def test_preprocessing_dtype(self):
# Dense
# np.float32
X_train, Y_train, X_test, Y_test = get_dataset("iris")
self.assertEqual(X_train.dtype, np.float32)
configuration_space = RandomTreesEmbedding.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
preprocessor = RandomTreesEmbedding(random_state=1,
**{hp_name: default[hp_name] for
hp_name in
default})
preprocessor.fit(X_train)
Xt = preprocessor.transform(X_train)
self.assertEqual(Xt.dtype, np.float32)
# np.float64
X_train, Y_train, X_test, Y_test = get_dataset("iris")
X_train = X_train.astype(np.float64)
configuration_space = RandomTreesEmbedding.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
preprocessor = RandomTreesEmbedding(random_state=1,
**{hp_name: default[hp_name] for
hp_name in
default})
preprocessor.fit(X_train, Y_train)
Xt = preprocessor.transform(X_train)
self.assertEqual(Xt.dtype, np.float64)
def test_predict_proba_batched(self):
# Multiclass
cls = SimpleClassificationPipeline(include={'classifier': ['sgd']})
X_train, Y_train, X_test, Y_test = get_dataset(dataset='digits')
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = cls.predict_proba(X_test_)
# The object behind the last step in the pipeline
cls_predict = unittest.mock.Mock(wraps=cls.steps[-1][1].predict_proba)
cls.steps[-1][-1].predict_proba = cls_predict
prediction = cls.predict_proba(X_test, batch_size=20)
self.assertEqual((1647, 10), prediction.shape)
self.assertEqual(84, cls_predict.call_count)
assert_array_almost_equal(prediction_, prediction)
# Multilabel
cls = SimpleClassificationPipeline(include={'classifier': ['lda']})
X_train, Y_train, X_test, Y_test = get_dataset(dataset='digits')
Y_train = np.array(list([(list([1 if i != y else 0 for i in range(10)]))
for y in Y_train]))
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = cls.predict_proba(X_test_)
# The object behind the last step in the pipeline
cls_predict = unittest.mock.Mock(wraps=cls.steps[-1][1].predict_proba)
cls.steps[-1][-1].predict_proba = cls_predict
prediction = cls.predict_proba(X_test, batch_size=20)
self.assertEqual((1647, 10), prediction.shape)
self.assertEqual(84, cls_predict.call_count)
assert_array_almost_equal(prediction_, prediction)
def test_predict_proba_batched(self):
cs = SimpleClassificationPipeline.get_hyperparameter_search_space()
default = cs.get_default_configuration()
# Multiclass
cls = SimpleClassificationPipeline(default)
X_train, Y_train, X_test, Y_test = get_dataset(dataset="digits")
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = cls.predict_proba(X_test_)
# The object behind the last step in the pipeline
cls_predict = mock.Mock(wraps=cls.pipeline_.steps[-1][1])
cls.pipeline_.steps[-1] = ("estimator", cls_predict)
prediction = cls.predict_proba(X_test, batch_size=20)
self.assertEqual((1647, 10), prediction.shape)
self.assertEqual(84, cls_predict.predict_proba.call_count)
assert_array_almost_equal(prediction_, prediction)
# Multilabel
cls = SimpleClassificationPipeline(default)
X_train, Y_train, X_test, Y_test = get_dataset(dataset="digits")
Y_train = np.array(list([(list([1 if i != y else 0 for i in range(10)])) for y in Y_train]))
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = cls.predict_proba(X_test_)
cls_predict = mock.Mock(wraps=cls.pipeline_.steps[-1][1])
cls.pipeline_.steps[-1] = ("estimator", cls_predict)
prediction = cls.predict_proba(X_test, batch_size=20)
self.assertIsInstance(prediction, np.ndarray)
self.assertEqual(prediction.shape, ((1647, 10)))
self.assertEqual(84, cls_predict.predict_proba.call_count)
assert_array_almost_equal(prediction_, prediction)
def test_predict_proba_batched(self):
cs = SimpleClassificationPipeline.get_hyperparameter_search_space()
default = cs.get_default_configuration()
# Multiclass
cls = SimpleClassificationPipeline(default)
X_train, Y_train, X_test, Y_test = get_dataset(dataset='digits')
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = cls.predict_proba(X_test_)
# The object behind the last step in the pipeline
cls_predict = mock.Mock(wraps=cls.pipeline_.steps[-1][1])
cls.pipeline_.steps[-1] = ("estimator", cls_predict)
prediction = cls.predict_proba(X_test, batch_size=20)
self.assertEqual((1647, 10), prediction.shape)
self.assertEqual(84, cls_predict.predict_proba.call_count)
assert_array_almost_equal(prediction_, prediction)
# Multilabel
cls = SimpleClassificationPipeline(default)
X_train, Y_train, X_test, Y_test = get_dataset(dataset='digits')
Y_train_ = np.zeros((Y_train.shape[0], 10))
for i, y in enumerate(Y_train):
Y_train_[i][y] = 1
Y_train = Y_train_
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = cls.predict_proba(X_test_)
cls_predict = mock.Mock(wraps=cls.pipeline_.steps[-1][1])
cls.pipeline_.steps[-1] = ("estimator", cls_predict)
prediction = cls.predict_proba(X_test, batch_size=20)
self.assertIsInstance(prediction, np.ndarray)
self.assertEqual(prediction.shape, ((1647, 10)))
self.assertEqual(84, cls_predict.predict_proba.call_count)
assert_array_almost_equal(prediction_, prediction)
def test_predict_batched(self):
cs = SimpleClassificationPipeline.get_hyperparameter_search_space()
default = cs.get_default_configuration()
cls = SimpleClassificationPipeline(default)
# Multiclass
X_train, Y_train, X_test, Y_test = get_dataset(dataset="digits")
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = cls.predict(X_test_)
cls_predict = mock.Mock(wraps=cls.pipeline_)
cls.pipeline_ = cls_predict
prediction = cls.predict(X_test, batch_size=20)
self.assertEqual((1647,), prediction.shape)
self.assertEqual(83, cls_predict.predict.call_count)
assert_array_almost_equal(prediction_, prediction)
# Multilabel
X_train, Y_train, X_test, Y_test = get_dataset(dataset="digits")
Y_train = np.array(list([(list([1 if i != y else 0 for i in range(10)])) for y in Y_train]))
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = cls.predict(X_test_)
cls_predict = mock.Mock(wraps=cls.pipeline_)
cls.pipeline_ = cls_predict
prediction = cls.predict(X_test, batch_size=20)
self.assertEqual((1647, 10), prediction.shape)
self.assertEqual(83, cls_predict.predict.call_count)
assert_array_almost_equal(prediction_, prediction)
def _test_preprocessing_dtype(self):
# Dense
# np.float32
X_train, Y_train, X_test, Y_test = get_dataset("iris")
self.assertEqual(X_train.dtype, np.float32)
configuration_space = Nystroem.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
preprocessor = Nystroem(random_state=1,
**{hp.hyperparameter.name: hp.value
for hp
in
default.values.values()})
preprocessor.fit(X_train)
Xt = preprocessor.transform(X_train)
self.assertEqual(Xt.dtype, np.float32)
# np.float64
X_train, Y_train, X_test, Y_test = get_dataset("iris")
X_train = X_train.astype(np.float64)
configuration_space = Nystroem.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
preprocessor = Nystroem(random_state=1,
**{hp.hyperparameter.name: hp.value
for hp
in
default.values.values()})
preprocessor.fit(X_train, Y_train)
Xt = preprocessor.transform(X_train)
self.assertEqual(Xt.dtype, np.float64)
# Sparse
# np.float32
X_train, Y_train, X_test, Y_test = get_dataset("iris", make_sparse=True)
self.assertEqual(X_train.dtype, np.float32)
configuration_space = Nystroem.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
preprocessor = Nystroem(random_state=1,
**{hp.hyperparameter.name: hp.value
for hp
in
default.values.values()})
preprocessor.fit(X_train)
Xt = preprocessor.transform(X_train)
self.assertEqual(Xt.dtype, np.float32)
# np.float64
X_train, Y_train, X_test, Y_test = get_dataset("iris", make_sparse=True)
X_train = X_train.astype(np.float64)
configuration_space = Nystroem.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
preprocessor = Nystroem(random_state=1,
**{hp.hyperparameter.name: hp.value
for hp
in
default.values.values()})
preprocessor.fit(X_train)
Xt = preprocessor.transform(X_train)
self.assertEqual(Xt.dtype, np.float64)
def test_predict_proba_batched_sparse(self):
cs = SimpleClassificationPipeline.get_hyperparameter_search_space(
dataset_properties={'sparse': True})
config = Configuration(cs,
values={"balancing:strategy": "none",
"classifier:__choice__": "random_forest",
"imputation:strategy": "mean",
"one_hot_encoding:minimum_fraction": 0.01,
"one_hot_encoding:use_minimum_fraction": 'True',
"preprocessor:__choice__": "no_preprocessing",
'classifier:random_forest:bootstrap': 'True',
'classifier:random_forest:criterion': 'gini',
'classifier:random_forest:max_depth': 'None',
'classifier:random_forest:min_samples_split': 2,
'classifier:random_forest:min_samples_leaf': 2,
'classifier:random_forest:min_weight_fraction_leaf': 0.0,
'classifier:random_forest:max_features': 0.5,
'classifier:random_forest:max_leaf_nodes': 'None',
'classifier:random_forest:n_estimators': 100,
"rescaling:__choice__": "min/max"})
# Multiclass
cls = SimpleClassificationPipeline(config)
X_train, Y_train, X_test, Y_test = get_dataset(dataset='digits',
make_sparse=True)
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = cls.predict_proba(X_test_)
# The object behind the last step in the pipeline
cls_predict = mock.Mock(wraps=cls.pipeline_.steps[-1][1])
cls.pipeline_.steps[-1] = ("estimator", cls_predict)
prediction = cls.predict_proba(X_test, batch_size=20)
self.assertEqual((1647, 10), prediction.shape)
self.assertEqual(84, cls_predict.predict_proba.call_count)
assert_array_almost_equal(prediction_, prediction)
# Multilabel
cls = SimpleClassificationPipeline(config)
X_train, Y_train, X_test, Y_test = get_dataset(dataset='digits',
make_sparse=True)
Y_train_ = np.zeros((Y_train.shape[0], 10))
for i, y in enumerate(Y_train):
Y_train_[i][y] = 1
Y_train = Y_train_
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = cls.predict_proba(X_test_)
cls_predict = mock.Mock(wraps=cls.pipeline_.steps[-1][1])
cls.pipeline_.steps[-1] = ("estimator", cls_predict)
prediction = cls.predict_proba(X_test, batch_size=20)
self.assertEqual(prediction.shape, ((1647, 10)))
self.assertIsInstance(prediction, np.ndarray)
self.assertEqual(84, cls_predict.predict_proba.call_count)
assert_array_almost_equal(prediction_, prediction)
def test_predict_proba_batched_sparse(self):
cs = SimpleClassificationPipeline.get_hyperparameter_search_space(dataset_properties={"sparse": True})
config = Configuration(
cs,
values={
"balancing:strategy": "none",
"classifier:__choice__": "random_forest",
"imputation:strategy": "mean",
"one_hot_encoding:minimum_fraction": 0.01,
"one_hot_encoding:use_minimum_fraction": "True",
"preprocessor:__choice__": "no_preprocessing",
"classifier:random_forest:bootstrap": "True",
"classifier:random_forest:criterion": "gini",
"classifier:random_forest:max_depth": "None",
"classifier:random_forest:min_samples_split": 2,
"classifier:random_forest:min_samples_leaf": 2,
"classifier:random_forest:min_weight_fraction_leaf": 0.0,
"classifier:random_forest:max_features": 0.5,
"classifier:random_forest:max_leaf_nodes": "None",
"classifier:random_forest:n_estimators": 100,
"rescaling:__choice__": "min/max",
},
)
# Multiclass
cls = SimpleClassificationPipeline(config)
X_train, Y_train, X_test, Y_test = get_dataset(dataset="digits", make_sparse=True)
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = cls.predict_proba(X_test_)
# The object behind the last step in the pipeline
cls_predict = mock.Mock(wraps=cls.pipeline_.steps[-1][1])
cls.pipeline_.steps[-1] = ("estimator", cls_predict)
prediction = cls.predict_proba(X_test, batch_size=20)
self.assertEqual((1647, 10), prediction.shape)
self.assertEqual(84, cls_predict.predict_proba.call_count)
assert_array_almost_equal(prediction_, prediction)
# Multilabel
cls = SimpleClassificationPipeline(config)
X_train, Y_train, X_test, Y_test = get_dataset(dataset="digits", make_sparse=True)
Y_train = np.array(list([(list([1 if i != y else 0 for i in range(10)])) for y in Y_train]))
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = cls.predict_proba(X_test_)
cls_predict = mock.Mock(wraps=cls.pipeline_.steps[-1][1])
cls.pipeline_.steps[-1] = ("estimator", cls_predict)
prediction = cls.predict_proba(X_test, batch_size=20)
self.assertEqual(prediction.shape, ((1647, 10)))
self.assertIsInstance(prediction, np.ndarray)
self.assertEqual(84, cls_predict.predict_proba.call_count)
assert_array_almost_equal(prediction_, prediction)
def test_preprocessing_dtype(self):
# Dense
# np.float32
X_train, Y_train, X_test, Y_test = get_dataset("iris")
self.assertEqual(X_train.dtype, np.float32)
configuration_space = SelectPercentileClassification.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
preprocessor = SelectPercentileClassification(random_state=1,
**{hp_name: default[hp_name]
for hp_name in default})
preprocessor.fit(X_train, Y_train)
Xt = preprocessor.transform(X_train)
self.assertEqual(Xt.dtype, np.float32)
# np.float64
X_train, Y_train, X_test, Y_test = get_dataset("iris")
X_train = X_train.astype(np.float64)
configuration_space = SelectPercentileClassification.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
preprocessor = SelectPercentileClassification(random_state=1,
**{hp_name: default[hp_name]
for hp_name in default})
preprocessor.fit(X_train, Y_train)
Xt = preprocessor.transform(X_train)
self.assertEqual(Xt.dtype, np.float64)
# Sparse
# np.float32
X_train, Y_train, X_test, Y_test = get_dataset("iris", make_sparse=True)
self.assertEqual(X_train.dtype, np.float32)
configuration_space = SelectPercentileClassification.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
preprocessor = SelectPercentileClassification(random_state=1,
**{hp_name: default[hp_name]
for hp_name in default})
preprocessor.fit(X_train, Y_train)
Xt = preprocessor.transform(X_train)
self.assertEqual(Xt.dtype, np.float32)
# np.float64
X_train, Y_train, X_test, Y_test = get_dataset("iris", make_sparse=True)
X_train = X_train.astype(np.float64)
configuration_space = SelectPercentileClassification.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
preprocessor = SelectPercentileClassification(random_state=1,
**{hp_name: default[hp_name]
for hp_name in default})
preprocessor.fit(X_train, Y_train)
Xt = preprocessor.transform(X_train)
self.assertEqual(Xt.dtype, np.float64)
def test_predict_batched_sparse(self):
cs = SimpleClassificationPipeline.get_hyperparameter_search_space(
dataset_properties={'sparse': True})
config = Configuration(cs,
values={"balancing:strategy": "none",
"classifier:__choice__": "random_forest",
"imputation:strategy": "mean",
"one_hot_encoding:minimum_fraction": 0.01,
"one_hot_encoding:use_minimum_fraction": "True",
"preprocessor:__choice__": "no_preprocessing",
'classifier:random_forest:bootstrap': 'True',
'classifier:random_forest:criterion': 'gini',
'classifier:random_forest:max_depth': 'None',
'classifier:random_forest:min_samples_split': 2,
'classifier:random_forest:min_samples_leaf': 2,
'classifier:random_forest:max_features': 0.5,
'classifier:random_forest:max_leaf_nodes': 'None',
'classifier:random_forest:n_estimators': 100,
'classifier:random_forest:min_weight_fraction_leaf': 0.0,
"rescaling:__choice__": "min/max"})
cls = SimpleClassificationPipeline(config)
# Multiclass
X_train, Y_train, X_test, Y_test = get_dataset(dataset='digits',
make_sparse=True)
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = cls.predict(X_test_)
cls_predict = mock.Mock(wraps=cls.pipeline_)
cls.pipeline_ = cls_predict
prediction = cls.predict(X_test, batch_size=20)
self.assertEqual((1647,), prediction.shape)
self.assertEqual(83, cls_predict.predict.call_count)
assert_array_almost_equal(prediction_, prediction)
# Multilabel
X_train, Y_train, X_test, Y_test = get_dataset(dataset='digits',
make_sparse=True)
Y_train = np.array([(y, 26 - y) for y in Y_train])
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = cls.predict(X_test_)
cls_predict = mock.Mock(wraps=cls.pipeline_)
cls.pipeline_ = cls_predict
prediction = cls.predict(X_test, batch_size=20)
self.assertEqual((1647, 2), prediction.shape)
self.assertEqual(83, cls_predict.predict.call_count)
assert_array_almost_equal(prediction_, prediction)
def test_default_configuration(self):
configuration_space = RidgeRegression.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
configuration_space_preproc = RandomKitchenSinks.get_hyperparameter_search_space()
default_preproc = configuration_space_preproc.get_default_configuration()
for i in range(10):
# This should be a bad results
predictions, targets = _test_regressor(RidgeRegression,)
self.assertAlmostEqual(0.32614416980439365,
sklearn.metrics.r2_score(y_true=targets, y_pred=predictions))
# This should be much more better
X_train, Y_train, X_test, Y_test = get_dataset(dataset='diabetes',
make_sparse=False)
preprocessor = RandomKitchenSinks(
random_state=1,
**{hp_name: default_preproc[hp_name] for hp_name in
default_preproc if default_preproc[hp_name] is not None})
transformer = preprocessor.fit(X_train, Y_train)
X_train_transformed = transformer.transform(X_train)
X_test_transformed = transformer.transform(X_test)
regressor = RidgeRegression(
random_state=1,
**{hp_name: default[hp_name] for hp_name in
default if default[hp_name] is not None})
predictor = regressor.fit(X_train_transformed, Y_train)
predictions = predictor.predict(X_test_transformed)
self.assertAlmostEqual(0.37183512452087852,
sklearn.metrics.r2_score(y_true=Y_test, y_pred=predictions))
def test_default_configuration(self):
transformation, original = _test_preprocessing(SelectPercentileClassification)
self.assertEqual(transformation.shape[0], original.shape[0])
self.assertEqual(transformation.shape[1], int(original.shape[1]/2))
self.assertFalse((transformation == 0).all())
transformation, original = _test_preprocessing(SelectPercentileClassification, make_sparse=True)
self.assertTrue(scipy.sparse.issparse(transformation))
self.assertEqual(transformation.shape[0], original.shape[0])
self.assertEqual(transformation.shape[1], int(original.shape[1]/2))
# Custon preprocessing test to check if clipping to zero works
X_train, Y_train, X_test, Y_test = get_dataset(dataset='digits')
original_X_train = X_train.copy()
ss = sklearn.preprocessing.StandardScaler()
X_train = ss.fit_transform(X_train)
configuration_space = SelectPercentileClassification.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
preprocessor = SelectPercentileClassification(random_state=1,
**{hp_name: default[hp_name] for hp_name in
default if default[hp_name] is not None})
transformer = preprocessor.fit(X_train, Y_train)
transformation, original = transformer.transform(X_train), original_X_train
self.assertEqual(transformation.shape[0], original.shape[0])
self.assertEqual(transformation.shape[1], int(original.shape[1] / 2))
def test_cv_results(self):
