search

NVIDIAGameWorks / kaolin-wisp

NVIDIA Kaolin Wisp is a PyTorch library powered by NVIDIA Kaolin Core to work with neural fields (including NeRFs, NGLOD, instant-ngp and VQAD).
Other
1.45k stars 132 forks source link

Codebook implementation is not training properly #135

Closed DomaradzkiMaciej closed 1 year ago

DomaradzkiMaciej commented 1 year ago

Hi, I'm trying to train codebook model with the following command: WISP_HEADLESS=1 python3 app/nerf/main_nerf.py --config app/nerf/configs/nerf_codebook.yaml --dataset-path ../data/blender/lego --dataset-num-workers 4 But after 50 epochs, PSNR is still around 9.

image

orperel commented 1 year ago

Hi @DomaradzkiMaciej ! The config you're using is for sparsely initialized grids (i.e. the RTMV dataset, which initializes the grid cells from RGBD data).

The lego data initializes a fully dense Codebook grid in this example, which takes way longer to optimize.

I just checked the codebook the V8 example and it seems to work fine: PSNR after 100 epochs should be ~27.75