In this vignette, we will state the main goal of the add/update/remove and adjust functions and describe what part of the processing each function is intended for. We will then demonstrate how to use the sets of add/update/remove functions, followed by the adjust functions, and end with a brief discussion on the tidy methods for recipe and frosting objects.
Main goal of the add/update/remove and adjust functions
The primary goal of the update and adjust functions is to allow the
user to modify a step, layer,
epi_recipe, frosting, or a part of an
epi_workflow so that they do not have to create a new
object each time they wish to make a change to the pre-processing,
fitting, or post-processing.
In the context of pre-processing, the goal of the update functions is
to add/remove/update an epi_recipe or a step in it. For
this, we have add_epi_recipe(),
update_epi_recipe(), and remove_epi_recipe()
to add/update/remove an entire epi_recipe in an
epi_workflow as well as adjust_epi_recipe() to
adjust a particular step in an epi_recipe or
epi_workflow by the step number or name. For a model, one
may Add_model(), Update_model(), or
Remove_model() in an epi_workflow.1 For
post-processing, where the goal is to update a frosting object or a
layer in it, we have add_frosting(),
remove_frosting(), and update_frosting() to
add/update/remove an entire frosting object in an
epi_workflow as well as adjust_frosting() to
adjust a particular layer in a frosting or
epi_workflow by its number or name. A summary of the
function uses by processing step is shown by the following table:
| Add/update/remove functions | adjust functions | |
|---|---|---|
| Pre-processing |
add_epi_recipe(), update_epi_recipe(),
remove_epi_recipe()
|
adjust_epi_recipe() |
| Model specification |
Add_model(), Update_model()
Remove_model()
|
|
| Post-processing |
add_frosting(), remove_frosting(),
update_frosting()
|
adjust_frosting() |
Since adding/removing/updating frosting as well as adjusting a layer
in a frosting object proceeds in the same way as performing
those tasks on an epi_recipe, we will focus on implementing
those for an epi_recipe in this vignette and only briefly
go through some examples for a frosting object.
Add/update/remove an epi_recipe in an
epi_workflow
We start with the built-in case_death_rate_subset
dataset that contains JHU daily COVID-19 cases and deaths by state and
take a subset of it from Nov. 1, 2021 to Dec. 31, 2021 for the four
states of Alaska, California, New York, and South Carolina.
jhu <- case_death_rate_subset %>%
dplyr::filter(time_value >= as.Date("2021-11-01"), geo_value %in% c("ak", "ca", "ny", "sc"))
jhu
#> An `epi_df` object, 244 x 4 with metadata:
#> * geo_type = state
#> * time_type = day
#> * as_of = 2022-05-31 19:08:25.791826
#>
#> # A tibble: 244 × 4
#> geo_value time_value case_rate death_rate
#> * <chr> <date> <dbl> <dbl>
#> 1 ak 2021-11-01 87.9 0.494
#> 2 ca 2021-11-01 15.6 0.239
#> 3 ny 2021-11-01 19.9 0.177
#> 4 sc 2021-11-01 16.0 0.531
#> 5 ak 2021-11-02 83.2 0.395
#> 6 ca 2021-11-02 15.5 0.201
#> 7 ny 2021-11-02 20.3 0.171
#> 8 sc 2021-11-02 15.6 0.550
#> 9 ak 2021-11-03 85.2 0.415
#> 10 ca 2021-11-03 15.4 0.186
#> # ℹ 234 more rowsThen, we construct a simple epi_recipe object named
r, where we lag the death rates by 0, 7, and 14 days, lead
the death rate by 14 days, omit NA values in all predictors and then in
all outcomes (and set skip = TRUE to skip over this
processing of the outcome variable when the recipe is baked).
r <- epi_recipe(jhu) %>%
step_epi_lag(death_rate, lag = c(0, 7, 14)) %>%
step_epi_ahead(death_rate, ahead = 14) %>%
step_naomit(all_predictors()) %>%
step_naomit(all_outcomes(), skip = TRUE)We add this recipe to an epi_workflow object by
inputting r into the add_epi_recipe()
function:
wf <- epi_workflow() %>%
add_epi_recipe(r)
wf
#>
#> ══ Epi Workflow ════════════════════════════════════════════════════════════════
#> Preprocessor: Recipe
#> Model: None
#> Postprocessor: None
#>
#> ── Preprocessor ────────────────────────────────────────────────────────────────
#>
#> 4 Recipe steps.
