r4ds/iteration.Rmd

586 lines
20 KiB
Plaintext
Raw Normal View History

2016-03-01 22:29:58 +08:00
# Iteration
```{r setup, include=FALSE}
library(purrr)
```
## For loops
Before we tackle the problem of rescaling each column, lets start with a simpler case. Imagine we want to summarise each column with its median. One way to do that is to use a for loop. Every for loop has three main components:
```{r}
df <- data.frame(
a = rnorm(10),
b = rnorm(10),
c = rnorm(10),
d = rnorm(10)
)
rescale01 <- function(x) {
rng <- range(x, na.rm = TRUE)
(x - rng[1]) / (rng[2] - rng[1])
}
results <- vector("numeric", ncol(df))
for (i in seq_along(df)) {
results[[i]] <- median(df[[i]])
}
results
```
There are three parts to a for loop:
1. The __results__: `results <- vector("integer", length(x))`.
This creates an integer vector the same length as the input. It's important
to enough space for all the results up front, otherwise you have to grow the
results vector at each iteration, which is very slow for large loops.
1. The __sequence__: `i in seq_along(df)`. This determines what to loop over:
each run of the for loop will assign `i` to a different value from
`seq_along(df)`, shorthand for `1:length(df)`. It's useful to think of `i`
as a pronoun.
1. The __body__: `results[i] <- median(df[[i]])`. This code is run repeatedly,
each time with a different value in `i`. The first iteration will run
`results[1] <- median(df[[2]])`, the second `results[2] <- median(df[[2]])`,
and so on.
This loop used a function you might not be familiar with: `seq_along()`. This is a safe version of the more familiar `1:length(l)`. There's one important difference in behaviour. If you have a zero-length vector, `seq_along()` does the right thing:
```{r}
y <- numeric(0)
seq_along(y)
1:length(y)
```
Lets go back to our original motivation:
```{r}
df$a <- rescale01(df$a)
df$b <- rescale01(df$b)
df$c <- rescale01(df$c)
df$d <- rescale01(df$d)
```
In this case the output is already present: we're modifying an existing object.
Think about a data frame as a list of columns (we'll make this definition precise later on). The length of a data frame is the number of columns. To extract a single column, you use `[[`.
That makes our for loop quite simple:
```{r, eval = FALSE}
for (i in seq_along(df)) {
df[[i]] <- rescale01(df[[i]])
}
```
For loops are not as important in R as they are in other languages as rather than writing your own for loops, you'll typically use prewritten functions that wrap up common for-loop patterns. You'll learn about those in the next chapter. These functions are important because they wrap up the book-keeping code related to the for loop, focussing purely on what's happening. For example the two for-loops we wrote above can be rewritten as:
```{r, eval = FALSE}
library(purrr)
map_dbl(df, median)
df[] <- map(df, rescale01)
```
The focus is now on the function doing the modification, rather than the apparatus of the for-loop.
### Looping patterns
There are three basic ways to loop over a vector:
1. Loop over the elements: `for (x in xs)`. Most useful for side-effects,
but it's difficult to save the output efficiently.
1. Loop over the numeric indices: `for (i in seq_along(xs))`. Most common
form if you want to know the element (`xs[[i]]`) and its position.
1. Loop over the names: `for (nm in names(xs))`. Gives you both the name
and the position. This is useful if you want to use the name in a
plot title or a file name.
The most general form uses `seq_along(xs)`, because from the position you can access both the name and the value:
```{r, eval = FALSE}
for (i in seq_along(x)) {
name <- names(x)[[i]]
value <- x[[i]]
}
```
### Exercises
1. Convert the song "99 bottles of beer on the wall" to a function. Generalise
to any number of any vessel containing any liquid on any surface.
1. Convert the nursey rhyme "ten in the bed" to a function. Generalise it
to any number of people in any sleeping structure.
1. It's common to see for loops that don't preallocate the output and instead
increase the length of a vector at each step:
```{r, eval = FALSE}
results <- vector("integer", 0)
for (i in seq_along(x)) {
results <- c(results, lengths(x[[i]]))
}
results
```
How does this affect performance?
## For loops vs functionals
Imagine you have a data frame and you want to compute the mean of each column. You might write code like this:
```{r}
df <- data.frame(
a = rnorm(10),
b = rnorm(10),
c = rnorm(10),
d = rnorm(10)
)
results <- numeric(length(df))
for (i in seq_along(df)) {
results[i] <- mean(df[[i]])
}
results
```
(Here we're taking advantage of the fact that a data frame is a list of the individual columns, so `length()` and `seq_along()` are useful.)
