This repo is for a submission under review

Flexible and Efficient Inference with Particles for the Variational Gaussian Approximation

Repo containing experiments for the ICML submission : "Flexible and Efficient Inference with Particles for the Variational Gaussian Approximation"

Installation

First of all install julia with version at least 1.5. Download this repository (there should be a gpf_icml.zip file) and unzip it somewhere. With a terminal go to the repo and run julia. Then make the following calls:

using Pkg
Pkg.activate(".")
Pkg.develop(;path="./AdvancedVI")
Pkg.instantiate()

To download the relevant datasets from the experiments section, you can simply call

include("scripts/download_and_convert.jl")

You are now all set!

Running the scripts

Once again open a Julia session and run on of the scripts present in scripts by calling include("scripts/{ name of file}"). You can set the desired parameters (they are commented) to try different setting.

Reproducing the plots

This is a bit more tricky, you will have to go to the analysis folder and play with the different parameters in place given the simulations you have run.

Exploring the code

The source code for each algorithm was directly included in an existing package AdvancedVI.jl from TuringLang of which a branch is locally copied in this repo. Each algorithm is contained in a file in AdvancedVI/src/, the relevant function to look at is optimize! and eventually grad!