differentiable cubify about pytorch3d HOT 4 CLOSED

Hi @haithemb! Cubify is not differentiable, explicitly defined here. This is because there is a thresholding of probabilities in the implementation. You could actually make it differentiable by writing approximate gradients for it. It's a great exercise, I recommend it!

from pytorch3d.

Thanks for the prompt response. Would replacing ge with softmax make it differentiable?

from pytorch3d.

The voxel inputs are already probabilities. So not sure how using a softmax will help. You could sample from the distribution and do something like gumbel-softmax. Or you could define approximate gradients. I will close this issue for now, but if you devise a solution I'd be happy to take a look!

from pytorch3d.

Hi!
@gkioxari has this been implemented or do you know of an implementation you can point me to?
@haithemb did you manage to implement it?

Thanks!

from pytorch3d.