我有一个数据集,其中包含两组天范围(一天被编码为一年中的天数)。对于每一行,我想计算这些范围总计每月对应的天数。
在我的示例数据中,列“ deb”和“ fin”是每一行中第一个子范围的开始和结束日期,而“ deb2”和“ fin2”是第二个子范围的界限。
d <- data.frame(deb = c(1, 32, 90, 91), fin = c(31, 59, 91, 91),
deb2 = c(50, 0, 0, 0), fin2 = c(60, 0, 0, 0))
d
# deb fin deb2 fin2
#1 1 31 50 60
#2 32 59 0 0
#3 90 91 0 0
#4 91 91 0 0
例如,对于第1行,第一个范围(从'deb'到'fin')从第1天到第31天,第二个范围从第50天到60天。
将两个范围的每月天数相加后,我希望得到类似的东西:
# jan feb mar
#[1,] 31 10 1
#[2,] 0 28 0
#[3,] 0 0 2
#[4,] 0 0 1
(NA而不是零不是问题)
我尝试了以下三个解决方案(第三个解决方案“ g3”是最快的),并且还尝试使用tidyverse,它显示出幅度较慢。我想知道是否有最快的选择,因为在现实生活中,我有无数行。问题似乎出在从参考范围到月参考列表的转换中,也许还在于计数的方式。
f1<-function(deb,fin,deb2,fin2,...) {
f<-factor(c(deb:fin,deb2:fin2))
levels(f)<-list(jan=1:31,feb=32:59,mar=60:91)
table(f)
}
g1 <- function() do.call(rbind,d %>% pmap(f1))
K <- vector(10,mode="character")
K[1:31] <- "jan"; K[32:59] <- "feb"; K[60:91] <- "mar"
f2 <- Vectorize(function(deb,fin,deb2,fin2) table(c(K[deb:fin],K[deb2:fin2])))
g2 <- function() do.call(bind_rows,f2(d$deb,d$fin,d$deb2,d$fin2))
L <- K
names(L) <- 1:91
f3 <- Vectorize(function(deb,fin,deb2,fin2) c(L[deb:fin],L[deb2:fin2]))
g3 <- function() {
as.matrix(do.call(bind_rows,f3(d$deb,d$fin,d$deb2,d$fin2))) -> m
z <- unlist(map(list("jan","feb","mar"),
function(y) apply(m,1,function(x) sum(x==y,na.rm=TRUE))))
dim(z)<-c(nrow(d),3)
z
}
已更新 遵循一些基准。我在试用版中添加了Chinsson12的解决方案,该解决方案在优雅的解决方案中表现良好。
firstOfMths <- seq(as.Date("2018-01-01"), as.Date("2019-01-01"), by="month")
daysPerMth <- c(1L, cumsum(as.integer(diff(firstOfMths))))
chinsoon12 <- function()
t(apply(d, 1, function(x)
table(cut(c(x["deb"]:x["fin"],x["deb2"]:x["fin2"]), daysPerMth, labels=month.abb, include.lowest=TRUE, right=TRUE))
))
N <- 500
d<-data.frame(deb=rep(c(1,32,90,91),N),fin=rep(c(31,59,91,91),N),deb2=rep(c(50,0,0,0),N),fin2=rep(c(60,0,0,0),N))
microbenchmark(g1(),g2(),g3(),chinsoon12())
#Unit: milliseconds
# expr min lq mean median uq max neval
# g1() 571.3890 615.1020 649.7619 639.6632 662.4808 976.9566 100
# g2() 306.7141 341.3056 360.9687 353.1227 373.8194 505.0882 100
# g3() 282.2767 304.4331 320.4908 314.2377 325.8846 543.4680 100
# chinsoon12() 429.7627 469.6998 500.6289 488.5176 512.0520 729.0995 100
答案 0 :(得分:1)
使用findInterval
,Map
和table
:
# create breaks to be used in findInterval
b <- <- as.numeric(format(seq(as.Date("2018-01-01"), as.Date("2018-12-31"), by = "month"), "%j"))
# use Map to expand the day of year ranges by row
# use findInterval to convert day of year to month number
# use the month numbers to index month.abb
l <- Map(function(from, to, from2, to2) month.abb[findInterval(c(from:to, from2:to2), b)], d$deb, d$fin, d$deb2, d$fin2)
# create a row index
i <- rep(1:nrow(d), lengths(l))
# use table to get a contigency table of row indices and months
table(i, factor(unlist(l), levels = month.abb))
