Multiple workers to send data to 1 gpu for faster preprocessing

Hi everyone,

I’m trying to send 3D data to GPU for augmentation since it is much faster. However, I encountered the error below while trying to execute “self.points = torch.from_numpy(self.points).float().cuda()”. This only happens when i set the number of workers > 0.

Thank you

This error is expected since you are trying to recreate CUDA contexts in a multiprocessing setting.
You could try to use the spawn method to avoid it.