added clean representation of objective

src/xgboost_distribution/model.py

@@ -99,6 +99,7 @@ def fit(
         self._distribution.check_target(y)
 
         params = self.get_xgb_params()
+        params["objective"] = None
         params["disable_default_eval_metric"] = True
         params["num_class"] = len(self._distribution.params)
 
@@ -136,7 +137,7 @@ def fit(
         evals_result = {}
         model, _, params = self._configure_fit(xgb_model, None, params)
 
-        # hack to suppress warnings from the extra distribution parameter
+        # Suppress warnings from unexpected distribution & natural_gradient params
         with config_context(verbosity=0):
             self._Booster = train(
                 params,

tests/test_model.py

@@ -41,6 +41,17 @@ def test_XGBDistribution_early_stopping_predict(small_train_test_data):
     np.testing.assert_array_equal(var, var_iter)
 
 
+def test_objective_overwrite_by_repeated_fit(small_X_y_data):
+    """In the fit step, we set the objective to distribution:normal (e.g.)
+
+    This needs to be re-set at each fit, as it's not a standard xgboost objective.
+    """
+    X, y = small_X_y_data
+    model = XGBDistribution()
+    model.fit(X, y)
+    model.fit(X, y)
+
+
 def test_distribution_set_param(small_X_y_data):
     """Check that updating the distribution params works"""
     X, y = small_X_y_data