safarzadeh-reza / STGCN-1

A PyTorch implementation of the paper https://arxiv.org/abs/1709.04875

STGCN

A PyTorch implementation of the paper https://arxiv.org/abs/1709.04875

Spatio-Temporal Graph Convolutional Networks (STGCN), author's code repo: https://github.com/VeritasYin/STGCN_IJCAI-18.

Requirements

PyTorch >= 1.1.0

DGL >= 0.4.3.post2

Other dependencies: Please refer to the code

Single-Step Ahead Forecasting

Run with following to train STGCN on the METR-LA dataset:

python train_stgcn.py

Run with following to train STGCN_WAVE on the METR-LA dataset:

python train_stgcn_wave.py

Implementation details:

As for layers:

• TemporalConvLayer_Residual is same as author's implementation
• SpatialConvLayer is based on DGL's GCN
• OutputLayer is same as author's implementation

As for network:

• STGCN has the same default structure of author's implementation with TSTN+TSTN+OutputLayer, i.e. Two ST-blocks with the output layer. The main difference is thatSpatialConvLayer has different in_channel and out_channel, i.e. c_in != c_out.
• STGCN_WAVE is an improved variation of STGCN where the main differences are:
  • Extra TemporalConvLayer and LayerNorm at the beginning, the default structure is TN+TSTN+TSTN+OutputLayer
  • TemporalConvLayer with increasing dilation like in TCN, i.e. 2, 4, 8, 16, etc
  • SpatialConvLayer does not change the channel size like in author's implementation, i.e. c_in=c_out

Parameter settings:

METR_LA: 34272 samples, 207 nodes, 5 mins sampling rate

PEMS-BAY: 52116 samples, 325 nodes, 5 mins sampling rate

STGCN paper default parameters: window = 12 (60 mins), horizon = 3, 6, 9 (15, 30, 45 mins), batch_size = 50, epoch = 50, lr = 0.001 with a decay rate of 0.7 after every 5 epochs, opt = RMSProp, temporal convolution kernel size = 3, graph convolution kernel size = 3, channels for the ST-Conv block are 64, 16, 64, dropout = 0.0

Our default parameters for STGCN : window = 12 (60 mins), horizon = 3, 6, 9 (15, 30, 45 mins), batch_size = 64, epoch = 50, lr = 0.001 with a decay rate of 0.7 after every 5 epochs, opt = Adam, Kt = 3, ST-Conv block with the channels = [64, 16, 64], dropout = 0.3

Results:

STGCN results on our default parameters (horizon=3) are:

+------------------------------------------------------------+
|                          Input                             |
+------------------------------------------------------------+
|  1D Conv, in_channel = 1, out_channel = 64, dilation = 1  |
+------------------------------------------------------------+
|       Graph Conv, in_channel = 64, out_channel = 16        |
+------------------------------------------------------------+
|  1D Conv, in_channel = 16, out_channel = 64, dilation = 1  |
+------------------------------------------------------------+
|               Layer Normalization + Dropout                |
+------------------------------------------------------------+
|  1D Conv, in_channel = 64, out_channel = 64, dilation = 1  |
+------------------------------------------------------------+
|       Graph Conv, in_channel = 64, out_channel = 16        |
+------------------------------------------------------------+
| 1D Conv, in_channel = 16, out_channel = 64, dilation = 1 |
+------------------------------------------------------------+
|               Layer Normalization + Dropout                |
+------------------------------------------------------------+
|                      OutputLayer                           |
+------------------------------------------------------------+

METR-LA: ~ MAE: 4.70 | RMSE: 9.09

PEMS-BAY: Upcoming...

To-Do:

• Masked metrics calculation