Introduction to R6 classes

The R6 package provides a type of class which is similar to R’s standard reference classes, but it is more efficient and doesn’t depend on S4 classes and the methods package.

R6 classes

R6 classes are similar to R’s standard reference classes, but are lighter weight, and avoid some issues that come along with using S4 classes (R’s reference classes are based on S4). For more information about speed and memory footprint, see the Performance vignette.

Unlike many objects in R, instances (objects) of R6 classes have reference semantics. R6 classes also support:

Why the name R6? When R’s reference classes were introduced, some users, following the names of R’s existing class systems S3 and S4, called the new class system R5 in jest. Although reference classes are not actually called R5, the name of this package and its classes takes inspiration from that name.

The name R5 was also a code-name used for a different object system started by Simon Urbanek, meant to solve some issues with S4 relating to syntax and performance. However, the R5 branch was shelved after a little development, and it was never released.

Basics

Here’s how to create a simple R6 class. The public argument is a list of items, which can be functions and fields (non-functions). Functions will be used as methods.

library(R6)

Person <- R6Class("Person",
  public = list(
    name = NULL,
    hair = NULL,
    initialize = function(name = NA, hair = NA) {
      self$name <- name
      self$hair <- hair
      self$greet()
    },
    set_hair = function(val) {
      self$hair <- val
    },
    greet = function() {
      cat(paste0("Hello, my name is ", self$name, ".\n"))
    }
  )
)

To instantiate an object of this class, use $new():

ann <- Person$new("Ann", "black")
#> Hello, my name is Ann.
ann
#> <Person>
#>   Public:
#>     clone: function (deep = FALSE) 
#>     greet: function () 
#>     hair: black
#>     initialize: function (name = NA, hair = NA) 
#>     name: Ann
#>     set_hair: function (val)

The $new() method creates the object and calls the initialize() method, if it exists.

Inside methods of the class, self refers to the object. Public members of the object (all you’ve seen so far) are accessed with self$x, and assignment is done with self$x <- y. Note that by default, self is required to access members, although for non-portable classes which we’ll see later, it is optional.

Once the object is instantiated, you can access values and methods with $:

ann$hair
#> [1] "black"
ann$greet()
#> Hello, my name is Ann.
ann$set_hair("red")
ann$hair
#> [1] "red"

Implementation note: The external face of an R6 object is basically an environment with the public members in it. This is also known as the public environment. An R6 object’s methods have a separate enclosing environment which, roughly speaking, is the environment they “run in”. This is where self binding is found, and it is simply a reference back to public environment.

Private members

In the previous example, all the members were public. It’s also possible to add private members:

Queue <- R6Class("Queue",
  public = list(
    initialize = function(...) {
      for (item in list(...)) {
        self$add(item)
      }
    },
    add = function(x) {
      private$queue <- c(private$queue, list(x))
      invisible(self)
    },
    remove = function() {
      if (private$length() == 0) return(NULL)
      # Can use private$queue for explicit access
      head <- private$queue[[1]]
      private$queue <- private$queue[-1]
      head
    }
  ),
  private = list(
    queue = list(),
    length = function() base::length(private$queue)
  )
)

q <- Queue$new(5, 6, "foo")

Whereas public members are accessed with self, like self$add(), private members are accessed with private, like private$queue.

The public members can be accessed as usual:

# Add and remove items
q$add("something")
q$add("another thing")
q$add(17)
q$remove()
#> [1] 5
q$remove()
#> [1] 6

However, private members can’t be accessed directly:

q$queue
#> NULL
q$length()
#> Error: attempt to apply non-function

A useful design pattern is for methods to return self (invisibly) when possible, because it makes them chainable. For example, the add() method returns self so you can chain them together:

q$add(10)$add(11)$add(12)

On the other hand, remove() returns the value removed, so it’s not chainable:

q$remove()
#> [1] "foo"
q$remove()
#> [1] "something"
q$remove()
#> [1] "another thing"
q$remove()
#> [1] 17

Active bindings

Active bindings look like fields, but each time they are accessed, they call a function. They are always publicly visible.

