[英]How to make my loop run faster in R?
I'm using a function to get p-values from multiple HWE chi square tests. 我正在使用一个函数来从多个HWE卡方检验中获取p值。 I'm looping through a large matrix called geno.data
which is (313 rows x 355232 columns) to do this. 我正在遍历一个名为geno.data
的大型矩阵,该矩阵是(313行x 355232列)来执行此操作。 I'm essentially looping two columns of the matrix at a time by row. 我本质上是一次一行地循环矩阵的两列。 It runs very slowly. 它运行非常缓慢。 How can I make it faster? 我怎样才能使其更快? Thanks 谢谢
library(genetics)
geno.data<-matrix(c("a","c"), nrow=313,ncol=355232)
Num_of_SNPs<-ncol(geno.data) /2
alleles<- vector(length = nrow(geno.data))
HWE_pvalues<-vector(length = Num_of_SNPs)
j<- 1
for (count in 1:Num_of_SNPs){
for (i in 1:nrow(geno.data)){
alleles[i]<- levels(genotype(paste(geno.data[i,c(2*j -1, 2*j)], collapse = "/")))
}
g2 <- genotype(alleles)
HWE_pvalues[count]<-HWE.chisq(g2)[3]
j = j + 2
}
First, note that the posted code will result in an index-out-of-bounds error, because after Num_of_SNPs
iterations of the main loop your j
value will be ncol(geno.data)-1
and you're accessing columns 2*j-1
and 2*j
. 首先,请注意,发布的代码将导致索引越界错误,因为在主循环的Num_of_SNPs
次迭代之后,您的j
值为ncol(geno.data)-1
并且您正在访问列2*j-1
和2*j
。 I'm assuming you instead want columns 2*count-1
and 2*count
and j
can be removed. 我假设您改为希望删除列2*count-1
和2*count
和j
。
Vectorization is extremely important for writing fast R code. 向量化对于编写快速的R代码极为重要。 In your code you're calling the paste
function 313 times, each time passing vectors of length 1. It's much faster in R to call paste
once passing vectors of length 313. Here are the original and vectorized interiors of the main for loop: 在您的代码中,每次传递paste
长度为1的向量时,都会调用paste
函数313次。在R中,传递传递长度为313的向量时,调用paste
速度要快得多。这是main for循环的原始内部矢量化:
# Original
get.pval1 <- function(count) {
for (i in 1:nrow(geno.data)){
alleles[i]<- levels(genotype(paste(geno.data[i,c(2*count -1, 2*count)], collapse = "/")))
}
g2 <- genotype(alleles)
HWE.chisq(g2)[3]
}
# Vectorized
get.pval2 <- function(count) {
g2 <- genotype(paste0(geno.data[,2*count-1], "/", geno.data[,2*count]))
HWE.chisq(g2)[3]
}
We get about a 20x speedup from the vectorization: 向量化可以使速度提高20倍:
library(microbenchmark)
all.equal(get.pval1(1), get.pval2(1))
# [1] TRUE
microbenchmark(get.pval1(1), get.pval2(1))
# Unit: milliseconds
# expr min lq mean median uq max neval
# get.pval1(1) 299.24079 304.37386 323.28321 307.78947 313.97311 482.32384 100
# get.pval2(1) 14.23288 14.64717 15.80856 15.11013 16.38012 36.04724 100
With the vectorized code, your code should finish in about 177616*.01580856 = 2807.853 seconds, or about 45 minutes (compared to 16 hours for the original code). 使用矢量化代码,您的代码应在大约177616 * .01580856 = 2807.853秒内完成,或大约45分钟(原始代码为16小时)。 If this is still not fast enough for you, then I would encourage you to look at the parallel
package in R. The mcmapply
should give a good speedup for you, since each iteration of the outer for
loop is independent. 如果这仍然不够快,那么我鼓励您看一下R中的parallel
包mcmapply
应该为您提供良好的加速,因为外部for
循环的每次迭代都是独立的。
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.