def test_cv_regression(tmp_dir, output_dir, dask_client):
"""
Makes sure that when using a cv strategy, we are able to fit
a regressor
"""
X_train, Y_train, X_test, Y_test = putil.get_dataset(
'boston', train_size_maximum=300)
automl = AutoSklearnRegressor(time_left_for_this_task=60,
per_run_time_limit=10,
resampling_strategy='cv',
tmp_folder=tmp_dir,
dask_client=dask_client,
output_folder=output_dir)
automl.fit(X_train, Y_train)
# Log file path
log_file_path = glob.glob(os.path.join(tmp_dir, 'AutoML*.log'))[0]
predictions = automl.predict(X_test)
assert predictions.shape == (206, )
score = r2(Y_test, predictions)
assert score >= 0.1, extract_msg_from_log(log_file_path)
assert count_succeses(
automl.cv_results_) > 0, extract_msg_from_log(log_file_path)
def test_regression_pandas_support(tmp_dir, output_dir, dask_client):
X, y = sklearn.datasets.fetch_openml(
data_id=41514, # diabetes
return_X_y=True,
as_frame=True,
)
# This test only make sense if input is dataframe
assert isinstance(X, pd.DataFrame)
assert isinstance(y, pd.Series)
automl = AutoSklearnRegressor(
time_left_for_this_task=40,
per_run_time_limit=5,
dask_client=dask_client,
tmp_folder=tmp_dir,
output_folder=output_dir,
)
# Make sure we error out because y is not encoded
automl.fit(X, y)
# Make sure that at least better than random.
# We use same X_train==X_test to test code quality
assert automl.score(X, y) >= 0.5, print_debug_information(automl)
automl.refit(X, y)
# Make sure that at least better than random.
assert r2(y, automl.predict(X)) > 0.5, print_debug_information(automl)
assert count_succeses(
automl.cv_results_) > 0, print_debug_information(automl)
def test_binary(tmp_dir, output_dir, dask_client):
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris',
make_binary=True)
automl = AutoSklearnClassifier(time_left_for_this_task=40,
per_run_time_limit=10,
tmp_folder=tmp_dir,
dask_client=dask_client,
output_folder=output_dir)
automl.fit(X_train,
Y_train,
X_test=X_test,
y_test=Y_test,
dataset_name='binary_test_dataset')
predictions = automl.predict(X_test)
assert predictions.shape == (50, ), print_debug_information(automl)
score = accuracy(Y_test, predictions)
assert score > 0.9, print_debug_information(automl)
assert count_succeses(
automl.cv_results_) > 0, print_debug_information(automl)
output_files = glob.glob(
os.path.join(output_dir, 'binary_test_dataset_test_*.predict'))
assert len(output_files) > 0, (output_files,
print_debug_information(automl))
def test_multilabel(tmp_dir, output_dir, dask_client):
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris',
make_multilabel=True)
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)
predictions = automl.predict(X_test)
assert predictions.shape == (50, 3), print_debug_information(automl)
assert count_succeses(
automl.cv_results_) > 0, print_debug_information(automl)
assert includes_train_scores(automl.performance_over_time_.columns) is True
assert performance_over_time_is_plausible(
automl.performance_over_time_) is True
score = f1_macro(Y_test, predictions)
assert score >= 0.9, print_debug_information(automl)
probs = automl.predict_proba(X_train)
assert np.mean(probs) == pytest.approx(0.33, rel=1e-1)
def test_cv_regression(tmp_dir, dask_client):
"""
Makes sure that when using a cv strategy, we are able to fit
a regressor
"""
X_train, Y_train, X_test, Y_test = putil.get_dataset(
'boston', train_size_maximum=300)
automl = AutoSklearnRegressor(
time_left_for_this_task=60,
per_run_time_limit=10,
resampling_strategy='cv',
tmp_folder=tmp_dir,
dask_client=dask_client,
)
automl.fit(X_train, Y_train)
predictions = automl.predict(X_test)
assert predictions.shape == (206, )
score = r2(Y_test, predictions)
assert score >= 0.1, print_debug_information(automl)
assert count_succeses(
automl.cv_results_) > 0, print_debug_information(automl)
assert includes_train_scores(automl.performance_over_time_.columns) is True
assert performance_over_time_is_plausible(
automl.performance_over_time_) is True
def test_binary(tmp_dir, dask_client):
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris',
make_binary=True)
