[python-package] early stopping min_delta (fixes #2526) (#4580)

* initial changes

* initial version

* better handling of cases

* warn only with positive threshold

* remove early_stopping_threshold from high-level functions

* remove remaining early_stopping_threshold

* update test to use callback

* better handling of cases

* rename threshold to min_delta

enhance parameter description

update tests

* Apply suggestions from code review

Co-authored-by: Nikita Titov <nekit94-08@mail.ru>

* reduce num_boost_round in tests

* Apply suggestions from code review

Co-authored-by: Nikita Titov <nekit94-08@mail.ru>

* trigger ci

Co-authored-by: Nikita Titov <nekit94-08@mail.ru>
Co-authored-by: Nikita Titov <nekit94-12@hotmail.com>

python-package/lightgbm/callback.py

@@ -1,12 +1,20 @@
 # coding: utf-8
 """Callbacks library."""
 import collections
-from operator import gt, lt
+from functools import partial
 from typing import Any, Callable, Dict, List, Union
 
 from .basic import _ConfigAliases, _log_info, _log_warning
 
 
+def _gt_delta(curr_score: float, best_score: float, delta: float) -> bool:
+    return curr_score > best_score + delta
+
+
+def _lt_delta(curr_score: float, best_score: float, delta: float) -> bool:
+    return curr_score < best_score - delta
+
+
 class EarlyStopException(Exception):
     """Exception of early stopping."""
 
@@ -181,11 +189,11 @@ def reset_parameter(**kwargs: Union[list, Callable]) -> Callable:
     return _callback
 
 
-def early_stopping(stopping_rounds: int, first_metric_only: bool = False, verbose: bool = True) -> Callable:
+def early_stopping(stopping_rounds: int, first_metric_only: bool = False, verbose: bool = True, min_delta: Union[float, List[float]] = 0.0) -> Callable:
     """Create a callback that activates early stopping.
 
     Activates early stopping.
-    The model will train until the validation score stops improving.
+    The model will train until the validation score doesn't improve by at least ``min_delta``.
     Validation score needs to improve at least every ``stopping_rounds`` round(s)
     to continue training.
     Requires at least one validation data and one metric.
@@ -203,6 +211,10 @@ def early_stopping(stopping_rounds: int, first_metric_only: bool = False, verbos
         Whether to log message with early stopping information.
         By default, standard output resource is used.
         Use ``register_logger()`` function to register a custom logger.
+    min_delta : float or list of float, optional (default=0.0)
+        Minimum improvement in score to keep training.
+        If float, this single value is used for all metrics.
+        If list, its length should match the total number of metrics.
 
     Returns
     -------
@@ -229,17 +241,43 @@ def early_stopping(stopping_rounds: int, first_metric_only: bool = False, verbos
         if verbose:
             _log_info(f"Training until validation scores don't improve for {stopping_rounds} rounds")
 
+        n_metrics = len(set(m[1] for m in env.evaluation_result_list))
+        n_datasets = len(env.evaluation_result_list) // n_metrics
+        if isinstance(min_delta, list):
+            if not all(t >= 0 for t in min_delta):
+                raise ValueError('Values for early stopping min_delta must be non-negative.')
+            if len(min_delta) == 0:
+                if verbose:
+                    _log_info('Disabling min_delta for early stopping.')
+                deltas = [0.0] * n_datasets * n_metrics
+            elif len(min_delta) == 1:
+                if verbose:
+                    _log_info(f'Using {min_delta[0]} as min_delta for all metrics.')
+                deltas = min_delta * n_datasets * n_metrics
+            else:
+                if len(min_delta) != n_metrics:
+                    raise ValueError('Must provide a single value for min_delta or as many as metrics.')
+                if first_metric_only and verbose:
+                    _log_info(f'Using only {min_delta[0]} as early stopping min_delta.')
+                deltas = min_delta * n_datasets
+        else:
+            if min_delta < 0:
+                raise ValueError('Early stopping min_delta must be non-negative.')
+            if min_delta > 0 and n_metrics > 1 and not first_metric_only and verbose:
+                _log_info(f'Using {min_delta} as min_delta for all metrics.')
+            deltas = [min_delta] * n_datasets * n_metrics
+
         # split is needed for "<dataset type> <metric>" case (e.g. "train l1")
         first_metric[0] = env.evaluation_result_list[0][1].split(" ")[-1]
-        for eval_ret in env.evaluation_result_list:
+        for eval_ret, delta in zip(env.evaluation_result_list, deltas):
             best_iter.append(0)
             best_score_list.append(None)
-            if eval_ret[3]:
+            if eval_ret[3]:  # greater is better
                 best_score.append(float('-inf'))
-                cmp_op.append(gt)
+                cmp_op.append(partial(_gt_delta, delta=delta))
             else:
                 best_score.append(float('inf'))
-                cmp_op.append(lt)
+                cmp_op.append(partial(_lt_delta, delta=delta))
 