#> 1. step_epi_lag()
#> 2. step_epi_ahead()
#> 3. step_naomit()
#> 4. step_naomit()
#> We may then go on to add the fitted linear model to our
epi_workflow:
# Fit a linear model
wf <- epi_workflow(r, parsnip::linear_reg()) %>% fit(jhu)
wf
#>
#> ══ Epi Workflow [trained] ══════════════════════════════════════════════════════
#> Preprocessor: Recipe
#> Model: linear_reg()
#> Postprocessor: None
#>
#> ── Preprocessor ────────────────────────────────────────────────────────────────
#>
#> 4 Recipe steps.
#> 1. step_epi_lag()
#> 2. step_epi_ahead()
#> 3. step_naomit()
#> 4. step_naomit()
#>
#> ── Model ───────────────────────────────────────────────────────────────────────
#>
#> Call:
#> stats::lm(formula = ..y ~ ., data = data)
#>
#> Coefficients:
#> (Intercept) lag_0_death_rate lag_7_death_rate lag_14_death_rate
#> 0.43505 -0.75576 0.02826 0.08960
#> At this stage, suppose we decide to overhaul our recipe so that we
have a different set of pre-processing steps or we want to make multiple
changes to existing steps, but we desire to keep the remainder of the
epi_workflow the same. We can use the
update_epi_recipe() function to trade our current recipe
r for another recipe r2 in wf as
follows:
r2 <- epi_recipe(jhu) %>%
step_epi_lag(death_rate, lag = c(0, 1, 7, 14)) %>%
step_epi_lag(case_rate, lag = c(0:7, 14)) %>%
step_epi_ahead(death_rate, ahead = 7) %>%
step_epi_naomit()
wf <- update_epi_recipe(wf, r2)
wf
#>
#> ══ Epi Workflow ════════════════════════════════════════════════════════════════
#> Preprocessor: Recipe
#> Model: linear_reg()
#> Postprocessor: None
#>
#> ── Preprocessor ────────────────────────────────────────────────────────────────
#>
#> 5 Recipe steps.
#> 1. step_epi_lag()
#> 2. step_epi_lag()
#> 3. step_epi_ahead()
#> 4. step_naomit()
#> 5. step_naomit()
#>
#> ── Model ───────────────────────────────────────────────────────────────────────
#>
#> Call:
#> stats::lm(formula = ..y ~ ., data = data)
#>
#> Coefficients:
#> (Intercept) lag_0_death_rate lag_7_death_rate lag_14_death_rate
#> 0.43505 -0.75576 0.02826 0.08960
#> You can see that the output of wf depicts the sequence
of steps in r2 instead of r, which indicates
that the update was successful.
A longer approach to achieve the same end is to use
remove_epi_recipe() to remove the old recipe and then
add_epi_recipe() to add the new one. Under the hood, the
update_epi_recipe() function operates in this way.
The add_epi_recipe() and
remove_epi_recipe() functions offload to the
workflows versions of the functions as much as possible.
The main reason for using the epipredict version is so
that we ensure that we retain the epi_workflow class.
To see this, let’s look at what happens if we remove our current
epi_recipe using workflows::remove_recipe()
and then inspect the class of wf:
wf %>% class() # class before
#> [1] "epi_workflow" "workflow"
workflows::remove_recipe(wf) %>% class() # class after removing recipe using workflows function
#> [1] "workflow"We can observe that wf is no longer an
epi_workflow and a workflow. It has been
demoted to only a workflow. While all
epi_workflows are workflows, not all
workflows are epi_workflows, meaning that
there may be compatibility issues and limitations to the tools that may
be used from the epipredict package on a plain
workflow object.
Now, while we checked what happens to the above
epi_recipe if we remove it, note that we did not actually
store that change to wf. Hence, our
epi_workflow remains unchanged.
wf
#>
#> ══ Epi Workflow ════════════════════════════════════════════════════════════════
#> Preprocessor: Recipe
#> Model: linear_reg()
#> Postprocessor: None
#>
#> ── Preprocessor ────────────────────────────────────────────────────────────────
#>
#> 5 Recipe steps.
#> 1. step_epi_lag()
#> 2. step_epi_lag()
#> 3. step_epi_ahead()
#> 4. step_naomit()
#> 5. step_naomit()
#>
#> ── Model ───────────────────────────────────────────────────────────────────────
#>
#> Call:
#> stats::lm(formula = ..y ~ ., data = data)
#>
#> Coefficients:
#> (Intercept) lag_0_death_rate lag_7_death_rate lag_14_death_rate
#> 0.43505 -0.75576 0.02826 0.08960
#> One thing to notice about this workflow output is that is that the
model fit remains the same as when we had r as the recipe.
