Implement end-to-end trainable Graph ConvRNN with PyTorch, by replacing 2D CNN in ConvRNN with GNNs.

This network can be used for time series prediction of correlated data in non Euclidean space (especially graph structure, e.g., metro system).

By far, The following GNNs and RNNs are supported and can be combined at will:

• Graph Neural Networks
  • GAT
  • GCN
• Recursive Neural Networks
  • LSTM
  • GRU

Require PyTorch 1.0+

git clone https://github.com/DarkstartsUp/Graph-ConvRNN.pytorch.git
cd Graph-ConvRNN.pytorch
pip install -r requirements.txt

This file contains the implementation of Graph ConvRNN in PyTorch.

The Graph_ConvRNN module derives from nn.Module so it can be used as any other PyTorch module. Example usage:

from graph_convrnn import Graph_ConvRNN

device = torch.device("cuda:0")
# Model initialization
model = Graph_ConvRNN(
    input_dim=64,
    num_layers=2,
    gnn_hidden_dim=32,
    rnn_hidden_dim=16,
    num_nodes=50,
    rnn_mode='lstm',
    gnn_mode='gcn',
).to(device)

# Test input output
x = torch.rand((128, 10, 50, 64)).to(device)  # [batch_size, sequence_length, num_nodes, feature_dim]
adj = torch.ones((50, 50)).to(device)
_, last_states = graph_convrnn(x, adj)
h = last_states[0][0]  # 0 for layer index, 0 for h index

Two real-world datasets were tested to demonstrate the effectiveness of the network. If you also want to train and test on these two datasets, please visit the download link below and place the unzipped dataset folder in the data folder at the root of the project.

• Metro dataset (detailed description)
• Seattle-Loop dataset (detailed description)

The unzipped file structure:

    - data
	- Metro Ridership Dataset
	- Seattle_Loop_Dataset

Execute the following code to train on the above datasets:

# train on Metro dataset
python metro_train.py

# train on Seattle-Loop dataset
python seattle_train.py

There are 2 sub-dataset in Metro dataset: Metro-Hangzhou and Metro-shanghai. In the paper, The author uses three adjacency matrices: Physical Graph (CONN), Similarity Graph (SML) and Correlation Graph (CORR). In addition, the identity matrix (generated by np.eye(), EYE) and all-1 matrix (generated by np.ones(), ONES) are also tested. Noted that the input data is normalized to mean 0 and standard deviation 1, and the experimental results were also calculated on this basis.

Metro-Hangzhou GCN+LSTM

Metro-Hangzhou GCN+GRU

Metro-Hangzhou GAT+LSTM

Metro-Hangzhou GAT+GRU

METRO-Shanghai GCN+LSTM

METRO-Shanghai GCN+GRU

In the paper, The author uses adjacency matrices including Physical Connection (A) and Loop Free-flow Reachability Matrix during X minites' drive (X in {5, 10, 15, 20, 25}). In addition, the identity matrix (generated by np.eye(), EYE) and all-1 matrix (generated by np.ones(), ONES) are also tested. Noted that the input data is normalized to mean 0 and standard deviation 1, and the experimental results were also calculated on this basis.

Seattle-Loop GCN+LSTM

Seattle-Loop GCN+GRU

This repo borrows some codes from:

• pygcn

• pyGAT

• ConvLSTM_pytorch

• ConvGRU-pytorch

GCN

@inproceedings{kipf2017semi,
  title={Semi-Supervised Classification with Graph Convolutional Networks},
  author={Kipf, Thomas N. and Welling, Max},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2017}
}

GAT

@article{
  velickovic2018graph,
  title="{Graph Attention Networks}",
  author={Veli{\v{c}}kovi{\'{c}}, Petar and Cucurull, Guillem and Casanova, Arantxa and Romero, Adriana and Li{\`{o}}, Pietro and Bengio, Yoshua},
  journal={International Conference on Learning Representations},
  year={2018},
  url={https://openreview.net/forum?id=rJXMpikCZ},
  note={accepted as poster},
}

ConvLSTM

@inproceedings{xingjian2015convolutional,
  title={Convolutional LSTM network: A machine learning approach for precipitation nowcasting},
  author={Xingjian, SHI and Chen, Zhourong and Wang, Hao and Yeung, Dit-Yan and Wong, Wai-Kin and Woo, Wang-chun},
  booktitle={Advances in neural information processing systems},
  pages={802--810},
  year={2015}
}

Metro Dataset

@article{liu2020physical,
  title={Physical-Virtual Collaboration Modeling for Intra-and Inter-Station Metro Ridership Prediction},
  author={Liu, Lingbo and Chen, Jingwen and Wu, Hefeng and Zhen, Jiajie and Li, Guanbin and Lin, Liang},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  year={2020},
  publisher={IEEE}
}

Seattle-Loop Dataset

@article{cui2019traffic,
  title={Traffic graph convolutional recurrent neural network: A deep learning framework for network-scale traffic learning and forecasting},
  author={Cui, Zhiyong and Henrickson, Kristian and Ke, Ruimin and Wang, Yinhai},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  volume={21},
  number={11},
  pages={4883--4894},
  year={2019},
  publisher={IEEE}
}