This repository contains the implementation of the Bayes-CATSI and Partial Bayes-CATSI models, designed for efficient medical time-series imputation with uncertainty quantification using variational inference. The project is built using PyTorch, ensuring modularity, scalability, and ease of integration into existing workflows
Requires Python 3.10 or later with PyTorch and related libraries. Please refer to requirements.txt for details of python packages required.
The folders /BayesCATSI and /partialBayesCATSI contain the code for Bayes-CATSI and partial Bayes-CATSI models. Place the dataset in a new folder and train the model as follows.
python main.py --input /path/to/training/data --testing /path/to/test/data
optional arguments:
--output Folder to save the results
--epochs number of epochs
--batch_size batch size for the model
--eval_batch_size evaluation batch size for the model
Data used in this project can be found at: Data
Dummy data has been uploaded in the \data folder to provide an idea of the data structure.
Computational Resources: In this project, we leverage Google Colab’s free tier, which provides a CPU-based execution environment for running our code.
@misc{kulkarni2024bayescatsivariationalbayesiandeep,
title={Bayes-CATSI: A variational Bayesian deep learning framework for medical time series data imputation},
author={Omkar Kulkarni and Rohitash Chandra},
year={2024},
eprint={2410.01847},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2410.01847},
}