Combining pipelines"
In pipeflow: Fast Interactive Data Analysis Pipelines

knitr::opts_chunk$set(
    comment = "#",
    prompt = FALSE,
    tidy = FALSE,
    cache = FALSE,
    collapse = TRUE
)

old <- options(width = 100L)

The possibility to combine pipelines basically allows to modularize the pipeline creation process. This is especially useful when you have a set of pipelines that are used in different contexts and you want to avoid code duplication.

Two pipelines

Let's define one pipeline that is used for data preprocessing and one that does the modeling.

Data preprocessing pipeline:

library(pipeflow)

pip1 <- pip_new("preprocessing") |>
    pip_add(
        "data",
        function(data = airquality) data
    ) |>
    pip_add(
        "data_prep",
        function(data = ~data) {
            replace(data, "Temp.Celsius", (data[, "Temp"] - 32) * 5 / 9)
        }
    ) |>
    pip_add(
        "standardize",
        function(
            data = ~data_prep,
            yVar = "Ozone"
        ) {
            data[, yVar] <- scale(data[, yVar])
            data
        }
    )

pip1

Modelling pipeline:

pip2 <- pip_new("modeling") |>
    pip_add(
        "data",
        function(data = airquality) data
    ) |>
    pip_add(
        "fit",
        function(
            data = ~data,
            xVar = "Temp",
            yVar = "Ozone"
        ) {
            lm(paste(yVar, "~", xVar), data = data)
        }
    ) |>
    pip_add(
        "plot",
        function(
            model = ~fit,
            data = ~data,
            xVar = "Temp",
            yVar = "Ozone",
            title = "Linear model fit"
        ) {
            require(ggplot2, quietly = TRUE)
            coeffs <- coefficients(model)
            ggplot(data) +
                geom_point(aes(.data[[xVar]], .data[[yVar]])) +
                geom_abline(intercept = coeffs[1], slope = coeffs[2]) +
                labs(title = title)
        }
    )

pip2

Combined pipeline

Next we combine the two pipelines using pip_bind().

pip <- pip_bind(pip1, pip2)

pip

First of all, note that the data step of the second pipeline has been renamed automatically to avoid name clashes. In particular, the first step of the second pipeline has been renamed from data to data2 (line 4 in the step column) and likewise the data-dependencies of the second pipeline have been updated (see lines 5-6 in the depends column).

That is, when binding two pipelines, {pipeflow} ensures that the step names remain unique in the resulting combined pipeline and therefore automatically renames duplicated step names if necessary.

Now, as can be also seen from the graphical representation of the pipeline,

library(visNetwork)
do.call(visNetwork, args = pip_get_graph(pip)) |>
    visHierarchicalLayout(direction = "LR")

library(visNetwork)
do.call(
    visNetwork,
    args = c(pip_get_graph(pip), list(height = 250, width = 500))
) |>
    visHierarchicalLayout(direction = "LR", sortMethod = "directed")

the two pipelines are not yet connected. To make actual use of the combined pipeline, we have to use the output of the first pipeline as input of the second pipeline, that is, we want to use the output of the standardize step as the data parameter input in the data2 step. To achieve this, we apply the replace function as described earlier in the vignette modify the pipeline:

pip |> pip_replace("data2", function(data = ~standardize) data)

pip

do.call(
    visNetwork,
    args = c(pip_get_graph(pip), list(height = 100, width = 700))
) |>
    visHierarchicalLayout(direction = "LR")

Relative indexing

Since the name of the re-routed step might not always be known^[A typical example would be appending several pipelines in a programmatic context.], the {pipeflow} package also provides a relative position indexing mechanism, which allows to rewrite the above command as follows:

pip |> pip_replace("data2", function(data = ~ -1) data)

pip

Generally speaking, the relative indexing mechanism allows to refer to steps positioned above the current step. The index ~-1 can be interpreted as "go one step back", ~-2 as "go two steps back", and so on.

Combined pipeline results

Let's now run the combined pipeline and inspect the plot.

pip_run(pip)

pip[["plot", "out"]]

As we can see, the plot shows the linear model fit of the standardized data. We can now go ahead and for example change the x-variable of the model and rerun the pipeline.

pip_set_params(pip, params = list(xVar = "Temp.Celsius"))

library(visNetwork)
do.call(
    visNetwork,
    args = c(pip_get_graph(pip), list(height = 100, width = 700))
) |>
    visHierarchicalLayout(direction = "LR", sortMethod = "directed")

pip_run(pip)

pip[["plot", "out"]]

Step cherry-picking

Another way to re-use steps from other pipelines is by cherry-picking, which can be done via pip_add_from, for example:

pip <- pip_new("cherry-picked-from-1-and-2") |>
    pip_add_from(pip1, "data") |>
    pip_add_from(pip1, "data_prep") |>
    pip_add_from(pip1, "standardize") |>
    pip_add_from(pip2, "fit") |>
    pip_add_from(pip2, "plot")

pip

Note that here the cherry-pick approach is not all that useful, because in contrast to the pip_bind command, which renames data to data2, the cherry-picked fit and plot steps still refer to the initial data step.

library(visNetwork)
do.call(
    visNetwork,
    args = c(pip_get_graph(pip), list(height = 250, width = 700))
) |>
    visHierarchicalLayout(direction = "LR", sortMethod = "directed")

In other scenarios, however, this might be exactly what you want.

Generally, when creating these pipelines, there will be a lot of steps calculating intermediate results and only a few steps contain the final output we are interested in (see e.g. the plot output in the above example). To see how {pipeflow} allows to conveniently tag, collect and possibly group those final outputs, see the next vignette Collecting and filtering output.