RepOPT in Yolov6: training in the hyper search (hs) mode appears SINE-LIKE mAP curve

[KidChou]

Thank you for reading.

In this experiment, I proposed to train yolov6s using the repopt method on the DOTA dataset. According to the official document, I firstly trained the model in hs mode to search the optimal hyper-parameters of optimizer, but found the weird val/mAP curves like a sine function. As seen in the figure, the orange curve refers to the yolov6s model trained after 400 epochs, the blue one is yolov6s in hs mode after 250 training epochs, and the red one represents the yolov6s trained in hs mode after 400 epochs.

As far as I know, in the hs mode, the Scales (hyper-parameters) are trained just as normal parameters together with other model parameters, everything should run like the orange curve. But what made the hs-mode mAP curves wave like sine function?