模式识别国家重点实验室知识库

Vision transformers (ViTs) have attracted considerable research attention recently, but the huge computational cost is still a severe issue. A mainstream paradigm for computation reduction aims to reduce the number of tokens given that the computation complexity of ViT is quadratic with respect to the input sequence length. Existing designs include structured spatial compression that uses a progressive shrinking pyramid to reduce the computations of large feature maps, and unstructured token pruning that dynamically drops redundant tokens. However, limitations of existing token pruning lie in the following aspects: 1) the incomplete spatial structure caused by pruning is incompatible with structured spatial compression that is commonly used in modern deep-narrow transformers; 2) it usually requires a time-consuming pretraining procedure. To address the limitations and expand the applicable scenario of token pruning, we present Evo-ViT, a self-motivated slow-fast token evolution approach for vision transformers. Specifically, we conduct unstructured instancewise token selection by taking advantage of the simple and effective global class attention that is native to vision transformers. Then, we propose to update the selected informative tokens and uninformative tokens with different computation paths, namely, slow-fast updating. Since slow-fast updating mechanism maintains the spatial structure and information flow, Evo-ViT can accelerate vanilla transformers of both flat and deep-narrow structures from the very beginning of the training process. Experimental results demonstrated that our method significantly reduces the computational cost of vision transformers while maintaining comparable performance on image classification. For example, our method accelerates DeiT-S by over 60% throughput while only sacrificing 0.4% top-1 accuracy on ImageNet-1K, outperforming current token pruning methods on both accuracy and efficiency. Code is available at https://github.com/YifanXu74/Evo-ViT.