Can't reimplement the result of nuscenes 3d detection result

[Wang-zipeng]

I used the script nuScenes_3Ddetection_e140.sh in the CenterTrack/experiments, but I can't reimplement the result in the model zoo with a mAP about 0.30~.When I run the command "python test.py ddd --exp_id nuScenes_3Ddetection_e140 --dataset nuscenes --resume" to evaluate the model, I got this error: Accumulating metric data... Traceback (most recent call last): File "/home/wangzp/git_projects/CenterTrack/src/tools/nuscenes-devkit/python-sdk/nuscenes/eval/detection/evaluate.py", line 302, in nusc_eval.main(plot_examples=plot_examples_, render_curves=render_curves_) File "/home/wangzp/git_projects/CenterTrack/src/tools/nuscenes-devkit/python-sdk/nuscenes/eval/detection/evaluate.py", line 204, in main metrics, metric_data_list = self.evaluate() File "/home/wangzp/git_projects/CenterTrack/src/tools/nuscenes-devkit/python-sdk/nuscenes/eval/detection/evaluate.py", line 116, in evaluate md = accumulate(self.gt_boxes, self.pred_boxes, class_name, self.cfg.dist_fcn_callable, dist_th) File "/home/wangzp/miniconda3/envs/mmdet3d/lib/python3.9/site-packages/nuscenes/eval/detection/algo.py", line 103, in accumulate match_data['scale_err'].append(1 - scale_iou(gt_box_match, pred_box)) File "/home/wangzp/miniconda3/envs/mmdet3d/lib/python3.9/site-packages/nuscenes/eval/common/utils.py", line 99, in scale_iou assert all(sr_size > 0), 'Error: sample_result sizes must be >0.' AssertionError: Error: sample_result sizes must be >0.

I think there is any class in any distance has an empty result result in this error. And my model's loss is bigger than the model provided in the model zoo. My trained model's loss is: hm 2.8014 |wh 6.9593 |reg 0.2466 |dep 8.8626 |dim 0.7961 |rot 2.1047 |amodel_offset 2.9173 The loss of provided model in the model zoo is: hm 1.0984 |wh 1.5407 |reg 0.2202 |dep 1.6663 |dim 0.2449 |rot 1.6251 |amodel_offset 0.9897 It's much smaller than my trained model. ps: i got your loss by load the model to train some step,and set the lr 2.5e-6. And i can reimplement the validation result with model provided in the model zoo.

@xingyizhou My question is if i can reimplement the model with the script nuScenes_3Ddetection_e140.sh, If i can't could you please share the script?