Fix handling of global bias for binary:logitraw objective of XGBoost

src/frontend/xgboost.cc

@@ -407,7 +407,7 @@ inline std::unique_ptr<treelite::Model> ParseStream(dmlc::Stream* fi) {
   // 1.0 it's the original value provided by user.
   const bool need_transform_to_margin = mparam_.major_version >= 1;
   if (need_transform_to_margin) {
-    treelite::details::xgboost::TransformGlobalBiasToMargin(name_obj_, &model->param);
+    treelite::details::xgboost::TransformGlobalBiasToMargin(&model->param);
   }
 
   // traverse trees

src/frontend/xgboost/xgboost.h

@@ -33,7 +33,7 @@ extern const std::vector<std::string> exponential_objectives;
 void SetPredTransform(const std::string& objective_name, ModelParam* param);
 
 // Transform the global bias parameter from probability into margin score
-void TransformGlobalBiasToMargin(const std::string& objective_name, ModelParam* param);
+void TransformGlobalBiasToMargin(ModelParam* param);
 
 enum FeatureType {
   kNumerical = 0,

src/frontend/xgboost_json.cc

@@ -432,8 +432,7 @@ bool XGBoostModelHandler::EndObject(std::size_t memberCount) {
   // 1.0 it's the original value provided by user.
   const bool need_transform_to_margin = (version[0] >= 1);
   if (need_transform_to_margin) {
-    treelite::details::xgboost::TransformGlobalBiasToMargin(
-        output.objective_name, &output.model->param);
+    treelite::details::xgboost::TransformGlobalBiasToMargin(&output.model->param);
   }
   return pop_handler();
 }

src/frontend/xgboost_util.cc

@@ -54,15 +54,8 @@ void SetPredTransform(const std::string& objective_name, ModelParam* param) {
 }
 
 // Transform the global bias parameter from probability into margin score
-void TransformGlobalBiasToMargin(const std::string& objective_name, ModelParam* param) {
+void TransformGlobalBiasToMargin(ModelParam* param) {
   std::string bias_transform{param->pred_transform};
-  if (objective_name == "binary:logitraw") {
-    // Special handling for 'logitraw', where the global bias is transformed with 'sigmoid',
-    // but the prediction is returned un-transformed.
-    CHECK_EQ(bias_transform, "identity");
-    bias_transform = "sigmoid";
-  }
-
   if (bias_transform == "sigmoid") {
     param->global_bias = ProbToMargin::Sigmoid(param->global_bias);
   } else if (bias_transform == "exponential") {

tests/python/test_xgboost_integration.py

@@ -114,18 +114,18 @@ def test_xgb_iris(tmpdir, toolchain, objective, model_format, expected_pred_tran
     np.testing.assert_almost_equal(out_pred, expected_pred, decimal=5)
 
 
-@pytest.mark.parametrize('toolchain', os_compatible_toolchains())
 @pytest.mark.parametrize('model_format', ['binary', 'json'])
 @pytest.mark.parametrize('objective,max_label,expected_global_bias',
                          [('binary:logistic', 2, 0),
                           ('binary:hinge', 2, 0.5),
-                          ('binary:logitraw', 2, 0),
+                          ('binary:logitraw', 2, 0.5),
                           ('count:poisson', 4, math.log(0.5)),
                           ('rank:pairwise', 5, 0.5),
                           ('rank:ndcg', 5, 0.5),
                           ('rank:map', 5, 0.5)],
                          ids=['binary:logistic', 'binary:hinge', 'binary:logitraw',
                               'count:poisson', 'rank:pairwise', 'rank:ndcg', 'rank:map'])
+@pytest.mark.parametrize('toolchain', os_compatible_toolchains())
 def test_nonlinear_objective(tmpdir, objective, max_label, expected_global_bias, toolchain,
                              model_format):
     # pylint: disable=too-many-locals,too-many-arguments