ayushchopra96 / covid-gnn-1mg

Spatio-Temporal Graph Convolutional Networks

About

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.

Paper

https://arxiv.org/abs/1709.04875

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. ChebyNet: 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. ChebyNet: 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 ChebyNet：

Model structure

Differents of code between mine and author's

1. Fix tons of 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
6. Add datasets METR-LA and PEMS-BAY
7. Using a different data preprocessing method

Requirements

To install requirements:

pip3 install -r requirements.txt

Experimental results

METR-LA (15/30/60 mins) (train: val: test = 70: 15: 15)

15 mins (H=3)

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%

30 mins (H=6)

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%

60 mins (H=12)

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%

PEMS-BAY (15/30/60 mins) (train: val: test = 70: 15: 15)

15 mins (H=3)

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%

30 mins (H=6)

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%

60 mins (H=12)

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%