def generate_noise_data(path):
with open(path, 'rb') as f:
audio_list = pickle.load(f)
print('read pkl from ', f)
spec_list = []
record_count = 0
for i, audio_name in enumerate(audio_list):
spec, y = process_stft(path_join(wave_noise_dir, audio_name))
if spec.shape[0] >= config.max_sequence_length:
spec_list.extend(
split_spectrogram(spec, config.max_sequence_length))
else:
spec_list.append(
expand_spectrogram(spec, config.max_sequence_length))
if len(spec_list) >= config.tfrecord_size:
fname = 'noise' + increment_id(record_count, 5) + '.tfrecords'
temp = spec_list[:config.tfrecord_size]
spec_list = spec_list[config.tfrecord_size:]
ex_list = [make_noise_example(spec) for spec in temp]
writer = tf.python_io.TFRecordWriter(
path_join(save_noise_dir, fname))
for ex in ex_list:
writer.write(ex.SerializeToString())
writer.close()
record_count += 1
print(fname, 'created')
print('save in %s' % save_noise_dir)
def generate_valid_data(path):
with open(path, 'rb') as f:
wav_list = pickle.load(f)
print('read pkl from %s' % f)
# each record should be (file.wav,((st,end),(st,end).....)))
file_list = [i[0] for i in wav_list]
label_list = [i[1] for i in wav_list]
tuple_list = []
counter = 0
record_count = 0
for i, audio_name in enumerate(file_list):
# print(audio_name)
# print(label_list[i])
spec, seq_len, labels = make_record(
path_join(wave_valid_dir, audio_name),
label_list[i])
counter += 1
tuple_list.append((spec, labels, seq_len))
if counter == config.tfrecord_size:
tuple_list = batch_padding(tuple_list)
fname = 'data' + increment_id(record_count, 5) + '.tfrecords'
ex_list = [make_example(spec, seq_len, labels) for
spec, labels, seq_len in tuple_list]
writer = tf.python_io.TFRecordWriter(
path_join(save_valid_dir, fname))
for ex in ex_list:
writer.write(ex.SerializeToString())
writer.close()
record_count += 1
counter = 0
tuple_list.clear()
print(fname, 'created')
print('save in %s' % save_valid_dir)
def generate_noise_data(path):
with open(path, 'rb') as f:
audio_list = pickle.load(f)
print('read pkl from ', f)
spec_list = []
counter = 0
record_count = 0
for i, audio_name in enumerate(audio_list):
spec, y = process_stft(path_join(wave_noise_dir, audio_name))
spec_list.append(spec)
counter += 1
if counter == config.tfrecord_size:
spec_list = [
expand_spectrogram(s, config.max_sequence_length)
for s in spec_list
]
fname = 'noise' + increment_id(record_count, 5) + '.tfrecords'
ex_list = [make_noise_example(spec) for spec in spec_list]
writer = tf.python_io.TFRecordWriter(
path_join(save_noise_dir, fname))
for ex in ex_list:
writer.write(ex.SerializeToString())
writer.close()
record_count += 1
counter = 0
spec_list.clear()
print(fname, 'created')
print('save in %s' % save_noise_dir)
def generate_trainning_data(path):
with open(path, 'rb') as f:
wav_list = pickle.load(f)
print('read pkl from %s' % f)
audio_list = [i[0] for i in wav_list]
label_list = [i[1] for i in wav_list]
text_list = [i[2] for i in wav_list]
assert len(audio_list) == len(text_list)
tuple_list = []
counter = 0
record_count = 0
for i, audio_name in enumerate(audio_list):
spec, seq_len, label_values, label_indices, label_shape = make_record(
path_join(wave_train_dir, audio_name), label_list[i])
# print(text_list[i])
# print(label_values)
if spec is not None:
counter += 1
tuple_list.append(
(spec, seq_len, label_values, label_indices, label_shape))
if counter == config.tfrecord_size:
tuple_list = batch_padding_trainning(tuple_list)
fname = 'data' + increment_id(record_count, 5) + '.tfrecords'
ex_list = [
make_trainning_example(spec, seq_len, label_values,
label_indices, label_shape) for spec,
seq_len, label_values, label_indices, label_shape in tuple_list
]
writer = tf.python_io.TFRecordWriter(
path_join(save_train_dir, fname))
for ex in ex_list:
writer.write(ex.SerializeToString())
writer.close()
record_count += 1
counter = 0
tuple_list.clear()
print(fname, 'created')
print('save in %s' % save_train_dir)
def generate_valid_data(pkl_path):
with open(pkl_path, 'rb') as f:
wav_list = pickle.load(f)
print('read pkl from %s' % f)
audio_list = [i[0] for i in wav_list]
correctness_list = [i[1] for i in wav_list]
label_list = [i[2] for i in wav_list]
assert len(audio_list) == len(correctness_list)
tuple_list = []
counter = 0
record_count = 0
for audio_name, correctness, label in zip(audio_list, correctness_list,
label_list):
spectrogram, wave = process_stft(path_join(wave_valid_dir, audio_name))
seq_len = spectrogram.shape[0]
label_values, _, _ = convert_label(label)
tuple_list.append(
(spectrogram, seq_len, correctness, label_values, audio_name))
counter += 1
if counter == config.tfrecord_size:
tuple_list = batch_padding_valid(tuple_list)
fname = 'valid' + increment_id(record_count, 5) + '.tfrecords'
ex_list = [
make_valid_example(spec, seq_len, correctness, label_values,
audio_name) for spec, seq_len, correctness,
label_values, audio_name in tuple_list
]
writer = tf.python_io.TFRecordWriter(
path_join(save_valid_dir, fname))
for ex in ex_list:
writer.write(ex.SerializeToString())
writer.close()
record_count += 1
counter = 0
tuple_list.clear()
print(fname, 'created')
print('save in %s' % save_valid_dir)