Precision, Recall and mAP seems incoherent

[yusiyoh]

Hello,

I am training YOLOR-D6, I obtain precision, recall and mAP results from test.py with following command using paper branch:
python3 test.py --weights ./runs/train/yolor-d6-1280size-multiGPU/weights/best.pt --img 1280 --verbose --data data/dtld_test.yaml --batch 32 --task test --conf 0.4 --iou 0.5

Here is the result:

I think the mAP0.5 is very high with those recall values. Is there a mistake in metrics.py? Or do I have to run test.py with different options?
Thank you.

[yusiyoh]

Yet another strange thing is that when I reduce conf-thres the mAP decreases significantly instead of increasing ( It is expected to increase with reducing conf-thres).
The first one is with default conf-thres (0.001):

Second one is with 0.4 conf-thres:

[yusiyoh]

I have changed the metrics.py according to https://github.com/ultralytics/yolov5/blob/master/utils/metrics.py and the mAP are very different.
Default YOLOR:

YOLOV5 metrics:

Which one is wrong and why?

[yusiyoh]

I also open an issue in YOLOv5 repo: ultralytics/yolov5#4546
Then, @glenn-jocher changed the metrics.py according to my suggestion: ultralytics/yolov5#4563
Could you please check them and consider changing the following lines:

yolor/utils/metrics.py

Lines 125 to 126 in 2fa3a31

	mrec = recall # np.concatenate(([0.], recall, [recall[-1] + 1E-3]))
	mpre = precision # np.concatenate(([0.], precision, [0.]))

To new metrics.py from YOLOv5: https://github.com/ultralytics/yolov5/blob/8b18b66304317276f4bfc7cc7741bd535dc5fa7a/utils/metrics.py#L94-L95

Thank you

[WongKinYiu]

thanks, i change these two lines with your suggestion, but i think two metrics.py will still output different results.