When I add the decode function in my network the weight didn’t update,the decode function as below:

1

when I add the decode function in my network the weight didn’t update,the decode function as below:

def decoder(x):
    colormax = x[:,0:192].data.cpu().numpy().reshape((-1,3,64))
    colormin = x[:,192:384].data.cpu().numpy().reshape((-1,3,64))
    modulationData = x[:,384:448].data.cpu().numpy()
    fresult = torch.ones((modulationData.shape[0], 3, 1024))
    fresult=fresult.view((modulationData.shape[0],3072))
    fresult=Variable(fresult, requires_grad=True)
    return fresult
my network is that:
    def forward(self, x):
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = F.relu(self.fc3(x))
        x = F.relu(self.fc4(x))
        x=decoder(x)
        x = F.relu(x)
        return x

what’s wrong with my decode function?

2

any operations done on .data are not backpropped through.

3

what sudo I do in decode function to make backpropped through

4

How about…
Replacing .data.cpu().numpy().reshape((-1,3,64)) with .view(-1, 3, 64) which has the advantage of leaving everything on the gpu.
Then

modulationData = x[:,384:448]
fresult = torch.ones((modulationData.size(0), 3, 1024)).view(-1, 3072))
return fresult

But unless I am much mistaken, your decoder (and this refactoring) returns a Variable that contains only 1s.