PipeOp Base Class

A PipeOp represents a transformation of a given "input" into a given "output", with two stages: "training" and "prediction". It can be understood as a generalized function that not only has multiple inputs, but also multiple outputs (as well as two stages). The "training" stage is used when training a machine learning pipeline or fitting a statistical model, and the "predicting" stage is then used for making predictions on new data.

To perform training, the $train() function is called which takes inputs and transforms them, while simultaneously storing information in its $state slot. For prediction, the $predict() function is called, where the $state information can be used to influence the transformation of the new data.

A PipeOp is usually used in a Graph object, a representation of a computational graph. It can have multiple input channels—think of these as multiple arguments to a function, for example when averaging different models—, and multiple output channels—a transformation may return different objects, for example different subsets of a Task. The purpose of the Graph is to connect different outputs of some PipeOps to inputs of other PipeOps.

Input and output channel information of a PipeOp is defined in the $input and $output slots; each channel has a name, a required type during training, and a required type during prediction. The $train() and $predict() function are called with a list argument that has one entry for each declared channel (with one exception, see next paragraph). The list is automatically type-checked for each channel against $input and then passed on to the private$.train() or private$.predict() functions. There the data is processed and a result list is created. This list is again type-checked for declared output types of each channel. The length and types of the result list is as declared in $output.

A special input channel name is "...", which creates a vararg channel that takes arbitrarily many arguments, all of the same type. If the $input table contains an "..."-entry, then the input given to $train() and $predict() may be longer than the number of declared input channels.

This class is an abstract base class that all PipeOps being used in a Graph should inherit from, and is not intended to be instantiated.

Format

Abstract R6Class.

Construction

PipeOp$new(id, param_set = ps(), param_vals = list(), input, output, packages = character(0), tags = character(0))

id :: character(1)
Identifier of resulting object. See $id slot.
param_set :: ParamSet | list of expression
Parameter space description. This should be created by the subclass and given to super$initialize(). If this is a ParamSet, it is used as the PipeOp's ParamSet directly. Otherwise it must be a list of expressions e.g. created by alist() that evaluate to ParamSets. These ParamSet are combined using a ParamSetCollection.
param_vals :: named list
List of hyperparameter settings, overwriting the hyperparameter settings given in param_set. The subclass should have its own param_vals parameter and pass it on to super$initialize(). Default list().
input :: data.table with columns name (character), train (character), predict (character)
Sets the $input slot of the resulting object; see description there.
output :: data.table with columns name (character), train (character), predict (character)
Sets the $output slot of the resulting object; see description there.
packages :: character
Set of all required packages for the PipeOp's $train and $predict methods. See $packages slot. Default is character(0).
tags ::character
A set of tags associated with the PipeOp. Tags describe a PipeOp's purpose. Can be used to filter as.data.table(mlr_pipeops). Default is "abstract", indicating an abstract PipeOp.

Internals

PipeOp is an abstract class with abstract functions private$.train() and private$.predict(). To create a functional PipeOp class, these two methods must be implemented. Each of these functions receives a named list according to the PipeOp's input channels, and must return a list (names are ignored) with values in the order of output channels in $output. The private$.train() and private$.predict() function should not be called by the user; instead, a $train() and $predict() should be used. The most convenient usage is to add the PipeOp to a Graph (possibly as singleton in that Graph), and using the Graph's $train() / $predict() methods.

private$.train() and private$.predict() should treat their inputs as read-only. If they are R6 objects, they should be cloned before being manipulated in-place. Objects, or parts of objects, that are not changed, do not need to be cloned, and it is legal to return the same identical-by-reference objects to multiple outputs.

