CVPR 2020 的一篇自注意力机制
Contributions
-
explore variations of self-attention and assess their effectiveness for image recognition; 按两类self-attention进行探讨:pairwise self-attention & patchwise self-attention
-
主要结论:
Methods
-
Pairwise Self-attention
乘在beta(xj)上的weight只由xi,xj决定。可以通过加position encoding让网络知晓xi,xj的位置关系。
-
Patch Self-attention
乘在be