Пример #1
0
class TrainableTest(unittest.TestCase):
    X_id = None
    y_id = None
    X = None
    y = None

    @classmethod
    def setUpClass(cls):
        ray.init(local_mode=True)
        X, y = make_classification(
            n_samples=50, n_features=50, n_informative=3, random_state=0)
        cls.X_id, cls.y_id = ray.put(X), ray.put(y)
        cls.y = y
        cls.X = X

    @classmethod
    def tearDownClass(cls):
        ray.shutdown()

    def base_params(self, estimator_list):
        config = {"estimator_list": estimator_list}
        cv = check_cv(
            cv=len(estimator_list), y=self.y, classifier=estimator_list[0])
        config["X_id"] = self.X_id
        config["y_id"] = self.y_id
        config["early_stopping"] = False
        config["early_stop_type"] = get_early_stop_type(
            estimator_list[0], False)
        config["max_iters"] = 1
        config["groups"] = None
        config["cv"] = cv
        config["fit_params"] = None
        config["scoring"], _ = _check_multimetric_scoring(
            estimator_list[0], scoring=None)
        config["return_train_score"] = False
        config["n_jobs"] = 1
        return config

    def test_basic_train(self):
        config = self.base_params(estimator_list=[SVC(), SVC()])
        trainable = _Trainable(config)
        trainable.train()
        trainable.stop()

    @unittest.skipIf(not has_xgboost(), "xgboost not installed")
    def testXGBoostEarlyStop(self):
        config = self.base_params(
            estimator_list=[create_xgboost(),
                            create_xgboost()])
        config["early_stopping"] = True
        config["early_stop_type"] = get_early_stop_type(
            config["estimator_list"][0], True)
        trainable = _Trainable(config)
        trainable.train()
        assert all(trainable.saved_models)
        trainable.train()
        assert all(trainable.saved_models)
        trainable.stop()

    @unittest.skipIf(not has_xgboost(), "xgboost not installed")
    def testXGBoostNoEarlyStop(self):
        config = self.base_params(
            estimator_list=[create_xgboost(),
                            create_xgboost()])
        config["early_stopping"] = False
        trainable = _Trainable(config)
        trainable.train()
        assert not any(trainable.saved_models)
        trainable.stop()

    @unittest.skipIf(not has_required_lightgbm_version(),
                     "lightgbm>=3.0.0 not installed")
    def testLGBMEarlyStop(self):
        config = self.base_params(
            estimator_list=[create_lightgbm(),
                            create_lightgbm()])
        config["early_stopping"] = True
        config["early_stop_type"] = get_early_stop_type(
            config["estimator_list"][0], True)
        trainable = _Trainable(config)
        trainable.train()
        assert all(trainable.saved_models)
        trainable.train()
        assert all(trainable.saved_models)
        trainable.stop()

    @unittest.skipIf(not has_required_lightgbm_version(),
                     "lightgbm>=3.0.0 not installed")
    def testLGBMNoEarlyStop(self):
        config = self.base_params(
            estimator_list=[create_lightgbm(),
                            create_lightgbm()])
        config["early_stopping"] = False
        trainable = _Trainable(config)
        trainable.train()
        assert not any(trainable.saved_models)
        trainable.stop()

    # @unittest.skipIf(not has_catboost(), "catboost not installed")
    @unittest.skip("Catboost needs to be updated.")
    def testCatboostEarlyStop(self):
        config = self.base_params(
            estimator_list=[create_catboost(),
                            create_catboost()])
        config["early_stopping"] = True
        config["early_stop_type"] = get_early_stop_type(
            config["estimator_list"][0], True)
        trainable = _Trainable(config)
        trainable.train()
        assert all(trainable.saved_models)
        trainable.train()
        assert all(trainable.saved_models)
        trainable.stop()

    # @unittest.skipIf(not has_catboost(), "catboost not installed")
    @unittest.skip("Catboost needs to be updated.")
    def testCatboostNoEarlyStop(self):
        config = self.base_params(
            estimator_list=[create_catboost(),
                            create_catboost()])
        config["early_stopping"] = False
        trainable = _Trainable(config)
        trainable.train()
        assert not any(trainable.saved_models)
        trainable.stop()

    def testPartialFit(self):
        config = self.base_params([SGDClassifier(), SGDClassifier()])
        config["early_stopping"] = True
        config["early_stop_type"] = get_early_stop_type(
            config["estimator_list"][0], True)
        trainable = _Trainable(config)
        trainable.train()
        assert trainable.estimator_list[0].t_ > 0
        previous_t = trainable.estimator_list[0].t_
        trainable.train()
        assert trainable.estimator_list[0].t_ > previous_t
        trainable.stop()

    def testNoPartialFit(self):
        config = self.base_params([SGDClassifier(), SGDClassifier()])
        config["early_stopping"] = False
        trainable = _Trainable(config)
        trainable.train()
        assert not hasattr(trainable.estimator_list[0], "t_")
        trainable.train()
        assert not hasattr(trainable.estimator_list[0], "t_")
        trainable.stop()

    def testWarmStart(self):
        # Hard to get introspection so we just test that it runs.
        config = self.base_params([LogisticRegression(), LogisticRegression()])
        config["early_stopping"] = True
        config["early_stop_type"] = get_early_stop_type(
            config["estimator_list"][0], True)
        trainable = _Trainable(config)
        trainable.train()
        trainable.train()
        trainable.stop()
Пример #2
0
class RandomizedSearchTest(unittest.TestCase):
    def test_random_search_cv_results(self):
        # Make a dataset with a lot of noise to get various kind of prediction
        # errors across CV folds and parameter settings
        X, y = make_classification(
            n_samples=50, n_features=50, n_informative=3, random_state=0)

        # scipy.stats dists now supports `seed` but we still support scipy 0.12
        # which doesn't support the seed. Hence the assertions in the test for
        # random_search alone should not depend on randomization.
        n_splits = 3
        n_search_iter = 30
        params = dict(C=expon(scale=10), gamma=expon(scale=0.1))
        random_search = TuneSearchCV(
            SVC(),
            n_trials=n_search_iter,
            cv=n_splits,
            param_distributions=params,
            return_train_score=True,
            n_jobs=2)
        random_search.fit(X, y)

        param_keys = ("param_C", "param_gamma")
        score_keys = (
            "mean_test_score",
            "mean_train_score",
            "rank_test_score",
            "rank_train_score",
            "split0_test_score",
            "split1_test_score",
            "split2_test_score",
            "split0_train_score",
            "split1_train_score",
            "split2_train_score",
            "std_test_score",
            "std_train_score",
            "time_total_s",
        )
        n_cand = n_search_iter

        def test_check_cv_results_array_types(cv_results, param_keys,
                                              score_keys):
            # Check if the search `cv_results`'s array are of correct types
            self.assertTrue(
                all(
                    isinstance(cv_results[param], np.ma.MaskedArray)
                    for param in param_keys))
            self.assertTrue(
                all(cv_results[key].dtype == object for key in param_keys))
            self.assertFalse(
                any(
                    isinstance(cv_results[key], np.ma.MaskedArray)
                    for key in score_keys))
            self.assertTrue(
                all(cv_results[key].dtype == np.float64 for key in score_keys
                    if not key.startswith("rank")))
            self.assertEquals(cv_results["rank_test_score"].dtype, np.int32)

        def test_check_cv_results_keys(cv_results, param_keys, score_keys,
                                       n_cand):
            # Test the search.cv_results_ contains all the required results
            assert_array_equal(
                sorted(cv_results.keys()),
                sorted(param_keys + score_keys + ("params", )))
            self.assertTrue(
                all(cv_results[key].shape == (n_cand, )
                    for key in param_keys + score_keys))

        cv_results = random_search.cv_results_
        # Check results structure
        test_check_cv_results_array_types(cv_results, param_keys, score_keys)
        test_check_cv_results_keys(cv_results, param_keys, score_keys, n_cand)
        # For random_search, all the param array vals should be unmasked
        self.assertFalse(
            any(cv_results["param_C"].mask)
            or any(cv_results["param_gamma"].mask))

    def test_local_dir(self):
        digits = datasets.load_digits()
        x = digits.data
        y = digits.target

        clf = SGDClassifier()
        parameter_grid = {
            "alpha": Real(1e-4, 1e-1, 1),
            "epsilon": Real(0.01, 0.1)
        }

        scheduler = MedianStoppingRule(grace_period=10.0)

