# 5 Pipelines

mlr3pipelines is a dataflow programming toolkit. This chapter focuses on the applicant’s side of the package. A more in-depth and technically oriented guide can be found in the In-depth look into mlr3pipelines chapter.

Machine learning workflows can be written as directed “Graphs”/“Pipelines” that represent data flows between preprocessing, model fitting, and ensemble learning units in an expressive and intuitive language. We will most often use the term “Graph” in this manual but it can interchangeably be used with “pipeline” or “workflow”.

Below you can examine an example for such a graph:

Single computational steps can be represented as so-called PipeOps, which can then be connected with directed edges in a Graph. The scope of mlr3pipelines is still growing. Currently supported features are:

• Data manipulation and preprocessing operations, e.g. PCA, feature filtering, imputation
• Task subsampling for speed and outcome class imbalance handling
• mlr3 Learner operations for prediction and stacking
• Ensemble methods and aggregation of predictions

Additionally, we implement several meta operators that can be used to construct powerful pipelines:

• Simultaneous path branching (data going both ways)
• Alternative path branching (data going one specific way, controlled by hyperparameters)

An extensive introduction to creating custom PipeOps (PO’s) can be found in the technical introduction.

Using methods from mlr3tuning, it is even possible to simultaneously optimize parameters of multiple processing units.

A predecessor to this package is the mlrCPO package, which works with mlr 2.x. Other packages that provide, to varying degree, some preprocessing functionality or machine learning domain specific language, are:

An example for a Pipeline that can be constructed using mlr3pipelines is depicted below:

## 5.3 Nodes, Edges and Graphs

POs are combined into Graphs. The manual way (= hard way) to construct a Graph is to create an empty graph first. Then one fills the empty graph with POs, and connects edges between the POs. Conceptually, this may look like this:

POs are identified by their $id. Note that the operations all modify the object in-place and return the object itself. Therefore, multiple modifications can be chained. For this example we use the pca PO defined above and a new PO named “mutate”. The latter creates a new feature from existing variables. Additionally, we use the filter PO again. mutate = po("mutate") filter = po("filter", filter = mlr3filters::flt("variance"), param_vals = list(filter.frac = 0.5)) graph = Graph$new()$add_pipeop(mutate)$
add_pipeop(filter)$add_edge("mutate", "variance") # add connection mutate -> filter The much quicker way is to use the %>>% operator to chain POs or Graph s. The same result as above can be achieved by doing the following: graph = mutate %>>% filter Now the Graph can be inspected using its $plot() function:

graph$plot() Chaining multiple POs of the same kind If multiple POs of the same kind should be chained, it is necessary to change the id to avoid name clashes. This can be done by either accessing the $id slot or during construction:

graph$add_pipeop(po("pca")) graph$add_pipeop(po("pca", id = "pca2"))

## 5.4 Modeling

The main purpose of a Graph is to build combined preprocessing and model fitting pipelines that can be used as mlr3 Learner.

Conceptually, the process may be summarized as follows:

In the following we chain two preprocessing tasks:

• mutate (creation of a new feature)
• filter (filtering the dataset)

Subsequently one can chain a PO learner to train and predict on the modified dataset.

mutate = po("mutate")
filter = po("filter",
filter = mlr3filters::flt("variance"),
param_vals = list(filter.frac = 0.5))

graph = mutate %>>%
filter %>>%
po("learner",
learner = lrn("classif.rpart"))

Until here we defined the main pipeline stored in Graph. Now we can train and predict the pipeline:

task = tsk("iris")
graph$train(task) ##$classif.rpart.output
## NULL
graph$predict(task) ##$classif.rpart.output
## <PredictionClassif> for 150 observations:
##     row_ids     truth  response
##           1    setosa    setosa
##           2    setosa    setosa
##           3    setosa    setosa
## ---
##         148 virginica virginica
##         149 virginica virginica
##         150 virginica virginica

