Skip to content

Latest commit

 

History

History
490 lines (314 loc) · 8.6 KB

00_pipelines.md

File metadata and controls

490 lines (314 loc) · 8.6 KB

Pipelines in R

autosize: true author: Tristan Mahr, @tjmahr date: March 18, 2015 css: assets/custom.css

Madison R Users Group

Repository for this talk: https://github.com/tjmahr/MadR_Pipelines

Scientific Computing

incremental: true

I'm interested in:

  • correctness
    • but not necessarily robustness against corner-cases
  • optimizing for human readers
    • collaborators, including me in the future
  • reproducibility
  • automation

Make bricks, not monoliths

I tackle these goals by building a problem-specific language from simpler, understandable pieces of code.

(But see also Best Practices for Scientific Computing for more tools and strategies.)

Working with bricks

incremental: true

  1. Develop a core vocabulary of functions.
    • including others' functions/packages.
  2. Construct your own functions on top of that core.
  3. Continue upwards.

Vocabulary

  • R Vocabulary
  • Awesome R
  • Packages that do one thing very well: dplyr, stringr, lubridate, broom, tidyr, rvest

Pause to demonstrate any requested functions

type: prompt

a_to_f <- head(letters)
a_to_f
tail(letters)
seq_along(a_to_f)
xs <- seq_len(10)
xs
ifelse(xs %% 2 == 0, xs, NA)

Functions Are Great

incremental: true

Solve a problem once, then re-use that solution elsewhere.

# Squish values into a range
squish <- function(xs, lower, upper) {
  xs[xs < lower] <- lower
  xs[upper < xs] <- upper
  xs
}
squish(rnorm(5), -.3, 1)
[1]  1.0000000 -0.3000000  0.7980856
[4] -0.3000000  0.2153025

Bootstrap from many smaller functions

incremental: true

Create my own problem-specific language.

# Insert values into the second-to-last
# position, in case last one is a delimiter
insert_line <- function(xs, ys) {
  c(but_last(xs), ys, last(xs))
}
but_last <- function(...) head(..., n = -1)
last <- function(...) tail(..., n = 1)
insert_line(c("x", "y", "z"), "&")
[1] "x" "y" "&" "z"

But readability quickly slips away.

incremental: true

Here's a function adapted from the strsplit help page.

mystery_func <- function(xs) {
  sapply(lapply(strsplit(xs, NULL), rev),
         paste, collapse = "")
}

Okay, maybe it would help if it were built out of understandable chunks.

With chunks!

incremental: true

"Extract function" refactoring

str_tokenize <- function(xs) {
  strsplit(xs, split = NULL)
}

str_collapse <- function(..., joiner = "") {
  paste(..., collapse = joiner)
}
mystery_func <- function(xs) {
  sapply(lapply(str_tokenize(xs), rev),
         str_collapse)
}

Okay, maybe it would help if it weren't a one-liner

Un-nest the function calls

incremental: true

Do one thing per line.

mystery_func <- function(xs) {
  char_sets <- str_tokenize(xs)
  char_sets_rev <- lapply(char_sets, rev)
  sapply(char_sets_rev, str_collapse)
}

Pretty good, but now we have these intermediate values we don't care about cluttering things up.

Pipelines to the rescue

type: section

A way to express successive data transformations.

Basic Idea

incremental: true

Use the value on the left-hand side as the first argument to the function on the right-hand side.

library("magrittr")

# Rule 1
f(xs)
xs %>% f

# Rule 2
g(xs, n = 5)
xs %>% g(n = 5)

Chaining pipes together

incremental: true

Do function composition by chaining pipes together.

# Rule 3
g(f(xs), n = 5)
xs %>% f %>% g(n = 5)

Mentally, read %>% as "then".

Take xs then do f then do g with n = 5.

Pipelines: Level 1

xs <- rnorm(5)
squish(sort(round(xs, 2)), -.3, 1)
[1] -0.30 -0.30 -0.27  0.20  0.68

xs %>% round(2) %>% sort %>% squish(-.3, 1)
[1] -0.30 -0.30 -0.27  0.20  0.68

You might already use pipelines

On the command line:

sort data.csv | uniq -u | wc -l
# 369 (number of unique lines)

Method chains are also like pipelines

incremental: true

Python

df.groupby(['letter','one']).sum()

Javascript

$("#p1")
  .css("color", "red")
  .slideUp(2000)
  .slideDown(2000);

Back to the mystery function

incremental: true

mystery_func <- function(xs) {
  str_tokenize(xs) %>%
    lapply(rev) %>%
    sapply(str_collapse)
}
  • Break each string into a vector of characters
  • THEN reverse each vector
  • THEN collapse each character vector together
words <- c("The", "quick", "brown", "fox")
mystery_func(words)
[1] "ehT"   "kciuq" "nworb" "xof"

Pause for questions

type: prompt

. - the placeholder

type: section

What if the input should not be the first argument?

incremental: true

Use . as an argument placeholder.

# Rule 4
f(y, x)
x %>% f(y, .)

# Rule 5
f(y, z = x)
x %>% f(y, z = .)

Placeholder says where the piped input should land.

Examples

words %>% paste0("~~", ., "~~") %>% toupper
[1] "~~THE~~"   "~~QUICK~~" "~~BROWN~~"
[4] "~~FOX~~"  

# As a named parameter
library("broom")
mtcars %>% lm(mpg ~ cyl * wt, data = .) %>%
  tidy %>% print(digits = 2)
         term estimate std.error statistic
1 (Intercept)    54.31      6.13       8.9
2         cyl    -3.80      1.01      -3.8
3          wt    -8.66      2.32      -3.7
4      cyl:wt     0.81      0.33       2.5
  p.value
1 1.3e-09
2 7.5e-04
3 8.6e-04
4 2.0e-02

Saving pipelines

incremental: true

The input to the pipeline can itself be a placeholder!

num_unique <- . %>% unique %>% length

In this case, the pipeline describes a function chain that can be saved and re-used. It also has a different print method.

num_unique
Functional sequence with the following components:

 1. unique(.)
 2. length(.)

Use 'functions' to extract the individual functions.

Final mystery_func

mystery_func <- . %>%
  str_tokenize %>%
  lapply(rev) %>%
  sapply(str_collapse)

mystery_func(words)
[1] "ehT"   "kciuq" "nworb" "xof"

That's most of it.

incremental: true

The pipe %>% and placeholder . covers 90% of magrittr.

What didn't I cover?

  • aliases (pipeline-friendly forms of functions like [, $, [[)
  • compound assignment %<>% (sugar for x <- x %>% ... )
  • tee %T>% (print, plot, save results during pipeline without interrupting the flow of data)
  • exposition %$% (like with as an infix)
  • braced expressions (arbitrary blocks of code in a pipeline)

See the magrittr vignette.

Next sections

Basic scheme:

  • Build or borrow a set of functions to work on a problem
  • Chain the functions together into an understandable pipeline

Next set of slides: dplyr for data-frames.