search

JuliaNeuralGraphics / NerfUtils.jl

Reusable NeRF components
MIT License
1 stars 0 forks source link

A question #2

Open vpuri3 opened 2 months ago

vpuri3 commented 2 months ago

Hi @pxl-th,

This looks like a great package. I'm wondering why you wrote your own Dense and Chain layers instead of using those in Lux/Flux?

pxl-th commented 2 months ago

Hi. I was writing this package targeting AMD GPU hardware as well. And back then Flux.jl did not support AMD GPU so it seemed easier to just implement couple of layers here than to add full Flux.jl support.

That said, once Flux.jl has gained AMDGPU support, I tried using it, but it was slower so I kept using this code.

vpuri3 commented 2 months ago

Understood. That's great that you're getting a speedup. I came across this package yesterday and I'd like to use this for image/SDF regression tasks. Do you have a MWE by chance for using this package for those tasks?

pxl-th commented 1 month ago

You mean you want to learn image based in input coordinates, like in the link? https://github.com/NVlabs/instant-ngp/tree/d78b422c53b925fd86c6fdfbcaa958a3d45ecc7d?tab=readme-ov-file#image-of-einstein

For SDF you can check Nerf.jl package, it implements instant-ngp NeRF example, but it should be easy to extend it to SDF.