print(sth)
tgt = torch.LongTensor(hypo).to(device).view(1, -1)
attn = model.get_attn(enc_out, mask, tgt)
attn = attn[0]
if len(sth) > len(stable_hypo):
cs = torch.cumsum(attn[head], dim=1)
ep = cs.le(1. - padding).sum(dim=1)
ep = ep.cpu().numpy()
for j in range(len(stable_hypo), len(sth)):
if ep[j - 1] * 4 + stable_time > end: break
if j > len(stable_hypo):
sth = sth[0:j]
latency += end * (len(sth) - len(stable_hypo))
stable_hypo = sth
latency += time_len * (len(hypo) - len(stable_hypo))
latency /= time_len * (len(hypo) - 1)
count += 1
total_latency += latency
print('Latency: %0.3f' % latency)
write_ctm([hypo[1:]], [score[1:]], fctm, [utt], dic, word_dic,
args.space)
fctm.close()
print('Final Latency: %0.3f' % (total_latency / count))
time_elapsed = time.time() - since
print(" Elapsed Time: %.0fm %.0fs" %
(time_elapsed // 60, time_elapsed % 60))
reader.initialize()
since = time.time()
batch_size = args.batch_size
fout = open(args.output, 'w')
while True:
src_seq, src_mask, utts = reader.read_batch_utt(batch_size)
if len(utts) == 0: break
with torch.no_grad():
src_seq, src_mask = src_seq.to(device), src_mask.to(device)
hypos, scores = beam_search(model,
src_seq,
src_mask,
device,
args.beam_size,
args.max_len,
len_norm=args.len_norm,
coverage=args.coverage,
lm=lm,
lm_scale=args.lm_scale)
hypos, scores = hypos.tolist(), scores.tolist()
if args.format == 'ctm':
write_ctm(hypos, scores, fout, utts, dic, word_dic, args.space)
else:
write_text(hypos, scores, fout, utts, dic, args.space)
fout.close()
time_elapsed = time.time() - since
print(" Elapsed Time: %.0fm %.0fs" %
(time_elapsed // 60, time_elapsed % 60))