        tune_search = TuneSearchCV(
            clf,
            parameter_grid,
            early_stopping=scheduler,
            max_iters=10,
            local_dir="./test-result")
        tune_search.fit(x, y)

        self.assertTrue(len(os.listdir("./test-result")) != 0)

    def test_local_mode(self):
        digits = datasets.load_digits()
        x = digits.data
        y = digits.target

        clf = SGDClassifier()
        parameter_grid = {
            "alpha": Real(1e-4, 1e-1, 1),
            "epsilon": Real(0.01, 0.1)
        }
        tune_search = TuneSearchCV(
            clf,
            parameter_grid,
            n_jobs=1,
            max_iters=10,
            local_dir="./test-result")
        import ray
        with patch.object(ray, "init", wraps=ray.init) as wrapped_init:
            tune_search.fit(x, y)
        self.assertTrue(wrapped_init.call_args[1]["local_mode"])

    def test_multi_best_classification(self):
        digits = datasets.load_digits()
        x = digits.data
        y = digits.target
        model = SGDClassifier()

        parameter_grid = {"alpha": [1e-4, 1e-1, 1], "epsilon": [0.01, 0.1]}
        scoring = ("accuracy", "f1_micro")
        search_methods = ["random", "bayesian", "hyperopt", "bohb"]
        for search_method in search_methods:

            tune_search = TuneSearchCV(
                model,
                parameter_grid,
                scoring=scoring,
                search_optimization=search_method,
                cv=2,
                n_trials=3,
                n_jobs=1,
                refit="accuracy")
            tune_search.fit(x, y)
            self.assertAlmostEqual(
                tune_search.best_score_,
                max(tune_search.cv_results_["mean_test_accuracy"]),
                places=10)

            p = tune_search.cv_results_["params"]
            scores = tune_search.cv_results_["mean_test_accuracy"]
            cv_best_param = max(
                list(zip(scores, p)), key=lambda pair: pair[0])[1]
            self.assertEqual(tune_search.best_params_, cv_best_param)

    def test_multi_best_classification_scoring_dict(self):
        digits = datasets.load_digits()
        x = digits.data
        y = digits.target
        model = SGDClassifier()

        parameter_grid = {"alpha": [1e-4, 1e-1, 1], "epsilon": [0.01, 0.1]}
        scoring = {"acc": "accuracy", "f1": "f1_micro"}
        search_methods = ["random", "bayesian", "hyperopt", "bohb"]
        for search_method in search_methods:

            tune_search = TuneSearchCV(
                model,
                parameter_grid,
                scoring=scoring,
                search_optimization=search_method,
                cv=2,
                n_trials=3,
                n_jobs=1,
                refit="acc")
            tune_search.fit(x, y)
            self.assertAlmostEqual(
                tune_search.best_score_,
                max(tune_search.cv_results_["mean_test_acc"]),
                places=10)

            p = tune_search.cv_results_["params"]
            scores = tune_search.cv_results_["mean_test_acc"]
            cv_best_param = max(
                list(zip(scores, p)), key=lambda pair: pair[0])[1]
            self.assertEqual(tune_search.best_params_, cv_best_param)

    def test_multi_best_regression(self):
        x, y = make_regression(n_samples=100, n_features=10, n_informative=5)
        model = SGDRegressor()
        parameter_grid = {"alpha": [1e-4, 1e-1, 1], "epsilon": [0.01, 0.1]}

        scoring = ("neg_mean_absolute_error", "neg_mean_squared_error")

        search_methods = ["random", "bayesian", "hyperopt", "bohb"]
        for search_method in search_methods:

            tune_search = TuneSearchCV(
                model,
                parameter_grid,
                scoring=scoring,
                search_optimization=search_method,
                cv=2,
                n_trials=3,
                n_jobs=1,
                refit="neg_mean_absolute_error")
            tune_search.fit(x, y)
            self.assertAlmostEqual(
                tune_search.best_score_,
                max(tune_search.cv_results_[
                    "mean_test_neg_mean_absolute_error"]),
                places=10)

            p = tune_search.cv_results_["params"]
            scores = tune_search.cv_results_[
                "mean_test_neg_mean_absolute_error"]
            cv_best_param = max(
                list(zip(scores, p)), key=lambda pair: pair[0])[1]
            self.assertEqual(tune_search.best_params_, cv_best_param)

