ST-ResNet

ST-ResNet (Deep Spatio-Temporal Residual Networks for Citywide Crowd Flows Prediction) implementaion with Pytorch, improved preprocessing speed over the original code.

Acknowledgement

Most of the code in the model definition is based on https://github.com/KL4805/STResNet-PyTorch, which is another ST-ResNet Pytorch implimentaion.

Data

Copy TaxiBJ and/or BikeNYC dataset under datasets dir.

Note TaxiBJ dataset is currently unavailable due to circumstances at the data provider.

See: TaxiBJ21: An open crowd flow dataset based on Beijing taxi GPS trajectories

Usage

python run.py [-h] [-s SEED] FILE

positional arguments:
  FILE                  path to config file

options:
  -h, --help            show this help message and exit
  -s SEED, --seed SEED  seed for initializing training

Example

python run.py examples/TaxiBJ/L4-C3-P1-T1/config.ini

Config

The following is a setting for TaxiBJ dataset and L4-C3-P1-T1, means four residual blocks, three closeness time steps, one period time step and one trend time step.

[dataset]
data_files = ["./datasets/TaxiBJ/BJ13_M32x32_T30_InOut.h5", "./datasets/TaxiBJ/BJ14_M32x32_T30_InOut.h5", "./datasets/TaxiBJ/BJ15_M32x32_T30_InOut.h5", "./datasets/TaxiBJ/BJ16_M32x32_T30_InOut.h5"]
holiday_file = ./datasets/TaxiBJ/BJ_Holiday.txt
meteorol_file = ./datasets/TaxiBJ/BJ_Meteorology.h5
T = 48			; time steps of the day. T=48 means 24 * 60 / 48 = 30 min = one time step
len_closeness = 3	; number of time steps used as closeness
len_period = 1		; number of time steps used as period
len_trend = 1		; number of time steps used as trend
period_interval = 1	; 1 specifies 1 day interval: 1 * T = 1 * 48 * 30 min = 24 hr = 1 day
trend_interval = 7	; 7 specifies 1 week interval: 7 * T = 7 * 48 * 30 min = 7 days = 1 wk
len_test = 1344		; number of test data
use_meta = true		; whether to use day of the week and weekend information
use_holiday = true	; use holiday information
use_meteorol = true	; use weather information

[model]
nb_flow = 2		; number of channels: 2 means the number of in/out-flows
map_height = 32		; grid height
map_width = 32		; grid width
nb_residual_unit = 4	; number of residual blocks

[learning]
epochs = 100
batch_size = 32
learning_rate = 0.0002

Reference

J. Zhang, Y. Zheng and D. Qi, "Deep spatio-temporal residual networks for citywide crowd flows prediction", AAAI, pp. 1655-1661, 2017.