# TODO restructure and actually use real SMAC output from a long run
# to do this unittest!
tmp = os.path.join(self.test_dir, '..', '.tmp_cv_results')
output = os.path.join(self.test_dir, '..', '.out_cv_results')
self._setUp(tmp)
self._setUp(output)
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
cls = AutoSklearnClassifier(time_left_for_this_task=20,
per_run_time_limit=5,
output_folder=output,
tmp_folder=tmp,
shared_mode=False,
seed=1,
initial_configurations_via_metalearning=0,
ensemble_size=0)
cls.fit(X_train, Y_train)
cv_results = cls.cv_results_
self.assertIsInstance(cv_results, dict)
self.assertIsInstance(cv_results['mean_test_score'], np.ndarray)
self.assertIsInstance(cv_results['mean_fit_time'], np.ndarray)
self.assertIsInstance(cv_results['params'], list)
self.assertIsInstance(cv_results['rank_test_scores'], np.ndarray)
self.assertTrue([isinstance(val, npma.MaskedArray) for key, val in
cv_results.items() if key.startswith('param_')])
del cls
self._tearDown(tmp)
self._tearDown(output)
def test_configurations_signed_data(self):
# Use a limit of ~4GiB
limit = 4000 * 1024 * 1024
resource.setrlimit(resource.RLIMIT_AS, (limit, limit))
cs = SimpleClassificationPipeline.get_hyperparameter_search_space(
dataset_properties={'signed': True})
print(cs)
for i in range(10):
config = cs.sample_configuration()
config._populate_values()
if config['classifier:passive_aggressive:n_iter'] is not None:
config._values['classifier:passive_aggressive:n_iter'] = 5
if config['classifier:sgd:n_iter'] is not None:
config._values['classifier:sgd:n_iter'] = 5
X_train, Y_train, X_test, Y_test = get_dataset(dataset='digits')
cls = SimpleClassificationPipeline(config, random_state=1)
print(config)
try:
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
predictions = cls.predict(X_test)
self.assertIsInstance(predictions, np.ndarray)
predicted_probabiliets = cls.predict_proba(X_test_)
self.assertIsInstance(predicted_probabiliets, np.ndarray)
except ValueError as e:
if "Floating-point under-/overflow occurred at epoch" in \
e.args[0] or \
"removed all features" in e.args[0] or \
"all features are discarded" in e.args[0]:
continue
else:
print(config)
print(traceback.format_exc())
raise e
except RuntimeWarning as e:
if "invalid value encountered in sqrt" in e.args[0]:
continue
elif "divide by zero encountered in" in e.args[0]:
continue
elif "invalid value encountered in divide" in e.args[0]:
continue
elif "invalid value encountered in true_divide" in e.args[0]:
continue
else:
print(config)
print(traceback.format_exc())
raise e
except UserWarning as e:
if "FastICA did not converge" in e.args[0]:
continue
else:
print(config)
print(traceback.format_exc())
raise e
except MemoryError as e:
continue
def test_default_configuration(self):
transformation, original = _test_preprocessing(SelectRates)
self.assertEqual(transformation.shape[0], original.shape[0])
self.assertEqual(transformation.shape[1], 3)
self.assertFalse((transformation == 0).all())
transformation, original = _test_preprocessing(SelectRates, make_sparse=True)
self.assertTrue(scipy.sparse.issparse(transformation))
self.assertEqual(transformation.shape[0], original.shape[0])
self.assertEqual(transformation.shape[1], int(original.shape[1] / 2))
# Custon preprocessing test to check if clipping to zero works
X_train, Y_train, X_test, Y_test = get_dataset(dataset="digits")
original_X_train = X_train.copy()
ss = sklearn.preprocessing.StandardScaler()
X_train = ss.fit_transform(X_train)
configuration_space = SelectRates.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
preprocessor = SelectRates(
random_state=1, **{hp_name: default[hp_name] for hp_name in default if default[hp_name] is not None}
)
transformer = preprocessor.fit(X_train, Y_train)
transformation, original = transformer.transform(X_train), original_X_train
self.assertEqual(transformation.shape[0], original.shape[0])
# I don't know why its 52 here and not 32 which would be half of the
# number of features. Seems to be related to a runtime warning raised
# by sklearn
self.assertEqual(transformation.shape[1], 52)
def get_regression_datamanager():
X_train, Y_train, X_test, Y_test = get_dataset('boston')
indices = list(range(X_train.shape[0]))
np.random.seed(1)
np.random.shuffle(indices)
X_train = X_train[indices]
Y_train = Y_train[indices]
X_valid = X_test[:200, ]
Y_valid = Y_test[:200, ]
X_test = X_test[200:, ]
Y_test = Y_test[200:, ]
D = Dummy()
D.info = {
'metric': R2_METRIC,
'task': REGRESSION,
'is_sparse': False,
'label_num': 1
}
D.data = {
'X_train': X_train,
'Y_train': Y_train.reshape((-1, 1)),
'X_valid': X_valid,
'Y_valid': Y_valid.reshape((-1, 1)),
'X_test': X_test,
'Y_test': Y_test.reshape((-1, 1))
}
D.feat_type = ['numerical', 'Numerical', 'numerical', 'numerical',
'numerical', 'numerical', 'numerical', 'numerical',
'numerical', 'numerical', 'numerical']
return D
def setUp(self):
self.X_train, self.Y_train, self.X_test, self.Y_test = \
get_dataset('iris')
eliminate_class_two = self.Y_train != 2
self.X_train = self.X_train[eliminate_class_two]
self.Y_train = self.Y_train[eliminate_class_two]
def get_multiclass_classification_datamanager():
X_train, Y_train, X_test, Y_test = get_dataset('iris')
indices = list(range(X_train.shape[0]))
np.random.seed(1)
np.random.shuffle(indices)
X_train = X_train[indices]
Y_train = Y_train[indices]
X_valid = X_test[:25, ]
Y_valid = Y_test[:25, ]
X_test = X_test[25:, ]
Y_test = Y_test[25:, ]
D = Dummy()
D.info = {
'metric': BAC_METRIC,
'task': MULTICLASS_CLASSIFICATION,
'is_sparse': False,
'label_num': 3
}
D.data = {
'X_train': X_train,
'Y_train': Y_train,
'X_valid': X_valid,
'Y_valid': Y_valid,
'X_test': X_test,
'Y_test': Y_test
}
D.feat_type = ['numerical', 'Numerical', 'numerical', 'numerical']
return D
def test_fit_pSMAC(self):
output = os.path.join(self.test_dir, '..', '.tmp_estimator_fit_pSMAC')
self._setUp(output)
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
automl = AutoSklearnClassifier(time_left_for_this_task=15,
per_run_time_limit=15,
output_folder=output,
tmp_folder=output,
shared_mode=True,
seed=1,
initial_configurations_via_metalearning=0,
ensemble_size=0)
automl.fit(X_train, Y_train)
# Create a 'dummy model' for the first run, which has an accuracy of
# more than 99%; it should be in the final ensemble if the ensemble
# building of the second AutoSklearn classifier works correct
true_targets_ensemble_path = os.path.join(output, '.auto-sklearn',
'true_targets_ensemble.npy')
true_targets_ensemble = np.load(true_targets_ensemble_path)
true_targets_ensemble[-1] = 1 if true_targets_ensemble[-1] != 1 else 0
probas = np.zeros((len(true_targets_ensemble), 3), dtype=float)
for i, value in enumerate(true_targets_ensemble):
probas[i, value] = 1.0
dummy_predictions_path = os.path.join(output, '.auto-sklearn',
