Question

我试图了解如何编写Rcpp汇总使用dplyr快速执行的函数。这样做的动机是dplyr似乎没有等效的函数，但是为了简单起见，我将使用仅仅取向量的最后一个元素的例子。

在下面的代码中，我考虑了三个不同的函数来获取向量的最后一个元素，并使用tapply和dplyr group_by / summarize来应用它们。

library(dplyr)
library(microbenchmark)
library(Rcpp)
n <- 5000
df <- data.frame(grp = factor(rep(1:n, 2)), valn = rnorm(2L*n), stringsAsFactors = F)

dplyr_num_last_element <- function() df %>% group_by(grp) %>% summarise(valn = last(valn))
dplyr_num_last_element_r <- function() df %>% group_by(grp) %>% summarise(valn = last_r(valn))
dplyr_num_last_element_rcpp <- function() df %>% group_by(grp) %>% summarise(val = last_rcpp(valn))
tapply_num_last_element <- function() tapply(df$valn, df$grp, FUN = last)
tapply_num_last_element_r <- function() tapply(df$valn, df$grp, FUN = last_r)
tapply_num_last_element_rcpp <- function() tapply(df$valn, df$grp, FUN = last_rcpp)

last_r <- function(x) {
  x[1]
}

cppFunction('double last_rcpp(NumericVector x) {
             int n = x.size();
             return x[n-1];
           }')

microbenchmark(dplyr_num_last_element(), dplyr_num_last_element_r(), dplyr_num_last_element_rcpp(), tapply_num_last_element(), tapply_num_last_element_r(), tapply_num_last_element_rcpp(), times = 10) 

Unit: milliseconds
                           expr        min         lq       mean     median         uq       max neval
       dplyr_num_last_element()   6.895850   7.088472   8.264270   7.766421   9.089424  11.00775    10
     dplyr_num_last_element_r() 205.375404 214.481520 220.995218 220.107130 225.971179 238.62544    10
  dplyr_num_last_element_rcpp() 211.593443 216.000009 222.247786 221.984289 228.801007 230.50220    10
      tapply_num_last_element()  97.082102  99.528712 101.955668 101.717887 104.370319 109.26982    10
    tapply_num_last_element_r()   6.101055   6.550065   7.386442   7.069754   7.589164   9.98025    10
 tapply_num_last_element_rcpp()  14.173171  15.145711  16.102816  15.400562  16.053229  22.00147    10

我的一般问题是：

1）为什么dplyr_num_last_element_r采用avg 220 ms，而tapply_num_last_element_r需要7 ms。

2）有没有办法编写我自己的最后一个与dplyr一起使用的函数，但它需要7ms的量级吗？

谢谢！

Answer 1

我有一些与你不同的结果。请注意，我更改了last_r以返回最后一个元素并使用dplyr::last（因为还有data.table::last）。

library(dplyr)
library(microbenchmark)
library(Rcpp)
n <- 5000
df <- data.frame(
  grp = factor(rep(1:n, 2)), 
  valn = rnorm(2L*n), 
  stringsAsFactors = FALSE
)

last_r <- function(x) {
  tail(x, 1)
}

cppFunction('double last_rcpp(NumericVector x) {
            int n = x.size();
            return x[n-1];
            }')

dplyr_num_last_element <- function() df %>% group_by(grp) %>% summarise(valn = dplyr::last(valn))
dplyr_num_last_element_r <- function() df %>% group_by(grp) %>% summarise(valn = last_r(valn))
dplyr_num_last_element_rcpp <- function() df %>% group_by(grp) %>% summarise(val = last_rcpp(valn))
tapply_num_last_element <- function() tapply(df$valn, df$grp, FUN = dplyr::last)
tapply_num_last_element_r <- function() tapply(df$valn, df$grp, FUN = last_r)
tapply_num_last_element_rcpp <- function() tapply(df$valn, df$grp, FUN = last_rcpp)


library(data.table) 
dt <- data.table(df)
DT_num_last_element_r <- function() {
  setkey(dt, grp)
  dt[, last_r(valn), grp]
}
microbenchmark(
  DT_num_last_element_r(), 
  dplyr_num_last_element(), 
  dplyr_num_last_element_r(), 
  dplyr_num_last_element_rcpp(), 
  tapply_num_last_element(), 
  tapply_num_last_element_r(), 
  tapply_num_last_element_rcpp(), 
  times = 20
)

基准：

Unit: milliseconds
                           expr        min        lq      mean    median        uq       max neval
        DT_num_last_element_r()  53.956258  55.76482  57.08700  57.33898  58.50556  59.03580    20
       dplyr_num_last_element() 224.289272 228.97531 235.87757 233.73353 237.56040 293.77219    20
     dplyr_num_last_element_r() 178.778382 182.11143 187.40303 184.34760 187.00788 246.64526    20
  dplyr_num_last_element_rcpp() 107.510245 109.64476 111.56974 112.50635 113.63999 114.92428    20
      tapply_num_last_element()  55.999728  58.68948  60.68782  59.78769  63.78408  66.06941    20
    tapply_num_last_element_r()  54.591615  57.31017  58.29962  58.16951  59.98568  63.08996    20
 tapply_num_last_element_rcpp()   9.558151  10.66994  14.76226  11.54004  12.64156  73.87743    20

我的结果更加连贯。你能用这些微小的变化进行测试吗？

这在Windows 10上，R 3.4.0（启用了JIT编译器）。

使用dplyr group_by定制rcpp最后一个函数，并与tapply

1 个答案: