ASR: 类注意力结构重参数化（英文版）

英文标题：ASR: Attention-alike Structural Re-parameterization中文摘要：通过结构重参数化技术 (SRP) 实现网络结构间的互相转换，本文提出一种 attention-alike SRP (ASR) 技术，使得 self-attention 模块也能被重构，从而可在不需要深度模型设计的情况下提高性能。英文摘要：The structural re-parameterization (SRP) technique is a novel deep learningtechnique that achieves interconversion between different network architecturesthrough equivalent parameter transformations. This technique enables themitigation of the extra costs for performance improvement during training, suchas parameter si