metric jitter on nuscene valset

[dwy927]

The metirc result is not deterministic.

cmd:
python tools/val_nusc_tracking.py examples/point_pillars/configs/nusc_all_pp_centernet_tracking.py --checkpoint model_zoo/simtrack_pillar.pth --work_dir word_dirs/baseline

For fast, I just use nuscenes v1.0-mini set to test.
It returns the different metric results when I run the above cmd.

[chenxuluo]

The model randomly drops some points during the voxelization step. We plan to release the dynamic voxelization version so the results will be deterministic then.

[dwy927]

Looking forward to it.

[liangchunlan]

python ./tools/val_nusc_tracking.py examples/point_pillars/configs/nusc_all_pp_centernet_tracking.py --checkpoint ./model_zoo/simtrack_pillar.pth  --work_dir ./save_dir/v1.0-mini/ --local_rank 0

I got the result. What could be the problem.

[zzm-hl]

The metirc result is not deterministic.

cmd: python tools/val_nusc_tracking.py examples/point_pillars/configs/nusc_all_pp_centernet_tracking.py --checkpoint model_zoo/simtrack_pillar.pth --work_dir word_dirs/baseline

For fast, I just use nuscenes v1.0-mini set to test. It returns the different metric results when I run the above cmd.

I want to know how to run the train and test code ,the data preparation is done