stark-t
commented
1 year ago | pt file | metric |
|---|---|
| best_ap50.pt | ap50%-90% |
| best_ap | ap90% |
| best_f | f1 score |
| best_overall | overall accuracy |
| best_p | precission |
| best | fitness (mean ap&ap50) |
| best_r | recall |
Closed valentinitnelav closed 1 year ago
stark-t
commented
1 year ago | pt file | metric |
|---|---|
| best_ap50.pt | ap50%-90% |
| best_ap | ap90% |
| best_f | f1 score |
| best_overall | overall accuracy |
| best_p | precission |
| best | fitness (mean ap&ap50) |
| best_r | recall |
stark-t
commented
1 year ago i think the simple best.pt file should be used since the other yolo version should use the same. But I will check if all yolo version use the same function to calculate the "best fitness".
stark-t
commented
1 year ago w = [0.0, 0.0, 0.1, 0.9] # weights for [P, R, mAP@0.5, mAP@0.5:0.95] return (x[:, :4] * w).sum(1)w = [0.0, 0.0, 0.1, 0.9] # weights for [P, R, mAP@0.5, mAP@0.5:0.95] return (x[:, :4] * w).sum(1)w = [0.0, 0.0, 0.1, 0.9] # weights for [P, R, mAP@0.5, mAP@0.5:0.95] return (x[:, :4] * w).sum(1)
valentinitnelav
commented
1 year ago We don't implement YOLOv4 any longer. But this remains generally valid.
@stark-t , feel free to close this issue if you think it should be closed.
Hi @stark-t ,
When setting a job run using detect.py on the test dataset I remembered again that for YOLOv4 we have multiple best weights options. There are "best" options for several epochs and then these options:
For now, I decided to use
best_overall.ptin detect.py, but without checking what this imply actually. Yolov7 & v5 have a best.pt , but not _ap, _ap50, _f, _p, _r, _overall options.Here are all the files containing the "best*" keyword in a runs/train folded for yolov4r: