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)
print("Number of times it was worse than random guessing:" + str(np.sum(err > 1)))
def test_5000_classes(self):
weights = ([0.0002] * 4750) + ([0.0001] * 250)
X, Y = sklearn.datasets.make_classification(
n_samples=10000,
n_features=20,
n_classes=5000,
n_clusters_per_class=1,
n_informative=15,
n_redundant=5,
n_repeated=0,
weights=weights,
flip_y=0,
class_sep=1.0,
hypercube=True,
shift=None,
scale=1.0,
shuffle=True,
random_state=1,
)
self.assertEqual(250, np.sum(np.bincount(Y) == 1))
D = Dummy()
D.info = {"metric": ACC_METRIC, "task": MULTICLASS_CLASSIFICATION, "is_sparse": False, "label_num": 1}
D.data = {"X_train": X, "Y_train": Y, "X_valid": X, "X_test": X}
D.feat_type = ["numerical"] * 5000
configuration_space = get_configuration_space(
D.info, include_estimators=["lda"], include_preprocessors=["no_preprocessing"]
)
configuration = configuration_space.sample_configuration()
D_ = copy.deepcopy(D)
evaluator = HoldoutEvaluator(D_, configuration)
evaluator.fit()
def test_with_abalone(self):
dataset = 'abalone'
dataset_path = os.path.join(os.path.dirname(__file__), '.datasets',
dataset)
D = CompetitionDataManager(dataset_path)
configuration_space = get_configuration_space(
D.info,
include_estimators=['extra_trees'],
include_preprocessors=['no_preprocessing'])
errors = []
for i in range(N_TEST_RUNS):
configuration = configuration_space.sample_configuration()
D_ = copy.deepcopy(D)
evaluator = NestedCVEvaluator(D_, configuration,
inner_cv_folds=2,
outer_cv_folds=2)
if not self._fit(evaluator):
continue
err = evaluator.predict()
self.assertLess(err, 0.99)
self.assertTrue(np.isfinite(err))
errors.append(err)
# This is a reasonable bound
self.assertEqual(10, len(errors))
self.assertLess(min(errors), 0.77)
def _create_search_space(tmp_dir, data_info, watcher, log_function):
task_name = 'CreateConfigSpace'
watcher.start_task(task_name)
config_space_path = os.path.join(tmp_dir, 'space.pcs')
configuration_space = paramsklearn.get_configuration_space(
data_info)
sp_string = pcs_parser.write(configuration_space)
_write_file_with_data(config_space_path, sp_string,
'Configuration space', log_function)
watcher.stop_task(task_name)
return configuration_space, config_space_path
def main():
parser = ArgumentParser()
parser.add_argument("configuration_directory",
metavar="configuration-directory")
parser.add_argument("output_directory", metavar="output-directory")
parser.add_argument("--cutoff", type=int, default=-1,
help="Only consider the validation performances up to "
"this time.")
parser.add_argument("--num-runs", type=int, default=1)
parser.add_argument("--only-best", type=bool, default=False,
help="Look only for the best configuration in the "
"validation files.")
args = parser.parse_args()
configuration_directory = args.configuration_directory
output_dir = args.output_directory
cutoff = int(args.cutoff)
num_runs = args.num_runs
for sparse, task in [(1, BINARY_CLASSIFICATION),
(1, MULTICLASS_CLASSIFICATION),
(0, BINARY_CLASSIFICATION),
(0, MULTICLASS_CLASSIFICATION)]:
for metric in ['acc_metric', 'auc_metric', 'bac_metric', 'f1_metric',
'pac_metric']:
output_dir_ = os.path.join(output_dir, '%s_%s_%s' % (
metric, TASK_TYPES_TO_STRING[task], 'sparse' if sparse else 'dense'))
configuration_space = paramsklearn.get_configuration_space(
{'is_sparse': sparse, 'task': task}
)
try:
os.makedirs(output_dir_)
except:
pass
outputs, configurations = retrieve_matadata(
validation_directory=configuration_directory,
num_runs=num_runs,
metric=metric,
cutoff=cutoff,
configuration_space=configuration_space,
only_best=args.only_best)
if len(outputs) == 0:
raise ValueError("Nothing found!")
