Huang Junjie

@lalaha1 Have you used a python version other than 3.7/3.6?

@lalaha1 here is the pip list of my environment, hope this is helpful for you absl-py 0.13.0 addict 2.4.0 anyio 3.6.1 apex 0.1 argcomplete 1.12.3 argon2-cffi 21.1.0 attrs 21.2.0 Babel...

@lalaha1 https://blog.csdn.net/qq_39266065/article/details/119173112 Maybe the torch version is inappropriate

@Divadi All differences between BEVDepth and BEVDet may lead to this problem……

EMMM, I think that I have found the main reasons: 1. have not used grad_clip for resnet-base configs 2. use warm-up lr_config when I change the optimizer of bevdepth4d-r50.py as...

It seem that you have use another version of mmdet3d to process the dataset. You have to rebuild the nuscenes dataset with the BEVDet code base.

@eezhang123 the main precision loss is in mASE and mAOE, you should process the dataset with BEVDet code base instead of other version of mmdet3d

@LigZhong have you pulled and used the latest code?

@LigZhong check this commit~ https://github.com/HuangJunJie2017/BEVDet/commit/196fed98adddf1d8b5989b5eb3028af5067157f2

@LigZhong If you use another version before this commit, you just need to test the trained model with the script of tools/test.py to get a more precise result