Data.table 元编程 [英] Data.table meta-programming

问题描述

我认为元编程在这里是正确的术语.

I think meta-programming is the right term here.

我希望能够像在 webapp 中使用 MySQL 一样使用 data.table.也就是说，Web 用户使用一些 Web 前端(例如 Shiny 服务器)来选择数据库、选择要过滤的列、选择要分组的列、选择要聚合的列和聚合功能.我想使用 R 和 data.table 作为后端进行查询、聚合等.假设前端存在并且 R 将这些变量作为字符串并且它们已经过验证等.

I want to be able to use data.table much like one would use MySQL in say a webapp. That is, web users use some web front-end (like Shiny server for example) to select a data-base, select columns to filter on, select columns to group-by, select columns to aggregate and aggregation functions. I want to use R and data.table as a backend for querying, aggregation etc. Assume that front end exists and R has these variables as character strings and they are validated etc.

我编写了以下函数来构建 data.table 表达式并使用 R 的 parse/eval 元编程功能来运行它.这是一个合理的方法吗?

I wrote the following function to build the data.table expression and use the parse/eval meta-programming functionality of R to run it. Is this a reasonable way to do this?

我包含所有相关代码来测试这一点.获取此代码(为安全起见阅读后！)和运行 test_agg_meta() 来测试它.这只是一个开始.我可以添加更多功能.

I includes all relevant code to test this. Source this code (after reading it for security!) and run test_agg_meta() to test it. It is just a start. I could add more functionality.

但我的主要问题是我是否严重过度考虑了这一点.当所有输入都事先未确定而不诉诸解析/评估元编程时，是否有更直接的方法来使用 data.table ?

But my main question is whether I am grossly over-thinking this. Is there is a more direct way to use data.table when all of the inputs are undetermined before hand without resorting to parse/eval meta-programming?

我也知道with"语句和其他一些无糖函数方法，但不知道它们是否可以处理所有情况.

I am also aware of the "with" statement and some of the other sugarless-functional methods but don't know if they can take care of all cases.

require(data.table)

fake_data<-function(num=12){
  #make some fake data
  x=1:num
  lets=letters[1:num]
  data=data.table(
    u=rep(c("A","B","C"),floor(num/3)),
    v=x %%2, w=lets, x=x, y=x^2, z=1-x)
  return(data)
}

data_table_meta<-function(
  #aggregate a data.table meta-programmatically
  data_in=fake_data(),
  filter_cols=NULL,
  filter_min=NULL,
  filter_max=NULL,
  groupby_cols=NULL,
  agg_cols=setdiff(names(data_in),groupby_cols),
  agg_funcs=NULL,
  verbose=F,
  validate=T,
  jsep="_"
){

  all_cols=names(data_in)

  if (validate) {
    stopifnot(length(filter_cols) == length(filter_min))
    stopifnot(length(filter_cols) == length(filter_max))
    stopifnot(filter_cols %in% all_cols)
    stopifnot(groupby_cols %in% all_cols)
    stopifnot(length(intersect(agg_cols,groupby_cols)) == 0)
    stopifnot((length(agg_cols) == length(agg_funcs))  | (length(agg_funcs)==1) | (length(agg_funcs)==0))
  }

  #build the command

  #defaults
  i_filter=""
  j_select=""
  n_agg_funcs=length(agg_funcs)
  n_agg_cols=length(agg_cols)
  n_groupby_cols=length(groupby_cols)
  if (n_agg_funcs == 0) {
    #NULL
    print("NULL")
    j_select=paste(agg_cols,collapse=",")
    j_select=paste("list(",j_select,")")
  } else {
    agg_names=paste(agg_funcs,agg_cols,sep=jsep)
    jsels=paste(agg_names,"=",agg_funcs,"(",agg_cols,")",sep="")
    if (n_groupby_cols>0) jsels=c(jsels,"N_Rows_Aggregated=.N")
    j_select=paste(jsels,collapse=",")
    j_select=paste("list(",j_select,")")
  }

  groupby=""

  if (n_groupby_cols>0) {
    groupby=paste(groupby_cols,collapse=",")
    groupby=paste("by=list(",groupby,")",sep="")
  }

  n_filter_cols=length(filter_cols)
  if (n_filter_cols > 0) {
    i_filters=rep("",n_filter_cols)
    for (i in 1:n_filter_cols) {
      i_filters[i]=paste(" (",filter_cols[i]," >= ",filter_min[i]," & ",filter_cols[i]," <= ",filter_max[i],") ",sep="")
    }
    i_filter=paste(i_filters,collapse="&")
  }

