Optimisation of Overparametrized Sum-Product Networks

This repository contains everything you need to reproduce the paper on "Overparameterization in Sum-Product Networks".

To run the experiments,

1. Install a running version of Julia
2. Add the SumProductNetworks package using pkg> add SumProductNetworks
3. Add the following additionally packages: Statistics, Plots, PGFPlots
4. Run the script using julia experiment.jl and wait for a while. Sorry for the slow code. :D

UPDATE

This repository now also contains an updated version of the experiments in the paper. The updated implementation uses a bijection to optimise the parameters in an unconstrainted space, reducing artifacts in the optimisation and providing more stable results. To run the updated script, make sure to use the provided Project.toml script to instantiate the julia environment. This is done by starting julia (version 1.3 or higher) inside the repository directory using julia --project=. and calling ] instantiate within the REPL.

To run the experiments, run the tpm2019.jl script.

Updated plots:

Dependencies (old script)

The following version for each package has been used.

Julia v1.1.0
SumProductNetworks v0.1.2
Plots v0.24.0
PGFPlots v3.0.3

New version may work too.

Citation

Martin Trapp, Robert Peharz, and Franz Pernkopf. "Optimisation of Overparametrized Sum-Product Networks", 3rd Workshop of Tractable Probabilistic Modeling at the International Conference on Machine Learning, 2019.