Fix application of best_ntree_limit to the entire list of estimators. Instead the limit is applied to per-class estimators split (#83)

Author: izeigerman

m2cgen/assemblers/boosting.py

@@ -9,14 +9,17 @@ class BaseBoostingAssembler(ModelAssembler):
 
     classifier_name = None
 
-    def __init__(self, model, trees, base_score=0):
+    def __init__(self, model, trees, base_score=0, tree_limit=None):
         super().__init__(model)
         self.all_trees = trees
         self._base_score = base_score
 
         self._output_size = 1
         self._is_classification = False
 
+        assert tree_limit is None or tree_limit > 0, "Unexpected tree limit"
+        self._tree_limit = tree_limit
+
         model_class_name = type(model).__name__
         if model_class_name == self.classifier_name:
             self._is_classification = True
@@ -34,6 +37,9 @@ def assemble(self):
                 self.all_trees, self._base_score)
 
     def _assemble_single_output(self, trees, base_score=0):
+        if self._tree_limit:
+            trees = trees[:self._tree_limit]
+
         trees_ast = [self._assemble_tree(t) for t in trees]
         result_ast = utils.apply_op_to_expressions(
             ast.BinNumOpType.ADD,
@@ -83,16 +89,14 @@ def __init__(self, model):
         }
 
         model_dump = model.get_booster().get_dump(dump_format="json")
-
-        # Respect XGBoost ntree_limit
-        ntree_limit = getattr(model, "best_ntree_limit", 0)
-
-        if ntree_limit > 0:
-            model_dump = model_dump[:ntree_limit]
-
         trees = [json.loads(d) for d in model_dump]
 
-        super().__init__(model, trees, base_score=model.base_score)
+        # Limit the number of trees that should be used for
+        # assembling (if applicable).
+        best_ntree_limit = getattr(model, "best_ntree_limit", None)
+
+        super().__init__(model, trees, base_score=model.base_score,
+                         tree_limit=best_ntree_limit)
 
     def _assemble_tree(self, tree):
         if "leaf" in tree:

tests/assemblers/test_xgboost.py

@@ -147,3 +147,84 @@ def test_regression_best_ntree_limit():
             ast.BinNumOpType.ADD))
 
     assert utils.cmp_exprs(actual, expected)
+
+
+def test_multi_class_best_ntree_limit():
+    base_score = 0.5
+    estimator = xgboost.XGBClassifier(n_estimators=100, random_state=1,
+                                      max_depth=1, base_score=base_score)
+
+    estimator.best_ntree_limit = 1
+
+    utils.train_model_classification(estimator)
+
+    assembler = assemblers.XGBoostModelAssembler(estimator)
+    actual = assembler.assemble()
+
+    estimator_exp_class1 = ast.ExpExpr(
+        ast.SubroutineExpr(
+            ast.BinNumExpr(
+                ast.NumVal(0.5),
+                ast.IfExpr(
+                    ast.CompExpr(
+                        ast.FeatureRef(2),
+                        ast.NumVal(2.5999999),
+                        ast.CompOpType.GTE),
+                    ast.NumVal(-0.0731707439),
+                    ast.NumVal(0.142857149)),
+                ast.BinNumOpType.ADD)),
+        to_reuse=True)
+
+    estimator_exp_class2 = ast.ExpExpr(
+        ast.SubroutineExpr(
+            ast.BinNumExpr(
+                ast.NumVal(0.5),
+                ast.IfExpr(
+                    ast.CompExpr(
+                        ast.FeatureRef(2),
+                        ast.NumVal(2.5999999),
+                        ast.CompOpType.GTE),
+                    ast.NumVal(0.0341463387),
+                    ast.NumVal(-0.0714285821)),
+                ast.BinNumOpType.ADD)),
+        to_reuse=True)
+
+    estimator_exp_class3 = ast.ExpExpr(
+        ast.SubroutineExpr(
+            ast.BinNumExpr(
+                ast.NumVal(0.5),
+                ast.IfExpr(
+                    ast.CompExpr(
+                        ast.FeatureRef(2),
+                        ast.NumVal(4.85000038),
+                        ast.CompOpType.GTE),
+                    ast.NumVal(0.129441619),
+                    ast.NumVal(-0.0681440532)),
+                ast.BinNumOpType.ADD)),
+        to_reuse=True)
+
+    exp_sum = ast.BinNumExpr(
+        ast.BinNumExpr(
+            estimator_exp_class1,
+            estimator_exp_class2,
+            ast.BinNumOpType.ADD),
+        estimator_exp_class3,
+        ast.BinNumOpType.ADD,
+        to_reuse=True)
+
+    expected = ast.VectorVal([
+        ast.BinNumExpr(
+            estimator_exp_class1,
+            exp_sum,
+            ast.BinNumOpType.DIV),
+        ast.BinNumExpr(
+            estimator_exp_class2,
+            exp_sum,
+            ast.BinNumOpType.DIV),
+        ast.BinNumExpr(
+            estimator_exp_class3,
+            exp_sum,
+            ast.BinNumOpType.DIV)
+    ])
+
+    assert utils.cmp_exprs(actual, expected)