The flow chart of Multi-Sub: Multi-Sub obtains a desired clustering based on the subspace spanned by reference words provided by GPT-4 using users' high-level interest.

Folder Structure

The project is organized as follows:

.
├── clip/                  
├── dataset/               # Contains datasets for training and evaluation
│   ├── fruit/             # Dataset for Fruit (Please download the dataset via the link provided in the Datasets section, then extract it and place corresponding folders in the specified directory.)
│   │   ├── color/         # Sub-dataset for fruit color
│   │   ├── instance/      # Sub-dataset for fruit instances
│   │   └── species/       # Sub-dataset for fruit species
│   ├── cifar10/           # Dataset for CIFAR-10 (Please download the dataset via the link provided in the Datasets section, then extract it and place corresponding folders in the specified directory.)
│   │   ├── type/          # Sub-dataset for CIFAR-10 type clustering (e.g., transportation, animals)
│   │   └── environment/   # Sub-dataset for CIFAR-10 environment clustering (e.g., land, air, water)
├── gpt.py                 # Sends a prompt to OpenAI's GPT model and retrieves generated reference words
├── main.py                # Main script to run training and evaluation
├── parse.py               # Argument parsing for command-line execution
├── README.md              # This is the README file
├── requirements.txt       # Dependencies are required for running the project
└── setup.py               # Installation setup

Requirements

To run this project, ensure you have the following dependencies installed:

pip install -r requirements.txt

Datasets

Please refer to Fruit and CIFAR-10 to download datasets, and create a dataset directory according to the folder structure and place the datasets in it.

Training and evaluation

Please change dataset_path_dict to adapt to different datasets

python main.py

Bibtex

Please cite our paper if you use this code in your own work:

@misc{yao2024customizedmultipleclusteringmultimodal,
      title={Customized Multiple Clustering via Multi-Modal Subspace Proxy Learning}, 
      author={Jiawei Yao and Qi Qian and Juhua Hu},
      year={2024},
      eprint={2411.03978},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2411.03978}, 
}