논문링크 : https://openaccess.thecvf.com/content/CVPR2022/papers/Dong_CSWin_Transformer_A_General_Vision_Transformer_Backbone_With_Cross-Shaped_Windo
Illustration of different self-attention mechanisms, our CSWin is fundamentally different from two aspects. First, we split multi-heads ({h 1 , . . .