Sum of a boolean tensor

Is it known that if you do torch.sum(y_hat == y), if the sum is larger than 255, the sum will be whatever - 255? I am seeing this behavior with the conda version of PyTorch. If you do torch.sum((y_hat == y).float()) then it is fine.

What is the your verson of PyTorch?


I tried it in PyTorch 0.1.12_2 and found both are fine.

>>> x = torch.autograd.Variable(torch.LongTensor([1]*300))
>>> y = torch.autograd.Variable(torch.LongTensor([1]*300))
>>> torch.sum(x == y)
Variable containing:
[torch.ByteTensor of size 1]

>>> print(torch.__version__)
>>> torch.sum((x == y).int())
Variable containing:
[torch.IntTensor of size 1]


As you observed, the comparison operators return ByteTensor. I would even recommend to use .long() to convert to a LongTensor. You are safer from overflow even if you do calculations but more importantly it is the type that is most often used in pytorch when things expect integers (e.g. indexing, .gather,…).
If you need to do things with other tensors .float() is often needed (e.g. for batch averaging) with a mask or so as you can’t use operators with different tensor types in general.

Best regards


1 Like

The point is that if the Variable wrapping is causing the different behavior. For example:

>>> x = torch.LongTensor([1]*300)
>>> y = torch.LongTensor([1]*300)
>>> x = x.cuda()
>>> y = y.cuda()
>>> torch.sum(x == y)

I am not entirely sure if they should have the same behavior with and without the Variable wrapper, but I am just trying to put it out there.

The difference is actually whether it becomes a python int or a Tensor again.
With (x==y).sum(1) you get the overflow with tensors. Now, Variables never are converted to python numbers (because it would lose autograd).

Best regards