fixing bug in linear_model.SGDClassifier for multi-class warm start

Author: sergeyk

sklearn/linear_model/stochastic_gradient.py

@@ -268,6 +268,7 @@ def fit_binary(est, i, X, y, alpha, C, learning_rate, n_iter,
     The i'th class is considered the "positive" class.
     """
     y_i, coef, intercept = _prepare_fit_binary(est, y, i)
+
     assert y_i.shape[0] == y.shape[0] == sample_weight.shape[0]
     dataset, intercept_decay = _make_dataset(X, y_i, sample_weight)
 
@@ -361,7 +362,7 @@ def _partial_fit(self, X, y, alpha, C,
                                                            self.classes_, y_ind)
         sample_weight = self._validate_sample_weight(sample_weight, n_samples)
 
-        if self.coef_ is None:
+        if self.coef_ is None or coef_init is not None:
             self._allocate_parameter_mem(n_classes, n_features,
                                          coef_init, intercept_init)
 

sklearn/linear_model/tests/test_sgd.py

@@ -104,7 +104,7 @@ def decision_function(self, X, *args, **kw):
 
 class CommonTest(object):
 
-    def _test_warm_start(self, lr):
+    def _test_warm_start(self, X, Y, lr):
         # Test that explicit warm restart...
         clf = self.factory(alpha=0.01, eta0=0.01, n_iter=5, shuffle=False,
                            learning_rate=lr)
@@ -131,13 +131,16 @@ def _test_warm_start(self, lr):
         assert_array_almost_equal(clf3.coef_, clf2.coef_)
 
     def test_warm_start_constant(self):
-        self._test_warm_start("constant")
+        self._test_warm_start(X, Y, "constant")
 
     def test_warm_start_invscaling(self):
-        self._test_warm_start("invscaling")
+        self._test_warm_start(X, Y, "invscaling")
 
     def test_warm_start_optimal(self):
-        self._test_warm_start("optimal")
+        self._test_warm_start(X, Y, "optimal")
+
+    def test_warm_start_multiclass(self):
+        self._test_warm_start(X2, Y2, "optimal")
 
     def test_multiple_fit(self):
         """Test multiple calls of fit w/ different shaped inputs."""