fiblab / deepstn Goto Github PK

Codes for AAAI 2019 DeepSTN+: Context-aware Spatial-Temporal Neural Network for Crowd Flow Prediction in Metropolis

Python 100.00%

crowd-flow-prediction traffic-flow-prediction aaai2019 spatial-temporal-forecasting

deepstn's Introduction

DeepSTN+

Keras implementation of DeepSTN+: Context-aware Spatial-Temporal Neural Network for Crowd Flow Prediction in Metropolis. Ziqian Lin^, Jie Feng^, Ziyang Lu, Yong Li, and Depeng Jin. AAAI 2019. (^ indicates equal contribution) PDF If our codes is helpful to your research, you can cite our work by:

@article{lin2019deepstn+:,
title={DeepSTN+: Context-aware Spatial Temporal Neural Network for Crowd Flow Prediction in Metropolis},
author={Lin, Ziqian and Feng, Jie and Lu, Ziyang and Li, Yong and Jin, Depeng},
booktitle={Thirty-Thrid AAAI Conference on Artificial Intelligence},
year={2019}
}

Datasets

Similar to ST-ResNet, our dataset is from the NYC Bike. Besides, we collect 9 types of PoIs for this dataset. The spatial map size of the dataset is 21x12. The dataset is in the folder /DATA/dataBikeNYC flow_data.npy ( TimeLenth x In&OutFlow x MapHeight x MapWidth = 4392 x 2 x 21 x 12 ) and poi_data.npy ( PoICategories x MapHeight x MapWidth = 9 x 21 x 12 ) for directly used.

Requirements

python 3.5
Keras 2.0
NumPy

Project Structure

File BikeNYC corresponds the Dataset BikeNYC in the Paper DeepSTN+.

/DATA
- dataBikeNYC contain dataset flow_data.npy and poi_data.npy
- lzq_read_data_time_poi.py transfer flow_data.npy and poi_data.npy to the input of the DeepSTN+ network
/DST_network baseline from ST-ResNet
- ilayer.py
- metrics.py
- STResNet.py
/DeepSTN_00/SCORE are used to save the results of DeepSTN
/DeepSTN_10/SCORE are used to save the results of DeepSTN+plus
/DeepSTN_01/SCORE are used to save the results of DeepSTN+PoI$*$time
/DeepSTN_11/SCORE are used to save the results of DeepSTN+plus+PoI$*$time
/DeepSTN_network
- DeepSTN_net.py model codes for DeepSTN+
- metrics.py contains the metric RMSE
/ComparisonBikeNYC.py you can run this file to get the results of ST-ResNet and DeepSTN in the paper.

Usage

python ComparisonBikeNYC.py

Other parameters:

Refer to ComparisonBikeNYC.py and DeepSTN_net.py

for training:
- epoch, batch_size, lr, days_test, iterate_num
- XDST, X11, X10, X01, X00, trainDST, train11, train10, train01, train00
model definition: H, W, channel, T, len_closeness, len_period, len_trend, T_closeness, T_period, T_trend, pre_F, conv_F, R_N, drop, is_plus, plus, rate, is_pt, P_N, T_F, PT_F, is_summary, kernel1, isPT_F

deepstn's People

Contributors

Stargazers

Watchers

deepstn's Issues

How to produce flow_data.npy and poi_data.npy from dataBikeNYC

Thank you very much for sharing this great model.
I am an AI student and trying to have a better understanding about data preprocessing step.
I have collected dataBikeNYC containting tripduration, starttime, stoptime, start station id, ....etc. Could I please know how to produce flow_data.npy and poi_data.npy from dataBikeNYC?
Thank you very much in advance.

多尺度融合的实现问题

您好，看了您的论文和代码，您在论文中提到multi-scale fusion mechanism at the end of DeepSTN+ to improve both the accuracy of prediction and the stability of the model.
但是在代码中，您是将ResPlus最后输出的结果进行了激活、正则化、卷积、正则化的操作，
` if is_plus:
for i in range(R_N):
cpt=Res_plus('Res_plus_'+str(i+1),conv_F,plus,rate,drop,H,W)(cpt)

else:  
    for i in range(R_N):
        cpt=Res_normal('Res_normal_'+str(i+1),conv_F,drop,H,W)(cpt)

cpt_conv2=Activation('relu')(cpt)
cpt_out2=BatchNormalization()(cpt_conv2)
cpt_conv1=Dropout(drop)(cpt_out2)
cpt_conv1=Conv2D(filters=channel,kernel_size=(1, 1),padding="same")(cpt_conv1)
cpt_out1=Activation('tanh')(cpt_conv1)`

并不是像论文架构图中将多个ResPlus Unit 的输出进行融合在输送到最后的卷积模块，而是只将ResPlus最后的输出送到了最后的卷积模块，是为什么呢？

model.load_weights()

Hi,I see you use keras==2.0 and i used keras==2.3.1 several days ago,i can run your programme.but,now i met the question(model_weights() IndexError: list index out of range).i think it is the question that i use keras==2.3.1,this version is too high?but i can not slove this question.

数据来源

您好，您位置的数据来源能不能分享一下，还有处理的方法

Dataset MobileBJ

Hi!

Could you share with the link I can access to the dataset MobileBJ used in your paper?

Thanks!