Ergodic-variational-flow-code

Code that reproduces experiments in MixFlows: principled variational inference via mixed flows

Package installation

We implement a practical version of the Ergodic variational flows---Hamiltonian ergodic flow---as a julia package ErgFlow.jl, which can be installed by executing the following code:

# in Julia REPL
using Pkg
Pkg.add("https://github.com/zuhengxu/ErgFlow.jl")

Code description

Examples run and generate output using Julia v1.8.1.

• inference/ provides implementation of our competitors (NUTS/standard HMC/three normalizing flow methods), as well as utility functions for performing inference (KSD estimation/mean-field Gaussian VI method/etc.)
• examples/ provides code to replicate results and figures for each example

How to run the code

Each experiment should be run in its own folder (e.g., example/1d_cauchy, example/banana/, and etc.). Specifically,

• To generate Figure 1 (effect of irrational shift), execute the following code in bash:

cd example/uniform/
julia beta_mixture.jl

• To generate plots for all synthetic examples, check the README.md files in each of the 1d_cauchy, 1d_gaussian, 1d_mixture, banana, cross, neals_funnel, and warped_gaussian folders.
• To generate plots for all real data examples: cd into each of the heavy_reg, linear_regression, lin_reg_heavy, logistic_reg, poiss,sparse_regression, and sp_reg_big folders, and then execute the following in bash:

./run.sh

Note that users should modify the number of execution threads in run.sh based on their own computational resources (currently set to be 18).