Package 'CRF'

Description

Library of Conditional Random Fields model

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

Author(s)

Ling-Yun Wu [email protected]

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.

Examples

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

Description

This data set gives a chain CRF example

Usage

data(Chain)

Format

A list containing two elements:

• crf The CRF

• answer A list of 4 elements:

  • decode The most likely configuration

  • node.bel The node belief

  • edge.bel The edge belief

  • logZ The logarithmic value of CRF normalization factor Z

Description

Generate clamped CRF by fixing the states of some nodes

Usage

clamp.crf(crf, clamped)

Arguments

Details

The function will generate a clamped CRF from a given CRF by fixing the states of some nodes. The vector clamped contains the desired state for each node while zero means the state is not fixed. The node and edge potentials are updated to the conditional potentials based on the clamped vector.

Value

The function will return a new CRF with additional components:

See Also

make.crf, sub.crf, clamp.reset

Examples

library(CRF)
data(Small)
crf <- clamp.crf(Small$crf, c(0, 0, 1, 1))

Description

Reset clamped CRF by changing the states of clamped nodes

Usage

clamp.reset(crf, clamped)

Arguments

Details

The function will reset a clamped CRF by changing the states of fixed nodes. The vector clamped contains the desired state for each node while zero means the state is not fixed. The node and edge potentials are updated to the conditional potentials based on the clamped vector.

Value

The function will return the same clamped CRF.

See Also

make.crf, clamp.crf

Examples

library(CRF)
data(Small)
crf <- clamp.crf(Small$crf, c(0, 0, 1, 1))
clamp.reset(crf, c(0,0,2,2))

Description

This data set gives a clique CRF example

Usage

data(Clique)

Format

A list containing two elements:

• crf The CRF

• answer A list of 4 elements:

  • decode The most likely configuration

  • node.bel The node belief

  • edge.bel The edge belief

  • logZ The logarithmic value of CRF normalization factor Z

Description

Calculate the negative log likelihood of CRF model

Usage

crf.nll(
  par,
  crf,
  instances,
  node.fea = NULL,
  edge.fea = NULL,
  node.ext = NULL,
  edge.ext = NULL,
  infer.method = infer.chain,
  ...
)

Arguments

Details

This function calculates the negative log likelihood of CRF model as well as the gradient. This function is intended to be called by optimization algorithm in training process.

In the training data matrix instances, each row is an instance and each column corresponds a node in CRF. The variables node.fea, edge.fea, node.ext, edge.ext are lists of length equal to the number of instances, and their elements are defined as in crf.update respectively.

Value

This function will return the value of CRF negative log-likelihood.

See Also

crf.update, train.crf

Description

Update node and edge potentials of CRF model

Usage

crf.update(
  crf,
  node.fea = NULL,
  edge.fea = NULL,
  node.ext = NULL,
  edge.ext = NULL
)

Arguments

Details

This function updates node.pot and edge.pot of CRF model by using the current values of parameters and features.

There are two ways to model the relationship between parameters and features. The first one exploits the special structure of features to reduce the memory usage. However it may not suitable for all circumstances. The other one is more straighforward by explicitly specifying the coefficients of each parameter to calculate the potentials, and may use much more memory. Two approaches can be used together.

The first way uses the objects node.par and edge.par to define the structure of features and provides the feature information in variables node.fea and edge.fea. The second way directly provides the feature information in variables node.ext and edge.ext without any prior assumption on feature structure. node.ext is a list and each element has the same structure as node.pot. edge.ext is a list and each element has the same structure as edge.pot.

In detail, the node potential is updated as follows:

$node.pot[n,i] = exp( \sum_{f} par[node.par[n,i,f]] * node.fea[f,n] + \sum_{k} par[k] * node.ext[[k]][n,i] )$

and the edge potential is updated as follows:

$edge.pot[[e]][i,j] = exp( \sum_{f} par[edge.par[[e]][i,j,f]] * edge.fea[f,e] + \sum_{k} par[k] * edge.ext[[k]][[e]][i,j] )$

Value

This function will directly modify the CRF and return the same CRF.

