Note: The latest version is in the latest branch, i.e. https://github.com/SYLan2019/MH-ASTIGCN/tree/latest

Multi-Head Self-Attention Based Spatial-Temporal Information Graph Convolutional Networks for Traffic Flow Forecasting

• python >= 3.5
• scipy
• tensorboard
• pytorch

Step 1: MH-ASTIGCN is implemented on those several public traffic datasets.

• PEMS04, PEMS07 and PEMS08 from STSGCN (AAAI-20). Download the data STSGCN_data.tar.gz with password: p72z and uncompress data file usingtar -zxvf data.tar.gz

Step 2: Process dataset

• on PEMS04 dataset

  python prepareData.py --config configurations/PEMS04.conf

If traffic data is available, TIG could also be generated by code:

cd ./data/
python Temporal_Graph_gen.py

The shape of input traffic data should be "(Total_Time_Steps, Node_Number)". For example, in PEMS08 dataset, it has 170 roads and 62 days data. Thus its shape is (62*288, 170).

The calculation uses CPU, which should be prepared for enough computation resources.

• on PEMS04 dataset (use our trained network parameters)

  python test.py --config configurations/PEMS04.conf   

The configuration file config.conf contains two parts: Data, Training:

• adj_filename: path of the adjacency matrix file
• graph_signal_matrix_filename: path of graph signal matrix file
• STIG_filename:path of the Spatial-Temporal Information Graph file
• TSG_filename:path of the Temporal Similarity Graph file
• num_of_vertices: number of vertices
• points_per_hour: points per hour, in our dataset is 12
• num_for_predict: points to predict, in our model is 12

• graph: select the graph structure, G or STIG, G stands for adjacency graph, STIG stands for Spatial-Temporal Information Graph
• ctx: set ctx = cpu, or set gpu-0, which means the first gpu device
• epochs: int, epochs to train
• learning_rate: float, like 0.0001
• batch_size: int
• num_of_weeks: int, how many weeks' data will be used
• num_of_days: int, how many days' data will be used
• num_of_hours: int, how many hours' data will be used
• n_heads: int, number of temporal att heads will be used
• d_k: int, the dimensions of the Q, K, and V vectors will be used
• d_model: int, d_E
• K: int, K-order chebyshev polynomials (number of spatial att heads) will be used