Fix discrepencies when xgboost trees are empty. (#447)

* Fix discrepencies when xgboost trees are empty.

Signed-off-by: xavier dupré <xavier.dupre@gmail.com>

* fix wrong converting function

Signed-off-by: xavier dupré <xavier.dupre@gmail.com>

* eol

Signed-off-by: xavier dupré <xavier.dupre@gmail.com>

Co-authored-by: xavier dupré <xavier.dupre@gmail.com>

onnxmltools/convert/xgboost/operator_converters/XGBoost.py

@@ -242,6 +242,8 @@ def convert(scope, operator, container):
             # See https://github.com/dmlc/xgboost/blob/master/src/common/math.h#L23.
             attr_pairs['post_transform'] = "LOGISTIC"
             attr_pairs['class_ids'] = [0 for v in attr_pairs['class_treeids']]
+            if js_trees[0].get('leaf', None) == 0:
+                attr_pairs['base_values'] = [0.5]
         else:
             # See https://github.com/dmlc/xgboost/blob/master/src/common/math.h#L35.
             attr_pairs['post_transform'] = "SOFTMAX"

tests/xgboost/test_xgboost_converters.py

@@ -6,6 +6,7 @@
 import os
 import unittest
 import numpy as np
+from numpy.testing import assert_almost_equal
 import pandas
 from sklearn.datasets import (
     load_diabetes, load_iris, make_classification, load_digits)
@@ -325,6 +326,22 @@ def test_xgboost_example_mnist(self):
             allow_failure="StrictVersion(onnx.__version__) < StrictVersion('1.3.0')",
             basename="XGBoostExample")
 
+    def test_xgb_empty_tree(self):
+        xgb = XGBClassifier(n_estimators=2, max_depth=2)
+
+        # simple dataset
+        X = [[0, 1], [1, 1], [2, 0]]
+        X = np.array(X, dtype=np.float32)
+        y = [0, 1, 0]
+        xgb.fit(X, y)
+        conv_model = convert_xgboost(
+            xgb, initial_types=[
+                ('input', FloatTensorType(shape=[None, X.shape[1]]))])
+        sess = InferenceSession(conv_model.SerializeToString())
+        res = sess.run(None, {'input': X.astype(np.float32)})
+        assert_almost_equal(xgb.predict_proba(X), res[1])
+        assert_almost_equal(xgb.predict(X), res[0])
+
 
 if __name__ == "__main__":
     unittest.main()