     def _final_iteration_check(env: CallbackEnv, eval_name_splitted: List[str], i: int) -> None:
         if env.iteration == env.end_iteration - 1:

tests/python_package_test/test_engine.py

@@ -643,6 +643,81 @@ def test_early_stopping():
     assert 'binary_logloss' in gbm.best_score[valid_set_name]
 
 
+@pytest.mark.parametrize('first_only', [True, False])
+@pytest.mark.parametrize('single_metric', [True, False])
+@pytest.mark.parametrize('greater_is_better', [True, False])
+def test_early_stopping_min_delta(first_only, single_metric, greater_is_better):
+    if single_metric and not first_only:
+        pytest.skip("first_metric_only doesn't affect single metric.")
+    metric2min_delta = {
+        'auc': 0.001,
+        'binary_logloss': 0.01,
+        'average_precision': 0.001,
+        'mape': 0.01,
+    }
+    if single_metric:
+        if greater_is_better:
+            metric = 'auc'
+        else:
+            metric = 'binary_logloss'
+    else:
+        if first_only:
+            if greater_is_better:
+                metric = ['auc', 'binary_logloss']
+            else:
+                metric = ['binary_logloss', 'auc']
+        else:
+            if greater_is_better:
+                metric = ['auc', 'average_precision']
+            else:
+                metric = ['binary_logloss', 'mape']
+
+    X, y = load_breast_cancer(return_X_y=True)
+    X_train, X_valid, y_train, y_valid = train_test_split(X, y, test_size=0.2, random_state=0)
+    train_ds = lgb.Dataset(X_train, y_train)
+    valid_ds = lgb.Dataset(X_valid, y_valid, reference=train_ds)
+
+    params = {'objective': 'binary', 'metric': metric, 'verbose': -1}
+    if isinstance(metric, str):
+        min_delta = metric2min_delta[metric]
+    elif first_only:
+        min_delta = metric2min_delta[metric[0]]
+    else:
+        min_delta = [metric2min_delta[m] for m in metric]
+    train_kwargs = dict(
+        params=params,
+        train_set=train_ds,
+        num_boost_round=50,
+        valid_sets=[train_ds, valid_ds],
+        valid_names=['training', 'valid'],
+    )
+
+    # regular early stopping
+    train_kwargs['callbacks'] = [lgb.callback.early_stopping(10, first_only, verbose=0)]
+    evals_result = {}
+    bst = lgb.train(evals_result=evals_result, **train_kwargs)
+    scores = np.vstack(list(evals_result['valid'].values())).T
+
+    # positive min_delta
+    train_kwargs['callbacks'] = [lgb.callback.early_stopping(10, first_only, verbose=0, min_delta=min_delta)]
+    delta_result = {}
+    delta_bst = lgb.train(evals_result=delta_result, **train_kwargs)
+    delta_scores = np.vstack(list(delta_result['valid'].values())).T
+
+    if first_only:
+        scores = scores[:, 0]
+        delta_scores = delta_scores[:, 0]
+
+    assert delta_bst.num_trees() < bst.num_trees()
+    np.testing.assert_allclose(scores[:len(delta_scores)], delta_scores)
+    last_score = delta_scores[-1]
+    best_score = delta_scores[delta_bst.num_trees() - 1]
+    if greater_is_better:
+        assert np.less_equal(last_score, best_score + min_delta).any()
+    else:
+        assert np.greater_equal(last_score, best_score - min_delta).any()
+
+
 def test_continue_train():
     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)