Algorithms for explaining machine learning models
This is a patch release fixing several bugs, updating dependencies and a change of license.
This is a patch release to support numpy >= 1.24 and scikit-learn >= 1.3.0, and drop official support for Python 3.7.
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 (#874).torch version to 2.x (#893).scikit-image version to 0.21.x (#928).numba version to 0.57.x (#922).sphinx version to 7.x (#921).This is a patch release fixing several bugs, updating dependencies and adding some small extensions.
IntegratedGradients layer selection to be specified with a custom callable (#894).reset_predictor method for PartialDependence explainer (#897).GradientSimilarity explainer to allow models of any input type (#912).AnchorText auto-regressive language model sampler updating input_ids tensor (#895).AnchorTabular length discrepancy between feature and names fields (#902).AnchorBaseBeam unintended coverage update during the multi-armed bandit run (#919, #914).tensorflow bumped to 2.12.x (#896).pandas bumped to >1.0.0, <3.0.0 (#899).pandas version 2.x deprecations (#908, #910).scikit-image bumped to 0.20.x (#882).attrs bumped to 23.x (#905).codecov to use Github Actions and don't fail CI on coverage report upload failure due to rate limiting (#901, #913).mypy version to >=1.0, <2.0 (#886).sphinx version to 6.x (#852).sphinx-design version to 0.4.1 (#904).nbsphinx version to 0.9.x (#889).myst-parser version to >=1.0, <2.0 (#887).twine version to 4.x (#620).pre-commit version to 3.x and update the config (#866).This is a patch release fixing several bugs.
np.object in the codebase which was causing errors with numpy>=1.24 (#872, #890).cfrl_base.py of the tensorflow backend (#891)..reset_predictor for KernelShap and TreeShap explainers (#880).KernelShap to avoid saving the internal _explainer object (#881).PermutationImportance explainer implementing the permutation feature importance global explanations. Also included is a plot_permutation_importance utility function for flexible plotting of the resulting feature importance scores (docs, #798).PartialDependenceVariance explainer implementing partial dependence variance global explanations. Also included is a plot_pd_variance utility function for flexible plotting of the resulting PD variance plots (docs, #758).GradientSimilarity explainer now automatically handles sparse tensors in the model by converting the gradient tensors to dense ones before calculating similarity. This used to be a source of bugs when calculating similarity for models with embedding layers for which gradients tensors are sparse by default. Additionally, it now filters any non-trainable parameters and doesn't consider those in the calculation as no gradients exist. A warning is raised if any non-trainable layers or parameters are detected (#829).ALE. The previous examples and documentation had some misleading claims; these have been removed and reworked with an emphasis on the mostly qualitative interpretation of ALE plots (#838, #846).ProtoSelect, allowing greater flexibility to the end user (#826).tensorflow models used in CounterfactualRL (#793).pytorch models used in CounterfactualRL (#799).ALE plotting functionality (#816).PartialDependence plotting functionality (#819).PartialDependenceVariance plotting functionality (#820).PermutationImportance plotting functionality (#824).ProtoSelect plotting functionality (#841).datasets subpackage (#814).pytest-cov bumped to 4.x (#794).pytest-xdist bumped to 3.x (#808).tox bumped to 4.x (#832).PartialDependence and TreePartialDependence explainers implementing partial dependence (PD) global explanations. Also included is a plot_pd utility function for flexible plotting of the resulting PD plots (docs, #721).exceptions.NotFittedError exception which is raised whenever a compulsory call to a fit method has not been carried out. Specifically, this is now raised in AnchorTabular.explain when AnchorTabular.fit has been skipped (#732).AnchorTabular where an error is raised during an explain call if the instance contains a categorical feature value not seen in the training data (#742).grid_points for the ALE explainer (#731).Alibi* prefix and standardised the *Error suffix. Concretely:
