Diffusion Attention Convolutional Recurrent Neural Network for Traffic Forecasting

This is a TensorFlow implementation of Diffusion Attention Convolutional Recurrent Neural Network. This code is based on DCRNN

Requirements

• scipy>=0.19.0
• numpy>=1.12.1
• pandas>=0.19.2
• tensorflow>=1.3.0
• pyaml

Dependency can be installed using the following command:

pip install -r requirements.txt

Data Preparation

The traffic data file for Los Angeles, i.e., METR-LA.h5, is available at Google Drive, Baidu Yun or DCRNN, and should be put into the data/ folder.

# METR-LA
python -m scripts.generate_training_data --output_dir=data/METR-LA --traffic_df_filename=data/metr-la.h5

# PEMS-BAY
python -m scripts.generate_training_data --output_dir=data/PEMS-BAY --traffic_df_filename=data/pems-bay.h5

The generated train/val/test dataset will be saved at data/METR-LA/{train,val,test}.npz or data/PEMS-BAY/{train,val,test}.npz.

Model Training

python dacrnn_train.py --config_filename=data/model/dcrnn_config.yaml

Each epoch takes about 7min~14min with a single GTX 1080 Ti.