Sampling implementation is very slow

[MrMois]

For anyone using this implementation, be aware the hierarchical sampling implementation used here is very slow compared to others.

A faster one can be found here (~20x faster in my benchmark): https://github.com/yenchenlin/nerf-pytorch/blob/63a5a630c9abd62b0f21c08703d0ac2ea7d4b9dd/run_nerf_helpers.py#L196

[bebeal]

Ah I didn't realize it was that bad, sheesh. Will fix soon when I get time

[massyzs]

can you share your modified code with us? @MrMois

[MrMois]

Unfortunately no, but as far as I remember it was basically the code from the repo linked above. Note the 20x speedup refers to the sampling method only and not to the overall training, so do not expect wonders to happen :D