def get_BLEU(self, batch_y_hat, batch_y):
"""Get the average smoothed BLEU of the predictions."""
hyps = batch_y_hat.tolist()
refs = batch_y.tolist()
bleus = []
for hyp, ref in zip(hyps, refs):
if 2 in hyp:
hyp = hyp[:hyp.index(2)]
if 2 in ref:
ref = ref[:ref.index(2)]
hyp = hyp[1:]
ref = ref[1:]
bleus.append(smoothed_bleu(hyp, ref))
return torch.tensor(np.mean(bleus) * 100.)
def _compute_bleu(sampled_tokens, tgt_seq):
"""Compute smoothed BLEU of sampled tokens
"""
bleus = []
tgt_seq = tgt_seq.cpu().numpy()
sampled_tokens = sampled_tokens.cpu().numpy()
tgt_mask = np.greater(tgt_seq, 0)
for i in xrange(tgt_seq.shape[0]):
target_len = int(tgt_mask[i].sum())
ref_tokens = tgt_seq[i, 1:target_len - 1]
out_tokens = list(sampled_tokens[i, 1:target_len - 1])
if not out_tokens:
bleus.append(0.)
else:
bleus.append(smoothed_bleu(out_tokens, ref_tokens))
return np.mean(bleus)