LGViT

Official PyTorch implementation of "LGViT: Dynamic Early Exiting for Accelerating Vision Transformer" (ACM MM 2023)

Usage

First, clone the repository locally:

git clone https://github.com/lostsword/LGViT

Then, install PyTorch and transformers 4.26.0

conda create -n lgvit python=3.9.13
pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113
pip install transformers==4.26.0 datasets==2.9.0 evaluate==0.4.0 timm==0.6.13 wandb==0.14.2 ipykernel scikit-learn

Enter the scripts folder to execute the scripts for training and evaluation

cd ./scripts

• train_base_deit.sh / train_base_swin.sh

  This is for fine-tuning base models.

• train_baseline_deit.sh / train_baseline_swin.sh

  This is for fine-tuning 1st stage LGViT models and baseline models.

• train_distillation_deit.sh / train_distillation_swin.sh

  This is for fine-tuning 2nd stage LGViT models.

• eval_highway_deit.sh / eval_highway_swin.sh

​ This is for evaluating fine-tuned models.

Before running the script, modify the path and model_path in the script to be appropriate.

Training

To fine-tune a ViT backbone, run:

source train_base_deit.sh

To fine-tune a LGViT models, run:

source train_baseline_deit.sh
source train_distillation_deit.sh

Evaluation

To evaluate a fine-tuned ViT, run:

source eval_highway_deit.sh

Some Hyperparameters Settings

• Exiting points settings

	ViT-EE	Others	LGViT
ViT-B/16	[6]	[1,2,3,4,5,6,7,8,9,10,11]	[4,5,6,7,8,9,10,11]
DeiT-B	[6]	[1,2,3,4,5,6,7,8,9,10,11]	[4,5,6,7,8,9,10,11]
Swin-B	[12]	[4,7,10,13,16,19,22,23]	[4,7,10,13,16,19,22,23]

Other hyperparameters are are kept unchanged from the original baselines.

Acknowledgments

This repository is built upon the transformers and DeeBERT library. Thanks for these awesome open-source projects!

Citation

If you find our work or this code useful, please consider citing the corresponding paper:

@article{xu2023lgvit,
  title={LGViT: Dynamic Early Exiting for Accelerating Vision Transformer},
  author={Xu, Guanyu and Hao, Jiawei and Shen, Li and Hu, Han and Luo, Yong and Lin, Hui and Shen, Jialie},
  journal={arXiv preprint arXiv:2308.00255},
  year={2023}
}