increase coverage

skore/src/skore/sklearn/_estimator/feature_importance_accessor.py

@@ -554,14 +554,17 @@ def _feature_permutation(
                 n_repeats = data.shape[1]
 
                 # Get score name
-                scoring_name = None
                 if scoring is None:
                     if is_classifier(self._parent.estimator_):
                         scoring_name = "accuracy"
                     elif is_regressor(self._parent.estimator_):
                         scoring_name = "r2"
-                if isinstance(scoring, str):
-                    scoring_name = scoring
+                else:
+                    # e.g. if scoring is a callable
+                    scoring_name = None
+
+                    # no other cases to deal with explicitly, because
+                    # scoring cannot possibly be a string at this stage
 
                 if scoring_name is None:
                     index = pd.Index(feature_names, name="Feature")

skore/tests/unit/sklearn/estimator/feature_importance/test_permutation_importance.py

@@ -5,7 +5,7 @@
 import pytest
 from sklearn.datasets import make_regression
 from sklearn.exceptions import NotFittedError
-from sklearn.linear_model import LinearRegression
+from sklearn.linear_model import LinearRegression, LogisticRegression
 from sklearn.metrics import make_scorer, r2_score, root_mean_squared_error
 from sklearn.model_selection import train_test_split
 from sklearn.pipeline import make_pipeline
@@ -70,7 +70,7 @@ def case_r2_numpy():
 def case_train_numpy():
     data = regression_data()
 
-    kwargs = {"data_source": "train", "random_state": 42}
+    kwargs = {"data_source": "train", "scoring": "r2", "random_state": 42}
 
     return data, kwargs, multiindex_numpy, repeat_columns
 
@@ -289,6 +289,26 @@ def test_cache_scoring_is_callable(regression_data, scoring):
     pd.testing.assert_frame_equal(cached_result, result)
 
 
+def test_classification(classification_data):
+    """If `scoring` is a callable then the result is cached properly."""
+
+    X, y = classification_data
+    report = EstimatorReport(LogisticRegression(), X_train=X, y_train=y)
+
+    result = report.feature_importance.feature_permutation(
+        data_source="train", random_state=42
+    )
+
+    pd.testing.assert_index_equal(
+        result.index,
+        pd.MultiIndex.from_product(
+            [["accuracy"], (f"Feature #{i}" for i in range(5))],
+            names=("Metric", "Feature"),
+        ),
+    )
+    pd.testing.assert_index_equal(result.columns, repeat_columns)
+
+
 def test_not_fitted(regression_data):
     """If estimator is not fitted, raise"""
     X, y = regression_data