glrn$param_set$values$variance.filter.frac = 0.25 cv3 = rsmp("cv", folds = 3) resample(task, glrn, cv3) ## <ResampleResult> of 3 iterations ## * Task: iris ## * Learner: mutate.variance.classif.rpart ## * Warnings: 0 in 0 iterations ## * Errors: 0 in 0 iterations ### 5.4.2 Tuning If you are unfamiliar with tuning in mlr3, we recommend to take a look at the section about tuning first. Here we define a ParamSet for the “rpart” learner and the “variance” filter which should be optimized during the tuning process. library("paradox") ps = ps( classif.rpart.cp = p_dbl(lower = 0, upper = 0.05), variance.filter.frac = p_dbl(lower = 0.25, upper = 1) ) After having defined the PerformanceEvaluator, a random search with 10 iterations is created. For the inner resampling, we are simply using holdout (single split into train/test) to keep the runtimes reasonable. library("mlr3tuning") instance = TuningInstanceSingleCrit$new(
learner = glrn,
resampling = rsmp("holdout"),
measure = msr("classif.ce"),
search_space = ps,
terminator = trm("evals", n_evals = 20)
)
tuner = tnr("random_search")
tuner$optimize(instance) The tuning result can be found in the respective result slots. instance$result_learner_param_vals
instance$result_y ## 5.5 Non-Linear Graphs The Graphs seen so far all have a linear structure. Some POs may have multiple input or output channels. These channels make it possible to create non-linear Graphs with alternative paths taken by the data. Possible types are: • Branching: Splitting of a node into several paths, e.g. useful when comparing multiple feature-selection methods (pca, filters). Only one path will be executed. • Copying: Splitting of a node into several paths, all paths will be executed (sequentially). Parallel execution is not yet supported. • Stacking: Single graphs are stacked onto each other, i.e. the output of one Graph is the input for another. In machine learning this means that the prediction of one Graph is used as input for another Graph ### 5.5.1 Branching & Copying The PipeOpBranch and PipeOpUnbranch POs make it possible to specify multiple alternative paths. Only one path is actually executed, the others are ignored. The active path is determined by a hyperparameter. This concept makes it possible to tune alternative preprocessing paths (or learner models). Below a conceptual visualization of branching: PipeOp(Un)Branch is initialized either with the number of branches, or with a character-vector indicating the names of the branches. If names are given, the “branch-choosing” hyperparameter becomes more readable. In the following, we set three options: 1. Doing nothing (“nop”) 2. Applying a PCA 3. Scaling the data It is important to “unbranch” again after “branching”, so that the outputs are merged into one result objects. In the following we first create the branched graph and then show what happens if the “unbranching” is not applied: graph = po("branch", c("nop", "pca", "scale")) %>>% gunion(list( po("nop", id = "null1"), po("pca"), po("scale") )) Without “unbranching” one creates the following graph: graph$plot(html = FALSE)

Now when “unbranching”, we obtain the following results:

(graph %>>% po("unbranch", c("nop", "pca", "scale")))$plot(html = FALSE) The same can be achieved using a shorter notation: # List of pipeops opts = list(po("nop", "no_op"), po("pca"), po("scale")) # List of po ids opt_ids = mlr3misc::map_chr(opts, [[, "id") po("branch", options = opt_ids) %>>% gunion(opts) %>>% po("unbranch", options = opt_ids) ## Graph with 5 PipeOps: ## ID State sccssors prdcssors ## branch <<UNTRAINED>> no_op,pca,scale ## no_op <<UNTRAINED>> unbranch branch ## pca <<UNTRAINED>> unbranch branch ## scale <<UNTRAINED>> unbranch branch ## unbranch <<UNTRAINED>> no_op,pca,scale ### 5.5.2 Model Ensembles We can leverage the different operations presented to connect POs. This allows us to form powerful graphs. Before we go into details, we split the task into train and test indices. task = tsk("iris") train.idx = sample(seq_len(task$nrow), 120)
test.idx = setdiff(seq_len(task$nrow), train.idx) #### 5.5.2.1 Bagging We first examine Bagging introduced by . The basic idea is to create multiple predictors and then aggregate those to a single, more powerful predictor. “… multiple versions are formed by making bootstrap replicates of the learning set and using these as new learning sets” Bagging then aggregates a set of predictors by averaging (regression) or majority vote (classification). The idea behind bagging is, that a set of weak, but different predictors can be combined in order to arrive at a single, better predictor. We can achieve this by downsampling our data before training a learner, repeating this e.g. 10 times and then performing a majority vote on the predictions. Graphically, it may be summarized as follows: First, we create a simple pipeline, that uses PipeOpSubsample before a PipeOpLearner is trained: single_pred = po("subsample", frac = 0.7) %>>% po("learner", lrn("classif.rpart")) We can now copy this operation 10 times using pipeline_greplicate. The pipeline_greplicate allows us to parallelize many copies of an operation by creating a Graph containing n copies of the input Graph. We can also create it using syntactic sugar via ppl(): pred_set = ppl("greplicate", single_pred, 10L) Afterwards we need to aggregate the 10 pipelines to form a single model: bagging = pred_set %>>% po("classifavg", innum = 10) Now we can plot again to see what happens: bagging$plot(html = FALSE)

This pipeline can again be used in conjunction with GraphLearner in order for Bagging to be used like a Learner:

baglrn = as_learner(bagging)
baglrn$train(task, train.idx) baglrn$predict(task, test.idx)
## <PredictionClassif> for 30 observations:
##     row_ids     truth  response prob.setosa prob.versicolor prob.virginica
##           8    setosa    setosa           1             0.0            0.0
##           9    setosa    setosa           1             0.0            0.0
##          18    setosa    setosa           1             0.0            0.0
## ---
##         134 virginica virginica           0             0.5            0.5
##         146 virginica virginica           0             0.0            1.0
##         147 virginica virginica           0             0.0            1.0

In conjunction with different Backends, this can be a very powerful tool. In cases when the data does not fully fit in memory, one can obtain a fraction of the data for each learner from a DataBackend and then aggregate predictions over all learners.

#### 5.5.2.2 Stacking

Stacking is another technique that can improve model performance. The basic idea behind stacking is the use of predictions from one model as features for a subsequent model to possibly improve performance.

Below an conceptual illustration of stacking:

As an example we can train a decision tree and use the predictions from this model in conjunction with the original features in order to train an additional model on top.

To limit overfitting, we additionally do not predict on the original predictions of the learner. Instead, we predict on out-of-bag predictions. To do all this, we can use PipeOpLearnerCV .

