Result of torch implementation different from CUDA implementation

nerfstudio-project / gsplat

CUDA accelerated rasterization of gaussian splatting

https://docs.gsplat.studio/

Apache License 2.0

1.76k stars 220 forks source link

Result of torch implementation different from CUDA implementation #157

Open wuzirui opened 5 months ago

wuzirui commented 5 months ago

Hi, thanks for the great work! I am currently using gsplat (and nerfstudio) to construct my own pipeline, and I want to use torch implementation first to check everything else is correct. But the results are not passing the assert_close API, do you think this is normal? (P.S The unit test passes (test_project_gaussians_forward...), but in my real data, the results from two implementations are different)

versions: nerfstudio 1.0.2 gsplat 0.1.9

Some screenshots for your info:

jb-ye commented 4 months ago

Could you try the latest gsplat and see if you still see those issues? I believe we fixed some gradient bugs.