Learning Effective and Efficient Embedding via an Adaptively-Masked Twins-based Layer