def forward(self, input, lengths=None, hidden=None, ent=None):
if self.training and (self.dropword > 0):
emb = embedded_dropout(self.embeddings,
input,
dropout=self.dropword)
else:
emb = self.embeddings(input)
if self.ent_embedding is not None:
emb_ent = self.ent_embedding(ent)
emb = torch.cat((emb, emb_ent), 2)
if self.word_dropout is not None:
emb = self.word_dropout(emb)
# s_len, batch, emb_dim = emb.size()
packed_emb = emb
need_pack = (lengths is not None) and (not self.no_pack_padded_seq)
if need_pack:
# Lengths data is wrapped inside a Variable.
if not isinstance(lengths, list):
lengths = lengths.view(-1).tolist()
packed_emb = pack(emb, lengths)
outputs, hidden_t = self.rnn(packed_emb, hidden)
if need_pack:
outputs = unpack(outputs)[0]
return hidden_t, outputs
def forward(self, inp, context, state, parent_index):
"""
Forward through the decoder.
Args:
input (LongTensor): a sequence of input tokens tensors
of size (len x batch x nfeats).
context (FloatTensor): output(tensor sequence) from the encoder
RNN of size (src_len x batch x hidden_size).
state (FloatTensor): hidden state from the encoder RNN for
initializing the decoder.
Returns:
outputs (FloatTensor): a Tensor sequence of output from the decoder
of shape (len x batch x hidden_size).
state (FloatTensor): final hidden state from the decoder.
attns (dict of (str, FloatTensor)): a dictionary of different
type of attention Tensor from the decoder
of shape (src_len x batch).
"""
# Args Check
assert isinstance(state, RNNDecoderState)
# END Args Check
if self.embeddings is not None:
if self.training and (self.dropword > 0):
emb = embedded_dropout(self.embeddings,
inp,
dropout=self.dropword)
else:
emb = self.embeddings(inp)
else:
emb = inp
if self.word_dropout is not None:
emb = self.word_dropout(emb)
# Run the forward pass of the RNN.
hidden, outputs, attns, rnn_output, concat_c = self._run_forward_pass(
emb, context, state, parent_index)
# Update the state with the result.
state.update_state(hidden)
# Concatenates sequence of tensors along a new dimension.
outputs = torch.stack(outputs)
attns = torch.stack(attns)
return outputs, state, attns, rnn_output, concat_c