# i Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
# 1 31 10 1 0 0 0 0 0 0 0 0 0
# 2 0 28 0 0 0 0 0 0 0 0 0 0
# 3 0 0 1 1 0 0 0 0 0 0 0 0
# 4 0 0 0 1 0 0 0 0 0 0 0 0
在较大的数据集(g3()
)上看起来比d <- d[rep(1:nrow(d), 1e4), ]
更快。
答案 1 :(得分:0)
OP要求找到最快的方法来重新编码一年中某天的范围并计数 ,并且mentioned的生产数据集包括1000万行。 OP进行了基准测试,问题大小为2000行,测试数据仅涵盖3个月而不是12个月。
尽管我想知道这个问题的答案是可以接受的
melt()
,foverlaps()
和dcast()
的{{3}}方法与其他答案的比较library(data.table)
library(magrittr)
cols <- c("deb", "fin")
# reshape sub ranges from wide to long format
long <- melt(setDT(d)[, rn := .I], id.vars = "rn", measure.vars = patterns(cols),
value.name = cols)[deb > 0]
# create data.table with monthly breaks and set keys
b <- seq(as.IDate("2018-01-01"), as.IDate("2019-01-01"), "month")
mDT <- data.table(abb = forcats::fct_inorder(month.abb),
deb = yday(head(b, 12L)),
fin = yday(tail(b, 12L) - 1L),
key = c("deb", "fin"))
# find overlaps between sub ranges and monthly breaks
foverlaps(long, mDT)[
# compute days in overlaps
, days := pmin(fin, i.fin) - pmax(deb, i.deb) + 1L] %>%
# reshape to wide format for final result
dcast(rn ~ abb, sum, value.var = "days", fill = 0L, drop = FALSE)
rn Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec 1: 1 31 10 1 0 0 0 0 0 0 0 0 0 2: 2 0 28 0 0 0 0 0 0 0 0 0 0 3: 3 0 0 1 1 0 0 0 0 0 0 0 0 4: 4 0 0 0 1 0 0 0 0 0 0 0 0
下面的基准代码进行比较
foverlaps()
方法,g1
,g2
和g3
(已修改为可处理12个月), chinsoon12's now deleted answer未被考虑,因为OP AntoniosK's tidyverse
answer (now deleted)认为这种方法比他自己的g2
方法慢30倍。
对于基准测试,问题大小有所不同(n_rows
= 100、1000、10000)以及具有两个范围的行所占的比例(p_sr_2 = 10%,50%,100%)。已针对某些方法的长期运行选择了最大的问题大小。
不同的方法在结果的类别(即data.table
,matrix
,tibble
和table
)和列数方面也不同。因此,重点放在编写适当的检查功能上。
library(bench)
library(data.table)
library(magrittr)
library(ggplot2)
bm2 <- press(
p_sr_2 = c(0.1, 0.5, 1), # share or rows with 2nd sub range ]0, 1]
n_rows = 10^seq(2, 4, 1),
{ # create test data
set.seed(1L)
d0 <- t(replicate(n_rows, sample(365L, 4L) %>% sort())) %>% data.table()
setnames(d0, c("deb", "fin", "deb2", "fin2"))
idx <- sample(nrow(d0), (1 - p_sr_2) * nrow(d0))
d0[idx, c("deb2", "fin2") := 0L]
str(d0)
mark(
foverlaps = {
d <- copy(d0)
cols <- c("deb", "fin")
long <- melt(setDT(d)[, rn := .I], id.vars = "rn", measure.vars = patterns(cols),
value.name = cols)[deb > 0]
b <- seq(as.IDate("2018-01-01"), as.IDate("2019-01-01"), "month")
mDT <- data.table(abb = forcats::fct_inorder(month.abb),
deb = yday(head(b, 12L)),
fin = yday(tail(b, 12L) - 1L),
key = c("deb", "fin"))
foverlaps(long, mDT)[, days := pmin(fin, i.fin) - pmax(deb, i.deb) + 1L] %>%
dcast(rn ~ abb, sum, value.var = "days", fill = 0L, drop = FALSE)
# returns a data.table with 13 columns and 0 for missing values
},
g1 = {
f1 <- function(deb, fin, deb2, fin2, ...) {
f <- factor(c(deb:fin, deb2:fin2))
