R：設置初始條件的for循環的dplyr解決方案

Question

我有一個數據，一年有40天和一些數據

set.seed(123)
df <- data.frame(day = 1:40,rain = runif(40,min = 0, max = 3), petc = runif(40, min = 0.3, max = 8),swc = runif(40, min = 27.01, max = 117.43))

我想計算每天另一個名為aetc的變量，計算方法如下：

SW.ini <- 2 # setting some initial values 
SW.max <- 5
SW.min <- 0

第1天，

1）確定一個名為PAW(day1) = SW.ini + rain(day1)的變量PAW(day1) = SW.ini + rain(day1)

2）如果PAW(day1) >= SWC(day1), aetc(day1) = petc(day1) ;

If `PAW(day1) < SWC(day1), aetc(day1) = PAW(day1)/SWC(day1) * petc(day1)`

3）檢查aetc(day1) > PAW(day1). If yes, aetc(day1) = paw(day1) aetc(day1) > PAW(day1). If yes, aetc(day1) = paw(day1)

4）更新SW(day1) = SW.ini + rain(day1) - aetc(day1)

5）如果SW(day1) > SW.max, SW(day1) = SW.max. Similarly if SW(day1) > SW.max, SW(day1) = SW.max. Similarly if SW（day1）<SW.min，SW（day1）= SW.min`

重復第2天

1）確定PAW(day2) = SW(day1) + rain(day2)
2）如果PAW(day2) >= SWC(day2), aetc(day2) = petc(day2) 2天） PAW(day2) >= SWC(day2), aetc(day2) = petc(day2) ; 如果PAW(day2) < SWC(day2), aetc(day2) = PAW(day2)/SWC(day2) * petc(day2)

3）檢查aetc(day2) > PAW(day2) 。 如果是， aetc(day2) = paw(day2)

4）更新SW(day2) = SW(day1) + rain(day2) - aetc(day2)

5）如果SW(day2) > SW.max, SW(day2) = SW.max. Similarly if SW(day2) > SW.max, SW(day2) = SW.max. Similarly if SW（第2天）<SW.min，SW（第2天）= SW.min`

這是我優雅的for循環來做到這一點：

      df$PAW <- NA
      df$aetc <- NA
      df$SW <- NA

      df$PAW[1] <- SW.ini + df$rain[1]

      df$aetc[1] <- ifelse(df$PAW[1] >= df$swc[1], df$petc[1],(df$PAW[1]/df$swc[1])*df$petc[1])
      df$aetc[1] <- ifelse(df$aetc[1] > df$PAW[1], df$PAW[1], df$aetc[1])
      df$SW[1] <- SW.ini + df$rain[1] -  df$aetc[1]
      df$SW[1] <- ifelse(df$SW[1] > SW.max, SW.max, ifelse(df$SW[1] < 0, 0,df$SW[1]))

      for (day in 2:nrow(df)){

        df$PAW[day] <- df$SW[day - 1] + df$rain[day]
        df$aetc[day] <- ifelse(df$PAW[day] >= df$swc[day], df$petc[day], (df$PAW[day]/df$swc[day]) * df$petc[day])
        df$aetc[day] <- ifelse(df$aetc[day] > df$PAW[day], df$PAW[day],df$aetc[day])
        df$SW[day] <- df$SW[day - 1] + df$rain[day] -  df$aetc[day]
        df$SW[day] <- ifelse(df$SW[day] > SW.max,SW.max, ifelse(df$SW[day] < 0, 0,df$SW[day]))
      }

我的問題是，這只是一年的數據，我想運行它多年。

      set.seed(123)
      df <- data.frame(year = 1980:2015, day = rep(1:40, each = 36),rain = 
      runif(40*36,min = 0, max = 3), petc = runif(40*36, min = 0.3, max = 8),swc = runif(40*36, min = 27.01, max = 117.43))

所以我想做點什么

                df %>% group_by(year) # and then run the above function for each year. 

是否有dplyr或任何其他解決方案？

謝謝

Answer 1

注意：我最初在你的后續問題上發布了這個答案， R：for fore循環中的循環 ，但在看到這個之后，似乎這個答案在這里更具相關性。 （我沒有解決與我的答案中的並行化相關的任何問題，這是您跟進的主題） 。

使用Rcpp和data.table

使用C ++編譯邏輯並使用data.table分組操作按組應用它可以使您的基線速度提高約2,000倍，遠遠超過您希望通過並行化獲得的速度。

在原始示例中，它有39,420,000行 ，這在我的機器上以1.883秒執行; 在修改后的28,800行中 ，這將在0.004秒內執行

library(data.table)
library(Rcpp)