'predictions_ensemble',
'predictions_ensemble_1_00030.npy')
with open(dummy_predictions_path, 'wb') as fh:
np.save(fh, probas)
probas_test = np.zeros((len(Y_test), 3), dtype=float)
for i, value in enumerate(Y_test):
probas_test[i, value] = 1.0
dummy = ArrayReturningDummyPredictor(probas_test)
backend = Backend(output, output)
backend.save_model(dummy, 30, 1)
automl = AutoSklearnClassifier(time_left_for_this_task=15,
per_run_time_limit=15,
output_folder=output,
tmp_folder=output,
shared_mode=True,
seed=2,
initial_configurations_via_metalearning=0,
ensemble_size=0)
automl.fit(X_train, Y_train)
automl.run_ensemble_builder(0, 1, 50).wait()
score = automl.score(X_test, Y_test)
self.assertEqual(len(os.listdir(os.path.join(output, '.auto-sklearn',
'ensembles'))), 1)
self.assertGreaterEqual(score, 0.90)
self.assertEqual(automl._task, MULTICLASS_CLASSIFICATION)
del automl
self._tearDown(output)
def test_default_configuration(self):
for i in range(2):
X_train, Y_train, X_test, Y_test = get_dataset(dataset='iris')
auto = SimpleClassificationPipeline()
auto = auto.fit(X_train, Y_train)
predictions = auto.predict(X_test)
self.assertAlmostEqual(0.94,
sklearn.metrics.accuracy_score(predictions, Y_test))
scores = auto.predict_proba(X_test)
def test_default_configuration_iterative_fit(self):
classifier = SimpleClassificationPipeline(
include={'classifier': ['random_forest'],
'preprocessor': ['no_preprocessing']})
X_train, Y_train, X_test, Y_test = get_dataset(dataset='iris')
XT = classifier.fit_transformer(X_train, Y_train)
for i in range(1, 11):
classifier.iterative_fit(X_train, Y_train)
self.assertEqual(classifier.steps[-1][-1].choice.estimator.n_estimators,
i)
def test_default_configuration_multilabel(self):
for i in range(2):
cs = SimpleClassificationPipeline.get_hyperparameter_search_space(dataset_properties={"multilabel": True})
default = cs.get_default_configuration()
X_train, Y_train, X_test, Y_test = get_dataset(dataset="iris", make_multilabel=True)
auto = SimpleClassificationPipeline(default)
auto = auto.fit(X_train, Y_train)
predictions = auto.predict(X_test)
self.assertAlmostEqual(0.9599999999999995, sklearn.metrics.accuracy_score(predictions, Y_test))
scores = auto.predict_proba(X_test)
def test_configurations_sparse(self):
# Use a limit of ~4GiB
limit = 4000 * 1024 * 1024
resource.setrlimit(resource.RLIMIT_AS, (limit, limit))
cs = SimpleRegressionPipeline.get_hyperparameter_search_space(
dataset_properties={'sparse': True})
print(cs)
for i in range(10):
config = cs.sample_configuration()
config._populate_values()
if 'classifier:passive_aggressive:n_iter' in config and \
config[
'classifier:passive_aggressive:n_iter'] is not None:
config._values['classifier:passive_aggressive:n_iter'] = 5
if 'classifier:sgd:n_iter' in config and \
config['classifier:sgd:n_iter'] is not None:
config._values['classifier:sgd:n_iter'] = 5
print(config)
X_train, Y_train, X_test, Y_test = get_dataset(dataset='boston',
make_sparse=True)
cls = SimpleRegressionPipeline(config, random_state=1)
try:
cls.fit(X_train, Y_train)
predictions = cls.predict(X_test)
except ValueError as e:
if "Floating-point under-/overflow occurred at epoch" in \
e.args[0] or \
"removed all features" in e.args[0] or \
"all features are discarded" in e.args[0]:
continue
else:
print(config)
traceback.print_tb(sys.exc_info()[2])
raise e
except RuntimeWarning as e:
if "invalid value encountered in sqrt" in e.args[0]:
continue
elif "divide by zero encountered in" in e.args[0]:
continue
elif "invalid value encountered in divide" in e.args[0]:
continue
elif "invalid value encountered in true_divide" in e.args[0]:
continue
else:
print(config)
raise e
except UserWarning as e:
if "FastICA did not converge" in e.args[0]:
continue
else:
print(config)
raise e
def test_default_configuration(self):
for i in range(2):
X_train, Y_train, X_test, Y_test = get_dataset(dataset='diabetes')
auto = SimpleRegressionPipeline()
auto = auto.fit(X_train, Y_train)
predictions = auto.predict(copy.deepcopy(X_test))
# The lower the worse
r2_score = sklearn.metrics.r2_score(Y_test, predictions)
self.assertAlmostEqual(0.339, r2_score, places=3)
model_score = auto.score(copy.deepcopy(X_test), Y_test)
self.assertAlmostEqual(model_score, r2_score, places=5)
def test_default_configuration(self):
for i in range(2):
cs = SimpleRegressionPipeline.get_hyperparameter_search_space()
default = cs.get_default_configuration()
X_train, Y_train, X_test, Y_test = get_dataset(dataset='diabetes')
auto = SimpleRegressionPipeline(default)
auto = auto.fit(X_train, Y_train)
predictions = auto.predict(copy.deepcopy(X_test))
# The lower the worse
r2_score = sklearn.metrics.r2_score(Y_test, predictions)
self.assertAlmostEqual(0.41732302035060087, r2_score)
model_score = auto.score(copy.deepcopy(X_test), Y_test)
self.assertEqual(model_score, r2_score)
def _test_helper(self, Preprocessor, dataset=None, make_sparse=False):
X_train, Y_train, X_test, Y_test = get_dataset(dataset=dataset,
make_sparse=make_sparse)
original_X_train = X_train.copy()
configuration_space = Preprocessor.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
preprocessor = Preprocessor(random_state=1,
**{hp_name: default[hp_name] for hp_name in
default if default[hp_name] is not None})
preprocessor = preprocessor.choice
transformer = preprocessor.fit(X_train, Y_train)
return transformer.transform(X_train), original_X_train
def test_fit(self):
output = os.path.join(self.test_dir, '..', '.tmp_test_fit')
self._setUp(output)
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
backend_api = backend.create(output, output)
automl = autosklearn.automl.AutoML(backend_api, 15, 5)
automl.fit(X_train, Y_train)
score = automl.score(X_test, Y_test)
self.assertGreaterEqual(score, 0.8)
self.assertEqual(automl._task, MULTICLASS_CLASSIFICATION)
del automl
self._tearDown(output)
def test_regression_methods_returns_self(self):
X_train, y_train, X_test, y_test = putil.get_dataset('boston')
automl = AutoSklearnRegressor(time_left_for_this_task=20,
per_run_time_limit=5,
ensemble_size=0)
automl_fitted = automl.fit(X_train, y_train)
self.assertIs(automl, automl_fitted)
automl_ensemble_fitted = automl.fit_ensemble(y_train, ensemble_size=5)
self.assertIs(automl, automl_ensemble_fitted)
automl_refitted = automl.refit(X_train.copy(), y_train.copy())
self.assertIs(automl, automl_refitted)
def test_evaluate_binary_classification(self):
X_train, Y_train, X_test, Y_test = get_dataset('iris')
eliminate_class_two = Y_train != 2
X_train = X_train[eliminate_class_two]
Y_train = Y_train[eliminate_class_two]
eliminate_class_two = Y_test != 2
X_test = X_test[eliminate_class_two]
Y_test = Y_test[eliminate_class_two]
X_valid = X_test[:25, ]
Y_valid = Y_test[:25, ]
X_test = X_test[25:, ]
Y_test = Y_test[25:, ]
D = Dummy()
D.info = {
