance optimization on GPU by @sheldon-roberts in #3525

New Features and Improvements

• Add proxies information to requests.head by @diego-morientez in #3535
• Allow specifying proxy information in TransformerEmbeddings by @diego-morientez in #3539
• Add use_tokenizer to JsonlDataset by @david-waterworth in #3486
• Use built-in version parsing from packaging by @adrianeboyd in #3502

Bugfixes

• TransformerDocumentEmbeddings: Fix error when cls_pooling="mean" or cls_pooling="max" by @fkdosilovic in #3558
• SequenceTagger : Fix the incorrect token prediction distribution from _all_scores_for_token() by @mdmotaharmahtab in #3449
• TransformerEmbeddings: Fix T5 tokenizer loading by @helpmefindaname in #3544
• TextPairRegressor: Fix: use proper eval default main eval metrics by @MattGPT-ai in #3538
• TextPairRegressor: Fix state dict key mismatch for embeddings by @MattGPT-ai in #3537
• Make onnx export work again by @helpmefindaname in #3530
• Fix support metric by @MattGPT-ai in #3510

Operations/Development

• Invalidate tars classifier and tars ner tests to save disk space by @helpmefindaname in #3527
• Ignore FutureWarning by @alanakbik in #3526
• Update SECURITY.md with current contact by @alanakbik in #3568

New Contributors

• @adrianeboyd made their first contribution in #3502
• @david-waterworth made their first contribution in #3486
• @diego-morientez made their first contribution in #3535
• @jeffpicard made their first contribution in #3548
• @mdmotaharmahtab made their first contribution in #3449
• @fkdosilovic made their first contribution in #3558

Full Changelog: v0.14.0...v0.15.0