Attentive Crowd Flow Machines

This is a PyTorch implementation of Attentive Crowd Flow Machines (ACFM) in ACM Multimedia 2018 and its journal version. ACFM is a a unified neural network which can effectively learn the spatial-temporal feature representations of crowd flow with an attention mechanism.

If you use this code for your research, please cite our papers （Conference Version and Journal Version):

@inproceedings{liu2018attentive,
  title={Attentive Crowd Flow Machines},
  author={Liu, Lingbo and Zhang, Ruimao and Peng, Jiefeng and Li, Guanbin and Du, Bowen and Lin, Liang},
  booktitle={2018 ACM Multimedia Conference on Multimedia Conference},
  pages={1553--1561},
  year={2018},
  organization={ACM}
}

@article{liu2019acfm,
  title={ACFM：A Dynamic Spatial-Temporal Network for Traffic Prediction},
  author={Liu, Lingbo and Zhen, Jiajie and  Li, Guanbin and Zhan, Geng and Lin, Liang},
  year={2019}
}

Requirements

• torch==0.4.1

Preprocessing

For Crowd Flow Prediction: download TaxiBJ / BikeNYC and put them into folder data/TaxiBJ and data/BikeNYC.

For Citywide Passenger Demand Prediction (CPDP): the dataset of CPDP has been in folder data/TaxiNYC.

Model Training

# TaxiBJ
python run_taxibj.py

# BikeNYC
python run_bikenyc.py

# TaxiNYC
python run_taxinyc.py

Testing

# TaxiBJ
python test_taxibj.py

# BikeNYC
python test_bikenyc.py

# TaxiNYC
python test_taxinyc.py