Mungebits2 is 4x slower than mungebits

[robertzk]

For simple operations... this might be due to the evasion of nonstandard evaluation in the train and predict function. Really need to figure out the discrepancy here.

[robertzk]

It might boil down to:

> data.frame(replicate(1000, 1:5, simplify=F)) -> foo
> fn1 <- function(df) {
+   for(i in colnames(df)) { df[[i]] <- df[[i]] + 1}
+   df
+ }
> fn2 <- function(df) {
+   df[colnames(df)] <- lapply(seq_along(df), function(i) { .subset2(df, i) + 1 })
+   df
+ }
> system.time(fn1(df))
   user  system elapsed
  0.050   0.005   0.056
> dim(df)
[1]    5 1000
> system.time(fn2(df))
   user  system elapsed
  0.032   0.003   0.035

TLDR: Use lapply?

[robertzk]

Solved by environment magic. https://github.com/robertzk/mungebits2/pull/12

[robertzk]

Latest speedups bring it within 3x of the most optimized base R code.

> load_all(); doubler <- mungebit$new(column_transformation(function(x) { 2 * x })); microbenchmark( doubler$run(foo, 1:1000), raw_double(foo, 1:1000), times = 5L)
Loading mungebits2
Unit: milliseconds
                     expr      min       lq    mean   median       uq      max
 doubler$run(foo, 1:1000) 3.366846 3.432370 4.00230 3.466510 3.610990 6.134785
  raw_double(foo, 1:1000) 1.093207 1.100542 1.16094 1.129146 1.151777 1.330030

[robertzk]

Additionally, train and predict are 1.5 faster than mungebits1

load_all(); column_transformation(function(x) x + 1) -> ct;  mp <- mungepiece$new(mungebit$new(ct)); system.time(mp$run(df))
   user  system elapsed
  0.007   0.000   0.006
> system.time(mp$run(df)) # predict
   user  system elapsed
  0.004   0.000   0.004

Compared to mungebits1:

> load_all(); column_transformation(function(x) x + 1) -> ct; mp <- mungepiece$new(mungebit$new(ct)); df2 <- mungeplane(data.frame(foo)); system.time(mp$run(df2))
Loading mungebits
   user  system elapsed
  0.008   0.000   0.008
> system.time(mp$run(df2)) # predict
   user  system elapsed
  0.009   0.000   0.009