CRF-package

CRF - Conditional Random Fields

Library of Conditional Random Fields model package

Details

CRF is R package for various computational tasks of conditional random fields as well as other probabilistic undirected graphical models of discrete data with pairwise and unary potentials. The decoding/inference/sampling tasks are implemented for general discrete undirected graphical models with pairwise potentials. The training task is less general, focusing on conditional random fields with log-linear potentials and a fixed structure. The code is written entirely in R and C++. The initial version is ported from UGM written by Mark Schmidt.

Decoding: Computing the most likely configuration

• decode.exact Exact decoding for small graphs with brute-force search
• decode.chain Exact decoding for chain-structured graphs with the Viterbi algorithm
• decode.tree Exact decoding for tree- and forest-structured graphs with max-product belief propagation
• decode.conditional Conditional decoding (takes another decoding method as input)
• decode.cutset Exact decoding for graphs with a small cutset using cutset conditioning
• decode.junction Exact decoding for low-treewidth graphs using junction trees
• decode.sample Approximate decoding using sampling (takes a sampling method as input)
• decode.marginal Approximate decoding using inference (takes an inference method as input)
• decode.lbp Approximate decoding using max-product loopy belief propagation
• decode.trbp Approximate decoding using max-product tree-reweighted belief propagtion
• decode.greedy Approximate decoding with greedy algorithm
• decode.icm Approximate decoding with the iterated conditional modes algorithm
• decode.block Approximate decoding with the block iterated conditional modes algorithm
• decode.ilp Exact decoding with an integer linear programming formulation and approximate using LP relaxation

Inference: Computing the partition function and marginal probabilities

• infer.exact Exact inference for small graphs with brute-force counting
• infer.chain Exact inference for chain-structured graphs with the forward-backward algorithm
• infer.tree Exact inference for tree- and forest-structured graphs with sum-product belief propagation
• infer.conditional Conditional inference (takes another inference method as input)
• infer.cutset Exact inference for graphs with a small cutset using cutset conditioning
• infer.junction Exact decoding for low-treewidth graphs using junction trees
• infer.sample Approximate inference using sampling (takes a sampling method as input)
• infer.lbp Approximate inference using sum-product loopy belief propagation
• infer.trbp Approximate inference using sum-product tree-reweighted belief propagation

Sampling: Generating samples from the distribution

• sample.exact Exact sampling for small graphs with brute-force inverse cumulative distribution
• sample.chain Exact sampling for chain-structured graphs with the forward-filter backward-sample algorithm
• sample.tree Exact sampling for tree- and forest-structured graphs with sum-product belief propagation and backward-sampling
• sample.conditional Conditional sampling (takes another sampling method as input)
• sample.cutset Exact sampling for graphs with a small cutset using cutset conditioning
• sample.junction Exact sampling for low-treewidth graphs using junction trees
• sample.gibbs Approximate sampling using a single-site Gibbs sampler

Training: Given data, computing the most likely estimates of the parameters

• train.crf Train CRF model
• train.mrf Train MRF model

Tools: Tools for building and manipulating CRF data

• make.crf Generate CRF from the adjacent matrix
• make.features Make the data structure of CRF features
• make.par Make the data structure of CRF parameters
• duplicate.crf Duplicate an existing CRF
• clamp.crf Generate clamped CRF by fixing the states of some nodes
• clamp.reset Reset clamped CRF by changing the states of clamped nodes
• sub.crf Generate sub CRF by selecting some nodes
• mrf.update Update node and edge potentials of MRF model
• crf.update Update node and edge potentials of CRF model

Examples

library(CRF)
data(Small)
decode.exact(Small$crf)
infer.exact(Small$crf)
sample.exact(Small$crf, 100)

References

J. Lafferty, A. McCallum, and F. Pereira. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In the proceedings of International Conference on Machine Learning (ICML), pp. 282-289, 2001.

Mark Schmidt. UGM: A Matlab toolbox for probabilistic undirected graphical models. http://www.cs.ubc.ca/~schmidtm/Software/UGM.html, 2007.

Author(s)

Ling-Yun Wu wulingyun@gmail.com