Attention-based transformer networks have demonstrated promising potential as their applications extend from natural language processing to vision. However, despite the recent improvements, such as sub-quadratic attention approximation and various training enhancements, the compact vision transformers to date using the regular attention still fall short in comparison with its convnet counterparts, in terms of \textit{accuracy,} \textit{model size}, \textit{and} \textit{throughput}. This paper introduces a compact self-attention mechanism that is fundamental and highly generalizable. The proposed method reduces redundancy and improves efficiency on top of the existing attention optimizations. We show its drop-in applicability for both the regular attention mechanism and some most recent variants in vision transformers. As a result, we produced smaller and faster models with the same or better accuracies.
View on arXiv