The import package

Introduction

One of the most important aspects of the R ecosystem is the ease with which extensions and new features can be developed and distributed in the form of packages. The main distribution channel is the Comprehensive R Archive Network, from which packages can be installed directly from R. Another popular option is using GitHub repositories from which packages can also be painlessly installed, e.g. using install_github from the devtools package.

The import package provides an alternative approach to using external functionality in R scripts; first however, it is useful to describe the standard approach to clarify how import may serve as improvement. The most common way to include the functionality provided by a package is to use the library function:

library(PackageA)
library(PackageB)

value1 <- function_a(...) # Supposedly this comes from PackageA, 
value2 <- function_b(...) # and this from PackageB, but who knows?!
...

In some situations this is fine; however there are some subtle shortcomings:

  1. Packages are attached and all of their exported objects are exposed,
  2. When using more packages this way, the order in which they are attached can be important,
  3. It quickly becomes unclear to the reader of a script which package certain functionality comes from, and
  4. the terms “library” and “package” are often used incorrectly (although a minor point, it seems to confuse somewhat).

The problem with (1) is that the search path is populated with more objects than are needed and it is not immediately clear whether name clashes will occur. Problem (2) refers to the case where packages export different objects with the same names, say if function_b is exported in both PackageA and PackageB above. In this case the name will point to the object from the package attached last. The earlier exposed objects are said to masked. Even if this is not a problem when writing the script, an update of packages may cause this problem later on when executing the script; and tracking down the resulting errors may be tough and time consuming. Problem (3) may appear unimportant, but it is not to be underestimated. Code snippets are very commonly shared and spending time figuring out where functionality comes from is not a very satisfying nor value-adding activity.

It is possible to unambiguously specify where a function comes from by prefixing it with :: every time it used, but this is often overly verbose and does not provide an easily accessible overview of what external functionality is used in a script. One may also import single exported objects, one at a time, using the (double) “colon syntax”,

function_a <- PackageA::function_a
function_b <- PackageB::function_b

The downside of this approach is that the object is placed in the user’s global work space, rather than being encapsulated somewhere else in the search path (when using library to load pkg, a namespace package:pkg will be attached in the search path which will contain the exported functions from pkg). Another minor point is that one can only import one object at a time using this approach.

While packages form the backbone of code distribution, another option comes in the form of scripts, but these are usually task specific and not commonly used to “bundle” functionality for use in other scripts. In particular, when source is used to include contents from one script in another, once again all objects produced by the script will be “exposed” and may “over populate” the working environment, masking other objects, if not only producing some mental clutter. Scope management is therefore not too comfortable when splitting functionality across files in a modular way.

The import package sets out to improve the way external functionality is included in your code by alleviating some of the concerns raised above by providing an expressive way of importing object from both packages and scripts. The latter provides a bridge between the package approach to distribution and simple stand-alone script files. This allows for the use of scripts as modules, a collection of related object definitions, each of which may be used at different places without exposing more than necessary.

The package is inspired in part by Python’s from some_module import some_function syntax, and solves the two issues raised above. It is also similar to roxygen2s @importFrom package function1 function2 for packages. While import will also work for package development, the intended use case is when using external functions R scripts.

In addition to being able to import objects from packages, import also allows you to import objects from other scripts (i.e. a kind of module). This allows a simple way to distribute and use functionality without the need to write a full package. One example is a Shiny app, where one can place definitions in a script and import only the needed objects where they are used. This avoids workspace clutter and name clashes. For more details see below.

Basic Usage

Importing from Packages

The most basic use case is to import a few functions from package (here the psych package):

import::from(psych, geometric.mean, harmonic.mean)
geometric.mean(trees$Volume)

The imported objects are placed in a separate entity in the search path which by default is named “imports”. It is therefore also easy to get rid of them again with detach("imports"). The main point is that it is clear which functions will be used and where they come from. It’s noteworthy that there is nothing special going on: the import::from function is only a convenient wrapper around getExportedValue (as is :: itself) and assign.

The import package itself should not to be attached (don’t include it via library, you will get a warning). Rather, it is designed to be expressive when using the colon syntax. To import non-exported objects one must use triple-colon syntax: import:::from(pkg, obj). If any of the import functions are called regularly, i.e. without preceding import:: or import:::, an error is raised. If import is attached, a startup message will inform that import should not be attached.

