When training, some strange phenomena appear after the learning rate decays.

[FightStone]

very outstanding job. when i training centermask2 on the coco datasets(only person class not 80 classes), everything(include map and loss) looks good in the first stage of training. when learning rate decays by 0.1, loss values significantly reduced, but map reduced. Do you know the reason for this? Looking forward to your reply.