Thanks for doing such a great job and for providing the network code for Nuscenes training. I noticed that the projection parameters in the code for loading the dataset for Nuscenes are self-generated. I then referenced this for the modelnet40 dataset, where the overlap handling is referenced to the RPMNet protocol. Regarding the projection parameters, I kept them the same. The result is that after epoch about 2~4, the loss is negative. What should I do and could you provide some good suggestions to help me please. Looking forward to your reply and thanks again for doing such a good job.