NVIDIAGameWorks/kaolin

Asking support for multiple-gpu training of structed point cloud class

Opened this issue · 1 comments

I want to use nn.DataParallel for multiple-gpu training of sparse convolution in structed point cloud. However, the spc class contain spc.octrees, which is of data type uint8. So It will say "Unconvertible NCCL type Short". This is bad if I want to scale the network size up. Currently I have no solution to this problem. Is there any suggestions?

Actually from this issue, it looks like the issue is from the int16. I'll look into it.