Compute gradient of bitwise OR

autograd
1

I have to calculate the Intersection over Union (IoU) for two segmentations. For that, at some point I have to calculate the bitwise OR of two tensors. I am doing this with (a + b) > 0.
Unfortunately this isn’t differentiable. Does someone know how I can calculate the bitwise OR and still be able to backprop?

2

What if you consider them as float and use multiplications, addition to calculate the IoU? That’s how its done in most codes if I remember correctly

3

I can’t think of a function that only uses non-logical operators
EDIT:
Actually, I just used torch.max(a, b)

4

What I meant is, to calculate the IoU, we can use something like below.

Consider A, B are two 2D tensors made up of 0 and 1s.

Intersection = (A * B).sum()
Union = (A+B - A*B).sum()

IoU = intersection / Union

1 Like