def pad_and_mask(self, tensors):
lengths = torch.tensor([len(s) for s in tensors], device=self.device)
max_length = torch.max(lengths)
pad_m = pad_mask(lengths, max_length=max_length, device=self.device)
sub_m = subsequent_mask(max_length)
tensors = (pad_sequence(tensors,
batch_first=True,
padding_value=self.pad_indx).to(self.device))
return tensors, pad_m, sub_m
def forward(self, x, lengths):
x = self.embed(x)
out, last_hidden, _ = self.rnn(x, lengths)
if self.attention is not None:
out, _ = self.attention(out,
attention_mask=pad_mask(
lengths, device=self.device))
out = out.sum(1)
else:
out = last_hidden
return out
def forward(self, x, lengths):
x = self.embed(x)
out, last_out, hidden = self.rnn(x, lengths)
if self.attention is not None:
out, _ = self.attention(out,
attention_mask=pad_mask(
lengths, device=self.device))
out = out.sum(1)
else:
out = last_out
"""
Word rnn returns as output the output of the last unpadded timestep
(attentions may have been applied)
"""
return out, hidden
