Skip to content

release raiwidgets and responsibleai v0.37.0

Latest

Choose a tag to compare

@imatiach-msft imatiach-msft released this 19 Nov 22:27
· 3 commits to main since this release

v0.37.0

  • educational materials

  • new features

  • breaking changes

    • Require Python 3.9 or later for responsibleai 0.37.0 and raiwidgets
      0.37.0; Python 3.7 and 3.8 are no longer supported. The independently
      released erroranalysis, raiutils, and rai-test-utils packages also require
      Python 3.9 or later.

    • Require Fairlearn >=0.8.0,<0.9 for raiwidgets Pandas 2 reductions
      compatibility; raiwidgets 0.36.0 required fairlearn==0.7.0.

    • Support NumPy 2 while retaining broad runtime ranges for coherent older
      environments. Key direct runtime constraints are below; exact tested
      combinations belong in CI and image constraints, not universal package
      pins. Other requirements and platform markers remain in package metadata.

      Dependency responsibleai 0.37.0 raiwidgets 0.37.0
      numpy >=1.17.2,<3 >=1.17.2,<3
      pandas >=0.25.1,<3 >=0.25.1,<3
      scipy >=1.4.1,<1.15 >=1.4.1,<1.14
      scikit-learn >=0.22.1,!=1.1,<=1.5.1 >=0.22.1,<=1.5.1
      lightgbm >=2.0.11,<4.6 >=2.0.11,<5
      dice-ml >=0.11,<0.13 Not direct
      econml >=0.15.0,<0.17 Not direct
      statsmodels <0.15.0 Not direct
      interpret-community >=0.31.0,<0.34 Not direct
      ml-wrappers >=0.5.4,<0.7 Not direct
      networkx <4 Not direct
      fairlearn Not direct >=0.8.0,<0.9
      erroranalysis >=0.6.0 >=0.6.0
      raiutils >=0.4.3 >=0.4.3
      rai-core-flask Not direct >=0.7.6
      responsibleai Not direct ==0.37.0

      "Not direct" does not mean unconstrained: raiwidgets depends on the matching
      responsibleai version and inherits its dependencies. Installing both uses
      the intersection of their constraints, including SciPy >=1.4.1,<1.14,
      LightGBM >=2.0.11,<4.6, and the scikit-learn !=1.1 exclusion.
      Interpret Core and SHAP are selected transitively, without direct exact pins.

      These are declared ranges, not a guarantee that every cross-product works:
      transitive requirements can impose higher effective minimums. DiCE 0.11
      requires Pandas <2, while DiCE 0.12 requires Pandas >=2. Keep older Pandas
      wheels on NumPy 1 even when their metadata lacks a NumPy upper bound.
      Official NumPy-2-compatible Pandas wheels start at 2.2.2; this is not a
      universal Pandas minimum for NumPy 1 environments.

    • Published responsibleai-vision 0.3.8 requires Pandas <2. The tabular
      Pandas range still includes Pandas 1.x and does not conflict with that
      cap at the metadata level. This does not establish universal runtime
      compatibility or NumPy 2 support for vision. Existing vision images need
      a coherent legacy NumPy 1/Pandas 1/DiCE 0.11 stack and image-level constraints.

  • bug fixes and tests

  • other

Full Changelog: v0.36.0...v0.37.0