重写一个永远耗时的for循环

Question

我在 R 中编写了一段代码，用于计算某些数据的累积和。有用。问题是，我有 25,000 个数字 X 12 个月需要“融化”，所以我最终有 300,000 行（每个月大约会有 2000x12 多行）。前六行是重新创建我的表格样本（一个巨大的 excel 文件）。然后做了一些魔法将事物转换为正确的格式，最后我有这个双 for 循环，它根据它是否是双“PDRcount”来计算每个月的累积总和。当我在我的真实数据上尝试时，循环需要 6 小时......我怎样才能更快地做到这一点？

library(reshape2)

PDR <- (c( 1,2,3,4,5,2))
START <-  as.Date(c("2008-01-01","2007-01-01","2010-01-01","2011-01-01","2017-02-01","2017-03-01"))
SWITCHOUT <- as.Date(c(NA, "2017-02-28", NA, NA, "2017-03-31",NA))
JAN17 <- (c(100,124,165,178,0,0))
FEB17 <- (c(101,125,133,178,170,0))
MAR17 <- (c(99,0,165,180,166,99))
APR17 <- (c(100,0,156,178,0,78))

alldata <- data.frame(PDR=PDR,
                  START=START,
                  SWITCHOUT=SWITCHOUT,
                  JAN17=JAN17,
                  FEB17=FEB17,
                  MAR17=MAR17,
                  APR17=APR17)

## count PDR occurrences    
alldata$PDRcount <- ave(alldata$PDR,alldata$PDR,FUN=length)
alldata$PDRcount <- as.numeric(alldata$PDRcount)

crossdata<-melt(alldata,id=(c("PDR", "START","SWITCHOUT","PDRcount" )))
colnames(crossdata) <- c("PDR","START","SWITCHOUT","PDRcount","MONTH","SMC")

## transform levels to date format
levels(crossdata$MONTH)[1] <- "2017-01-01"
levels(crossdata$MONTH)[2] <- "2017-02-01"
levels(crossdata$MONTH)[3] <- "2017-03-01"
levels(crossdata$MONTH)[4] <- "2017-04-01"
crossdata$MONTH <- as.Date(crossdata$MONTH,format = "%Y-%m-%d" )


for (pdr in crossdata[,"PDR"]){

maxPDR <- max(crossdata$PDRcount[crossdata$PDR == pdr])
dates <- unique(crossdata$START[crossdata$PDR == pdr])

for (i in 1:maxPDR) {

CumSum <- cumsum( crossdata$SMC[crossdata$PDR == pdr & crossdata$START == dates[i]] )

    crossdata$SMCcum[crossdata$PDR == pdr & crossdata$START == dates[i] & crossdata$MONTH == "2017-01-01"] <- CumSum[1]
    crossdata$SMCcum[crossdata$PDR == pdr & crossdata$START == dates[i] & crossdata$MONTH == "2017-02-01"] <- CumSum[2]
    crossdata$SMCcum[crossdata$PDR == pdr & crossdata$START == dates[i]  & crossdata$MONTH == "2017-03-01"] <- CumSum[3]
    crossdata$SMCcum[crossdata$PDR == pdr & crossdata$START == dates[i] & crossdata$MONTH == "2017-04-01"] <- CumSum[4]     
}
}

已编辑：抱歉出现错误...

Answer 1

您不断地覆盖您的结果。 一个明显的改进是循环 unique(crossdata[,"PDR"]) 而不是为每一行调用循环。

我不确定您的内部循环是否为 maxPDR > 1 提供了预期的结果，您不断地覆盖 START与 maxPDR'th dates 条目匹配的值 - 请注意，您没有对 @987654326 进行排序@ 并因此不能保证 dates[maxPDR] 是最大的（最近的）条目。

我在 dplyr 中编写了一个替代解决方案，其中包含两个步骤来简化转换为所需格式的过程。

alldata <- data.frame(PDR=PDR, START=START, SWITCHOUT=SWITCHOUT, JAN17=JAN17,
                      FEB17=FEB17, MAR17=MAR17, APR17=APR17)

library(dplyr)
library(tidyr) # to reshape the data

crossdata_2 <- alldata %>% 
  gather(MONTH,SMC,ends_with("17")) %>%
  mutate(MONTH = as.character(strptime(paste0(MONTH,"-01"), format = "%b%y-%d"))) %>% 
  # the following line adds your PDRcount but is unnecessary for further computation
  group_by(PDR) %>% mutate(PDRcount = n_distinct(START)) %>% 
  group_by(PDR,START) %>% mutate(SMCcum = cumsum(SMC))

请注意，我为每个PDR 和START 计算cumsum()。如果您只想为每个 PDR 提供一个结果，则只需添加适当的过滤器即可。

我想指出，strptime 中的缩写月份转换 %b 是特定于语言环境的。要正常工作，您可能需要更改LC_TIME。

Answer 2

这是部分答案。我不明白“......基于它是否是双“PDRcount”的部分。

这里是 PDR !=2 使用 dplyr 库的情况的部分答案。我还通过在任何计算之前对交叉数据变量使用 dput 来简化数据输入。

crossdata1<-structure(list(PDR = c(1, 2, 3, 4, 5, 2, 1, 2, 3, 4, 5, 2, 1, 
                                   2, 3, 4, 5, 2, 1, 2, 3, 4, 5, 2),
                           START = structure(c(13879, 13514, 14610, 14975, 17198, 17226, 13879, 13514, 14610, 14975, 
                                    17198, 17226, 13879, 13514, 14610, 14975, 17198, 17226, 13879, 
                                    13514, 14610, 14975, 17198, 17226), class = "Date"), 
                           SWITCHOUT = structure(c(NA, 17225, NA, NA, 17256, NA, NA, 17225, NA, NA, 17256, NA, NA, 17225, 
                                      NA, NA, 17256, NA, NA, 17225, NA, NA, 17256, NA), class = "Date"), 
                           PDRcount = c(1, 2, 1, 1, 1, 2, 1, 2, 1, 1, 1, 2, 1, 2, 1, 1, 1, 2, 1, 2, 1, 1, 1, 2), 
                           MONTH = structure(c(17167, 17167, 
                                     17167, 17167, 17167, 17167, 17198, 17198, 17198, 17198, 17198, 
                                     17198, 17226, 17226, 17226, 17226, 17226, 17226, 17257, 17257, 
                                     17257, 17257, 17257, 17257), class = "Date"), 
                           SMC = c(100, 124, 165, 178, 0, 0, 101, 125, 133, 178, 170, 0, 99, 0, 165, 
                                     180, 166, 99, 100, 0, 156, 178, 0, 78)), 
                      row.names = c(NA,  -24L), .Names = c("PDR", "START", "SWITCHOUT", "PDRcount", "MONTH", "SMC"),
                      class = "data.frame")   

#test to see if starting data is the same
identical(crossdata, crossdata1)
library(dplyr)

#group by and add the cumsum column to answer dataframe
ans<-group_by(crossdata1, PDR) %>%
  mutate(SMCcum = cumsum(SMC))

#rows where the 2 final dataframes do not match
crossdata[-which(crossdata$SMCcum== ans$SMCcum),]

如果应用额外的过滤器来删除“...double "PDRcount" or not”的情况，上述行很可能会起作用。适用。

我发现这篇文章很有帮助：cumsum in grouped data with dplyr

祝你好运。