levels(f) <- list(jan = 1:31,
feb = 32:59,
mar = 60:90,
apr = 91:120,
may = 121:151,
jun = 152:181,
jul = 182:212,
aug = 213:243,
sep = 244:273,
oct = 274:304,
nov = 305:334,
dec = 335:365)
table(f)
}
do.call(rbind, d %>% purrr::pmap(f1))
# returns a matrix with 12 named columns and 0 for missing values
},
g2 = {
K <- vector(10, mode = "character")
K[1:31] <- "jan"
K[32:59] <- "feb"
K[60:90] <- "mar"
K[91:120] <- "apr"
K[121:151] <- "may"
K[152:181] <- "jun"
K[182:212] <- "jul"
K[213:243] <- "aug"
K[244:273] <- "sep"
K[274:304] <- "oct"
K[305:334] <- "nov"
K[335:365] <- "dec"
f2 <-
Vectorize(function(deb, fin, deb2, fin2)
table(c(K[deb:fin], K[deb2:fin2])))
template <- matrix(
integer(0), ncol = 12L,
dimnames = list(NULL, c("jan", "feb", "mar", "apr", "may", "jun",
"jul", "aug", "sep", "oct", "nov", "dec"))) %>%
tibble::as.tibble()
do.call(dplyr::bind_rows, c(list(template), f2(d$deb, d$fin, d$deb2, d$fin2)))
# returns a tibble with 12 columns and NA for missing values
},
g3 = {
K <- vector(10, mode = "character")
K[1:31] <- "jan"
K[32:59] <- "feb"
K[60:90] <- "mar"
K[91:120] <- "apr"
K[121:151] <- "may"
K[152:181] <- "jun"
K[182:212] <- "jul"
K[213:243] <- "aug"
K[244:273] <- "sep"
K[274:304] <- "oct"
K[305:334] <- "nov"
K[335:365] <- "dec"
names(K) <- 1:365
f3 <-
Vectorize(function(deb, fin, deb2, fin2)
c(K[deb:fin], K[deb2:fin2]))
m <- as.matrix(do.call(dplyr::bind_rows, f3(d$deb, d$fin, d$deb2, d$fin2)))
z <- unlist(purrr::map(list("jan", "feb", "mar", "apr", "may", "jun",
"jul", "aug", "sep", "oct", "nov", "dec"),
function(y)
apply(m, 1, function(x)
sum(x == y, na.rm = TRUE))))
dim(z) <- c(nrow(d), 12)
z
# returns a matrix with 12 columns and 0 for missing values
},
henrik = {
d <- copy(d0)
b <- as.numeric(format(seq(as.Date("2018-01-01"), as.Date("2018-12-31"),
by = "month"), "%j"))
l <- Map(
function(from, to, from2, to2) month.abb[findInterval(c(from:to, from2:to2), b)],
d$deb, d$fin, d$deb2, d$fin2)
i <- rep(1:nrow(d), lengths(l))
table(i, factor(unlist(l), levels = month.abb))
# returns an object of class table with 12 columns and 0 for missing values
},
chinsoon12 = {
d <- copy(d0)
firstOfMths <- seq(as.Date("2018-01-01"), as.Date("2019-01-01"), by="month")
daysPerMth <- c(1L, cumsum(as.integer(diff(firstOfMths))))
g <- ceiling(seq(1, ncol(d)) / 2)
t(apply(d, 1, function(x) {
x <- unlist(by(x, g, function(k) seq(k[1L], k[2L])), use.names=FALSE)
table(cut(x, daysPerMth, labels=month.abb, include.lowest=TRUE, right=TRUE))
}))
# returns a matrix with 12 named columns and 0 for missing values
},
check = function(x, y) {
cat("Run check: ")
xdt <- as.data.table(x) %>% .[, .SD, .SDcols = tail(seq_len(ncol(.)), 12L)]
if (tibble::is_tibble(y)) {
y <- dplyr::mutate_all(y, function(x) dplyr::if_else(is.na(x), 0L, x))
