Compute blocks' std and backpropagation

vision
1

I want to compute a block-based std on the image, but I don’t know which way is efficient and whether the gradient of std can be backpropagated.

For example, I have images: 10x1x20x20 (NCHW), I want to compute the std within 3x3 window on the images and compute each std in this 3x3 matrix and do backpropagation. Can anyone help me? Thanks!

2

You could use F.unfold to create these 3x3 windows from the image and compute the std for each of these windows.
You should be able to backpropagate through torch.std.

3

Thank you. I tried to use unfold, and then computed torch.std on each patch, but got nan…

4

Could you check the input values of this particular patch, which is creating the NaN output?
torch.std should return a zero, if the input is constant.