DeepST-Rebuild

Introduction

This repo is a tensorflow rebuilding from the work of

Junbo Zhang, Yu Zheng, Dekang Qi. Deep Spatio-Temporal Residual Networks for Citywide Crowd Flows Prediction. In AAAI 2017. https://github.com/lucktroy/DeepST

The purpose of this modified script is for interview of JD and practice of building deep neural network.

All rights belong to the researchers and institutes listed in above paper.

Details of this model please refer to above paper and repo.

Modified subroutine

There are two subroutine modified from original work, exptBikeNYC.py and /deepst/models/STResNet1.py, and one added deepst/utils/random_mini_batches.py.

In the original work, STResNet.py and exptBikeNYC.py are developed based on Keras 1.2. Here I rebuild them using python3 and tensorflow 1.8. random_mini_batches.py is added for model traning purpose.

I didn't change other subroutines and modules, since they are focusing on the input and preprocessing of data.

Hyperparameters and settings

All the hyperparameters are same as original work in the repo above.

Main packages including tensorflow, numpy, pandas.

Run

In command line:

export DATAPATH=[path_to_your_data]
python exptBikeNYC.py

Result

First I need to mention is long running time of the program for lacking of computational power. For 500 epoches, it took 10.5 hours to finish. RMSE is used for evaluation of the model, and result as below:

Train Accuracy: 0.011191

Test Accuracy: 0.042504

Cost of every 5 epoches is recored in output.txt.

It looks like there is a potential overfitting of the model. One possible reason is the limited number of data points. Training set has 3480 data points and Test has 240. Though shuffle and random mini batch are used for training, the dataset is still not very large. The other potential reason is the hyperparameters setting of the model. Like number of residual units, learning rate, etc. I will dig more about the model and update.