If one of the function names conflicts with an existing function (such as filter from the dplyr package) it is simple to rename it:

import::from(dplyr, select, arrange, keep_when = filter)
keep_when(mtcars, hp>250)

This does pretty much what it says: three functions are imported from dplyr, two of which will keep their original name, and one which is renamed, e.g. to avoid name clash with stats::filter.

You can use .all=TRUE to import all functions from a package, but rename one of them:

import::from(dplyr, keep_when = filter, .all=TRUE)

To omit a function from the import, use .except (which takes a character vector):

import::from(dplyr, .except=c("filter", "lag"))

Note that import tries to be smart about this and assumes that if you are using the .except parameter, you probably want to import everything you are not explicitly omitting, and sets the .all parameter to TRUE. You can override this in exceptional cases, but you seldom need to.

Finally, a more complex example, combining a few different import statements:

import::from(magrittr, "%>%")
import::from(dplyr, starwars, select, mutate, keep_when = filter)
import::from(tidyr, unnest)
import::from(broom, tidy)

ready_data <-
  starwars %>% 
  keep_when(mass < 100) %>% 
  select(name, height, mass, films) %>%
  unnest(films) %>%
  mutate( log_mass = log(mass), films=factor(films))

linear_model <- 
  lm(log_mass ~ height + films, data = ready_data) %>% 
  tidy

In the above, it is clear which package provides which functions (one could e.g. otherwise be tempted to think that tidy belonged to tidyr). Note that ordering is irrelevant, even if tidyr at some point exposes a function tidy after an update, as import is explicit about importing.

It also shows that one can import multiple objects in a single statement, and even rename objects if desired; for example, in the above one can imagine that filter from stats is needed later on, and so dplyr’s filter is renamed to avoid confusion. Sometimes, it is not at all clear what purpose a package has; e.g. the name magrittr does not immediately reveal that it’s main purpose is to provide the pipe operator, %>%.

Importing Functions from “Module” Scripts

The import package allows for importing objects defined in script files, which we will here refer to as “modules”. The module will be fully evaluated by import when an import is requested, after which objects such as functions or data can be imported. Such modules should be side-effect free, but this is not enforced.

Attachments are detached (e.g. packages attached by library) but loaded namespaces remain loaded. This means that values created by functions in an attached namespace will work with import, but functions to be exported should not rely on such functions (use function importing in the modules instead).

For example, the file sequence_module.R contains several functions calculating terms of mathematical sequences. It is possible to import from such files, just as one imports from packages:

import::from(sequence_module.R, fibonacci, square, triangular)

Renaming, the .all, and the .except parameters work in the same way as for packages:

import::from(sequence_module.R, fib=fibonacci, .except="square")

If a module is modified, import will realize this and reload the script if further imports are executed or re-executed; otherwise additional imports will not cause the script to be reloaded for efficiency. As the script is loaded in its own environment (maintained by import) dependencies are kept (except those exposed through attachment), as the following small example shows.

Contents of “some_module.R”:

## Do not use library() inside a module. This results in a warning, 
## and functions relying on ggplot2 will not work.
#library(ggplot2)

## This is also not recommended, because it is not clear wether recursively 
## imported functions should be available after the module is imported
#import::here(qplot, .from = ggplot2)

## This is the recommended way to recursively import functions on which
## module functions depend. The qplot function will be available to 
## module functions, but will not itself be available after import
import::here(qplot, .from = ggplot2)

## Note this operator overload is not something you want to `source`!
`+` <- function(e1, e2)
  paste(e1, e2)

## Some function relying on the above overload:
a <- function(s1, s2)
  s1 + rep(s2, 3)

## Another value.
b <- head(iris, 10)

## A value created using a recursively imported function
p <- qplot(Sepal.Length, Sepal.Width, data = iris, color = Species)

## A function relying on a function exposed through attachment:
plot_it <- function()
  qplot(Sepal.Length, Sepal.Width, data = iris, color = Species)

Usage:

import::from(some_module.R, a, b, p, plot_it)

## Works:
a("cool", "import")

## The `+` is not affecting anything here, so this won't work:
# "cool" + "import"

# Works:
b
p
plot_it()

Suppose that you have some related functionality that you wish to bundle, and that authoring a full package seems excessive or inappropriate for the specific task, for example bundling related user interface components for a shiny application. One option with import is to author a module (script), say as outlined below:

# File: foo.R
# Desc: Functionality related to foos.
# Imports from other_resources.R
# When recursively importing from another module or package for use by 
# your module functions, you should always use import::here() rather 
# than import::from() or library()
import::here(fun_a, fun_b, .from = "other_resources.R")

internal_fun <- function(...) ...

fun_c <- function(...) 
{
  ...
  a <- fun_a(...)
  i <- internal_fun(...)
  ...
}

fun_d <- function(...) ...