'metric': AUC_METRIC,
'task': BINARY_CLASSIFICATION,
'is_sparse': False,
'label_num': 2
}
D.data = {
'X_train': X_train,
'Y_train': Y_train,
'X_valid': X_valid,
'X_test': X_test
}
D.feat_type = ['numerical', 'Numerical', 'numerical', 'numerical']
configuration_space = get_configuration_space(
D.info,
include_estimators=['lda'],
include_preprocessors=['pca'])
err = np.zeros([N_TEST_RUNS])
for i in range(N_TEST_RUNS):
print('Evaluate configuration: %d; result:' % i)
configuration = configuration_space.sample_configuration()
D_ = copy.deepcopy(D)
evaluator = HoldoutEvaluator(D_, configuration)
if not self._fit(evaluator):
continue
err[i] = evaluator.predict()
self.assertTrue(np.isfinite(err[i]))
print(err[i])
self.assertGreaterEqual(err[i], 0.0)
def test_predict_proba_binary_classification(self):
X_train, Y_train, X_test, Y_test = get_dataset('iris')
eliminate_class_two = Y_train != 2
X_train = X_train[eliminate_class_two]
Y_train = Y_train[eliminate_class_two]
eliminate_class_two = Y_test != 2
X_test = X_test[eliminate_class_two]
Y_test = Y_test[eliminate_class_two]
X_valid = X_test[:25, ]
Y_valid = Y_test[:25, ]
X_test = X_test[25:, ]
Y_test = Y_test[25:, ]
class Dummy2(object):
def predict_proba(self, y, batch_size=200):
return np.array([[0.1, 0.9], [0.7, 0.3]])
model = Dummy2()
task_type = BINARY_CLASSIFICATION
D = Dummy()
D.info = {
'metric': BAC_METRIC,
'task': task_type,
'is_sparse': False,
'label_num': 3
}
D.data = {
'X_train': X_train,
'Y_train': Y_train,
'X_valid': X_valid,
'X_test': X_test
}
D.feat_type = ['numerical', 'Numerical', 'numerical', 'numerical']
configuration_space = get_configuration_space(
D.info,
include_estimators=['lda'],
include_preprocessors=['select_rates'])
configuration = configuration_space.sample_configuration()
evaluator = HoldoutEvaluator(D, configuration)
pred = evaluator.predict_proba(None, model, task_type)
expected = [[0.9], [0.3]]
for i in range(len(expected)):
self.assertEqual(expected[i], pred[i])
def test_evaluate_multiclass_classification_all_metrics(self):
X_train, Y_train, X_test, Y_test = get_dataset('iris')
X_valid = X_test[:25, ]
Y_valid = Y_test[:25, ]
X_test = X_test[25:, ]
Y_test = Y_test[25:, ]
D = Dummy()
D.info = {
'metric': BAC_METRIC,
'task': MULTICLASS_CLASSIFICATION,
'is_sparse': False,
'label_num': 3
}
D.data = {
'X_train': X_train,
'Y_train': Y_train,
'X_valid': X_valid,
'X_test': X_test
}
D.feat_type = ['numerical', 'Numerical', 'numerical', 'numerical']
configuration_space = get_configuration_space(
D.info, include_estimators=['lda'], include_preprocessors=['pca'])
# Test all scoring functions
err = []
for i in range(N_TEST_RUNS):
print('Evaluate configuration: %d; result:' % i)
configuration = configuration_space.sample_configuration()
D_ = copy.deepcopy(D)
evaluator = HoldoutEvaluator(D_,
configuration,
all_scoring_functions=True)
if not self._fit(evaluator):
continue
err.append(evaluator.predict())
print(err[-1])
self.assertIsInstance(err[-1], dict)
for key in err[-1]:
self.assertEqual(len(err[-1]), 5)
self.assertTrue(np.isfinite(err[-1][key]))
self.assertGreaterEqual(err[-1][key], 0.0)
def test_do_dummy_prediction(self):
datasets = {
'breast_cancer': BINARY_CLASSIFICATION,
'wine': MULTICLASS_CLASSIFICATION,
'diabetes': REGRESSION,
}
for name, task in datasets.items():
backend_api = self._create_backend('test_do_dummy_prediction')
X_train, Y_train, X_test, Y_test = putil.get_dataset(name)
datamanager = XYDataManager(
X_train, Y_train,
X_test, Y_test,
task=task,
dataset_name=name,
feat_type=None,
)
auto = autosklearn.automl.AutoML(
backend_api, 20, 5,
initial_configurations_via_metalearning=25,
metric=accuracy,
)
setup_logger()
auto._logger = get_logger('test_do_dummy_predictions')
auto._backend.save_datamanager(datamanager)
D = backend_api.load_datamanager()
# Check if data manager is correcly loaded
self.assertEqual(D.info['task'], datamanager.info['task'])
auto._do_dummy_prediction(D, 1)
# Ensure that the dummy predictions are not in the current working
# directory, but in the temporary directory.
self.assertFalse(os.path.exists(os.path.join(os.getcwd(),
'.auto-sklearn')))
self.assertTrue(os.path.exists(os.path.join(
backend_api.temporary_directory, '.auto-sklearn', 'predictions_ensemble',
'predictions_ensemble_1_1_0.0.npy')))
del auto
self._tearDown(backend_api.temporary_directory)
self._tearDown(backend_api.output_directory)
def test_evaluate_multilabel_classification(self):
X_train, Y_train, X_test, Y_test = get_dataset('iris')
Y_train = np.array(convert_to_bin(Y_train, 3))
Y_train[:, -1] = 1
Y_test = np.array(convert_to_bin(Y_test, 3))
Y_test[:, -1] = 1
X_valid = X_test[:25, ]
Y_valid = Y_test[:25, ]
X_test = X_test[25:, ]
Y_test = Y_test[25:, ]
D = Dummy()
D.info = {
'metric': F1_METRIC,
'task': MULTILABEL_CLASSIFICATION,
'is_sparse': False,
'label_num': 3
}
D.data = {
'X_train': X_train,
'Y_train': Y_train,
'X_valid': X_valid,
'X_test': X_test
}
D.feat_type = ['numerical', 'Numerical', 'numerical', 'numerical']
configuration_space = get_configuration_space(
D.info,
include_estimators=['extra_trees'],
include_preprocessors=['no_preprocessing'])
err = np.zeros([N_TEST_RUNS])
for i in range(N_TEST_RUNS):
print('Evaluate configuration: %d; result:' % i)
configuration = configuration_space.sample_configuration()
D_ = copy.deepcopy(D)
evaluator = HoldoutEvaluator(D_, configuration)
if not self._fit(evaluator):
continue
err[i] = evaluator.predict()
print(err[i])
self.assertTrue(np.isfinite(err[i]))
self.assertGreaterEqual(err[i], 0.0)
def test_can_pickle_classifier(tmp_dir, output_dir, dask_client):
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
automl = AutoSklearnClassifier(time_left_for_this_task=30,
per_run_time_limit=5,
tmp_folder=tmp_dir,
dask_client=dask_client,
output_folder=output_dir)
automl.fit(X_train, Y_train)
initial_predictions = automl.predict(X_test)
initial_accuracy = sklearn.metrics.accuracy_score(Y_test,
initial_predictions)
assert initial_accuracy >= 0.75
assert count_succeses(automl.cv_results_) > 0
assert includes_train_scores(automl.performance_over_time_.columns) is True
assert performance_over_time_is_plausible(
automl.performance_over_time_) is True
# Test pickle
dump_file = os.path.join(output_dir, 'automl.dump.pkl')
with open(dump_file, 'wb') as f:
pickle.dump(automl, f)
with open(dump_file, 'rb') as f:
restored_automl = pickle.load(f)
restored_predictions = restored_automl.predict(X_test)
restored_accuracy = sklearn.metrics.accuracy_score(Y_test,
restored_predictions)
assert restored_accuracy >= 0.75
assert initial_accuracy == restored_accuracy
# Test joblib
dump_file = os.path.join(output_dir, 'automl.dump.joblib')
joblib.dump(automl, dump_file)
restored_automl = joblib.load(dump_file)
restored_predictions = restored_automl.predict(X_test)
