FIX bug in nested set_params usage

jnothman · jnothman · commit c890fd1c9871 · 2017-10-25T20:27:34.000+11:00
Issue where estimator is changed as well as its parameter: scikit-learn#9945 (comment)
diff --git a/sklearn/base.py b/sklearn/base.py
@@ -250,6 +250,7 @@ def set_params(self, **params):
             return self
         valid_params = self.get_params(deep=True)
 
+        changed = False
         nested_params = defaultdict(dict)  # grouped by prefix
         for key, value in params.items():
             key, delim, sub_key = key.partition('__')
@@ -262,8 +263,12 @@ def set_params(self, **params):
             if delim:
                 nested_params[key][sub_key] = value
             else:
+                changed = True
                 setattr(self, key, value)
 
+        if changed and nested_params:
+            # still need deep because Pipeline steps are deep
+            valid_params = self.get_params(deep=True)
         for key, sub_params in nested_params.items():
             valid_params[key].set_params(**sub_params)
 
diff --git a/sklearn/tests/test_base.py b/sklearn/tests/test_base.py
@@ -246,6 +246,13 @@ def set_params(self, **kwargs):
                        estimator__min_samples_leaf=2)
 
 
+def test_set_params_updates_valid_params():
+    # Check that set_params tries to set SVC().C, not
+    # DecisionTreeClassifier().C
+    pipe = GridSearchCV(DecisionTreeClassifier(), {})
+    pipe.set_params(estimator=SVC(), estimator__C=1.0)
+
+
 def test_score_sample_weight():
 
     rng = np.random.RandomState(0)
