yolov7
yolov7 copied to clipboard
Maybe there's a bug here
YOLO v7
YOLOv5
Hi, this code differs from yolov5. I think m.anchors should not be divided by the m.stride before running check_anchor_order, and that sorting should be done on the original size
I think you are correct. Thanks for point out of this.
@WongKinYiu
Thank you for your reply. I see that you have not corrected the code error. Besides, I wonder if the training curve is normal.
val loss_obj's curve changed during training. I think it is caused by the different loss functions of training and testing (ComputeLoss and ComputeLossOTA). I hope to get your confirmation