[Code] Enable gradient in renderer.py

[NK-CS-ZZL]

An amazing work and congrats! I have a question in the code and hope for a reply if you're convenient. In the file renderer.py Line 37, the TensoIR model is warped by torch.enable_grad(), so I wonder why we need the gradient here?

[Haian-Jin]

Hi,

This is because the code warped will compute derived normals as done here. The process of computing derived normals will call torch.autograd.grad, for which gradient computing is necessary. Since I warp the evaluation code with torch.no_grad (such as here ), an error will arise when evaluating if I don't warp the code you mentioned with torch.enable_grad() because there will be no gradient by default when evaluating.

Hope my answer can solve your problem.

[NK-CS-ZZL]

Since the function compute_derived_normals has been decorated by @torch.enable_grad(), we do have the gradient in this func whether we use torch.enable_grad or torch.no_grad outside it. Besides, using torch.no_grad prevents the gradient accumulation outside the func, which is more vram-friendly. In my environment, these two settings provide the same results, but using torch.no_grad leads to less memory consumption.

[Haian-Jin]

You are right. The code you mentioned is unnecessary. I will delete it later. Thanks for your suggestion!

[NK-CS-ZZL]

You're welcome and thanks for the reply.