Re: [筆記] R 的設定-平行運算 (CPU)

[celestialgod (攸藍)]

分享另一個平行運算的package: snowfall

以下是一個例子 for cross validation

### Example: Cross validation for choosing
### the tuning parameter of ridge regression

## Data
data(longley)
X <- as.matrix(longley[,1:5])
Y <- as.matrix(longley[,7])

## Normalization
norm <- function(a) (a-mean(a))/sd(a)
Y.N <- apply(Y, 2, norm)
X.N <- apply(X, 2, norm)
D <- cbind(Y.N, X.N)
n <- nrow(X.N)
p <- ncol(X.N)

## Ridge regression, given lambda
beta.R <- function(Y, X, nc, lambda) solve(t(X)%*%X +
lambda*diag(nc),t(X)%*%Y)

## cross validation, single CPU version
m <- 1001
lambda <- seq(0, 1, length=m)
CV <- sapply(1:m, function(j){
        out <- sapply(1:n, function(i)
        beta.hat <- beta.R(Y.N[-i,], X.N[-i,], p, lambda[j]))
        out <- mean((Y.N - diag(X.N%*%out))^2)
})

plot(lambda, CV, type='l', xlab=expression(lambda), ylab='CV')
lambda.opt <- lambda[which(CV == min(CV))]
beta.opt <- beta.R(Y.N, X.N, p, lambda.opt)

## cross validation, parallel version
##   in this case, the sample size is small but the number of lambda is large
##   so we replace the outer loop by parallel computing

## prepare for parallel computing
library(snowfall)  # parallel computing
sfInit(parallel = TRUE, cpus = 8, type="SOCK") # built four nodes

## Tell all nodes that following objects are going to be used
sfExport('Y.N', 'X.N','m','n','p','beta.R', 'lambda')
inner.loop <- function(j){
        A <- apply(matrix(1:m,m,1), 1, function(i)
        beta.hat <- beta.R(Y.N[-i], X.N[-i,], p, lambda[j]))
        out <- mean((Y.N - diag(X.N%*%A))^2)
        return(out)
}
CV <- sfSapply(seq(along=Y.N), inner.loop)

## evaluate the performance
wallTime <- function(expr) system.time(expr)[3]
wallTime(sapply(1:m, function(j){
        out <- sapply(1:n, function(i)
        beta.hat <- beta.R(Y.N[-i,], X.N[-i,], p, lambda[j]))
        out <- mean((Y.N - diag(X.N%*%out))^2)
}))

wallTime(sfSapply(seq(along=Y.N), inner.loop))
sfStop()


--
※ 發信站: 批踢踢實業坊(ptt.cc) 
◆ From: 218.164.73.129
推 Wush978:謝謝分享，有空來看看
推 gsuper: