err = AverageMeter('Loss/test')
ent = AverageMeter('Entropy/test')
cer = AverageCER(blank, space)
wer = AverageWER(blank, space)
with torch.no_grad():
temperature = 3
prediction = []
prior = 0
for xs, ys, xn, yn in dev:
xs, xn = model.forward_acoustic(xs, xn)
xs = model.greedy_decode(xs, argmax=False)
xs = xs.exp().view(-1, len(labels))
prediction.append(xs.argmax(1).cpu())
prior += xs.sum(dim=0)
dev.set_description('Epoch %d Prior %.5f' % (epoch, prior.std().item()))
prediction = torch.cat(prediction)
prior = (prior / prediction.size(0)).log() / temperature
writer.add_histogram('Prediction', prediction[prediction != labels.blank()], epoch)
writer.add_histogram('Prior', prior, epoch)
for xs, ys, xn, yn in test:
ys = ys.cuda(non_blocking=True)
xn = xn.cuda(non_blocking=True)
yn = yn.cuda(non_blocking=True)
zs, xs, xn = model(xs, ys, xn, yn)
ys = ys.t().contiguous()
loss = rnnt_loss(zs,
ys,
xn,
yn,
average_frames=False,
reduction="mean")
xs = model.greedy_decode(xs)
err.update(loss.item())
ent.update(entropy(xs))
hypothesis = decoder.unpad(xs, xn, labels)
references = decoder.unpad(ys, yn, labels)
for h, r in zip(hypothesis, references):
cer.update(decoder.cer(h, r))
wer.update(decoder.wer(h, r))
progress.set_description('epoch %d %s %s %s %s' %
(epoch + 1, err, cer, wer, ent))
sys.stderr.write('\n')
step = 0
writer = SummaryWriter(comment='_rl_bs16x4_beta0.01_lr3e-6')
model.eval()
with torch.no_grad():
temperature = 3
prediction = []
prior = 0
for xs, ys, xn, yn in dev:
xs, xn = model.forward_acoustic(xs, xn)
xs = model.greedy_decode(xs, argmax=False)
xs = xs.exp().view(-1, len(labels))
prediction.append(xs.argmax(1).cpu())
prior += xs.sum(dim=0)
dev.set_description('Prior %.5f' % (prior.std().item()))
prediction = torch.cat(prediction)
prior = (prior / prediction.size(0)).log() / temperature
for epoch in range(1, 11):
train.shuffle(epoch)