In the train process , the point cloud loss will be nan, and causing the loss of total_loss also is a nan. How to solve this problem?