search

ayushjain1144 / odin

Code for the paper: "ODIN: A Single Model for 2D and 3D Segmentation" (CVPR 2024)
https://odin-seg.github.io
MIT License
125 stars 10 forks source link

CUDA OOM on scannet200 inference #27

Closed Boyu-Mi closed 3 weeks ago

Boyu-Mi commented 3 weeks ago

Hi. Thanks for the great work. I run scripts/scannet200/scannet200_swin.sh using scannet200_swin_31.5_76k_5.5k.pth checkpoint for scannet200 segmentation results on 2 4090GPU. However I found that it cost large GPU memory. Is this normal? Thanks.

image
ayushjain1144 commented 3 weeks ago

I think I was running it on 30G gpus -- but you can perhaps set MAX_FRAME_NUM -1 to 80 and that should work (with potentially some small drop in performance).

esentially, this will load at max 80 images per scene during inference -- you can play around with this number.

Boyu-Mi commented 3 weeks ago

It works on A100. Thanks.