equivalent of pytorch tensor detach ?

[ninobone-oz]

Very interested in using LichtFeld Studio.
In looking at porting some pytorch based code, I am not aware how to do the equivalent of pytorch tensor.detach() or .no_grad().
I have looked at the implementation of pytorch free tensor code in LichtFeld, doesn't seem to have it as far as I can see.

Any ideas?

[MrNeRF]

Itdoes not have autograd. All gradients are straight in cuda implemented. So there is no equivalent.

[ninobone-oz]

Thanks for the reply.

I want to use a weighted loss function w(r) * |r - gt|, where the weights are per pixel function of the rendered output r (gt is ground truth).
Looks like I need to write my own loss function in cuda. Easiest seems to me to modify:

__global__ void fused_l1_kernel(
     const float* __restrict__ img1,
     const float* __restrict__ img2,
     float* __restrict__ grad_out,
     float* __restrict__ partial_sums,
     size_t N,
     float grad_scale) {

and replace the scalar grad_scale with a tensor containing w(r).
Hope I have got that right! Correct me if I am wrong...

[ninobone-oz]

the other option I was thinking of was to use an on-the-fly generated mask with compute_photometric_loss_with_mask