You realise that you're going to want to compute the means of every column pretty frequently, so you extract it out into a function:
```{r}
col_mean <- function(df) {
results <- numeric(length(df))
for (i in seq_along(df)) {
results[i] <- mean(df[[i]])
}
results
}
```
But then you think it'd also be helpful to be able to compute the median or the standard deviation:
```{r}
col_median <- function(df) {
results <- numeric(length(df))
for (i in seq_along(df)) {
results[i] <- median(df[[i]])
}
results
}
col_sd <- function(df) {
results <- numeric(length(df))
for (i in seq_along(df)) {
results[i] <- sd(df[[i]])
}
results
}
```
I've now copied-and-pasted this function three times, so it's time to think about how to generalise it. Most of the code is for-loop boilerplate and it's hard to see the one piece (`mean()`, `median()`, `sd()`) that differs.
What would you do if you saw a set of functions like this:
```{r}
f1 <- function(x) abs(x - mean(x)) ^ 1
f2 <- function(x) abs(x - mean(x)) ^ 2
f3 <- function(x) abs(x - mean(x)) ^ 3
```
Hopefully, you'd notice that there's a lot of duplication, and extract it out into an additional argument:
```{r}
f <- function(x, i) abs(x - mean(x)) ^ i
```
You've reduce the chance of bugs (because you now have 1/3 less code), and made it easy to generalise to new situations. We can do exactly the same thing with `col_mean()`, `col_median()` and `col_sd()`, by adding an argument that contains the function to apply to each column:
```{r}
col_summary <- function(df, fun) {
out <- vector("numeric", length(df))
for (i in seq_along(df)) {
out[i] <- fun(df[[i]])
}
out
}
col_summary(df, median)
col_summary(df, min)
```
The idea of using a function as an argument to another function is extremely powerful. It might take you a while to wrap your head around it, but it's worth the investment. In the rest of the chapter, you'll learn about and use the purrr package which provides a set of functions that eliminate the need for for-loops for many common scenarios.
### Exercises
1. Read the documentation for `apply()`. In the 2d case, what two for loops
does it generalise?
1. Adapt `col_summary()` so that it only applies to numeric columns
You might want to start with an `is_numeric()` function that returns
a logical vector that has a TRUE corresponding to each numeric column.
## The map functions
This pattern of looping over a list and doing something to each element is so common that the purrr package provides a family of functions to do it for you. Each function always returns the same type of output so there are six variations based on what sort of result you want:
* `map()` returns a list.
* `map_lgl()` returns a logical vector.
* `map_int()` returns a integer vector.
* `map_dbl()` returns a double vector.
* `map_chr()` returns a character vector.
* `map_df()` returns a data frame.
* `walk()` returns nothing. Walk is a little different to the others because
it's called exclusively for its side effects, so it's described in more detail
later in [walk](#walk).
Each function takes a list as input, applies a function to each piece, and then returns a new vector that's the same length as the input. The type of the vector is determined by the specific map function. Usually you want to use the most specific available, using `map()` only as a fallback when there is no specialised equivalent available.
We can use these functions to perform the same computations as the previous for loops:
```{r}
map_int(df, length)
map_dbl(df, mean)
map_dbl(df, median)
```
Compared to using a for loop, focus is on the operation being performed (i.e. `length()`, `mean()`, or `median()`), not the book-keeping required to loop over every element and store the results.
There are a few differences between `map_*()` and `compute_summary()`:
* All purrr functions are implemented in C. This means you can't easily
understand their code, but it makes them a little faster.
* The second argument, `.f`, the function to apply, can be a formula, a
character vector, or an integer vector. You'll learn about those handy
shortcuts in the next section.
* Any arguments after `.f` will be passed on to it each time it's called:
```{r}
map_dbl(df, mean, trim = 0.5)
```
* The map functions also preserve names:
```{r}
z <- list(x = 1:3, y = 4:5)
map_int(z, length)
```
### Shortcuts
There are a few shortcuts that you can use with `.f` in order to save a little typing. Imagine you want to fit a linear model to each group in a dataset. The following toy example splits the up the `mtcars` dataset in to three pieces (one for each value of cylinder) and fits the same linear model to each piece:
```{r}
models <- mtcars %>%
split(.$cyl) %>%
map(function(df) lm(mpg ~ wt, data = df))
```
The syntax for creating an anonymous function in R is quite verbose so purrr provides a convenient shortcut: a one-sided formula.
```{r}
models <- mtcars %>%
split(.$cyl) %>%
map(~lm(mpg ~ wt, data = .))
```
Here I've used `.` as a pronoun: it refers to the current list element (in the same way that `i` referred to the current index in the for loop). You can also use `.x` and `.y` to refer to up to two arguments. If you want to create a function with more than two arguments, do it the regular way!
