Question

我需要将大矩阵转换为特定格式以与libsvm一起使用。格式包括使用标签（1或-1）开始每一行，后跟0： row_number ，行值为1： value_at_row_number_1st_column

下面给出的简单方法太慢了，

require(microbenchmark) 
nR = 100; nC = 500
kMat = matrix(runif(nR*nC), nrow=nR)
yLab = sample(c(1, -1), nR, replace = T)

# Simple method
met1 = function() {
  lines = c()
  for(ix in 1:nrow(kMat))
    lines = c(lines, 
          paste(yLab[ix],
                paste0("0:", ix),
                paste0(1:ncol(kMat), ":", kMat[ix, ], collapse=" ")))
  lines
}

我的版本速度提高了约50％（虽然方式更加丑陋），

# Sprintf
met2 = function() {
  fmt = c("%i", "0:%i", paste0(1:ncol(kMat), ":%f")) 
  kMat = cbind(yLab, 1:nrow(kMat), kMat)
  # Unfortunately sprintf cannot handle more than 100 arguments
  splts = lapply(seq(1, length(fmt), 99L),
                 function(ix) {
                   r = ix:min(ncol(kMat), ix+98L)
                   list(range = r, fmt = list(paste(fmt[r], collapse =     " ")) )
             })
  lines = sapply(1:nrow(kMat),
                 function(ix) {
                   Reduce(function(a, b) sprintf("%s %s", a, b), 
                          sapply(splts, 
                                 function(s){
                                   do.call(sprintf, c(s$fmt, kMat[ix, s$range]))
                                 }),
                          "")
                 })
  lines
}
print(microbenchmark(met1(), met2())) 

Unit: milliseconds
   expr      min       lq     mean   median       uq      max neval
 met1() 85.83051 88.00289 92.01948 88.61834 90.31918 175.3362   100
 met2() 44.81729 45.61020 56.12835 54.75313 56.65249 108.7218   100

是否有更快（或更整洁）的方式来处理这种格式？

Answer 1

这既短又快。似乎sprintf从double到character的转换比隐式as.character更快，而stringi::stri_join比paste加快了一些速度1}}和paste0。我也尝试过删除转置的变体，但下面的代码更快。

library(stringi)

met3 <- function() {   
    s <- stri_join("0:", seq_len(nC), " ", sprintf("%f", t(kMat)))
    m <- matrix(s, nC)
    stri_join(yLab, apply(m, 2, stri_join, collapse = " "), sep = " ")
}

，并提供：

> microbenchmark(met3(), met2(), times = 10)
Unit: milliseconds
   expr      min       lq     mean   median       uq      max neval
 met3() 236.6127 255.1396 264.7797 256.6331 292.1037 296.6377    10
 met2() 307.6371 322.1467 354.7281 332.0041 388.2474 464.2259    10

加快矩阵格式化

1 个答案: