def test_with_abalone(self):
dataset = "abalone"
dataset_dir = os.path.join(os.path.dirname(__file__), ".datasets")
D = CompetitionDataManager(dataset, dataset_dir)
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):
print
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 test_with_abalone(self):
dataset = "abalone"
dataset_dir = os.path.join(os.path.dirname(__file__), ".datasets")
D = CompetitionDataManager(dataset, dataset_dir)
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):
print
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 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, '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 = NestedCVEvaluator(D_, configuration,
with_predictions=True,
all_scoring_functions=True)
if not self._fit(evaluator):
print
continue
e_, Y_optimization_pred, Y_valid_pred, Y_test_pred = \
evaluator.predict()
err[i] = e_['acc_metric']
print err[i], configuration['classifier']
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 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!
"""
if mode != "test":
num_run = get_new_run_num()
for key in params:
try:
params[key] = int(params[key])
except:
try:
params[key] = float(params[key])
except:
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':
evaluator = HoldoutEvaluator(D,
configuration,
with_predictions=True,
all_scoring_functions=True,
output_y_test=True,
seed=seed,
num_run=num_run)
evaluator.fit()
signal.signal(15, empty_signal_handler)
evaluator.finish_up()
model_directory = os.path.join(os.getcwd(), "models_%d" % seed)
if os.path.exists(model_directory):
model_filename = os.path.join(model_directory,
"%s.model" % num_run)
with open(model_filename, "w") as fh:
pickle.dump(evaluator.model, fh, -1)
elif mode == 'test':
evaluator = TestEvaluator(D,
configuration,
all_scoring_functions=True,
seed=seed)
evaluator.fit()
scores = evaluator.predict()
duration = time.time() - evaluator.starttime
score = scores[metric]
additional_run_info = ";".join(
["%s: %s" % (m_, value) for m_, value in scores.items()])
additional_run_info += ";" + "duration: " + str(duration)
print "Result for ParamILS: %s, %f, 1, %f, %d, %s" % (
"SAT", abs(duration), score, evaluator.seed, additional_run_info)
# CV on the whole training set
elif mode == 'cv':
evaluator = CVEvaluator(D,
configuration,
with_predictions=True,
all_scoring_functions=True,
output_y_test=True,
cv_folds=mode_args['folds'],
seed=seed,
num_run=num_run)
evaluator.fit()
signal.signal(15, empty_signal_handler)
evaluator.finish_up()
elif mode == 'partial_cv':
evaluator = CVEvaluator(D,
configuration,
all_scoring_functions=True,
cv_folds=mode_args['folds'],
seed=seed,
num_run=num_run)
evaluator.partial_fit(mode_args['fold'])
scores = evaluator.predict()
duration = time.time() - evaluator.starttime
score = scores[metric]
additional_run_info = ";".join(
["%s: %s" % (m_, value) for m_, value in scores.items()])
additional_run_info += ";" + "duration: " + str(duration)
print "Result for ParamILS: %s, %f, 1, %f, %d, %s" % (
"SAT", abs(duration), score, evaluator.seed, additional_run_info)
elif mode == 'nested-cv':
evaluator = NestedCVEvaluator(D,
configuration,
with_predictions=True,
inner_cv_folds=mode_args['inner_folds'],
outer_cv_folds=mode_args['outer_folds'],
all_scoring_functions=True,
output_y_test=True,
seed=seed,
num_run=num_run)
evaluator.fit()
signal.signal(15, empty_signal_handler)
evaluator.finish_up()
else:
raise ValueError("Must choose a legal mode.")
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,
'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 = NestedCVEvaluator(D_,
configuration,
with_predictions=True,
all_scoring_functions=True)
if not self._fit(evaluator):
print
continue
e_, Y_optimization_pred, Y_valid_pred, Y_test_pred = \
evaluator.predict()
err[i] = e_['acc_metric']
print err[i], configuration['classifier']
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)