PipeOpLearnerCV performs nested cross-validation on the training data, fitting a model in each fold. Each of the models is then used to predict on the out-of-fold data. As a result, we obtain predictions for every data point in our input data.

We first create a “level 0” learner, which is used to extract a lower level prediction. Additionally, we clone() the learner object to obtain a copy of the learner. Subsequently, one sets a custom id for the PipeOp .

lrn = lrn("classif.rpart")
lrn_0 = po("learner_cv", lrn$clone()) lrn_0$id = "rpart_cv"

We use PipeOpNOP in combination with gunion, in order to send the unchanged Task to the next level. There it is combined with the predictions from our decision tree learner.

level_0 = gunion(list(lrn_0, po("nop")))

Afterwards, we want to concatenate the predictions from PipeOpLearnerCV and the original Task using PipeOpFeatureUnion :

combined = level_0 %>>% po("featureunion", 2)

Now we can train another learner on top of the combined features:

stack = combined %>>% po("learner", lrn$clone()) stack$plot(html = FALSE)
stacklrn = as_learner(stack)
stacklrn$train(task, train.idx) stacklrn$predict(task, test.idx)
## <PredictionClassif> for 30 observations:
##     row_ids     truth   response
##           8    setosa     setosa
##           9    setosa     setosa
##          18    setosa     setosa
## ---
##         134 virginica versicolor
##         146 virginica  virginica
##         147 virginica  virginica

In this vignette, we showed a very simple use-case for stacking. In many real-world applications, stacking is done for multiple levels and on multiple representations of the dataset. On a lower level, different preprocessing methods can be defined in conjunction with several learners. On a higher level, we can then combine those predictions in order to form a very powerful model.

#### 5.5.2.3 Multilevel Stacking

In order to showcase the power of mlr3pipelines, we will show a more complicated stacking example.

In this case, we train a glmnet and 2 different rpart models (some transform its inputs using PipeOpPCA ) on our task in the “level 0” and concatenate them with the original features (via gunion). The result is then passed on to “level 1”, where we copy the concatenated features 3 times and put this task into an rpart and a glmnet model. Additionally, we keep a version of the “level 0” output (via PipeOpNOP) and pass this on to “level 2”. In “level 2” we simply concatenate all “level 1” outputs and train a final decision tree.

In the following examples, use <lrn>$param_set$values$<param_name> = <param_value> to set hyperparameters for the different learner. library("magrittr") library("mlr3learners") # for classif.glmnet rprt = lrn("classif.rpart", predict_type = "prob") glmn = lrn("classif.glmnet", predict_type = "prob") # Create Learner CV Operators lrn_0 = po("learner_cv", rprt, id = "rpart_cv_1") lrn_0$param_set$values$maxdepth = 5L
lrn_1 = po("pca", id = "pca1") %>>% po("learner_cv", rprt, id = "rpart_cv_2")
lrn_1$param_set$values$rpart_cv_2.maxdepth = 1L lrn_2 = po("pca", id = "pca2") %>>% po("learner_cv", glmn) # Union them with a PipeOpNULL to keep original features level_0 = gunion(list(lrn_0, lrn_1, lrn_2, po("nop", id = "NOP1"))) # Cbind the output 3 times, train 2 learners but also keep level # 0 predictions level_1 = level_0 %>>% po("featureunion", 4) %>>% po("copy", 3) %>>% gunion(list( po("learner_cv", rprt, id = "rpart_cv_l1"), po("learner_cv", glmn, id = "glmnt_cv_l1"), po("nop", id = "NOP_l1") )) # Cbind predictions, train a final learner level_2 = level_1 %>>% po("featureunion", 3, id = "u2") %>>% po("learner", rprt, id = "rpart_l2") # Plot the resulting graph level_2$plot(html = FALSE)
task = tsk("iris")
lrn = as_learner(level_2)

And we can again call .$train and .$predict:

lrn$train(task, train.idx)$
predict(task, test.idx)$score() ## classif.ce ## 0.1 ## 5.6 Special Operators This section introduces some special operators, that might be useful in numerous further applications. ### 5.6.1 Imputation: PipeOpImpute An often occurring setting is the imputation of missing data. Imputation methods range from relatively simple imputation using either mean, median or histograms to way more involved methods including using machine learning algorithms in order to predict missing values. The following PipeOps, PipeOpImpute: • Impute numeric values from a histogram • Adds a new level for factors • Add a column marking whether a value for a given feature was missing or not (numeric only) • We use po("featureunion") to cbind the missing indicator features. pom = po("missind") pon = po("imputehist", id = "imputer_num", affect_columns = is.numeric) pof = po("imputeoor", id = "imputer_fct", affect_columns = is.factor) imputer = pom %>>% pon %>>% pof A learner can thus be equipped with automatic imputation of missing values by adding an imputation Pipeop. polrn = po("learner", lrn("classif.rpart")) lrn = as_learner(imputer %>>% polrn) ### 5.6.2 Feature Engineering: PipeOpMutate New features can be added or computed from a task using PipeOpMutate . The operator evaluates one or multiple expressions provided in an alist. In this example, we compute some new features on top of the iris task. Then we add them to the data as illustrated below: pom = po("mutate") # Define a set of mutations mutations = list( Sepal.Sum = ~ Sepal.Length + Sepal.Width, Petal.Sum = ~ Petal.Length + Petal.Width, Sepal.Petal.Ratio = ~ (Sepal.Length / Petal.Length) ) pom$param_set$values$mutation = mutations

If outside data is required, we can make use of the env parameter. Moreover, we provide an environment, where expressions are evaluated (env defaults to .GlobalEnv).

