search

ziyc / drivestudio

A 3DGS framework for omni urban scene reconstruction and simulation.
https://ziyc.github.io/omnire/
MIT License
595 stars 51 forks source link

Modify MAX_JOBS parameter #17

Closed MatteoMarengo closed 2 months ago

MatteoMarengo commented 2 months ago

Hi,

How to modify MAX_JOBS parameter so it is not 10 anymore when running gsplat ? Thanks !

ziyc commented 2 months ago

@MatteoMarengo Can you share a screenshot or output to help locate the issue?

MatteoMarengo commented 2 months ago

@ziyc

for sure, this is done on a A100 - 40GB GPU for NuScenes Scene 3 (same output with nuscenes/3cams and nuscenes/6cams and omnire.yaml)

issue

ziyc commented 2 months ago

It seems that you installed gsplat with pip install gsplat, this way requires compilation for the first run. I recommend compiling form source by: pip install git+https://github.com/nerfstudio-project/gsplat.git@v1.3.0 to install gsplat

MatteoMarengo commented 2 months ago

No I have installed gsplat with your method (I have rerun the command line to be certain) and I still have the same error, would you see any points where I can downgrade the required requirements for example ?

my command line is this one:

python tools/train.py --config_file configs/omnire.yaml --output_root $output_root --project $project --run_name $expname dataset=nuscenes/6cams data.scene_idx=$scene_idx data.start_timestep=$start_timestep data.end_timestep=$end_timestep seed=31

ziyc commented 2 months ago

Is this happening with all scenes or just the one you showed? If it's all scenes, it might be an environment issue. Can you share your CUDA version?

Also, what values did you use for scene_idx, start_timestep, and end_timestep in your command? These are just placeholders, what's the actual value?

ziyc commented 2 months ago

Can you try the examples of gsplat to check if gsplat can successfully run on your side?

MatteoMarengo commented 2 months ago

It now works better, the issue was coming from not enough CPU Ram, it is now better with an increased value.