def _decode(self, source_lang, target_lang, segments):
prefix_lang = target_lang if self._checkpoint.multilingual_target else None
batch, input_indexes, sentence_len = self._make_decode_batch(segments, prefix_lang=prefix_lang)
# Compute translation
self._translator.max_len_b = self._checkpoint.decode_length(source_lang, target_lang, sentence_len)
translations = self._translator.generate([self._model], batch)
# Decode translation
sub_dict = self._checkpoint.subword_dictionary
results = []
for i, hypo in enumerate(translations):
hypo = hypo[0] # (top-1 best nbest)
hypo_score = math.exp(hypo['score'])
hypo_tokens = hypo['tokens']
hypo_indexes = sub_dict.indexes_of(hypo_tokens)
hypo_str = sub_dict.string(hypo_tokens)
hypo_attention = np.asarray(hypo['attention'].data.cpu())
# Make alignment
if len(hypo_indexes) > 0:
hypo_alignment = make_alignment(input_indexes[i], hypo_indexes, hypo_attention,
prefix_lang=prefix_lang is not None)
hypo_alignment = clean_alignment(hypo_alignment, segments[i], hypo_str)
else:
hypo_alignment = []
results.append(Translation(hypo_str, alignment=hypo_alignment, score=hypo_score))
return results
def _force_decode(self, target_lang, segments, translations):
prefix_lang = target_lang if self._checkpoint.multilingual_target else None
batch = self._make_force_decode_batch(segments,
translations,
prefix_lang=prefix_lang)
src_tokens = batch['src_tokens']
tgt_tokens = batch['trg_tokens']
src_indexes = batch['src_indexes']
tgt_indexes = batch['trg_indexes']
src_lengths = batch['src_lengths']
if self._device is not None:
src_tokens = src_tokens.cuda(self._device)
src_lengths = src_lengths.cuda(self._device)
tgt_tokens = tgt_tokens.cuda(self._device)
self._model.eval()
_, attn = self._model(src_tokens, src_lengths, tgt_tokens)
if type(attn) is dict:
attn = attn['attn']
results = []
for i, hypo_attention in enumerate(
attn): # for each entry of the original batch
hypo_attention = hypo_attention.transpose(0, 1).cpu()
hypo_attention = hypo_attention[hypo_attention.size(0) -
(len(src_indexes[i]) + 1):,
hypo_attention.size(1) -
(len(tgt_indexes[i]) + 1):]
# Make alignment
hypo_alignment = make_alignment(src_indexes[i],
tgt_indexes[i],
hypo_attention.data.numpy(),
prefix_lang=prefix_lang
is not None)
hypo_alignment = clean_alignment(hypo_alignment, segments[i],
translations[i])
results.append(
Translation(translations[i], alignment=hypo_alignment))
return results