Branching
Last updated on 2024-07-09 | Edit this page
Overview
Questions
- How can we specify many targets without typing everything out?
Objectives
- Be able to specify targets using branching
Why branching?
One of the major strengths of targets is the ability to
define many targets from a single line of code (“branching”). This not
only saves you typing, it also reduces the risk of
errors since there is less chance of making a typo.
Types of branching
There are two types of branching, dynamic branching
and static branching. “Branching” refers to the idea
that you can provide a single specification for how to make targets (the
“pattern”), and targets generates multiple targets from it
(“branches”). “Dynamic” means that the branches that result from the
pattern do not have to be defined ahead of time—they are a dynamic
result of the code.
In this workshop, we will only cover dynamic branching since it is
generally easier to write (static branching requires use of meta-programming,
an advanced topic). For more information about each and when you might
want to use one or the other (or some combination of the two), see the
targets package manual.
Example without branching
To see how this works, let’s continue our analysis of the
palmerpenguins dataset.
Our hypothesis is that bill depth decreases with bill length. We will test this hypothesis with a linear model.
For example, this is a model of bill depth dependent on bill length:
R
lm(bill_depth_mm ~ bill_length_mm, data = penguins_data)
We can add this to our pipeline. We will call it the
combined_model because it combines all the species together
without distinction:
R
source("R/packages.R")
source("R/functions.R")
tar_plan(
# Load raw data
tar_file_read(
penguins_data_raw,
path_to_file("penguins_raw.csv"),
read_csv(!!.x, show_col_types = FALSE)
),
# Clean data
penguins_data = clean_penguin_data(penguins_data_raw),
# Build model
combined_model = lm(
bill_depth_mm ~ bill_length_mm,
data = penguins_data
)
)
OUTPUT
✔ skipped target penguins_data_raw_file
✔ skipped target penguins_data_raw
✔ skipped target penguins_data
▶ dispatched target combined_model
● completed target combined_model [0.005 seconds]
▶ ended pipeline [0.36 seconds]Let’s have a look at the model. We will use the glance()
function from the broom package. Unlike base R
summary(), this function returns output as a tibble (the
tidyverse equivalent of a dataframe), which as we will see later is
quite useful for downstream analyses.
R
library(broom)
tar_load(combined_model)
glance(combined_model)
OUTPUT
# A tibble: 1 × 12
r.squared adj.r.squared sigma statistic p.value df logLik AIC BIC deviance df.residual nobs
<dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <int> <int>
1 0.0552 0.0525 1.92 19.9 0.0000112 1 -708. 1422. 1433. 1256. 340 342Notice the small P-value. This seems to indicate that the model is highly significant.
But wait a moment… is this really an appropriate model? Recall that there are three species of penguins in the dataset. It is possible that the relationship between bill depth and length varies by species.
We should probably test some alternative models. These could include models that add a parameter for species, or add an interaction effect between species and bill length.
Now our workflow is getting more complicated. This is what a workflow
for such an analysis might look like without branching
(make sure to add library(broom) to
packages.R):
R
source("R/packages.R")
source("R/functions.R")
tar_plan(
# Load raw data
tar_file_read(
penguins_data_raw,
path_to_file("penguins_raw.csv"),
read_csv(!!.x, show_col_types = FALSE)
),
# Clean data
penguins_data = clean_penguin_data(penguins_data_raw),
# Build models
combined_model = lm(
bill_depth_mm ~ bill_length_mm,
data = penguins_data
),
species_model = lm(
bill_depth_mm ~ bill_length_mm + species,
data = penguins_data
),
interaction_model = lm(
bill_depth_mm ~ bill_length_mm * species,
data = penguins_data
),
# Get model summaries
combined_summary = glance(combined_model),
species_summary = glance(species_model),
interaction_summary = glance(interaction_model)
)
OUTPUT
✔ skipped target penguins_data_raw_file
✔ skipped target penguins_data_raw
✔ skipped target penguins_data
✔ skipped target combined_model
▶ dispatched target interaction_model
● completed target interaction_model [0.004 seconds]
▶ dispatched target species_model
● completed target species_model [0.001 seconds]
▶ dispatched target combined_summary
● completed target combined_summary [0.007 seconds]
▶ dispatched target interaction_summary
● completed target interaction_summary [0.003 seconds]
▶ dispatched target species_summary
● completed target species_summary [0.003 seconds]
▶ ended pipeline [0.475 seconds]Let’s look at the summary of one of the models:
R
tar_read(species_summary)
OUTPUT
# A tibble: 1 × 12
r.squared adj.r.squared sigma statistic p.value df logLik AIC BIC deviance df.residual nobs
<dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <int> <int>
1 0.769 0.767 0.953 375. 3.65e-107 3 -467. 944. 963. 307. 338 342So this way of writing the pipeline works, but is repetitive: we have
to call glance() each time we want to obtain summary
statistics for each model. Furthermore, each summary target
(combined_summary, etc.) is explicitly named and typed out
manually. It would be fairly easy to make a typo and end up with the
wrong model being summarized.
Example with branching
First attempt
Let’s see how to write the same plan using dynamic branching:
R
source("R/packages.R")
source("R/functions.R")
tar_plan(
# Load raw data
tar_file_read(
penguins_data_raw,
path_to_file("penguins_raw.csv"),
read_csv(!!.x, show_col_types = FALSE)
),
# Clean data
penguins_data = clean_penguin_data(penguins_data_raw),
# Build models
models = list(
combined_model = lm(
bill_depth_mm ~ bill_length_mm, data = penguins_data),
species_model = lm(
bill_depth_mm ~ bill_length_mm + species, data = penguins_data),
interaction_model = lm(
bill_depth_mm ~ bill_length_mm * species, data = penguins_data)
),
# Get model summaries
tar_target(
model_summaries,
glance(models[[1]]),
pattern = map(models)
)
)
What is going on here?
First, let’s look at the messages provided by
tar_make().
OUTPUT
✔ skipped target penguins_data_raw_file
✔ skipped target penguins_data_raw
✔ skipped target penguins_data
▶ dispatched target models
● completed target models [0.005 seconds]
▶ dispatched branch model_summaries_812e3af782bee03f
● completed branch model_summaries_812e3af782bee03f [0.007 seconds]
▶ dispatched branch model_summaries_2b8108839427c135
● completed branch model_summaries_2b8108839427c135 [0.003 seconds]
▶ dispatched branch model_summaries_533cd9a636c3e05b
● completed branch model_summaries_533cd9a636c3e05b [0.003 seconds]
● completed pattern model_summaries
▶ ended pipeline [0.462 seconds]There is a series of smaller targets (branches) that are each named
like model_summaries_812e3af782bee03f, then one overall
model_summaries target. That is the result of specifying
targets using branching: each of the smaller targets are the “branches”
that comprise the overall target. Since targets has no way
of knowing ahead of time how many branches there will be or what they
represent, it names each one using this series of numbers and letters
(the “hash”). targets builds each branch one at a time,
then combines them into the overall target.
Next, let’s look in more detail about how the workflow is set up, starting with how we defined the models:
R
# Build models
models = list(
combined_model = lm(
bill_depth_mm ~ bill_length_mm, data = penguins_data),
species_model = lm(
bill_depth_mm ~ bill_length_mm + species, data = penguins_data),
interaction_model = lm(
bill_depth_mm ~ bill_length_mm * species, data = penguins_data)
),Unlike the non-branching version, we defined the models in a
list (instead of one target per model). This is because dynamic
branching is similar to the base::apply() or purrrr::map()
method of looping: it applies a function to each element of a list. So
we need to prepare the input for looping as a list.
Next, take a look at the command to build the target
model_summaries.
R
# Get model summaries
tar_target(
model_summaries,
glance(models[[1]]),
pattern = map(models)
)
As before, the first argument is the name of the target to build, and the second is the command to build it.
Here, we apply the glance() function to each element of
models (the [[1]] is necessary because when
the function gets applied, each element is actually a nested list, and
we need to remove one layer of nesting).
Finally, there is an argument we haven’t seen before,
pattern, which indicates that this target should be built
using dynamic branching. map means to apply the command to
each element of the input list (models) sequentially.
Now that we understand how the branching workflow is constructed, let’s inspect the output:
R
tar_read(model_summaries)
OUTPUT
# A tibble: 3 × 12
r.squared adj.r.squared sigma statistic p.value df logLik AIC BIC deviance df.residual nobs
<dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <int> <int>
1 0.0552 0.0525 1.92 19.9 1.12e- 5 1 -708. 1422. 1433. 1256. 340 342
2 0.769 0.767 0.953 375. 3.65e-107 3 -467. 944. 963. 307. 338 342
3 0.770 0.766 0.955 225. 8.52e-105 5 -466. 947. 974. 306. 336 342The model summary statistics are all included in a single dataframe.
But there’s one problem: we can’t tell which row came from which model! It would be unwise to assume that they are in the same order as the list of models.
This is due to the way dynamic branching works: by default, there is no information about the provenance of each target preserved in the output.
How can we fix this?
Second attempt
The key to obtaining useful output from branching pipelines is to
include the necessary information in the output of each individual
branch. Here, we want to know the kind of model that corresponds to each
row of the model summaries. To do that, we need to write a
custom function. You will need to write custom
functions frequently when using targets, so it’s good to
get used to it!
Here is the function. Save this in R/functions.R:
R
glance_with_mod_name <- function(model_in_list) {
model_name <- names(model_in_list)
model <- model_in_list[[1]]
glance(model) |>
mutate(model_name = model_name)
}
Our new pipeline looks almost the same as before, but this time we
use the custom function instead of glance().
