search

Tianxiaomo / pytorch-YOLOv4

PyTorch ,ONNX and TensorRT implementation of YOLOv4
Apache License 2.0
4.47k stars 1.49k forks source link

loss_xy and loss_wh do not converge. #479

Open rgw117 opened 3 years ago

rgw117 commented 3 years ago

image Hi I`m trying to train the model on my own dataset. However, for some reason, the losses for wh and xy do not converge but just oscillate. Also, the AP score is too low like 10%ish after 50 epochs when there is only one class in my own dataset.

I made a train.txt and val.txt file following the instruction in this github, a.k.a xyxy format.

I tried to check the loss function and it seems like the annotations for all bounding bboxes are changed to xywh format where xy refer to the center point in a normalized scale.

I also checked the iou calculation function and did not really find any error in there.

However, even though I tried those kinds of attempts, I could not solve the aforementioned issue(loss_xy and loss_wh).

Does anybody know why?

shoudlee commented 2 years ago

same problem. Did you find it out？