Rcpp Version 1.0.14
Loading...
Searching...
No Matches
RcppGibbs_Updated.R
Go to the documentation of this file.
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/
4##
5## Sanjog Misra and Dirk Eddelbuettel, June-July 2011
6## Updated by Dirk Eddelbuettel, Mar 2020
7
8suppressMessages({
9 library(Rcpp)
10 library(rbenchmark)
11})
12
13
14## Actual joint density -- the code which follow implements
15## a Gibbs sampler to draw from the following joint density f(x,y)
16fun <- function(x,y) {
17 x*x * exp(-x*y*y - y*y + 2*y - 4*x)
18}
19
20## Note that the full conditionals are propotional to
21## f(x|y) = (x^2)*exp(-x*(4+y*y)) : a Gamma density kernel
22## f(y|x) = exp(-0.5*2*(x+1)*(y^2 - 2*y/(x+1)) : Normal Kernel
23
24## There is a small typo in Darrens code.
25## The full conditional for the normal has the wrong variance
26## It should be 1/sqrt(2*(x+1)) not 1/sqrt(1+x)
27## This we can verify ...
28## The actual conditional (say for x=3) can be computed as follows
29## First - Construct the Unnormalized Conditional
30fy.unnorm <- function(y) fun(3,y)
31
32## Then - Find the appropriate Normalizing Constant
33K <- integrate(fy.unnorm,-Inf,Inf)
34
35## Finally - Construct Actual Conditional
36fy <- function(y) fy.unnorm(y)/K$val
37
38## Now - The corresponding Normal should be
39fy.dnorm <- function(y) {
40 x <- 3
41 dnorm(y,1/(1+x),sqrt(1/(2*(1+x))))
42}
43
44## and not ...
45fy.dnorm.wrong <- function(y) {
46 x <- 3
47 dnorm(y,1/(1+x),sqrt(1/((1+x))))
48}
49
50if (interactive()) {
51 ## Graphical check
52 ## Actual (gray thick line)
53 curve(fy,-2,2,col='grey',lwd=5)
54
55 ## Correct Normal conditional (blue dotted line)
56 curve(fy.dnorm,-2,2,col='blue',add=T,lty=3)
57
58 ## Wrong Normal (Red line)
59 curve(fy.dnorm.wrong,-2,2,col='red',add=T)
60}
61
62## Here is the actual Gibbs Sampler
63## This is Darren Wilkinsons R code (with the corrected variance)
64## But we are returning only his columns 2 and 3 as the 1:N sequence
65## is never used below
66Rgibbs <- function(N,thin) {
67 mat <- matrix(0,ncol=2,nrow=N)
68 x <- 0
69 y <- 0
70 for (i in 1:N) {
71 for (j in 1:thin) {
72 x <- rgamma(1,3,y*y+4)
73 y <- rnorm(1,1/(x+1),1/sqrt(2*(x+1)))
74 }
75 mat[i,] <- c(x,y)
76 }
77 mat
78}
79
80## Now for the Rcpp version -- Notice how easy it is to code up!
81
82cppFunction("NumericMatrix RcppGibbs(int N, int thn){
83 NumericMatrix mat(N, 2); // Setup storage
84 double x = 0, y = 0; // The rest follows the R version
85 for (int i = 0; i < N; i++) {
86 for (int j = 0; j < thn; j++) {
87 x = R::rgamma(3.0,1.0/(y*y+4));
88 y = R::rnorm(1.0/(x+1),1.0/sqrt(2*x+2));
89 }
90 mat(i,0) = x;
91 mat(i,1) = y;
92 }
93 return mat; // Return to R
94}")
95
96
97## Use of the sourceCpp() is preferred for users who wish to source external
98## files or specify their headers and Rcpp attributes within their code.
99## Code here is able to easily be extracted and placed into its own C++ file.
