Update XGBoost max supported version to 3.0.2

coremltools/_deps/__init__.py

@@ -94,7 +94,7 @@ def __get_sklearn_version(version):
 
 # ---------------------------------------------------------------------------------------
 _HAS_XGBOOST = True
-_XGBOOST_MAX_VERSION = "1.4.2"
+_XGBOOST_MAX_VERSION = "3.0.2"
 try:
     import xgboost
     _warn_if_above_max_supported_version("XGBoost", xgboost.__version__, _XGBOOST_MAX_VERSION)

coremltools/converters/xgboost/_tree_ensemble.py

@@ -159,6 +159,7 @@ def convert_tree_ensemble(
     import os
 
     feature_map = None
+    base_score = None
     if isinstance(
         model, (_xgboost.core.Booster, _xgboost.XGBRegressor, _xgboost.XGBClassifier)
     ):
@@ -199,6 +200,8 @@ def convert_tree_ensemble(
                 model = model.get_booster()
             else:
                 model = model.booster()
+        config = json.loads(model.save_config())
+        base_score = float(config['learner']['learner_model_param']['base_score'])
 
         # Xgboost sometimes has feature names in there. Sometimes does not.
         if (feature_names is None) and (model.feature_names is None):
@@ -212,6 +215,8 @@ def convert_tree_ensemble(
             # When XGboost model artifact does not have feature names
             # (seems to be the default in new Xgboost releases),
             # but the user provides them, use them as they are expecting later.
+            if not isinstance(feature_names, list):
+                feature_names = list(feature_names)
             model.feature_names=feature_names
 
         xgb_model_str = model.get_dump(with_stats=True, dump_format="json")
@@ -247,9 +252,10 @@ def convert_tree_ensemble(
             class_labels = range(n_classes)
         if n_classes == 2:
             # if we have only 2 classes we only have one sequence of estimators
-            base_prediction = [0.0]
+            base_prediction = [base_score if base_score is not None else 0.0]
         else:
-            base_prediction = [0.0 for c in range(n_classes)]
+            bs = base_score if base_score is not None else 0.0
+            base_prediction = [bs for c in range(n_classes)]
         # target here is the equivalent of output_features in scikit learn
         mlkit_tree = TreeEnsembleClassifier(feature_names, class_labels, target)
         mlkit_tree.set_default_prediction_value(base_prediction)
@@ -259,7 +265,7 @@ def convert_tree_ensemble(
             mlkit_tree.set_post_evaluation_transform("Classification_SoftMax")
     else:
         mlkit_tree = _TreeEnsembleRegressor(feature_names, target)
-        mlkit_tree.set_default_prediction_value(0.5)
+        mlkit_tree.set_default_prediction_value(base_score if base_score is not None else 0.5)
 
     for xgb_tree_id, xgb_tree_str in enumerate(xgb_model_str):
         if mode == "classifier" and n_classes > 2:

coremltools/test/xgboost_tests/test_boosted_trees_regression_numeric.py

@@ -307,3 +307,42 @@ def test_boston_housing_parameter_stress_test(self):
         print("Testing a total of %s cases. This could take a while" % len(args))
         for it, arg in enumerate(args):
             self._train_convert_evaluate_assert(arg)
+
+@unittest.skipIf(not _HAS_XGBOOST, "Missing xgboost. Skipping")
+class XGBoostBaseScoreTest(unittest.TestCase):
+    @classmethod
+    def setUpClass(self):
+        scikit_data = load_boston()
+        self.X = scikit_data["data"]
+        self.target = scikit_data["target"]
+        self.feature_names = scikit_data["feature_names"]
+        self.output_name = "target"        
+
+    def test_base_score_extraction(self):
+        """Verify that the converter reads base_score from the model config
+        instead of using the hardcoded default (0.5)."""
+        import json
+
+        # Train a model - XGBoost 3.x auto-estimates base_score from data
+        xgb_model = xgboost.XGBRegressor(n_estimators=5)
+        xgb_model.fit(self.X, self.target)
+
+        # Extract the actual base_score from the booster config
+        booster = xgb_model.get_booster()
+        config = json.loads(booster.save_config())
+        actual_base_score = float(
+            config['learner']['learner_model_param']['base_score']
+        )
+
+        # Convert and check the default_prediction_value matches
+        spec = xgb_converter.convert(
+            xgb_model, self.feature_names, self.output_name, force_32bit_float=False
+        )
+        tree_ensemble = spec.get_spec().treeEnsembleRegressor.treeEnsemble
+        default_value = tree_ensemble.basePredictionValue[0]
+        self.assertAlmostEqual(
+            default_value,
+            actual_base_score,
+            places=5,
+            msg="Converter should use model's actual base_score, not hardcoded 0.5",
+        )

reqs/test.pip

@@ -11,7 +11,7 @@ scipy==1.8.1; python_version >= '3.8' and python_version < '3.11' and platform_m
 scipy==1.9.2; python_version == '3.11' and platform_machine != "arm64"
 scipy; platform_machine == "arm64"
 
-xgboost==1.4.2; platform_machine != "arm64"
+xgboost==3.0.2; platform_machine != "arm64"
 
 # coremltools.optimize.torch
 filelock==3.6.0