This project focuses on enhancing diffusion models for time series data, specifically improving the Conditional Score-based Diffusion Model (CSDI). Our goal was to test the potential of improving performance or speed by applying methods from other papers to CSDI and to deepen our understanding of diffusion models.
- Exploration of noise variation (Simplex noise vs. Gaussian noise) in diffusion models.
- Implementation of an ODE solver for handling long sampling distances, a known limitation in DDPM.
- Comparative analysis of DDPM and DDIM methodologies for efficiency and performance.
- Development and analysis of enhanced models using a healthcare dataset from the PhysioNet Challenge 2012.
- Experimental code for model development and analysis.
- Data preprocessing and analysis scripts.
- Documentation of methodologies and results.
- We plan to extend our research to generative models for time series data and explore reinforcement learning using data augmented by these generative models.
This is the github repository for the NeurIPS 2021 paper "CSDI: Conditional Score-based Diffusion Models for Probabilistic Time Series Imputation".
Please install the packages in requirements.txt
python download.py physiopython exe_physio.py --testmissingratio [missing ratio] --nsample [number of samples]python exe_physio.py --modelfolder pretrained --testmissingratio [missing ratio] --nsample [number of samples]python exe_pm25.py --nsample [number of samples]'visualize_examples.ipynb' is a notebook for visualizing results.
A part of the codes is based on BRITS and DiffWave
If you use this code for your research, please cite our paper:
@inproceedings{tashiro2021csdi,
title={CSDI: Conditional Score-based Diffusion Models for Probabilistic Time Series Imputation},
author={Tashiro, Yusuke and Song, Jiaming and Song, Yang and Ermon, Stefano},
booktitle={Advances in Neural Information Processing Systems},
year={2021}
}