Fitness Landscape Analysis of Graph Neural Network Architecture Search Spaces

The Genetic and Evolutionary Computation Conference (GECCO) - 2021

Matheus Nunes, Paulo M. Fraga (github) and Gisele L. Pappa

Overview

This repository contains code from the original GraphNAS repository and from the [Nunes & Pappa, 2020] fork.

In this paper, we extend the work of Nunes & Pappa, 2020 by:

1. Expanding the search space for each dataset (as done here)
2. Encoding the architectures in two different ways (here)
3. Calculating FLA metrics for both representations (here and here)

The results can be found in this notebook.

Requirements

Recent versions of PyTorch, numpy, pandas, scipy, dgl, and torch_geometric are required.

We have provided a utility script that installs the dependencies, considering the usage of CUDA 10.1. If this is not your CUDA version, follow the instructions on the script.

Example run:

./virtualenv_script.sh  

After executing this script, you will have an Anaconda powered virtual environment called py37 with the dependencies necessary to run the code in this repository.

Running the code

The execution of the code for this paper is divided into two parts: a server that evaluates the architectures, and the client that requests the evaluations. In order to run the server, just cd into the src directory and run:

python -u -m server_client.evaluator_server -p 12345    

This command initiates a server on port 12345.

In order to execute the script that runs the architecture evaluations in this work, just run the codes in the first item of the list in section Overview.

Acknowledgements

This repo is modified based on DGL, PYG and GraphNAS.