This is the SSIM model for SSIM—A Deep Learning Approach for Recovering Missing Time Series Sensor Data
Considering the dataset we are using in the paper is not public available, we use a different open dataset for demo.
The original PM2.5 data can be download from: PM2.5
The Pytorch implementation has not been fully tested. Bugs may be fixed later.
Update 2021
Datasets from the same WQ monitoring stations can be found on Kaggle as Real Time Water Quality Data. The measurements may be different from the paper because of sensor replacement.
Please check our improved SSIM model. The paper "A Dual-Head Attention Model for Time Series Data Imputation" has been submitted to "computers and electronics in agriculture". The code can be found in repo DualHeadSSIM.
Code structure:
/checkpoints ------- store trained model
/data ------- data set
/model ------- SSIM model: encoder, decoder, attention
/utils
/prepare_PM2.5 ------------ prepare train/test for PM2.5 data. 2010-2013 for train, 2014 for test
/VLSM --------------- VLSM algorithm to generate variable length samples (with 0 pad)
Three branches:
- master: PyTorch version
- newest: PyTorch version (same model as master, training functions slightly changed for other papers )
- MXnet: MXnet version
Link the model's architecture with the equations in the paper
