This repository contains the code for reimplementing our ICLR 2024 work, METR.
A Simple Romance Between Multi-Exit Vision Transformer and Token Reduction
Dongyang Liu, Meina Kan, Shiguang Shan, Xilin Chen
The implementation is built upon the EViT codebase (many thanks for their excellent work!). All of their original code is preserved in this repository for easy reuse and comparison, while the code for our method is primarily located in the metr directory.
@inproceedings{
liu2024a,
title={A Simple Romance Between Multi-Exit Vision Transformer and Token Reduction},
author={Dongyang Liu and Meina Kan and Shiguang Shan and Xilin CHEN},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=gJeYtRuguR}
}
Please refer to the guidance in the official EViT repository for environment and data setup.
The code is based on EViT and DiffRate, many thanks to their great work!