Simplifying Architecture Search for Graph Neural Network

Overview

This is the code for our paper Simplified Neural Architecture search for Graph Neural Networks, publised in CSSA-CIKM 2020. It is a neural architecture search (NAS) for graph neural network (GNN). To obtain optimal data-specific GNN architectures, we propose the SNAG framework, consisting of a simpler yet more expressive search space and a RL-based search algorithm.

The framewwork is implemented on top of GraphNAS and PyG. The main difference compared with GraphNAS:

1. We provide the implementation of weight sharing strategy.
2. The finetuning stage of GNN.
3. The implementations of Random and Bayesian search algorithms.

Requirements

Latest version of Pytorch-geometric(PyG) is required. More details can be found in here

Python == 3.7.4   Pytorch-geometric>=1.6.3   PyTorch == 1.6.0 

Architecture Search

Search a 3-layer GNN on Cora dataset based on the designed search space, please run:

python -m rlctr.main  --dataset Cora   --layers_of_child_model 3  --shared_initial_step 10   --shared_params True  #SNAG-WS
python -m rlctr.main  --dataset Cora   --layers_of_child_model 3  #SNAG

Other NAS methods, e.g., Random, Bayesian and GraphNAS in Section 4, please run:

python -m rlctr.main  --dataset Cora   --layers_of_child_model 3  --search_mode graphnas    #GraphNAS
python -m rlctr.main  --dataset Cora   --layers_of_child_model 3  --search_mode graphnas  --shared_initial_step 10   --shared_params True  #GraphNAS-WS
python -m rlctr.main  --dataset Cora   --layers_of_child_model 3  --mode random #Random
python -m rlctr.main  --dataset Cora   --layers_of_child_model 3  --mode bayes  #Bayesian

Cite

Please kindly cite our paper if you use this code:

@Technicalreport{zhao2020simplifying,
title={Simplifying Architecture Search for Graph Neural Network},
author={Zhao, Huan and Wei, Lanning and Yao, Quanming},
journal={arXiv preprint arXiv:2008.11652},
year={2020}
}

Misc

If you have any questions about this project, you can open issues, thus it can help more people who are interested in this project. We will reply to your issues as soon as possible. You are also welcomed to reach us by weilanning@4paradigm.com