asarigun / GraphMixerNetworks

Official Implementation of Graph Mixer Networks

Graph Nasreddin Networks

In this model, we try to use MLP-Mixer in Graph Neural Networks. The motivation for this project is to replace Transformers and Message Passing with MLP-Mixers in Graph Neural Nets

Check out our paper below for more details

Graph Mixer Networks,
Ahmet Sarıgün
Arxiv, 2023

Overview

• geometric_linear.py: Linear Layer from PyG Source Code for Graph Nasreddin Networks
• gmn_layer.py: Graph Nasreddin Layer
• gmn_train_zinc.py: Graph Nasreddin Network Training on ZINC Dataset

Usage

python gmn_train_zinc.py

License

MIT

Acknowledgement

The name of the Nasreddin coming from Anatolian figure Nasreddin Hodja's story called 'What if it happens?'. Also, while doing benchmarking, we use the PNA paper implementation in PyTorch Geometric. Special thanks to authors for sharing code!

You can find the story behind the Graph Mixer Nets here!

Citation

@article{sarigun2023graph,
  title={Graph Mixer Networks},
  author={Sar{\i}g{\"u}n, Ahmet},
  journal={arXiv preprint arXiv:2301.12493},
  year={2023}
}