Abstract: To transfer the representation capacity of large pre-trained models to lightweight models, knowledge distillation has been widely explored. However, conventional single-stage distillation ...