exceptions.AlibiPredictorCallException is now exceptions.PredictorCallError
exceptions.AlibiPredictorReturnTypeError is now exceptions.PredictorReturnTypeError. Backwards compatibility has been maintained by subclassing the new exception classes by the old ones, but these will likely be removed in a future version (#733).TreeShap is used with more than 100 samples in the background dataset which is due to a limitation in the upstream shap package (#710).scikit-learn bumped to 1.0.0 mainly due to upcoming deprecations (#776).scikit-image bumped to 0.17.2 to fix a possible bug when using the slic segmentation function with AnchorImage (#753).attrs bumped to 22.x (#727).tensorflow bumped to 2.10.x (#745).ray bumped to 2.x (#740).numba bumped to 0.56.x (#724).shap bumped to 0.41.x (#702).shap example notebooks to recommend installing matplotlib==3.5.3 due to failure of shap plotting functions with matplotlib==3.6.0 (#776).pytest-custom_exit_code to let notebook CI pass when no notebooks are selected for tests (#728).README.md in setup.py to avoid a possible failure of installation for some users (#744).ray tests (#756).3.10.x version on CI has been pinned to 3.10.6 due to typechecking failures, pending a new release of mypy (#761).test_changed_notebooks workflow can now be triggered manually and is run on push/PR for any branch (#762).codecov flags for more granular reporting of code coverage (#759).sphinx bumped to 5.x but capped at <5.1.0 to avoid CI failures (#722).myst-parser bumped to 0.18.x (#693).flake8 bumped to 5.x (#729).ipykernel bumped to 6.x (#431).ipython bumped to 8.x (#572).pytest bumped to 7.x (#591).sphinx-design bumped to 0.3.0 (#739).nbconvert bumped to 7.x (#738).This release introduces two new methods, a GradientSimilarity explainer and a ProtoSelect data summarisation algorithm.
GradientSimilarity explainer for explaining predictions of gradient-based (PyTorch and TensorFlow) models by returning the most similar training data points from the point of view of the model (docs).alibi.prototypes which contains the ProtoSelect algorithm for summarising datasets with a representative set of "prototypes" (docs).ALE explainer now can take a custom grid-point per feature to evaluate the ALE on. This can help in certain situations when grid-points defined by quantiles might not be the best choice (docs).IntegratedGradients method target selection to handle explaining any scalar dimension of tensors of any rank (previously only rank-1 and rank-2 were supported). See #635.PyTorch at the time of writing doesn't support Python 3.10 on Windows.CounterfactualProto (#646).CounterfactualRLTabular (#651).tensorflow is now an optional dependency. To use methods that require tensorflow you can install alibi using pip install alibi[tensorflow] which will pull in a supported version. For full instructions for the recommended way of installing optional dependencies please refer to Installation docs.sklearn version bounds to scikit-learn>=0.22.0, <2.0.0.tensorflow maximum allowed version to 2.9.x.MissingDependency object. For full details on developing alibi with optional dependencies see Contributing: Optional Dependencies.Explainer base class into Base and Explainer to facilitate reusability and better class hierarchy semantics with introducing methods that are not explainers (#649).mypy has been updated to ~=0.900 which requires additional development dependencies for type stubs, currently only types-requests has been necessary to add to requirements/dev.txt.doc/ and examples/ from the source distribution (by adding prune directives in MANIFEST.in). This results in considerably smaller file sizes for the source distribution.This is a patch release to correct a regression in CounterfactualProto introduced in v0.6.3.
v0.6.3 which prevented CounterfactualProto working with categorical features (#612).LanguageModelSampler where it would sometimes sample punctuation (#585).tensorflow version has been bumped from 2.7 to 2.8 (#588).This is a patch release to correct a regression in AnchorImage introduced in v0.6.3.
v0.6.3 where AnchorImage would ignore user segmentation_kwargs (#581).Pillow and scikit-image have been bumped to 9.x and 0.19.x respectively.