def forward(self,
batch,
task,
tokenizer,
compute_loss: bool = False,
loss_weights: torch.Tensor = None):
if loss_weights is not None:
raise NotImplementedError()
encoder_output = get_output_from_encoder_and_batch(
encoder=self.encoder, batch=batch)
logits = self.span_prediction_head(unpooled=encoder_output.unpooled)
# Ensure logits in valid range is at least self.offset_margin higher than others
logits_offset = logits.max() - logits.min() + self.offset_margin
logits = logits + logits_offset * batch.selection_token_mask.unsqueeze(
dim=2)
if compute_loss:
loss_fct = nn.CrossEntropyLoss()
loss = loss_fct(
logits.transpose(dim0=1, dim1=2).flatten(end_dim=1),
batch.gt_span_idxs.flatten(),
)
return LogitsAndLossOutput(logits=logits,
loss=loss,
other=encoder_output.other)
else:
return LogitsOutput(logits=logits, other=encoder_output.other)
def forward(self,
batch,
task,
tokenizer,
compute_loss: bool = False,
loss_weights: torch.Tensor = None):
input_ids = batch.input_ids
segment_ids = batch.segment_ids
input_mask = batch.input_mask
choice_score_list = []
encoder_outputs_other_ls = []
for i in range(self.num_choices):
encoder_output = get_output_from_encoder(
encoder=self.encoder,
input_ids=input_ids[:, i],
segment_ids=segment_ids[:, i],
input_mask=input_mask[:, i],
)
choice_score = self.choice_scoring_head(
pooled=encoder_output.pooled)
choice_score_list.append(choice_score)
encoder_outputs_other_ls.append(encoder_output.other)
reshaped_outputs = []
if encoder_outputs_other_ls[0]:
for j in range(len(encoder_outputs_other_ls[0])):
reshaped_outputs.append([
torch.stack([
misc[j][layer_i] for misc in encoder_outputs_other_ls
],
dim=1)
for layer_i in range(len(encoder_outputs_other_ls[0][0]))
])
reshaped_outputs = tuple(reshaped_outputs)
logits = torch.cat([
choice_score.unsqueeze(1).squeeze(-1)
for choice_score in choice_score_list
],
dim=1)
if compute_loss:
if loss_weights is not None:
loss_fct = nn.CrossEntropyLoss(reduction='none')
loss = loss_fct(logits.view(-1, self.num_choices),
batch.label_id.view(-1))
loss = (loss * loss_weights).mean()
else:
loss_fct = nn.CrossEntropyLoss()
loss = loss_fct(logits.view(-1, self.num_choices),
batch.label_id.view(-1))
return LogitsAndLossOutput(logits=logits,
loss=loss,
other=reshaped_outputs)
else:
return LogitsOutput(logits=logits, other=reshaped_outputs)