assert torch.isnan(output_voxels).sum().item() == 0

[xioatian1]

非常感谢你们开源了这个优秀的工作！ 我尝试复现多模态baseline：Multimodal-R50_img1600_128x128x10.py 但遇到了图片所示问题，损失为零。我用的是两卡3090进行训练，并尝试将lr调整至3e-5，可是仍然报这个错误，请问您能提一些建议解决这个问题吗？感谢！ @JeffWang987

Thank you very much for open-sourcing this excellent work! I have attempted to reproduce the multimodal baseline: Multimodal-R50_img1600_128x128x10.py, but encountered the issue shown in the image, where the loss remains zero. I am using two 3090 GPUs for training and have tried adjusting the learning rate to 3e-5, but the error persists. Could you kindly provide some suggestions to resolve this issue? Thank you!