  command=paste("data_in[",i_filter,",",j_select,",",groupby,"]",sep="")

  if (verbose == 2) {
    print("all_cols:")
    print(all_cols)
    print("filter_cols:")
    print(filter_cols)
    print("agg_cols:")
    print(agg_cols)
    print("filter_min:")
    print(filter_min)
    print("filter_max:")
    print(filter_max)
    print("groupby_cols:")
    print(groupby_cols)
    print("agg_cols:")
    print(agg_cols)
    print("agg_funcs:")
    print(agg_funcs)
    print("i_filter")
    print(i_filter)
    print("j_select")
    print(j_select)
    print("groupby")
    print(groupby)
    print("command")
    print(command)
  }
  print(paste("evaluating command:",command))
  eval(parse(text=command))
}

my_agg<-function(data=fake_data()){
  data_out=data[
    i=x<=5,
    j=list(
      mean_x=mean(x),
      mean_y=mean(y),
      sum_z=sum(z),
      N_Rows_Aggregated=.N
    ),
    by=list(u,v)]
  return(data_out)
}

my_agg_meta<-function(data=fake_data()){
  #should give same results as my_agg
  data_out=data_table_meta(data,
      filter_cols=c("x"),
      filter_min=c(-10000),
      filter_max=c(5),
      groupby_cols=c("u","v"),
      agg_cols=c("x","y","z"),
      agg_funcs=c("mean","mean","sum"),
      verbose=T,
      validate=T,
      jsep="_")
  return(data_out)
}

test_agg_meta<-function(){
  stopifnot(all(my_agg()==my_agg_meta()))
  print("Congrats, you passed the test")
}

推荐答案

虽然您的函数看起来确实很有趣，但我相信您是在问是否有其他方法可以实现它.
就个人而言，我喜欢使用这样的东西:

While your functions certainly look interesting, I believe you are asking if there are other ways to go about it.
Personally, I like to use something like this:

## SAMPLE DATA
DT1 <- data.table(id=sample(LETTERS[1:4], 20, TRUE), Col1=1:20, Col2=rnorm(20))
DT2 <- data.table(id=sample(LETTERS[3:8], 20, TRUE), Col1=sample(100:500, 20), Col2=rnorm(20))
DT3 <- data.table(id=sample(LETTERS[19:20], 20, TRUE), Col1=sample(100:500, 20), Col2=rnorm(20))

通过引用表名访问表:

这很简单，很像 R

# use strings to select the table
tablesSelected <- "DT3"

# use get to access them 
get(tablesSelected)

# and we can perform operations:
get(tablesSelected)[, list(C1mean=mean(Col1), C2mean=mean(Col2))]

通过引用选择列

要通过引用名称来选择列，请使用 .SDcols 参数.给定一个列名向量:

SELECTING COLUMNS BY REFERENCE

To select columns by reference to their names, use the .SDcols argument. Given a vector of column names:

columnsSelected <- c("Col1", "Col2")

将该向量分配给 .SDcols 参数:

Assign that vector to the .SDcols argument:

## Here we are simply accessing those columns
DT3[, .SD, .SDcols = columnsSelected]

我们还可以对字符串向量中命名的每一列应用一个函数:

We can also apply a function to each column named in the string vector:

## apply a function to each column
DT3[, lapply(.SD, mean), .SDcols = columnsSelected]

请注意，如果我们的目标只是输出列，我们可以关闭 with:

Note that if our goal is simply to output the columns we can turn off with:

# This works for displaying
DT3[, columnsSelected, with=FALSE]

注意:更现代"的做法是使用 .. 快捷方式从上一级"访问 columnsSelected:

Note: a more "modern" way of doing this is to use the .. shortcut to access columnsSelected from "up one level":

DT3[ , ..columnsSelected]

但是，如果使用 with=FALSE，我们就不能以通常的方式直接对列进行操作

However, if using with=FALSE, we cannot then operate directly on the columns in the usual fashion

## This does NOT work: 
DT3[, someFunc(columnsSelected), with=FALSE]

## This DOES work: 
DT3[, someFunc(.SD), .SDcols=columnsSelected]

## This also works, but is less ideal, ie assigning to new columns is more cumbersome
DT3[, columnsSelected, with=FALSE][, someFunc(.SD)]

我们也可以使用get，但它有点棘手.我把它留在这里供参考，但 .SDcols 是要走的路

We can also use get, but it is a bit trickier. I am leaving it here for reference, but .SDcols is the way to go

