def eval_step(self, batch, decoding_strategy='score', dump=False):
xs, ys, use_packed = batch.text_vecs, batch.label_vecs, batch.use_packed
xs_lens, ys_lens = batch.text_lens, batch.label_lens
self.eval_mode()
encoder_states = self.encoder(xs, xs_lens, use_packed=use_packed)
if decoding_strategy == 'score':
assert ys is not None
_ = self.compute_loss(encoder_states, xs_lens, ys)
if decoding_strategy == 'greedy':
scores, preds, attn_w_log = self.decode_greedy(
encoder_states, batch.text_vecs.size(0))
preds = torch.stack(preds, dim=1)
scores = torch.stack(scores, dim=1)
#import ipdb; ipdb.set_trace()
pred_lengths = (scores < 0).sum(dim=1).to(scores.device)
length_penalties = torch.Tensor([
Beam.get_length_penalty(i) for i in pred_lengths.tolist()
]).to(scores.device)
scores_length_penalized = scores.sum(dim=1) / length_penalties
pred_scores = tuple(
(p, s) for p, s in zip(preds, scores_length_penalized))
if dump is True:
_dump = [attn_w_log]
return pred_scores, _dump
else:
return pred_scores
if 'beam' in decoding_strategy:
beams = self.decode_beam(int(decoding_strategy.split(':')[-1]),
len(batch.text_lens), encoder_states)
pred_scores = beams
return pred_scores
