[python][sklearn] respect objective aliases

python-package/lightgbm/sklearn.py

@@ -578,21 +578,32 @@ def fit(self, X, y,
             feature_name='auto', categorical_feature='auto',
             callbacks=None, init_model=None):
         """Docstring is set after definition, using a template."""
+        params = self.get_params()
+
+        params.pop('objective', None)
+        for alias in _ConfigAliases.get('objective'):
+            if alias in params:
+                self._objective = params.pop(alias)
+                _log_warning(f"Found '{alias}' in params. Will use it instead of 'objective' argument")
         if self._objective is None:
             if isinstance(self, LGBMRegressor):
                 self._objective = "regression"
             elif isinstance(self, LGBMClassifier):
-                self._objective = "binary"
+                if self._n_classes > 2:
+                    self._objective = "multiclass"
+                else:
+                    self._objective = "binary"
             elif isinstance(self, LGBMRanker):
                 self._objective = "lambdarank"
             else:
                 raise ValueError("Unknown LGBMModel type.")
         if callable(self._objective):
             self._fobj = _ObjectiveFunctionWrapper(self._objective)
+            params['objective'] = 'None'  # objective = nullptr for unknown objective
         else:
             self._fobj = None
+            params['objective'] = self._objective
 
-        params = self.get_params()
         # user can set verbose with kwargs, it has higher priority
         if self.silent != "warn":
             _log_warning("'silent' argument is deprecated and will be removed in a future release of LightGBM. "
@@ -603,13 +614,13 @@ def fit(self, X, y,
         if not any(verbose_alias in params for verbose_alias in _ConfigAliases.get("verbosity")) and silent:
             params['verbose'] = -1
         params.pop('silent', None)
+
         params.pop('importance_type', None)
         params.pop('n_estimators', None)
         params.pop('class_weight', None)
+
         if isinstance(params['random_state'], np.random.RandomState):
             params['random_state'] = params['random_state'].randint(np.iinfo(np.int32).max)
-        for alias in _ConfigAliases.get('objective'):
-            params.pop(alias, None)
         if self._n_classes is not None and self._n_classes > 2:
             for alias in _ConfigAliases.get('num_class'):
                 params.pop(alias, None)
@@ -621,9 +632,6 @@ def fit(self, X, y,
                     _log_warning(f"Found '{alias}' in params. Will use it instead of 'eval_at' argument")
                     eval_at = params.pop(alias)
             params['eval_at'] = eval_at
-        params['objective'] = self._objective
-        if self._fobj:
-            params['objective'] = 'None'  # objective = nullptr for unknown objective
 
         # Do not modify original args in fit function
         # Refer to https://github.com/microsoft/LightGBM/pull/2619
@@ -930,12 +938,6 @@ def fit(self, X, y,
         self._classes = self._le.classes_
         self._n_classes = len(self._classes)
 
-        if self._n_classes > 2:
-            # Switch to using a multiclass objective in the underlying LGBM instance
-            ova_aliases = {"multiclassova", "multiclass_ova", "ova", "ovr"}
-            if self._objective not in ova_aliases and not callable(self._objective):
-                self._objective = "multiclass"
-
         if not callable(eval_metric):
             if isinstance(eval_metric, (str, type(None))):
                 eval_metric = [eval_metric]

tests/python_package_test/test_sklearn.py

@@ -157,13 +157,41 @@ def test_eval_at_aliases():
     X_test, y_test = load_svmlight_file(str(rank_example_dir / 'rank.test'))
     q_train = np.loadtxt(str(rank_example_dir / 'rank.train.query'))
     q_test = np.loadtxt(str(rank_example_dir / 'rank.test.query'))
-    for alias in ('eval_at', 'ndcg_eval_at', 'ndcg_at', 'map_eval_at', 'map_at'):
+    for alias in lgb.basic._ConfigAliases.get('eval_at'):
         gbm = lgb.LGBMRanker(n_estimators=5, **{alias: [1, 2, 3, 9]})
         with pytest.warns(UserWarning, match=f"Found '{alias}' in params. Will use it instead of 'eval_at' argument"):
             gbm.fit(X_train, y_train, group=q_train, eval_set=[(X_test, y_test)], eval_group=[q_test])
         assert list(gbm.evals_result_['valid_0'].keys()) == ['ndcg@1', 'ndcg@2', 'ndcg@3', 'ndcg@9']
 
 
+@pytest.mark.parametrize("custom_objective", [True, False])
+def test_objective_aliases(custom_objective):
+    X, y = load_boston(return_X_y=True)
+    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.1, random_state=42)
+    if custom_objective:
+        obj = custom_dummy_obj
+        metric_name = 'l2'  # default one
+    else:
+        obj = 'mape'
+        metric_name = 'mape'
+    evals = []
+    for alias in lgb.basic._ConfigAliases.get('objective'):
+        gbm = lgb.LGBMRegressor(n_estimators=5, **{alias: obj})
+        if alias != 'objective':
+            with pytest.warns(UserWarning, match=f"Found '{alias}' in params. Will use it instead of 'objective' argument"):
+                gbm.fit(X_train, y_train, eval_set=[(X_test, y_test)])
+        else:
+            gbm.fit(X_train, y_train, eval_set=[(X_test, y_test)])
+        assert list(gbm.evals_result_['valid_0'].keys()) == [metric_name]
+        evals.append(gbm.evals_result_['valid_0'][metric_name])
+    evals_t = np.array(evals).T
+    for i in range(evals_t.shape[0]):
+        np.testing.assert_allclose(evals_t[i], evals_t[i][0])
+    # check that really dummy objective was used and estimator didn't learn anything
+    if custom_objective:
+        np.testing.assert_allclose(evals_t, evals_t[0][0])
+
+
 def test_regression_with_custom_objective():
     X, y = load_boston(return_X_y=True)
     X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.1, random_state=42)
@@ -910,10 +938,8 @@ def test_metrics():
     assert 'multi_logloss' in gbm.evals_result_['training']
     assert 'multi_error' in gbm.evals_result_['training']
 
-    # invalid objective is replaced with default multiclass one
-    # and invalid binary metric is replaced with multiclass alternative
-    gbm = lgb.LGBMClassifier(objective='invalid_obj',
-                             **params).fit(eval_metric='binary_error', **params_fit)
+    # invalid binary metric is replaced with multiclass alternative
+    gbm = lgb.LGBMClassifier(**params).fit(eval_metric='binary_error', **params_fit)
     assert gbm.objective_ == 'multiclass'
     assert len(gbm.evals_result_['training']) == 2
     assert 'multi_logloss' in gbm.evals_result_['training']