FIX fix bug when warm starting with early stopping in Hist GBDT (#15624)

NicolasHug · adrinjalali · commit a631424ae9cd · 2019-11-19T11:54:01.000+01:00
diff --git a/sklearn/ensemble/_hist_gradient_boosting/gradient_boosting.py b/sklearn/ensemble/_hist_gradient_boosting/gradient_boosting.py
@@ -270,6 +270,8 @@ def fit(self, X, y):
 
             # Compute raw predictions
             raw_predictions = self._raw_predict(X_binned_train)
+            if self.do_early_stopping_ and self._use_validation_data:
+                raw_predictions_val = self._raw_predict(X_binned_val)
 
             if self.do_early_stopping_ and self.scoring != 'loss':
                 # Compute the subsample set
diff --git a/sklearn/ensemble/_hist_gradient_boosting/tests/test_warm_start.py b/sklearn/ensemble/_hist_gradient_boosting/tests/test_warm_start.py
@@ -93,14 +93,15 @@ def test_warm_start_max_depth(GradientBoosting, X, y):
     (HistGradientBoostingClassifier, X_classification, y_classification),
     (HistGradientBoostingRegressor, X_regression, y_regression)
 ])
-def test_warm_start_early_stopping(GradientBoosting, X, y):
+@pytest.mark.parametrize('scoring', (None, 'loss'))
+def test_warm_start_early_stopping(GradientBoosting, X, y, scoring):
     # Make sure that early stopping occurs after a small number of iterations
     # when fitting a second time with warm starting.
 
     n_iter_no_change = 5
     gb = GradientBoosting(
         n_iter_no_change=n_iter_no_change, max_iter=10000,
-        random_state=42, warm_start=True, tol=1e-3
+        random_state=42, warm_start=True, tol=1e-3, scoring=scoring,
     )
     gb.fit(X, y)
     n_iter_first_fit = gb.n_iter_