write_output(outputs, configurations, output_dir_,
configuration_space, metric)
def test_evaluate_multiclass_classification(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=["extra_trees"], include_preprocessors=["select_rates"]
)
err = np.zeros([N_TEST_RUNS])
num_models_better_than_random = 0
for i in range(N_TEST_RUNS):
print("Evaluate configuration: %d; result:" % i)
configuration = configuration_space.sample_configuration()
D_ = copy.deepcopy(D)
evaluator = CVEvaluator(D_, configuration, with_predictions=True)
if not self._fit(evaluator):
print()
continue
e_, Y_optimization_pred, Y_valid_pred, Y_test_pred = evaluator.predict()
err[i] = e_
print(err[i], configuration["classifier:__choice__"])
num_targets = len(np.unique(Y_train))
self.assertTrue(np.isfinite(err[i]))
self.assertGreaterEqual(err[i], 0.0)
# Test that ten models were trained
self.assertEqual(len(evaluator.models), 10)
self.assertEqual(Y_optimization_pred.shape[0], Y_train.shape[0])
self.assertEqual(Y_optimization_pred.shape[1], num_targets)
self.assertEqual(Y_valid_pred.shape[0], Y_valid.shape[0])
self.assertEqual(Y_valid_pred.shape[1], num_targets)
self.assertEqual(Y_test_pred.shape[0], Y_test.shape[0])
self.assertEqual(Y_test_pred.shape[1], num_targets)
# Test some basic statistics of the dataset
if err[i] < 0.5:
self.assertTrue(0.3 < Y_valid_pred.mean() < 0.36666)
self.assertGreaterEqual(Y_valid_pred.std(), 0.01)
self.assertTrue(0.3 < Y_test_pred.mean() < 0.36666)
self.assertGreaterEqual(Y_test_pred.std(), 0.01)
num_models_better_than_random += 1
self.assertGreater(num_models_better_than_random, 5)
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_metalearning(self):
dataset_name = 'digits'
initial_challengers = {
ACC_METRIC: "--initial-challengers \" "
"-balancing:strategy 'weighting' "
"-classifier:__choice__ 'proj_logit'",
AUC_METRIC: "--initial-challengers \" "
"-balancing:strategy 'none' "
"-classifier:__choice__ 'random_forest'",
BAC_METRIC: "--initial-challengers \" "
"-balancing:strategy 'weighting' "
"-classifier:__choice__ 'proj_logit'",
F1_METRIC: "--initial-challengers \" "
"-balancing:strategy 'weighting' "
"-classifier:__choice__ 'proj_logit'",
PAC_METRIC: "--initial-challengers \" "
"-balancing:strategy 'none' "
"-classifier:__choice__ 'random_forest'"
}
for metric in initial_challengers:
configuration_space = get_configuration_space(
{
'metric': metric,
'task': MULTICLASS_CLASSIFICATION,
'is_sparse': False
},
include_preprocessors=['no_preprocessing'])
X_train, Y_train, X_test, Y_test = get_dataset(dataset_name)
categorical = [False] * X_train.shape[1]
meta_features_label = calc_meta_features(X_train, Y_train,
categorical, dataset_name)
meta_features_encoded_label = calc_meta_features_encoded(X_train,
Y_train,
categorical,
dataset_name)
initial_configuration_strings_for_smac = \
create_metalearning_string_for_smac_call(
meta_features_label,
meta_features_encoded_label,
configuration_space, dataset_name, metric,
MULTICLASS_CLASSIFICATION, False, 1, None)
print(metric)
print(initial_configuration_strings_for_smac[0])
self.assertTrue(initial_configuration_strings_for_smac[
0].startswith(initial_challengers[metric]))
def _create_search_space(
tmp_dir, data_info, backend, watcher, logger, include_estimators=None, include_preprocessors=None
):
task_name = "CreateConfigSpace"
watcher.start_task(task_name)
configspace_path = os.path.join(tmp_dir, "space.pcs")
configuration_space = paramsklearn.get_configuration_space(
data_info, include_estimators=include_estimators, include_preprocessors=include_preprocessors
)
sp_string = pcs_parser.write(configuration_space)
backend.write_txt_file(configspace_path, sp_string, "Configuration space")
