R功能效率

Question

对以下功能更有效地实现的任何想法？

prod_A_B <- function(A, B, i = NULL, j = NULL) {
  if (is.null(i) & is.null(j)) {
    A * B
  } else if (!is.null(i) & is.null(j)) {
    A[i, ] * B[i, ]
  } else if (!is.null(i) & !is.null(j)) {
    A[i, j] * B[i, j]
  }
}

尤其是，是否有可能对条件语句进行更清洁的实现，从而减少运行时？

这是一些基准：

n <- 1e5
d <- 200
A <- matrix(rnorm(n*d), nrow = n, ncol = d)
B <- matrix(rnorm(n*d), nrow = n, ncol = d)
pr1 <- pr2 <- matrix(0, nrow=n, ncol =d)
tm <- microbenchmark(
  {for(i in 1:n) pr1[i,] <- prod_A_B(A, B, i)},
  {for(i in 1:n) pr2[i, ] <- A[i,] * B[i,]}, 
  times=100L)
print(tm)
Unit: milliseconds
       min       lq     mean   median       uq      max neval
 1164.4078 1208.348 1234.607 1224.097 1254.059 1370.098   100
  969.9961 1004.821 1036.738 1021.025 1056.182 1278.767   100

鉴于矩阵尺寸的差异并不大，但是我仍然想知道我是否可以提高性能...

谢谢！

Answer 1

您可以通过简化逻辑谓词并消除多个函数调用，如下所示：

：

prod2_A_B <- function(A, B, i = NULL, j = NULL) {
  ni <- is.null(i)
  nj <- is.null(j)
  if (ni & nj) {
    A * B
  } else if (nj) { # must be !ni
    A[i, ] * B[i, ]
  } else {         # must be !ni & !nj
    A[i, j] * B[i, j]
  }
}

n <- 1e5
d <- 200
pr1 <- pr2 <- pr <- matrix(0, nrow=n, ncol =d)
A <- matrix(rnorm(n*d), nrow = n, ncol = d)
B <- matrix(rnorm(n*d), nrow = n, ncol = d)
library(microbenchmark)
tm <- microbenchmark(
  {for(i in 1:n) pr1[i,] <- prod_A_B(A, B, i)},
  {for(i in 1:n) pr2[i,] <- prod2_A_B(A, B, i)},
  {for(i in 1:n) pr[i,] <- A[i,] * B[i,]}, 
  times=100L)
print(tm)
Unit: milliseconds
                                           expr      min       lq     mean   median       uq      max neval cld
  {for (i in 1:n) pr1[i, ] <- prod_A_B(A, B, i)}  937.9470 944.6690 969.7894 952.2308 964.4701 1390.467   100   c
  {for (i in 1:n) pr2[i, ] <- prod2_A_B(A, B, i)} 898.6802 908.3323 929.7343 914.6826 929.4356 1211.623   100  b 
  {for (i in 1:n) pr[i, ] <- A[i, ] * B[i, ]}     661.2350 666.7071 688.8127 672.6218 679.9028 1005.342   100 a