Then in another file we need fun_c:

# File: bar.R
# Desc: Functionality related to bars. 
# Imports from foo.R
import::here(fun_c, .from = "foo.R")
...

In the above, only fun_c is visible inside bar.R. The functions on which it depends exist, but are not exposed. Also, note that imported scripts may themselves import.

Since the desired effect of import::from inside a module script is ambiguous, this results in a warning (but the functions will still be imported into the local environment of the script, just as with import::here which only imports are only exposed to the module itself.

When importing from a module, it is sourced into an environment managed by import, and will not be sourced again upon subsequent imports (unless the file has changed). For example, in a shiny application, importing some objects in server.R and others in ui.R from the same module will not cause it to be sourced twice.

Choosing where import looks for packages or modules

The import package will by default use the current set of library paths, i.e. the result of .libPaths(). It is, however, possible to specify a different set of library paths using the .library argument in any of the import functions, for example to import packages installed in a custom location, or to remove any ambiguity as to where imports come from.

Note that in versions up to and including 1.3.0 this defaulted to use only the first entry in the library paths, i.e. .library=.libPaths()[1L]. We believe the new default is applicable in a broader set of circumstances, but if this change causes any issues, we would very much appreciate hearing about it.

When importing from a module (.R file), the directory where import looks for the module script can be specified with the with .directory parameter. The default is . (the current working directory).

Choosing where the imported functions are placed

One can also specify which names to use in the search path and use several to group imports. Names can be specified either as character literals or as variables of type character (for example if the environment needs to be determined dynamically).

import::from(magrittr, "%>%", "%$%", .into = "operators")
import::from(dplyr, arrange, .into = "datatools")
import::from(psych, describe, .into=month.name[1]) # Uses env: "January"

The import::into and import::from accept the same parameters and achieve the same result. The the choice between them a matter of preference). If using custom search path entities actively, one might prefer the alternative syntax (which does the same but reverses the argument order):

import::into("operators", "%>%", "%$%", .from = magrittr)
import::into("datatools", arrange, .from = dplyr)
import::into(month.name[1], describe, .from=psych)

Be aware that beginning in version 1.3.0 hidden objects (those with names prefixed by a period) are supported. Take care to avoid name clashes with argument names.

If it is desired to import objects directly into the current environment, this can be accomplished by import::here. This is particularly useful when importing inside a function definition, or module scripts as described here.

import::here("%>%", "%$%", .from = magrittr)
import::here(arrange, .from = dplyr)

Instead of specifying a named environment on the search path, by passing a character to the .into parameter, it is possible to directly specify an environment. The function automatically determines which use case is involved, based on the mode() of the .into parameter (either character or environment).

Prior to version 1.3.0, non-standard evaluation (NSE) was applied to the .into parameter, and it was necessary to surround it with {}, in order for it to be treated as an environment. This is no longer needed, although it is still allowed (curly brackets are simply ignored).

Examples include:

# Import into the local environment
import::into(environment(), "%>%", .from = magrittr)

# Import into the global environment, curlies are optional
import::into({.GlobalEnv}, "%>%", "%$%", .from = magrittr)

# Import into a new environment, mainly useful for python-style imports
# (see below)
x = import::into(new.env(), "%<>%", .from = magrittr)

Advanced usage

Advanced usage and the .character_only parameter

The import package uses non-standard evaluation (NSE) on the .from and ... parameters, allowing the names of packages and functions to be listed without quoting them. This makes some common use-cases very straightforward, but can get in the way of more programmatic usages.

This is where the .character_only parameter comes in handy. By setting .character_only=TRUE, the non-standard evaluation of the .from and the ... parameters is disabled. Instead, the parameters are processed as character vectors containing the relevant values.

