python-crfsuite

Support for CPython 3.11

What's Changed

• Support for 3.10 added

• Python 3.4 is no longer supported (it may work, but CI is disabled)

• Python 3.8 support

• fixed installation issues on OS X (thanks @kvinwang)

• make it easier for distributions to have a reproducible build (thanks @bmwiedemann)

• Python 3.7 support (thanks @fgregg, @danmacnaughtan and @fuhrysteve).

• Python 3.3 support is dropped.

• new Tagger.open_inmemory method which allows to load tagger data without having a file on-disk (thanks @lucywang000).

• license information is added to setup.py (thanks @nils-werner).

• Python 3.6 wheels for Windows (thanks @fgregg).

• Packaging fix (thanks @fgregg).

• binary wheels for OS X and Linux (thanks @jeancochrane).

This is a release without changes in functionality.

• Repository is moved to https://github.com/scrapinghub/python-crfsuite;

• We're now providing Windows wheels for Python 2.7, 3.3. and 3.4.

• Python 2.6 support is dropped;

• CRFSuite C++ library is updated to a more recent commit;

• improved Windows support (thanks @fgregg);

• fixed building with gcc < 5.0.0 (thanks @kantan2015);

• extension is rebuilt with Cython 0.25.1; this improves PyPy compatibility (but we're not quite there yet).

• docs: trainer.logparser example is added to the notebook (thanks @samgalen).

• the wrapper is rebuilt with Cython 0.23.4;

• declared Python 3.5 compatibility;

• fixed an issue with feature names ending with white spaces.

• fix build on Windows. (thanks @fgregg)

• memory leak is fixed by updating the bundled CRFsuite C++ library;

• the wrapper is rebuilt with Cython 0.21.2.

• fix packaging issues with 0.8 release.

• ~ItemSequence wrapper is added;

• tox tests are fixed.

• More data formats for xseq: {"prefix": {feature_dict}} and {"key": set(["key1",...])} feature dicts are now accepted by pycrfsuite.Trainer and pycrfsuite.Tagger;

• feature separator changed from "=" to ":" (it looks better in case of multi-level features);

• small doc and README fixes.

• Switch to setuptools;

• wheels are uploaded to pypi for faster installation.

• More data formats for xseq: {"key": "value"} and {"key": bool_value} feature dicts are now accepted by pycrfsuite.Trainer and pycrfsuite.Tagger.

• Exceptions in logging message handlers are now propagated and raised. This allows, for example, to stop training earlier by pressing Ctrl-C.

• It is now possible to customize pycrfsuite.Trainer logging more easily by overriding the following methods: pycrfsuite.Trainer.on_start, pycrfsuite.Trainer.on_featgen_progress, pycrfsuite.Trainer.on_featgen_end, pycrfsuite.Trainer.on_prepared, pycrfsuite.Trainer.on_prepare_error, pycrfsuite.Trainer.on_iteration, pycrfsuite.Trainer.on_optimization_end pycrfsuite.Trainer.on_end. The feature is implemented by parsing CRFsuite log. There is pycrfsuite.BaseTrainer that is not doing this.

• pycrfsuite.Tagger.info() is fixed.

• (backwards-incompatible) training parameters are now passed using params argument of pycrfsuite.Trainer constructor instead of **kwargs;

• (backwards-incompatible) logging support is dropped;

• verbose argument for pycrfsuite.Trainer constructor;

• pycrfsuite.Trainer.get_params and pycrfsuite.Trainer.set_params for getting/setting multiple training parameters at once;

• string handling in Python 3.x is fixed by rebuilding the wrapper with Cython 0.21dev;

• algorithm names are normalized to support names used by crfsuite console utility and documented in crfsuite manual;

• type conversion for training parameters is fixed: feature.minfreq now works, and boolean arguments become boolean.

python-crfsuite now detects the feature format (dict vs list of strings) automatically - it turns out the performance overhead is negligible.

• Trainer.append_stringslists and Trainer.append_dicts methods are replaced with a single pycrfsuite.Trainer.append method;

• Tagger.set_stringlists and Tagger.set_dicts methods are removed in favor of pycrfsuite.Tagger.set method;

• feature_format arguments in pycrfsuite.Tagger methods and constructor are dropped.

• pycrfsuite.Tagger.dump() and pycrfsuite.Tagger.info() methods for model debugging;

• a memory leak in Trainer is fixed (trainer instances were never garbage collected);

• documentation and testing improvements.

Many changes; python-crfsuite is almost rewritten.