### 5.6.3 Training on data subsets: PipeOpChunk

In cases, where data is too big to fit into the machine’s memory, an often-used technique is to split the data into several parts. Subsequently, the parts are trained on each part of the data.

After undertaking these steps, we aggregate the models. In this example, we split our data into 4 parts using PipeOpChunk . Additionally, we create 4 PipeOpLearner POS, which are then trained on each split of the data.

chks = po("chunk", 4)
lrns = ppl("greplicate", po("learner", lrn("classif.rpart")), 4)

Afterwards we can use PipeOpClassifAvg to aggregate the predictions from the 4 different models into a new one.

mjv = po("classifavg", 4)

We can now connect the different operators and visualize the full graph:

pipeline = chks %>>% lrns %>>% mjv
pipeline$plot(html = FALSE) task = tsk("iris") train.idx = sample(seq_len(task$nrow), 120)
test.idx = setdiff(seq_len(task$nrow), train.idx) pipelrn = as_learner(pipeline) pipelrn$train(task, train.idx)$predict(task, train.idx)$
score()
## classif.ce
##        0.3

### 5.6.4 Feature Selection: PipeOpFilter and PipeOpSelect

The package mlr3filters contains many different mlr3filters::Filters that can be used to select features for subsequent learners. This is often required when the data has a large amount of features.

A PipeOp for filters is PipeOpFilter:

po("filter", mlr3filters::flt("information_gain"))
## PipeOp: <information_gain> (not trained)
## values: <list()>
## Input channels <name [train type, predict type]>:
## Output channels <name [train type, predict type]>:
##   output [Task,Task]

How many features to keep can be set using filter_nfeat, filter_frac and filter_cutoff.

Filters can be selected / de-selected by name using PipeOpSelect.

## 5.7 In-depth look into mlr3pipelines

This vignette is an in-depth introduction to mlr3pipelines, the dataflow programming toolkit for machine learning in R using mlr3. It will go through basic concepts and then give a few examples that both show the simplicity as well as the power and versatility of using mlr3pipelines.

### 5.7.1 What’s the Point

Machine learning toolkits often try to abstract away the processes happening inside machine learning algorithms. This makes it easy for the user to switch out one algorithm for another without having to worry about what is happening inside it, what kind of data it is able to operate on etc. The benefit of using mlr3, for example, is that one can create a Learner, a Task, a Resampling etc. and use them for typical machine learning operations. It is trivial to exchange individual components and therefore use, for example, a different Learner in the same experiment for comparison.

task = as_task_classif(iris, target = "Species")
lrn = lrn("classif.rpart")
rsmp = rsmp("holdout")
resample(task, lrn, rsmp)
## <ResampleResult> of 1 iterations
## * Learner: classif.rpart
## * Warnings: 0 in 0 iterations
## * Errors: 0 in 0 iterations

However, this modularity breaks down as soon as the learning algorithm encompasses more than just model fitting, like data preprocessing, ensembles or other meta models. mlr3pipelines takes modularity one step further than mlr3: it makes it possible to build individual steps within a Learner out of building blocks called PipeOps.

### 5.7.2PipeOp: Pipeline Operators

The most basic unit of functionality within mlr3pipelines is the PipeOp, short for “pipeline operator”, which represents a trans-formative operation on input (for example a training dataset) leading to output. It can therefore be seen as a generalized notion of a function, with a certain twist: PipeOps behave differently during a “training phase” and a “prediction phase”. The training phase will typically generate a certain model of the data that is saved as internal state. The prediction phase will then operate on the input data depending on the trained model.

An example of this behavior is the principal component analysis operation (“PipeOpPCA”): During training, it will transform incoming data by rotating it in a way that leads to uncorrelated features ordered by their contribution to total variance. It will also save the rotation matrix to be used during for new data. This makes it possible to perform “prediction” with single rows of new data, where a row’s scores on each of the principal components (the components of the training data!) is computed.

po = po("pca")
po$train(list(task))[[1]]$data()
##        Species    PC1      PC2      PC3       PC4
##   1:    setosa -2.684  0.31940 -0.02791 -0.002262
##   2:    setosa -2.714 -0.17700 -0.21046 -0.099027
##   3:    setosa -2.889 -0.14495  0.01790 -0.019968
##   4:    setosa -2.745 -0.31830  0.03156  0.075576
##   5:    setosa -2.729  0.32675  0.09008  0.061259
##  ---
## 146: virginica  1.944  0.18753  0.17783 -0.426196
## 147: virginica  1.527 -0.37532 -0.12190 -0.254367
## 148: virginica  1.764  0.07886  0.13048 -0.137001
## 149: virginica  1.901  0.11663  0.72325 -0.044595
## 150: virginica  1.390 -0.28266  0.36291  0.155039
single_line_task = task$clone()$filter(1)
po$predict(list(single_line_task))[[1]]$data()
##    Species    PC1    PC2      PC3       PC4
## 1:  setosa -2.684 0.3194 -0.02791 -0.002262
po$state ## Standard deviations (1, .., p=4): ## [1] 2.0563 0.4926 0.2797 0.1544 ## ## Rotation (n x k) = (4 x 4): ## PC1 PC2 PC3 PC4 ## Petal.Length 0.85667 -0.17337 0.07624 0.4798 ## Petal.Width 0.35829 -0.07548 0.54583 -0.7537 ## Sepal.Length 0.36139 0.65659 -0.58203 -0.3155 ## Sepal.Width -0.08452 0.73016 0.59791 0.3197 This shows the most important primitives incorporated in a PipeOp: * $train(), taking a list of input arguments, turning them into a list of outputs, meanwhile saving a state in $state * $predict(), taking a list of input arguments, turning them into a list of outputs, making use of the saved $state * $state, the “model” trained with $train() and utilized during $predict().

