MapTR
MapTR copied to clipboard
hustvl
Can not reproduce the performance of maptr with bevpool
Hi，the training of maptr with bevpool has some errors that some parameters are not used to compute loss, such as position_encoding. Althrough I have fixed them, the performance of maptr with bevpool still has a large gap with the reported one and the training is crashed caused by
File "MapTR/projects/mmdet3d_plugin/maptr/assigners/maptr_assigner.py", line 182, in assign 1950 【2023-08-28 13:03:46】 matched_row_inds, matched_col_inds = linear_sum_assignment(cost) 1951 【2023-08-28 13:03:46】ValueError: matrix contains invalid numeric entries
The following shows the perfomance of the checkpoint of the 2nd epoch before the model training crash.
I noticed that
NuscMap_chamfer/boundary_AP in the 2nd epoch is 0.1950 in the log you provided.
Emmm, it is strange. Have you prepared the pretrained resnet50 backbone? If yes , and can you successfully reproduced the experiments with BEVFormer and GKT?
Emmm, it is strange. Have you prepared the pretrained resnet50 backbone? If yes , and can you successfully reproduced the experiments with BEVFormer and GKT?
I used the pretrained resnet50 backbone and I have successfully reproduce the experiments with BEVFormer and GKT. If possible, you could run the training using the config of MapTR with bevpool in this repo and I guess the same problems will arise.
I rerun the bevpool experiment with the released code,which is waiting in a queue. Once I have the intermediate results, I will inform in this thread.
My intermediate result seems fine. I guess that the problem may lie in the mmdet3d op compilation.
@JingweiZhang12 Which type of GPU do you use for training? It may be attributed to the mmdet3d compilation flag.
@LegendBC Could you provide the version of the environment? In my environment(A100*8), the effect using maptr with LSS is very poor. my environment: mmdet3d is same as https://github.com/mit-han-lab/bevfusion/tree/db75150717a9462cb60241e36ba28d65f6908607 mmdet==2.20.0 mmcv-full==1.4.0 mmsegmentation==0.14.1
By the way, the bev features extracted by my model can be used normally in detection tasks. Strangely, when using MapTR head, there is a significant difference in performance compared to the original metric