Errors when running 3D object detection about bevcontrast

[CSautier]

Hi, sorry for the delay. Could you please tell me what you are trying to run?

[CSautier]

It seems to me you are trying to run the spconv object detection backbone on multiple GPUs, which is something we didn't do in our experiments. I don't recall why we ran this experiment on a single GPU, but given that our other experiments were run on multiple GPUs, it's possible our spconv's code does not work with multiple GPUs. I will try that on my side.

[CSautier]

I can confirm that this issue appears when using spconv in multiGPU. If you need multiGPU, you can fix it by replacing in bevlab/dataloader.py l11 by from spconv.pytorch.utils import PointToVoxel as VoxelGenerator and l169 by voxel_output = self._voxel_generator(torch.from_numpy(points), empty_mean=True).

Use at your own risk. Neither multiprocessing nor this version of VoxelGenerator have been tested.