Super Vision Transformer

TL;DR

SuperViT is a versatile vision transformer training paradigm that reduces computational costs and improves accuracy by enabling a single model to adapt to various input sizes and token retention rates, enhancing efficiency and performance.

Contribution

We introduce SuperViT, a novel training method for vision transformers that allows a single model to adapt to different computational budgets and input configurations, outperforming existing efficient transformer models.

Findings

Reduces 2x FLOPs of DeiT-S with increased accuracy

Outperforms SOTA models at the same FLOPs level

Provides a versatile transformer model adaptable to hardware constraints

Abstract

We attempt to reduce the computational costs in vision transformers (ViTs), which increase quadratically in the token number. We present a novel training paradigm that trains only one ViT model at a time, but is capable of providing improved image recognition performance with various computational costs. Here, the trained ViT model, termed super vision transformer (SuperViT), is empowered with the versatile ability to solve incoming patches of multiple sizes as well as preserve informative tokens with multiple keeping rates (the ratio of keeping tokens) to achieve good hardware efficiency for inference, given that the available hardware resources often change from time to time. Experimental results on ImageNet demonstrate that our SuperViT can considerably reduce the computational costs of ViT models with even performance increase. For example, we reduce 2x FLOPs of DeiT-S while increasing the Top-1 accuracy by 0.2% and 0.7% for 1.5x reduction. Also, our SuperViT significantly outperforms existing studies on efficient vision transformers. For example, when consuming the same amount of FLOPs, our SuperViT surpasses the recent state-of-the-art (SOTA) EViT by 1.1% when using DeiT-S as their backbones. The project of this work is made publicly available at https://github.com/lmbxmu/SuperViT.