This illustrates an important point - Any operations performed using the
old recipe are not updated automatically. So we should be careful to fit
the model using the new recipe, r2. Similarly, if
predictions were made using the old recipe, then they should be
re-generated using the version epi_workflow that contains
the updated recipe. We can use Update_model() to replace
the model used in wf, and then fit as before:
# fit linear model
wf <- Update_model(wf, parsnip::linear_reg()) %>% fit(jhu)
wf
#>
#> ══ Epi Workflow [trained] ══════════════════════════════════════════════════════
#> Preprocessor: Recipe
#> Model: linear_reg()
#> Postprocessor: None
#>
#> ── Preprocessor ────────────────────────────────────────────────────────────────
#>
#> 5 Recipe steps.
#> 1. step_epi_lag()
#> 2. step_epi_lag()
#> 3. step_epi_ahead()
#> 4. step_naomit()
#> 5. step_naomit()
#>
#> ── Model ───────────────────────────────────────────────────────────────────────
#>
#> Call:
#> stats::lm(formula = ..y ~ ., data = data)
#>
#> Coefficients:
#> (Intercept) lag_0_death_rate lag_1_death_rate lag_7_death_rate
#> 0.293291 -0.158682 0.055165 -0.349197
#> lag_14_death_rate lag_0_case_rate lag_1_case_rate lag_2_case_rate
#> -0.305136 0.009031 -0.009463 -0.005372
#> lag_3_case_rate lag_4_case_rate lag_5_case_rate lag_6_case_rate
#> -0.006244 0.004840 0.005537 -0.013347
#> lag_7_case_rate lag_14_case_rate
#> 0.011286 0.011721
#> Alternatively, we may use the Remove_model() followed by
Add_model() combination for the same effect.
Add/update/remove a frosting object in an
epi_workflow
We will now generate and create a frosting object for
post-processing predictions. In our initial frosting object,
f, we simply implement predictions on the fitted
epi_workflow:
f <- frosting() %>%
layer_predict()
wf1 <- wf %>% add_frosting(f)
p1 <- forecast(wf1)
p1
#> An `epi_df` object, 4 x 3 with metadata:
#> * geo_type = state
#> * time_type = day
#> * as_of = 2022-05-31 19:08:25.791826
#>
#> # A tibble: 4 × 3
#> geo_value time_value .pred
#> * <chr> <date> <dbl>
#> 1 ak 2021-12-31 -0.206
#> 2 ca 2021-12-31 0.0989
#> 3 ny 2021-12-31 0.289
#> 4 sc 2021-12-31 0.352Suppose we decide to augment our post-processing to include a threshold to enforce that the predictions are at least 0. As well, let’s include the forecast and target dates as separate columns.
To update the frosting while leaving the remainder of
the epi_workflow the same, we can use the
update_frosting() function as follows:
# Update frosting in a workflow and predict
f2 <- frosting() %>%
layer_predict() %>%
layer_threshold(.pred) %>%
layer_add_forecast_date() %>%
layer_add_target_date()
wf2 <- wf1 %>% update_frosting(f2)
p2 <- forecast(wf2)
p2
#> An `epi_df` object, 4 x 5 with metadata:
#> * geo_type = state
#> * time_type = day
#> * as_of = 2022-05-31 19:08:25.791826
#>
#> # A tibble: 4 × 5
#> geo_value time_value .pred forecast_date target_date
#> * <chr> <date> <dbl> <date> <date>
#> 1 ak 2021-12-31 0 2021-12-31 2022-01-07
#> 2 ca 2021-12-31 0.0989 2021-12-31 2022-01-07
#> 3 ny 2021-12-31 0.289 2021-12-31 2022-01-07
#> 4 sc 2021-12-31 0.352 2021-12-31 2022-01-07Internally, this works by removing the old frosting followed by adding the new frosting, just like when we update a recipe or model.
update_frosting
#> function (x, frosting, ...)
