hazdzz / STGCN

The PyTorch implementation of STGCN.

Spatio-Temporal Graph Convolutional Networks

About

The PyTorch implementation of STGCN from the paper Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting.

Paper

https://arxiv.org/abs/1709.04875

Citation

@inproceedings{10.5555/3304222.3304273,
author = {Yu, Bing and Yin, Haoteng and Zhu, Zhanxing},
title = {Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting},
year = {2018},
isbn = {9780999241127},
publisher = {AAAI Press},
booktitle = {Proceedings of the 27th International Joint Conference on Artificial Intelligence},
pages = {3634–3640},
numpages = {7},
series = {IJCAI'18}
}

Related works

1. TCN: An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling
2. GLU and GTU: Language Modeling with Gated Convolutional Networks
3. ChebNet: Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering
4. GCN: Semi-Supervised Classification with Graph Convolutional Networks

Related code

1. TCN: https://github.com/locuslab/TCN
2. ChebNet: https://github.com/mdeff/cnn_graph
3. GCN: https://github.com/tkipf/pygcn

Dataset

Source

1. METR-LA: DCRNN author's Google Drive
2. PEMS-BAY: DCRNN author's Google Drive
3. PeMSD7(M): STGCN author's GitHub repository

Preprocessing

Using the formula from ChebNet：

Model structure

Differents of code between mine and author's

1. Fix bugs
2. Add Early Stopping approach
3. Add Dropout approach
4. Offer a different set of hyperparameters
5. Offer config files for two different categories graph convolution (ChebyGraphConv and GraphConv)
6. Add datasets METR-LA and PEMS-BAY
7. Adopt a different data preprocessing method

Requirements

To install requirements:

pip3 install -r requirements.txt