R
source("R/functions.R")
source("R/packages.R")
tar_plan(
# Load raw data
tar_file_read(
penguins_data_raw,
path_to_file("penguins_raw.csv"),
read_csv(!!.x, show_col_types = FALSE)
),
# Clean data
penguins_data = clean_penguin_data(penguins_data_raw),
# Build models
models = list(
combined_model = lm(
bill_depth_mm ~ bill_length_mm, data = penguins_data),
species_model = lm(
bill_depth_mm ~ bill_length_mm + species, data = penguins_data),
interaction_model = lm(
bill_depth_mm ~ bill_length_mm * species, data = penguins_data)
),
# Get model summaries
tar_target(
model_summaries,
glance_with_mod_name(models),
pattern = map(models)
)
)
OUTPUT
✔ skipped target penguins_data_raw_file
✔ skipped target penguins_data_raw
✔ skipped target penguins_data
✔ skipped target models
▶ dispatched branch model_summaries_812e3af782bee03f
● completed branch model_summaries_812e3af782bee03f [0.018 seconds]
▶ dispatched branch model_summaries_2b8108839427c135
● completed branch model_summaries_2b8108839427c135 [0.007 seconds]
▶ dispatched branch model_summaries_533cd9a636c3e05b
● completed branch model_summaries_533cd9a636c3e05b [0.005 seconds]
● completed pattern model_summaries
▶ ended pipeline [0.58 seconds]And this time, when we load the model_summaries, we can
tell which model corresponds to which row (you may need to scroll to the
right to see it).
R
tar_read(model_summaries)
OUTPUT
# A tibble: 3 × 13
r.squared adj.r.squared sigma statistic p.value df logLik AIC BIC deviance df.residual nobs model_name
<dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <int> <int> <chr>
1 0.0552 0.0525 1.92 19.9 1.12e- 5 1 -708. 1422. 1433. 1256. 340 342 combined_model
2 0.769 0.767 0.953 375. 3.65e-107 3 -467. 944. 963. 307. 338 342 species_model
3 0.770 0.766 0.955 225. 8.52e-105 5 -466. 947. 974. 306. 336 342 interaction_modelNext we will add one more target, a prediction of bill depth based on
each model. These will be needed for plotting the models in the report.
Such a prediction can be obtained with the augment()
function of the broom package.
R
tar_load(models)
augment(models[[1]])
OUTPUT
# A tibble: 342 × 8
bill_depth_mm bill_length_mm .fitted .resid .hat .sigma .cooksd .std.resid
<dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
1 18.7 39.1 17.6 1.14 0.00521 1.92 0.000924 0.594
2 17.4 39.5 17.5 -0.127 0.00485 1.93 0.0000107 -0.0663
3 18 40.3 17.5 0.541 0.00421 1.92 0.000168 0.282
4 19.3 36.7 17.8 1.53 0.00806 1.92 0.00261 0.802
5 20.6 39.3 17.5 3.06 0.00503 1.92 0.00641 1.59
6 17.8 38.9 17.6 0.222 0.00541 1.93 0.0000364 0.116
7 19.6 39.2 17.6 2.05 0.00512 1.92 0.00293 1.07
8 18.1 34.1 18.0 0.114 0.0124 1.93 0.0000223 0.0595
9 20.2 42 17.3 2.89 0.00329 1.92 0.00373 1.50
10 17.1 37.8 17.7 -0.572 0.00661 1.92 0.000296 -0.298
# ℹ 332 more rowsDefine the new function as augment_with_mod_name(). It
is the same as glance_with_mod_name(), but use
augment() instead of glance():
R
augment_with_mod_name <- function(model_in_list) {
model_name <- names(model_in_list)
model <- model_in_list[[1]]
augment(model) |>
mutate(model_name = model_name)
}
Add the step to the workflow:
R
source("R/functions.R")
source("R/packages.R")
tar_plan(
# Load raw data
tar_file_read(
penguins_data_raw,
path_to_file("penguins_raw.csv"),
read_csv(!!.x, show_col_types = FALSE)
),
# Clean data
penguins_data = clean_penguin_data(penguins_data_raw),
# Build models
models = list(
combined_model = lm(
bill_depth_mm ~ bill_length_mm, data = penguins_data),
species_model = lm(
bill_depth_mm ~ bill_length_mm + species, data = penguins_data),
interaction_model = lm(
bill_depth_mm ~ bill_length_mm * species, data = penguins_data)
),
# Get model summaries
tar_target(
model_summaries,
glance_with_mod_name(models),
pattern = map(models)
),
# Get model predictions
tar_target(
model_predictions,
augment_with_mod_name(models),
pattern = map(models)
)
)
Some other ways of applying branching patterns include:
- crossing: one branch per combination of elements
(
cross()function) - slicing: one branch for each of a manually selected set of elements
(
slice()function) - sampling: one branch for each of a randomly selected set of elements
(
sample()function)
You can find out
more about different branching patterns in the targets
manual.