This is a pytorch implementation of VLDB 25 paper [Scalable Pre-Training of Compact Urban Spatio-Temporal Predictive Models on Large-Scale Multi-Domain Data]

Requirements

CompactST has been tested using Python 3.10 and CUDA Version: 12.1

To have consistent libraries and their versions, you can install the needed dependencies for this project by running the following command:

pip install -r requirements.txt

Data

The datasets used in our paper will be publicly released soon.

Run the Model

1. Pre-training

python pretrain.py

2. Fine-tuning

python few_shot.py --dataset {dataset} --unseen {dataset}

• dataset: the dataset name.