Add XGBoost integration

README.md

@@ -31,6 +31,7 @@ Optuna-Integration API reference is [here](https://optuna-integration.readthedoc
 * [skorch](https://optuna-integration.readthedocs.io/en/stable/reference/index.html#skorch)  ([example](https://github.com/optuna/optuna-examples/tree/main/pytorch/skorch_simple.py))
 * [TensorBoard](https://optuna-integration.readthedocs.io/en/stable/reference/index.html#tensorboard)  ([example](https://github.com/optuna/optuna-examples/tree/main/tensorboard/tensorboard_simple.py))
 * [tf.keras](https://optuna-integration.readthedocs.io/en/stable/reference/index.html#tensorflow)  ([example](https://github.com/optuna/optuna-examples/tree/main/tfkeras/tfkeras_integration.py))
+* [XGBoost](https://optuna-integration.readthedocs.io/en/stable/reference/index.html#xgboost)  ([example](https://github.com/optuna/optuna-examples/tree/main/xgboost/xgboost_integration.py))
 
 ## Installation
 

docs/source/reference/index.rst

@@ -136,4 +136,13 @@ TensorFlow
    :toctree: generated/
    :nosignatures:
 
-   optuna_integration.TFKerasPruningCallback
+   optuna.integration.TFKerasPruningCallback
+
+XGBoost
+----------
+
+.. autosummary::
+   :toctree: generated/
+   :nosignatures:
+
+   optuna.integration.XGBoostPruningCallback

optuna_integration/xgboost.py

@@ -0,0 +1,123 @@
+from typing import Any
+
+import optuna
+
+
+use_callback_cls = True
+
+with optuna_integration._imports.try_import() as _imports:
+    import xgboost as xgb
+
+    xgboost_version = xgb.__version__.split(".")
+    xgboost_major_version = int(xgboost_version[0])
+    xgboost_minor_version = int(xgboost_version[1])
+    use_callback_cls = (
+        xgboost_major_version >= 1 and xgboost_minor_version >= 3
+    ) or xgboost_major_version >= 2
+
+_doc = """Callback for XGBoost to prune unpromising trials.
+
+    See `the example <https://github.com/optuna/optuna-examples/blob/main/
+    xgboost/xgboost_integration.py>`__
+    if you want to add a pruning callback which observes validation accuracy of
+    a XGBoost model.
+
+    Args:
+        trial:
+            A :class:`~optuna.trial.Trial` corresponding to the current evaluation of the
+            objective function.
+        observation_key:
+            An evaluation metric for pruning, e.g., ``validation-error`` and
+            ``validation-merror``. When using the Scikit-Learn API, the index number of
+            ``eval_set`` must be included in the ``observation_key``, e.g.,
+            ``validation_0-error`` and ``validation_0-merror``. Please refer to ``eval_metric``
+            in `XGBoost reference <https://xgboost.readthedocs.io/en/latest/parameter.html>`_
+            for further details.
+    """
+
+if _imports.is_successful() and use_callback_cls:
+
+    class XGBoostPruningCallback(xgb.callback.TrainingCallback):
+        __doc__ = _doc
+
+        def __init__(self, trial: optuna.trial.Trial, observation_key: str) -> None:
+            self._trial = trial
+            self._observation_key = observation_key
+            self._is_cv = False
+
+        def before_training(self, model: Any) -> Any:
+            # The use of Any type is due to _PackedBooster is not yet being exposed
+            # to public interface as of xgboost 1.3.
+            if isinstance(model, xgb.Booster):
+                self._is_cv = False
+            else:
+                self._is_cv = True
+            return model
+
+        def after_iteration(self, model: Any, epoch: int, evals_log: dict) -> bool:
+            evaluation_results = {}
+            # Flatten the evaluation history to `{dataset-metric: score}` layout.
+            for dataset, metrics in evals_log.items():
+                for metric, scores in metrics.items():
+                    assert isinstance(scores, list), scores
+                    key = dataset + "-" + metric
+                    if self._is_cv:
+                        # Remove stddev of the metric across the cross-validation
+                        # folds.
+                        evaluation_results[key] = scores[-1][0]
+                    else:
+                        evaluation_results[key] = scores[-1]
+
+            current_score = evaluation_results[self._observation_key]
+            self._trial.report(current_score, step=epoch)
+            if self._trial.should_prune():
+                message = "Trial was pruned at iteration {}.".format(epoch)
+                raise optuna.TrialPruned(message)
+            # The training should not stop.
+            return False
+
+elif _imports.is_successful():
+
+    def _get_callback_context(env: "xgb.core.CallbackEnv") -> str:  # type: ignore
+        """Return whether the current callback context is cv or train.
+
+        .. note::
+            `Reference
+            <https://github.com/dmlc/xgboost/blob/master/python-package/xgboost/callback.py>`_.
+        """
+
+        if env.model is None and env.cvfolds is not None:
+            context = "cv"
+        else:
+            context = "train"
+        return context
+
+    class XGBoostPruningCallback:  # type: ignore
+        __doc__ = _doc
+
+        def __init__(self, trial: optuna.trial.Trial, observation_key: str) -> None:
+            self._trial = trial
+            self._observation_key = observation_key
+
+        def __call__(self, env: "xgb.core.CallbackEnv") -> None:  # type: ignore
+            context = _get_callback_context(env)
+            evaluation_result_list = env.evaluation_result_list
+            if context == "cv":
+                # Remove a third element: the stddev of the metric across the
+                # cross-validation folds.
+                evaluation_result_list = [
+                    (key, metric) for key, metric, _ in evaluation_result_list
+                ]
+            current_score = dict(evaluation_result_list)[self._observation_key]
+            self._trial.report(current_score, step=env.iteration)
+            if self._trial.should_prune():
+                message = "Trial was pruned at iteration {}.".format(env.iteration)
+                raise optuna.TrialPruned(message)
+
+else:
+
+    class XGBoostPruningCallback:  # type: ignore
+        __doc__ = _doc
+
+        def __init__(self, trial: optuna.trial.Trial, observation_key: str) -> None:
+            _imports.check()

