Download 2 - R

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Algorithm characterizations wikipedia , lookup

Mathematical optimization wikipedia , lookup

Eigenvalues and eigenvectors wikipedia , lookup

Smith–Waterman algorithm wikipedia , lookup

Pierre-Simon Laplace wikipedia , lookup

Computational electromagnetics wikipedia , lookup

Operational transformation wikipedia , lookup

Expectation–maximization algorithm wikipedia , lookup

Laplace transform wikipedia , lookup

Multidimensional empirical mode decomposition wikipedia , lookup

Linear algebra wikipedia , lookup

Non-negative matrix factorization wikipedia , lookup

Four-vector wikipedia , lookup

Transcript
Package ‘bcs’
April 5, 2017
Type Package
Title Bayesian Compressive Sensing Using Laplace Priors
Version 1.0.0
Date 2017-03-30
Description A Bayesian method for solving the compressive sensing problem.
In particular, this package implements the algorithm 'Fast Laplace' found
in the paper 'Bayesian Compressive Sensing Using Laplace Priors' by Babacan,
Molina, Katsaggelos (2010) <DOI:10.1109/TIP.2009.2032894>.
License GPL-3 | file LICENSE
LazyData TRUE
LinkingTo Rcpp, RcppArmadillo
Imports fda (>= 2.4.4), stats (>= 3.3.2), Rcpp (>= 0.12.5), wmtsa (>=
2.0-1)
RoxygenNote 5.0.1
Suggests testthat
NeedsCompilation yes
Author Matt Goodwin [aut, cre]
Maintainer Matt Goodwin <[email protected]>
Repository CRAN
Date/Publication 2017-04-04 22:44:15 UTC
R topics documented:
BSplineBasis
FastLaplace .
FindSparse .
FourierBasis .
WaveletBasis
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Index
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2
2
4
5
6
8
1
2
FastLaplace
BSplineBasis
Finds the Transformation Matrix for the B-spline Basis
Description
Uses the package fda to find a transformation matrix where each column is a basis function from
the B-spline basis, evaluated at specified points along the rows of the matrix.
Usage
BSplineBasis(tlist, N, train = NULL)
Arguments
tlist
an array of the specific points where the basis functions are evaluated at.
N
number of basis functions in the matrix.
train
indices corresponding to which rows of the matrix to keep. Default is to keep
all rows.
Value
A P x N discrete Fourier transformation matrix where P is equal to the length of train and N is the
number of basis. If train is NULL then P equals the length of tlist.
Examples
# Points to evaluate the basis functions at
points <- c(0,1,2,3,4)
# Find first 8 B-spline basis functions evaluted at "points"
b <- BSplineBasis(points, 8)
# Find first 8 B-spline basis functions evaluated at "points" but only keep
# last two rows
b <- BSplineBasis(points, 8, train = c(4,5))
FastLaplace
Implements the Fast Laplace Algorithm
Description
Implements the fast Laplace algorithm in Rcpp. For a more user friendly implementation of this
function that makes things more convenient see FindSparse.
Usage
FastLaplace(PHI, y, sigma2, eta, roundit = 0L, verbose = 0L)
FastLaplace
3
Arguments
PHI
typically equals the product of a measurment matrix and basis representation
matrix, such as the wavelet basis. The solution vector w is assumed to be sparse
in the chosen basis.
y
CS measurements, samples from the signal or function.
sigma2
initial noise variance.
eta
threshold in determining convergence of marginal likelihood.
roundit
whether or not to round the marginal likelihood, in order to avoid machine precision error when comparing across platforms. 0 is False, 1 is True.
verbose
print which basis are added, re-estimated, or deleted. 0 is False, 1 is True.
Details
This code implements the fast Laplace algorithm from [1], which is based on [2]. The fast Laplace
algorithm is a method used to solve the compressive sensing problem, or in general, a highly underdetermined system of equations. It does this by taking the system of equations
y = Φw + n
and converting it into a minimization problem where we minimize the error with a constraint on
w (the vector we are solving for) that enforces sparsity. The fast Laplace method uses a Bayesian
framework, and in particular, uses a Laplace prior to enforce sparsity on w. See [1] for more
information.
Value
A list containing the following elements:
weights
used
sigma2
errbars
alpha
sparse weights, the non-zero values of the sparse vector w.
the positions of the sparse weights or non-zero values.
re-estimated noise variance.
one standard deviation around the sparse weights.
sparse hyperparameters (1/gamma).
References
[1] S. D. Babacan, R. Molina and A. K. Katsaggelos, "Bayesian Compressive Sensing Using
Laplace Priors," in IEEE Transactions on Image Processing, vol. 19, no. 1, pp. 53-63, Jan. 2010.
[2] S. Ji, Y. Xue, L. Carin, "Bayesian Compressive Sensing," IEEE Trans. Signal Processing, vol.
56, no. 6, June 2008.
[3] M. Tipping and A. Faul, "Fast marginal likelihood maximisation for sparse Bayesian models,"
in Proc. 9th Int. Workshop Artificial Intelligence and Statistics, C. M. Bishop and B. J. Frey, Eds.,
