torch.cuda.OutOfMemoryError: CUDA out of memory

[bc129697]

When i run with "bash scripts/run_train_infer.sh" ，This error will occur when using 60 frames of images, but it will not occur when using 25 frames of images

[kairunwen]

Hi, the OOM problem is caused by the memory limitations of DUSt3R: naver/dust3r#1, which should be solved by InstantSplat_MASt3R.

You can modify these parameters to reduce GPU memory usage and accelerate model training, but the model performance will be slightly degraded: https://github.com/NVlabs/InstantSplat/blob/main/coarse_init_infer.py#L70
(1) scene_graph='complete' --> scene_graph='swin'
(2) symmetrize=True --> symmetrize=False

[alpergel]

Hi, where can we find InstantSplat_MAST3R?