writer.add_histogram('Prior', prior, epoch)
for xs, ys, xn, yn in test:
zs, xs, xn = model(xs, ys.t(), xn, yn)
loss1 = rnnt_loss(zs, ys, xn, yn, average_frames=False, reduction="mean")
loss2 = -(zs.exp() * zs).sum(dim=-1).mean()
xs = model.greedy_decode(xs, prior)
err.update(loss1.item())
ent.update(loss2.item())
remove_blank(xs, xn, blank)
cer.update(xs, ys, xn, yn)
wer.update(xs, ys, xn, yn)
test.set_description('Epoch %d %s %s %s %s' % (epoch, err, ent, cer, wer))
sys.stderr.write('\n')
err.summary(writer, epoch)
ent.summary(writer, epoch)
cer.summary(writer, epoch)
wer.summary(writer, epoch)
writer.flush()
step += 1
model.eval()
with torch.no_grad():
hs, hn = model.forward_acoustic(xs, xn)
hs_k = hs.repeat(K, 1, 1)
hn_k = hn.repeat(K)
ys_k = ys.repeat(K, 1)
yn_k = yn.repeat(K)
hs_k = model.greedy_decode(hs_k, prior, sampled=True)
remove_blank(hs_k, hn_k, blank)
WER = compute_wer(hs_k, ys_k, hn_k, yn_k, blank, space)
SymAcc = 1 - 0.5 * WER * (1 + yn_k.float() / hn_k.clamp_min(1).float())
rewards = relu(SymAcc).reshape(K, -1).cuda()
rewards_mean = rewards.mean().item()
rewards -= rewards.mean(dim=0)
elu(rewards, alpha=gamma, inplace=True)
hs_k = hs_k.reshape(K, len(xs), -1)
hn_k = hn_k.reshape(K, len(xs))