scikit-learn · ogrisel · Feb 25, 2019 · Feb 12, 2019 · Feb 14, 2019 · Feb 14, 2019
diff --git a/doc/whats_new/v0.21.rst b/doc/whats_new/v0.21.rst
@@ -291,6 +291,15 @@ Support for Python 3.4 and below has been officially dropped.
   affects all ensemble methods using decision trees.
   :issue:`12344` by :user:`Adrin Jalali <adrinjalali>`.
 
+:mod:`sklearn.mixture`
+......................
+
+- |Fix| Fixed a bug in :class:`mixture.BaseMixture` and therefore on estimators
+  based on it, i.e. :class:`mixture.GaussianMixture` and
+  :class:`mixture.BayesianGaussianMixture`, where ``fit_predict`` and
+  ``fit.predict`` were not equivalent. :issue:`13142` by
+  :user:`Jérémie du Boisberranger <jeremiedbb>`.
+
 
 Multiple modules
 ................

diff --git a/sklearn/mixture/base.py b/sklearn/mixture/base.py
@@ -257,11 +257,6 @@ def fit_predict(self, X, y=None):
                 best_params = self._get_parameters()
                 best_n_iter = n_iter
 
-        # Always do a final e-step to guarantee that the labels returned by
-        # fit_predict(X) are always consistent with fit(X).predict(X)
-        # for any value of max_iter and tol (and any random_state).
-        _, log_resp = self._e_step(X)
-
         if not self.converged_:
             warnings.warn('Initialization %d did not converge. '
                           'Try different init parameters, '
@@ -273,6 +268,11 @@ def fit_predict(self, X, y=None):
         self.n_iter_ = best_n_iter
         self.lower_bound_ = max_lower_bound
 
+        # Always do a final e-step to guarantee that the labels returned by
+        # fit_predict(X) are always consistent with fit(X).predict(X)
+        # for any value of max_iter and tol (and any random_state).
+        _, log_resp = self._e_step(X)
+
         return log_resp.argmax(axis=1)
 
     def _e_step(self, X):

diff --git a/sklearn/mixture/tests/test_bayesian_mixture.py b/sklearn/mixture/tests/test_bayesian_mixture.py
@@ -451,6 +451,15 @@ def test_bayesian_mixture_fit_predict(seed, max_iter, tol):
         assert_array_equal(Y_pred1, Y_pred2)
 
 
+def test_bayesian_mixture_fit_predict_n_init():
+    # Check that fit_predict is equivalent to fit.predict, when n_init > 1
+    X = np.random.RandomState(0).randn(1000, 5)
+    gm = BayesianGaussianMixture(n_components=5, n_init=10, random_state=0)
+    y_pred1 = gm.fit_predict(X)
+    y_pred2 = gm.predict(X)
+    assert_array_equal(y_pred1, y_pred2)
+
+
 def test_bayesian_mixture_predict_predict_proba():
     # this is the same test as test_gaussian_mixture_predict_predict_proba()
     rng = np.random.RandomState(0)

diff --git a/sklearn/mixture/tests/test_gaussian_mixture.py b/sklearn/mixture/tests/test_gaussian_mixture.py
@@ -598,6 +598,15 @@ def test_gaussian_mixture_fit_predict(seed, max_iter, tol):
         assert_greater(adjusted_rand_score(Y, Y_pred2), .95)
 
 
+def test_gaussian_mixture_fit_predict_n_init():
+    # Check that fit_predict is equivalent to fit.predict, when n_init > 1
+    X = np.random.RandomState(0).randn(1000, 5)
+    gm = GaussianMixture(n_components=5, n_init=5, random_state=0)
+    y_pred1 = gm.fit_predict(X)
+    y_pred2 = gm.predict(X)
+    assert_array_equal(y_pred1, y_pred2)
+
+
 def test_gaussian_mixture_fit():
     # recover the ground truth
     rng = np.random.RandomState(0)