def validation_step(self, batch, batch_nb):
unvalid = ['[B]', '[S]', '[N]', '[T]', '[P]']
feature, feature_length, target, target_length = batch[0], batch[
1], batch[2], batch[3]
model_output, output_token, spec_output, feature_length, ori_token, ori_token_length, ce_loss, switch_loss = self.forward(
feature, feature_length, target, target_length, True)
result_string_list = [
' '.join([j for j in tokenize(i) if j not in unvalid])
for i in self.transformer.inference(feature, feature_length)
]
target_string_list = [
' '.join([
j
for j in tokenize(self.transformer.vocab.id2string(i.tolist()))
if j not in unvalid
]) for i in output_token
]
print(result_string_list[0])
print(target_string_list[0])
mers = [
cal_wer(i[0], i[1])
for i in zip(target_string_list, result_string_list)
]
mer = np.mean(mers)
ctc_loss = self.transformer.cal_ctc_loss(spec_output, feature_length,
ori_token, ori_token_length)
loss = self.hparams.loss_lambda * ce_loss + (
1 - self.hparams.loss_lambda) * ctc_loss + switch_loss / 2
tqdm_dict = {
'loss': loss,
'ce': ce_loss,
'switch': switch_loss,
'mer': mer,
'lr': self.lr
}
output = OrderedDict({
'loss': loss,
'ce': ce_loss,
# 'ctc_loss': ctc_loss,
'switch': switch_loss,
'mer': mer,
'progress_bar': tqdm_dict,
'log': tqdm_dict
})
return output
def merge(wav_list, target_dict, extract_name_fn, manifest_csv_path):
wav_df = pd.DataFrame(wav_list, columns=['wav_file'])
wav_df.index = wav_df.wav_file.apply(extract_name_fn)
target_df = pd.DataFrame.from_dict(target_dict,
orient='index',
columns=['target'])
merged_df = pd.merge(left=wav_df,
right=target_df,
left_index=True,
right_index=True)
merged_df['duration'] = merged_df['wav_file'].apply(cal_duration)
merged_df['target'] = merged_df['target'].apply(
lambda x: combine(tokenize(x)))
try:
merged_df.to_csv(manifest_csv_path, encoding='utf8')
except:
merged_df.to_csv(manifest_csv_path)
print(f'manifest saved to {manifest_csv_path}')
return 'done'
def extract_corpus_from_target_dict(target_dict, write_to):
with open(write_to, 'w', encoding='utf8') as writer:
for name, target in target_dict.items():
writer.write(combine(tokenize(target.strip())) + '\n')
print('done')
