CUDA out of memory while training on FlyingThings3D

[starsong98]

Thank you for releasing your model code! I am really appreciating it.

I am trying to train FlowFormer on FlyingThings3D dataset, but even with batch size reduced to 1 and enabling --mixed_precision I keep getting CUDA out of memory errors.

I am using NVIDIA Titan GPUs which have 12 GiB memory. Is there a way to further reduce memory usage such that I can train on these GPUs without using the flowformer-small model?