See Also

crf.nll, train.crf

Description

Computing the most likely configuration for CRF

Usage

decode.block(
  crf,
  blocks,
  decode.method = decode.tree,
  restart = 0,
  start = apply(crf$node.pot, 1, which.max),
  ...
)

Arguments

Details

Approximate decoding with the block iterated conditional modes algorithm

Value

This function will return the most likely configuration, which is a vector of length crf$n.nodes.

Examples

library(CRF)
data(Small)
d <- decode.block(Small$crf, list(c(1,3), c(2,4)))

Description

Computing the most likely configuration for CRF

Usage

decode.chain(crf)

Arguments

Details

Exact decoding for chain-structured graphs with the Viterbi algorithm.

Value

This function will return the most likely configuration, which is a vector of length crf$n.nodes.

Examples

library(CRF)
data(Small)
d <- decode.chain(Small$crf)

Description

Computing the most likely configuration for CRF

Usage

decode.conditional(crf, clamped, decode.method, ...)

Arguments

Details

Conditional decoding (takes another decoding method as input)

Value

This function will return the most likely configuration, which is a vector of length crf$n.nodes.

Examples

library(CRF)
data(Small)
d <- decode.conditional(Small$crf, c(0,1,0,0), decode.exact)

Description

Computing the most likely configuration for CRF

Usage

decode.cutset(
  crf,
  cutset,
  engine = "default",
  start = apply(crf$node.pot, 1, which.max)
)

Arguments

Details

Exact decoding for graphs with a small cutset using cutset conditioning

Value

This function will return the most likely configuration, which is a vector of length crf$n.nodes.

Examples

library(CRF)
data(Small)
d <- decode.cutset(Small$crf, c(2))

Description

Computing the most likely configuration for CRF

Usage

decode.exact(crf)

Arguments

Details

Exact decoding for small graphs with brute-force search

Value

This function will return the most likely configuration, which is a vector of length crf$n.nodes.

Examples

library(CRF)
data(Small)
d <- decode.exact(Small$crf)

Description

Computing the most likely configuration for CRF

Usage

decode.greedy(crf, restart = 0, start = apply(crf$node.pot, 1, which.max))

Arguments

Details

Approximate decoding with greedy algorithm

Value

This function will return the most likely configuration, which is a vector of length crf$n.nodes.

Examples

library(CRF)
data(Small)
d <- decode.greedy(Small$crf)

Description

Computing the most likely configuration for CRF

Usage

decode.icm(crf, restart = 0, start = apply(crf$node.pot, 1, which.max))

Arguments

Details

Approximate decoding with the iterated conditional modes algorithm

Value

This function will return the most likely configuration, which is a vector of length crf$n.nodes.

Examples

library(CRF)
data(Small)
d <- decode.icm(Small$crf)

Description

Computing the most likely configuration for CRF

Usage

decode.ilp(crf, lp.rounding = FALSE)

Arguments

Details

Exact decoding with an integer linear programming formulation and approximate using LP relaxation

Value

This function will return the most likely configuration, which is a vector of length crf$n.nodes.

Examples

## Not run: 
library(CRF)
data(Small)
d <- decode.ilp(Small$crf)

## End(Not run)

Description

Computing the most likely configuration for CRF

Usage

decode.junction(crf)

Arguments

Details

Exact decoding for low-treewidth graphs using junction trees

Value

This function will return the most likely configuration, which is a vector of length crf$n.nodes.

Examples

library(CRF)
data(Small)
d <- decode.junction(Small$crf)

Description

Computing the most likely configuration for CRF

Usage

decode.lbp(crf, max.iter = 10000, cutoff = 1e-04, verbose = 0)

Arguments

Details

Approximate decoding using max-product loopy belief propagation

Value

This function will return the most likely configuration, which is a vector of length crf$n.nodes.