automl = AutoSklearnClassifier(
time_left_for_this_task=40,
delete_tmp_folder_after_terminate=False,
per_run_time_limit=10,
tmp_folder=tmp_dir,
dask_client=dask_client,
)
automl.fit(X_train,
Y_train,
X_test=X_test,
y_test=Y_test,
dataset_name='binary_test_dataset')
predictions = automl.predict(X_test)
assert predictions.shape == (50, ), print_debug_information(automl)
score = accuracy(Y_test, predictions)
assert score > 0.9, print_debug_information(automl)
assert count_succeses(
automl.cv_results_) > 0, print_debug_information(automl)
assert includes_all_scores(automl.performance_over_time_.columns) is True
assert performance_over_time_is_plausible(
automl.performance_over_time_) is True
def test_can_pickle_classifier(tmp_dir, dask_client):
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
automl = AutoSklearnClassifier(
time_left_for_this_task=30,
delete_tmp_folder_after_terminate=False,
per_run_time_limit=5,
tmp_folder=tmp_dir,
dask_client=dask_client,
)
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(tmp_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(tmp_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_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_fit_roar(dask_client_single_worker, backend):
def get_roar_object_callback(
scenario_dict,
seed,
ta,
ta_kwargs,
dask_client,
n_jobs,
**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,
dask_client=dask_client,
n_jobs=n_jobs,
)
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,
initial_configurations_via_metalearning=0,
get_smac_object_callback=get_roar_object_callback,
metric=accuracy,
dask_client=dask_client_single_worker,
)
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 includes_train_scores(automl.performance_over_time_.columns) is True
assert automl._task == MULTICLASS_CLASSIFICATION
del automl
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
# 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_regression(tmp_dir, output_dir, dask_client):
X_train, Y_train, X_test, Y_test = putil.get_dataset('boston')
automl = AutoSklearnRegressor(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)
predictions = automl.predict(X_test)
assert predictions.shape == (356, )
score = mean_squared_error(Y_test, predictions)
# On average np.sqrt(30) away from the target -> ~5.5 on average
# Results with select rates drops avg score to a range of -32.40 to -37, on 30 seconds
# constraint. With more time_left_for_this_task this is no longer an issue
assert score >= -37, print_debug_information(automl)
assert count_succeses(automl.cv_results_) > 0
def test_fit(dask_client):
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
automl = autosklearn.automl.AutoML(
seed=0,
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 includes_train_scores(automl.performance_over_time_.columns) is True
assert performance_over_time_is_plausible(
automl.performance_over_time_) is True
assert automl._task == MULTICLASS_CLASSIFICATION
del automl
def test_classification_pandas_support(tmp_dir, output_dir, dask_client):
X, y = sklearn.datasets.fetch_openml(
data_id=2, # cat/num dataset
return_X_y=True,
as_frame=True,
)
# Drop NAN!!
X = X.dropna('columns')
# This test only make sense if input is dataframe
assert isinstance(X, pd.DataFrame)
assert isinstance(y, pd.Series)
automl = AutoSklearnClassifier(
time_left_for_this_task=30,
per_run_time_limit=5,
exclude_estimators=['libsvm_svc'],
dask_client=dask_client,
seed=5,
tmp_folder=tmp_dir,
output_folder=output_dir,
)
automl.fit(X, y)
# Make sure that at least better than random.
# We use same X_train==X_test to test code quality
assert automl.score(X, y) > 0.555, print_debug_information(automl)
automl.refit(X, y)
# Make sure that at least better than random.
# accuracy in sklearn needs valid data
# It should be 0.555 as the dataset is unbalanced.
prediction = automl.predict(X)
assert accuracy(y, prediction) > 0.555
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
def test_classification_pandas_support(tmp_dir, output_dir, dask_client):
X, y = sklearn.datasets.fetch_openml(
data_id=2, # cat/num dataset
return_X_y=True,
as_frame=True,
)
