/TokenMix

TokenMix: Rethinking Image Mixing for Data Augmentation in Vision Transformers (ECCV 2022)

Primary LanguagePythonMIT LicenseMIT

Pytorch implementation of TokenMix (ECCV 2022)

tenser

This repo is the offcial implementation of the paper TokenMix: Rethinking Image Mixing for Data Augmentation in Vision Transformers

@article{UniNet,
  author  = {Jihao Liu, Boxiao Liu, Hang Zhou, Yu Liu, Hongsheng Li},
  journal = {arXiv:2207.08409},
  title   = {TokenMix: Rethinking Image Mixing for Data Augmentation in Vision Transformers},
  year    = {2022},
}

Update

8/9/2022 Update the source code.

Preparation

Data

Following TokenLabeling to prepare ImageNet data and label maps generated with NFNet-F6.

Environment

The code is tested with torch==1.11 and timm==0.5.4.

Run experiments

Currently, we supporting running experiments with slurm. You can reproduce the results of Deit-small as follows:

sh exp/deit_small/run.sh partition

Models

Model epochs Top-1 Acc. Ckpt
Deit-tiny 300 73.2 -
Deit-small 300 80.8 -
Deit-base 300 82.9 -