search

ayaanzhaque / instruct-nerf2nerf

Instruct-NeRF2NeRF: Editing 3D Scenes with Instructions (ICCV 2023)
https://instruct-nerf2nerf.github.io/
MIT License
792 stars 70 forks source link

The training speed is a bit slow, how to make it faster #40

Closed 870223050 closed 1 year ago

870223050 commented 1 year ago

Thanks for your work, I train very slowly when training the bear dataset, I would like to ask if this is normal? My graphics card is two Quadro RTX 6000, CUDA version is 11.8, and the training command is ns-train in2n-tiny --data \bear --load-dir \instruct-nerf2nerf\outputs\bear\nerfacto\2023-06-06_111321\nerfstudio_models --pipeline.prompt "Turn the bear into a grizzly bear" --pipeline.guidance-scale 7.5 --pipeline.image-guidance-scale 1.5 --pipeline.ip2p-device cuda:1 nerfstudio-data --downscale-factor 8

44530 (148.43%) 618.377 ms 21 h, 30 m, 15 s 44.08 K 44540 (148.47%) 603.640 ms 21 h, 33 m, 44 s 44.73 K 44550 (148.50%) 623.457 ms 21 h, 28 m, 49 s 43.03 K 44560 (148.53%) 638.317 ms 21 h, 25 m, 7 s 43.84 K 44570 (148.57%) 656.085 ms 21 h, 20 m, 41 s 42.92 K https://img1.imgtp.com/2023/06/06/mPALwpCR.png How much time will you need to train, and what percentage of steps will you take after the training is completed?

ayaanzhaque commented 1 year ago

This result is very surprising. The scene should not diverge. Can you show some results from your original nerfacto to make sure that it looks correct? Also, use a downscale factor of 2 instead of 8.

ayaanzhaque commented 1 year ago

Will close this issue as there hasn't been a response in a while.