Fixed GradientBoostingRegression model conversion failure with init=z…

…ero (#164)

* Fixed GradientBoostingRegression model conversion failure with init=zero

* Loop on fit n_estimators_ instead of passed n_estimators

* Added corresponding unit test

* Fix error in scikit < 0.21

skl2onnx/operator_converters/gradient_boosting.py

@@ -96,17 +96,22 @@ def convert_sklearn_gradient_boosting_regressor(scope, operator, container):
     attrs['name'] = scope.get_unique_operator_name(op_type)
     attrs['n_targets'] = 1
 
-    # constant_ was introduced in scikit-learn 0.21.
-    if hasattr(op.init_, 'constant_'):
-        cst = [float(x) for x in op.init_.constant_]
-    elif op.loss == 'ls':
-        cst = [op.init_.mean]
+    if op.init == 'zero':
+        cst = np.zeros((operator.inputs[0].type.shape[0], op.loss_.K))
     else:
-        cst = [op.init_.quantile]
+        # constant_ was introduced in scikit-learn 0.21.
+        if hasattr(op.init_, 'constant_'):
+            cst = [float(x) for x in op.init_.constant_]
+        elif op.loss == 'ls':
+            cst = [op.init_.mean]
+        else:
+            cst = [op.init_.quantile]
     attrs['base_values'] = [float(x) for x in cst]
 
     tree_weight = op.learning_rate
-    for i in range(op.n_estimators):
+    n_est = (op.n_estimators_ if hasattr(op, 'n_estimators_') else
+             op.n_estimators)
+    for i in range(n_est):
         tree = op.estimators_[i][0].tree_
         tree_id = i
         add_tree_to_attribute_pairs(attrs, False, tree, tree_id, tree_weight,

tests/test_sklearn_gradient_boosting_converters.py

@@ -163,6 +163,25 @@ def test_gradient_boosting_regressor_int(self):
                           " <= StrictVersion('0.2.1')"
         )
 
+    def test_gradient_boosting_regressor_zero_init(self):
+        model, X = fit_regression_model(
+            GradientBoostingRegressor(n_estimators=30, init="zero",
+                                      random_state=42))
+        model_onnx = convert_sklearn(
+            model,
+            "gradient boosting regression",
+            [("input", FloatTensorType([1, X.shape[1]]))],
+        )
+        self.assertIsNotNone(model_onnx)
+        dump_data_and_model(
+            X,
+            model,
+            model_onnx,
+            basename="SklearnGradientBoostingRegressionZeroInit-Dec4",
+            allow_failure="StrictVersion(onnxruntime.__version__)"
+                          " <= StrictVersion('0.2.1')"
+        )
+
 
 if __name__ == "__main__":
     unittest.main()