通过向量乘以矩阵行？

Question

我正在优化一个函数,我想摆脱缓慢的for循环.我正在寻找一种更快的方法来将矩阵的每一行乘以一个向量.

有任何想法吗？

编辑:

我不是在寻找'经典'乘法.

例如.我有一个有23列和25行的矩阵和一个长度为23的向量.结果我想要矩阵25x23,每行乘以向量.

Answer 1

我想你在找sweep().

# Create example data and vector
mat <- matrix(rep(1:3,each=5),nrow=3,ncol=5,byrow=TRUE)
     [,1] [,2] [,3] [,4] [,5]
[1,]    1    1    1    1    1
[2,]    2    2    2    2    2
[3,]    3    3    3    3    3

vec <- 1:5

# Use sweep to apply the vector with the multiply (`*`) function
#  across columns (See ?apply for an explanation of MARGIN) 
sweep(mat, MARGIN=2, vec, `*`)
     [,1] [,2] [,3] [,4] [,5]
[1,]    1    2    3    4    5
[2,]    2    4    6    8   10
[3,]    3    6    9   12   15

这是R的核心功能之一,虽然多年来已经对其进行了改进.

Answer 2

> MyMatrix <- matrix(c(1,2,3, 11,12,13), nrow = 2, ncol=3, byrow=TRUE)
> MyMatrix
     [,1] [,2] [,3]
[1,]    1    2    3
[2,]   11   12   13
> MyVector <- c(1:3)
> MyVector
[1] 1 2 3

你可以使用:

> t(t(MyMatrix) * MyVector)
     [,1] [,2] [,3]
[1,]    1    4    9
[2,]   11   24   39

要么:

> MyMatrix %*% diag(MyVector)
     [,1] [,2] [,3]
[1,]    1    4    9
[2,]   11   24   39

Answer 3

实际上,sweep这不是我电脑上最快的选择:

MyMatrix <- matrix(c(1:1e6), ncol=1e4, byrow=TRUE)
MyVector <- c(1:1e4)

Rprof(tmp <- tempfile(),interval = 0.001)
t(t(MyMatrix) * MyVector) # first option
Rprof()
MyTimerTranspose=summaryRprof(tmp)$sampling.time
unlink(tmp)

Rprof(tmp <- tempfile(),interval = 0.001)
MyMatrix %*% diag(MyVector) # second option
Rprof()
MyTimerDiag=summaryRprof(tmp)$sampling.time
unlink(tmp)

Rprof(tmp <- tempfile(),interval = 0.001)
sweep(MyMatrix ,MARGIN=2,MyVector,`*`)  # third option
Rprof()
MyTimerSweep=summaryRprof(tmp)$sampling.time
unlink(tmp)

Rprof(tmp <- tempfile(),interval = 0.001)
t(t(MyMatrix) * MyVector) # first option again, to check order 
Rprof()
MyTimerTransposeAgain=summaryRprof(tmp)$sampling.time
unlink(tmp)

MyTimerTranspose
MyTimerDiag
MyTimerSweep
MyTimerTransposeAgain

这会产生:

> MyTimerTranspose
[1] 0.04
> MyTimerDiag
[1] 40.722
> MyTimerSweep
[1] 33.774
> MyTimerTransposeAgain
[1] 0.043

除了是最慢的选项之外,第二个选项达到内存限制(2046 MB).然而,考虑到剩下的选择,双转置似乎比sweep我看来要好很多.

编辑

只是重复尝试较小的数据:

MyMatrix <- matrix(c(1:1e3), ncol=1e1, byrow=TRUE)
MyVector <- c(1:1e1)
n=100000

[...]

for(i in 1:n){
# your option
}

[...]

> MyTimerTranspose
[1] 5.383
> MyTimerDiag
[1] 6.404
> MyTimerSweep
[1] 12.843
> MyTimerTransposeAgain
[1] 5.428

Answer 4

为了速度，可以在乘法之前从向量创建矩阵

mat <-  matrix(rnorm(1e6), ncol=1e4)
vec <- c(1:1e4)
mat * matrix(vec, dim(mat)[1], length(vec))

library(microbenchmark)
microbenchmark(
  transpose = t(t(mat) * vec), 
  make_matrix = mat * matrix(vec, dim(mat)[1], length(vec), byrow = TRUE),
  sweep = sweep(mat,MARGIN=2,vec,`*`))
#Unit: milliseconds
#       expr      min        lq     mean    median       uq      max neval cld
#  transpose 9.940555 10.480306 14.39822 11.210735 16.19555 77.67995   100   b
#make_matrix 5.556848  6.053933  9.48699  6.662592 10.74121 74.14429   100   a 
#      sweep 8.033019  8.500464 13.45724 12.331015 14.14869 77.00371   100   b