Numbers <- R6Class("Numbers",
  public = list(
    x = 100
  ),
  active = list(
    x2 = function(value) {
      if (missing(value)) return(self$x * 2)
      else self$x <- value/2
    },
    rand = function() rnorm(1)
  )
)

n <- Numbers$new()
n$x
#> [1] 100

When an active binding is accessed as if reading a value, it calls the function with value as a missing argument:

n$x2
#> [1] 200

When it’s accessed as if assigning a value, it uses the assignment value as the value argument:

n$x2 <- 1000
n$x
#> [1] 500

If the function takes no arguments, it’s not possible to use it with <-:

n$rand
#> [1] 0.2648
n$rand
#> [1] 2.171
n$rand <- 3
#> Error: unused argument (quote(3))

Implementation note: Active bindings are bound in the public environment. The enclosing environment for these functions is also the public environment.

Inheritance

One R6 class can inherit from another. In other words, you can have super- and sub-classes.

Subclasses can have additional methods, and they can also have methods that override the superclass methods. In this example of a queue that retains its history, we’ll add a show() method and override the remove() method:

# Note that this isn't very efficient - it's just for illustrating inheritance.
HistoryQueue <- R6Class("HistoryQueue",
  inherit = Queue,
  public = list(
    show = function() {
      cat("Next item is at index", private$head_idx + 1, "\n")
      for (i in seq_along(private$queue)) {
        cat(i, ": ", private$queue[[i]], "\n", sep = "")
      }
    },
    remove = function() {
      if (private$length() - private$head_idx == 0) return(NULL)
      private$head_idx <<- private$head_idx + 1
      private$queue[[private$head_idx]]
    }
  ),
  private = list(
    head_idx = 0
  )
)

hq <- HistoryQueue$new(5, 6, "foo")
hq$show()
#> Next item is at index 1 
#> 1: 5
#> 2: 6
#> 3: foo
hq$remove()
#> [1] 5
hq$show()
#> Next item is at index 2 
#> 1: 5
#> 2: 6
#> 3: foo
hq$remove()
#> [1] 6

Superclass methods can be called with super$xx(). The CountingQueue (example below) keeps a count of the total number of objects that have ever been added to the queue. It does this by overriding the add() method – it increments a counter and then calls the superclass’s add() method, with super$add(x):

CountingQueue <- R6Class("CountingQueue",
  inherit = Queue,
  public = list(
    add = function(x) {
      private$total <<- private$total + 1
      super$add(x)
    },
    get_total = function() private$total
  ),
  private = list(
    total = 0
  )
)

cq <- CountingQueue$new("x", "y")
cq$get_total()
#> [1] 2
cq$add("z")
cq$remove()
#> [1] "x"
cq$remove()
#> [1] "y"
cq$get_total()
#> [1] 3

Fields containing reference objects

If your R6 class contains any fields that also have reference semantics (e.g., other R6 objects, and environments), those fields should be populated in the initialize method. If the field set to the reference object directly in the class definition, that object will be shared across all instances of the R6 objects. Here’s an example:

SimpleClass <- R6Class("SimpleClass",
  public = list(x = NULL)
)

SharedField <- R6Class("SharedField",
  public = list(
    e = SimpleClass$new()
  )
)

s1 <- SharedField$new()
s1$e$x <- 1

s2 <- SharedField$new()
s2$e$x <- 2

# Changing s2$e$x has changed the value of s1$e$x
s1$e$x
#> [1] 2

To avoid this, populate the field in the initialize method:

NonSharedField <- R6Class("NonSharedField",
  public = list(
    e = NULL,
    initialize = function() self$e <- SimpleClass$new()
  )
)

n1 <- NonSharedField$new()
n1$e$x <- 1

n2 <- NonSharedField$new()
n2$e$x <- 2

# n2$e$x does not affect n1$e$x
n1$e$x
#> [1] 1

Portable and non-portable classes

In R6 version 1.0.1, the default was to create non-portable classes. In subsequent versions, the default is to create portable classes. The two most noticeable differences are that portable classes:

The implementation of the first point is such that it makes the second point necessary.