Schematically we can represent the PipeOp like so:

#### 5.7.2.1 Why the $state It is important to take a moment and notice the importance of a $state variable and the $train() / $predict() dichotomy in a PipeOp. There are many preprocessing methods, for example scaling of parameters or imputation, that could in theory just be applied to training data and prediction / validation data separately, or they could be applied to a task before resampling is performed. This would, however, be fallacious:

• The preprocessing of each instance of prediction data should not depend on the remaining prediction dataset. A prediction on a single instance of new data should give the same result as prediction performed on a whole dataset.
• If preprocessing is performed on a task before resampling is done, information about the test set can leak into the training set. Resampling should evaluate the generalization performance of the entire machine learning method, therefore the behavior of this entire method must only depend only on the content of the training split during resampling.

Each PipeOp is an instance of an “R6” class, many of which are provided by the mlr3pipelines package itself. They can be constructed explicitly (“PipeOpPCA$new()”) or retrieved from the mlr_pipeops dictionary: po("pca"). The entire list of available PipeOps, and some meta-information, can be retrieved using as.data.table(): as.data.table(mlr_pipeops)[, c("key", "input.num", "output.num")] ## key input.num output.num ## 1: boxcox 1 1 ## 2: branch 1 NA ## 3: chunk 1 NA ## 4: classbalancing 1 1 ## 5: classifavg NA 1 ## 6: classweights 1 1 ## 7: colapply 1 1 ## 8: collapsefactors 1 1 ## 9: colroles 1 1 ## 10: copy 1 NA ## 11: datefeatures 1 1 ## 12: encode 1 1 ## 13: encodeimpact 1 1 ## 14: encodelmer 1 1 ## 15: featureunion NA 1 ## 16: filter 1 1 ## 17: fixfactors 1 1 ## 18: histbin 1 1 ## 19: ica 1 1 ## 20: imputeconstant 1 1 ## 21: imputehist 1 1 ## 22: imputelearner 1 1 ## 23: imputemean 1 1 ## 24: imputemedian 1 1 ## 25: imputemode 1 1 ## 26: imputeoor 1 1 ## 27: imputesample 1 1 ## 28: kernelpca 1 1 ## 29: learner 1 1 ## 30: learner_cv 1 1 ## 31: missind 1 1 ## 32: modelmatrix 1 1 ## 33: multiplicityexply 1 NA ## 34: multiplicityimply NA 1 ## 35: mutate 1 1 ## 36: nmf 1 1 ## 37: nop 1 1 ## 38: ovrsplit 1 1 ## 39: ovrunite 1 1 ## 40: pca 1 1 ## 41: proxy NA 1 ## 42: quantilebin 1 1 ## 43: randomprojection 1 1 ## 44: randomresponse 1 1 ## 45: regravg NA 1 ## 46: removeconstants 1 1 ## 47: renamecolumns 1 1 ## 48: replicate 1 1 ## 49: scale 1 1 ## 50: scalemaxabs 1 1 ## 51: scalerange 1 1 ## 52: select 1 1 ## 53: smote 1 1 ## 54: spatialsign 1 1 ## 55: subsample 1 1 ## 56: targetinvert 2 1 ## 57: targetmutate 1 2 ## 58: targettrafoscalerange 1 2 ## 59: textvectorizer 1 1 ## 60: threshold 1 1 ## 61: tunethreshold 1 1 ## 62: unbranch NA 1 ## 63: vtreat 1 1 ## 64: yeojohnson 1 1 ## key input.num output.num When retrieving PipeOps from the mlr_pipeops dictionary, it is also possible to give additional constructor arguments, such as an id or parameter values. po("pca", rank. = 3) ## PipeOp: <pca> (not trained) ## values: <rank.=3> ## Input channels <name [train type, predict type]>: ## input [Task,Task] ## Output channels <name [train type, predict type]>: ## output [Task,Task] ### 5.7.3 PipeOp Channels #### 5.7.3.1 Input Channels Just like functions, PipeOps can take multiple inputs. These multiple inputs are always given as elements in the input list. For example, there is a PipeOpFeatureUnion that combines multiple tasks with different features and “cbind()s” them together, creating one combined task. When two halves of the iris task are given, for example, it recreates the original task: iris_first_half = task$clone()$select(c("Petal.Length", "Petal.Width")) iris_second_half = task$clone()$select(c("Sepal.Length", "Sepal.Width")) pofu = po("featureunion", innum = 2) pofu$train(list(iris_first_half, iris_second_half))[[1]]$data() ## Species Petal.Length Petal.Width Sepal.Length Sepal.Width ## 1: setosa 1.4 0.2 5.1 3.5 ## 2: setosa 1.4 0.2 4.9 3.0 ## 3: setosa 1.3 0.2 4.7 3.2 ## 4: setosa 1.5 0.2 4.6 3.1 ## 5: setosa 1.4 0.2 5.0 3.6 ## --- ## 146: virginica 5.2 2.3 6.7 3.0 ## 147: virginica 5.0 1.9 6.3 2.5 ## 148: virginica 5.2 2.0 6.5 3.0 ## 149: virginica 5.4 2.3 6.2 3.4 ## 150: virginica 5.1 1.8 5.9 3.0 Because PipeOpFeatureUnion effectively takes two input arguments here, we can say it has two input channels. An input channel also carries information about the type of input that is acceptable. The input channels of the pofu object constructed above, for example, each accept a Task during training and prediction. This information can be queried from the $input slot:

