Less is More: Pay Less Attention in Vision Transformers

Official PyTorch implementation of Less is More: Pay Less Attention in Vision Transformers.

By Zizheng Pan, Bohan Zhuang, Haoyu He, Jing Liu and Jianfei Cai.

In our paper, we present a novel Less attention vIsion Transformer (LIT), building upon the fact that convolutions, fully-connected (FC) layers, and self-attentions have almost equivalent mathematical expressions for processing image patch sequences. LIT uses pure multi-layer perceptrons (MLPs) to encode rich local patterns in the early stages while applying self-attention modules to capture longer dependencies in deeper layers. Moreover, we further propose a learned deformable token merging module to adaptively fuse informative patches in a non-uniform manner.

If you use this code for a paper please cite:

@article{pan2021less,
  title={Less is More: Pay Less Attention in Vision Transformers},
  author={Pan, Zizheng and Zhuang, Bohan and He, Haoyu and Liu, Jing and Cai, Jianfei},
  journal={arXiv preprint arXiv:2105.14217},
  year={2021}
}

Usage

First, clone this repository.

git clone https://github.com/MonashAI/LIT

Next, create a conda virtual environment.

# Make sure you have a NVIDIA GPU.
cd LIT/
bash setup_env.sh [conda_install_path] [env_name]

# For example
bash setup_env.sh /home/anaconda3 lit

Note: We use PyTorch 1.7.1 with CUDA 10.1 for all experiments. The setup_env.sh has illustrated all dependencies we used in our experiments. You may want to edit this file to install a different version of PyTorch or any other packages.

Data Preparation

Download the ImageNet 2012 dataset from here, and prepare the dataset based on this script. The file structure should look like:

imagenet
├── train
│   ├── class1
│   │   ├── img1.jpeg
│   │   ├── img2.jpeg
│   │   └── ...
│   ├── class2
│   │   ├── img3.jpeg
│   │   └── ...
│   └── ...
└── val
    ├── class1
    │   ├── img4.jpeg
    │   ├── img5.jpeg
    │   └── ...
    ├── class2
    │   ├── img6.jpeg
    │   └── ...
    └── ...

Model Zoo

We provide baseline LIT models pretrained on ImageNet 2012.

Training and Evaluation

In our implementation, we have different training strategies for LIT-Ti and other LIT models. Therefore, we provide two codebases.

For LIT-Ti, please refer to code_for_lit_ti.

For LIT-S, LIT-M, LIT-B, please refer to code_for_lit_s_m_b.

License

This repository is released under the Apache 2.0 license as found in the LICENSE file.

Acknowledgement

This repository has adopted codes from DeiT, PVT and Swin, we thank the authors for their open-sourced code.