When you're looking at many models, you might want to extract a summary statistic like the $R^2$. To do that we need to first run `summary()` and then extract the component called `r.squared`. We could do that using the shorthand for anonymous functions:
```{r}
models %>%
map(summary) %>%
map_dbl(~.$r.squared)
```
But extracting named components is a common operation, so purrr provides an even shorter shortcut: you can use a string.
```{r}
models %>%
map(summary) %>%
map_dbl("r.squared")
```
You can also use a numeric vector to select elements by position:
```{r}
x <- list(list(1, 2, 3), list(4, 5, 6), list(7, 8, 9))
x %>% map_dbl(2)
```
### Base R
If you're familiar with the apply family of functions in base R, you might have noticed some similarities with the purrr functions:
* `lapply()` is basically identical to `map()`. There's no advantage to using
`map()` over `lapply()` except that it's consistent with all the other
functions in purrr.
* The base `sapply()` is a wrapper around `lapply()` that automatically tries
to simplify the results. This is useful for interactive work but is
problematic in a function because you never know what sort of output
you'll get:
```{r}
x1 <- list(
c(0.27, 0.37, 0.57, 0.91, 0.20),
c(0.90, 0.94, 0.66, 0.63, 0.06),
c(0.21, 0.18, 0.69, 0.38, 0.77)
)
x2 <- list(
c(0.50, 0.72, 0.99, 0.38, 0.78),
c(0.93, 0.21, 0.65, 0.13, 0.27),
c(0.39, 0.01, 0.38, 0.87, 0.34)
)
threshold <- function(x, cutoff = 0.8) x[x > cutoff]
str(sapply(x1, threshold))
str(sapply(x2, threshold))
```
* `vapply()` is a safe alternative to `sapply()` because you supply an additional
argument that defines the type. The only problem with `vapply()` is that
it's a lot of typing: `vapply(df, is.numeric, logical(1))` is equivalent to
`map_lgl(df, is.numeric)`.
One of advantage of `vapply()` over the map functions is that it can also
produce matrices - the map functions only ever produce vectors.
* `map_df(x, f)` is effectively the same as `do.call("rbind", lapply(x, f))`
but under the hood is much more efficient.
### Exercises
1. How can you determine which columns in a data frame are factors?
(Hint: data frames are lists.)
1. What happens when you use the map functions on vectors that aren't lists?
What does `map(1:5, runif)` do? Why?
1. What does `map(-2:2, rnorm, n = 5)` do. Why?
1. Rewrite `map(x, function(df) lm(mpg ~ wt, data = df))` to eliminate the
anonymous function.
## Dealing with failure
When you do many operations on a list, sometimes one will fail. When this happens, you'll get an error message, and no output. This is annoying: why does one failure prevent you from accessing all the other successes? How do you ensure that one bad apple doesn't ruin the whole barrel?
In this section you'll learn how to deal this situation with a new function: `safely()`. `safely()` is an adverb: it takes a function (a verb) and returns a modified version. In this case, the modified function will never throw an error. Instead, it always returns a list with two elements:
1. `result` is the original result. If there was an error, this will be `NULL`.
1. `error` is an error object. If the operation was successful this will be
`NULL`.
(You might be familiar with the `try()` function in base R. It's similar, but because it sometimes returns the original result and it sometimes returns an error object it's more difficult to work with.)
Let's illustrate this with a simple example: `log()`:
```{r}
safe_log <- safely(log)
str(safe_log(10))
str(safe_log("a"))
```
When the function succeeds the `result` element contains the result and the `error` element is `NULL`. When the function fails, the `result` element is `NULL` and the `error` element contains an error object.
`safely()` is designed to work with map:
```{r}
x <- list(1, 10, "a")
y <- x %>% map(safely(log))
str(y)
```
This would be easier to work with if we had two lists: one of all the errors and one of all the results. That's easy to get with `transpose()`.
```{r}
y <- y %>% transpose()
str(y)
```
It's up to you how to deal with the errors, but typically you'll either look at the values of `x` where `y` is an error or work with the values of y that are ok:
```{r}
is_ok <- y$error %>% map_lgl(is_null)
x[!is_ok]
y$result[is_ok] %>% flatten_dbl()
```
Purrr provides two other useful adverbs:
* Like `safely()`, `possibly()` always succeeds. It's simpler than `safely()`,
because you give it a default value to return when there is an error.
```{r}
x <- list(1, 10, "a")
x %>% map_dbl(possibly(log, NA_real_))
```
* `quietly()` performs a similar role to `safely()`, but instead of capturing
errors, it captures printed output, messages, and warnings:
```{r}
x <- list(1, -1)
x %>% map(quietly(log)) %>% str()
```
### Exercises
1. Challenge: read all the csv files in this directory. Which ones failed
and why?