}
if (is.table(y)) y <- matrix(y, ncol = 12L)
ydt <- as.data.table(y) %>% .[, .SD, .SDcols = tail(seq_len(ncol(.)), 12L)]
result <- all.equal(xdt, ydt, check.attributes = FALSE)
if (!isTRUE(result)) {
print(result)
} else cat("OK\n")
return(result)
}
)
}
)
可以绘制基准测试的时间:
library(ggplot2)
autoplot(bm)
请注意对数时间刻度。
显然,与行数相比,具有两个范围的行份额对性能没有明显影响。对于较小的问题,henrik
的方法是确认OP观察结果的最快方法。但是,对于1000行及更多行的问题大小,foverlaps
方法要快得多。对于1万行,foverlaps
比其他方法快一到两个数量级。
此外,内存要求也有很大差异:
bm %>%
tidyr::unnest() %>%
ggplot(aes(expression, mem_alloc, color = gc)) +
ggbeeswarm::geom_quasirandom() +
coord_flip() +
facet_grid(p_sr_2 ~ n_rows, labeller = label_both)
再次,请注意对数刻度。
foverlaps
方法分配的内存比其他方法少一到两个数量级。
由于运行时间长(以及我的不耐烦),因此仅对最多1万行进行了高于基准测试。 也仅测试了40 k行。因此,我想知道foverlaps
方法是否能够在合理的时间内处理1000万行(OP生产数据集的大小)。
不幸的是,对于1M行或更大的问题,我创建输入数据的代码太慢了。因此,我不得不分别优化(和基准测试)这部分。
只有foverlaps
方法的基准是OP规定的固定为25%的第二范围的份额。
library(bench)
library(data.table)
library(magrittr)
library(ggplot2)
bm5 <- press(
n_rows = 10^seq(2, 7, 1),
{ # create test data
cat("Start to create test data", format(Sys.time()), "\n")
p_sr_2 <- 0.25 # share or rows with 2nd sub range ]0, 1]
set.seed(1L)
long <- data.table(rn = rep(seq_len(n_rows), each = 4L),
day = sample(365L, 4L * n_rows, replace = TRUE))
setorder(long, rn, day)
dups <- long[, which(anyDuplicated(day) > 0), by = rn]$rn
if (length(dups) > 0) long[
rn %in% dups,
day := replicate(length(dups), sample(365L, 4L) %>% sort(), simplify = FALSE) %>% unlist()]
d0 <- dcast(long, rn ~ rowid(rn), value.var = "day")[, rn := NULL]
setnames(d0, c("deb", "fin", "deb2", "fin2"))
idx <- sample(nrow(d0), (1 - p_sr_2) * nrow(d0))
d0[idx, c("deb2", "fin2") := 0L]
str(d0)
rm(long) # free the memory
tables()
cat("Test data created", format(Sys.time()), "\n")
mark(
foverlaps = {
d <- copy(d0)
cols <- c("deb", "fin")
long <- melt(setDT(d)[, rn := .I], id.vars = "rn", measure.vars = patterns(cols),
value.name = cols)[deb > 0]
b <- seq(as.IDate("2018-01-01"), as.IDate("2019-01-01"), "month")
mDT <- data.table(abb = forcats::fct_inorder(month.abb),
deb = yday(head(b, 12L)),
fin = yday(tail(b, 12L) - 1L),
key = c("deb", "fin"))
foverlaps(long, mDT)[, days := pmin(fin, i.fin) - pmax(deb, i.deb) + 1L] %>%
dcast(rn ~ abb, sum, value.var = "days", fill = 0L, drop = FALSE)
# returns a data.table with 13 columns and 0 for missing values
},
min_time = 2
)
}
)
在我的系统上,1000万行案例的运行时间约为23.7秒。对于1000行以上的问题,运行时间几乎呈线性增长。 1000万行的轻微向上弯曲可能是由于我的系统上的内存限制所致。
bm4 %>%
tidyr::unnest() %>%
ggplot(aes(n_rows, time, colour = expression)) +
geom_smooth(data = . %>% dplyr::filter(n_rows > 10^3),
method = "lm", se = FALSE, colour = "blue", size = 1) +
geom_point(size = 2) +
scale_x_log10() +
stat_summary(fun.y = median, geom = "line", size = 1) +
ggtitle("time vs n_rows")
请注意双对数刻度。