Skip to content

Novel technique to fit a target distribution with a class of distributions using SVI (via NumPyro). Unlike standard SVI, our "data" is a distribution rather than a finite collection of samples.

License

Notifications You must be signed in to change notification settings

edwarddramirez/svi-dist-fit

Repository files navigation

Binder License: MIT Python Repo Size

svi-dist-fit

Summary notebooks showing how to fit a target distribution with a class of distributions using SVI (via NumPyro). Unlike standard SVI, our "data" is a distribution rather than a finite collection of samples. This is the first time fitting distributions with other distributions is performed using SVI-type optimization.

Notebooks

  1. 01_kl_simple.ipynb - Fitting a parametric model with another parametric model
  2. 02_kl_rate.ipynb - Fitting a non-parametric model with a parametric model
  3. 03_kl_poisson.ipynb - Fitting a non-parametric Poisson model with a parametric Poisson model

Installation

Run the environment.yml file by running the following command on the main repo directory:

conda env create

The installation works for conda==4.12.0. This will install all packages needed to run the code on a CPU with jupyter.

If you want to run this code with a CUDA GPU, you will need to download the appropriate jaxlib==0.4.13 version. For example, for my GPU running on CUDA==12.3, I would run:

pip install jaxlib==0.4.13+cuda12.cudnn89

The key to using this code directly would be to retain the jax and jaxlib versions.

About

Novel technique to fit a target distribution with a class of distributions using SVI (via NumPyro). Unlike standard SVI, our "data" is a distribution rather than a finite collection of samples.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published