pofu$input ## name train predict ## 1: input1 Task Task ## 2: input2 Task Task Other PipeOps may have channels that take different types during different phases. The backuplearner PipeOp, for example, takes a NULL and a Task during training, and a Prediction and a Task during prediction: ## TODO this is an important case to handle here, do not delete unless there is a better example. ## po("backuplearner")$input

#### 5.7.3.2 Output Channels

Unlike the typical notion of a function, PipeOps can also have multiple output channels. $train() and $predict() always return a list, so certain PipeOps may return lists with more than one element. Similar to input channels, the information about the number and type of outputs given by a PipeOp is available in the $output slot. The chunk PipeOp, for example, chunks a given Task into subsets and consequently returns multiple Task objects, both during training and prediction. The number of output channels must be given during construction through the outnum argument. po("chunk", outnum = 3)$output
##       name train predict
## 3: output3  Task    Task

Note that the number of output channels during training and prediction is the same. A schema of a PipeOp with two output channels:

#### 5.7.3.3 Channel Configuration

Most PipeOps have only one input channel (so they take a list with a single element), but there are a few with more than one; In many cases, the number of input or output channels is determined during construction, e.g. through the innum / outnum arguments. The input.num and output.num columns of the mlr_pipeops-table above show the default number of channels, and NA if the number depends on a construction argument.

The default printer of a PipeOp gives information about channel names and types:

## po("backuplearner")

### 5.7.4Graph: Networks of PipeOps

#### 5.7.4.1 Basics

What is the advantage of this tedious way of declaring input and output channels and handling in/output through lists? Because each PipeOp has a known number of input and output channels that always produce or accept data of a known type, it is possible to network them together in Graphs. A Graph is a collection of PipeOps with “edges” that mandate that data should be flowing along them. Edges always pass between PipeOp channels, so it is not only possible to explicitly prescribe which position of an input or output list an edge refers to, it makes it possible to make different components of a PipeOp’s output flow to multiple different other PipeOps, as well as to have a PipeOp gather its input from multiple other PipeOps.

A schema of a simple graph of PipeOps:

A Graph is empty when first created, and PipeOps can be added using the $add_pipeop() method. The $add_edge() method is used to create connections between them. While the printer of a Graph gives some information about its layout, the most intuitive way of visualizing it is using the $plot() function. gr = Graph$new()
gr$add_pipeop(po("scale")) gr$add_pipeop(po("subsample", frac = 0.1))
gr$add_edge("scale", "subsample") print(gr) ## Graph with 2 PipeOps: ## ID State sccssors prdcssors ## scale <<UNTRAINED>> subsample ## subsample <<UNTRAINED>> scale gr$plot(html = FALSE)

A Graph itself has a $train() and a $predict() method that accept some data and propagate this data through the network of PipeOps. The return value corresponds to the output of the PipeOp output channels that are not connected to other PipeOps.

gr$train(task)[[1]]$data()
##        Species Petal.Length Petal.Width Sepal.Length Sepal.Width
##  1:     setosa      -1.3358     -1.3111      -1.1392    -0.13154
##  2:     setosa      -1.3924     -1.3111      -1.3807     0.32732
##  3:     setosa      -1.3924     -1.0487      -0.5354     1.93331
##  4:     setosa      -1.3358     -1.1799      -0.8977     1.01560
##  5:     setosa      -1.2791     -1.3111      -1.1392     0.09789
##  6:     setosa      -1.2225     -0.7863      -1.0184     1.01560
##  7: versicolor      -0.2594     -0.2615      -1.1392    -1.50811
##  8: versicolor       0.4203      0.3945       0.4307    -1.96696
##  9: versicolor       0.3637      0.2633       0.9138    -0.13154
## 10: versicolor       0.5903      0.2633       1.1553    -0.59040
## 11: versicolor       0.3637      0.1321       0.5515    -1.73754
## 12:  virginica       1.3267      1.7064       1.6384     1.24503
## 13:  virginica       1.7799      1.4440       2.2422    -1.04925
## 14:  virginica       0.7036      0.3945       0.1892    -1.96696
## 15:  virginica       0.6469      0.7880       0.5515    -0.81982
gr$predict(single_line_task)[[1]]$data()
##    Species Petal.Length Petal.Width Sepal.Length Sepal.Width
## 1:  setosa       -1.336      -1.311      -0.8977       1.016

The collection of PipeOps inside a Graph can be accessed through the $pipeops slot. The set of edges in the Graph can be inspected through the $edges slot. It is possible to modify individual PipeOps and edges in a Graph through these slots, but this is not recommended because no error checking is performed and it may put the Graph in an unsupported state.