2003.
4
FindSparse
FindSparse
Implements the Fast Laplace Algorithm
Description
Implements the fast Laplace algorithm given a measurement matrix and samples from a signal or
function. Note that this function is a convenient wrapper for the function FastLaplace.
Usage
FindSparse(PHI, y, eta = 1e-08, roundit = FALSE, verbose = FALSE)
Arguments
PHI
typically equals the product of a measurment matrix and basis representation
matrix, such as the wavelet basis. The solution vector w (see Details below) is
assumed to be sparse in the chosen basis.
y
CS measurements, samples from the signal or function.
eta
tolerance level in determining convergence of marginal likelihood.
roundit
whether or not to round the marginal likelihood, in order to avoid machine precision error when comparing across platforms.
verbose
print to screen which basis are added, re-estimated, or deleted.
Details
This code implements the fast Laplace algorithm. The fast Laplace algorithm is a method used to
solve the compressive sensing problem, or in general, a highly underdetermined system of equations. It does this by taking the system of equations
y = Φw + n
and converting it into a minimization problem where we minimize the error with a constraint on
w (the vector we are solving for) that enforces sparsity. The fast Laplace method uses a Bayesian
framework, and in particular, uses a Laplace prior to enforce sparsity on w. See [1] for more
information.
Value
The sparse signal w as found by the fast Laplace algorithm.
References
[1] S. D. Babacan, R. Molina and A. K. Katsaggelos, "Bayesian Compressive Sensing Using
Laplace Priors," in IEEE Transactions on Image Processing, vol. 19, no. 1, pp. 53-63, Jan. 2010.
FourierBasis
5
Examples
# size of the basis function expansion
N <- 64
# generate sparse coefficient vector
w <- rep(0,N)
w[sample(1:N,10)] <- runif(10,-1,1)
# create wavelet basis trasform matrix
wavelet.basis <- WaveletBasis(N)
# generate actual signal
signal <- wavelet.basis%*%w
# now we try and recover 'w' and 'signal' from samples
num_samps <- 25
# create random measurement matrix
measure.mat <- matrix(runif(num_samps*N),num_samps,N)
measure.mat <- measure.mat/matrix(rep(sqrt(apply(measure.mat^2,2,sum)),
num_samps),num_samps,N,byrow=TRUE);
PHI <- measure.mat%*%wavelet.basis
# actual samples we see
y <- measure.mat%*%signal
# use fast Laplace algorithm
w_est <- FindSparse(PHI, y)
# compare plots of the sparse vector and the estimated sparse vector
plot(w,type='h')
lines(w_est,type='h',col='red')
# estimate signal
signal_est <- wavelet.basis%*%w_est
# Root mean squared error of estimate
error <- sqrt(mean((signal - signal_est)^2))
FourierBasis
Finds the Discrete Fourier Transformation Matrix
Description
Uses the package fda to find a transformation matrix where the columns are the different Fourier
basis functions, evaluated at specified points along the rows of the matrix.
Usage
FourierBasis(tlist, N, train = NULL)
6
WaveletBasis
Arguments
tlist
an array of the specific points where the basis functions are evaluated at.
N
number of basis functions in the matrix.
train
indices corresponding to which rows of the matrix to keep. Default is to keep
all rows.
Value
A P x N discrete Fourier transformation matrix where P is equal to the length of train and N is the
number of basis. If train is NULL then P equals the length of tlist.
Examples
# Points to evaluate the basis functions at
points <- c(0,1,2,3,4)
# Find first 8 Fourier basis functions evaluted at "points"
f <- FourierBasis(points, 8)
# Find first 8 Fourier basis functions evaluated at "points" but only keep
# last two rows
f <- FourierBasis(points, 8, train = c(4,5))
WaveletBasis
Finds the Discrete Wavelet Transform Matrix
Description
Uses the functions wavDWT and reconstruct from the wmtsa package to find the transformation
matrix of the given wavelet basis type. Each column of the matrix is a wavelet basis function.
Usage
WaveletBasis(N, train = NULL, wavelet = "Haar")
Arguments
N
number of wavelet basis functions to include in matrix. Note that N must be a
power of 2, otherwise the matrix will include NA’s. The reason for this has to
do with how the wavelet basis is defined.
train
indices corresponding to which rows of the matrix to keep. Default is to keep
all rows.
wavelet
the type of wavelet basis to use. See wavDaubechies from wmtsa for types.
Value
A P x N discrete wavelet transform matrix, where P is equal to the length of train and N is the
number of basis. If train is NULL then P equals N.
WaveletBasis
Examples
# Find first 8 basis functions of the Haar wavelet type
w.Haar <- WaveletBasis(8)
# Find first 8 basis functions of the d4 wavelet type, keeping the first
# half of the rows
w.d4 <- WaveletBasis(8, 1:4, wavelet='d4')
7
Index
BSplineBasis, 2
FastLaplace, 2, 4
fda, 2, 5
FindSparse, 2, 4
FourierBasis, 5
reconstruct, 6
wavDaubechies, 6
wavDWT, 6
WaveletBasis, 6
8