This repository provides a reference implementation of paper: Overcoming Forgetting in Fine-Grained Urban Flow Inference via Adaptive Knowledge Replay, Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI), 2023.

The codes include the implementation of CUFAR and other FUFI methods (UrbanFM, DeepLGR, FODE, UrbanODE, UrbanPy).

We implement CUFAR and other FUFI methods with following dependencies:

• python 3.7.12
• pytorch 1.8.0
• numpy
• einops
• scikit-learn

For Linux, install the enviroment via Anaconda:

bash install_env.sh

For Windows, follow the step in the install_env.sh.

TaxiBJ datasets can be obtained in baseline UrbanFM's repo.

Before you run the code, you may need to ensure the package structure of CUFAR is as follows:

.
├── buffers
├── datasets
│   └── TaxiBJ
│       ├── P1
│       ├── P2
│       ├── P3
│       └── P4
├── model
├── src
└── README.md

We also provide the training approches of all protocols, they are train_single_task.py, train_finetune.py, train_continual.py and train_joint.py. You can change the backbone through model argument.

# Run single-task protocol of CUFAR
python train_single_task.py --model=CUFAR

# Run joint protocol of CUFAR
python train_joint.py --model=CUFAR

# Run fine-tune protocol of CUFAR
python train_finetune.py --model=CUFAR --initial_train

# Run continual protocol to evaluate our AKR
python train_continual.py --model=CUFAR --initial_train

If you find CUFAR useful in your research, please cite the following paper:

@inproceedings{yu2023Overcoming,
  title={Overcoming Forgetting in Fine-Grained Urban Flow Inference via Adaptive Knowledge Replay},
  author={Yu, Haoyang and Xu, Xovee and Zhong, Ting and Zhou, Fan},
  booktitle={AAAI},
  year={2023}
}