Synchronous Spatio-Temporal Graph Transformer: A New Framework for Traffic Data Prediction
[Paper]
Tensorflow > 2.0
This repo was tested with Ubuntu 18.04.1 LTS, Python 3.8, tensorflow 2.5.0, and CUDA 11. But it should be runnable with recent tensoflow versions > 2.0
The dataset is PEMS data for traffic prediction. Download and extract them under S2TAT/data, and make them look like this:
S2TAT/
├──data/
├──PEMS03/
├──PEMS04/
├──PEMS07/
├──PeMSD7L/
└──PeMSD7M/
├──main.py
├──Model.py
├──Layers.py
├──utils.py
└── ...
cd S2TAT
python main.py --config $config_path$
The results of our
| Dataset | MAE | MAPE(%) | MSE | #Parameters(M) |
|---|---|---|---|---|
| PEMS03 | 15.12 | 15.38 | 25.98 | 3.2 |
| PEMS04 | 19.08 | 12.58 | 30.79 | 2.9 |
| PEMS07 | 21.06 | 8.72 | 34.02 | 6.8 |
| PEMS08 | 15.41 | 9.85 | 24.36 | 2.0 |
| PeMSD7M | 2.67 | 6.61 | 5.41 | 2.4 |
| PeMSD7L | 2.85 | 7.25 | 5.83 | 7.7 |
For any questions, please contact Jiahui Chen ([email protected]).
If you find this repo useful for your research, please consider citing the paper
@article{wang2022synchronous,
title={Synchronous Spatiotemporal Graph Transformer: A New Framework for Traffic Data Prediction},
author={Wang, Tian and Chen, Jiahui and L{\"u}, Jinhu and Liu, Kexin and Zhu, Aichun and Snoussi, Hichem and Zhang, Baochang},
journal={IEEE Transactions on Neural Networks and Learning Systems},
year={2022},
publisher={IEEE}
}