GraphNAS

Overview

Graph Neural Architecture Search (GraphNAS for short) enables automatic design of the best graph neural architecture based on reinforcement learning. This directory contains code necessary to run GraphNAS. Specifically, GraphNAS first uses a recurrent network to generate variable-length strings that describe the architectures of graph neural networks, and then trains the recurrent network with a policy gradient algorithm to maximize the expected accuracy of the generated architectures on a validation data set. An illustration of GraphNAS is shown as follows:

A recurrent network (Controller RNN) generates descriptions of graph neural architectures (Child model GNNs). Once an architecture is generated by the controller, GraphNAS trains the architecture m on a given graph and test on a validate set . The validation result is taken as the reward of the recurrent network.

To improve the search efficiency of GraphNAS, we restrict the search space from an entire architecture to a concatenation of the best search results built on each single architecture layer. An example of GraphNAS constructing a single GNN layer (the right-hand side) is shown as follows:

In the above example, the layer has two input states and , two intermediate states and , and an output state . The controller at the left-hand side samples from and takes as input of , and then samples "GAT" for processing . The output state collects information from and , and the controller assigns a readout operator "add" and an activation operator "relu" for . As a result, this layer can be described as a list of operators: [0, gcn, 1, gat, add, relu].

Requirements

Recent versions of PyTorch, numpy, scipy, sklearn, dgl, torch_geometric and networkx are required. Ensure that PyTorch 1.1.0 and CUDA 9.0 are installed. Then run:

pip install torch==1.1.0 -f https://download.pytorch.org/whl/cu90/torch_stable.html
pip install -r requirements.txt

If you want to run in docker, you can run:

docker build -t graphnas -f DockerFile . 
docker run -it -v $(pwd):/GraphNAS graphnas python -m eval_scripts.semi.eval_designed_gnn

Running the code

Architecture evaluation

To evaluate our best architecture designed on semi-supervised experiments by training from scratch, run

python -m eval_scripts.semi.eval_designed_gnn

To evaluate our best architecture designed on semi-supervised experiments by training from scratch, run

python -m eval_scripts.sup.eval_designed_gnn

Results

Semi-supervised node classification w.r.t. accuracy

Model	Cora	Citeseer	Pubmed
GCN	81.5+/-0.4	70.9+/-0.5	79.0+/-0.4
SGC	81.0+/-0.0	71.9+/-0.1	78.9+/-0.0
GAT	83.0+/-0.7	72.5+/-0.7	79.0+/-0.3
LGCN	83.3+/-0.5	73.0+/-0.6	79.5+/-0.2
DGCN	82.0+/-0.2	72.2+/-0.3	78.6+/-0.1
ARMA	82.8+/-0.6	72.3+/-1.1	78.8+/-0.3
APPNP	83.3+/-0.6	71.8+/-0.4	80.2+/-0.2
simple-NAS	81.4+/-0.6	71.7+/-0.6	79.5+/-0.5
GraphNAS	84.3+/-0.4	73.7+/-0.2	80.6+/-0.2

Supervised node classification w.r.t. accuracy

Model	Cora	Citeseer	Pubmed
GCN	90.2+/-0.0	80.0+/-0.3	87.8+/-0.2
SGC	88.8+/-0.0	80.6+/-0.0	86.5+/-0.1
GAT	89.5+/-0.3	78.6+/-0.3	86.5+/-0.6
LGCN	88.7+/-0.5	79.2+/-0.4	OOM
DGCN	88.4+/-0.2	78.0+/-0.2	88.0+/-0.9
ARMA	89.8+/-0.1	79.9+/-0.6	88.1+/-0.2
APPNP	90.4+/-0.2	79.2+/-0.4	87.4+/-0.3
random-NAS	90.0+/-0.3	81.1+/-0.3	90.7+/-0.6
simple-NAS	90.1+/-0.3	79.6+/-0.5	88.5+/-0.2
GraphNAS	90.6+/-0.3	81.3+/-0.4	91.3+/-0.3

Architectures designed in supervised learning are showed as follow:

The architecture G-Cora designed by GraphNAS on Cora is [0, gat6, 0, gcn, 0, gcn, 2, arma, tanh, concat], the architecture G-Citeseer designed by GraphNAS on Citeseer is [0, identity, 0, gat6, linear, concat], the architecture G-Pubmed designed by GraphNAS on Pubmed is [1, gat8, 0, arma, tanh, concat].

Searching for new architectures

To design an entire graph neural architecture based on the search space described in Section 3.2, please run:

python -m graphnas.main --dataset Citeseer

To design an entire graph neural architecture based on the search space described in Section 3.4, please run:

python -m graphnas.main --dataset Citeseer --supervised True --search_mode micro

Be aware that different runs would end up with different local minimum.

Acknowledgements

This repo is modified based on DGL and PYG.