pyproject.toml

@@ -68,6 +68,7 @@ all = [
   "skorch",
   "tensorboard",
   "tensorflow",
+  "xgboost",
 ]
 
 [tool.setuptools.packages.find]

tests/test_xgboost.py

@@ -0,0 +1,79 @@
+import numpy as np
+import pytest
+
+import optuna
+from optuna_integration._imports import try_import
+from optuna_integration.xgboost import XGBoostPruningCallback
+from optuna.testing.pruners import DeterministicPruner
+
+
+with try_import():
+    import xgboost as xgb
+
+pytestmark = pytest.mark.integration
+
+
+def test_xgboost_pruning_callback_call() -> None:
+    # The pruner is deactivated.
+    study = optuna.create_study(pruner=DeterministicPruner(False))
+    trial = study.ask()
+    pruning_callback = XGBoostPruningCallback(trial, "validation-logloss")
+    pruning_callback.after_iteration(
+        model=None, epoch=1, evals_log={"validation": {"logloss": [1.0]}}
+    )
+
+    # The pruner is activated.
+    study = optuna.create_study(pruner=DeterministicPruner(True))
+    trial = study.ask()
+    pruning_callback = XGBoostPruningCallback(trial, "validation-logloss")
+    with pytest.raises(optuna.TrialPruned):
+        pruning_callback.after_iteration(
+            model=None, epoch=1, evals_log={"validation": {"logloss": [1.0]}}
+        )
+
+
+def test_xgboost_pruning_callback() -> None:
+    def objective(trial: optuna.trial.Trial) -> float:
+        dtrain = xgb.DMatrix(np.asarray([[1.0]]), label=[1.0])
+        dtest = xgb.DMatrix(np.asarray([[1.0]]), label=[1.0])
+
+        pruning_callback = XGBoostPruningCallback(trial, "validation-logloss")
+        xgb.train(
+            {"objective": "binary:logistic"},
+            dtrain,
+            1,
+            evals=[(dtest, "validation")],
+            verbose_eval=False,
+            callbacks=[pruning_callback],
+        )
+        return 1.0
+
+    study = optuna.create_study(pruner=DeterministicPruner(True))
+    study.optimize(objective, n_trials=1)
+    assert study.trials[0].state == optuna.trial.TrialState.PRUNED
+
+    study = optuna.create_study(pruner=DeterministicPruner(False))
+    study.optimize(objective, n_trials=1)
+    assert study.trials[0].state == optuna.trial.TrialState.COMPLETE
+    assert study.trials[0].value == 1.0
+
+
+def test_xgboost_pruning_callback_cv() -> None:
+    def objective(trial: optuna.trial.Trial) -> float:
+        dtrain = xgb.DMatrix(np.ones((2, 1)), label=[1.0, 1.0])
+        params = {
+            "objective": "binary:logistic",
+        }
+
+        pruning_callback = optuna.integration.XGBoostPruningCallback(trial, "test-logloss")
+        xgb.cv(params, dtrain, callbacks=[pruning_callback], nfold=2)
+        return 1.0
+
+    study = optuna.create_study(pruner=DeterministicPruner(True))
+    study.optimize(objective, n_trials=1)
+    assert study.trials[0].state == optuna.trial.TrialState.PRUNED
+
+    study = optuna.create_study(pruner=DeterministicPruner(False))
+    study.optimize(objective, n_trials=1)
+    assert study.trials[0].state == optuna.trial.TrialState.COMPLETE
+    assert study.trials[0].value == 1.0