Fixes early stopping with XGBoost 2.0 (#597)

* add one unit test to investigate a bug Signed-off-by: xadupre <xadupre@microsoft.com> * add one more test Signed-off-by: xadupre <xadupre@microsoft.com> * remove unnecessary print Signed-off-by: xadupre <xadupre@microsoft.com> * update CI Signed-off-by: Xavier Dupre <xadupre@microsoft.com> * ci Signed-off-by: Xavier Dupre <xadupre@microsoft.com> * remove removed files Signed-off-by: Xavier Dupre <xadupre@microsoft.com> * update test Signed-off-by: Xavier Dupre <xadupre@microsoft.com> * fix early stopping Signed-off-by: Xavier Dupre <xadupre@microsoft.com> * fix rf models Signed-off-by: Xavier Dupre <xadupre@microsoft.com> * remaining merge issue Signed-off-by: Xavier Dupre <xadupre@microsoft.com> --------- Signed-off-by: xadupre <xadupre@microsoft.com> Signed-off-by: Xavier Dupre <xadupre@microsoft.com>
onnx · Dec 16, 2023 · 180e733 · 180e733
1 parent 7858f9f
commit 180e733
Show file tree

Hide file tree

Showing 8 changed files with 517 additions and 17 deletions.
diff --git a/.azure-pipelines/linux-conda-CI.yml b/.azure-pipelines/linux-conda-CI.yml
@@ -25,13 +25,13 @@ jobs:
         numpy.version: ''
         scipy.version: ''
 
-      Python311-1150-RT1160-xgb175-lgbm40:
+      Python311-1150-RT1163-xgb175-lgbm40:
         python.version: '3.11'
         ONNX_PATH: 'onnx==1.15.0'
-        ONNXRT_PATH: 'onnxruntime==1.16.2'
+        ONNXRT_PATH: 'onnxruntime==1.16.3'
         COREML_PATH: NONE
         lightgbm.version: '>=4.0'
-        xgboost.version: '==1.7.5'
+        xgboost.version: '>=1.7.5,<2'
         numpy.version: ''
         scipy.version: ''
 
@@ -41,7 +41,7 @@ jobs:
         ONNXRT_PATH: 'onnxruntime==1.16.2'
         COREML_PATH: NONE
         lightgbm.version: '>=4.0'
-        xgboost.version: '==1.7.5'
+        xgboost.version: '>=1.7.5,<2'
         numpy.version: ''
         scipy.version: ''
 
@@ -51,7 +51,7 @@ jobs:
         ONNXRT_PATH: 'onnxruntime==1.15.1'
         COREML_PATH: NONE
         lightgbm.version: '<4.0'
-        xgboost.version: '==1.7.5'
+        xgboost.version: '>=1.7.5,<2'
         numpy.version: ''
         scipy.version: ''
 
@@ -61,7 +61,7 @@ jobs:
         ONNXRT_PATH: 'onnxruntime==1.14.0'
         COREML_PATH: NONE
         lightgbm.version: '<4.0'
-        xgboost.version: '==1.7.5'
+        xgboost.version: '>=1.7.5,<2'
         numpy.version: ''
         scipy.version: ''
 
@@ -71,7 +71,7 @@ jobs:
         ONNXRT_PATH: 'onnxruntime==1.15.1'
         COREML_PATH: NONE
         lightgbm.version: '>=4.0'
-        xgboost.version: '==1.7.5'
+        xgboost.version: '>=1.7.5,<2'
         numpy.version: ''
         scipy.version: '==1.8.0'
 

diff --git a/.azure-pipelines/win32-conda-CI.yml b/.azure-pipelines/win32-conda-CI.yml
@@ -15,6 +15,14 @@ jobs:
   strategy:
     matrix:
 
+      Python311-1150-RT1163:
+        python.version: '3.11'
+        ONNX_PATH: 'onnx==1.15.0'
+        ONNXRT_PATH: 'onnxruntime==1.16.3'
+        COREML_PATH: NONE
+        numpy.version: ''
+        xgboost.version: '2.0.2'
+
       Python311-1150-RT1162:
         python.version: '3.11'
         ONNX_PATH: 'onnx==1.15.0'

diff --git a/CHANGELOGS.md b/CHANGELOGS.md
@@ -2,6 +2,8 @@
 
 ## 1.12.0
 
+* Fix early stopping for XGBClassifier and xgboost > 2
+  [#597](https://github.com/onnx/onnxmltools/pull/597)
 * Fix discrepancies with XGBRegressor and xgboost > 2
   [#670](https://github.com/onnx/onnxmltools/pull/670)
 * Support count:poisson for XGBRegressor