在R腳本中內聯定義並編譯C++函數CalcSW() ：

一個注意事項： C / C++計數從0開始，與R不同，從1開始 - 這就是為什么這里的指數不同

Rcpp::cppFunction('
List CalcSW(NumericVector SW_ini,
            NumericVector SW_max,
            NumericVector rain,
            NumericVector swc,
            NumericVector PETc) {

  int n = SW_ini.length();
  NumericVector SW(n);
  NumericVector PAW(n);
  NumericVector aetc(n);

  double SW_ini_glob = SW_ini[0];
  double SW_max_glob = SW_max[0];

  SW[0] = SW_ini_glob;
  PAW[0] = SW[0] + rain[0];

  if (PAW[0] > swc[0]){
    aetc[0] = PETc[0];
  } else {
    aetc[0] = PAW[0]/swc[0]*PETc[0];
  }

  if (aetc[0] > PAW[0]){
    aetc[0] = PAW[0];
  }

  SW[0] = SW[0] + rain[0] - aetc[0];

  if(SW[0] > SW_max_glob){
    SW[0] = SW_max_glob;
  }

  if(SW[0] < 0){
    SW[0] = 0;
  }

  for (int i = 1; i < n; i++) {

    PAW[i] = SW[i-1] + rain[0];

    if (PAW[i] > swc[i]){
      aetc[i] = PETc[i];
    } else {
      aetc[i] = PAW[i]/swc[i]*PETc[i];
    }

    if (aetc[i] > PAW[i]){
      aetc[i] = PAW[i];
    }

    SW[i] = SW[i-1] + rain[i] - aetc[i];

    if(SW[i] > SW_max_glob){
      SW[i] = SW_max_glob;
    }

    if(SW[i] < 0){
     SW[i] = 0;
    }
  }
  return Rcpp::List::create(Rcpp::Named("SW") = SW,
                            Rcpp::Named("PAW") = PAW,
                            Rcpp::Named("aetc") = aetc);
}')

創建data.table

df <- data.table(loc.id = rep(1:10, each = 80*36), 
                 year = rep(rep(1980:2015, each = 80), times = 10),
                 day = rep(rep(1:80, times = 36),times = 10),
                 rain = runif(10*36*80, min = 0 , max = 5),
                 swc = runif(10*36*80,min = 0, max = 50),
                 SW_max = rep(runif(10, min = 100, max = 200), each = 80*36),
                 SW_ini = runif(10*36*80),
                 PETc = runif(10*36*80, min = 0 , max = 1.3),
                 SW = as.numeric(NA),
                 PAW = as.numeric(NA), 
                 aetc = as.numeric(NA))

setkey(df, loc.id, year, day)

對於loc.id和year每個組合，在df上執行函數CalcSW() ，同時將返回值分配給三列：

system.time({
  df[,  c("SW","PAW","aetc") := CalcSW(SW_ini,
                                       SW_max,
                                       rain,
                                       swc,
                                       PETc), keyby = .(loc.id, year)]
})

...

   user  system elapsed 
  0.004   0.000   0.004 

結果：

head(df)

...

   loc.id year day       rain       swc   SW_max     SW_ini      PETc       SW      PAW       aetc
1:      1 1980   1 0.35813251 28.360715 177.3943 0.69116310 0.2870478 1.038675 1.049296 0.01062025
2:      1 1980   2 1.10331116 37.013022 177.3943 0.02742273 0.4412420 2.125335 1.396808 0.01665171
3:      1 1980   3 1.76680011 32.509970 177.3943 0.66273062 1.1071233 3.807561 2.483467 0.08457420
4:      1 1980   4 3.20966558  8.252797 177.3943 0.12220454 0.3496968 6.840713 4.165693 0.17651342
5:      1 1980   5 1.32498191 14.784203 177.3943 0.66381497 1.2168838 7.573160 7.198845 0.59253503
6:      1 1980   6 0.02547458 47.903637 177.3943 0.21871598 1.0864713 7.418750 7.931292 0.17988449

我並非100％肯定我完全實現了你的邏輯，但是在調整我可能錯過的東西的邏輯應該非常簡單，我以非常類似的方式實現它。

---

另一個注意事項：如果你創建一個單獨的文件，如下面的TestCode.cpp那樣命名，那么用自動縮進和代碼突出顯示（無論你是使用RStudio還是Emacs）來編寫C++更容易。

然后，你可以使用Rcpp::sourceCpp("TestCode.cpp")在你的R腳本中編譯你的函數，或者你可以將除了前三行之外的所有內容復制並粘貼為字符串作為Rcpp::cppFunction()的參數Rcpp::cppFunction()就像我上面做的那樣。

#include <Rcpp.h>
using namespace Rcpp;

