Retrieval-Augmented Primitive Representations for Compositional Zero-Shot Learning

• C. Jing, Y. Li, H. Chen, C. Shen, Retrieval-Augmented Primitive Representations for Compositional Zero-Shot Learning. in AAAI 2024 (PDF)

Setup

conda create --name rapr python=3.8
conda activate rapr
pip install git+https://github.com/openai/CLIP.git
pip install -r requirements.txt

Download Dataset

We conduct experiments on commonly used MIT-States, UT-Zappos, and C-GQA. The datasets can be downloaded via

sh download_data/download_data.sh

The databases for each dataset are in the directory of "data/database".

Training

python train.py --dataset mit-states

Evaluation

The evaluations are conducted in two settings: closed-world and open-world.

python test.py --dataset mit-states
python test.py --dataset mit-states --open_world True

Citation

@inproceedings{jing2024retrieval,
  title={Retrieval-Augmented Primitive Representations for Compositional Zero-Shot Learning},
  author={Jing, Chenchen and Li, Yukun and Chen, Hao and Shen, Chunhua},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={3},
  pages={2652--2660},
  year={2024}
}

Acknowledgement

The implementation of our method is partly based on the following codebases, DFSP and CZSL. We gratefully thank the authors for their wonderful works.