def forward(self, t, length):
mask = torch.gt(torch.unsqueeze(t, 2), 0).type(
torch.cuda.FloatTensor) # (batch_size,sent_len,1)
mask.requires_grad = False
SEQ = mask.size(1)
emb = t = self.embeds(t)
t = self.fc1_dropout(t)
t = nn.utils.rnn.pack_padded_sequence(t,
lengths=length,
batch_first=True)
# t = t.mul(mask) # (batch_size,sent_len,char_size)
t1, (h_n, c_n) = self.lstm1(t, None)
# t1, (h_n, c_n) = self.lstm2(t1, None)
t1, _ = nn.utils.rnn.pad_packed_sequence(t1,
batch_first=True,
total_length=SEQ)
t_max, t_max_index = seq_max_pool([t1, mask])
t_dim = list(t1.size())[-1]
o = seq_and_vec([t1, t_max])
o = o.permute(0, 2, 1)
o = self.conv1(o)
o = o.permute(0, 2, 1)
h_n = torch.cat((h_n[0], h_n[1]), dim=-1).unsqueeze(0)
c_n = torch.cat((c_n[0], c_n[1]), dim=-1).unsqueeze(0)
return o, (h_n, c_n)
def forward(self, t):
mask = torch.gt(torch.unsqueeze(t, 2), 0).type(
torch.cuda.FloatTensor
) # (batch_size,sent_len,1)
mask.requires_grad = False
t = self.embeds(t)
t = self.fc1_dropout(t)
t = t.mul(mask) # (batch_size,sent_len,char_size)
t, (h_n, c_n) = self.lstm1(t, None)
t, (h_n, c_n) = self.lstm2(t, None)
t_max, t_max_index = seq_max_pool([t, mask])
t_dim = list(t.size())[-1]
h = seq_and_vec([t, t_max])
h = h.permute(0, 2, 1)
h = self.conv1(h)
h = h.permute(0, 2, 1)
ps1 = self.fc_ps1(h)
ps2 = self.fc_ps2(h)
return [ps1, ps2, t, t_max, mask]
def to_rel(self, input, h, encoder_o, mask):
output, attn, h = self.forward_step(input, h, encoder_o)
new_encoder_o = seq_and_vec([encoder_o, output.squeeze(1)])
new_encoder_o = new_encoder_o.permute(0, 2, 1)
new_encoder_o = self.conv2_to_1_rel(new_encoder_o)
new_encoder_o = new_encoder_o.permute(0, 2, 1)
output = self.dropout(new_encoder_o)
output = activation(output)
output = self.rel(output)
output, _ = seq_max_pool([output, mask])
return output, h, new_encoder_o, attn