Hypergraph Vision Transformers: Images are More than Nodes, More than Edges

CVPR 2025

   

Official PyTorch implementation of "Hypergraph Vision Transformers: Images are More than Nodes, More than Edges" [2025].

Introduction

We proposed HgVT, a vision model that incorporates hypergraph-based structures into vision transformers, improving classification accuracy and enabling implicit feature interpretability.

Dependencies and Installation

• Python >= 3.9.0
• PyTorch >= 2.0.1+cu12.1

git clone https://github.com/RGenDiff/hgvt.git
cd hgvt
pip install -e .

How to Train HgVT

PRECISION="bf16" && \
export ATTN_PRECISION="fp32" && \
export USE_XFORMERS=1 && \
export USE_APEX=0 && \
python train.py \
    --precision "$PRECISION" \
    --gpus 2 \
    --name name_of_run \
    --logdir path/to/log/dir \
    --config config.yaml 

Documentation

1. Architecture Overview
2. ?

To-Do List

• Network code
• Document architecture
• Training code
• Evaluation code
• Feature visualization code

Acknowledgements

This repository uses code from the following open-sourced codebases:

• timm
• ViG

BibTeX

@InProceedings{Fixelle_2025_CVPR,
    author    = {Fixelle, Joshua},
    title     = {Hypergraph Vision Transformers: Images are More than Nodes, More than Edges},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {9751-9761}
}