ENH Raises warning when getting non-finite score in SearchCV

doc/whats_new/v0.24.rst

@@ -353,6 +353,11 @@ Changelog
   all distributions are lists and `n_iter` is more than the number of unique
   parameter combinations. :pr:`18222` by `Nicolas Hug`_.
 
+- |Fix| A fix to raise warning when one or more CV splits of
+  :class:`GridSearchCV` results in inf or -inf.
+  :pr:`18266` by :user:`Subrat Sahu <subrat93>`,
+  :user:`Nirvan <Nirvan101>` and :user:`Arthur Book <ArthurBook>`.
+
 :mod:`sklearn.multiclass`
 .........................
 
@@ -364,7 +369,7 @@ Changelog
 - |Enhancement| :class:`multiclass.OneVsOneClassifier` now accepts
   the inputs with missing values. Hence, estimators which can handle
   missing values (may be a pipeline with imputation step) can be used as
-  a estimator for multiclass wrappers. 
+  a estimator for multiclass wrappers.
   :pr:`17987` by :user:`Venkatachalam N <venkyyuvy>`.
 
 :mod:`sklearn.multioutput`
@@ -375,11 +380,11 @@ Changelog
   :pr:`18124` by :user:`Gus Brocchini <boldloop>` and
   :user:`Amanda Dsouza <amy12xx>`.
 
-- |Enhancement| :class:`multioutput.MultiOutputClassifier` and 
+- |Enhancement| :class:`multioutput.MultiOutputClassifier` and
   :class:`multioutput.MultiOutputRegressor` now accepts the inputs
   with missing values. Hence, estimators which can handle missing
   values (may be a pipeline with imputation step, HistGradientBoosting
-  estimators) can be used as a estimator for multiclass wrappers. 
+  estimators) can be used as a estimator for multiclass wrappers.
   :pr:`17987` by :user:`Venkatachalam N <venkyyuvy>`.
 
 :mod:`sklearn.naive_bayes`

sklearn/model_selection/_search.py

@@ -863,6 +863,15 @@ def _store(key_name, array, weights=None, splits=False, rank=False):
 
             array_means = np.average(array, axis=1, weights=weights)
             results['mean_%s' % key_name] = array_means
+
+            if (key_name.startswith(("train_", "test_")) and
+                    np.any(~np.isfinite(array_means))):
+                warnings.warn(
+                    f"One or more of the {key_name.split('_')[0]} scores "
+                    f"are non-finite: {array_means}",
+                    category=UserWarning
+                )
+
             # Weighted std is not directly available in numpy
             array_stds = np.sqrt(np.average((array -
                                              array_means[:, np.newaxis]) ** 2,

sklearn/model_selection/tests/test_search.py

@@ -25,6 +25,8 @@
 from sklearn.utils._mocking import CheckingClassifier, MockDataFrame
 
 from scipy.stats import bernoulli, expon, uniform
+from scipy.stats.distributions import norm
+
 
 from sklearn.base import BaseEstimator, ClassifierMixin
 from sklearn.base import clone
@@ -1750,6 +1752,42 @@ def get_n_splits(self, *args, **kw):
         ridge.fit(X[:train_size], y[:train_size])
 
 
+@pytest.mark.parametrize(
+    "return_train_score, expected_msgs",
+    [(False, ('One or more of the test scores are non-finite')),
+     (True, ("One or more of the test scores are non-finite",
+             "One or more of the train scores are non-finite"))]
+)
+def test_gridsearchcv_raise_warning_with_non_finite_score(
+        return_train_score, expected_msgs):
+    # Non-regression test for:
+    # https://github.com/scikit-learn/scikit-learn/issues/10529
+    # Check that we raise a UserWarning when a non-finite score is
+    # computed in the GridSearchCV
+    X = norm(-1, 0.5).rvs(100, random_state=np.random.RandomState(28))
+    kernel = 'epanechnikov'
+    steps = 10
+    lower = 0.0194867441113
+    upper = 0.0974337205567
+    bandwidth_range = np.linspace(lower, upper, steps)
+    grid = GridSearchCV(
+        KernelDensity(kernel=kernel),
+        param_grid={'bandwidth': bandwidth_range},
+        cv=20,
+        return_train_score=return_train_score
+    )
+
+    with pytest.warns(UserWarning) as warnings:
+        grid.fit(X[:, np.newaxis])
+
+    warnings = list(map(lambda warning: str(warning.message), warnings))
+    warnings = ",".join(warnings)
+    assert expected_msgs[0] in warnings
+
+    if return_train_score:
+        assert expected_msgs[1] in warnings
+
+
 def test_callable_multimetric_confusion_matrix():
     # Test callable with many metrics inserts the correct names and metrics
     # into the search cv object