xianggebenben / pdADMM-G

pdADMM-G: parallel graph deep learning Alternating Direction Method of Multipliers

pdADMM-G: parallel graph deep learning Alternating Direction Method of Multipliers

This is an implementation of ADMM to achieve communication-efficient model parallelism for the Graph Augmented Multi-Layer Perceptron (GA-MLP) model, as described in our paper:

Junxiang Wang, Hongyi Li(first-coauthor), Zheng Chai, Yongchao Wang, Yue Cheng, and Liang Zhao. Towards Quantized Model Parallelism for Graph-Augmented MLPs Based on Gradient-Free ADMM Framework. IEEE Transactions on Neural Networks and Learning Systems 2022. Paper

serial pdADMM-G is the source code of serial implementation of the pdADMM-G and pdADMM-G-Q algorithms.

parallel pdADMM-G is the source code of parallel implementation of the pdADMM-G and pdADMM-G-Q algorithms.

Requirements

The codebase is implemented in Python 3.8.10. package versions used for development are just below.

torch                1.8.1
torch-cluster        1.5.9
torch-geometric      1.7.1
torch-scatter        2.0.7
torch-sparse         0.6.10
torch-spline-conv    1.2.1
pyarrow              3.0.0
tornado              6.1

Cite

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

@article{wang2022towards,

title={Towards Quantized Model Parallelism for Graph-Augmented MLPs Based on Gradient-Free ADMM framework},

author={Wang, Junxiang and Li, Hongyi and Chai, Zheng and Wang, Yongchao and Cheng, Yue and Zhao, Liang},

journal={IEEE Transactions on Neural Networks and Learning Systems},

year={2022} }