NeuroDataDesign
diff --git a/‎doc/whats_new/v0.22.rst‎
Lines changed: 4 additions & 0 deletions b/‎doc/whats_new/v0.22.rst‎
Lines changed: 4 additions & 0 deletions
diff --git a/‎sklearn/linear_model/logistic.py‎
Lines changed: 18 additions & 3 deletions
10000 b/‎sklearn/linear_model/logistic.py‎
Lines changed: 18 additions & 3 deletions
10000
diff --git a/‎sklearn/linear_model/tests/test_logistic.py‎
Lines changed: 31 additions & 0 deletions b/‎sklearn/linear_model/tests/test_logistic.py‎
Lines changed: 31 additions & 0 deletions
@@ -343,6 +343,10 @@ Changelog
   now correctly scores when `cv=None`.
   :pr:`14864` by :user:`Venkatachalam N <venkyyuvy>`.
 
+- |FIX| Fixed a bug in :class:`linear_model.LogisticRegressionCV` where the
+  ``scores_``, ``n_iter_`` and ``coefs_paths_`` attribute would have a wrong
+  ordering with ``penalty='elastic-net'``. :pr:`15044` by `Nicolas Hug`_
+
 - |FIX| :class:`linear_model.MultiTaskLassoCV` and
   :class:`linear_model.MultiTaskElasticNetCV` with X of dtype int
   and `fit_intercept=True`.
 
@@ -2201,14 +2201,29 @@ def fit(self, X, y, sample_weight=None):
         # if elasticnet was used, add the l1_ratios dimension to some
         # attributes
         if self.l1_ratios is not None:
+            # with n_cs=2 and n_l1_ratios=3
+            # the layout of scores is
+            # [c1, c2, c1, c2, c1, c2]
+            #   l1_1 ,  l1_2 ,  l1_3
+            # To get a 2d array with the following layout
+            #      l1_1, l1_2, l1_3
+            # c1 [[ .  ,  .  ,  .  ],
+            # c2  [ .  ,  .  ,  .  ]]
+            # We need to first reshape and then transpose.
+            # The same goes for the other arrays
             for cls, coefs_path in self.coefs_paths_.items():
                 self.coefs_paths_[cls] = coefs_path.reshape(
-                    (len(folds), self.Cs_.size, self.l1_ratios_.size, -1))
+                    (len(folds), self.l1_ratios_.size, self.Cs_.size, -1))
+                self.coefs_paths_[cls] = np.transpose(self.coefs_paths_[cls],
+                                                      (0, 2, 1, 3))
             for cls, score in self.scores_.items():
                 self.scores_[cls] = score.reshape(
-                    (len(folds), self.Cs_.size, self.l1_ratios_.size))
+                    (len(folds), self.l1_ratios_.size, self.Cs_.size))
+                self.scores_[cls] = np.transpose(self.scores_[cls], (0, 2, 1))
+
             self.n_iter_ = self.n_iter_.reshape(
-                (-1, len(folds), self.Cs_.size, self.l1_ratios_.size))
+                (-1, len(folds), self.l1_ratios_.size, self.Cs_.size))
+            self.n_iter_ = np.transpose(self.n_iter_, (0, 1, 3, 2))
 
         return self
 
 
@@ -13,6 +13,7 @@
 from sklearn.model_selection import StratifiedKFold
 from sklearn.model_selection import GridSearchCV
 from sklearn.model_selection import train_test_split
+from sklearn.model_selection import cross_val_score
 from sklearn.preprocessing import LabelEncoder, StandardScaler
 from sklearn.utils import compute_class_weight, _IS_32BIT
 from sklearn.utils.testing import assert_almost_equal
@@ -1760,3 +1761,33 @@ def test_penalty_none(solver):
         "LogisticRegressionCV",
         lr.fit, X, y
     )
+
+
+def test_scores_attribute_layout_elasticnet():
+    # Non regression test for issue #14955.
+    # when penalty is elastic net the scores_ attribute has shape
+    # (n_classes, n_Cs, n_l1_ratios)
+    # We here make sure that the second dimension indeed corresponds to Cs and
+    # the third dimension corresponds to l1_ratios.
+
+    X, y = make_classification(n_samples=1000, random_state=0)
+    cv = StratifiedKFold(n_splits=5, shuffle=False)
+
+    l1_ratios = [.1, .9]
+    Cs = [.1, 1, 10]
+
+    lrcv = LogisticRegressionCV(penalty='elasticnet', solver='saga',
+                                l1_ratios=l1_ratios, Cs=Cs, cv=cv,
+                                random_state=0)
+    lrcv.fit(X, y)
+
+    avg_scores_lrcv = lrcv.scores_[1].mean(axis=0)  # average over folds
+
+    for i, C in enumerate(Cs):
+        for j, l1_ratio in enumerate(l1_ratios):
+
+            lr = LogisticRegression(penalty='elasticnet', solver='saga', C=C,
+                                    l1_ratio=l1_ratio, random_state=0)
+
+            avg_score_lr = cross_val_score(lr, X, y, cv=cv).mean()
+            assert avg_scores_lrcv[i, j] == pytest.approx(avg_score_lr)