# Drop NAN!!
X = X.dropna('columns')
# This test only make sense if input is dataframe
assert isinstance(X, pd.DataFrame)
assert isinstance(y, pd.Series)
automl = AutoSklearnClassifier(
time_left_for_this_task=30,
per_run_time_limit=5,
exclude_estimators=['libsvm_svc'],
dask_client=dask_client,
seed=5,
tmp_folder=tmp_dir,
output_folder=output_dir,
)
automl.fit(X, y)
log_file_path = glob.glob(os.path.join(tmp_dir, 'AutoML*.log'))[0]
# Make sure that at least better than random.
# We use same X_train==X_test to test code quality
assert automl.score(X, y) > 0.555, extract_msg_from_log(log_file_path)
automl.refit(X, y)
# Make sure that at least better than random.
# accuracy in sklearn needs valid data
# It should be 0.555 as the dataset is unbalanced.
y = automl.automl_.InputValidator.encode_target(y)
prediction = automl.automl_.InputValidator.encode_target(automl.predict(X))
assert accuracy(y, prediction) > 0.555
assert count_succeses(automl.cv_results_) > 0
def test_multilabel(tmp_dir, output_dir, dask_client):
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris',
make_multilabel=True)
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)
# Log file path
log_file_path = glob.glob(os.path.join(tmp_dir, 'AutoML*.log'))[0]
predictions = automl.predict(X_test)
assert predictions.shape == (50, 3), extract_msg_from_log(log_file_path)
assert count_succeses(
automl.cv_results_) > 0, extract_msg_from_log(log_file_path)
score = f1_macro(Y_test, predictions)
assert score >= 0.9, extract_msg_from_log(log_file_path)
probs = automl.predict_proba(X_train)
assert np.mean(probs) == pytest.approx(0.33, rel=1e-1)
def test_fit_n_jobs(tmp_dir, output_dir):
X_train, Y_train, X_test, Y_test = putil.get_dataset('breast_cancer')
# test parallel Classifier to predict classes, not only indices
Y_train += 1
Y_test += 1
class get_smac_object_wrapper:
def __call__(self, *args, **kwargs):
self.n_jobs = kwargs['n_jobs']
smac = get_smac_object(*args, **kwargs)
self.dask_n_jobs = smac.solver.tae_runner.n_workers
self.dask_client_n_jobs = len(
smac.solver.tae_runner.client.scheduler_info()['workers'])
return smac
get_smac_object_wrapper_instance = get_smac_object_wrapper()
automl = AutoSklearnClassifier(
time_left_for_this_task=30,
per_run_time_limit=5,
output_folder=output_dir,
tmp_folder=tmp_dir,
seed=1,
initial_configurations_via_metalearning=0,
ensemble_size=5,
n_jobs=2,
include_estimators=['sgd'],
include_preprocessors=['no_preprocessing'],
get_smac_object_callback=get_smac_object_wrapper_instance,
max_models_on_disc=None,
)
automl.fit(X_train, Y_train)
# Test that the argument is correctly passed to SMAC
assert getattr(get_smac_object_wrapper_instance, 'n_jobs') == 2
assert getattr(get_smac_object_wrapper_instance, 'dask_n_jobs') == 2
assert getattr(get_smac_object_wrapper_instance, 'dask_client_n_jobs') == 2
available_num_runs = set()
for run_key, run_value in automl.automl_.runhistory_.data.items():
if run_value.additional_info is not None and 'num_run' in run_value.additional_info:
available_num_runs.add(run_value.additional_info['num_run'])
available_predictions = set()
predictions = glob.glob(
os.path.join(automl.automl_._backend.get_runs_directory(), '*',
'predictions_ensemble*.npy'))
seeds = set()
for prediction in predictions:
prediction = os.path.split(prediction)[1]
match = re.match(MODEL_FN_RE,
prediction.replace("predictions_ensemble", ""))
if match:
num_run = int(match.group(2))
available_predictions.add(num_run)
seed = int(match.group(1))
seeds.add(seed)
# Remove the dummy prediction, it is not part of the runhistory
available_predictions.remove(1)
assert available_num_runs.issubset(available_predictions)
assert len(seeds) == 1
ensemble_dir = automl.automl_._backend.get_ensemble_dir()
ensembles = os.listdir(ensemble_dir)
seeds = set()
for ensemble_file in ensembles:
seeds.add(int(ensemble_file.split('.')[0].split('_')[0]))
assert len(seeds) == 1
assert count_succeses(automl.cv_results_) > 0
# For travis-ci it is important that the client no longer exists
assert automl.automl_._dask_client is None
