def forward(self, inputs):
input_len = inputs.shape
embed = self._embedding(inputs.view([-1, input_len[2]]))
embed = self._dropout(embed)
hidden = init_const(2,
input_len[0] * input_len[1],
self._hidden_size,
use_cuda=self._use_cuda)
word_rnn_output, _ = self._word_gru(embed, hidden)
word_attn_output = self._word_attn(word_rnn_output)
sen_input = word_attn_output.view(
[-1, input_len[1], self._hidden_size * 2])
sen_input = self._dropout(sen_input)
hidden = init_const(2,
input_len[0],
self._hidden_size,
use_cuda=self._use_cuda)
sen_rnn_output, _ = self._sen_gru(sen_input, hidden)
sen_attn_output = self._sen_attn(sen_rnn_output)
logits = self._fc(sen_attn_output)
return logits
def forward(self, encoder_hidden, encoder_output, targets=None):
batch_size = encoder_output.shape[1]
max_len = self.max_output_len
if targets is not None:
max_len = targets.shape[1]
decoder_input = init_const(batch_size,
value=TOKEN.START,
dtype=torch.long,
use_cuda=self.use_cuda)
hidden = encoder_hidden[1].unsqueeze(0)
decoder_outputs = []
use_teacher_forcing = False if (targets is None
or random.random() < 0.5) else True
for i in range(max_len):
decoder_output, hidden = self.forward_step(decoder_input, hidden,
encoder_output)
decoder_outputs.append(decoder_output)
if use_teacher_forcing:
decoder_input = targets[:, i]
else:
_, top_index = decoder_output.topk(1)
decoder_input = top_index
return decoder_outputs
def forward(self, inputs):
batch_size = inputs.shape[0]
embedded = self.embedding(inputs).transpose(0, 1)
embedded = self.dropout(embedded)
hidden = init_const(2,
batch_size,
self.hidden_size,
use_cuda=self.use_cuda)
output, hidden = self.gru(embedded, hidden)
return output, hidden
def forward(self, inputs):
input_len = inputs.shape
embed = self._embedding(inputs)
embed = self._dropout(embed)
hidden = init_const(2,
input_len[0],
self._hidden_size,
use_cuda=self._use_cuda)
rnn_output, _ = self._gru(embed, hidden)
rnn_output = rnn_output.mean(dim=1)
logits = self._fc(rnn_output)
return logits
def forward(self, inputs):
input_len = inputs.shape
embed = self._embedding(inputs)
embed = self._dropout(embed)
hidden = init_const(2,
input_len[0],
self._hidden_size,
use_cuda=self._use_cuda)
rnn_output, _ = self._gru(embed, hidden)
rnn_output = rnn_output.transpose(1, 2)
conv_output = self._conv1(rnn_output).view([input_len[0], -1])
logits = self._fc(conv_output)
return logits