def forward(self, input, hidden_state):
f_output, h_output = self.gru(input, hidden_state)
output = self.weight(f_output)
output = self.context(output).squeeze().permute(1, 0)
output = F.softmax(output, dim=-1)
output = element_wise_mul(f_output, output.permute(1, 0)).squeeze(0)
output = self.fc(output)
return output, h_output
def forward(self, x, hidden_state):
# output = self.attention_model(x,hidden_state)
# return output
f_output, h_output = self.gru(x.float(),hidden_state)
output = matrix_mul(f_output, self.sent_weight, self.sent_bias)
output = matrix_mul(output, self.context_weight).permute(1,0)
output = F.softmax(output)
output = element_wise_mul(f_output,output.permute(1,0))
return output,h_output
def forward(self, x, hidden_state):
f_output, h_output = self.gru(x, hidden_state)
output = matrix_mul(f_output, self.comment_weight, self.comment_bias)
output = matrix_mul(output, self.context_weight).permute(1, 0)
output = F.softmax(output)
output = element_wise_mul(f_output, output.permute(1, 0)).squeeze(0)
output = self.fc(output)
return output, h_output
def forward(self, input, hidden_state):
output = self.lookup(input)
# feature output and hidden state output
f_output, h_output = self.gru(output.float(), hidden_state)
output = matrix_mul(f_output, self.word_weight, self.word_bias)
output = matrix_mul(output, self.context_weight).permute(1, 0)
output = F.softmax(output, dim=-1)
output = element_wise_mul(f_output, output.permute(1, 0))
return output, h_output
def forward(self, input, hidden_state):
f_output, h_output = self.gru(input, hidden_state)
output = matrix_mul(f_output, self.sent_weight, self.sent_bias)
output_2 = matrix_mul(output, self.context_weight).permute(1, 0)
if torch.isnan(output_2[0][0]):
print('nan detected')
return None
output = F.softmax(output_2, dim=-1)
output = element_wise_mul(f_output, output.permute(1, 0)).squeeze(0)
output = self.dropout(self.fc(output))
return output, h_output
def forward(self, input, hidden_state,
doc_len): # input [Doc_len,batch_size,2*word_hidden_size]
mask = torch.arange(input.shape[0]).expand(
len(doc_len), input.shape[0]).to(
self.device) < doc_len.unsqueeze(1)
f_output, h_output = self.gru(
input,
hidden_state) # f_output [Doc_len,batch_size,2*sent_hidden_size]
output = self.sent_weight(f_output)
output = self.context_weight(output).squeeze().permute((1, 0))
#output = mat_mul(f_output,self.sent_weight,self.sent_bias) # out_put [Doc_len,batch_size,2*hidden_size]
#output = mat_mul(output,self.context_weight).permute((1,0)) # [batch_size,Doc_len]
exps = torch.exp(output) * mask.float()
output = exps / (torch.sum(exps, dim=1).unsqueeze(1)
) #[batch_size,Doc_len]
#output[~mask] = float('-inf')
#output = F.softmax(output,dim=1) #[batch_size,Doc_len]
output = element_wise_mul(f_output, output.permute(
(1, 0))).squeeze(0) # [batch_size,2*sent_hidden_size]
output = self.fc(output) # [batch_size,num_classes]
return output, h_output
def forward(self, input, hidden_state, sent_len):
mask = torch.arange(input.shape[0]).expand(
len(sent_len), input.shape[0]).to(
self.device) < sent_len.unsqueeze(1)
output = self.lookup(input) #[Seq_len,batch_size,emb_size]
f_output, h_output = self.gru(
output.float(),
hidden_state) # f_output [Seq_len,batch_size,2*hidden_size]
output = self.word_weight(f_output)
output = self.context_weight(output).squeeze().permute((1, 0))
#output = mat_mul(f_output,self.word_weight,self.word_bias) # out_put [Seq_len,batch_size,2*hidden_size]
#output = mat_mul(output,self.context_weight).permute((1,0)) # [batch_size,Seq_len]
#print(torch.isnan(output).any())
exps = torch.exp(output) * mask.float()
output = exps / (torch.sum(exps, dim=1).unsqueeze(1) + 1e-8
) #[batch_size,Seq_len]
#output[~mask] = float('-inf')
#output = F.softmax(output,dim=1) #[batch_size,Seq_len]
output = element_wise_mul(f_output, output.permute(
(1, 0))) # [1,batch_size,2*hidden_size]
return output, h_output