watcher.stop_task(task_name)
return configuration_space, configspace_path
def test_file_output(self):
output_dir = os.path.join(os.getcwd(), '.test')
try:
shutil.rmtree(output_dir)
except Exception:
pass
X_train, Y_train, X_test, Y_test = get_dataset('boston')
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': R2_METRIC,
'task': REGRESSION,
'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']
D.name = 'test'
configuration_space = get_configuration_space(D.info)
while True:
configuration = configuration_space.sample_configuration()
evaluator = HoldoutEvaluator(D, configuration,
with_predictions=True,
all_scoring_functions=True,
output_dir=output_dir,
output_y_test=True)
if not self._fit(evaluator):
continue
evaluator.predict()
evaluator.file_output()
self.assertTrue(os.path.exists(os.path.join(
output_dir, '.auto-sklearn', 'true_targets_ensemble.npy')))
break
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_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_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)
print("Number of times it was worse than random guessing:" + str(np.sum(err > 1)))
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_file_output(self):
output_dir = os.path.join(os.getcwd(), ".test")
try:
shutil.rmtree(output_dir)
except Exception:
pass
X_train, Y_train, X_test, Y_test = get_dataset("boston")
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": R2_METRIC, "task": REGRESSION, "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"]
D.name = "test"
configuration_space = get_configuration_space(D.info)
while True:
configuration = configuration_space.sample_configuration()
evaluator = HoldoutEvaluator(
D,
configuration,
with_predictions=True,
all_scoring_functions=True,
output_dir=output_dir,
output_y_test=True,
)
if not self._fit(evaluator):
continue
evaluator.predict()
evaluator.file_output()
self.assertTrue(os.path.exists(os.path.join(output_dir, ".auto-sklearn", "true_targets_ensemble.npy")))
break
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)
print("Number of times it was worse than random guessing:" + str(np.sum(err > 1)))
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_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)
print("Number of times it was worse than random guessing:" + str(np.sum(err > 1)))
def test_evaluate_multiclass_classification(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': ACC_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'])
err = np.zeros([N_TEST_RUNS])
num_models_better_than_random = 0
for i in range(N_TEST_RUNS):
print('Evaluate configuration: %d; result:' % i)
configuration = configuration_space.sample_configuration()
D_ = copy.deepcopy(D)
evaluator = NestedCVEvaluator(D_, configuration,
with_predictions=True,
all_scoring_functions=True)
if not self._fit(evaluator):
continue
e_, Y_optimization_pred, Y_valid_pred, Y_test_pred = \
evaluator.predict()
err[i] = e_[ACC_METRIC]
print(err[i], configuration['classifier:__choice__'])
print(e_['outer:bac_metric'], e_[BAC_METRIC])
# Test the outer CV
num_targets = len(np.unique(Y_train))
self.assertTrue(np.isfinite(err[i]))
self.assertGreaterEqual(err[i], 0.0)
# Test that ten models were trained
self.assertEqual(len(evaluator.outer_models), 5)
self.assertTrue(all([model is not None
for model in evaluator.outer_models]))
self.assertEqual(Y_optimization_pred.shape[0], Y_train.shape[0])
self.assertEqual(Y_optimization_pred.shape[1], num_targets)
self.assertEqual(Y_valid_pred.shape[0], Y_valid.shape[0])
self.assertEqual(Y_valid_pred.shape[1], num_targets)
self.assertEqual(Y_test_pred.shape[0], Y_test.shape[0])
self.assertEqual(Y_test_pred.shape[1], num_targets)
# Test some basic statistics of the predictions
if err[i] < 0.5:
