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, '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._fit(evaluator):
print
continue
e_, Y_optimization_pred, Y_valid_pred, Y_test_pred = \
evaluator.predict()
err[i] = e_
print err[i], configuration['classifier']
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_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 = CVEvaluator(D_, configuration, cv_folds=5)
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)