Using self and <<-

With reference classes, you can access the field without self, and assign to fields using <<-. For example:

RC <- setRefClass("RC",
  fields = list(x = 'ANY'),
  methods = list(
    getx = function() x,
    setx = function(value) x <<- value
  )
)

rc <- RC$new()
rc$setx(10)
rc$getx()
#> [1] 10

The same is true for non-portable R6 classes:

NP <- R6Class("NP",
  portable = FALSE,
  public = list(
    x = NA,
    getx = function() x,
    setx = function(value) x <<- value
  )
)

np <- NP$new()
np$setx(10)
np$getx()
#> [1] 10

But for portable R6 classes (this is the default), you must use self and/or private, and <<- assignment doesn’t work – unless you use self, of course:

P <- R6Class("P",
  portable = TRUE,  # This is default
  public = list(
    x = NA,
    getx = function() self$x,
    setx = function(value) self$x <- value
  )
)

p <- P$new()
p$setx(10)
p$getx()
#> [1] 10

For more information, see the Portable vignette.

Other topics

Adding members to an existing class

It is sometimes useful to add members to a class after the class has already been created. This can be done using the $set() method on the generator object.

Simple <- R6Class("Simple",
  public = list(
    x = 1,
    getx = function() self$x
  )
)

Simple$set("public", "getx2", function() self$x*2)

# To replace an existing member, use overwrite=TRUE
Simple$set("public", "x", 10, overwrite = TRUE)

s <- Simple$new()
s$x
#> [1] 10
s$getx2()
#> [1] 20

The new members will be present only in instances that are created after $set() has been called.

To prevent modification of a class, you can use lock_class=TRUE when creating the class. You can also lock and unlock a class as follows:

# Create a locked class
Simple <- R6Class("Simple",
  public = list(
    x = 1,
    getx = function() self$x
  ),
  lock_class = TRUE
)

# This would result in an error
# Simple$set("public", "y", 2)

# Unlock the class
Simple$unlock()

# Now it works
Simple$set("public", "y", 2)

# Lock the class again
Simple$lock()

Cloning objects

By default, R6 objects have method named clone for making a copy of the object.

Simple <- R6Class("Simple",
  public = list(
    x = 1,
    getx = function() self$x
  )
)

s <- Simple$new()

# Create a clone
s1 <- s$clone()
# Modify it
s1$x <- 2
s1$getx()
#> [1] 2

# Original is unaffected by changes to the clone
s$getx()
#> [1] 1

If you don’t want a clone method to be added, you can use cloneable=FALSE when creating the class. If any loaded R6 object has a clone method, that function uses 44 kB, but for each additional object, the clone method costs a trivial amount of space (112 bytes).

Deep cloning

If there are any fields which are objects with reference sematics (environments, R6 objects, reference class objects), the copy will get a reference to the same object. This is sometimes desirable, but often it is not.

For example, we’ll create an object c1 which contains another R6 object, s, and then clone it. Because the original’s and the clone’s s fields both refer to the same object, modifying it from one results in a change that is reflect in the other.

Simple <- R6Class("Simple", public = list(x = 1))

Cloneable <- R6Class("Cloneable",
  public = list(
    s = NULL,
    initialize = function() self$s <- Simple$new()
  )
)

c1 <- Cloneable$new()
c2 <- c1$clone()

# Change c1's `s` field
c1$s$x <- 2

# c2's `s` is the same object, so it reflects the change
c2$s$x
#> [1] 2

To make it so the clone receives a copy of s, we can use the deep=TRUE option:

c3 <- c1$clone(deep = TRUE)

# Change c1's `s` field
c1$s$x <- 3

# c2's `s` is different
c3$s$x
#> [1] 2

The default behavior of clone(deep=TRUE) is to copy fields which are R6 objects, but not copy fields which are environments, reference class objects, or other data structures which contain other reference-type objects (for example, a list with an R6 object).

