Site icon R-bloggers

Easier way to chain commands using Pipe function

[This article was first published on The blog of Kun Ren, and kindly contributed to R-bloggers]. (You can report issue about the content on this page here)
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

In pipeR 0.4 version, one of the new features is Pipe() function. The function basically creates a Pipe object that allows command chaining with $, and thus makes it easier to perform operations in pipeline without any external operator.

In this post, I will introduce how to use this function and some basic knowledge about how it works. But before that, I would like to make clear that you don't have to learn a whole new thing if you are familiar with magrittr's %>% operator or pipeR's %>>% operator. If you are not, you can go ahead without hesitation. After all, the tools are made to be easier to work with.

Introducing Pipe()

Consider a task we plot the log differences of 100 normally distributed random numbers with mean 10. The traditional code can be written as

plot(diff(log(rnorm(100, mean = 10))),col = "red")

magrittr's %>% and pipeR's %>>% are designed to chain these commands in a human readable way. With %>% operator, the code can be restructured like

library(magrittr)
rnorm(100, mean = 10) %>%
  log %>%
  diff %>%
  plot(col="red")

In this case, %>% and %>>% are interchangeable which produce similar output. The operator does nothing special but hack the expression so that the left-hand side object is inserted into the function call on the right-hand side of the operator.

library(pipeR)
rnorm(100, mean = 10) %>>%
  log %>>%
  diff %>>%
  plot(col="red")

From the examples above, it seems that %>% and %>>% are exactly the same. In fact, they are not. I wrote an article Difference between magrittr and pipeR to explain their differences.

Both operators can solve the problem above by building a pipeline to avoid deeply nested code and make the operations readable. But is there an even easier way? The answer is Yes.

With Pipe() function introduced in pipeR 0.4, the code can be more simplified, even without any weird user-defined operator that has to be enclosed by % %. It goes like

library(pipeR)
Pipe(rnorm(100, mean = 10))$
  log()$
  diff()$
  plot(col="red")

You may have noticed that the pipeline starts with Pipe() function. This function basically creates a Pipe object which, in essence, is an environment which stores a value and whose $ is specially defined to perform first-argument piping. If a function name that follows $ is called, then the resulted value will be stored in the next-level Pipe object.

Pipe(c(1,2,3))$
  mean()

$value : numeric 
------
[1] 2

Note that the output indicates that the result is not a simple numeric vector but a box that contains that numeric vector as an element $value.

To see the difference, try to run

Pipe(c(1,2,3))$mean() + 1

Error: non-numeric argument to binary operator

If the pipeline returns a numeric value 2, it should add 1 and return 3 as a result. Clearly, this is not the case. It is the box containing the value that allows $ to perform more levels of piping. In fact, The pipeline construction does not stop until the value is extracted by $value.

Pipe(c(1,2,3))$
  mean()$
  value

[1] 2

or simply [] as a shortcut.

Pipe(c(1,2,3))$
  mean() []

[1] 2

Once the value is extracted from the box (or Pipe environment), the pipeline is ended with the stored value returned.

Having known these features, Pipe() function can be used to work with pipeline-friendly packages such as dplyr, ggvis, and rlist. Here are some simple examples.

Pipe() works with dplyr functions.

library(dplyr)
Pipe(mtcars)$
  filter(mpg <= mean(mpg))$
  select(mpg, cyl, wt)$
  group_by(cyl)$
  do(Pipe(.)$
      arrange(wt)$
      head(1)$
      value)$
  value

Source: local data frame [2 x 3]
Groups: cyl

   mpg cyl   wt
1 19.7   6 2.77
2 15.8   8 3.17

Pipe() works with ggvis.

library(ggvis)
Pipe(mtcars)$
  ggvis(~ mpg, ~ wt)$
  layer_points()$
  layer_smooths()

Pipe() also works with rlist.

library(rlist)
Pipe(1:10)$
  list.filter(x ~ x <= 5)$
  list.mapv(letters[.])

$value : character 
------
[1] "a" "b" "c" "d" "e"

More features

As I mentioned in Introducing pipeR 0.4, pipeR's %>>% operator is able to

The same features are supported with .() function used with Pipe(). For example,

Pipe(mtcars)$
  .(lm(mpg ~ cyl + wt, data = .))$
  summary()$
  .(coefficients)

$value : matrix 
------
            Estimate Std. Error t value  Pr(>|t|)
(Intercept)   39.686     1.7150  23.141 3.043e-20
cyl           -1.508     0.4147  -3.636 1.064e-03
wt            -3.191     0.7569  -4.216 2.220e-04

You can regard the above code as evaluated in the following steps:

m <- lm(mpg ~ cyl + wt, data = mtcars)
msum <- summary(m)
msum$coefficients

A noteworthy difference between the results produced by the two cases is that the final result produced by Pipe() is still stored in the Pipe object (the box), and you can extract the value or build longer pipeline with it. For example,

model <- Pipe(mtcars)$
  .(lm(mpg ~ cyl + wt, data = .))

Then model is a Pipe object in which the value is a linear model and can be used for further piping.

model$summary()$.(r.squared)

$value : numeric 
------
[1] 0.8302

model$predict(list(cyl = 6, wt = 2.9))

$value : numeric 
------
    1 
21.39 

Another interesting feature of Pipe object is about creating easy-to-use closures (roughly, a function created runtime within a context). For example, we can create a closure that generates 10 uniformly distributed numbers but its range is undecided.

rnd <- Pipe(10)$runif

A function rnd(...) has been created an it can be used to generate 10 uniformly distributed random numbers with different settings of range.

rnd(min = 1, max = 2)

$value : numeric 
------
 [1] 1.258 1.552 1.056 1.469 1.484 1.812 1.370 1.547 1.170 1.625

rnd(min = 10, max = 20)

$value : numeric 
------
 [1] 18.82 12.80 13.98 17.63 16.69 12.05 13.58 13.59 16.90 15.36

Performance

The overhead of Pipe() function is very low. Its performance is very close to %>>%. In intensive iterations, using Pipe() may also save some time. For more details, see pipeR's vignette Performance.

Conclusion

While %>% and %>>% implements operator-based pipeline like in F#, Pipe() function implements an object-like pipeline mechanism like the implementation in jQuery in JavaScript and LINQ in C#.

It dynamically creates closures as if the object had the child function to operate with it. It is more light-weight and easier to type than operator approach especially in R which requires user-defined operators take a name enclosed by % %.

If you like this idea, just install pipeR with

install.packages("pipeR")

and try Pipe().

To leave a comment for the author, please follow the link and comment on their blog: The blog of Kun Ren.

R-bloggers.com offers daily e-mail updates about R news and tutorials about learning R and many other topics. Click here if you're looking to post or find an R/data-science job.
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.