Skip to content

Latest commit

 

History

History
88 lines (72 loc) · 3.01 KB

README.md

File metadata and controls

88 lines (72 loc) · 3.01 KB
Raw

Distribution-Aware Prompt Tuning for Vision-Language Models

Official pytorch implementation of "Distribution-Aware Prompt Tuning for Vision-Language Models" (ICCV 2023).

Setup

Clone repository

git clone https://github.com/mlvlab/DAPT.git
cd DAPT

Prepare dataset

Follow DATASET.md to install the datasets.

Setup conda environment

Before creating the environment, you should modify appropriate conda path in env.yaml

conda env create —-file env.yaml
conda activate dapt
pip install torch==1.10.0+cu111 torchvision==0.11.0+cu111 torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html

Setup Dassl.pytorch package

cd Dassl.pytorch
python setup.py develop
cd ..

Run

Dataset path setting

Modify the data path $DATA in main.sh, gen_prototype.sh, and eval.sh to match the path to the dataset you downloaded.

Generate prototype

When the dataset is ready, you can generate the prototype as follows.

bash scripts/gen_prototype.sh [gpu_id]

Few-shot image classification

Below is an example of Caltech101 for each shot.

Note that for ImageNet, we use configs/trainers/DAPT/vit_b16_ep50.yaml for all settings following CoOp.

# 1shot
bash scripts/main.sh caltech101 1 [gpu_id]
# 2shots
bash scripts/main.sh caltech101 2 [gpu_id]
# 4shots
bash scripts/main.sh caltech101 4 [gpu_id]
# 8shots
bash scripts/main.sh caltech101 8 [gpu_id]
# 16shots
bash scripts/main.sh caltech101 16 [gpu_id]

Domain generalization

Before domain generalization, you should completed few-shot image classification on ImageNet.

After the few-shot image classification experiment on ImageNet is finished, you can load the model learned on ImageNet using --eval-only command to conduct domain generalization on imagenetv2, imagenet-sketch, imagenet-a, and imagenet-r.

bash scripts/eval.sh [gpu_id]

Acknowledgement

This repository is built upon Dassl.pytorch, CoOp, and VPT. We thank the authors for their code.

Citation

If you use this code in your research, please kindly cite the following paper:

@InProceedings{Cho_2023_ICCV,
    author    = {Cho, Eulrang and Kim, Jooyeon and Kim, Hyunwoo J},
    title     = {Distribution-Aware Prompt Tuning for Vision-Language Models},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {22004-22013}
}

License

Licensed under MIT License

Copyright (c) 2023 MLV Lab (Machine Learning and Vision Lab at Korea University)