Examples

library(CRF)
data(Small)
d <- decode.lbp(Small$crf)

Description

Computing the most likely configuration for CRF

Usage

decode.marginal(crf, infer.method, ...)

Arguments

Details

Approximate decoding using inference (takes an inference method as input)

Value

This function will return the most likely configuration, which is a vector of length crf$n.nodes.

Examples

library(CRF)
data(Small)
d <- decode.marginal(Small$crf, infer.exact)

Description

Computing the most likely configuration for CRF

Usage

decode.rbp(crf, max.iter = 10000, cutoff = 1e-04, verbose = 0)

Arguments

Details

Approximate decoding using max-product residual belief propagation

Value

This function will return the most likely configuration, which is a vector of length crf$n.nodes.

Examples

library(CRF)
data(Small)
d <- decode.rbp(Small$crf)

Description

Computing the most likely configuration for CRF

Usage

decode.sample(crf, sample.method, ...)

Arguments

Details

Approximate decoding using sampling (takes a sampling method as input)

Value

This function will return the most likely configuration, which is a vector of length crf$n.nodes.

Examples

library(CRF)
data(Small)
d <- decode.sample(Small$crf, sample.exact, 10000)

Description

Computing the most likely configuration for CRF

Usage

decode.trbp(crf, max.iter = 10000, cutoff = 1e-04, verbose = 0)

Arguments

Details

Approximate decoding using max-product tree-reweighted belief propagtion

Value

This function will return the most likely configuration, which is a vector of length crf$n.nodes.

Examples

library(CRF)
data(Small)
d <- decode.trbp(Small$crf)

Description

Computing the most likely configuration for CRF

Usage

decode.tree(crf)

Arguments

Details

Exact decoding for tree- and forest-structured graphs with max-product belief propagation

Value

This function will return the most likely configuration, which is a vector of length crf$n.nodes.

Examples

library(CRF)
data(Small)
d <- decode.tree(Small$crf)

Description

Duplicate an existing CRF

Usage

duplicate.crf(crf)

Arguments

Details

This function will duplicate an existing CRF. Since CRF is implemented as an environment, normal assignment will only copy the pointer instead of the real data. This function will generate a new CRF and really copy all data.

Value

The function will return a new CRF with copied data

See Also

make.crf

Description

Calculate the logarithmic potential of a CRF with given configuration

Usage

get.logPotential(crf, configuration)

Arguments

Details

The function will calculate the logarithmic potential of a CRF with given configuration, i.e., the assigned states of nodes in the CRF.

Value

The function will return the log-potential of CRF with given configuration

See Also

get.potential

Description

Calculate the potential of a CRF with given configuration

Usage

get.potential(crf, configuration)

Arguments

Details

The function will calculate the potential of a CRF with given configuration, i.e., the assigned states of nodes in the CRF.

Value

The function will return the potential of CRF with given configuration

See Also

get.logPotential

Description

Computing the partition function and marginal probabilities

Usage

infer.chain(crf)

Arguments

Details

Exact inference for chain-structured graphs with the forward-backward algorithm

Value

This function will return a list with components:

Examples

library(CRF)
data(Small)
i <- infer.chain(Small$crf)

Description

Computing the partition function and marginal probabilities

Usage

infer.conditional(crf, clamped, infer.method, ...)