Fields

id :: character
ID of the PipeOp. IDs are user-configurable, and IDs of PipeOps must be unique within a Graph. IDs of PipeOps must not be changed once they are part of a Graph, instead the Graph's $set_names() method should be used.
packages :: character
Packages required for the PipeOp. Functions that are not in base R should still be called using :: (or explicitly attached using require()) in private$.train() and private$.predict(), but packages declared here are checked before any (possibly expensive) processing has started within a Graph.
param_set :: ParamSet
Parameters and parameter constraints. Parameter values that influence the functioning of $train and / or $predict are in the $param_set$values slot; these are automatically checked against parameter constraints in $param_set.
state :: any | NULL
Method-dependent state obtained during training step, and usually required for the prediction step. This is NULL if and only if the PipeOp has not been trained. The $state is the only slot that can be reliably modified during $train(), because private$.train() may theoretically be executed in a different R-session (e.g. for parallelization). $state should furthermore always be set to something with copy-semantics, since it is never cloned. This is a limitation not of PipeOp or mlr3pipelines, but of the way the system as a whole works, together with GraphLearner and mlr3.
input :: data.table with columns name (character), train (character), predict (character)
Input channels of PipeOp. Column name gives the names (and order) of values in the list given to $train() and $predict(). Column train is the (S3) class that an input object must conform to during training, column predict is the (S3) class that an input object must conform to during prediction. Types are checked by the PipeOp itself and do not need to be checked by private$.train() / private$.predict() code.
A special name is "...", which creates a vararg input channel that accepts a variable number of inputs.
If a row has both train and predict values enclosed by square brackets ("[", "]), then this channel is Multiplicity-aware. If the PipeOp receives a Multiplicity value on these channels, this Multiplicity is given to the .train() and .predict() functions directly. Otherwise, the Multiplicity is transparently unpacked and the .train() and .predict() functions are called multiple times, once for each Multiplicity element. The type enclosed by square brackets indicates that only a Multiplicity containing values of this type are accepted. See Multiplicity for more information.
output :: data.table with columns name (character), train (character), predict (character)
Output channels of PipeOp, in the order in which they will be given in the list returned by $train and $predict functions. Column train is the (S3) class that an output object must conform to during training, column predict is the (S3) class that an output object must conform to during prediction. The PipeOp checks values returned by private$.train() and private$.predict() against these types specifications.
If a row has both train and predict values enclosed by square brackets ("[", "]), then this signals that the channel emits a Multiplicity of the indicated type. See Multiplicity for more information.
innum :: numeric(1)
Number of input channels. This equals nrow($input).
outnum :: numeric(1)
Number of output channels. This equals nrow($output).
is_trained :: logical(1)
Indicate whether the PipeOp was already trained and can therefore be used for prediction.
tags ::character
A set of tags associated with the PipeOp. Tags describe a PipeOp's purpose. Can be used to filter as.data.table(mlr_pipeops). PipeOp tags are inherited and child classes can introduce additional tags.
hash :: character(1)
Checksum calculated on the PipeOp, depending on the PipeOp's class and the slots $id and $param_set$values. If a PipeOp's functionality may change depending on more than these values, it should inherit the $hash active binding and calculate the hash as digest(list(super$hash, <OTHER THINGS>), algo = "xxhash64").
phash :: character(1)
Checksum calculated on the PipeOp, depending on the PipeOp's class and the slots $id but ignoring $param_set$values. If a PipeOp's functionality may change depending on more than these values, it should inherit the $hash active binding and calculate the hash as digest(list(super$hash, <OTHER THINGS>), algo = "xxhash64").
.result :: list
If the Graph's $keep_results flag is set to TRUE, then the intermediate Results of $train() and $predict() are saved to this slot, exactly as they are returned by these functions. This is mainly for debugging purposes and done, if requested, by the Graph backend itself; it should not be done explicitly by private$.train() or private$.predict().
man :: character(1)
Identifying string of the help page that shows with help().
properties :: character()
The properties of the pipeop. Currently supported values are:
- "validation": the PipeOp can make use of the $internal_valid_task of an mlr3::Task. This is for example used for PipeOpLearners that wrap a Learner with this property, see mlr3::Learner. PipeOps that have this property, also have a $validate field, which controls whether to use the validation task, as well as a $internal_valid_scores field, which allows to access the internal validation scores after training.
- "internal_tuning": the PipeOp is able to internally optimize hyperparameters. This works analogously to the internal tuning implementation for mlr3::Learner. PipeOps with that property also implement the standardized accessor $internal_tuned_values and have at least one parameter tagged with "internal_tuning". An example for such a PipeOp is a PipeOpLearner that wraps a Learner with the "internal_tuning" property.
Programatic access to all available properties is possible via mlr_reflections$pipeops$properties.