If your R6 object contains these types of objects and you want to make a deep clone of them, you must provide your own function for deep cloning, in a private method named deep_clone. Below is an example of an R6 object with two fields, a and b, both of which which are environments, and both of which contain a value x. It also has a field v which is a regular (non-reference) value, and a private deep_clone method.

The deep_clone method is be called once for each field. It is passed the name and value of the field, and the value it returns is be used in the clone.

CloneEnv <- R6Class("CloneEnv",
  public = list(
    a = NULL,
    b = NULL,
    v = 1,
    initialize = function() {
      self$a <- new.env(parent = emptyenv())
      self$b <- new.env(parent = emptyenv())
      self$a$x <- 1
      self$b$x <- 1
    }
  ),
  private = list(
    deep_clone = function(name, value) {
      # With x$clone(deep=TRUE) is called, the deep_clone gets invoked once for
      # each field, with the name and value.
      if (name == "a") {
        # `a` is an environment, so use this quick way of copying
        list2env(as.list.environment(value, all.names = TRUE),
                 parent = emptyenv())
      } else {
        # For all other fields, just return the value
        value
      }
    }
  )
)

c1 <- CloneEnv$new()
c2 <- c1$clone(deep = TRUE)

When c1$clone(deep=TRUE) is called, the deep_clone method is called for each field in c1, and is passed the name of the field and value. In our version, the a environment gets copied, but b does not, nor does v (but that doesn’t matter since v is not a reference object). We can test out the clone:

# Modifying c1$a doesn't affect c2$a, because they're separate objects
c1$a$x <- 2
c2$a$x
#> [1] 1

# Modifying c1$b does affect c2$b, because they're the same object
c1$b$x <- 3
c2$b$x
#> [1] 3

# Modifying c1$v doesn't affect c2$v, because they're not reference objects
c1$v <- 4
c2$v
#> [1] 1

In the example deep_clone method above, we checked the name of each field to determine what to do with it, but we could also check the value, by using inherits(value, "R6"), or is.environment(), and so on.

Printing R6 objects to the screen

R6 objects have a default print method that lists all members of the object. If a class defines a print method, then it overrides the default one.

PrettyCountingQueue <- R6Class("PrettyCountingQueue",
  inherit = CountingQueue,
  public = list(
    print = function(...) {
      cat("<PrettyCountingQueue> of ", self$get_total(), " elements\n", sep = "")
      invisible(self)
    }
  )
)
pq <- PrettyCountingQueue$new(1, 2, "foobar")
pq
#> <PrettyCountingQueue> of 3 elements

Finalizers

Sometimes it’s useful to run a function when the object is garbage collected. For example, you may want to make sure a file or database connection gets closed. To do this, you can call the reg.finalizer() function in your initialize method, and pass it self as the object that will trigger the function when it is garbage collected.

A <- R6Class("A", public = list(
  initialize = function() {
    reg.finalizer(self,
                  function(e) print("Finalizer has been called!"),
                  onexit = TRUE)
  }
))


# Instantiate an object, but don't save a reference to it
A$new()
#> <A>
#>   Public:
#>     clone: function (deep = FALSE) 
#>     initialize: function ()

# Run something else to get rid of any references to the object, because the
# last returned value is saved in .Last.value
1+1
#> [1] 2

# Force immediate garbage collection (normally this will happen automatically
# from time to time)
gc()
#> [1] "Finalizer has been called!"
#>          used (Mb) gc trigger (Mb) max used (Mb)
#> Ncells 400276 21.4     750400 40.1   592000 31.7
#> Vcells 616985  4.8    1308461 10.0   909818  7.0

In the example above, we used onexit=TRUE, so that the finalizer will also be called when R exits. This is useful in some cases, like database connections, but it isn’t necessary for others, like open files, since they will be closed anyway when the R process exits.

Summary

R6 classes provide capabilities that are common in other object-oriented programming languages. They’re similar to R’s built-in reference classes, but are simpler, smaller, and faster, and they allow inheritance across packages.