renamed supported_loss to _SUPPORTED_LOSS (constants)

add test for deprecated warning

Author: pprett

sklearn/ensemble/gradient_boosting.py

@@ -498,10 +498,14 @@ def _check_params(self):
         if self.learning_rate <= 0.0:
             raise ValueError("learning_rate must be greater than 0")
 
-        if (self.loss not in self.supported_loss or
+        if (self.loss not in self._SUPPORTED_LOSS or
             self.loss not in LOSS_FUNCTIONS):
             raise ValueError("Loss '{0:s}' not supported. ".format(self.loss))
 
+        if self.loss in ('mdeviance', 'bdeviance'):
+            warn(("Loss '{0:s}' is deprecated as of version 0.14. "
+                 "Use 'deviance' instead. ").format(self.loss))
+
         if self.loss == 'deviance':
             loss_class = (MultinomialDeviance
                           if len(self.classes_) > 2
@@ -854,7 +858,7 @@ class GradientBoostingClassifier(BaseGradientBoosting, ClassifierMixin):
     Elements of Statistical Learning Ed. 2, Springer, 2009.
     """
 
-    supported_loss = ('deviance', 'mdeviance', 'bdeviance')
+    _SUPPORTED_LOSS = ('deviance', 'mdeviance', 'bdeviance')
 
     def __init__(self, loss='deviance', learning_rate=0.1, n_estimators=100,
                  subsample=1.0, min_samples_split=2, min_samples_leaf=1,
@@ -1095,7 +1099,7 @@ class GradientBoostingRegressor(BaseGradientBoosting, RegressorMixin):
     Elements of Statistical Learning Ed. 2, Springer, 2009.
     """
 
-    supported_loss = ('ls', 'lad', 'huber', 'quantile')
+    _SUPPORTED_LOSS = ('ls', 'lad', 'huber', 'quantile')
 
     def __init__(self, loss='ls', learning_rate=0.1, n_estimators=100,
                  subsample=1.0, min_samples_split=2, min_samples_leaf=1,

sklearn/ensemble/tests/test_gradient_boosting.py

@@ -103,18 +103,6 @@ def test_parameter_checks():
     assert_raises(ValueError,
                   lambda: GradientBoostingClassifier().feature_importances_)
 
-    # binomial deviance requires ``n_classes == 2``.
-    assert_raises(ValueError,
-                  lambda X, y: GradientBoostingClassifier(
-                      loss='bdeviance').fit(X, y),
-                  X, [0, 0, 1, 1, 2, 2])
-
-    # multinomial deviance requires ``n_classes > 2``.
-    assert_raises(ValueError,
-                  lambda X, y: GradientBoostingClassifier(
-                      loss='mdeviance').fit(X, y),
-                  X, [0, 0, 1, 1, 1, 0])
-
     # deviance requires ``n_classes >= 2``.
     assert_raises(ValueError,
                   lambda X, y: GradientBoostingClassifier(
@@ -133,10 +121,6 @@ def test_loss_function():
                   GradientBoostingClassifier(loss='huber').fit, X, y)
     assert_raises(ValueError,
                   GradientBoostingRegressor(loss='deviance').fit, X, y)
-    assert_raises(ValueError,
-                  GradientBoostingRegressor(loss='bdeviance').fit, X, y)
-    assert_raises(ValueError,
-                  GradientBoostingRegressor(loss='mdeviance').fit, X, y)
 
 
 def test_classification_synthetic():
@@ -596,3 +580,21 @@ def test_more_verbose_output():
     n_lines = sum(1 for l in verbose_output.readlines())
     # 100 lines for n_estimators==100
     assert_equal(100, n_lines)
+
+
+def test_warn_deviance():
+    """Test if mdeviance and bdeviance give deprecated warning. """
+    for loss in ('bdeviance', 'mdeviance'):
+        with warnings.catch_warnings(record=True) as w:
+            # This will raise a DataConversionWarning that we want to
+            # "always" raise, elsewhere the warnings gets ignored in the
+            # later tests, and the tests that check for this warning fail
+            warnings.simplefilter("always", DataConversionWarning)
+            clf = GradientBoostingClassifier(loss=loss)
+            try:
+                clf.fit(X, y)
+            except:
+                # mdeviance will raise ValueError because only 2 classes
+                pass
+            # deprecated warning for bdeviance and mdeviance
+            assert len(w) == 1