def forward(self, input, encoder_state, incr_state=None):
attention_mask = None
if incr_state is None:
# first step
if (not self.add_start_token and input.size(1) == 1
and int(input[0][0]) == self.start_idx):
# generating: ignore the start token
model_input = encoder_state
else:
# forced decoding: concatenate the context
# with the labels
model_input, _ = concat_without_padding(encoder_state,
input,
use_cuda=self.use_cuda,
null_idx=self.null_idx)
attention_mask = model_input != self.null_idx
else:
# generation: get the last token input
model_input = input[:, -1].unsqueeze(1)
transformer_outputs = self.transformer(model_input,
past=incr_state,
attention_mask=attention_mask)
hidden_states = transformer_outputs[0]
new_incr_state = transformer_outputs[1]
return hidden_states, new_incr_state
def score_candidates(self, batch, cand_vecs, cand_encs=None):
if cand_encs is not None:
raise Exception('Candidate pre-computation is impossible on the '
'crossencoder')
num_cands_per_sample = cand_vecs.size(1)
bsz = cand_vecs.size(0)
text_idx = (batch.text_vec.unsqueeze(1).expand(
-1, num_cands_per_sample,
-1).contiguous().view(num_cands_per_sample * bsz, -1))
cand_idx = cand_vecs.view(num_cands_per_sample * bsz, -1)
tokens, segments = concat_without_padding(text_idx, cand_idx,
self.use_cuda, self.NULL_IDX)
scores = self.model(tokens, segments)
scores = scores.view(bsz, num_cands_per_sample)
return scores
def score_candidates(self, batch, cand_vecs, cand_encs=None):
# concatenate text and candidates (not so easy)
# unpad and break
nb_cands = cand_vecs.size()[1]
size_batch = cand_vecs.size()[0]
text_vec = batch.text_vec
tokens_context = (text_vec.unsqueeze(1).expand(
-1, nb_cands, -1).contiguous().view(nb_cands * size_batch, -1))
# remove the start token ["CLS"] from candidates
tokens_cands = cand_vecs.view(nb_cands * size_batch, -1)
all_tokens, all_segments = concat_without_padding(
tokens_context, tokens_cands, self.use_cuda, self.NULL_IDX)
all_mask = all_tokens != self.NULL_IDX
all_tokens *= all_mask.long()
scores = self.model(all_tokens, all_segments, all_mask)
return scores.view(size_batch, nb_cands)