1## Simple Gibbs Sampler Example
2## Adapted from Darren Wilkinson's post at
3## http://darrenjw.wordpress.com/2010/04/28/mcmc-programming-in-r-python-java-and-c/
5## Sanjog Misra and Dirk Eddelbuettel, June-July 2011
7suppressMessages(library(Rcpp))
8suppressMessages(library(inline))
9suppressMessages(library(compiler))
10suppressMessages(library(rbenchmark))
13## Actual joint density -- the code which follow implements
14## a Gibbs sampler to draw from the following joint density f(x,y)
16 x*x * exp(-x*y*y - y*y + 2*y - 4*x)
19## Note that the full conditionals are propotional to
20## f(x|y) = (x^2)*exp(-x*(4+y*y)) : a Gamma density kernel
21## f(y|x) = exp(-0.5*2*(x+1)*(y^2 - 2*y/(x+1)) : Normal Kernel
23## There is a small typo in Darrens code.
24## The full conditional for the normal has the wrong variance
25## It should be 1/sqrt(2*(x+1)) not 1/sqrt(1+x)
26## This we can verify ...
27## The actual conditional (say for x=3) can be computed as follows
28## First - Construct the Unnormalized Conditional
29fy.unnorm <- function(y) fun(3,y)
31## Then - Find the appropriate Normalizing Constant
32K <- integrate(fy.unnorm,-Inf,Inf)
34## Finally - Construct Actual Conditional
35fy <- function(y) fy.unnorm(y)/K$val
37## Now - The corresponding Normal should be
38fy.dnorm <- function(y) {
40 dnorm(y,1/(1+x),sqrt(1/(2*(1+x))))
44fy.dnorm.wrong <- function(y) {
46 dnorm(y,1/(1+x),sqrt(1/((1+x))))
51 ## Actual (gray thick line)
52 curve(fy,-2,2,col='grey',lwd=5)
54 ## Correct Normal conditional (blue dotted line)
55 curve(fy.dnorm,-2,2,col='blue',add=T,lty=3)
57 ## Wrong Normal (Red line)
58 curve(fy.dnorm.wrong,-2,2,col='red',add=T)
61## Here is the actual Gibbs Sampler
62## This is Darren Wilkinsons R code (with the corrected variance)
63## But we are returning only his columns 2 and 3 as the 1:N sequence
65Rgibbs <- function(N,thin) {
66 mat <- matrix(0,ncol=2,nrow=N)
71 x <- rgamma(1,3,y*y+4)
72 y <- rnorm(1,1/(x+1),1/sqrt(2*(x+1)))
79## We can also try the R compiler on this R function
80RCgibbs <- cmpfun(Rgibbs)
83## mat <- Rgibbs(10000,10); dim(mat)
84## would give: [1] 10000 2
86## Now for the Rcpp version -- Notice how easy it is to code up!
88## NOTE: This is the old way to compile Rcpp code inline.
89## The code here has left as a historical artifact and tribute to the old way.
90## Please use the code under the "new" inline compilation section.
94 using namespace Rcpp; // inline does that for us already
96 // n and thin are SEXPs which the Rcpp::as function maps to C++ vars
98 int thn = as<int>(thin);
101 NumericMatrix mat(N, 2);
103 RNGScope scope; // Initialize Random number generator. Not needed when Attributes used.
105 // The rest of the code follows the R version
108 for (i=0; i<N; i++) {
109 for (j=0; j<thn; j++) {
110 x = ::Rf_rgamma(3.0,1.0/(y*y+4));
111 y = ::Rf_rnorm(1.0/(x+1),1.0/sqrt(2*x+2));
117 return mat; // Return to R
121RcppGibbs_old <- cxxfunction(signature(n="int", thin = "int"),
122 gibbscode, plugin="Rcpp")
126 #include <gsl/gsl_rng.h>
127 #include <gsl/gsl_randist.h>
129 using namespace Rcpp; // just to be explicit
134 int thin = as<int>(thns);
136 gsl_rng *r = gsl_rng_alloc(gsl_rng_mt19937);
138 NumericMatrix mat(N, 2);
139 for (i=0; i<N; i++) {
140 for (j=0; j<thin; j++) {
141 x = gsl_ran_gamma(r,3.0,1.0/(y*y+4));
142 y = 1.0/(x+1)+gsl_ran_gaussian(r,1.0/sqrt(2*x+2));
149 return mat; // Return to R
153GSLGibbs_old <- cxxfunction(signature(ns="int", thns = "int"),
154 body=gslgibbscode, includes=gslgibbsincl,
157## without RcppGSL, using cfunction()
158#GSLGibbs <- cfunction(signature(ns="int", thns = "int"),
159# body=gslgibbscode, includes=gslgibbsincl,
161# cppargs="-I/usr/include",
162# libargs="-lgsl -lgslcblas")
165## NOTE: Within this section, the new way to compile Rcpp code inline has been
166## written. Please use the code next as a template for your own project.
