def generate_train(self):
'''Generate mini-batch data for training.
Returns:
batch_data_dict: dict containing audio_name, feature and target
'''
while True:
# Reset pointer
if self.pointer >= len(self.train_audio_indexes):
self.pointer = 0
self.random_state.shuffle(self.train_audio_indexes)
# Get batch audio_indexes
batch_audio_indexes = self.train_audio_indexes[self.pointer:self.
pointer +
self.batch_size]
self.pointer += self.batch_size
batch_data_dict = {}
batch_data_dict['filename'] = \
self.data_dict['filename'][batch_audio_indexes]
batch_feature = self.data_dict['feature'][batch_audio_indexes]
# batch_feature = self.transform(batch_feature)
batch_data_dict['feature'] = batch_feature
sparse_target = self.data_dict['target'][batch_audio_indexes]
batch_data_dict['target'] = sparse_to_categorical(
sparse_target, self.in_domain_classes_num)
yield batch_data_dict
def generate_validate(self, data_type, max_iteration=None):
'''Generate mini-batch data for training.
Args:
data_type: 'train' | 'validate'
max_iteration: int, maximum iteration to validate to speed up validation
Returns:
batch_data_dict: dict containing audio_name, feature and target
'''
batch_size = self.batch_size
if data_type == 'train':
audio_indexes = np.array(self.train_audio_indexes)
elif data_type == 'validate':
audio_indexes = np.array(self.validate_audio_indexes)
else:
raise Exception('Incorrect argument!')
iteration = 0
pointer = 0
while True:
if iteration == max_iteration:
break
# Reset pointer
if pointer >= len(audio_indexes):
break
# Get batch audio_indexes
batch_audio_indexes = audio_indexes[pointer:pointer + batch_size]
pointer += batch_size
iteration += 1
batch_data_dict = {}
batch_data_dict['filename'] = \
self.data_dict['filename'][batch_audio_indexes]
batch_feature = self.data_dict['feature'][batch_audio_indexes]
batch_data_dict['feature'] = batch_feature
batch_start = self.data_dict['start'][batch_audio_indexes]
batch_data_dict['start'] = batch_start
batch_end = self.data_dict['end'][batch_audio_indexes]
batch_data_dict['end'] = batch_end
sparse_target = []
for i in range(len(batch_audio_indexes)):
sparse_target.append(
self.data_dict['target'][batch_audio_indexes[i]])
# sparse_target = self.data_dict['target'][batch_audio_indexes]
batch_data_dict['target'] = sparse_to_categorical(
sparse_target, self.in_domain_classes_num)
yield batch_data_dict