Fix gibbs sampling behavior in RBM with integral random_state.

Author: amueller

sklearn/neural_network/rbm.py

@@ -210,10 +210,10 @@ def gibbs(self, v):
             Values of the visible layer after one Gibbs step.
         """
         check_is_fitted(self, "components_")
-
-        rng = check_random_state(self.random_state)
-        h_ = self._sample_hiddens(v, rng)
-        v_ = self._sample_visibles(h_, rng)
+        if not hasattr(self, "random_state_"):
+            self.random_state_ = check_random_state(self.random_state)
+        h_ = self._sample_hiddens(v, self.random_state_)
+        v_ = self._sample_visibles(h_, self.random_state_)
 
         return v_
 

sklearn/neural_network/tests/test_rbm.py

@@ -131,14 +131,16 @@ def test_fit_gibbs_sparse():
 
 
 def test_gibbs_smoke():
-    """Check if we don't get NaNs sampling the full digits dataset."""
-    rng = np.random.RandomState(42)
+    """Check if we don't get NaNs sampling the full digits dataset.
+    Also check that sampling again will yield different results."""
     X = Xdigits
     rbm1 = BernoulliRBM(n_components=42, batch_size=40,
-                        n_iter=20, random_state=rng)
+                        n_iter=20, random_state=42)
     rbm1.fit(X)
     X_sampled = rbm1.gibbs(X)
     assert_all_finite(X_sampled)
+    X_sampled2 = rbm1.gibbs(X)
+    assert_true(np.all((X_sampled != X_sampled2).max(axis=1)))
 
 
 def test_score_samples():