Classwise mAP during validation

[guarin]

Hi, thanks a lot for your amazing work on mmyolo!

Is it possible to show classwise mAP during validation? I currently only get the following output in the log file:

2023/05/03 14:58:52 - mmengine - INFO - Evaluating bbox...
2023/05/03 14:59:03 - mmengine - INFO - bbox_mAP_copypaste: 0.002 0.010 0.000 0.000 0.001 0.003
2023/05/03 14:59:04 - mmengine - INFO - Epoch(val) [1][572/572]    coco/bbox_mAP: 0.0020  coco/bbox_mAP_50: 0.0100  coco/bbox_mAP_75: 0.0000  coco/bbox_mAP_s: 0.0000  coco/bbox_mAP_m: 0.0010  coco/bbox_mAP_l: 0.0030  data_time: 0.0112  time: 0.0358

And in scalars.json:

{"coco/bbox_mAP": 0.002, "coco/bbox_mAP_50": 0.01, "coco/bbox_mAP_75": 0.0, "coco/bbox_mAP_s": 0.0, "coco/bbox_mAP_m": 0.001, "coco/bbox_mAP_l": 0.003, "data_time": 0.011173534393310546, "time": 0.035815372467041015, "step": 1}

I use val_evaluator config from yolov7_l_syncbn_fast_8x16b-300e_coco.py.

[hhaAndroid]

可以通过在配置中将 test_evaluator.classwise 设置为 True，或者在运行 test.py 时增加 --cfg-options test_evaluator.classwise=True 来打印每个类的 mAP。

[guarin]

Thanks a lot!

Just for completeness, it also works for the val_evaluator and can be used in validation during training.

If you save the best checkpoints, make sure to set the target metric in the checkpoint hook: default_hooks.checkpoint.save_best="coco/bbox_mAP". Otherwise it will take the mAP of the first class in the dataset as classwise metrics are reported before overall metrics.