Algorithms for outlier, adversarial and drift detection
This is a patch release fixing several bugs, updating dependencies and a change of license.
This is a patch release to officially enable support for Python 3.11. This is the last release with official support for Python 3.7.
mypy ignore commands (#794).scikit-image version to 0.21.x (#803).numba version to 0.57.x (#783).sphinx version to 7.x (#782).plot_feature_outlier_image utility function when no outliers are detected (#774 - thanks @signupatgmx !).tensorflow optimizers to work with the new optimizers introduced in 2.11 (#739).tensorflow bumped to 2.12.x (#764).tensorflow-probability version to 0.19.x (#687).pandas bumped to >1.0.0, <3.0.0 (#765).scikit-image bumped to 0.20.x (#751).codecov to use Github Actions and don't fail CI on coverage report upload failure due to rate limiting (#768, #776).mypy version to >=1.0, <2.0 (#754).sphinx version to 6.x (#709).sphinx-design version to 0.4.1 (#769).nbsphinx version to 0.9.x (#757).myst-parser version to >=1.0, <2.0 (#756).twine version to 4.x (#511).pre-commit version to 3.x and update the config (#731).preprocess_fn's (#752):
preprocess_fn was a custom Python function wrapped in a partial, kwarg's were not serialized. This has now been fixed.preprocess_fn's, the filenames for kwargs saved to .dill files are now prepended with the kwarg name. This avoids files being overwritten if multiple kwargs are saved to .dill.UAE preprocessing utility function (#656, (#705).ClassifierDrift and SpotTheDiffDrift detectors, we can also return the out-of-fold instances of the reference and test sets. When using train_size for training the detector, this allows to associate the returned prediction probabilities with the correct instances (#665).prophet version bumped to 1.1.0 (used by OutlierProphet). This upgrade removes the dependency on pystan as cmdstanpy is used instead. This version also comes with pre-built wheels for all major platforms and Python versions, making both installation and testing easier (#627).config_spec has been removed. In order to load detectors serialized from previous Alibi Detect versions, the field will need to be deleted from the detector's config.toml file. However, in any case, serialization compatibility across Alibi Detect versions is not currently guranteed. (#641).optimizer kwarg would also be set when a detector was loaded with load_detector, regardless of the optimizer given to the original detector (#656).flavour backend is now validated whilst taking into account the optional dependencies. For example, a ValidationError will be raised if flavour='pytorch' is given but PyTorch is not installed (#656).categories_per_feature dictionary is not passed to TabularDrift, a warning is now raised to inform the user that all features are assumed to be numerical (#606).tensorflow version has been bumped from 2.9 to 2.10 (#608).torch version has been bumped from 1.12 to 1.13 (#669).kernel_a and kernel_b in DeepKernel's (#656).device='cpu' was passed to PyTorch based detectors (#698).IndexError's to be raised in the TensorFlow MMDDriftOnline detector when older numpy versions were installed (#710).README.md is opened by setup.py. This is to prevent pip install errors on systems with PYTHONIOENCODING set to use other encoders (#605).test/ directories are now ignored when measuring testing code coverage. This has a side-effect of lowering the reported test coverage (#614).<v0.10.0
(#729 and #732). This bug also meant that detectors
saved with save_detector(..., legacy=True) in >=v0.10.0 did not properly obey the legacy file format. The config.toml file format used by default in >=v0.10.0 is unaffected.