def create_batch(self, flag='train', shuffle=True):
data_num = len(self.wav2py[flag])
idxs = list(range(data_num))
if shuffle:
random.shuffle(idxs)
wavs = []
labels = []
for i, idx in enumerate(idxs):
wav_path, pys = self.wav2py[flag][idx]
fbank = extract_feature(wav_path)
label = np.array([self.py2id[py] for py in pys])
while ((fbank.shape[0] >= self.audio_len)
or (len(label) >= self.label_len)):
temp = random.randint(len(idxs) // 4, len(idxs) // 2)
wav_path, pys = self.wav2py[flag][temp]
fbank = extract_feature(wav_path)
label = np.array([self.py2id[py] for py in pys],
dtype=np.int32)
assert len(wavs) == len(labels)
if len(wavs) == self.batch_size:
the_inputs, input_length = self.wav_padding(wavs)
the_labels, label_length = self.label_padding(labels)
inputs = [the_inputs, the_labels, input_length, label_length]
outputs = np.zeros([self.batch_size, 1], dtype=np.float32)
yield inputs, outputs
wavs, labels = [], []
wavs.append(fbank)
labels.append(label)
if len(wavs) != 0:
the_inputs, input_length = self.wav_padding(wavs)
the_labels, label_length = self.label_padding(labels)
inputs = [the_inputs, the_labels, input_length, label_length]
outputs = np.zeros([len(wavs), 1], dtype=np.float32)
yield inputs, outputs
def create_batch(self, flag='train', shuffle=True):
data_num = len(self.wav2py[flag])
idxs = list(range(data_num))
if shuffle:
random.shuffle(idxs)
wavs = []
labels = []
for i, idx in enumerate(idxs):
wav_path, pys = self.wav2py[flag][idx]
fbank = extract_feature(wav_path)
label = np.array([self.py2id[py] for py in pys], dtype=np.int16)
assert len(wavs) == len(labels)
if len(wavs) == self.batch_size:
the_inputs, input_length = self.wav_padding(wavs)
the_labels, label_length = self.label_padding(labels)
for i in range(self.batch_size):
while input_length[i] <= max(label_length):
r = random.randint(1, self.batch_size - 1)
the_inputs[i] = the_inputs[r]
the_labels[i] = the_labels[r]
input_length[i] = input_length[r]
label_length[i] = label_length[r]
assert min(input_length) > max(label_length), str(
min(input_length)) + ":" + str(max(label_length))
inputs = [the_inputs, the_labels, input_length, label_length]
outputs = np.zeros([self.batch_size, 1], dtype=np.float32)
yield inputs, outputs
wavs, labels = [], []
wavs.append(fbank)
labels.append(label)
def check_input(wav_path, label):
fbank = extract_feature(wav_path)
if (fbank.shape[0] // 8) < len(label):
with open('log/check_input.txt', '+a') as file:
file.write(wav_path + '\t' + 'fbank_len:' + str(fbank.shape[0]) +
'\tinput_len' + str(fbank.shape[0] // 8) +
'\tlabel_len' + str(len(label)) + '\n')
def create_batch_dev(self):
wavs = []
labels = []
data_num = len(self.wav2py['dev'])
wavs = []
labels = []
for i in enumerate(data_num):
wav_path, pys = self.wav2py['dev'][i]
fbank = extract_feature(wav_path)
label = [self.py2id[py] for py in pys]
pass
def create_online(self, wav_path):
assert os.path.exists(wav_path)
fbank = extract_feature(wav_path)
assert fbank.shape[0] < self.audio_len
wav, wav_len = self.wav_padding([fbank])
yield wav, wav_len
def report_data(wav_path, label):
fbank = extract_feature(wav_path)
with open('log/data_report.txt', '+a') as file:
file.write(wav_path + '\t' + 'fbank_len:' + str(fbank.shape[0]) +
'\tinput_len:' + str(fbank.shape[0] // 8) + '\tlabel_len:' +
str(len(label)) + '\n')
def check_audio(wav_path):
fbank = extract_feature(wav_path)
if fbank.shape[0] > config.audio_len:
with open('log/check_audio.txt', '+a') as file:
file.write(wav_path + '\n')