Arguments

Details

Conditional inference (takes another inference method as input)

Value

This function will return a list with components:

Examples

library(CRF)
data(Small)
i <- infer.conditional(Small$crf, c(0,1,0,0), infer.exact)

Description

Computing the partition function and marginal probabilities

Usage

infer.cutset(crf, cutset, engine = "default")

Arguments

Details

Exact inference for graphs with a small cutset using cutset conditioning

Value

This function will return a list with components:

Examples

library(CRF)
data(Small)
i <- infer.cutset(Small$crf, c(2))

Description

Computing the partition function and marginal probabilities

Usage

infer.exact(crf)

Arguments

Details

Exact inference for small graphs with brute-force counting

Value

This function will return a list with components:

Examples

library(CRF)
data(Small)
i <- infer.exact(Small$crf)

Description

Computing the partition function and marginal probabilities

Usage

infer.junction(crf)

Arguments

Details

Exact decoding for low-treewidth graphs using junction trees

Value

This function will return a list with components:

Examples

library(CRF)
data(Small)
i <- infer.junction(Small$crf)

Description

Computing the partition function and marginal probabilities

Usage

infer.lbp(crf, max.iter = 10000, cutoff = 1e-04, verbose = 0, maximize = FALSE)

Arguments

Details

Approximate inference using sum-product loopy belief propagation

Value

This function will return a list with components:

Examples

library(CRF)
data(Small)
i <- infer.lbp(Small$crf)

Description

Computing the partition function and marginal probabilities

Usage

infer.rbp(crf, max.iter = 10000, cutoff = 1e-04, verbose = 0, maximize = FALSE)

Arguments

Details

Approximate inference using sum-product residual belief propagation

Value

This function will return a list with components:

Examples

library(CRF)
data(Small)
i <- infer.rbp(Small$crf)

Description

Computing the partition function and marginal probabilities

Usage

infer.sample(crf, sample.method, ...)

Arguments

Details

Approximate inference using sampling (takes a sampling method as input)

Value

This function will return a list with components:

Examples

library(CRF)
data(Small)
i <- infer.sample(Small$crf, sample.exact, 10000)

Description

Computing the partition function and marginal probabilities

Usage

infer.trbp(
  crf,
  max.iter = 10000,
  cutoff = 1e-04,
  verbose = 0,
  maximize = FALSE
)

Arguments

Details

Approximate inference using sum-product tree-reweighted belief propagation

Value

This function will return a list with components:

Examples

library(CRF)
data(Small)
i <- infer.trbp(Small$crf)

Description

Computing the partition function and marginal probabilities

Usage

infer.tree(crf)

Arguments

Details

Exact inference for tree- and forest-structured graphs with sum-product belief propagation

Value

This function will return a list with components:

Examples

library(CRF)
data(Small)
i <- infer.tree(Small$crf)

Description

This data set gives a loop CRF example

Usage

data(Loop)

Format

A list containing two elements:

• crf The CRF

• answer A list of 4 elements:

  • decode The most likely configuration

  • node.bel The node belief

  • edge.bel The edge belief

  • logZ The logarithmic value of CRF normalization factor Z

Description

Generate CRF from the adjacent matrix

Usage

make.crf(adj.matrix = NULL, n.states = 2, n.nodes = 2)

Arguments

Details

The function will generate an empty CRF from a given adjacent matrix. If the length of nstates is less than n.nodes, it will be used repeatly. All node and edge potentials are initilized as 1.

Since the CRF data are often very huge, CRF is implemented as an environment. The assignment of environments will only copy the addresses instead of real data, therefore the variables using normal assignment will refer to the exactly same CRF. For complete duplication of the data, please use duplicate.crf.

Value

The function will return a new CRF, which is an environment with components:

See Also

duplicate.crf, clamp.crf, sub.crf

Examples

library(CRF)

nNodes <- 4
nStates <- 2

adj <- matrix(0, nrow=nNodes, ncol=nNodes)
for (i in 1:(nNodes-1))
{
	adj[i,i+1] <- 1
	adj[i+1,i] <- 1
}

crf <- make.crf(adj, nStates)

crf$node.pot[1,] <- c(1, 3)
crf$node.pot[2,] <- c(9, 1)
crf$node.pot[3,] <- c(1, 3)
crf$node.pot[4,] <- c(9, 1)

for (i in 1:crf$n.edges)
{
   crf$edge.pot[[i]][1,] <- c(2, 1)
   crf$edge.pot[[i]][2,] <- c(1, 2)
}

Description

Make the data structure of CRF features

Usage

make.features(crf, n.nf = 1, n.ef = 1)

Arguments

Details

This function makes the data structure of features need for modeling and training CRF.

