search

a1600012888 / PhysDreamer

Code for PhysDreamer
437 stars 20 forks source link

How to run this job with 40G VRAM？ #4

Closed 2017211801 closed 2 months ago

2017211801 commented 2 months ago

Thank you for your excellent work! Since I only have roughly 40G VRAM, I can't run this awesome job on my local machine, is it possible to lower the video memory by modifying the config file? If not can you provide some examples of low VRAM, looking forward to your reply!

a1600012888 commented 2 months ago

Hi, 40GB memory machine is enough.

After downloading the data and model from huggingface, please go to folder projects/inference/, and run python3 demo.py --scene_name carnation --apply_force --force_id 1 --point_id 0 --force_mag 2.0 --cam_id 0.

I think 40 GB is fine for this scene, if you encounter OOM at any scenes, you can change to downsampling function following: https://github.com/a1600012888/PhysDreamer/tree/main/projects/inference

It's the pointcloud downsampling function that bottleneck the memory now. (a GPU implementation of Kmeans)

2017211801 commented 2 months ago

Thank you! It works!