Add option to use YOLOv5 AP metric

[kivanctezoren]

As stated in this issue, YOLOv7 uses the AP metric from an earlier version of YOLOv5. The metric has since been updated and according to the developer of v5, this new metric is the same one used by the Detectron2 and MMDetection repositories.

The new v5 metric assumes the maximum recall in the precision-recall curve as 1.0, which results in a higher numeric value than the v7 metric for the same performance. If other repositories use a similar calculation, it may provide a more fair comparison.

Thus, I have added the optional flag v5-metric to test using this metric. It can also be used in training for choosing the best model after training.

[AlexeyAB]

@WongKinYiu I think we can add it as option --v5-metric
Even if we think that this mAP calculation in YOLOv5/Detectron/MMDetection is incorrect, it allows for a fair comparison for custom datasets.

[akashAD98]

@AlexeyAB how can i get map of each class? like yolov4 darknet is showing map of each class

I tried

running the script but not able to get map of each class

I used this commnaad

!python test.py --data /content/gdrive/MyDrive/yolov7/data/custom_yolov7_data.yaml --img 640 --batch 24 --conf 0.50 --iou 0.65 --device 0 --weights /content/gdrive/MyDrive/yolov7/weights/best.pt --save-txt --save-hybrid --save-json --save-conf --augment --task val

[TusharChauhaan]

Bro go into the test.py file and make the changes.