scikit-learn
diff --git a/‎doc/whats_new/v0.20.rst
Lines changed: 6 additions & 0 deletions b/‎doc/whats_new/v0.20.rst
Lines changed: 6 additions & 0 deletions
diff --git a/‎sklearn/linear_model/logistic.py
Lines changed: 5 additions & 3 deletions b/‎sklearn/linear_model/logistic.py
Lines changed: 5 additions & 3 deletions
diff --git a/‎sklearn/linear_model/tests/test_logistic.py
Lines changed: 20 additions & 0 deletions b/‎sklearn/linear_model/tests/test_logistic.py
Lines changed: 20 additions & 0 deletions
@@ -587,6 +587,12 @@ Classifiers and regressors
   considered all samples to be of equal weight importance.
   :issue:`11464` by :user:`John Stott <JohnStott>`.
 
+- Fixed a bug in :func:`logistic.logistic_regression_path` to ensure that the
+  returned coefficients are correct when ``multiclass='multinomial'``.
+  Previously, some of the coefficients would override each other, leading to
+  incorrect results in :class:`logistic.LogisticRegressionCV`. :issue:`11724`
+  by :user:`Nicolas Hug <NicolasHug>`.
+
 Decomposition, manifold learning and clustering
 
 - Fix for uninformative error in :class:`decomposition.IncrementalPCA`:
 
@@ -572,7 +572,9 @@ def logistic_regression_path(X, y, pos_class=None, Cs=10, fit_intercept=True,
     coefs : ndarray, shape (n_cs, n_features) or (n_cs, n_features + 1)
         List of coefficients for the Logistic Regression model. If
-        n_features + 1, where the last item represents the intercept.
+        n_features + 1, where the last item represents the intercept. For
+        ``multiclass='multinomial'``, the shape is (n_classes, n_cs,
+        n_features) or (n_classes, n_cs, n_features + 1).
 
     Cs : ndarray
         Grid of Cs used for cross-validation.
@@ -762,13 +764,13 @@ def logistic_regression_path(X, y, pos_class=None, Cs=10, fit_intercept=True,
             multi_w0 = np.reshape(w0, (classes.size, -1))
             if classes.size == 2:
                 multi_w0 = multi_w0[1][np.newaxis, :]
-            coefs.append(multi_w0)
+            coefs.append(multi_w0.copy())
         else:
             coefs.append(w0.copy())
 
         n_iter[i] = n_iter_i
 
-    return coefs, np.array(Cs), n_iter
+    return np.array(coefs), np.array(Cs), n_iter
 
 
 # helper function for LogisticCV
 
@@ -1307,3 +1307,23 @@ def test_warm_start_converge_LR():
         lr_ws.fit(X, y)
     lr_ws_loss = log_loss(y, lr_ws.predict_proba(X))
     assert_allclose(lr_no_ws_loss, lr_ws_loss, rtol=1e-5)
+
+
+def test_logistic_regression_path_coefs_multinomial():
+    # Make sure that the returned coefs by logistic_regression_path when
+    # multi_class='multinomial' don't override each other (used to be a
+    # bug).
+    X, y = make_classification(n_samples=200, n_classes=3, n_informative=2,
+                               n_redundant=0, n_clusters_per_class=1,
+                               random_state=0, n_features=2)
+    Cs = [.00001, 1, 10000]
+    coefs, _, _ = logistic_regression_path(X, y, penalty='l1', Cs=Cs,
+                      
647F
                     solver='saga', random_state=0,
+                                           multi_class='multinomial')
+
+    with pytest.raises(AssertionError):
+        assert_array_almost_equal(coefs[0], coefs[1], decimal=1)
+    with pytest.raises(AssertionError):
+        assert_array_almost_equal(coefs[0], coefs[2], decimal=1)
+    with pytest.raises(AssertionError):
+        assert_array_almost_equal(coefs[1], coefs[2], decimal=1)