All notable changes to this project will be documented in this file.
- Add HAP/detoxify module to Python TrustyAI (#197)
- Update CI Ubuntu version from 20.04 to 22.04 (#199)
- feat: Publish using PyPi Trusted Providers (#183)
- Update README.md (#184)
- Update pyarrow and pillow dependencies (#187)
- Add WER metric (#189)
- Change WER signature (#190)
- Add Levenshtein distance (#191)
- fix typo in tutorial.rst (#193)
This is the first version of Python TrustyAI to include support for external explainability algorithms.
In this release we've included AIX360's TSICE, TSLime and TSSaliency time-series explainers.
To use these explainers TrustyAI's extra dependencies must be installed with pip install trustyai[extras].
- Refactor TrustyAI fairness metrics namespaces (#156)
- Upgrade pyarrow dependency (#159)
- Initial support for external algorithms (#160)
- Replace deprecated requirements file (#166)
- Update date, logo and favicon (#168)
- Support TSICE explanations as plots (#169)
- Add TSSaliency to Python TrustyAI (#172)
- Add TSLime explainer by (#174)
- Update TSSaliency plots to include smoothing (#176)
- Fix invalid Development Status (#178)
- Update direct dependency to released dependency (#179)
- Fixed broken gridding after bokeh 3.0.x update (#153)
- Fixed null pointer error for Numpy inputs to Tyrus (#149)
- Add feature and output name specification to models (#130)
- Python benchmarks failing since namespace migration (#82)
- Tyrus dashboard (#97)
- Move arrow converters into exp-core (#101)
- Fix failing RtD build after to pip migration (#100)
- Add Java 9+ Arrow compatibility flags (#103)
- Sync version with exp-core (#107)
- Move explainers into separate files (#110)
- Upgrade Pylint version (#112)
- Add kwargs to explainers (#113)
- Add SHAP background generators to bindings (#117)
- Unified input/output types, conversion functions, and docstrings (#116)
- Generalize ExplanationResults (#115)
- Allow non-string categorical feature domains (#118)
- Tyrus TempDirectory (#121)
- Make test plots non-blocking by default (#120)
- Doc standardization and cleanup (#126)
- Add bokeh dependency (#123)
- Added feature domain argument to counterfactuals (#128)
- Increased input/ouput conversion flexibility (#127)
- Add group fairness Python bindings (#129)
- Improve API to declare selections in Python fairness (#132)
- Add error message capturing within Python models (#137)
- Implement custom counterfactual goal criteria (#140)
- Change dependencies to align with ODH workbench images (#143)
-
Fixed linting, adding doc build reqs to requirements-dev
-
Fixed line-too-long linting errors
-
Fixed distutils import formatting
-
Fixed conflicts with FAI-806, added build output
-
Fixed incorrect LIME typing
-
Fixed typo in Output field
-
Fixed rtd.yaml python version inconsistency
-
Fixed path in rtd install config
-
Fixed custom.css path
-
Fixed failing tests
-
Fixed black-induced line-too-long
-
Fixed failing shapresults.getSaliencies call
-
Fixed linting errors re. import order
-
Fixed failing initializer tests
-
Fixed unrelated explainers change
-
Fixed incorrect ShapResults.get_saliency output type
-
Fixed broken favicon
-
Fixed linting issue
-
Fixed enumeration tuple issue
-
Fixed failing SHAP tests, improved output name imputation in Model
-
Fixed overzealous find and replace of trustyai
-
Fixed overzealous find and replace of trustyai part 2
-
Fixed broken output casting for 1d output arrays
- Include doc link in README
- Fixed missing final newline
-
Fixed linting issues
-
Fixed linting issues re. the glob checker for wildcard paths
-
Fixed line-too-long linting issue in model/init
- Refactored arrow inclusion; mvn pulls dependencies, then includes precompiled arrowconverters jar