This repository includes the official project for the paper: TransMix: Attend to Mix for Vision Transformers, CVPR 2022

Improve your Vision Transformer (ViT) by ~1% on ImageNet with minimal computational cost and a simple --transmix.

First, clone the repo:

git clone https://github.com/Beckschen/TransMix.git

Then, you need to install the required packages including: Pytorch version 1.7.1, torchvision version 0.8.2, Timm version 0.5.4 and pyyaml. To install all these packages, simply run

pip3 install -r requirements.txt

Download and extract the ImageNet dataset to data folder. Suppose you're using 8 GPUs for training, then simply run

bash ./distributed_train.sh 8 data/ --config $YOUR_CONFIG_PATH_HERE

By default, all our config files have enabled the training with TransMix. If you want to enable TransMix during the training of your own model, you can add a --transmix in your training script. For example:

python3 -m torch.distributed.launch --nproc_per_node=8 train.py data/ --config $YOUR_CONFIG_PATH_HERE --transmix

Or you can simply specify transmix: True in your yaml config file like what we did in deit_s_transmix.

To evaluate your model trained with TransMix, please refer to timm. You can also find your validation accuracy during training.

Coming soon!

This repository is built using the Timm library and the DeiT repository.

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

If you find our code helpful for your research, please using the following bibtex to cite our paper:

@InProceedings{transmix,
title = {TransMix: Attend to Mix for Vision Transformers},
author = {Chen, Jie-Neng and Sun, Shuyang and He, Ju and Torr, Philip and Yuille, Alan and Bai, Song},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2022}
}