old_warn_on_fallback <- options(`vctrs:::warn_on_fallback` = FALSE)
knitr_defer(options(old_warn_on_fallback))

This guide illustrates how to implement vec_ptype2() and vec_cast() methods for existing classes. Related topics:

The natural number class

We'll illustrate how to implement coercion methods with a simple class that represents natural numbers. In this scenario we have an existing class that already features a constructor and methods for print() and subset.

#' @export
new_natural <- function(x) {
  if (is.numeric(x) || is.logical(x)) {
    stopifnot(is_whole(x))
    x <- as.integer(x)
  } else {
    stop("Can't construct natural from unknown type.")
  }
  structure(x, class = "my_natural")
}
is_whole <- function(x) {
  all(x %% 1 == 0 | is.na(x))
}

#' @export
print.my_natural <- function(x, ...) {
  cat("<natural>\n")
  x <- unclass(x)
  NextMethod()
}
#' @export
`[.my_natural` <- function(x, i, ...) {
  new_natural(NextMethod())
}
# Necessary because includeRmd() evaluated in a child of global
knitr_local_registration("base::print", "my_natural")
knitr_local_registration("base::[", "my_natural")
new_natural(1:3)
new_natural(c(1, NA))

Roxygen workflow


Implementing vec_ptype2()

The self-self method

The first method to implement is the one that signals that your class is compatible with itself:

#' @export
vec_ptype2.my_natural.my_natural <- function(x, y, ...) {
  x
}

vec_ptype2(new_natural(1), new_natural(2:3))
# Necessary because includeRmd() evaluated in a child of global
knitr_local_registration("vctrs::vec_ptype2", "my_natural.my_natural")

vec_ptype2() implements a fallback to try and be compatible with simple classes, so it may seem that you don't need to implement the self-self coercion method. However, you must implement it explicitly because this is how vctrs knows that a class that is implementing vctrs methods (for instance this disable fallbacks to base::c()). Also, it makes your class a bit more efficient.

The parent and children methods

Our natural number class is conceptually a parent of <logical> and a child of <integer>, but the class is not compatible with logical, integer, or double vectors yet:

vec_ptype2(TRUE, new_natural(2:3))

vec_ptype2(new_natural(1), 2:3)

We'll specify the twin methods for each of these classes, returning the richer class in each case.

#' @export
vec_ptype2.my_natural.logical <- function(x, y, ...) {
  # The order of the classes in the method name follows the order of
  # the arguments in the function signature, so `x` is the natural
  # number and `y` is the logical
  x
}
#' @export
vec_ptype2.logical.my_natural <- function(x, y, ...) {
  # In this case `y` is the richer natural number
  y
}

Between a natural number and an integer, the latter is the richer class:

#' @export
vec_ptype2.my_natural.integer <- function(x, y, ...) {
  y
}
#' @export
vec_ptype2.integer.my_natural <- function(x, y, ...) {
  x
}
# Necessary because includeRmd() evaluated in a child of global
knitr_local_registration("vctrs::vec_ptype2", "my_natural.logical")
knitr_local_registration("vctrs::vec_ptype2", "my_natural.integer")
knitr_local_registration("vctrs::vec_ptype2", "integer.my_natural")
knitr_local_registration("vctrs::vec_ptype2", "logical.my_natural")

We no longer get common type errors for logical and integer:

vec_ptype2(TRUE, new_natural(2:3))

vec_ptype2(new_natural(1), 2:3)

We are not done yet. Pairwise coercion methods must be implemented for all the connected nodes in the coercion hierarchy, which include double vectors further up. The coercion methods for grand-parent types must be implemented separately:

#' @export
vec_ptype2.my_natural.double <- function(x, y, ...) {
  y
}
#' @export
vec_ptype2.double.my_natural <- function(x, y, ...) {
  x
}
# Necessary because includeRmd() evaluated in a child of global
knitr_local_registration("vctrs::vec_ptype2", "my_natural.double")
knitr_local_registration("vctrs::vec_ptype2", "double.my_natural")

Incompatible attributes

Most of the time, inputs are incompatible because they have different classes for which no vec_ptype2() method is implemented. More rarely, inputs could be incompatible because of their attributes. In that case incompatibility is signalled by calling stop_incompatible_type().