self.assertTrue(0.3 < Y_valid_pred.mean() < 0.36666)
self.assertGreaterEqual(Y_valid_pred.std(), 0.1)
self.assertTrue(0.3 < Y_test_pred.mean() < 0.36666)
self.assertGreaterEqual(Y_test_pred.std(), 0.1)
num_models_better_than_random += 1
# Test the inner CV
self.assertEqual(len(evaluator.inner_models), 5)
for fold in range(5):
self.assertEqual(len(evaluator.inner_models[fold]), 5)
self.assertTrue(all([model is not None
for model in evaluator.inner_models[fold]
]))
self.assertGreaterEqual(len(evaluator.outer_indices[fold][0]),
75)
for inner_fold in range(5):
self.assertGreaterEqual(
len(evaluator.inner_indices[fold][inner_fold][0]), 60)
self.assertGreater(num_models_better_than_random, 9)
def test_metalearning(self):
dataset_name = 'digits'
initial_challengers = {
'acc_metric': ["--initial-challengers \" "
"-adaboost:algorithm 'SAMME.R' "
"-adaboost:learning_rate '0.400363929326' "
"-adaboost:max_depth '5' "
"-adaboost:n_estimators '319' "
"-balancing:strategy 'none' "
"-classifier 'adaboost' "
"-imputation:strategy 'most_frequent' "
"-preprocessor 'no_preprocessing' "
"-rescaling:strategy 'min/max'\""],
'auc_metric': ["--initial-challengers \" "
"-adaboost:algorithm 'SAMME.R' "
"-adaboost:learning_rate '0.966883114819' "
"-adaboost:max_depth '5' "
"-adaboost:n_estimators '412' "
"-balancing:strategy 'weighting' "
"-classifier 'adaboost' "
"-imputation:strategy 'median' "
"-preprocessor 'no_preprocessing' "
"-rescaling:strategy 'min/max'\""],
'bac_metric': ["--initial-challengers \" "
"-adaboost:algorithm 'SAMME.R' "
"-adaboost:learning_rate '0.400363929326' "
"-adaboost:max_depth '5' "
"-adaboost:n_estimators '319' "
"-balancing:strategy 'none' "
"-classifier 'adaboost' "
"-imputation:strategy 'most_frequent' "
"-preprocessor 'no_preprocessing' "
"-rescaling:strategy 'min/max'\""],
'f1_metric': ["--initial-challengers \" "
"-adaboost:algorithm 'SAMME.R' "
"-adaboost:learning_rate '0.966883114819' "
"-adaboost:max_depth '5' "
"-adaboost:n_estimators '412' "
"-balancing:strategy 'weighting' "
"-classifier 'adaboost' "
"-imputation:strategy 'median' "
"-preprocessor 'no_preprocessing' "
"-rescaling:strategy 'min/max'\""],
'pac_metric': ["--initial-challengers \" "
"-adaboost:algorithm 'SAMME.R' "
"-adaboost:learning_rate '0.400363929326' "
"-adaboost:max_depth '5' "
"-adaboost:n_estimators '319' "
"-balancing:strategy 'none' "
"-classifier 'adaboost' "
"-imputation:strategy 'most_frequent' "
"-preprocessor 'no_preprocessing' "
"-rescaling:strategy 'min/max'\""]
}
for metric in initial_challengers:
configuration_space = get_configuration_space(
{
'metric': metric,
'task': MULTICLASS_CLASSIFICATION,
'is_sparse': False
},
include_preprocessors=['no_preprocessing'])
X_train, Y_train, X_test, Y_test = get_dataset(dataset_name)
categorical = [False] * X_train.shape[1]
meta_features_label = calc_meta_features(X_train, Y_train,
categorical, dataset_name)
meta_features_encoded_label = calc_meta_features_encoded(X_train,
Y_train,
categorical,
dataset_name)
initial_configuration_strings_for_smac = \
create_metalearning_string_for_smac_call(
meta_features_label,
meta_features_encoded_label,
configuration_space, dataset_name, metric,
MULTICLASS_CLASSIFICATION, False, 1, None)
print(metric)
self.assertEqual(initial_challengers[metric],
initial_configuration_strings_for_smac)
def main(dataset_info, mode, seed, params, mode_args=None):
"""This command line interface has three different operation modes:
* CV: useful for the Tweakathon
* 1/3 test split: useful to evaluate a configuration
* cv on 2/3 train split: useful to optimize hyperparameters in a training
mode before testing a configuration on the 1/3 test split.
