Spatio-Temporal Joint Convolutional Network for Traffic Forecasting

Structure

• config: configurations of STJGCN
• data: includes the PEMSD4 and PEMSD8 datasets used in our experiments
• logs: logs during training or testing
• model: saved models (we also provide pre-trained models for both datasets)
• model.py: implement of our STJGCN model
• utils.py: tools, including data processing, evaluation metrics, etc.
• tf_utils.py: tensorflow-based tools
• train.py: code of training STJGCN
• test.py: code of testing STJGCN

Requirements

Python 3.7.10, tensorflow 1.14.0, numpy 1.16.4, scipy 1.2.1, argparse and configparser

Training

To train STJGCN on the PeMSD4 or PeMSD8 dataset, run:

python train.py --config config/STJGCN_PeMSD4.conf
python train.py --config config/STJGCN_PeMSD8.conf

Evaluation

To evaluate STJGCN on the PeMSD4 or PeMSD8 dataset, run:

python test.py --config config/STJGCN_PeMSD4.conf
python test.py --config config/STJGCN_PeMSD8.conf

Results

We provide pre-trained models on both datasets, which achieve the following performance:

Dataset	MAE	RMSE	MAPE
PeMSD4	18.79	30.38	11.87%
PeMSD8	14.50	23.66	9.07%

Note that this result is different to (better than) Table 1 in the paper, because we report the average error over 10 runs in Table 1.