DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification

Created by Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, Cho-Jui Hsieh

This repository contains PyTorch implementation for DynamicViT (NeurIPS 2021).

DynamicViT is a dynamic token sparsification framework to prune redundant tokens in vision transformers progressively and dynamically based on the input. Our method can reduces over 30% FLOPs and improves the throughput by over 40% while the drop of accuracy is within 0.5% for various vision transformers

Our code is based on pytorch-image-models, DeiT and LV-ViT

[Project Page] [arXiv]

Model Zoo

We provide our DynamicViT models pretrained on ImageNet:

name	arch	rho	acc@1	acc@5	FLOPs	url
DynamicViT-256/0.7	`deit_256`	0.7	76.532	93.118	1.3G	Google Drive / Tsinghua Cloud
DynamicViT-384/0.7	`deit_small`	0.7	79.316	94.676	2.9G	Google Drive / Tsinghua Cloud
DynamicViT-LV-S/0.5	`lvvit_s`	0.5	81.970	95.756	3.7G	Google Drive / Tsinghua Cloud
DynamicViT-LV-S/0.7	`lvvit_s`	0.7	83.076	96.252	4.6G	Google Drive / Tsinghua Cloud
DynamicViT-LV-M/0.7	`lvvit_m`	0.7	83.816	96.584	8.5G	Google Drive / Tsinghua Cloud

Usage

Requirements

torch>=1.7.0
torchvision>=0.8.1
timm==0.4.5

Data preparation: download and extract ImageNet images from http://image-net.org/. The directory structure should be

│ILSVRC2012/
├──train/
│  ├── n01440764
│  │   ├── n01440764_10026.JPEG
│  │   ├── n01440764_10027.JPEG
│  │   ├── ......
│  ├── ......
├──val/
│  ├── n01440764
│  │   ├── ILSVRC2012_val_00000293.JPEG
│  │   ├── ILSVRC2012_val_00002138.JPEG
│  │   ├── ......
│  ├── ......

Model preparation: download pre-trained DeiT and LV-ViT models for training DynamicViT:

model	url
DeiT-Small	link
LVViT-S	Google Drive / Tsinghua Cloud
LVViT-M	Google Drive / Tsinghua Cloud

Since the original links of LVViT models are not available now (#4), we provide the models used in our experiments on Google Drive and Tsinghua Cloud.

Demo

You can try DynamicViT on Colab . Thank @dirtycomputer for the contribution.

We also provide a Jupyter notebook where you can run the visualization of DynamicViT.

To run the demo, you need to install matplotlib.

Evaluation

To evaluate a pre-trained DynamicViT model on the ImageNet validation set with a single GPU, run:

python infer.py --data-path /path/to/ILSVRC2012/ --arch arch_name --model-path /path/to/model --base_rate 0.7

Training

To train DynamicViT models on ImageNet, run:

DeiT-small

python -m torch.distributed.launch --nproc_per_node=8 --use_env main_dynamic_vit.py  --output_dir logs/dynamic-vit_deit-small --arch deit_small --input-size 224 --batch-size 96 --data-path /path/to/ILSVRC2012/ --epochs 30 --dist-eval --distill --base_rate 0.7

LV-ViT-S

python -m torch.distributed.launch --nproc_per_node=8 --use_env main_dynamic_vit.py  --output_dir logs/dynamic-vit_lvvit-s --arch lvvit_s --input-size 224 --batch-size 64 --data-path /path/to/ILSVRC2012/ --epochs 30 --dist-eval --distill --base_rate 0.7

LV-ViT-M

python -m torch.distributed.launch --nproc_per_node=8 --use_env main_dynamic_vit.py  --output_dir logs/dynamic-vit_lvvit-m --arch lvvit_m --input-size 224 --batch-size 48 --data-path /path/to/ILSVRC2012/ --epochs 30 --dist-eval --distill --base_rate 0.7

You can train models with different keeping ratio by adjusting base_rate. DynamicViT can also achieve comparable performance with only 15 epochs training (around 0.1% lower accuracy compared to 30 epochs).

License

MIT License

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{rao2021dynamicvit,
  title={DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification},
  author={Rao, Yongming and Zhao, Wenliang and Liu, Benlin and Lu, Jiwen and Zhou, Jie and Hsieh, Cho-Jui},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year = {2021}
}

Yu-An-Shih / DynamicViT