Adaptive Token Sampling For Efficient Vision Transformers

TL;DR

This paper introduces a differentiable, parameter-free Adaptive Token Sampler (ATS) that can be integrated into vision transformers to dynamically select significant tokens, reducing computational costs by 2X while maintaining accuracy.

Contribution

The novel ATS module enables adaptive token sampling in vision transformers, improving efficiency without additional training or parameters, and can be added to existing models as a plug-and-play component.

Findings

Reduces GFLOPs by 2X on multiple vision transformers

Maintains state-of-the-art accuracy on ImageNet and Kinetics datasets

Can be integrated into pre-trained models without retraining

Abstract

While state-of-the-art vision transformer models achieve promising results in image classification, they are computationally expensive and require many GFLOPs. Although the GFLOPs of a vision transformer can be decreased by reducing the number of tokens in the network, there is no setting that is optimal for all input images. In this work, we therefore introduce a differentiable parameter-free Adaptive Token Sampler (ATS) module, which can be plugged into any existing vision transformer architecture. ATS empowers vision transformers by scoring and adaptively sampling significant tokens. As a result, the number of tokens is not constant anymore and varies for each input image. By integrating ATS as an additional layer within the current transformer blocks, we can convert them into much more efficient vision transformers with an adaptive number of tokens. Since ATS is a parameter-free module, it can be added to the off-the-shelf pre-trained vision transformers as a plug and play module, thus reducing their GFLOPs without any additional training. Moreover, due to its differentiable design, one can also train a vision transformer equipped with ATS. We evaluate the efficiency of our module in both image and video classification tasks by adding it to multiple SOTA vision transformers. Our proposed module improves the SOTA by reducing their computational costs (GFLOPs) by 2X, while preserving their accuracy on the ImageNet, Kinetics-400, and Kinetics-600 datasets.