```{r, eval = FALSE}
files <- dir("data", pattern = "\\.csv$")
files %>%
set_names(., basename(.)) %>%
map_df(safely(readr::read_csv), .id = "filename") %>%
```
## Parallel maps
So far we've mapped along a single list. But often you have multiple related lists that you need iterate along in parallel. That's the job of the `map2()` and `pmap()` functions. For example, imagine you want to simulate some random normals with different means. You know how to do that with `map()`:
```{r}
mu <- list(5, 10, -3)
mu %>% map(rnorm, n = 10)
```
What if you also want to vary the standard deviation? You need to iterate along a vector of means and a vector of standard deviations in parallel. That's a job for `map2()` which works with two parallel sets of inputs:
```{r}
sigma <- list(1, 5, 10)
map2(mu, sigma, rnorm, n = 10)
```
`map2()` generates this series of function calls:
```{r, echo = FALSE, out.width = "75%"}
knitr::include_graphics("diagrams/lists-map2.png")
```
The arguments that vary for each call come before the function name, and arguments that are the same for every function call come afterwards.
Like `map()`, `map2()` is just a wrapper around a for loop:
```{r}
map2 <- function(x, y, f, ...) {
out <- vector("list", length(x))
for (i in seq_along(x)) {
out[[i]] <- f(x[[i]], y[[i]], ...)
}
out
}
```
You could also imagine `map3()`, `map4()`, `map5()`, `map6()` etc, but that would get tedious quickly. Instead, purrr provides `pmap()` which takes a list of arguments. You might use that if you wanted to vary the mean, standard deviation, and number of samples:
```{r}
n <- list(1, 3, 5)
args1 <- list(n, mu, sigma)
args1 %>% pmap(rnorm) %>% str()
```
That looks like:
```{r, echo = FALSE, out.width = "75%"}
knitr::include_graphics("diagrams/lists-pmap-unnamed.png")
```
However, instead of relying on position matching, it's better to name the arguments. This is more verbose, but it makes the code clearer.
```{r}
args2 <- list(mean = mu, sd = sigma, n = n)
args2 %>% pmap(rnorm) %>% str()
```
That generates longer, but safer, calls:
```{r, echo = FALSE, out.width = "75%"}
knitr::include_graphics("diagrams/lists-pmap-named.png")
```
Since the arguments are all the same length, it makes sense to store them in a data frame:
```{r}
params <- dplyr::data_frame(mean = mu, sd = sigma, n = n)
params$result <- params %>% pmap(rnorm)
params
```
As soon as your code gets complicated, I think a data frame is a good approach because it ensures that each column has a name and is the same length as all the other columns. We'll come back to this idea when we explore the intersection of dplyr, purrr, and model fitting.
### Invoking different functions
There's one more step up in complexity - as well as varying the arguments to the function you might also vary the function itself:
```{r}
f <- c("runif", "rnorm", "rpois")
param <- list(
list(min = -1, max = 1),
list(sd = 5),
list(lambda = 10)
)
```
To handle this case, you can use `invoke_map()`:
```{r}
invoke_map(f, param, n = 5) %>% str()
```
```{r, echo = FALSE}
knitr::include_graphics("diagrams/lists-invoke.png")
```
The first argument is a list of functions or character vector of function names. The second argument is a list of lists giving the arguments that vary for each function. The subsequent arguments are passed on to every function.
You can use `dplyr::frame_data()` to make creating these matching pairs a little easier:
```{r, eval = FALSE}
# Needs dev version of dplyr
sim <- dplyr::frame_data(
~f, ~params,
"runif", list(min = -1, max = -1),
"rnorm", list(sd = 5),
"rpois", list(lambda = 10)
)
sim %>% dplyr::mutate(
samples = invoke_map(f, params, n = 10)
)
```
## Walk {#walk}
Walk is an alternative to map that you use when you want to call a function for its side effects, rather than for its return value. You typically do this because you want to render output to the screen or save files to disk - the important thing is the action, not the return value. Here's a very simple example:
```{r}
x <- list(1, "a", 3)
x %>%
walk(print)
```
`walk()` is generally not that useful compared to `walk2()` or `pwalk()`. For example, if you had a list of plots and a vector of file names, you could use `pwalk()` to save each file to the corresponding location on disk:
```{r}
library(ggplot2)
plots <- mtcars %>%
split(.$cyl) %>%
map(~ggplot(., aes(mpg, wt)) + geom_point())
paths <- paste0(names(plots), ".pdf")
pwalk(list(paths, plots), ggsave, path = tempdir())
```
`walk()`, `walk2()` and `pwalk()` all invisibly return the `.x`, the first argument. This makes them suitable for use in the middle of pipelines.