BakedSDF-MLP PyTorch Memory Allocation

autonomousvision / sdfstudio

A Unified Framework for Surface Reconstruction

Apache License 2.0

1.99k stars 185 forks source link

BakedSDF-MLP PyTorch Memory Allocation #57

Open tsugg opened 1 year ago

tsugg commented 1 year ago

OS : Ubuntu-18.04 LTS (WSL2) CUDA : 11.3 Pytorch : 1.12.1 GPU : 3090ti

I've had no issue training on other models like neus-facto, but when I try baskedsdf or bakedsdf-mlp it's saying that PyTorch is reserving almost every single byte of my GPU memory. Is this an actual bug or does this model require more than 24GBs?

niujinshuchong commented 1 year ago

Hi, you could try to reduce the batch size for training e.g. --pipeline.datamanager.train-num-rays-per-batch 2048.

The default setting of bakedsdf should work on 3090ti.