    def test_multi_refit_false(self):
        digits = datasets.load_digits()
        x = digits.data
        y = digits.target
        model = SGDClassifier()

        parameter_grid = {"alpha": [1e-4, 1e-1, 1], "epsilon": [0.01, 0.1]}
        scoring = ("accuracy", "f1_micro")

        tune_search = TuneSearchCV(
            model,
            parameter_grid,
            scoring=scoring,
            search_optimization="random",
            cv=2,
            n_trials=3,
            n_jobs=1,
            refit=False)
        tune_search.fit(x, y)
        with self.assertRaises(ValueError) as exc:
            tune_search.best_score_
        self.assertTrue(("instance was initialized with refit=False. "
                         "For multi-metric evaluation,") in str(exc.exception))
        with self.assertRaises(ValueError) as exc:
            tune_search.best_index_
        self.assertTrue(("instance was initialized with refit=False. "
                         "For multi-metric evaluation,") in str(exc.exception))
        with self.assertRaises(ValueError) as exc:
            tune_search.best_params_
        self.assertTrue(("instance was initialized with refit=False. "
                         "For multi-metric evaluation,") in str(exc.exception))

    def test_warm_start_detection(self):
        parameter_grid = {"alpha": Real(1e-4, 1e-1, 1)}
        from sklearn.ensemble import VotingClassifier, RandomForestClassifier
        clf = VotingClassifier(estimators=[(
            "rf", RandomForestClassifier(n_estimators=50, random_state=0))])
        tune_search = TuneSearchCV(
            clf,
            parameter_grid,
            n_jobs=1,
            max_iters=10,
            local_dir="./test-result")
        self.assertEqual(tune_search.early_stop_type,
                         EarlyStopping.NO_EARLY_STOP)

        from sklearn.tree import DecisionTreeClassifier
        clf = DecisionTreeClassifier(random_state=0)
        tune_search2 = TuneSearchCV(
            clf,
            parameter_grid,
            n_jobs=1,
            max_iters=10,
            local_dir="./test-result")
        self.assertEqual(tune_search2.early_stop_type,
                         EarlyStopping.NO_EARLY_STOP)

        from sklearn.linear_model import LogisticRegression
        clf = LogisticRegression()
        tune_search3 = TuneSearchCV(
            clf,
            parameter_grid,
            n_jobs=1,
            max_iters=10,
            local_dir="./test-result")

        self.assertEqual(tune_search3.early_stop_type,
                         EarlyStopping.NO_EARLY_STOP)

        tune_search4 = TuneSearchCV(
            clf,
            parameter_grid,
            early_stopping=True,
            n_jobs=1,
            max_iters=10,
            local_dir="./test-result")
        self.assertEqual(tune_search4.early_stop_type,
                         EarlyStopping.WARM_START_ITER)

        clf = RandomForestClassifier()
        tune_search5 = TuneSearchCV(
            clf,
            parameter_grid,
            early_stopping=True,
            n_jobs=1,
            max_iters=10,
            local_dir="./test-result")
        self.assertEqual(tune_search5.early_stop_type,
                         EarlyStopping.WARM_START_ENSEMBLE)

    def test_warm_start_error(self):
        parameter_grid = {"alpha": Real(1e-4, 1e-1, 1)}
        from sklearn.ensemble import VotingClassifier, RandomForestClassifier
        clf = VotingClassifier(estimators=[(
            "rf", RandomForestClassifier(n_estimators=50, random_state=0))])
        tune_search = TuneSearchCV(
            clf,
            parameter_grid,
            n_jobs=1,
            early_stopping=False,
            max_iters=10,
            local_dir="./test-result")
        self.assertFalse(tune_search._can_early_stop())
        with self.assertRaises(ValueError):
            tune_search = TuneSearchCV(
                clf,
                parameter_grid,
                n_jobs=1,
                early_stopping=True,
                max_iters=10,
                local_dir="./test-result")

        from sklearn.linear_model import LogisticRegression
        clf = LogisticRegression()
        with self.assertRaises(ValueError):
            parameter_grid = {"max_iter": [1, 2]}
            TuneSearchCV(
                clf,
                parameter_grid,
                early_stopping=True,
                n_jobs=1,
                max_iters=10,
                local_dir="./test-result")

