search

dcharatan / pixelsplat

[CVPR 2024 Oral, Best Paper Runner-Up] Code for "pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction" by David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann
http://davidcharatan.com/pixelsplat/
MIT License
830 stars 56 forks source link

Same CUDA OOM problem with pytorch1.8.0 + cu11.1 on RTX A6000 #45

Closed railgun526 closed 6 months ago

railgun526 commented 6 months ago 
          Hi @kevinYitshak , I still encounter the same error following the link. My environment is pytorch1.8.0 + cu11.1 on RTX 3090Ti.

Originally posted by @Pixie8888 in https://github.com/dcharatan/pixelsplat/issues/8#issuecomment-1937408010

railgun526 commented 6 months ago

File "/home/miniconda3/envs/gs18_copy/lib/python3.8/site-packages/diff_gaussian_rasterization/init.py", line 92, in forward num_rendered, color, radii, geomBuffer, binningBuffer, imgBuffer = _C.rasterize_gaussians(*args) RuntimeError: CUDA out of memory. Tried to allocate 87964.63 GiB (GPU 0; 44.56 GiB total capacity; 1.60 GiB already allocated; 41.18 GiB free; 1.66 GiB reserved in total by PyTorch) Hi @dcharatan I encountered OOM when doing the forward process in pixelsplat. However, with the same environment, it works well in original gaussian splatting. Could you please help me for this problem?

dcharatan commented 6 months ago

I'm still not sure what causes this issue. I would recommend trying out the environments that work for other people in #8. As far as I can tell, there's no reason why our modifications to diff-gaussian-rasterization should cause this issue. It might be worth trying to include our modified diff-gaussian-rasterization as a submodule and then doing pip install . (like in the original Gaussian Splatting repo) in case the build process this invokes is somehow different than what doing pip install git+<link> uses.