Decoupling long- and short-term patterns in spatiotemporal inference

Requirements

Anaconda is highly recommended, and the code is tested under:

• Python = 3.6
• Numpy = 1.19
• Pytorch = 1.7
• Pandas = 1.1
• tqdm = 4.49

Datasets

The datasets adopted in our work are accessible at:

• METR-LA: https://github.com/liyaguang/DCRNN.
• PeMS-Bay: https://github.com/liyaguang/DCRNN.
• NREL: https://www.nrel.gov/grid/solar-power-data.html.
• Beijing-Air: https://www.biendata.xyz/competition/kdd_2018/data. (Register is required)

Run the program

Arguments

Key arguments are listed below:

setting	default	values	help
--name	DualSTN	DualSTN	name of the experiment. It partially decides checkpoint file name.
--gpu_ids	-1	int	index of GPU. Multiple GPUs are not supported.
--file_time	N.A.	str	time when the model was trained. specify during evaluation.
--dataset_mode	metrla	metrla, nrel, bjair, pemsbay	choose what dataset is used.
--epoch	N.A.	int	which epoch to load during evaluation.

Model Training and Evaluation

Here are the command for training the model on METR-LA. The framework will automatically evaluate the best checkpoint on the validation/testing set and save all related results to a .pkl file.

python train.py --name DualSTN --dataset_mode metrla --gpu_ids 0

Alternatively you can use train.sh to train the model with different initialization seeds.

train.sh DualSTN metrla 0 2030

Assume the timestamp of running the code is 20210725T160246, the framework will store configurations and checkpoints at

DualSTN/
│   
└───checkpoints/
   │
   └───metrla/
       │   
       └──DualSTN_20210725T160246 

For other functions like continue training, please refer the code at (base_options.py)

Pretrained checkpoints

We provide the pretrained checkpoints for METR-LA, PeMS-Bay, NREL, and Beijing-Air. The checkpoints are stored at checkpoints. The checkpoints are named as DualSTN_{dataset}_{timestamp}. For example, the checkpoint for METR-LA is named as DualSTN_metrla_20210725T160246. The checkpoints are trained with the same hyperparameters as in the paper.

Acknowledge

This framework is largely based on CycleGAN. Some parts of the code are adopted from IGNNK. Thanks for their great works! :)