GM: Graphical Models for Torch/LuaJIT

This package provides standard functions to create (arbitrary) undirected graphical models, using adjacency matrices.

A graph is described by an adjacency matrix, node potentials and edge potentials.

Inference, Decoding, Sampling, Parameter estimation

Four common tasks are implemented:

• Decoding: finding the joint configuration of the variables with the highest joint probability;

• Inference: computing the normalization constant Z (partition function), as well as the probabilities of each variable taking each possible state (the marginal probabilities);

• Sampling: given a model, sampling generates likely configurations for each node of the graph;

• Training (or parameter estimation): the task of computing the potentials that maximize the likelihood of a set of data (MAP estimation).

Note 1: parameter estimation is implemented for CRF and MRF objectives.

Note 2: this code is heavily based on UGM, from Mark Schmidt.

Install

$ [sudo] luarocks install gm

Use

First run torch, and load gm:

$ th

> require 'gm'

Once loaded, see and run the examples:

> gm.examples.simple()
> gm.examples.trainMRF()
> gm.examples.trainCRF()