Extract CNN's output with gather, convert in Variable; lost gradient?

autograd
1

Hi guys,

Is more a clarification question more then technical.

I need extract two class from my features extractors (CNN).
For doing that i have used

dataLoad.data.gather(1, idsDim.view(-1,1)

Where: dataLoad are my CNN output prediction and ids a zero tensor of the
same length of output class.

Now, I need to backward those independent outputs with a two different targets and different loss function as well.
Then, I have converted the specific sequence output (from gather) in autograd Variable and i set the flag “requires_grad=True”.

==> The question is: This it’s correct or i have lost the gradient of CNN?

I am sorry for the question.

Best,

Nico

2

You have lost it when you used .data. If you use gather on the variable (no .data), yoy get to keep it.

Best regards

Thomas

1 Like