data.table |组内更快的逐行递归更新

时间:2017-01-27 03:17:02

标签: r recursion data.table shift

我必须执行以下递归逐行操作才能获得z

myfun = function (xb, a, b) {

z = NULL

for (t in 1:length(xb)) {

    if (t >= 2) { a[t] = b[t-1] + xb[t] }
    z[t] = rnorm(1, mean = a[t])
    b[t] = a[t] + z[t]

}

return(z)

}

set.seed(1)

n_smpl = 1e6 
ni = 5

id = rep(1:n_smpl, each = ni)

smpl = data.table(id)
smpl[, time := 1:.N, by = id]

a_init = 1; b_init = 1
smpl[, ':=' (a = a_init, b = b_init)]
smpl[, xb := (1:.N)*id, by = id]

smpl[, z := myfun(xb, a, b), by = id]

我想得到这样的结果:

      id time a b  xb            z
  1:   1    1 1 1   1    0.3735462
  2:   1    2 1 1   2    2.7470924
  3:   1    3 1 1   3    8.4941848
  4:   1    4 1 1   4   20.9883695
  5:   1    5 1 1   5   46.9767390
 ---                              
496: 100    1 1 1 100    0.3735462
497: 100    2 1 1 200  200.7470924
498: 100    3 1 1 300  701.4941848
499: 100    4 1 1 400 1802.9883695
500: 100    5 1 1 500 4105.9767390

这确实有效,但需要时间:

system.time(smpl[, z := myfun(xb, a, b), by = id])
   user  system elapsed 
 33.646   0.994  34.473

考虑到我的实际数据(超过200万次观察),我需要让它更快。我猜do.call(myfun, .SD), .SDcols = c('xb', 'a', 'b')by = .(id, time)会更快,避免myfun内的for循环。但是,当我在b中运行这个逐行操作时,我不确定如何更新shift及其滞后(可能使用data.table)。有什么建议吗?

1 个答案:

答案 0 :(得分:36)

很棒的问题!

从一个全新的R会话开始,显示500万行的演示数据,这是您在笔记本电脑上的问题和时间的功能。有一些内联评论。

require(data.table)   # v1.10.0
n_smpl = 1e6
ni = 5
id = rep(1:n_smpl, each = ni)
smpl = data.table(id)
smpl[, time := 1:.N, by = id]
a_init = 1; b_init = 1
smpl[, ':=' (a = a_init, b = b_init)]
smpl[, xb := (1:.N)*id, by = id]

myfun = function (xb, a, b) {

  z = NULL
  # initializes a new length-0 variable

  for (t in 1:length(xb)) {

      if (t >= 2) { a[t] = b[t-1] + xb[t] }
      # if() on every iteration. t==1 could be done before loop

      z[t] = rnorm(1, mean = a[t])
      # z vector is grown by 1 item, each time

      b[t] = a[t] + z[t]
      # assigns to all of b vector when only really b[t-1] is
      # needed on the next iteration 
  }
  return(z)
}

set.seed(1); system.time(smpl[, z := myfun(xb, a, b), by = id][])
   user  system elapsed 
 19.216   0.004  19.212

smpl
              id time a b      xb            z
      1:       1    1 1 1       1 3.735462e-01
      2:       1    2 1 1       2 3.557190e+00
      3:       1    3 1 1       3 9.095107e+00
      4:       1    4 1 1       4 2.462112e+01
      5:       1    5 1 1       5 5.297647e+01
     ---                                      
4999996: 1000000    1 1 1 1000000 1.618913e+00
4999997: 1000000    2 1 1 2000000 2.000000e+06
4999998: 1000000    3 1 1 3000000 7.000003e+06
4999999: 1000000    4 1 1 4000000 1.800001e+07
5000000: 1000000    5 1 1 5000000 4.100001e+07

所以 19.2s 是时候击败。在所有这些时间中,我在本地运行命令3次,以确保它是一个稳定的时间。时间差异在这个任务中是微不足道的,所以我只报告一个时间来更快地阅读答案。

myfun()

中处理上面的评论
myfun2 = function (xb, a, b) {

  z = numeric(length(xb))
  # allocate size up front rather than growing

  z[1] = rnorm(1, mean=a[1])
  prevb = a[1]+z[1]
  t = 2L
  while(t<=length(xb)) {
    at = prevb + xb[t]
    z[t] = rnorm(1, mean=at)
    prevb = at + z[t]
    t = t+1L
  }
  return(z)
}
set.seed(1); system.time(smpl[, z2 := myfun2(xb, a, b), by = id][])
   user  system elapsed 
 13.212   0.036  13.245 
smpl[,identical(z,z2)]
[1] TRUE

那是相当不错的(19.2秒到13.2秒),但它仍然是R级别的for循环。乍一看它无法进行矢量化,因为rnorm()调用取决于之前的值。实际上,它可以通过使用m+sd*rnorm(mean=0,sd=1) == rnorm(mean=m, sd=sd)并调用矢量化rnorm(n=5e6)一次而不是5e6次的属性进行矢量化。但是可能会有一个cumsum()来处理这些团体。所以,我们不要去那里,因为这可能会使代码更难以阅读,并且将特定于这个精确的问题。

所以让我们试试Rcpp,它看起来与你写的风格非常相似,并且适用范围更广:

require(Rcpp)   # v0.12.8
cppFunction(
'NumericVector myfun3(IntegerVector xb, NumericVector a, NumericVector b) {
  NumericVector z = NumericVector(xb.length());
  z[0] = R::rnorm(/*mean=*/ a[0], /*sd=*/ 1);
  double prevb = a[0]+z[0];
  int t = 1;
  while (t<xb.length()) {
    double at = prevb + xb[t];
    z[t] = R::rnorm(at, 1);
    prevb = at + z[t];
    t++;
  }
  return z;
}')

set.seed(1); system.time(smpl[, z3 := myfun3(xb, a, b), by = id][])
   user  system elapsed 
  1.800   0.020   1.819 
smpl[,identical(z,z3)]
[1] TRUE

好多了: 19.2秒降至1.8秒。但每次调用该函数都会调用第一行(NumericVector()),该行分配一个新的向量,只要该组中的行数即可。然后填写并返回,然后将其复制到该组的正确位置的最终列(:=),仅发布。所有这100万个小临时载体(每组一个)的分配和管理都有点复杂。

为什么我们不能一次性完成整个专栏?你已经用for循环风格编写了它并没有错。我们调整C函数以接受id列,并在它到达新组时添加if

cppFunction(
'NumericVector myfun4(IntegerVector id, IntegerVector xb, NumericVector a, NumericVector b) {

  // ** id must be pre-grouped, such as via setkey(DT,id) **

  NumericVector z = NumericVector(id.length());
  int previd = id[0]-1;  // initialize to anything different than id[0]
  for (int i=0; i<id.length(); i++) {
    double prevb;
    if (id[i]!=previd) {
      // first row of new group
      z[i] = R::rnorm(a[i], 1);
      prevb = a[i]+z[i];
      previd = id[i];
    } else {
      // 2nd row of group onwards
      double at = prevb + xb[i];
      z[i] = R::rnorm(at, 1);
      prevb = at + z[i];
    }
  }
  return z;
}')

system.time(setkey(smpl,id))  # ensure grouped by id
   user  system elapsed
  0.028   0.004   0.033
set.seed(1); system.time(smpl[, z4 := myfun4(id, xb, a, b)][])
   user  system elapsed 
  0.232   0.004   0.237 
smpl[,identical(z,z4)]
[1] TRUE

那更好: 19.2s降至0.27s