100
101## Compile and Load
102sourceCpp(code="
103#include <RcppGSL.h>
104#include <gsl/gsl_rng.h>
105#include <gsl/gsl_randist.h>
106
107using namespace Rcpp; // just to be explicit
108
109// [[Rcpp::depends(RcppGSL)]]
110
111// [[Rcpp::export]]
112NumericMatrix GSLGibbs(int N, int thin){
113 gsl_rng *r = gsl_rng_alloc(gsl_rng_mt19937);
114 double x = 0, y = 0;
115 NumericMatrix mat(N, 2);
116 for (int i = 0; i < N; i++) {
117 for (int j = 0; j < thin; j++) {
118 x = gsl_ran_gamma(r,3.0,1.0/(y*y+4));
119 y = 1.0/(x+1)+gsl_ran_gaussian(r,1.0/sqrt(2*x+2));
120 }
121 mat(i,0) = x;
122 mat(i,1) = y;
123 }
124 gsl_rng_free(r);
125
126 return mat; // Return to R
127}")
128
129
130
131## Now for some tests
132## You can try other values if you like
133## Note that the total number of interations are N*thin!
134Ns <- c(1000,5000,10000,20000)
135thins <- c(10,50,100,200)
136tim_R <- rep(0,4)
137tim_Rgsl <- rep(0,4)
138tim_Rcpp <- rep(0,4)
139
140for (i in seq_along(Ns)) {
141 tim_R[i] <- system.time(mat <- Rgibbs(Ns[i],thins[i]))[3]
142 tim_Rgsl[i] <- system.time(gslmat <- GSLGibbs(Ns[i],thins[i]))[3]
143 tim_Rcpp[i] <- system.time(rcppmat <- RcppGibbs(Ns[i],thins[i]))[3]
144 cat("Replication #", i, "complete \n")
145}
146
147## Comparison
148speedup <- round(tim_R/tim_Rcpp,2);
149speedup2 <- round(tim_R/tim_Rgsl,2);
150summtab <- round(rbind(tim_R, tim_Rcpp,tim_Rgsl,speedup,speedup2),3)
151colnames(summtab) <- c("N=1000","N=5000","N=10000","N=20000")
152rownames(summtab) <- c("Elasped Time (R)","Elapsed Time (Rcpp)", "Elapsed Time (Rgsl)",
153 "SpeedUp Rcpp", "SpeedUp GSL")
154print(summtab)
155
156## Contour Plots -- based on Darren's example
157if (interactive() && require(KernSmooth)) {
158 op <- par(mfrow=c(4,1),mar=c(3,3,3,1))
159 x <- seq(0,4,0.01)
160 y <- seq(-2,4,0.01)
161 z <- outer(x,y,fun)
162 contour(x,y,z,main="Contours of actual distribution",xlim=c(0,2), ylim=c(-2,4))
163 fit <- bkde2D(as.matrix(mat),c(0.1,0.1))
164 contour(drawlabels=T, fit$x1, fit$x2, fit$fhat, xlim=c(0,2), ylim=c(-2,4),
165 main=paste("Contours of empirical distribution:",round(tim_R[4],2)," seconds"))
166 fitc <- bkde2D(as.matrix(rcppmat),c(0.1,0.1))
167 contour(fitc$x1,fitc$x2,fitc$fhat,xlim=c(0,2), ylim=c(-2,4),
168 main=paste("Contours of Rcpp based empirical distribution:",round(tim_Rcpp[4],2)," seconds"))
169 fitg <- bkde2D(as.matrix(gslmat),c(0.1,0.1))
170 contour(fitg$x1,fitg$x2,fitg$fhat,xlim=c(0,2), ylim=c(-2,4),
171 main=paste("Contours of GSL based empirical distribution:",round(tim_Rgsl[4],2)," seconds"))
172 par(op)
173}
174
175
176## also use rbenchmark package
177N <- 20000
178thn <- 200
179res <- benchmark(Rgibbs(N, thn),
180 RcppGibbs(N, thn),
181 GSLGibbs(N, thn),
182 columns=c("test", "replications", "elapsed",
183 "relative", "user.self", "sys.self"),
184 order="relative",
185 replications=10)
186print(res)
187
188
189## And we are done