restored_accuracy = sklearn.metrics.accuracy_score(Y_test,
restored_predictions)
assert restored_accuracy >= 0.75
assert initial_accuracy == restored_accuracy
def test_evaluate_regression(self):
X_train, Y_train, X_test, Y_test = get_dataset('boston')
X_valid = X_test[:200, ]
Y_valid = Y_test[:200, ]
X_test = X_test[200:, ]
Y_test = Y_test[200:, ]
D = Dummy()
D.info = {
'metric': R2_METRIC,
'task': REGRESSION,
'is_sparse': False,
'label_num': 1
}
D.data = {
'X_train': X_train,
'Y_train': Y_train,
'X_valid': X_valid,
'X_test': X_test
}
D.feat_type = [
'numerical', 'Numerical', 'numerical', 'numerical', 'numerical',
'numerical', 'numerical', 'numerical', 'numerical', 'numerical',
'numerical'
]
configuration_space = get_configuration_space(
D.info,
include_estimators=['extra_trees'],
include_preprocessors=['no_preprocessing'])
err = np.zeros([N_TEST_RUNS])
for i in range(N_TEST_RUNS):
print('Evaluate configuration: %d; result:' % i)
configuration = configuration_space.sample_configuration()
D_ = copy.deepcopy(D)
evaluator = HoldoutEvaluator(D_, configuration)
if not self._fit(evaluator):
continue
err[i] = evaluator.predict()
self.assertTrue(np.isfinite(err[i]))
print(err[i])
self.assertGreaterEqual(err[i], 0.0)
def test_delete_non_candidate_models(self):
backend_api = self._create_backend(
'test_delete', delete_tmp_folder_after_terminate=False)
seed = 555
X, Y, _, _ = putil.get_dataset('iris')
automl = autosklearn.automl.AutoML(
backend_api,
time_left_for_this_task=30,
per_run_time_limit=5,
ensemble_nbest=3,
seed=seed,
initial_configurations_via_metalearning=0,
resampling_strategy='holdout',
include_estimators=['sgd'],
include_preprocessors=['no_preprocessing'],
metric=accuracy,
)
automl.fit(X, Y, task=MULTICLASS_CLASSIFICATION,
X_test=X, y_test=Y)
# Assert at least one model file has been deleted and that there were no
# deletion errors
log_file_path = glob.glob(os.path.join(
backend_api.temporary_directory, 'AutoML(' + str(seed) + '):*.log'))
with open(log_file_path[0]) as log_file:
log_content = log_file.read()
self.assertIn('Deleted files of non-candidate model', log_content)
self.assertNotIn('Failed to delete files of non-candidate model', log_content)
self.assertNotIn('Failed to lock model', log_content)
# Assert that the files of the models used by the ensemble weren't deleted
model_files = backend_api.list_all_models(seed=seed)
model_files_idx = set()
for m_file in model_files:
# Extract the model identifiers from the filename
m_file = os.path.split(m_file)[1].replace('.model', '').split('.', 2)
model_files_idx.add((int(m_file[0]), int(m_file[1]), float(m_file[2])))
ensemble_members_idx = set(automl.ensemble_.identifiers_)
self.assertTrue(ensemble_members_idx.issubset(model_files_idx))
del automl
self._tearDown(backend_api.temporary_directory)
self._tearDown(backend_api.output_directory)
def test_default_configuration_predict_proba_individual(self):
# Leave this additional test here
for i in range(2):
predictions, targets = _test_classifier_predict_proba(
LibSVM_SVC,
sparse=True,
dataset='digits',
train_size_maximum=500)
self.assertAlmostEqual(
5.4706296711768925,
sklearn.metrics.log_loss(targets, predictions))
for i in range(2):
predictions, targets = _test_classifier_predict_proba(
LibSVM_SVC, sparse=True, dataset='iris')
self.assertAlmostEqual(
0.84333924656905945,
sklearn.metrics.log_loss(targets, predictions))
# 2 class
for i in range(2):
X_train, Y_train, X_test, Y_test = get_dataset(dataset='iris')
remove_training_data = Y_train == 2
remove_test_data = Y_test == 2
X_train = X_train[~remove_training_data]
Y_train = Y_train[~remove_training_data]
X_test = X_test[~remove_test_data]
Y_test = Y_test[~remove_test_data]
ss = sklearn.preprocessing.StandardScaler()
X_train = ss.fit_transform(X_train)
configuration_space = LibSVM_SVC.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
cls = LibSVM_SVC(random_state=1,
**{
hp_name: default[hp_name]
for hp_name in default
if default[hp_name] is not None
})
cls = cls.fit(X_train, Y_train)
prediction = cls.predict_proba(X_test)
self.assertAlmostEqual(
sklearn.metrics.log_loss(Y_test, prediction),
0.69323680119641773)
def test_do_dummy_prediction(dask_client, datasets):
name, task = datasets
X_train, Y_train, X_test, Y_test = putil.get_dataset(name)
datamanager = XYDataManager(
X_train, Y_train,
X_test, Y_test,
task=task,
dataset_name=name,
feat_type={i: 'numerical' for i in range(X_train.shape[1])},
)
auto = autosklearn.automl.AutoML(
20, 5,
initial_configurations_via_metalearning=25,
metric=accuracy,
dask_client=dask_client,
delete_tmp_folder_after_terminate=False,
)
auto._backend = auto._create_backend()
# Make a dummy logger
auto._logger_port = 9020
auto._logger = unittest.mock.Mock()
auto._logger.info.return_value = None
auto._backend.save_datamanager(datamanager)
D = auto._backend.load_datamanager()
# Check if data manager is correcly loaded
assert D.info['task'] == datamanager.info['task']
auto._do_dummy_prediction(D, 1)
# Ensure that the dummy predictions are not in the current working
# directory, but in the temporary directory.
assert not os.path.exists(os.path.join(os.getcwd(), '.auto-sklearn'))
assert os.path.exists(os.path.join(
auto._backend.temporary_directory, '.auto-sklearn', 'runs', '1_1_0.0',
'predictions_ensemble_1_1_0.0.npy')
)
auto._clean_logger()
del auto
def test_can_pickle_classifier(self):
output = os.path.join(self.test_dir, '..', '.tmp_can_pickle')
self._setUp(output)
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
automl = AutoSklearnClassifier(time_left_for_this_task=20,
per_run_time_limit=5,
tmp_folder=output,
output_folder=output)
automl.fit(X_train, Y_train)
initial_predictions = automl.predict(X_test)
initial_accuracy = sklearn.metrics.accuracy_score(
Y_test, initial_predictions)
self.assertGreaterEqual(initial_accuracy, 0.75)
# Test pickle
dump_file = os.path.join(output, 'automl.dump.pkl')
with open(dump_file, 'wb') as f:
pickle.dump(automl, f)
with open(dump_file, 'rb') as f:
restored_automl = pickle.load(f)
restored_predictions = restored_automl.predict(X_test)
restored_accuracy = sklearn.metrics.accuracy_score(
Y_test, restored_predictions)
self.assertGreaterEqual(restored_accuracy, 0.75)
self.assertEqual(initial_accuracy, restored_accuracy)
# Test joblib
dump_file = os.path.join(output, 'automl.dump.joblib')
sklearn.externals.joblib.dump(automl, dump_file)
restored_automl = sklearn.externals.joblib.load(dump_file)
restored_predictions = restored_automl.predict(X_test)
restored_accuracy = sklearn.metrics.accuracy_score(
Y_test, restored_predictions)
self.assertGreaterEqual(restored_accuracy, 0.75)
self.assertEqual(initial_accuracy, restored_accuracy)
def test_delete_non_candidate_models(backend, dask_client):
seed = 555
X, Y, _, _ = putil.get_dataset('iris')