(Previously, NSE was also applied to the .into parameter, but as of version 1.3.0 this is no longer the case, and all parameters except .from and ... are always evaluated in a standard way.)

It is useful to examine some examples of how specifying .character_only=TRUE can be helpful.

Programmatic selection of objects to import

It is not always know in advance which objects to import from a given package. For example, assume we have a list of objects from the broom package that we need to import, we can do it as follows:

objects <- c("tidy", "glance", "augment")
import::from("broom", objects, .character_only=TRUE)

This will import the three functions specified in the objects vector. It is worth noting that because .character_only disables non-standard evaluation on all parameters, the name of the package must now be quoted.

One common use case is when one wants to import all objects except one or a few, because of conflicts with other packages. Should one, for example, want to use the stats versions of the filter() and lag() functions, but import all the other functions in the dplyr package, one could do it like this:

objects <- setdiff(getNamespaceExports("dplyr"), c("filter","lag"))
import::from("dplyr", objects, .character_only=TRUE)

Programmatic selection of module location

The same approach can be used when the directory of the source file for a module is not known in advance. This can be useful when the original source file is not always run with the original working directory, but one still does not want to specify a hard-coded absolute path, but to determine it at run time:

mymodule <- file.path(mypath, "module.R")
import::from(mymodule, "myfunction", .character_only=TRUE)

Again, note that now the name of the function must be quoted because non-standard evaluation is disabled on all parameters.

The here package is useful in many circumstances like this; it allows the setting of a “root” directory for a project and by using the here::here() function to figure out the correct directory, regardless of the working directory.

import::from(here::here("src/utils/module.R")), "myfunction", .character_only=TRUE)

Alternatively, if the file name is always the same and it is only the directory that differs, you could use the .directory parameter, which always expects standard evaluation arguments.

import::from(module.R, "myfunction", here::here("src/utils"))

Note that here::here() has no relation to import::here() despite the similarity in names.

Importing from a URL

Another case where .character_only comes in handy is when one wants to import some functions from a URL. While import does not allow direct importing from a URL (because of difficult questions about when a URL target has changes, whether to download a file and other things), it easy to achieve the desired result by using the pins package (whose main purpose is to resolve such difficult questions). A simple example follows, which directly imports the myfunc() function, which is defined in the plusone_module.R:

url <- "https://raw.githubusercontent.com/rticulate/import/master/man/examples/plusone_module.R"
import::from(pins::pin(url), "myfunc", .character_only=TRUE)
myfunc(3)
#> [1] 4

Python-like imports

A frequent pattern in python imports packages under an alias; all subsequent use of the imported objects then explicitly includes the alias:

import pandas as pd
import numpy as np
import math as m

print(m.pi)
print(m.e)

In order to achieve this functionality with the import package, use .into={new.env()} which assign to a new environment without attaching it. import::from() returns this environment, so it can be assigned to a variable:

# Import into a new namespace, use $ to access
td <- import::from(tidyr, spread, pivot_wider, .into={new.env()})
dp <- import::from(dplyr, .all=TRUE, .into={new.env()})
dp$select(head(cars),dist)
#>   dist
#> 1    2
#> 2   10
#> 3    4
#> 4   22
#> 5   16
#> 6   10

# Note that functions are not visible without dp$ prefix
select(head(cars),dist)
#> Error in select(head(cars), dist): could not find function "select"

Importing S3 methods

lifecyclelifecycleexperimentalexperimental

S3 methods work well in local context, but when the method is called from a different environment, it must be registered (in packages, this is done in the NAMESPACE file). import can now register methods of form generic.class or generic.class.name automatically using the new .S3 argument. By specifying .S3=TRUE, import will automatically detect methods for existing or new generics. No need to export and/or register them manually!

Consider following script foo.r with a generic and two methods:

# foo.r
# functions with great foonctionality
foo = function(x){
  UseMethod("foo", x)
}

foo.numeric <- function(x){
  x + 1
}

foo.character <- function(x){
  paste0("_", x, "_")
}

Now, all we need is to import the foo generic:

import::from("foo.r", foo, .S3=TRUE)

foo(0) # 1
foo("bar") # _bar_

This is an experimental feature. We think it should work well and you are encouraged to use it and report back – but the syntax and semantics may change in the future to improve the feature.