        from sklearn.ensemble import RandomForestClassifier
        clf = RandomForestClassifier()
        with self.assertRaises(ValueError):
            parameter_grid = {"n_estimators": [1, 2]}
            TuneSearchCV(
                clf,
                parameter_grid,
                early_stopping=True,
                n_jobs=1,
                max_iters=10,
                local_dir="./test-result")

    def test_warn_reduce_maxiters(self):
        parameter_grid = {"alpha": Real(1e-4, 1e-1, 1)}
        from sklearn.ensemble import RandomForestClassifier
        clf = RandomForestClassifier(max_depth=2, random_state=0)
        with self.assertWarnsRegex(UserWarning, "max_iters is set"):
            TuneSearchCV(
                clf, parameter_grid, max_iters=10, local_dir="./test-result")
        with self.assertWarnsRegex(UserWarning, "max_iters is set"):
            TuneSearchCV(
                SGDClassifier(),
                parameter_grid,
                max_iters=10,
                local_dir="./test-result")

    def test_warn_early_stop(self):
        with self.assertWarnsRegex(UserWarning, "max_iters = 1"):
            TuneSearchCV(
                LogisticRegression(), {"C": [1, 2]}, early_stopping=True)
        with self.assertWarnsRegex(UserWarning, "max_iters = 1"):
            TuneSearchCV(
                SGDClassifier(), {"epsilon": [0.1, 0.2]}, early_stopping=True)

    @unittest.skipIf(not has_xgboost(), "xgboost not installed")
    def test_early_stop_xgboost_warn(self):
        from xgboost.sklearn import XGBClassifier
        with self.assertWarnsRegex(UserWarning, "github.com"):
            TuneSearchCV(
                XGBClassifier(), {"C": [1, 2]},
                early_stopping=True,
                max_iters=10)
        with self.assertWarnsRegex(UserWarning, "max_iters"):
            TuneSearchCV(
                XGBClassifier(), {"C": [1, 2]},
                early_stopping=True,
                max_iters=1)

    @unittest.skipIf(not has_required_lightgbm_version(),
                     "lightgbm not installed")
    def test_early_stop_lightgbm_warn(self):
        from lightgbm import LGBMClassifier
        with self.assertWarnsRegex(UserWarning, "lightgbm"):
            TuneSearchCV(
                LGBMClassifier(), {"learning_rate": [0.1, 0.5]},
                early_stopping=True,
                max_iters=10)
        with self.assertWarnsRegex(UserWarning, "max_iters"):
            TuneSearchCV(
                LGBMClassifier(), {"learning_rate": [0.1, 0.5]},
                early_stopping=True,
                max_iters=1)

    @unittest.skipIf(not has_catboost(), "catboost not installed")
    def test_early_stop_catboost_warn(self):
        from catboost import CatBoostClassifier
        with self.assertWarnsRegex(UserWarning, "Catboost"):
            TuneSearchCV(
                CatBoostClassifier(), {"learning_rate": [0.1, 0.5]},
                early_stopping=True,
                max_iters=10)
        with self.assertWarnsRegex(UserWarning, "max_iters"):
            TuneSearchCV(
                CatBoostClassifier(), {"learning_rate": [0.1, 0.5]},
                early_stopping=True,
                max_iters=1)

    def test_pipeline_early_stop(self):
        digits = datasets.load_digits()
        x = digits.data
        y = digits.target

        pipe = Pipeline([("reduce_dim", PCA()), ("classify", SGDClassifier())])
        parameter_grid = [
            {
                "classify__alpha": [1e-4, 1e-1, 1],
                "classify__epsilon": [0.01, 0.1]
            },
        ]

        with self.assertRaises(ValueError) as exc:
            TuneSearchCV(
                pipe,
                parameter_grid,
                early_stopping=True,
                pipeline_auto_early_stop=False,
                max_iters=10)
        self.assertTrue((
            "Early stopping is not supported because the estimator does "
            "not have `partial_fit`, does not support warm_start, or "
            "is a tree classifier. Set `early_stopping=False`."
        ) in str(exc.exception))

        tune_search = TuneSearchCV(
            pipe, parameter_grid, early_stopping=True, max_iters=10)
        tune_search.fit(x, y)