This is a TensorFlow implementation of DDP-GCN in the following paper: https://arxiv.org/abs/1905.12256. Our codes are mostly built upon the codes of https://github.com/VeritasYin/STGCN_IJCAI-18.
Our code is based on Python3. There are a few dependencies to run the code. We list the major libraries as below.
- tensorflow >= 1.9.0
- numpy >= 1.15
- pandas >= 0.23
- scipy >= 1.1.0
Dependency can be installed using the following command:
pip install -r requirements.txt
We provide the raw dataset in https://github.com/snu-adsl/ddpgcn-dataset.
We defined three types of weighted graphs based on the distances and the direction among the link vectors in the traffic network. For each graph, adjacency matrix W can be computed as below.
We provide the pre-defined weighted graphs in dataset directory. Individual files in a repository are limited to a 100MB maximum size limit, we zip speed files. Please unzip the files when you use the speed data.
Similar with the previous studies, we normalize the input data by z-score method.
We provide the default hyperparamter setups for DDP-GCN(Single), DDP-GCN(Parallel), and DDP-GCN(Stacked) in model-Urban1.json. The results are also available at output directory.
If you need more detailed information about implementation, please read and cite the following paper:
@article{lee2019ddp,
title={DDP-GCN: Multi-graph convolutional network for spatiotemporal traffic forecasting},
author={Lee, Kyungeun and Rhee, Wonjong},
journal={arXiv preprint arXiv:1905.12256},
year={2019}
}