It must by no means be used for the Auto part of the competition!
"""
debug_log("Run script")
num_run = None
if mode != 'test':
num_run = get_new_run_num()
for key in params:
try:
params[key] = int(params[key])
except Exception:
try:
params[key] = float(params[key])
except Exception:
pass
if seed is not None:
seed = int(float(seed))
else:
seed = 1
output_dir = os.getcwd()
D = store_and_or_load_data(dataset_info=dataset_info, outputdir=output_dir)
cs = get_configuration_space(D.info)
configuration = configuration_space.Configuration(cs, params)
metric = D.info['metric']
global evaluator
# Train/test split
if mode == 'holdout':
make_mode_holdout(
D,
seed,
configuration,
num_run)
elif mode == 'test':
make_mode_holdout(
D,
seed,
configuration, metric)
elif mode == 'cv':
make_mode_cv(
D,
seed,
configuration,
num_run,
mode_args['folds'])
elif mode == 'partial_cv':
make_mode_partial_cv(
D,
seed,
configuration,
num_run,
metric,
mode_args['folds'],
mode_args['fold']),
elif mode == 'nested_cv':
make_mode_nested_cv(
D,
seed,
configuration,
num_run,
mode_args['inner_folds'],
mode_args['outer_folds']),
else:
raise ValueError('Must choose a legal mode.')
def test_evaluate_multiclass_classification_partial_fit(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,
'target_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=['ridge'],
include_preprocessors=['select_rates'])
err = np.zeros([N_TEST_RUNS])
num_models_better_than_random = 0
for i in range(N_TEST_RUNS):
print('Evaluate configuration: %d; result:' % i)
configuration = configuration_space.sample_configuration()
D_ = copy.deepcopy(D)
evaluator = CVEvaluator(D_, configuration, with_predictions=True)
if not self._partial_fit(evaluator, fold=i % 10):
print()
continue
e_, Y_optimization_pred, Y_valid_pred, Y_test_pred = \
evaluator.predict()
err[i] = e_
print(err[i], configuration['classifier'])
self.assertTrue(np.isfinite(err[i]))
self.assertGreaterEqual(err[i], 0.0)
# Test that only one model was trained
self.assertEqual(len(evaluator.models), 10)
self.assertEqual(1, np.sum([True if model is not None else False
for model in evaluator.models]))
self.assertLess(Y_optimization_pred.shape[0], 13)
self.assertEqual(Y_valid_pred.shape[0], Y_valid.shape[0])
self.assertEqual(Y_test_pred.shape[0], Y_test.shape[0])
# Test some basic statistics of the dataset
if err[i] < 0.5:
self.assertTrue(0.3 < Y_valid_pred.mean() < 0.36666)
self.assertGreaterEqual(Y_valid_pred.std(), 0.01)
self.assertTrue(0.3 < Y_test_pred.mean() < 0.36666)
self.assertGreaterEqual(Y_test_pred.std(), 0.01)
num_models_better_than_random += 1
self.assertGreaterEqual(num_models_better_than_random, 5)