一种“非 tidyverse”方式:
data.frame(
x = c('a', 'b', 'c'),
y = c(4, 5, 6),
from = c(1, 2, 3),
to = c(2, 4, 6),
stringsAsFactors = FALSE
) -> xdf
do.call(rbind.data.frame, lapply(1:nrow(xdf), function(i) {
data.frame(x = xdf$x[i], y=xdf$y[i], z=xdf$from[i]:xdf$to[i], stringsAsFactors=FALSE)
}))
一种“tidyverse”方式:
library(tidyverse)
data_frame(
x = c('a', 'b', 'c'),
y = c(4, 5, 6),
from = c(1, 2, 3),
to = c(2, 4, 6)
) -> xdf
rowwise(xdf) %>%
do(data_frame(x = .$x, y=.$y, z=.$from:.$to))
另一种“tidyverse”方式not已进行以下基准测试:
xdf %>%
rowwise() %>%
do( merge( as_tibble(.), tibble(z=.$from:.$to), by=NULL) ) %>%
select( -from, -to ) # Omit this line if you want to keep all original columns.
既然你问了 abt 性能:
library(microbenchmark)
data.table::data.table(
x = c('a','b','c'),
y = c(4,5,6),
from = c(1,2,3),
to = c(2,4,6)
) -> xdt1
data.frame(
x = c('a', 'b', 'c'),
y = c(4, 5, 6),
from = c(1, 2, 3),
to = c(2, 4, 6),
stringsAsFactors = FALSE
) -> xdf1
data.table
操作经常就地修改,因此保持公平竞争环境并在执行操作之前复制每个数据帧/表。
那个时间惩罚是~100纳秒在大多数现代系统上。
microbenchmark(
data.table = {
xdt2 <- xdt1
xdt2[, diff:= (to - from) + 1]
xdt2 <- xdt2[rep(1:.N, diff)]
xdt2[,z := seq(from,to), by=.(x,y,from,to)]
xdt2[,c("x", "y", "z")]
},
base = {
xdf2 <- xdf1
do.call(rbind.data.frame, lapply(1:nrow(xdf2), function(i) {
data.frame(x = xdf2$x[i], y=xdf2$y[i], z=xdf2$from[i]:xdf2$to[i], stringsAsFactors=FALSE)
}))
},
tidyverse = {
xdf2 <- xdf1
dplyr::rowwise(xdf2) %>%
dplyr::do(dplyr::data_frame(x = .$x, y=.$y, z=.$from:.$to))
},
plyr = {
xdf2 <- xdf1
plyr::mdply(xdf2, function(x,y,from,to) data.frame(x,y,z=seq(from,to)))[c("x","y","z")]
},
times = 1000
)
## Unit: microseconds
## expr min lq mean median uq max neval
## data.table 920.361 1072.9265 1257.2321 1178.832 1280.2660 10628.552 1000
## base 677.069 761.3145 884.4136 825.472 915.8985 5366.515 1000
## tidyverse 15926.127 17231.5015 19201.4798 17994.919 20014.4140 166901.570 1000
## plyr 1938.838 2196.4205 2448.5314 2322.949 2501.5075 5735.255 1000