Training Code Issues

[G-1nOnly]

Congratulations and thanks for such a great work！
I found that for the DTU dataset (probably other datasets as well?), the norm of normals may be 0 when calculating the normal error for supervising, which causes the code stop running, so I add the following line:
norm = torch.where(torch.eq(norm,0.0), torch.tensor(1.), norm)
in the render_core function in ./models/renderer.py to prevent from such issues and the code works now.
But I'm not sure that the modification of code is correct, so I just raise this issue.

[Runsong123]

Hello, I also met the similar "Division by zero" error. I just add the small number (1e-5) to avoid it.

Bests,
Runsong

[G-1nOnly]

Good to know !

[Lagwein]

Hi @G-1nOnly @G-1nOnly, sorry for the late reply. The normal loss can be just set to 0. It is just a naive attempt to smooth the output mesh for visual pleasure, and it tends out to be unnecessary.
Hi @baekhyun77, I didn't reproduce this error. I guess it is the issue of cuda device, likely some tensors on cpu and some on gpu. Could you please try specify like --gpu 0 in your command line, and see if it may help.

[baekhyun77]

Thank you very much for your answer. I have noticed that it may be normal_ Loss encountered a situation with a value of NaN.

[G-1nOnly]

Thanks for the reply! Okay, setting the loss to be 0 works definitely and thanks for the clarification.