In the following example, we implement a self-self ptype2 method for a hypothetical subclass of <factor> that has stricter combination semantics. The method throws when the levels of the two factors are not compatible.

#' @export
vec_ptype2.my_strict_factor.my_strict_factor <- function(x, y, ..., x_arg = "", y_arg = "") {
  if (!setequal(levels(x), levels(y))) {
    stop_incompatible_type(x, y, x_arg = x_arg, y_arg = y_arg)
  }

  x
}

Note how the methods need to take x_arg and y_arg parameters and pass them on to stop_incompatible_type(). These argument tags help create more informative error messages when the common type determination is for a column of a data frame. They are part of the generic signature but can usually be left out if not used.

Implementing vec_cast()

Corresponding vec_cast() methods must be implemented for all vec_ptype2() methods. The general pattern is to convert the argument x to the type of to. The methods should validate the values in x and make sure they conform to the values of to.

Please note that for historical reasons, the order of the classes in the method name is in reverse order of the arguments in the function signature. The first class represents to, whereas the second class represents x.

The self-self method is easy in this case, it just returns the target input:

#' @export
vec_cast.my_natural.my_natural <- function(x, to, ...) {
  x
}

The other types need to be validated. We perform input validation in the new_natural() constructor, so that's a good fit for our vec_cast() implementations.

#' @export
vec_cast.my_natural.logical <- function(x, to, ...) {
  # The order of the classes in the method name is in reverse order
  # of the arguments in the function signature, so `to` is the natural
  # number and `x` is the logical
  new_natural(x)
}
vec_cast.my_natural.integer <- function(x, to, ...) {
  new_natural(x)
}
vec_cast.my_natural.double <- function(x, to, ...) {
  new_natural(x)
}
# Necessary because includeRmd() evaluated in a child of global
knitr_local_registration("vctrs::vec_cast", "my_natural.my_natural")
knitr_local_registration("vctrs::vec_cast", "my_natural.logical")
knitr_local_registration("vctrs::vec_cast", "my_natural.integer")
knitr_local_registration("vctrs::vec_cast", "my_natural.double")

With these methods, vctrs is now able to combine logical and natural vectors. It properly returns the richer type of the two, a natural vector:

vec_c(TRUE, new_natural(1), FALSE)

Because we haven't implemented conversions from natural, it still doesn't know how to combine natural with the richer integer and double types:

vec_c(new_natural(1), 10L)
vec_c(1.5, new_natural(1))

This is quick work which completes the implementation of coercion methods for vctrs:

#' @export
vec_cast.logical.my_natural <- function(x, to, ...) {
  # In this case `to` is the logical and `x` is the natural number
  attributes(x) <- NULL
  as.logical(x)
}
#' @export
vec_cast.integer.my_natural <- function(x, to, ...) {
  attributes(x) <- NULL
  as.integer(x)
}
#' @export
vec_cast.double.my_natural <- function(x, to, ...) {
  attributes(x) <- NULL
  as.double(x)
}
# Necessary because includeRmd() evaluated in a child of global
knitr_local_registration("vctrs::vec_cast", "logical.my_natural")
knitr_local_registration("vctrs::vec_cast", "integer.my_natural")
knitr_local_registration("vctrs::vec_cast", "double.my_natural")

And we now get the expected combinations.

vec_c(new_natural(1), 10L)

vec_c(1.5, new_natural(1))


Try the vctrs package in your browser

Any scripts or data that you put into this service are public.

vctrs documentation built on April 29, 2021, 5:08 p.m.