The parameters n.nf and n.ef specify the number of node and edge features, respectively.

The objects node.par and edge.par define the corresponding parameters used with each feature. node.par is a 3-dimensional arrays, and element node.par[n,i,f] is the index of parameter associated with the corresponding node potential node.pot[n,i] and node feature f. edge.par is a list of 3-dimensional arrays, and element edge.par[[e]][i,j,f] is the index of parameter associated with the corresponding edge potential edge.pot[[e]][i,j] and edge feature f. The value 0 is used to indicate the corresponding node or edge potential does not depend on that feature.

For detail of calculation of node and edge potentials from features and parameters, please see crf.update.

Value

This function will directly modify the CRF and return the same CRF.

See Also

crf.update, make.par, make.crf

Description

Make the data structure of CRF parameters

Usage

make.par(crf, n.par = 1)

Arguments

Details

This function makes the data structure of parameters need for modeling and training CRF. The parameters are stored in par, which is a numeric vector of length n.par.

Value

This function will directly modify the CRF and return the same CRF.

See Also

crf.update, make.features, make.crf

Description

Calculate the negative log-likelihood of MRF model

Usage

mrf.nll(par, crf, instances, infer.method = infer.chain, ...)

Arguments

Details

This function calculates the negative log-likelihood of MRF model as well as the gradient. This function is intended to be called by optimization algorithm in training process. Before calling this function, the MRF sufficient statistics must be calculated and stored in object par.stat of CRF.

In the training data matrix instances, each row is an instance and each column corresponds a node in CRF.

Value

This function will return the value of MRF negative log-likilihood.

See Also

mrf.stat, mrf.update, train.mrf

Description

Calculate the sufficient statistics of MRF model

Usage

mrf.stat(crf, instances)

Arguments

Details

This function calculates the sufficient statistics of MRF model. This function much be called before the first calling to mrf.nll. In the training data matrix instances, each row is an instance and each column corresponds a node in CRF.

Value

This function will return the value of MRF sufficient statistics.

See Also

mrf.nll, train.mrf

Description

Update node and edge potentials of MRF model

Usage

mrf.update(crf)

Arguments

Details

The function updates node.pot and edge.pot of MRF model.

Value

This function will directly modify the CRF and return the same CRF.

See Also

mrf.nll, train.mrf

Description

This data set gives an example of rain data used to train CRF and MRF models

Usage

data(Rain)

Format

A list containing two elements:

• rain A matrix of 28 columns containing raining data (1: rain, 2: sunny). Each row is an instance of 28 days for one month.

• months A vector containing the months of each instance.

References

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

Description

Generating samples from the distribution

Usage

sample.chain(crf, size)

Arguments

Details

Exact sampling for chain-structured graphs with the forward-filter backward-sample algorithm

Value

This function will return a matrix with size rows and crf$n.nodes columns, in which each row is a sampled configuration.

Examples

library(CRF)
data(Small)
s <- sample.chain(Small$crf, 100)

Description

Generating samples from the distribution

Usage

sample.conditional(crf, size, clamped, sample.method, ...)

Arguments

Details

Conditional sampling (takes another sampling method as input)

Value

This function will return a matrix with size rows and crf$n.nodes columns, in which each row is a sampled configuration.

Examples

library(CRF)
data(Small)
s <- sample.conditional(Small$crf, 100, c(0,1,0,0), sample.exact)

Description

Generating samples from the distribution

Usage

sample.cutset(crf, size, cutset, engine = "default")

Arguments

Details

Exact sampling for graphs with a small cutset using cutset conditioning

Value

This function will return a matrix with size rows and crf$n.nodes columns, in which each row is a sampled configuration.

