Closed Chi-Zaozao closed 4 years ago
I meet the same warning, and i found the loss is too big, I'm not sure whether the warning is normal. But the validation results is 0.
Does the validation result mean there is no problem? The loss gets smaller and smaller as the training progresses. @chanyn
But when PointPillars trained on Kitti, the initial loss is lower than 10. If you get the right performance on Nuscenes? Actually, I can't run Nuscenes in Det3d project, I met data preprocess problem when i start the train. My problem is recorded in https://github.com/poodarchu/Det3D/issues/85. But I can run Nuscenes in SECOND.pytorch project. And the log is the same with you. The final result is far lower than the official reported results.
I tried kitti today, but I met this error: ''' No such file or directory: '/data/Datasets/KITTI/Kitti/object/training/planes/000000.txt' ''' Do you know how to get the ground planes file? @chanyn
@Chi-Zaozao You need to generate ground plane files before. The code is in det3d/datasets/utils/ground_plane_detection.py, but needs some small changes, not very difficult.
Or you can run without ground planes. Just modify with_gp=True
to with_gp=False
in the __getitem__
method of Dataset Class.
Hi @zwqnju, I found that hey create pkl file for the plane but couldn't find where they generate the txt files, do you have any idea?
thanks!
@MeyLavie Ground plane txt files should be saved in training/planes
I reuse the code of det3d/datasets/utils/ground_plane_detection.py, and write a script to generate ground planes for KITTI dataset, which may help you.
It can work, but the speed is very slow. I do not use it because I don't need ground planes when training PointPillars. If you need ground planes, maybe you can use some methods like multiprocessing to make it fast.
Does it normal that several numbawarning happens when I train PointPillars on the nuscenes mini? whereas it produce the final result. what I observed:
I don't know what the normal result should be. Could you give a tutorial please? Maybe a jupyter notebook? Thank you very much.