TGOpt: Redundancy-Aware Optimizations for Temporal Graph Attention Networks

This repository is the artifact for our TGOpt paper (PPoPP 2023). For the archived artifact, see our Zenodo record.

[Paper, Zenodo]

Requirements

• Python 3.7, PyTorch 1.12, dependencies in requirements.txt
• g++ >= 7.2, OpenMP >= 201511, Intel TBB >= 2020.1 (for the C++ extension)

This repo provides scripts for setting up docker (see Dockerfile) or an AWS GPU instance (see scripts/setup-aws-ec2.sh). Otherwise, we recommend using conda to create a virtual Python environment. If needed, manually compile the extension after installing dependencies:

$ cd extension && make && cd ..

Datasets

The datasets we used in our experiments are provided in our Zenodo artifact under the data/ subdirectory. You can prepare the datasets yourself using the data-* scripts:

$ ./data-download.sh <name>
$ python data-reformat.py -d <name>
$ python data-process.py -d <name>

Trained models

The models used in our experiments are provided in our Zenodo artifact under the saved_models/ subdirectory. You can train different models using the train.py script (be sure to use the same --model flag in later inference commands):

$ python train.py -d <name> --model <prefix> ... [--gpu 0]

Running experiments

The inference.py script is the main entry-point for our experiments. To kick the tires, run a small experiment:

$ python inference.py -d snap-msg --model tgat --prefix test --opt-all [--gpu 0]

Use the -h flag to see all available options. To reproduce our experiments, see Appendix A in our paper.

Citation

You can cite our paper with the following:

@inproceedings{tgopt:ppopp23,
  author = {Wang, Yufeng and Mendis, Charith},
  title = {{TGOpt}: Redundancy-Aware Optimizations for Temporal Graph Attention Networks},
  year = {2023},
  publisher = {Association for Computing Machinery},
  booktitle = {Proceedings of the 28th ACM SIGPLAN Annual Symposium on Principles and Practice of Parallel Programming},
  pages = {354–368},
  numpages = {15},
  series = {PPoPP '23}
}