[python-package] ignore training set on early stopping callback (fixes #5354) (#5412)

* ignore training set on early stopping callback

* fixes

* lint

* Apply suggestions from code review
Co-authored-by: James Lamb <jaylamb20@gmail.com>

* trigger ci
Co-authored-by: James Lamb <jaylamb20@gmail.com>

python-package/lightgbm/callback.py

@@ -258,11 +258,24 @@ class _EarlyStoppingCallback:
     def _lt_delta(self, curr_score: float, best_score: float, delta: float) -> bool:
         return curr_score < best_score - delta
 
+    def _is_train_set(self, ds_name: str, eval_name: str, train_name: str) -> bool:
+        return (ds_name == "cv_agg" and eval_name == "train") or ds_name == train_name
+
     def _init(self, env: CallbackEnv) -> None:
-        self.enabled = not any(env.params.get(boost_alias, "") == 'dart' for boost_alias
-                               in _ConfigAliases.get("boosting"))
+        is_dart = any(env.params.get(alias, "") == 'dart' for alias in _ConfigAliases.get("boosting"))
+        only_train_set = (
+            len(env.evaluation_result_list) == 1
+            and self._is_train_set(
+                ds_name=env.evaluation_result_list[0][0],
+                eval_name=env.evaluation_result_list[0][1].split(" ")[0],
+                train_name=env.model._train_data_name)
+        )
+        self.enabled = not is_dart and not only_train_set
         if not self.enabled:
-            _log_warning('Early stopping is not available in dart mode')
+            if is_dart:
+                _log_warning('Early stopping is not available in dart mode')
+            elif only_train_set:
+                _log_warning('Only training set found, disabling early stopping.')
             return
         if not env.evaluation_result_list:
             raise ValueError('For early stopping, '
@@ -339,9 +352,7 @@ class _EarlyStoppingCallback:
             eval_name_splitted = env.evaluation_result_list[i][1].split(" ")
             if self.first_metric_only and self.first_metric != eval_name_splitted[-1]:
                 continue  # use only the first metric for early stopping
-            if ((env.evaluation_result_list[i][0] == "cv_agg" and eval_name_splitted[0] == "train"
-                    or env.evaluation_result_list[i][0] == env.model._train_data_name)):
-                self._final_iteration_check(env, eval_name_splitted, i)
+            if self._is_train_set(env.evaluation_result_list[i][0], eval_name_splitted[0], env.model._train_data_name):
                 continue  # train data for lgb.cv or sklearn wrapper (underlying lgb.train)
             elif env.iteration - self.best_iter[i] >= self.stopping_rounds:
                 if self.verbose:

tests/python_package_test/test_engine.py

@@ -765,6 +765,43 @@ def test_early_stopping():
     assert 'binary_logloss' in gbm.best_score[valid_set_name]
 
 
+@pytest.mark.parametrize('use_valid', [True, False])
+def test_early_stopping_ignores_training_set(use_valid):
+    x = np.linspace(-1, 1, 100)
+    X = x.reshape(-1, 1)
+    y = x**2
+    X_train, X_valid = X[:80], X[80:]
+    y_train, y_valid = y[:80], y[80:]
+    train_ds = lgb.Dataset(X_train, y_train)
+    valid_ds = lgb.Dataset(X_valid, y_valid)
+    valid_sets = [train_ds]
+    valid_names = ['train']
+    if use_valid:
+        valid_sets.append(valid_ds)
+        valid_names.append('valid')
+    eval_result = {}
+
+    def train_fn():
+        return lgb.train(
+            {'num_leaves': 5},
+            train_ds,
+            num_boost_round=2,
+            valid_sets=valid_sets,
+            valid_names=valid_names,
+            callbacks=[lgb.early_stopping(1), lgb.record_evaluation(eval_result)]
+        )
+    if use_valid:
+        bst = train_fn()
+        assert bst.best_iteration == 1
+        assert eval_result['train']['l2'][1] < eval_result['train']['l2'][0]  # train improved
+        assert eval_result['valid']['l2'][1] > eval_result['valid']['l2'][0]  # valid didn't
+    else:
+        with pytest.warns(UserWarning, match='Only training set found, disabling early stopping.'):
+            bst = train_fn()
+        assert bst.current_iteration() == 2
+        assert bst.best_iteration == 0
+
+
 @pytest.mark.parametrize('first_metric_only', [True, False])
 def test_early_stopping_via_global_params(first_metric_only):
     X, y = load_breast_cancer(return_X_y=True)

tests/python_package_test/test_sklearn.py

@@ -1124,11 +1124,6 @@ def test_first_metric_only():
     iter_min = min([iter_min_l1, iter_min_l2])
     iter_min_valid1 = min([iter_valid1_l1, iter_valid1_l2])
 
