my dataset is about 640videos,so the lr is low,i just increase the lr simply.but in the training,the lr are increasing still, is that alright?