diff --git a/onnxmltools/convert/xgboost/common.py b/onnxmltools/convert/xgboost/common.py
@@ -32,6 +32,14 @@ def get_xgb_params(xgb_node):
         bs = float(config["learner"]["learner_model_param"]["base_score"])
         # xgboost >= 2.0
         params["base_score"] = bs
+
+    bst = xgb_node.get_booster()
+    if hasattr(bst, "best_ntree_limit"):
+        params["best_ntree_limit"] = bst.best_ntree_limit
+    if "gradient_booster" in config["learner"]:
+        gbp = config["learner"]["gradient_booster"]["gbtree_model_param"]
+        if "num_trees" in gbp:
+            params["best_ntree_limit"] = int(gbp["num_trees"])
     return params
 
 

diff --git a/onnxmltools/convert/xgboost/operator_converters/XGBoost.py b/onnxmltools/convert/xgboost/operator_converters/XGBoost.py
@@ -40,14 +40,20 @@ def common_members(xgb_node, inputs):
         params = XGBConverter.get_xgb_params(xgb_node)
         objective = params["objective"]
         base_score = params["base_score"]
+        if hasattr(xgb_node, "best_ntree_limit"):
+            best_ntree_limit = xgb_node.best_ntree_limit
+        elif hasattr(xgb_node, "best_iteration"):
+            best_ntree_limit = xgb_node.best_iteration + 1
+        else:
+            best_ntree_limit = params.get("best_ntree_limit", None)
         if base_score is None:
             base_score = 0.5
         booster = xgb_node.get_booster()
         # The json format was available in October 2017.
         # XGBoost 0.7 was the first version released with it.
         js_tree_list = booster.get_dump(with_stats=True, dump_format="json")
         js_trees = [json.loads(s) for s in js_tree_list]
-        return objective, base_score, js_trees
+        return objective, base_score, js_trees, best_ntree_limit
 
     @staticmethod
     def _get_default_tree_attribute_pairs(is_classifier):
@@ -231,17 +237,17 @@ def _get_default_tree_attribute_pairs():
     def convert(scope, operator, container):
         xgb_node = operator.raw_operator
         inputs = operator.inputs
-        objective, base_score, js_trees = XGBConverter.common_members(xgb_node, inputs)
+        objective, base_score, js_trees, best_ntree_limit = XGBConverter.common_members(
+            xgb_node, inputs
+        )
 
         if objective in ["reg:gamma", "reg:tweedie"]:
             raise RuntimeError("Objective '{}' not supported.".format(objective))
 
         attr_pairs = XGBRegressorConverter._get_default_tree_attribute_pairs()
         attr_pairs["base_values"] = [base_score]
 
-        bst = xgb_node.get_booster()
-        best_ntree_limit = getattr(bst, "best_ntree_limit", len(js_trees))
-        if best_ntree_limit < len(js_trees):
+        if best_ntree_limit and best_ntree_limit < len(js_trees):
             js_trees = js_trees[:best_ntree_limit]
 
         XGBConverter.fill_tree_attributes(
@@ -289,7 +295,9 @@ def convert(scope, operator, container):
         xgb_node = operator.raw_operator
         inputs = operator.inputs
 
-        objective, base_score, js_trees = XGBConverter.common_members(xgb_node, inputs)
+        objective, base_score, js_trees, best_ntree_limit = XGBConverter.common_members(
+            xgb_node, inputs
+        )
 
         params = XGBConverter.get_xgb_params(xgb_node)
         n_estimators = get_n_estimators_classifier(xgb_node, params, js_trees)
@@ -305,8 +313,9 @@ def convert(scope, operator, container):
         else:
             ncl = (max(attr_pairs["class_treeids"]) + 1) // n_estimators
 
-        bst = xgb_node.get_booster()
-        best_ntree_limit = getattr(bst, "best_ntree_limit", len(js_trees)) * ncl
+        best_ntree_limit = best_ntree_limit or len(js_trees)
+        if ncl > 0:
+            best_ntree_limit *= ncl
         if 0 < best_ntree_limit < len(js_trees):
             js_trees = js_trees[:best_ntree_limit]
             attr_pairs = XGBClassifierConverter._get_default_tree_attribute_pairs()