## we need to use `get`, but inside `j`
##   AND IN A WRAPPER FUNCTION     <~~~~~ THIS IS VITAL

DT3[, lapply(columnsSelected, function(.col) get(.col))]

## We can execute functions on the columns:
DT3[, lapply(columnsSelected, function(.col) mean( get(.col) ))]


## And of course, we can use more involved-functions, much like any *ply call:
# using .SDcols 
DT3[, lapply(.SD, function(.col) c(mean(.col) + 2*sd(.col), mean(.col) - 2*sd(.col))), .SDcols = columnsSelected]

# using `get` and assigning the value to a var.  
#   Note that this method has memory drawbacks, so using .SDcols is preferred
DT3[, lapply(columnsSelected, function(.col) {TheCol <- get(.col); c(mean(TheCol) + 2*sd(TheCol), mean(TheCol) - 2*sd(TheCol))})]

作为参考，如果您尝试以下操作，您会发现它们不会产生我们所追求的结果.

For reference, if you try the following, you will notice that they do not produce the results we are after.

    ## this DOES NOT work (need ..columnsSelected)
    DT3[, columnsSelected]

    ## netiher does this
    DT3[, eval(columnsSelected)]

    ## still does not work: 
    DT3[, lapply(columnsSelected, get)]

如果要更改列的名称:

# Using the `.SDcols` method:  change names using `setnames`  (lowercase "n")
DT3[, setnames(.SD, c("new.Name1", "new.Name2")), .SDcols =columnsSelected]

# Using the `get` method:  
##  The names of the new columns will be the names of the `columnsSelected` vector
##  Thus, if we want to preserve the names, use the following: 
names(columnsSelected) <- columnsSelected    
DT3[, lapply(columnsSelected, function(.col) get(.col))]

## we can also use this trick to give the columns new names
names(columnsSelected) <- c("new.Name1", "new.Name2")
DT3[, lapply(columnsSelected, function(.col) get(.col))]

显然，使用 .SDcols 更简单、更优雅.

# `by` is straight forward, you can use a vector of strings in the `by` argument. 

# lets add another column to show how to use two columns in `by`
DT3[, secondID := sample(letters[1:2], 20, TRUE)]

# here is our string vector: 
byCols <- c("id", "secondID")

# and here is our call
DT3[, lapply(columnsSelected, function(.col) mean(get(.col))), by=byCols]

<小时>

把它们放在一起

我们可以通过引用它的名字来访问data.table，然后也可以通过名字来选择它的列:

---

PUTTING IT ALL TOGETHER

We can access the data.table by reference to its name and then select its columns also by name:

get(tablesSelected)[, .SD, .SDcols=columnsSelected]

## OR WITH MULTIPLE TABLES
tablesSelected <- c("DT1", "DT3")
lapply(tablesSelected, function(.T) get(.T)[, .SD, .SDcols=columnsSelected])

# we may want to name the vector for neatness, since
#  the resulting list inherits the names. 
names(tablesSelected) <- tablesSelected

这是最好的部分:

由于 data.table 中的很多内容都是通过引用传递的，因此很容易拥有一个表列表、一个要添加的单独列列表以及另一个要操作的列列表, 并将所有内容放在一起以在所有表​​上添加执行类似的操作 - 但输入不同.与使用 data.frame 做类似的事情相反，不需要重新分配最终结果.

THIS IS THE BEST PART:

Since so much in data.table is pass-by-reference, it is easy to have a list of tables, a separate list of columns to add and yet another list of columns to operate on, and put all together to add perform similar operations -- but with different inputs -- on all your tables. As opposed to doing something similar with data.frame, there is no need to reassign the end result.

newColumnsToAdd <- c("UpperBound", "LowerBound") 
FunctionToExecute <- function(vec) c(mean(vec) - 2*sd(vec), mean(vec) + 2*sd(vec))

# note the list of column names per table! 
columnsUsingPerTable <- list("DT1" = "Col1", DT2 = "Col2", DT3 = "Col1")
tablesSelected <- names(columnsUsingPerTable)
byCols <- c("id")

# TADA: 
dummyVar <- # I use `dummyVar` because I do not want to display the  output
lapply(tablesSelected, function(.T) 
  get(.T)[, c(newColumnsToAdd) := lapply(.SD, FunctionToExecute), .SDcols=columnsUsingPerTable[[.T]], by=byCols ]  )

# Take a look at the tables now: 
DT1
DT2
DT3

这篇关于Data.table 元编程的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！