/MSSTN

MSSTN: Multi-Scale Spatial-Temporal Network for Air Pollution Prediction

Primary LanguagePythonApache License 2.0Apache-2.0

MSSTN

This repo is the code for our paper MSSTN: Multi-Scale Spatial-Temporal Network for Air Pollution Prediction. We provide pre-processed data and trained models that can reproduce main result listed in our paper. Please contact us at wu-zy18@mails.tsinghua.edu.cn if you have any question.

Requirement

We have tested our code under centos7, python3, and tensorflow 1.8.0. Similiar environment and later versions may also work but we didn't test that.

Data

We provide pre-processed data on Baidu NetDisk (Secure Code: 057p). Download data and replace the '/data/' folder before use.

Usage

Train

Use following command to train models from scratch:

python3 main.py train
Test

Use following command to load trained models and show result on test set:

python3 main.py test [InferenceModel]

where InferenceModel can be found below in config part.

Config

You may modify config.yaml to tune the training process by yourself. For example, item 'Target_City' decide which city to optimize/test when 'city_number' is set to 1, and 'InferenceModel' claim the trained model to load. More specifically, this table shows the relationship between them:

City Index InferenceModel
Beijing 0 MSSTN20190802_225240
Shijiazhuang 1 MSSTN20190803_085115
Taiyuan 2 MSSTN20190803_090722
Huhot 3 MSSTN20190803_092719
Dalian 4 MSSTN20190803_094349