Examples

library(CRF)
data(Small)
s <- sample.cutset(Small$crf, 100, c(2))

Description

Generating samples from the distribution

Usage

sample.exact(crf, size)

Arguments

Details

Exact sampling for small graphs with brute-force inverse cumulative distribution

Value

This function will return a matrix with size rows and crf$n.nodes columns, in which each row is a sampled configuration.

Examples

library(CRF)
data(Small)
s <- sample.exact(Small$crf, 100)

Description

Generating samples from the distribution

Usage

sample.gibbs(
  crf,
  size,
  burn.in = 1000,
  start = apply(crf$node.pot, 1, which.max)
)

Arguments

Details

Approximate sampling using a single-site Gibbs sampler

Value

This function will return a matrix with size rows and crf$n.nodes columns, in which each row is a sampled configuration.

Examples

library(CRF)
data(Small)
s <- sample.gibbs(Small$crf, 100)

Description

Generating samples from the distribution

Usage

sample.junction(crf, size)

Arguments

Details

Exact sampling for low-treewidth graphs using junction trees

Value

This function will return a matrix with size rows and crf$n.nodes columns, in which each row is a sampled configuration.

Examples

library(CRF)
data(Small)
s <- sample.junction(Small$crf, 100)

Description

Generating samples from the distribution

Usage

sample.tree(crf, size)

Arguments

Details

Exact sampling for tree- and forest-structured graphs with sum-product belief propagation and backward-sampling

Value

This function will return a matrix with size rows and crf$n.nodes columns, in which each row is a sampled configuration.

Examples

library(CRF)
data(Small)
s <- sample.tree(Small$crf, 100)

Description

This data set gives a small CRF example

Usage

data(Small)

Format

A list containing two elements:

• crf The CRF

• answer A list of 4 elements:

  • decode The most likely configuration

  • node.bel The node belief

  • edge.bel The edge belief

  • logZ The logarithmic value of CRF normalization factor Z

Description

Generate sub CRF by selecting some nodes

Usage

sub.crf(crf, subset)

Arguments

Details

The function will generate a new CRF from a given CRF by selecting some nodes. The vector subset contains the node ids selected to generate the new CRF. Unlike clamp.crf, the potentials of remainning nodes and edges are untouched.

Value

The function will return a new CRF with additional components:

See Also

make.crf, clamp.crf

Examples

library(CRF)
data(Small)
crf <- sub.crf(Small$crf, c(2, 3))

Description

Train the CRF model to estimate the parameters

Usage

train.crf(
  crf,
  instances,
  node.fea = NULL,
  edge.fea = NULL,
  node.ext = NULL,
  edge.ext = NULL,
  nll = crf.nll,
  infer.method = infer.chain,
  ...,
  trace = 0
)

Arguments

Details

This function train the CRF model.

In the training data matrix instances, each row is an instance and each column corresponds a node in CRF. The variables node.fea, edge.fea, node.ext, edge.ext are lists of length equal to the number of instances, and their elements are defined as in crf.update respectively.

Value

This function will directly modify the CRF and return the same CRF.

See Also

crf.update, crf.nll, make.crf

Description

Train the MRF model to estimate the parameters

Usage

train.mrf(
  crf,
  instances,
  nll = mrf.nll,
  infer.method = infer.chain,
  ...,
  trace = 0
)

Arguments

Details

This function trains the Markov Random Fields (MRF) model, which is a simple variant of CRF model.

In the training data matrix instances, each row is an instance and each column corresponds a node in CRF.

Value

This function will directly modify the CRF and return the same CRF.

See Also

mrf.update, mrf.stat, mrf.nll, make.crf

Description

This data set gives a tree CRF example

Usage

data(Tree)

Format

A list containing two elements:

• crf The CRF

• answer A list of 4 elements:

  • decode The most likely configuration

  • node.bel The node belief

  • edge.bel The edge belief

  • logZ The logarithmic value of CRF normalization factor Z