168## Use of the cppFunction() gives the ability to immediately compile embed C++
169## without having to worry about header specification or Rcpp attributes.
172NumericMatrix RcppGibbs(int N, int thn){
173 // Note: n and thin are SEXPs which the Rcpp automatically converts to ints
175 // Setup storage matrix
176 NumericMatrix mat(N, 2);
178 // The rest of the code follows the R version
181 for (int i = 0; i < N; i++) {
182 for (int j = 0; j < thn; j++) {
183 x = R::rgamma(3.0,1.0/(y*y+4));
184 y = R::rnorm(1.0/(x+1),1.0/sqrt(2*x+2));
190 return mat; // Return to R
194## Use of the sourceCpp() is preferred for users who wish to source external
195## files or specify their headers and Rcpp attributes within their code.
196## Code here is able to easily be extracted and placed into its own C++ file.
201#include <gsl/gsl_rng.h>
202#include <gsl/gsl_randist.h>
204using namespace Rcpp; // just to be explicit
206// [[Rcpp::depends(RcppGSL)]]
209NumericMatrix GSLGibbs(int N, int thin){
210 gsl_rng *r = gsl_rng_alloc(gsl_rng_mt19937);
212 NumericMatrix mat(N, 2);
213 for (int i = 0; i < N; i++) {
214 for (int j = 0; j < thin; j++) {
215 x = gsl_ran_gamma(r,3.0,1.0/(y*y+4));
216 y = 1.0/(x+1)+gsl_ran_gaussian(r,1.0/sqrt(2*x+2));
223 return mat; // Return to R
229## You can try other values if you like
230## Note that the total number of interations are N*thin!
231Ns <- c(1000,5000,10000,20000)
232thins <- c(10,50,100,200)
238for (i in seq_along(Ns)) {
239 tim_R[i] <- system.time(mat <- Rgibbs(Ns[i],thins[i]))[3]
240 tim_RC[i] <- system.time(cmat <- RCgibbs(Ns[i],thins[i]))[3]
241 tim_Rgsl[i] <- system.time(gslmat <- GSLGibbs(Ns[i],thins[i]))[3]
242 tim_Rcpp[i] <- system.time(rcppmat <- RcppGibbs(Ns[i],thins[i]))[3]
243 cat("Replication #", i, "complete \n")
247speedup <- round(tim_R/tim_Rcpp,2);
248speedup2 <- round(tim_R/tim_Rgsl,2);
249speedup3 <- round(tim_R/tim_RC,2);
250summtab <- round(rbind(tim_R,tim_RC, tim_Rcpp,tim_Rgsl,speedup3,speedup,speedup2),3)
251colnames(summtab) <- c("N=1000","N=5000","N=10000","N=20000")
252rownames(summtab) <- c("Elasped Time (R)","Elasped Time (RC)","Elapsed Time (Rcpp)", "Elapsed Time (Rgsl)",
253 "SpeedUp Rcomp.","SpeedUp Rcpp", "SpeedUp GSL")
257## Contour Plots -- based on Darren's example
258if (interactive() && require(KernSmooth)) {
259 op <- par(mfrow=c(4,1),mar=c(3,3,3,1))
263 contour(x,y,z,main="Contours of actual distribution",xlim=c(0,2), ylim=c(-2,4))
264 fit <- bkde2D(as.matrix(mat),c(0.1,0.1))
265 contour(drawlabels=T, fit$x1, fit$x2, fit$fhat, xlim=c(0,2), ylim=c(-2,4),
266 main=paste("Contours of empirical distribution:",round(tim_R[4],2)," seconds"))
267 fitc <- bkde2D(as.matrix(rcppmat),c(0.1,0.1))
268 contour(fitc$x1,fitc$x2,fitc$fhat,xlim=c(0,2), ylim=c(-2,4),
269 main=paste("Contours of Rcpp based empirical distribution:",round(tim_Rcpp[4],2)," seconds"))
270 fitg <- bkde2D(as.matrix(gslmat),c(0.1,0.1))
271 contour(fitg$x1,fitg$x2,fitg$fhat,xlim=c(0,2), ylim=c(-2,4),
272 main=paste("Contours of GSL based empirical distribution:",round(tim_Rgsl[4],2)," seconds"))
277## also use rbenchmark package
280res <- benchmark(Rgibbs(N, thn),
284 columns=c("test", "replications", "elapsed",
285 "relative", "user.self", "sys.self"),