The PyTorch version of STGCN implemented by the paper Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting with tons of bugs fixed.
https://arxiv.org/abs/1709.04875
- TCN: An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling
- GLU and GTU: Language Modeling with Gated Convolutional Networks
- ChebyNet: Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering
- GCN: Semi-Supervised Classification with Graph Convolutional Networks
- TCN: https://github.com/locuslab/TCN
- ChebyNet: https://github.com/mdeff/cnn_graph
- GCN: https://github.com/tkipf/pygcn
- METR-LA: DCRNN author's Google Drive
- PEMS-BAY: DCRNN author's Google Drive
- PeMSD7(M): STGCN author's GitHub repository
Using the formula from ChebyNet：
- Fix tons of bugs
- Add Early Stopping approach
- Add Dropout approach
- Offer a different set of hyperparameters
- Offer config files for two different categories graph convolution
- Add datasets METR-LA and PEMS-BAY
- Using a different data preprocessing method
To install requirements:
pip3 install -r requirements.txt| Model (paper) | Model (code) | Laplacian matrix type | Gated activation function | MAE | RMSE | WMAPE |
|---|---|---|---|---|---|---|
| STGCN (Cheb) | STGCN_ChebConv (Ks=3, Kt=3) | sym | GLU | 3.825249 | 7.949693 | 7.530186% |
| STGCN (1st) | STGCN_GCNConv (Kt=3) | sym | GLU | 3.703660 | 7.685864 | 7.290832% |
| Model (paper) | Model (code) | Laplacian matrix type | Gated activation function | MAE | RMSE | WMAPE |
|---|---|---|---|---|---|---|
| STGCN (Cheb) | STGCN_ChebConv (Ks=3, Kt=3) | sym | GLU | 4.789775 | 9.501917 | 9.430166% |
| STGCN (1st) | STGCN_GCNConv (Kt=3) | sym | GLU | 4.518740 | 8.863177 | 8.896550% |
| Model (paper) | Model (code) | Laplacian matrix type | Gated activation function | MAE | RMSE | WMAPE |
|---|---|---|---|---|---|---|
| STGCN (Cheb) | STGCN_ChebConv (Ks=3, Kt=3) | sym | GLU | 6.047641 | 11.888628 | 11.909882% |
| STGCN (1st) | STGCN_GCNConv (Kt=3) | sym | GLU | 5.997484 | 11.498759 | 11.811108% |
| Model (paper) | Model (code) | Laplacian matrix type | Gated activation function | MAE | RMSE | WMAPE |
|---|---|---|---|---|---|---|
| STGCN (Cheb) | STGCN_ChebConv (Ks=3, Kt=3) | sym | GLU | 1.504175 | 3.031081 | 2.420486% |
| STGCN (1st) | STGCN_GCNConv (Kt=3) | sym | GLU | 1.472308 | 2.987471 | 2.369206% |
| Model (paper) | Model (code) | Laplacian matrix type | Gated activation function | MAE | RMSE | WMAPE |
|---|---|---|---|---|---|---|
| STGCN (Cheb) | STGCN_ChebConv (Ks=3, Kt=3) | sym | GLU | 1.919455 | 3.964940 | 3.088833% |
| STGCN (1st) | STGCN_GCNConv (Kt=3) | sym | GLU | 1.910708 | 3.948517 | 3.074757% |
| Model (paper) | Model (code) | Laplacian matrix type | Gated activation function | MAE | RMSE | WMAPE |
|---|---|---|---|---|---|---|
| STGCN (Cheb) | STGCN_ChebConv (Ks=3, Kt=3) | sym | GLU | 2.308847 | 4.690512 | 3.715672% |
| STGCN (1st) | STGCN_GCNConv (Kt=3) | sym | GLU | 2.306092 | 4.701984 | 3.711238% |