#### 5.7.4.2 Networks

The example above showed a linear preprocessing pipeline, but it is in fact possible to build true “graphs” of operations, as long as no loops are introduced1. PipeOps with multiple output channels can feed their data to multiple different subsequent PipeOps, and PipeOps with multiple input channels can take results from different PipeOps. When a PipeOp has more than one input / output channel, then the Graph’s $add_edge() method needs an additional argument that indicates which channel to connect to. This argument can be given in the form of an integer, or as the name of the channel. The following constructs a Graph that copies the input and gives one copy each to a “scale” and a “pca” PipeOp. The resulting columns of each operation are put next to each other by “featureunion”. gr = Graph$new()$add_pipeop(po("copy", outnum = 2))$
add_pipeop(po("scale"))$add_pipeop(po("pca"))$

gr$add_edge("copy", "scale", src_channel = 1)$        # designating channel by index
add_edge("copy", "pca", src_channel = "output2")$# designating channel by name add_edge("scale", "featureunion", dst_channel = 1)$

#### 5.7.4.4PipeOp IDs and ID Name Clashes

PipeOps within a graph are addressed by their $id-slot. It is therefore necessary for all PipeOps within a Graph to have a unique $id. The $id can be set during or after construction, but it should not directly be changed after a PipeOp was inserted in a Graph. At that point, the $set_names()-method can be used to change PipeOp ids.

po1 = po("scale")
po2 = po("scale")
po1 %>>% po2  ## name clash
## Error in gunion(list(g1, g2)): Assertion on 'ids of pipe operators' failed: Must have unique names, but element 2 is duplicated.
po2$id = "scale2" gr = po1 %>>% po2 gr ## Graph with 2 PipeOps: ## ID State sccssors prdcssors ## scale <<UNTRAINED>> scale2 ## scale2 <<UNTRAINED>> scale ## Alternative ways of getting new ids: po("scale", id = "scale2") ## PipeOp: <scale2> (not trained) ## values: <robust=FALSE> ## Input channels <name [train type, predict type]>: ## input [Task,Task] ## Output channels <name [train type, predict type]>: ## output [Task,Task] po("scale", id = "scale2") ## PipeOp: <scale2> (not trained) ## values: <robust=FALSE> ## Input channels <name [train type, predict type]>: ## input [Task,Task] ## Output channels <name [train type, predict type]>: ## output [Task,Task] ## sometimes names of PipeOps within a Graph need to be changed gr2 = po("scale") %>>% po("pca") gr %>>% gr2 ## Error in gunion(list(g1, g2)): Assertion on 'ids of pipe operators' failed: Must have unique names, but element 3 is duplicated. gr2$set_names("scale", "scale3")
gr %>>% gr2
## Graph with 4 PipeOps:
##      ID         State sccssors prdcssors
##   scale <<UNTRAINED>>   scale2
##  scale2 <<UNTRAINED>>   scale3     scale
##  scale3 <<UNTRAINED>>      pca    scale2
##     pca <<UNTRAINED>>             scale3

### 5.7.5 Learners in Graphs, Graphs in Learners

The true power of mlr3pipelines derives from the fact that it can be integrated seamlessly with mlr3. Two components are mainly responsible for this:

• PipeOpLearner, a PipeOp that encapsulates a mlr3 Learner and creates a PredictionData object in its $predict() phase • GraphLearner, a mlr3 Learner that can be used in place of any other mlr3 Learner, but which does prediction using a Graph given to it Note that these are dual to each other: One takes a Learner and produces a PipeOp (and by extension a Graph); the other takes a Graph and produces a Learner. #### 5.7.5.1PipeOpLearner The PipeOpLearner is constructed using a mlr3 Learner and will use it to create PredictionData in the $predict() phase. The output during $train() is NULL. It can be used after a preprocessing pipeline, and it is even possible to perform operations on the PredictionData, for example by averaging multiple predictions or by using the “PipeOpBackupLearner” operator to impute predictions that a given model failed to create. The following is a very simple Graph that performs training and prediction on data after performing principal component analysis. gr = po("pca") %>>% po("learner", lrn("classif.rpart")) gr$train(task)
## $classif.rpart.output ## NULL gr$predict(task)
## $classif.rpart.output ## <PredictionClassif> for 150 observations: ## row_ids truth response ## 1 setosa setosa ## 2 setosa setosa ## 3 setosa setosa ## --- ## 148 virginica virginica ## 149 virginica virginica ## 150 virginica virginica #### 5.7.5.2GraphLearner Although a Graph has $train() and $predict() functions, it can not be used directly in places where mlr3 Learners can be used like resampling or benchmarks. For this, it needs to be wrapped in a GraphLearner object, which is a thin wrapper that enables this functionality. The resulting Learner is extremely versatile, because every part of it can be modified, replaced, parameterized and optimized over. Resampling the graph above can be done the same way that resampling of the Learner was performed in the introductory example. lrngrph = as_learner(gr) resample(task, lrngrph, rsmp) ## <ResampleResult> of 1 iterations ## * Task: iris ## * Learner: pca.classif.rpart ## * Warnings: 0 in 0 iterations ## * Errors: 0 in 0 iterations ### 5.7.6 Hyperparameters mlr3pipelines relies on the paradox package to provide parameters that can modify each PipeOp’s behavior. paradox parameters provide information about the parameters that can be changed, as well as their types and ranges. They provide a unified interface for benchmarks and parameter optimization (“tuning”). For a deep dive into paradox, see the tuning chapter or the in-depth paradox chapter. The ParamSet, representing the space of possible parameter configurations of a PipeOp, can be inspected by accessing the $param_set slot of a PipeOp or a Graph.

