MNT Remove DeprecationWarning for scipy.sparse.linalg.cg tol vs rtol argument

[lesteve]

Fix most of scipy-dev issues seen in #26791

Looking at
https://github.com/scipy/scipy/pull/18488/files#diff-b65f85e6dd5881e4611b34a584512f8df370e520758c30e8d297e70c61dcf48f, it does not seem there is a way to keep the atol="legacy" behaviour. This means that potentially there will be a behaviour change for our users ...

[github-actions]

❌ Linting issues

There was an issue running the linter job. Please update with upstream/main (link) and push the changes. If you already have done that, please send an empty commit with git commit --allow-empty and push the changes to trigger the CI.

_{Generated for commit: 8d43a31. Link to the linter CI: here}

[lesteve]

OK there is still one issue with np.byte_bounds being deprecated and used in joblib, this was already noted in joblib/joblib#1454, and can be tackled in a separate PR.

[thomasjpfan]

Looks like this PR is conflicting with #26813

[ogrisel]

I wonder if the new tol criterion has an impact on the ongoing PR that aims to check good convergence tolerance #25948.

[lesteve]

I took test_ridge.py from #25948 and enabled the tests with fit_intercept=False and ran the tests for sparse_cg, the intercept=False seems still fail:

pytest sklearn/linear_model/tests/test_ridge.py -k sparse_cg -ra

FAILED sklearn/linear_model/tests/test_ridge.py::test_ridge_regression_unpenalized[wide-42-False-sparse_cg] - AssertionError: 
FAILED sklearn/linear_model/tests/test_ridge.py::test_ridge_regression_unpenalized_hstacked_X[wide-42-False-sparse_cg] - AssertionError: 
FAILED sklearn/linear_model/tests/test_ridge.py::test_ridge_regression_unpenalized_vstacked_X[wide-42-False-sparse_cg] - AssertionError: 

Let me know in case you had something else in mind!

[thomasjpfan]

Thank you for the PR! I do not see a better alternative to this PR. LGTM

[lesteve]

Note the remaining errors all look like some formatting changes for numpy scalars in numpy development version, better to leave them for a separate PR. This is due to numpy/numpy#22449 that implements NEP 51.

For example:

E       AssertionError: Regex pattern did not match.
E        Regex: "The classes, \\['dog'\\], are not in class_weight"
E        Input: "The classes, [np.str_('dog')], are not in class_weight"

[glemaitre]

LGTM as well.

[glemaitre]

I think it is worth to acknowledge the changes in the Changed models section.
I agree with @thomasjpfan that we might not have a better way to do handle this case.

[lesteve]

Merging my own PR with two approvals 😅

[lesteve]

It looks like the remaining failures may be fixed by #27042