🌷 PromptHash: Affinity-Prompted Collaborative Cross-Modal Learning for Adaptive Hashing Retrieval [CVPR 2025]

Overview 🌟

PromptHash (Affinity-Prompted Collaborative Cross-Modal Learning for Adaptive Hashing Retrieval) is a method for adaptive hashing retrieval through collaborative cross-modal learning.
This repository contains the code associated with the paper:

"PromptHash: Affinity-Prompted Collaborative Cross-Modal Learning for Adaptive Hashing Retrieval" (accepted to CVPR 2025)
Qiang Zou, Shuli Cheng✉️, and Jiayi Chen

---

Usage 🛠️

1. Clone the repository

   git clone https://github.com/ShiShuMo/PromptHash.git
   cd PromptHash

2. Create and activate the environment

   conda create -n prompthash python=3.12 -y
   conda activate prompthash

3. Install dependencies

   # Install PyTorch 2.7.1
   pip3 install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu128
   
   # Install mamba_ssm
   # See: https://github.com/state-spaces/mamba
   # For CUDA 12.8, download pre-built .whl from the release page
   
   # Install causal-conv1d
   # See: https://github.com/Dao-AILab/causal-conv1d
   # For CUDA 12.8, download pre-built .whl from the release page

---

Data 🗂️

Directory Structure:

dataset
├── coco
│   ├── caption.mat
│   ├── prompt_caption.npz 
│   ├── index.mat
│   └── label.mat 
├── flickr25k
│   ├── caption.mat
│   ├── prompt_caption.npz
│   ├── index.mat
│   └── label.mat
├── nuswide
│   ├── caption.txt
│   ├── prompt_caption.npz
│   ├── index.mat 
│   └── label.mat

• Please preprocess the datasets to the required input format.
• All .mat files have already been uploaded in the ./dataset directory. Use the unzip command to extract them.
• Cleaned datasets (MIRFLICKR25K, MSCOCO, NUSWIDE), including all files required for experiments, are available via Baidu Netdisk.

🔗 Baidu Netdisk Download
Password: kdq7
(Source: CMCL)

---

About prompt_caption.npz 📝

The prompt_caption.npz file for each dataset contains:

• Pseudo-label embeddings generated by large language models.
• Template sentence embeddings for each dataset sample.

These files are generated and preprocessed for direct use in training and evaluation.
You do not need to generate them yourself.
The prompt_caption.npz files are included in the Baidu Netdisk dataset below.

🔗 Baidu Netdisk Download
Password: krua

---

Training 🏋️‍♂️

After preparing the Python environment, pretrained MetaCLIP model, and dataset, you can train the PromptHash model.

• The script will automatically download the required weights. Ensure your network connection is stable.
• See PromptHash_$DATASET$.sh for training scripts.
• For hyperparameter settings, refer to the paper.

---

Evaluation 📊

• See PromptHash_$DATASET$_TEST.sh for evaluation scripts.
• Hash code files generated by our algorithm are included in the repository (test.zip), which can be used directly for verification.

---

Citation 📜

If you find the PromptHash paper and code useful for your research and applications, please cite using this BibTeX:

@InProceedings{Zou_2025_CVPR,
    author    = {Zou, Qiang and Cheng, Shuli and Chen, Jiayi},
    title     = {PromptHash: Affinity-Prompted Collaborative Cross-Modal Learning for Adaptive Hashing Retrieval},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {19649-19658}
}

---

Any Question ✉️

If you have any questions, please feel free to contact:

• Shuli Cheng (cslxju@xju.edu.cn)
• Qiang Zou (admin@52mss.cn)

---

Acknowledgements 🌸

This project is based on open_clip and CMCL — special thanks to all contributors.

---