Inconsistent reproducing results on the nuscene dataset!

[zyqww]

I downloaded the latest code and checkpoint files When testing the Lidar-only and Camera-only models, the results obtained were mAP=0.6468, NDS=0.6924 and mAP=0.3554, NDS=0.4121, which are very similar to the paper results, using the commands respectively:

• torchpack dist-run -np 1 python tools/test.py configs/nuscenes/det/transfusion/secfpn/lidar/voxelnet_0p075.yaml pretrained/lidar-only-det .pth --eval bbox
• torchpack dist-run -np 1 python tools/test.py configs/nuscenes/det/centerhead/lssfpn/camera/256x704/swint/default.yaml pretrained/camera- only-det.pth --eval bbox However, when testing the fusion model the results obtained are mAP=0.6728 and NDS=0.7069, which is inconsistent with the paper results mAP=0.6852 and NDS=0.7138, why does this problem arise? The command used is: torchpack dist-run -np 1 python tools/test.py configs/nuscenes/det/transfusion/secfpn/camera+lidar/swint_v0p075/convfuser.yaml pretrained/bevfusion-det.pth --eval bbox

[zyqww]

It is worth mentioning that I can only train on one RTX3090 now, samples_per_gpu and workers_per_gpu are set to 2 and 0 respectively, and there will be an error when training with the original parameters, that is "ValueError: matrix contains invalid numeric entries". I will try to change the learning rate for this problem, but I want to ask if this training method is feasible at present? In addition, the training time is also very long, how can I change it to train on four RTX3090?

[zhijian-liu]

Thank you for your interest in our project. This repository is no longer actively maintained, so we will be closing this issue. Please refer to the amazing implementation at MMDetection3D. Thank you again!