IGLU: EffIcient GCN Training via Lazy Updates

This repository contains code accompanying the paper IGLU: EffIcient GCN Training via Lazy Updates.

The repository is subdivided into two key directories:

• src - This directory contains the main runner scripts, along with dataset specific architectures. Further details are presented within the directory.
• makedata - This directory contains instructions for creating data in the format IGLU uses.

Dependencies

• For OGB Datasets (Proteins and Products): Compatible version of Pytorch Geometric is needed. Installation instructions can be used from this link.

• For other datasets, we use standard python packages - NumPy, SciPy, Scikit-Learn, Json, NetworkX (Older Version 1.x might be required).

• Tensorflow Version Used: 1.15.2

Queries

In case of any questions, feel free to raise an issue.

Citing our work

To cite our work, kindly use the BibTeX below.

@inproceedings{
narayanan2022iglu,
title={{IGLU}: Efficient {GCN} Training via Lazy Updates},
author={S Deepak Narayanan and Aditya Sinha and Prateek Jain and Purushottam Kar and Sundararajan Sellamanickam},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=5kq11Tl1z4}
}