Token-Label Alignment for Vision Transformers (ICCV 2023)

Paper

This is the pytorch implementation for the paper: Token-Label Alignment for Vision Transformers.

Han Xiao*, Wenzhao Zheng*, Zheng Zhu, Jie Zhou, and Jiwen Lu

Highlights

• Improve your ViTs by ~0.7% with a simple --tl-align command.
• Improvements in accuracy, generalization, and robustness without additional computation during inference.
• Efficient alignment of token labels without distillation.

Results

Model	Image Size	Params	FLOPs	Top-1 Acc.(%)	Top-5 Acc.(%)
DeiT-T	$224^2$	5.7M	1.6G	72.2	91.3
+TL-Align	$224^2$	5.7M	1.6G	73.2	91.7
DeiT-S	$224^2$	22M	4.6G	79.8	95.0
+TL-Align	$224^2$	22M	4.6G	80.6	95.0
DeiT-B	$224^2$	86M	17.5G	81.8	95.5
+TL-Align	$224^2$	86M	17.5G	82.3	95.8
Swin-T	$224^2$	29M	4.5G	81.2	95.5
+TL-Align	$224^2$	29M	4.5G	81.4	95.7
Swin-S	$224^2$	50M	8.8G	83.0	96.3
+TL-Align	$224^2$	50M	8.8G	83.4	96.5
Swin-B	$224^2$	88M	15.4G	83.5	96.4
+TL-Align	$224^2$	88M	15.4G	83.7	96.5

Usage

Prerequisites

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

You need to install PyTorch 1.7.0+ and torchvision 0.8.1+ and pytorch-image-models 0.3.2:

conda install -c pytorch pytorch torchvision
pip install timm==0.3.2

Data preparation

Download and extract ImageNet train and val images from http://image-net.org/. The directory structure is the standard layout for the torchvision datasets.ImageFolder, and the training and validation data are expected to be in the train/ folder and val folder respectively:

/path/to/imagenet/
  train/
    class1/
      img1.jpeg
    class2/
      img2.jpeg
  val/
    class1/
      img3.jpeg
    class/2
      img4.jpeg

Training by token-label alignment

To enable token-label alignment during training, you can simply add a --tl-align in your training script. For example, for DeiT-small, run:

python3 -m torch.distributed.launch --nproc_per_node=8 --use_env main_tla.py \
--model deit_small_patch16_224 \
--batch-size 128    \
--mixup 0.0 \
--tl-align \
--data-path /path/to/imagenet  \
--output_dir /path/to/output  \

or

bash train_deit_small_tla.sh

This should give 80.6% top-1 accuracy after 300 epochs of training.

Evaluation

The evaluation of models trained by our token-label alignment is the same as timm. You can also find your validation accuracy during training.

For Deit-small, run:

python main_tla.py --eval --resume checkpoint.pth --model deit_small_patch16_224 --data-path /path/to/imagenet

Citation

If you find this project useful in your research, please cite:

@article{xiao2022token,
    title={Token-Label Alignment for Vision Transformers},
    author={Xiao, Han and Zheng, Wenzhao and Zhu, Zheng and Zhou, Jie and Lu, Jiwen},
    journal={arXiv preprint arXiv:2210.06455},
    year={2022}
}