for i, sample in enumerate(samples):
wave = sample['wave']
trn = sample['trn']
copyfile(wave, 'audios/audio_{}.wav'.format(i))
feature = extract_feature(input_file=wave,
feature='fbank',
dim=input_dim,
cmvn=True)
feature = build_LFR_features(feature, m=LFR_m, n=LFR_n)
# feature = np.expand_dims(feature, axis=0)
input = torch.from_numpy(feature).to(device)
input_length = [input[0].shape[0]]
input_length = torch.LongTensor(input_length).to(device)
nbest_hyps = model.recognize(input, input_length, char_list, args)
out_list = []
for hyp in nbest_hyps:
out = hyp['yseq']
out = [char_list[idx] for idx in out]
out = ''.join(out)
out_list.append(out)
print('OUT_LIST: {}'.format(out_list))
gt = [char_list[idx] for idx in trn]
gt = ''.join(gt)
print('GT: {}\n'.format(gt))
results.append({
'out_list_{}'.format(i): out_list,
'gt_{}'.format(i): gt
samples = random.sample(samples, 10)
for sample in samples:
sentence_in = sample['in']
sentence_out = sample['out']
input = torch.from_numpy(np.array(sentence_in,
dtype=np.long)).to(device)
input_length = torch.LongTensor([len(sentence_in)]).to(device)
sentence_in = ' '.join([src_idx2char[idx] for idx in sentence_in])
sentence_out = ''.join([tgt_idx2char[idx] for idx in sentence_out])
sentence_out = sentence_out.replace('<sos>', '').replace('<eos>', '')
print('< ' + sentence_in)
print('= ' + sentence_out)
with torch.no_grad():
nbest_hyps = model.recognize(input=input,
input_length=input_length,
char_list=tgt_idx2char)
# print(nbest_hyps)
for hyp in nbest_hyps:
out = hyp['yseq']
out = [tgt_idx2char[idx] for idx in out]
out = ''.join(out)
out = out.replace('<sos>', '').replace('<eos>', '')
print('> {}'.format(out))
for i in [0, 50]:
input_id = samples['input_corpus'][i]
output_id = samples['output_corpus'][i]
input_sent = torch.from_numpy(np.array(input_id, dtype=np.long)).to(
Config.device)
input_len = torch.LongTensor([len(output_id)]).to(Config.device)
output_sent = torch.from_numpy(np.array(output_id, dtype=np.long)).to(
Config.device)
sent = [id2vocab.get(i, 'unk') for i in input_id]
sent = ''.join(sent)
c, q = sent.split('q:')
ans = [id2vocab.get(i, 'unk') for i in output_id]
ans = ''.join(ans)
ans = ans.replace('<sos>', '').replace('<eos>', '')
print("问题:", q)
print("真实答案:", ans)
with torch.no_grad():
nbest_hyps = model.recognize(input=input_sent,
input_length=input_len,
char_list=id2vocab)
for hyp in nbest_hyps:
out = hyp['yseq']
out = [id2vocab[idx] for idx in out]
out = ''.join(out)
out = out.replace('<sos>', '').replace('<eos>', '')
print("预测答案:", out)
