search

skhu101 / GauHuman

Code for our CVPR'2024 paper "GauHuman: Articulated Gaussian Splatting from Monocular Human Videos"
Other
339 stars 29 forks source link

CUDA OUT OF MEMORY #23

Closed ZhaoLizz closed 8 months ago

ZhaoLizz commented 8 months ago

Hi! I encountered an issue while running the training code using the provided train_zju_mocap_refine.sh script on a single Nvidia RTX 4090 GPU. It appears that the rasterizer is demanding 72 GB GPU memory.

I'm reaching out to inquire if any specific parameters within the script or configuration can be adjusted to reduce the GPU memory requirements to a more manageable level.

Thank you for your assistance. image

skhu101 commented 8 months ago

Hi, I did not encounter this problem during my training. Do you check the dimensions of the parameters for rasterizer function?

ZhaoLizz commented 8 months ago

Currently, I'm working on other GS-based Avatar projects. I have encountered that if the init GS scale is too large, it leads to CUDA out of memory. I'm not sure whether GauHuman met the same reason, but for anyone who encounters similar error, reducing scale for the first few epochs worth a try.