automl = autosklearn.automl.AutoML(
backend,
time_left_for_this_task=60,
per_run_time_limit=5,
ensemble_nbest=3,
seed=seed,
initial_configurations_via_metalearning=0,
resampling_strategy='holdout',
include_estimators=['sgd'],
include_preprocessors=['no_preprocessing'],
metric=accuracy,
dask_client=dask_client,
# Force model to be deleted. That is, from 50 which is the
# default to 3 to make sure we delete models.
max_models_on_disc=3,
)
automl.fit(X, Y, task=MULTICLASS_CLASSIFICATION,
X_test=X, y_test=Y)
# Assert at least one model file has been deleted and that there were no
# deletion errors
log_file_path = glob.glob(os.path.join(
backend.temporary_directory, 'AutoML(' + str(seed) + '):*.log'))
with open(log_file_path[0]) as log_file:
log_content = log_file.read()
assert 'Deleted files of non-candidate model' in log_content, log_content
assert 'Failed to delete files of non-candidate model' not in log_content, log_content
assert 'Failed to lock model' not in log_content, log_content
# Assert that the files of the models used by the ensemble weren't deleted
model_files = backend.list_all_models(seed=seed)
model_files_idx = set()
for m_file in model_files:
# Extract the model identifiers from the filename
m_file = os.path.split(m_file)[1].replace('.model', '').split('.', 2)
model_files_idx.add((int(m_file[0]), int(m_file[1]), float(m_file[2])))
ensemble_members_idx = set(automl.ensemble_.identifiers_)
assert ensemble_members_idx.issubset(model_files_idx), (ensemble_members_idx, model_files_idx)
del automl
def test_fit(self):
backend_api = self._create_backend('test_fit')
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
automl = autosklearn.automl.AutoML(backend_api, 20, 5)
automl.fit(
X_train,
Y_train,
metric=accuracy,
task=MULTICLASS_CLASSIFICATION,
)
score = automl.score(X_test, Y_test)
self.assertGreaterEqual(score, 0.8)
self.assertEqual(automl._task, MULTICLASS_CLASSIFICATION)
del automl
self._tearDown(backend_api.temporary_directory)
self._tearDown(backend_api.output_directory)
def test_regression(self):
tmp = os.path.join(self.test_dir, '..', '.tmp_regression_fit')
output = os.path.join(self.test_dir, '..', '.out_regression_fit')
self._setUp(tmp)
self._setUp(output)
X_train, Y_train, X_test, Y_test = putil.get_dataset('boston')
automl = AutoSklearnRegressor(time_left_for_this_task=20,
per_run_time_limit=5,
tmp_folder=tmp,
output_folder=output)
automl.fit(X_train, Y_train)
predictions = automl.predict(X_test)
self.assertEqual(predictions.shape, (356, ))
score = mean_squared_error(Y_test, predictions)
# On average np.sqrt(30) away from the target -> ~5.5 on average
self.assertGreaterEqual(score, -30)
def test_produce_zero_scaling(self):
from autosklearn.pipeline.classification import SimpleClassificationPipeline
from autosklearn.pipeline import util as putil
p = SimpleClassificationPipeline(configuration={
'balancing:strategy': 'weighting',
'classifier:__choice__': 'qda',
'classifier:qda:reg_param': 2.992955287687101,
'imputation:strategy': 'most_frequent',
'one_hot_encoding:use_minimum_fraction': 'False',
'preprocessor:__choice__': 'gem',
'preprocessor:gem:N': 18,
'preprocessor:gem:precond': 0.12360249797270745,
'rescaling:__choice__': 'none'})
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
self.assertRaisesRegexp(ValueError, 'Numerical problems in '
'QDA. QDA.scalings_ contains '
'values <= 0.0',
p.fit, X_train, Y_train)
def test_fit(self):
output = os.path.join(self.test_dir, '..', '.tmp_estimator_fit')
self._setUp(output)
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
automl = AutoSklearnClassifier(time_left_for_this_task=15,
per_run_time_limit=5,
tmp_folder=output,
output_folder=output)
automl.fit(X_train, Y_train)
score = automl.score(X_test, Y_test)
self.assertGreaterEqual(score, 0.8)
self.assertEqual(automl._automl._automl._task, MULTICLASS_CLASSIFICATION)
del automl
self._tearDown(output)
def test_multilabel(self):
output = os.path.join(self.test_dir, '..', '.tmp_multilabel_fit')
self._setUp(output)
X_train, Y_train, X_test, Y_test = putil.get_dataset(
'iris', make_multilabel=True)
automl = AutoSklearnClassifier(time_left_for_this_task=20,
per_run_time_limit=5,
tmp_folder=output,
output_folder=output)
automl.fit(X_train, Y_train)
predictions = automl.predict(X_test)
self.assertEqual(predictions.shape, (50, 3))
score = f1_macro(Y_test, predictions)
self.assertGreaterEqual(score, 0.9)
probs = automl.predict_proba(X_train)
self.assertAlmostEqual(np.mean(probs), 0.33333333333333331)
def test_predict_batched_sparse(self):
cs = SimpleRegressionPipeline.get_hyperparameter_search_space(
dataset_properties={'sparse': True},
include={'regressor': ['decision_tree']})
default = cs.get_default_configuration()
cls = SimpleRegressionPipeline(default)
X_train, Y_train, X_test, Y_test = get_dataset(dataset='diabetes',
make_sparse=True)
cls.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = cls.predict(X_test_)
cls_predict = unittest.mock.Mock(wraps=cls.pipeline_)
cls.pipeline_ = cls_predict
prediction = cls.predict(X_test, batch_size=20)
self.assertEqual((292, ), prediction.shape)
self.assertEqual(15, cls_predict.predict.call_count)
assert_array_almost_equal(prediction_, prediction)
def test_predict_batched(self):
include = {'regressor': ['decision_tree']}
cs = SimpleRegressionPipeline(
include=include).get_hyperparameter_search_space()
default = cs.get_default_configuration()
regressor = SimpleRegressionPipeline(default, include=include)
X_train, Y_train, X_test, Y_test = get_dataset(dataset='boston')
regressor.fit(X_train, Y_train)
X_test_ = X_test.copy()
prediction_ = regressor.predict(X_test_)
mock_predict = unittest.mock.Mock(
wraps=regressor.steps[-1][-1].predict)
regressor.steps[-1][-1].predict = mock_predict
prediction = regressor.predict(X_test, batch_size=20)
self.assertEqual((356, ), prediction.shape)
self.assertEqual(18, mock_predict.call_count)
np.testing.assert_array_almost_equal(prediction_, prediction)
def test_trials_callback_execution(self):
trials_summary_fname = os.path.join(tempfile.gettempdir(),
"trials.csv")
X_train, Y_train, X_test, Y_test = putil.get_dataset('breast_cancer')
cls = AutoSklearnClassifier(
time_left_for_this_task=30,
initial_configurations_via_metalearning=0,
per_run_time_limit=10,
memory_limit=1024,
delete_tmp_folder_after_terminate=False,
n_jobs=1,
include_estimators=["sgd"],
include_preprocessors=["no_preprocessing"],
get_trials_callback=AutoMLTrialsCallBack(trials_summary_fname))
cls.fit(X_train, Y_train, X_test, Y_test)
trials = pd.read_csv(trials_summary_fname)
assert trials.shape[
0] > 0, f"Auto-Sklearn explored {trials.shape[0] - 1} trials"
def test_fit_roar(self):
output = os.path.join(self.test_dir, '..', '.tmp_test_fit_roar')
self._setUp(output)
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
backend_api = backend.create(output, output)
automl = autosklearn.automl.AutoML(backend_api, 20, 5,
initial_configurations_via_metalearning=0,
configuration_mode='ROAR')
automl.fit(X_train, Y_train, metric=accuracy)
# print(automl.show_models(), flush=True)
# print(automl.cv_results_, flush=True)
score = automl.score(X_test, Y_test)
self.assertGreaterEqual(score, 0.8)
self.assertEqual(automl._task, MULTICLASS_CLASSIFICATION)
del automl
self._tearDown(output)
def test_fit(self):
if self.travis:
self.skipTest('This test does currently not run on travis-ci. '
'Make sure it runs locally on your machine!')
