There is error between eval mAP during training and inference mAP after training

[Senwang98]

@SPengLiang
Sorry to disturb you again, but I indeed meet one problem and could you help me?

When I training DID-M3D, I let model eval per 10 epoches, and then I get mAP=1.82925 at epoch10.

After training, I use the command CUDA_VISIBLE_DEVICES=0,1,2,3 python tools/train_val.py --config config/kitti.yaml -e to see the performance of checkpoint_10.pth.

Unfortunately, two mAPs are different.

(即，训练中的eval结果与我用-e指令得到的结果有误差，不知道这是不是正常，好奇到底是什么地方有问题，因为GUPNet没出现这个问题，怀疑是mmdet3d的eval.py？)

[SPengLiang]

Sorry for the late reply! This slight difference is caused by the flag in the training config: drop_last_val, which sets "drop_last" of the val dataloader.

[Senwang98]

@SPengLiang
Okay, I got it.