op_pca = po("pca")
op_pca$param_set ## <ParamSet:pca> ## id class lower upper nlevels default value ## 1: center ParamLgl NA NA 2 TRUE ## 2: scale. ParamLgl NA NA 2 FALSE ## 3: rank. ParamInt 1 Inf Inf ## 4: affect_columns ParamUty NA NA Inf <Selector[1]> To set or retrieve a parameter, the $param_set$values slot can be accessed. Alternatively, the param_vals value can be given during construction. op_pca$param_set$values$center = FALSE
op_pca$param_set$values
## $center ## [1] FALSE op_pca = po("pca", center = TRUE) op_pca$param_set$values ##$center
## [1] TRUE

Each PipeOp can bring its own individual parameters which are collected together in the Graph’s $param_set. A PipeOp’s parameter names are prefixed with its $id to prevent parameter name clashes.

gr = op_pca %>>% po("scale")
gr$param_set ## <ParamSetCollection> ## id class lower upper nlevels default value ## 1: pca.center ParamLgl NA NA 2 TRUE TRUE ## 2: pca.scale. ParamLgl NA NA 2 FALSE ## 3: pca.rank. ParamInt 1 Inf Inf ## 4: pca.affect_columns ParamUty NA NA Inf <Selector[1]> ## 5: scale.center ParamLgl NA NA 2 TRUE ## 6: scale.scale ParamLgl NA NA 2 TRUE ## 7: scale.robust ParamLgl NA NA 2 <NoDefault[3]> FALSE ## 8: scale.affect_columns ParamUty NA NA Inf <Selector[1]> gr$param_set$values ##$pca.center
## [1] TRUE
##
## $scale.robust ## [1] FALSE Both PipeOpLearner and GraphLearner preserve parameters of the objects they encapsulate. op_rpart = po("learner", lrn("classif.rpart")) op_rpart$param_set
## <ParamSet:classif.rpart>
##                 id    class lower upper nlevels        default value
##  1:             cp ParamDbl     0     1     Inf           0.01
##  2:     keep_model ParamLgl    NA    NA       2          FALSE
##  3:     maxcompete ParamInt     0   Inf     Inf              4
##  4:       maxdepth ParamInt     1    30      30             30
##  5:   maxsurrogate ParamInt     0   Inf     Inf              5
##  6:      minbucket ParamInt     1   Inf     Inf <NoDefault[3]>
##  7:       minsplit ParamInt     1   Inf     Inf             20
##  8: surrogatestyle ParamInt     0     1       2              0
##  9:   usesurrogate ParamInt     0     2       3              2
## 10:           xval ParamInt     0   Inf     Inf             10     0
glrn = as_learner(gr %>>% op_rpart)
glrn\$param_set
## <ParamSetCollection>
##                               id    class lower upper nlevels        default
##  1:                   pca.center ParamLgl    NA    NA       2           TRUE
##  2:                   pca.scale. ParamLgl    NA    NA       2          FALSE
##  3:                    pca.rank. ParamInt     1   Inf     Inf
##  4:           pca.affect_columns ParamUty    NA    NA     Inf  <Selector[1]>
##  5:                 scale.center ParamLgl    NA    NA       2           TRUE
##  6:                  scale.scale ParamLgl    NA    NA       2           TRUE
##  7:                 scale.robust ParamLgl    NA    NA       2 <NoDefault[3]>
##  8:         scale.affect_columns ParamUty    NA    NA     Inf  <Selector[1]>
##  9:             classif.rpart.cp ParamDbl     0     1     Inf           0.01
## 10:     classif.rpart.keep_model ParamLgl    NA    NA       2          FALSE
## 11:     classif.rpart.maxcompete ParamInt     0   Inf     Inf              4
## 12:       classif.rpart.maxdepth ParamInt     1    30      30             30
## 13:   classif.rpart.maxsurrogate ParamInt     0   Inf     Inf              5
## 14:      classif.rpart.minbucket ParamInt     1   Inf     Inf <NoDefault[3]>
## 15:       classif.rpart.minsplit ParamInt     1   Inf     Inf             20
## 16: classif.rpart.surrogatestyle ParamInt     0     1       2              0
## 17:   classif.rpart.usesurrogate ParamInt     0     2       3              2
## 18:           classif.rpart.xval ParamInt     0   Inf     Inf             10
##     value
##  1:  TRUE
##  2:
##  3:
##  4:
##  5:
##  6:
##  7: FALSE
##  8:
##  9:
## 10:
## 11:
## 12:
## 13:
## 14:
## 15:
## 16:
## 17:
## 18:     0