How to remove the grad_fn=<SelectBackward> in output array

autograd
1

I am trying to put all my model’s output in an array, though I am using it in eval mode,
computing the output with torch.no_grad(): and detaching it the array is showing
grad_fn=<SelectBackward>
My code

m.eval() # m is my model
for vec,ind in loaderx:

    with torch.no_grad():
        opp,_,_ = m(vec)
        opp = opp.detach().cpu()
    for i in range(ind.shape[0]):
        ar[i] = torch.cat((torch.tensor([ind[i]]),opp[i]),dim =0)
    break

when I try to output the array where my outputs are
ar[0][0] #shown only one element since its a big array
output →
tensor(3239., grad_fn=<SelectBackward>)

2

Hi,

The detach() in the no_grad block is not needed.
You will need to move all the ops into the no_grad block though to make sure no gradient is tracked :slight_smile:

1 Like
3

Hi!
I did it like this, it worked. Thanks

m.eval()
for vec,ind in loaderx:
    with torch.no_grad():
        opp,_,_ = m(vec)    
    for i in range(ind.shape[0]):
        ans = torch.cat((torch.tensor([ind[i]]),opp[i]),dim =0)
        ar[ind[i]] = ans

But can you solve another query, detach() should have been cut it in the computational graph then why it didn’t work?

4

At the place where you detach() is, there is no computational graph because you are running in no_grad mode. So it is just a no-op where it is.

1 Like