Changing in place operation

autograd
1

Hi,

I am getting an error: RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [1, 5]], which is output 0 of SoftmaxBackward, is at version 1; expected version 0 instead.

My operation on the tensor of dimensions (1,5) is this:

probs = net(*input_var)
probs[-1] = 0.
probs = probs/torch.sum(probs)

To remove the in place operation, I replaced it with:

probs = net(*input_var)
probs_new = probs.clone()
probs_new[-1] = 0.
probs_new = probs_new/torch.sum(probs_new)

But this still gives the error. What is the correct way to avoid the in place operation?

2

Hi,

Do you get the exact same error? Is it still an output of Softmax?
maybe something inside your net changes the output of the Softmax inplace before returning it?

3

Seems like I forgot to do the same at one more place that piece of code was used. It works now. Thanks.

1 Like