API prepare change of default solver QuantileRegressor #23637
Conversation
…_quantile_transformer
|
When will scipy 1.6 be our minimal scipy version? |
|
https://scipy.org/news/#releases and in particular
From NEP29 which is a minimum requirement for scikit-learn:
scikit-learn 1.1.0 was released in 2022-05, so if we anticipate to release scikit-learn 1.2 in 2022-11, that means that we need to at least support 2022-11 - 24 == 2020-11 which includes SciPy 1.6 but not necessarily SciPy 1.5. |
|
But we can safely anticipate that for scikit-learn 1.4 the scipy dependency will be at least 1.6 (more likely 1.7). So I think we can directly change this PR to skip the "auto" option and warn that in scikit-learn 1.4 |
|
Works for me. I will update this PR. |
| @@ -120,7 +126,7 @@ def test_quantile_estimates_calibration(q): | |||
| quant = QuantileRegressor( | |||
| quantile=q, | |||
| alpha=0, | |||
| solver_options={"lstsq": False}, | |||
| solver="highs", | |||
There was a problem hiding this comment.
This tests breaks on old scipy versions (<1.6.0) that we still support for now.
There was a problem hiding this comment.
There are other tests to be updated similarly. We could probably have a help to chose the solver to use in the test based on the installed version of scipy.
There was a problem hiding this comment.
I did a small fixture to get the default value to use depending of the version.
There was a problem hiding this comment.
LGTM.
Should update update the example examples/linear_model/plot_quantile_regression.py to avoid raising the warning?
Since we need to run on the oldest supported scipy we might need to do the same ternary operation:
solver = "highs" if sp_version >= parse_version("1.6.0") else "interior-point"in the example itself. But maybe we can just live with the warning. As you prefer.
Co-authored-by: Olivier Grisel <olivier.grisel@ensta.org>
|
I think this is nicer to not have the warning. I add the if/else condition with a short explanation for our reader advising to use |
…23637) Co-authored-by: Olivier Grisel <olivier.grisel@ensta.org>
Related to #23626
"interior-point"solver will be removed from SciPy 1.11.We need to introduce an
"auto"mode to switch from one solver to another and announce a change of default.