long time for train

[chensh1127]

hey, it's a nice work. I am interested at your job. some custom datasets have been trained by your codes. However, the time for train is almost 4 hours. Is this normal?

[LingzheZhao]

Hi, the training time mostly depends on your hardware, image resolution and an important hyper-parameter densify_grad_thresh. For example, on the deblur-nerf synthetic datasets, it usually takes about 0.5 hours on my workstation (AMD Ryzen 7950X CPU + NVIDIA RTX 4090 GPU) and about 1 hour on our server (AMD Epyc 9554 CPU + NVIDIA RTX 4090 GPU).

If your training images are in high resolution, I would first suggest trying to add this argument --downscale_factor 4 after nerfstudio-data to speed up the training;

Also note that if you have a lot of training images, you may change the --eval-mode "all" to something else, like --eval-mode "interval" --eval-interval 20, to reduce the evaluation time:

ns-train bad-gaussians \
    --data path/to/custom_data \
    --vis viewer+tensorboard \
    nerfstudio-data --downscale_factor 4 --eval_mode "interval" --eval-interval 20