def get_data(dataset_path, normalization, unlabeled_percentage, seed=None):
train_set, _ = timit.load_data(dataset_path)
train_set, val_set = timit.split_validation(train_set, seed=seed)
train_set, val_set = normalize(train_set, val_set, mode=normalization)
train_set = unlabel(train_set, unlabeled_percentage, seed=seed)
x_train_labeled = np.array([utterance['features'] for utterance in train_set if 'labels' in utterance])
x_train_unlabeled = np.array([utterance['features'] for utterance in train_set if 'labels' not in utterance])
y_train_labeled = np.array([utterance['labels'] for utterance in train_set if 'labels' in utterance])
x_val = np.array([utterance['features'] for utterance in val_set])
y_val = np.array([utterance['labels'] for utterance in val_set])
return x_train_labeled, x_train_unlabeled, y_train_labeled, x_val, y_val
def test_normalize(self):
dataset_path = get_root_dir() / 'data' / 'timit'
train_set, _ = timit.load_data(dataset_path)
# test normalization on whole dataset
normalized_train_set, _ = normalize(train_set, mode='full')
x_train = np.concatenate(
[utterance['features'] for utterance in normalized_train_set])
mean = x_train.mean(axis=0)
var = x_train.var(axis=0)
for i in range(x_train.shape[1]):
self.assertAlmostEqual(mean[i], 0)
self.assertAlmostEqual(var[i], 1)
def test_unlabel(self):
dataset_path = get_root_dir() / 'data' / 'timit'
train_set, _ = timit.load_data(dataset_path)
n_total = len(train_set)
unlabel(train_set, 0.7, seed=1)
n_labeled = len(
[utterance for utterance in train_set if 'labels' in utterance])
n_unlabeled = len([
utterance for utterance in train_set if 'labels' not in utterance
])
self.assertEqual(n_labeled + n_unlabeled, n_total)
self.assertTrue(n_labeled < n_unlabeled)
self.assertEqual(n_labeled, 1104)
self.assertEqual(n_unlabeled, 2592)
def main():
args = get_command_line_arguments()
dataset_path = args.data
checkpoint_path = args.checkpoint
version = args.version
model_path = Path(args.output)
model_path.parent.mkdir(parents=True, exist_ok=True)
model_path = str(model_path)
train_set, _ = timit.load_data(dataset_path)
model = DualStudent(n_classes=get_number_of_classes(),
n_hidden_layers=N_HIDDEN_LAYERS,
n_units=N_UNITS,
padding_value=PADDING_VALUE,
version=version)
model.build(input_shape=(None, ) + train_set[0]['features'].shape)
checkpoint = tf.train.Checkpoint(model=model)
checkpoint.restore(checkpoint_path)
model.save_weights(model_path)
def get_data(dataset_path, normalization):
train_set, test_set = timit.load_data(dataset_path)
_, test_set = normalize(train_set, test_set, mode=normalization)
x_test = np.array([utterance['features'] for utterance in test_set])
y_test = np.array([utterance['labels'] for utterance in test_set])
return x_test, y_test