Question regarding implementation of the variable out_of_bound about sfmlearner-pytorch HOT 2 CLOSED

the ref_img_warped == 0 variable is not differentiable since it's binary operation (it's either 0 or 1, so it's either constant or not even continuous).
So the gradients are not propagated through this variable. However, you're right that the gradient can propagate at least until this binary tensor, while we already know it's a dead en from out_of_bound, so a few computation is wasted, but I'd say that the overhead is very low.

To make sure you don't compute any gradient wrt to out_of_bound, you can use the context manager with torch.no_grad(): before setting this variable, but I doubt that it will be very useful.

from sfmlearner-pytorch.

Yes. That makes sense. It doesn't put too much overhead.

Thank you very much.

from sfmlearner-pytorch.