Skip to content

Latest commit

 

History

History
85 lines (59 loc) · 4.4 KB

README.md

File metadata and controls

85 lines (59 loc) · 4.4 KB
Raw

ESGNN

Code for "Echo state graph neural networks with analogue random resistive memory arrays."

DOI:10.1038/s42256-023-00609-5 License PEP8

Abstract

Recent years have witnessed a surge of interest in learning representations of graph-structured data, with applications from social networks to drug discovery. However, graph neural networks, the machine learning models for handling graph-structured data, face significant challenges when running on conventional digital hardware, including the von Neumann bottleneck in efficiency incurred by physically separated memory and processing units, and a high training cost. Here we present a hardware-software co-design to address these challenges, by designing an echo state graph neural network based on random resistor arrays, which are built from low-cost, nanoscale and stackable resistors for efficient in-memory computing. The approach leverages the intrinsic stochasticity of dielectric breakdown in the resistors to implement random projections in hardware for an echo state network that effectively minimizes the training complexity thanks to its fixed and random weights. The system demonstrates state-of-the-art performance on both graph classification using the MUTAG and COLLAB datasets and node classification using the CORA dataset, achieving 2.16×, 35.42×, and 40.37× improvement of energy efficiency for a projected random resistor-based hybrid analogue-digital system over a state-of-the-art graphics processing unit and 99.35%, 99.99%, and 91.40% reduction of backward pass complexity compared to conventional graph learning. The results point to a promising direction for the next generation AI system for graph learning.

Requirements

The codes are tested on Ubuntu 20.04, CUDA 11.1 with the following packages:

torch == 1.9.0
torch-geometric == 1.7.2
tensorboad == 2.5.0
scipy == 1.7.0
numpy == 1.20.2

Note: The graph classification tasks can be executed without torch-geometric (PyG).

Installation

You can install the required dependencies with the following code.

conda create -n ESGNN python=3.8
conda activate ESGNN
conda install pytorch=1.9.0 cudatoolkit=11.1 -c pytorch -c conda-forge --yes
pip install tensorboard=2.5.0
CUDA=cu111
TORCH=1.9.0
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-${TORCH}+${CUDA}.html 
pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-${TORCH}+${CUDA}.html 
pip install torch-geometric==1.7.2

Demo

For graph classification on MUTAG dataset, run the following line in terminal:

bash run_mutag.sh

For graph classification on COLLAB dataset, run the following line in terminal:

bash run_collab.sh

For node classification on Cora dataset, run the following line in terminal:

bash run_cora.sh

Note: The code for graph classification simulation provided in this demo is the same with that using the random resistive memory hardware. The only difference is the weight multiplication function WeightMultiplication performs hardware calls to a Xilinx FPGA in the customized system via pynq.dma.

Dataset

Both the processed MUTAG and COLLAB datasets are provided in data folder. The raw data can be downloaded here.

The Cora dataset can be automatically downloaded in the code via PyG. The raw data can also be downloaded here.

The experimental and simulation measured source data are provided in source_data subfolder.

CItation

If you find this work interesting, welcome to cite our paper!

@article{wang2023echo,
  title={Echo state graph neural networks with analogue random resistive memory arrays},
  author={Wang, Shaocong and Li, Yi and Wang, Dingchen and Zhang, Woyu and Chen, Xi and Dong, Danian and Wang, Songqi and Zhang, Xumeng and Lin, Peng and Gallicchio, Claudio and others},
  journal={Nature Machine Intelligence},
  pages={1--10},
  year={2023},
  publisher={Nature Publishing Group UK London}
}