KRNet: Towards Efficient Knowledge Replay