output = os.path.join(self.test_dir, '..', '.tmp_test_fit')
self._setUp(output)
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
backend_api = backend.create(output, output)
automl = autosklearn.automl.AutoML(backend_api, 15, 15)
automl.fit(X_train, Y_train)
score = automl.score(X_test, Y_test)
self.assertGreaterEqual(score, 0.8)
self.assertEqual(automl._task, MULTICLASS_CLASSIFICATION)
del automl
self._tearDown(output)
def _test_helper(self, Preprocessor, dataset=None, make_sparse=False):
X_train, Y_train, X_test, Y_test = get_dataset(
dataset=dataset,
make_sparse=make_sparse,
)
dataset_properties = {'sparse': make_sparse}
original_X_train = X_train.copy()
configuration_space = Preprocessor(dataset_properties).\
get_hyperparameter_search_space(dataset_properties)
default = configuration_space.get_default_configuration()
preprocessor = Preprocessor(dataset_properties, random_state=1)
preprocessor.set_hyperparameters(default)
preprocessor = preprocessor.choice
transformer = preprocessor.fit(X_train, Y_train)
return transformer.transform(X_train), original_X_train
def test_default_configuration_predict_proba(self):
for i in range(10):
predictions, targets = _test_classifier_predict_proba(
LibSVM_SVC,
sparse=True,
dataset='digits',
train_size_maximum=500)
self.assertAlmostEqual(
4.6680593525563063,
sklearn.metrics.log_loss(targets, predictions))
for i in range(10):
predictions, targets = _test_classifier_predict_proba(
LibSVM_SVC, sparse=True, dataset='iris')
self.assertAlmostEqual(
0.8649665185853217,
sklearn.metrics.log_loss(targets, predictions))
# 2 class
for i in range(10):
X_train, Y_train, X_test, Y_test = get_dataset(dataset='iris')
remove_training_data = Y_train == 2
remove_test_data = Y_test == 2
X_train = X_train[~remove_training_data]
Y_train = Y_train[~remove_training_data]
X_test = X_test[~remove_test_data]
Y_test = Y_test[~remove_test_data]
ss = sklearn.preprocessing.StandardScaler()
X_train = ss.fit_transform(X_train)
configuration_space = LibSVM_SVC.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
cls = LibSVM_SVC(random_state=1,
**{
hp_name: default[hp_name]
for hp_name in default
if default[hp_name] is not None
})
cls = cls.fit(X_train, Y_train)
prediction = cls.predict_proba(X_test)
self.assertAlmostEqual(
sklearn.metrics.log_loss(Y_test, prediction),
0.69323680119641773)
def test_exceptions_inside_log_in_smbo(self, smbo_run_mock):
# Make sure that any exception during the AutoML fit due to
# SMAC are properly captured in a log file
backend_api = self._create_backend('test_exceptions_inside_log')
self._tearDown(backend_api.temporary_directory)
self._tearDown(backend_api.output_directory)
automl = autosklearn.automl.AutoML(
backend_api,
20,
5,
metric=accuracy,
)
output_file = 'test_exceptions_inside_log.log'
setup_logger(output_file=output_file)
logger = get_logger('test_exceptions_inside_log')
# Create a custom exception to prevent other errors to slip in
class MyException(Exception):
pass
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
# The first call is on dummy predictor failure
message = str(np.random.randint(100)) + '_run_smbo'
smbo_run_mock.side_effect = MyException(message)
with unittest.mock.patch(
'autosklearn.automl.AutoML._get_logger') as mock:
mock.return_value = logger
with self.assertRaises(MyException):
automl.fit(
X_train,
Y_train,
task=MULTICLASS_CLASSIFICATION,
)
with open(output_file) as f:
self.assertTrue(message in f.read())
# Cleanup
os.unlink(output_file)
self._tearDown(backend_api.temporary_directory)
self._tearDown(backend_api.output_directory)
def test_fit_roar(self):
def get_roar_object_callback(
scenario_dict,
seed,
ta,
ta_kwargs,
**kwargs
):
"""Random online adaptive racing.
http://ml.informatik.uni-freiburg.de/papers/11-LION5-SMAC.pdf"""
scenario = Scenario(scenario_dict)
return ROAR(
scenario=scenario,
rng=seed,
tae_runner=ta,
tae_runner_kwargs=ta_kwargs,
)
backend_api = self._create_backend('test_fit_roar')
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
automl = autosklearn.automl.AutoML(
backend=backend_api,
time_left_for_this_task=20,
per_run_time_limit=5,
initial_configurations_via_metalearning=0,
get_smac_object_callback=get_roar_object_callback,
metric=accuracy,
)
setup_logger()
automl._logger = get_logger('test_fit_roar')
automl.fit(
X_train, Y_train, task=MULTICLASS_CLASSIFICATION,
)
score = automl.score(X_test, Y_test)
self.assertGreaterEqual(score, 0.8)
self.assertGreater(self._count_succeses(automl.cv_results_), 0)
self.assertEqual(automl._task, MULTICLASS_CLASSIFICATION)
del automl
self._tearDown(backend_api.temporary_directory)
self._tearDown(backend_api.output_directory)
def test_default_configuration_negative_values(self):
# Custon preprocessing test to check if clipping to zero works
X_train, Y_train, X_test, Y_test = get_dataset(dataset='digits')
ss = sklearn.preprocessing.StandardScaler()
X_train = ss.fit_transform(X_train)
configuration_space = MultinomialNB.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
cls = MultinomialNB(random_state=1,
**{
hp_name: default[hp_name]
for hp_name in default
if default[hp_name] is not None
})
cls = cls.fit(X_train, Y_train)
prediction = cls.predict(X_test)
self.assertAlmostEqual(np.nanmean(prediction == Y_test),
0.88888888888888884)
def test_default_configuration_classify(self):
for i in range(2):
X_train, Y_train, X_test, Y_test = get_dataset(dataset='digits',
make_sparse=False)
configuration_space = ExtraTreesPreprocessor.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
preprocessor = ExtraTreesPreprocessor(random_state=1,
**{hp_name: default[hp_name]
for hp_name in default})
preprocessor.fit(X_train, Y_train)
X_train_trans = preprocessor.transform(X_train)
X_test_trans = preprocessor.transform(X_test)
# fit a classifier on top
classifier = RidgeClassifier()
predictor = classifier.fit(X_train_trans, Y_train)
predictions = predictor.predict(X_test_trans)
accuracy = sklearn.metrics.accuracy_score(predictions, Y_test)
self.assertAlmostEqual(accuracy, 0.87310261080752882, places=2)
def test_fit(dask_client, backend):
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
automl = autosklearn.automl.AutoML(
backend=backend,
time_left_for_this_task=30,
per_run_time_limit=5,
metric=accuracy,
dask_client=dask_client,
)
automl.fit(
X_train, Y_train, task=MULTICLASS_CLASSIFICATION,
)
score = automl.score(X_test, Y_test)
assert score > 0.8
assert count_succeses(automl.cv_results_) > 0
assert automl._task == MULTICLASS_CLASSIFICATION
del automl
def test_load_best_individual_model(metric, backend, dask_client):
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
automl = autosklearn.automl.AutoML(
backend=backend,
time_left_for_this_task=30,
per_run_time_limit=5,
metric=metric,
dask_client=dask_client,
)
# We cannot easily mock a function sent to dask
# so for this test we create the whole set of models/ensembles
# but prevent it to be loaded
automl.fit(
X_train,
Y_train,
task=MULTICLASS_CLASSIFICATION,
)
automl._backend.load_ensemble = unittest.mock.MagicMock(return_value=None)
# A memory error occurs in the ensemble construction
assert automl._backend.load_ensemble(automl._seed) is None
# The load model is robust to this and loads the best model
automl._load_models()
assert automl.ensemble_ is not None
# Just 1 model is there for ensemble and all weight must be on it
get_models_with_weights = automl.get_models_with_weights()
assert len(get_models_with_weights) == 1
assert get_models_with_weights[0][0] == 1.0
# Match a toy dataset
if metric.name == 'balanced_accuracy':
assert automl.score(X_test, Y_test) > 0.9
elif metric.name == 'log_loss':
# Seen values in github actions of 0.6978304740364537
assert automl.score(X_test, Y_test) < 0.7
else:
raise ValueError(metric.name)
del automl
def test_default_configuration_classify(self):
for i in range(3):
X_train, Y_train, X_test, Y_test = get_dataset(dataset='digits',
make_sparse=False)
configuration_space = FeatureAgglomeration.get_hyperparameter_search_space()
default = configuration_space.get_default_configuration()
preprocessor = FeatureAgglomeration(random_state=1,
**{hp_name: default[hp_name] for
hp_name in default})
preprocessor.fit(X_train, Y_train)
X_train_trans = preprocessor.transform(X_train)
X_test_trans = preprocessor.transform(X_test)
# fit a classifier on top
classifier = RandomForestClassifier(random_state=1)
predictor = classifier.fit(X_train_trans, Y_train)
predictions = predictor.predict(X_test_trans)
accuracy = sklearn.metrics.accuracy_score(predictions, Y_test)
self.assertAlmostEqual(accuracy, 0.8026715)
def get_multilabel_classification_datamanager():
X_train, Y_train, X_test, Y_test = get_dataset('iris')
indices = list(range(X_train.shape[0]))
np.random.seed(1)
np.random.shuffle(indices)
X_train = X_train[indices]
Y_train = Y_train[indices]
Y_train = np.array(convert_to_bin(Y_train, 3))
# for i in range(Y_train_.shape[0]):
# Y_train_[:, Y_train[i]] = 1
# Y_train = Y_train_
Y_test = np.array(convert_to_bin(Y_test, 3))
# for i in range(Y_test_.shape[0]):
# Y_test_[:, Y_test[i]] = 1
# Y_test = Y_test_
X_valid = X_test[:25, ]
Y_valid = Y_test[:25, ]
X_test = X_test[25:, ]
Y_test = Y_test[25:, ]
D = Dummy()
D.info = {
'task': MULTILABEL_CLASSIFICATION,
'is_sparse': False,
'label_num': 3
}
D.data = {
'X_train': X_train,
'Y_train': Y_train,
'X_valid': X_valid,
'Y_valid': Y_valid,
'X_test': X_test,
'Y_test': Y_test
}
D.feat_type = {
0: 'numerical',
1: 'Numerical',
2: 'numerical',
3: 'numerical'
}
return D
