[FIX] Fix compatibility with scikit-learn 0.23

[PrimozGodec]

Issue

The new scikit-learn release yesterday caused orange learners to start failing:

• Learners parameters cannot be retrieved with learner.init.code anymore since init is not called directly.
• KMeans is significantly slower: KMeans singnificantly slower on 0.23 scikit-learn/scikit-learn#17208

Description of changes

• Use inspect.signature instead of __code__ to get parameters.
• I suggest temporary pin scikit-learn to != 0.23.0.

Discussion

Should we release orange with these fixes after it is merged? Since I think we will not make a major release soon, people installing Orange with pip/conda will still get the newest version of scikit-learn. The same is for tests on add-ons. We just need a release for pip and condo. The installed do not need to be updated since it has packages packed together with orange.

Includes

• Code changes
• Tests
• Documentation

[codecov]

Codecov Report

Merging #4768 into master will decrease coverage by 0.05%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #4768      +/-   ##
==========================================
- Coverage   83.86%   83.81%   -0.06%     
==========================================
  Files         281      276       -5     
  Lines       56804    55900     -904     
==========================================
- Hits        47638    46850     -788     
+ Misses       9166     9050     -116     

[janezd]

Doesn't pinning to 0.22 already do the job of fixing the test? Would it make sense to have a (temporary) pinning in a separate PR, so tests we'll pass, and then you can take your time fixing the pickling issue?

[PrimozGodec]

Doesn't pinning to 0.22 already do the job of fixing the test?

Pinning to >=0.22 does not help since the problematic version is 0.23

Would it make sense to have a (temporary) pinning in a separate PR, so tests we'll pass, and then you can take your time fixing the pickling issue?

The main reason why I pinned is k-means issue scikit-learn/scikit-learn#17208 K-means is extremely slow with the latest release. I think it should be pinned until they fix the issue and make a new release. If that will be soon maybe it is not required to merge pinning, but meanwhile, tests on PRs will fail.

For pickling, it is funny since I cannot reproduce the issue at my computer anymore. I will try on some other machine.

[markotoplak]

@PrimozGodec, so you suggest releasing 3.25.1 and building pypi/conda packages, which would be 3.25 with frozen scikit-learn dependency?

On conda forge we could fix requirements without doing a new release (there we can release new builds), but for pip we would need a new release.

[PrimozGodec]

Yep, I think we need to fix versions on both conda and pip (for people who directly download from pip and for CI on addons). I agree with you that on conda we can easily make a new build without rising the version. On pip we can make either 3.25.1 or 3.25.0.post1.