Backward engine computes unnecessary dependencies

[apaszke]

import torch
from torch.autograd import Variable

x = Variable(torch.randn(5, 5), requires_grad=True)
y = Variable(torch.randn(5, 5), requires_grad=True)

a = x + y
b = torch.max(a, 1)[1].repeat(1, 5).float()
o = (b + a).sum()
o.backward()

The error is raised, because the non-differentiable path is included in dependency calculation (we only check requires_grad locally at the moment). It doesn't affect correctness and should never cause problems in most use cases. I'm putting this on hold until #662 is merged.

[danielamassiceti]

I ran into this issue as well. A temporary work around could be to re-package the tensor of argmax b into a new Variable.

Something like:

b = torch.max(a, 1)[1].repeat(1,5).float().data
o = (Variable(b) + a).sum()

This does not raise any error for me!

[apaszke]

Yeah, that's what I proposed on the forums. The issue was originally reported in this thread. Thanks for help though!

[apaszke]

Fixed in #752.