DOC Update rst doctests to be compatible with numpy >= 2

[lesteve]

Similar to #29613 but for doctests inside rst files.

Let's see what the CI has to say about it.

Edit: turns out we were not running the rst doctests in any of our CI build (because the combination numpy<2 and matplotlib and maybe other things to run doctests were not inside any CI build) ... oh well there was only one failure to fix.

[github-actions]

✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: 51ada5b. Link to the linter CI: here}

[lesteve]

CI is green and the rst doctests ran in a few CI builds e.g. this one and this one

[lesteve]

I had to skip this one because the estimator is defined interactively and does not pickle ...

____________________________________________________________________________________________________________ [doctest] develop.rst _____________________________________________________________________________________________________________
341       ...         # Input validation
342       ...         X = validate_data(self, X, reset=False)
343       ...
344       ...         closest = np.argmin(euclidean_distances(X, self.X_), axis=1)
345       ...         return self.y_[closest]
346 
347 And you can check that the above estimator passes all common checks::
348 
349     >>> from sklearn.utils.estimator_checks import check_estimator
350     >>> check_estimator(TemplateClassifier())  # passes
UNEXPECTED EXCEPTION: PicklingError("Can't pickle <class 'TemplateClassifier'>: attribute lookup TemplateClassifier on builtins failed")
Traceback (most recent call last):
  File "/home/lesteve/micromamba/envs/scikit-learn-dev/lib/python3.13/doctest.py", line 1395, in __run
    exec(compile(example.source, filename, "single",
    ~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                 compileflags, True), test.globs)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "<doctest develop.rst[10]>", line 1, in <module>
  File "/home/lesteve/dev/scikit-learn/sklearn/utils/_param_validation.py", line 216, in wrapper
    return func(*args, **kwargs)
  File "/home/lesteve/dev/scikit-learn/sklearn/utils/estimator_checks.py", line 857, in check_estimator
    check(estimator)
    ~~~~~^^^^^^^^^^^
  File "/home/lesteve/dev/scikit-learn/sklearn/utils/_testing.py", line 147, in wrapper
    return fn(*args, **kwargs)
  File "/home/lesteve/dev/scikit-learn/sklearn/utils/estimator_checks.py", line 2338, in check_estimators_pickle
    pickled_estimator = pickle.dumps(estimator)
_pickle.PicklingError: Can't pickle <class 'TemplateClassifier'>: attribute lookup TemplateClassifier on builtins failed
/home/lesteve/dev/scikit-learn/doc/developers/develop.rst:350: UnexpectedException
=====================================================================================================

[lesteve]

Not sure what to show here maybe only ... would be enough?

[lesteve]

I ended up using ..., I don't think it is useful to try to show part of the output ...

[thomasjpfan]

It's unfortunate that the output is so much more verbose now. But I'm okay with moving into the future with NumPy's new printing style.

[jeremiedbb]

Reading #27339, it looks like there's a consensus for always returning python scalars from scorers. So I'm afraid that all the changes here will have to be reverted once this is done. Shouldn't we do the "always return a python scalar" first ?

[lesteve]

This PR is a simple one that makes sure that we are actually testing our rst file doctests because we are currently not testing them and we never realized ...

When #27339 happens, I am more than happy to be pinged to fix the doctests 😉.

[lesteve]

Side-comment, there is also #30496 that could help because scipy-doctests allows you to ignore the difference between 0.214 and np.float64(0.214), see #30495 (comment).

But same thing I am not very keen on making a "Quick Review" PR depend on a potentially more controversial one 😉

[jeremiedbb]

LGTM.

beyond the scorers output repr, there are also a few ones from different origins that I find really not user friendly, like

    >>> list(le.classes_)
    [np.str_('amsterdam'), np.str_('paris'), np.str_('tokyo')]

that we might consider dealing with at some point as well.

[lesteve]

beyond the scorers output repr, there are also a few ones from different origins that I find really not user friendly

Yep agreed. IIRC sometimes we get this because we do things like list(array) rather than array.tolist() in the doctest code but in other cases this is nested deeper inside the scikit-learn code ...