Methods

train(input)
(list) -> named list
Train PipeOp on inputs, transform it to output and store the learned $state. If the PipeOp is already trained, already present $state is overwritten. Input list is typechecked against the $input train column. Return value is a list with as many entries as $output has rows, with each entry named after the $output name column and class according to the $output train column. The workhorse function for training each PipeOp is the private .train(input)
: (named list) -> list
function. It's an Abstract function that must be implemented by concrete subclasses. private$.train() is called by $train() after typechecking. It must change the $state value to something non-NULL and return a list of transformed data according to the $output train column. Names of the returned list are ignored.
The private$.train() method should not be called by a user; instead, the $train() method should be used which does some checking and possibly type conversion.
predict(input)
(list) -> named list
Predict on new data in input, possibly using the stored $state. Input and output are specified by $input and $output in the same way as for $train(), except that the predict column is used for type checking. The workhorse function for predicting in each using each PipeOp is .predict(input)
(named list) -> list
Abstract function that must be implemented by concrete subclasses. private$.predict() is called by $predict() after typechecking and works analogously to private$.train(). Unlike private$.train(), private$.predict() should not modify the PipeOp in any way.
Just as private$.train(), private$.predict() should not be called by a user; instead, the $predict() method should be used.
print()
() -> NULL
Prints the PipeOps most salient information: $id, $is_trained, $param_set$values, $input and $output.
help(help_type)
(character(1)) -> help file
Displays the help file of the concrete PipeOp instance. help_type is one of "text", "html", "pdf" and behaves as the help_type argument of R's help().

Inheriting

To create your own PipeOp, you need to overload the private$.train() and private$.test() functions. It is most likely also necessary to overload the $initialize() function to do additional initialization. The $initialize() method should have at least the arguments id and param_vals, which should be passed on to super$initialize() unchanged. id should have a useful default value, and param_vals should have the default value list(), meaning no initialization of hyperparameters.

If the $initialize() method has more arguments, then it is necessary to also overload the private$.additional_phash_input() function. This function should return either all objects, or a hash of all objects, that can change the function or behavior of the PipeOp and are independent of the class, the id, the $state, and the $param_set$values. The last point is particularly important: changing the $param_set$values should not change the return value of private$.additional_phash_input().

Examples

# example (bogus) PipeOp that returns the sum of two numbers during $train()
# as well as a letter of the alphabet corresponding to that sum during $predict().

PipeOpSumLetter = R6::R6Class("sumletter",
  inherit = PipeOp,  # inherit from PipeOp
  public = list(
    initialize = function(id = "posum", param_vals = list()) {
      super$initialize(id, param_vals = param_vals,
        # declare "input" and "output" during construction here
        # training takes two 'numeric' and returns a 'numeric';
        # prediction takes 'NULL' and returns a 'character'.
        input = data.table::data.table(name = c("input1", "input2"),
          train = "numeric", predict = "NULL"),
        output = data.table::data.table(name = "output",
          train = "numeric", predict = "character")
      )
    }
  ),
  private = list(
    # PipeOp deriving classes must implement .train and
    # .predict; each taking an input list and returning
    # a list as output.
    .train = function(input) {
      sum = input[[1]] + input[[2]]
      self$state = sum
      list(sum)
    },
    .predict = function(input) {
      list(letters[self$state])
    }
  )
)
posum = PipeOpSumLetter$new()

print(posum)
#> PipeOp: <posum> (not trained)
#> values: <list()>
#> Input channels <name [train type, predict type]>:
#>   input1 [numeric,NULL], input2 [numeric,NULL]
#> Output channels <name [train type, predict type]>:
#>   output [numeric,character]

posum$train(list(1, 2))
#> $output
#> [1] 3
#> 
# note the name 'output' is the name of the output channel specified
# in the $output data.table.

posum$predict(list(NULL, NULL))
#> $output
#> [1] "c"
#>