-    # training data as eval_set
-    params_fit['eval_set'] = (X_train, y_train)
-    fit_and_check(['training'], ['l2'], 30, False)
-    fit_and_check(['training'], ['l2'], 30, True)
-
     # feval
     params['metric'] = 'None'
     params_fit['eval_metric'] = lambda preds, train_data: [decreasing_metric(preds, train_data),

tests/python_package_test/test_utilities.py

@@ -29,17 +29,18 @@ def test_register_logger(tmp_path):
                   [1, 2, 3]],
                  dtype=np.float32)
     y = np.array([0, 1, 1, 0])
-    lgb_data = lgb.Dataset(X, y)
+    lgb_train = lgb.Dataset(X, y)
+    lgb_valid = lgb.Dataset(X, y)  # different object for early-stopping
 
     eval_records = {}
     callbacks = [
         lgb.record_evaluation(eval_records),
         lgb.log_evaluation(2),
-        lgb.early_stopping(4)
+        lgb.early_stopping(10)
     ]
     lgb.train({'objective': 'binary', 'metric': ['auc', 'binary_error']},
-              lgb_data, num_boost_round=10, feval=dummy_metric,
-              valid_sets=[lgb_data], categorical_feature=[1], callbacks=callbacks)
+              lgb_train, num_boost_round=10, feval=dummy_metric,
+              valid_sets=[lgb_valid], categorical_feature=[1], callbacks=callbacks)
 
     lgb.plot_metric(eval_records)
 
@@ -51,32 +52,32 @@ INFO | [LightGBM] [Info] Number of data points in the train set: 4, number of us
 INFO | [LightGBM] [Info] [binary:BoostFromScore]: pavg=0.500000 -> initscore=0.000000
 INFO | [LightGBM] [Warning] Stopped training because there are no more leaves that meet the split requirements
 DEBUG | In dummy_metric
-INFO | Training until validation scores don't improve for 4 rounds
+INFO | Training until validation scores don't improve for 10 rounds
 INFO | [LightGBM] [Warning] Stopped training because there are no more leaves that meet the split requirements
 DEBUG | In dummy_metric
-INFO | [2]	training's auc: 0.5	training's binary_error: 0.5	training's dummy_metric: 1
+INFO | [2]	valid_0's auc: 0.5	valid_0's binary_error: 0.5	valid_0's dummy_metric: 1
 INFO | [LightGBM] [Warning] Stopped training because there are no more leaves that meet the split requirements
 DEBUG | In dummy_metric
 INFO | [LightGBM] [Warning] Stopped training because there are no more leaves that meet the split requirements
 DEBUG | In dummy_metric
-INFO | [4]	training's auc: 0.5	training's binary_error: 0.5	training's dummy_metric: 1
+INFO | [4]	valid_0's auc: 0.5	valid_0's binary_error: 0.5	valid_0's dummy_metric: 1
 INFO | [LightGBM] [Warning] Stopped training because there are no more leaves that meet the split requirements
 DEBUG | In dummy_metric
 INFO | [LightGBM] [Warning] Stopped training because there are no more leaves that meet the split requirements
 DEBUG | In dummy_metric
-INFO | [6]	training's auc: 0.5	training's binary_error: 0.5	training's dummy_metric: 1
+INFO | [6]	valid_0's auc: 0.5	valid_0's binary_error: 0.5	valid_0's dummy_metric: 1
 INFO | [LightGBM] [Warning] Stopped training because there are no more leaves that meet the split requirements
 DEBUG | In dummy_metric
 INFO | [LightGBM] [Warning] Stopped training because there are no more leaves that meet the split requirements
 DEBUG | In dummy_metric
-INFO | [8]	training's auc: 0.5	training's binary_error: 0.5	training's dummy_metric: 1
+INFO | [8]	valid_0's auc: 0.5	valid_0's binary_error: 0.5	valid_0's dummy_metric: 1
 INFO | [LightGBM] [Warning] Stopped training because there are no more leaves that meet the split requirements
 DEBUG | In dummy_metric
 INFO | [LightGBM] [Warning] Stopped training because there are no more leaves that meet the split requirements
 DEBUG | In dummy_metric
-INFO | [10]	training's auc: 0.5	training's binary_error: 0.5	training's dummy_metric: 1
+INFO | [10]	valid_0's auc: 0.5	valid_0's binary_error: 0.5	valid_0's dummy_metric: 1
 INFO | Did not meet early stopping. Best iteration is:
-[1]	training's auc: 0.5	training's binary_error: 0.5	training's dummy_metric: 1
+[1]	valid_0's auc: 0.5	valid_0's binary_error: 0.5	valid_0's dummy_metric: 1
 WARNING | More than one metric available, picking one to plot.
 """.strip()