Neural Posterior Estimation for Stochastic Epidemic Modeling

Code for calibrating stochastic infectious disease models to data through simulation-based inference and deep learning.

How to Use

Dependencies are managed through Conda and Poetry. To create a Conda environment with Pytorch, run conda env create --name envname --file=environments.yml. You can install all other necessary Python packages with poetry install.

This repository uses Lightning to improve code readability and modularity and Hydra to manage configurations for deep learning experiments. For example, python -m run.py simulator=si-model model=gdn simulator.d_model=16,32,64 trains a Gaussian Density Network on data simulated from a (homogeneous) Susceptible-Infected model, sweeping over three different network widths.

Implemented Simulators

Simulation Experiments

• Normal/Normal conjugate model (for testing purposes)
• Bayesian Linear Regression model (for testing purposes)
• Susceptible-Infected (SI) transmission model (homogeneous or heterogeneous transmission rates, complete or partial observation of cases)

Empirical Models

• SI model for carbapenem-resistant Klebsiella pneumonia (CRKP) transmission; requires confidential data

Implemented Posterior Estimators

• Gaussian Density Network
• Normalizing Flow (RealNVP)
• Approximate Bayesian Computation