// [[Rcpp::export]]
List CalcSW(NumericVector SW_ini,
                     NumericVector SW_max,
                     NumericVector rain,
                     NumericVector swc,
                     NumericVector PETc) {

  int n = SW_ini.length();
  NumericVector SW(n);
  NumericVector PAW(n);
  NumericVector aetc(n);

  double SW_ini_glob = SW_ini[0];
  double SW_max_glob = SW_max[0];

  SW[0] = SW_ini_glob;
  PAW[0] = SW[0] + rain[0];

  if (PAW[0] > swc[0]){
    aetc[0] = PETc[0];
  } else {
    aetc[0] = PAW[0]/swc[0]*PETc[0];
  }

  if (aetc[0] > PAW[0]){
    aetc[0] = PAW[0];
  }

  SW[0] = SW[0] + rain[0] - aetc[0];

  if(SW[0] > SW_max_glob){
    SW[0] = SW_max_glob;
  }

  if(SW[0] < 0){
    SW[0] = 0;
  }

  for (int i = 1; i < n; i++) {

    PAW[i] = SW[i-1] + rain[0];

    if (PAW[i] > swc[i]){
      aetc[i] = PETc[i];
    } else {
      aetc[i] = PAW[i]/swc[i]*PETc[i];
    }

    if (aetc[i] > PAW[i]){
      aetc[i] = PAW[i];
    }

    SW[i] = SW[i-1] + rain[i] - aetc[i];

    if(SW[i] > SW_max_glob){
      SW[i] = SW_max_glob;
    }

    if(SW[i] < 0){
      SW[i] = 0;
    }
  }
  return Rcpp::List::create(Rcpp::Named("SW") = SW,
                            Rcpp::Named("PAW") = PAW,
                            Rcpp::Named("aetc") = aetc);
}

Answer 2

您可以將代碼包裝在另一個for循環中，並將每年df保存在列表中：

library(tidyverse)
lst <- vector("list", length(unique(df$year)))
for (i in seq_along(unique(df$year))) {
    df_year <- df %>% filter(year == unique(df$year)[[i]])

    # rest of code with df_year replacing df

    lst[[i]] <- df_year
}
final_df <- bind_rows(lst)

Answer 3

來自Matt的data.table插圖很好地說明了data.table速度有多快，因為它可以在沒有副本和移動數據的情況下進行計算。

但是，要回答關於使用管道的問題的關鍵，你可以使用group_by和do來完成你的目標（盡管比data.table慢得多）

下面我設置了Matt所做的相同虛擬數據。 然后我使用你的功能（但在PETc固定的情況下）。 它並不快，但它很容易遵循。

df <- data.frame(loc.id = rep(1:10, each = 80*36), 
                 year = rep(rep(1980:2015, each = 80), times = 10),
                 day = rep(rep(1:80, times = 36),times = 10),
                 rain = runif(10*36*80, min = 0 , max = 5),
                 swc = runif(10*36*80,min = 0, max = 50),
                 SW_max = rep(runif(10, min = 100, max = 200), each = 80*36),
                 SW_ini = runif(10*36*80),
                 PETc = runif(10*36*80, min = 0 , max = 1.3) 
                 )

my_fun <- function(df){
  SW.ini <- 2 # setting some initial values 
  SW.max <- 5
  SW.min <- 0

  df$PAW <- NA
  df$aetc <- NA
  df$SW <- NA

  df$PAW[1] <- SW.ini + df$rain[1]

  df$aetc[1] <- ifelse(df$PAW[1] >= df$swc[1], df$PETc[1],(df$PAW[1]/df$swc[1])*df$PETc[1])
  df$aetc[1] <- ifelse(df$aetc[1] > df$PAW[1], df$PAW[1], df$aetc[1])
  df$SW[1] <- SW.ini + df$rain[1] -  df$aetc[1]
  df$SW[1] <- ifelse(df$SW[1] > SW.max, SW.max, ifelse(df$SW[1] < 0, 0,df$SW[1]))

  for (day in 2:nrow(df)){

    df$PAW[day] <- df$SW[day - 1] + df$rain[day]
    df$aetc[day] <- ifelse(df$PAW[day] >= df$swc[day], df$PETc[day], (df$PAW[day]/df$swc[day]) * df$PETc[day])
    df$aetc[day] <- ifelse(df$aetc[day] > df$PAW[day], df$PAW[day],df$aetc[day])
    df$SW[day] <- df$SW[day - 1] + df$rain[day] -  df$aetc[day]
    df$SW[day] <- ifelse(df$SW[day] > SW.max,SW.max, ifelse(df$SW[day] < 0, 0,df$SW[day]))
  }
  return(df)
}


library(tictoc)
library(tidyverse)


tic()
df  %>% 
  group_by(year) %>%
  do(my_fun(.)) -> 
  out
toc()
#> 5.075 sec elapsed