Bug in the training code that causes validation losses to be stuck to 0 and infinite
robertastellino opened this issue · comments
robertastellino commented
I trained a yolov8 model for object detection, but when looking at the results, I saw that the training losses are correct (they decrease with the training process), but the validation losses are not computed properly: val/box_losses and val/dfl_loss are always equal to zero (from the beginning to the end), while val/cls_loss is always "inf".
I also trained my model with the official github and I don't get this error, so I think there is a bug in the code that trains the network.
Thank you in advance for your attention, I hope this error will be fixed soon.