ViT-LSLA: Vision Transformer with Light Self-Limited-Attention

원문 초록

Transformers have demonstrated a competitive performance across a wide range of vision tasks, while it is very expensive to compute the global self-attention. Many methods limit the range of attention within a local window to reduce computation complexity. However, their approaches cannot save the number of parameters; meanwhile, the self-attention and inner position bias (inside the softmax function) cause each query to focus on similar and close patches. Consequently, this paper presents a light self-limited-attention (LSLA) consisting of a light self-attention mechanism (LSA) to save the computation cost and the number of parameters, and a self-limited-attention mechanism (SLA) to improve the performance. Firstly, the LSA replaces the K (Key) and V (Value) of self-attention with the X(origin input). Applying it in vision Transformers which have encoder architecture and self-attention mechanism, can simplify the computation. Secondly, the SLA has a positional information module and a limited-attention module. The former contains a dynamic scale and an inner position bias to adjust the distribution of the self-attention scores and enhance the positional information. The latter uses an outer position bias after the softmax function to limit some large values of attention weights. Finally, a hierarchical Vision Transformer with Light self-Limited-attention (ViT-LSLA) is presented. The experiments show that ViT-LSLA achieves 71.6% top-1 accuracy on IP102 (2.4% absolute improvement of Swin-T); 87.2% top-1 accuracy on Mini-ImageNet (3.7% absolute improvement of Swin-T). Furthermore, it greatly reduces FLOPs (3.5GFLOPs vs. 4.5GFLOPs of Swin-T) and parameters (18.9M vs. 27.6M of Swin-T).