#> {
#> rlang::check_dots_empty()
#> x <- remove_frosting(x)
#> add_frosting(x, frosting)
#> }
#> <bytecode: 0x55e127e46440>
#> <environment: namespace:epipredict>If we decide that we do not want the frosting
post-processing at all, we can remove the frosting object
from the workflow and make predictions as follows:
wf3 <- wf2 %>% remove_frosting()
p3 <- forecast(wf3)
p3
#> An `epi_df` object, 4 x 3 with metadata:
#> * geo_type = state
#> * time_type = day
#> * as_of = 2022-05-31 19:08:25.791826
#>
#> # A tibble: 4 × 3
#> geo_value time_value .pred
#> * <chr> <date> <dbl>
#> 1 ak 2021-12-31 -0.206
#> 2 ca 2021-12-31 0.0989
#> 3 ny 2021-12-31 0.289
#> 4 sc 2021-12-31 0.352You can see that the above results from p3 are the same
as from p1, when we simply have a prediction layer in the
frosting post-processing container.
Adjust a single step of an epi_recipe
Suppose that we just want to change a single step in an
epi_recipe (that is either standalone or a part of an
epi_workflow). Instead of replacing an entire
epi_recipe, we can use the adjust_epi_recipe()
function. In this function, the step to be adjusted is indicated either
the step number or name in the which_step parameter. Then,
the parameter name and update value must be inputted as
....
For instance, suppose that we decide to lead the
death_rate by 14 days instead of 7. We may adjust this step
in wf recipe by setting which_step to the step
number in the order of operations, which can be obtained by inspecting
r2 or the tidy summary of it:
workflows::extract_preprocessor(wf) # step_epi_ahead is the third step in r2
#>
#> ── Epi Recipe ──────────────────────────────────────────────────────────────────
#>
#> ── Inputs
#> Number of variables by role
#> raw: 2
#> geo_value: 1
#> time_value: 1
#>
#> ── Operations
#> 1. Lagging: death_rate by 0, 1, 7, 14
#> 2. Lagging: case_rate by 0, 1, 2, 3, 4, 5, 6, 7, 14
#> 3. Leading: death_rate by 7
#> 4. • Removing rows with NA values in: all_predictors()
#> 5. • Removing rows with NA values in: all_outcomes()
tidy(workflows::extract_preprocessor(wf)) # tidy tibble summary of r2
#> # A tibble: 5 × 6
#> number operation type trained skip id
#> <int> <chr> <chr> <lgl> <lgl> <chr>
#> 1 1 step epi_lag FALSE FALSE epi_lag_9QeuR
#> 2 2 step epi_lag FALSE FALSE epi_lag_8HfdV
#> 3 3 step epi_ahead FALSE FALSE epi_ahead_HIyvQ
#> 4 4 step naomit FALSE FALSE naomit_vjF0H
#> 5 5 step naomit FALSE TRUE naomit_gF0fi
wf <- wf %>% adjust_epi_recipe(which_step = 3, ahead = 14)Alternatively, we may adjust that step by name by specifying the full
name of the step, step_epi_ahead, in
which_step:
wf %>% adjust_epi_recipe(which_step = "step_epi_ahead", ahead = 14) # not overwrite r2 because same result
#>
#> ══ Epi Workflow ════════════════════════════════════════════════════════════════
#> Preprocessor: Recipe
#> Model: linear_reg()
#> Postprocessor: None
#>
#> ── Preprocessor ────────────────────────────────────────────────────────────────
#>
#> 5 Recipe steps.
#> 1. step_epi_lag()
#> 2. step_epi_lag()
#> 3. step_epi_ahead()
#> 4. step_naomit()
#> 5. step_naomit()
#>
#> ── Model ───────────────────────────────────────────────────────────────────────
#>
#> Call:
#> stats::lm(formula = ..y ~ ., data = data)
#>
#> Coefficients:
#> (Intercept) lag_0_death_rate lag_1_death_rate lag_7_death_rate
#> 0.293291 -0.158682 0.055165 -0.349197
#> lag_14_death_rate lag_0_case_rate lag_1_case_rate lag_2_case_rate
#> -0.305136 0.009031 -0.009463 -0.005372
#> lag_3_case_rate lag_4_case_rate lag_5_case_rate lag_6_case_rate
#> -0.006244 0.004840 0.005537 -0.013347
#> lag_7_case_rate lag_14_case_rate
#> 0.011286 0.011721
#> If there are at least two steps in a recipe that share the same name,
specifying the name in which_step will throw an error as
adjust_epi_recipe() is not intended to be used to modify
multiple steps at once. The way, then, to modify a step that has the
same name as another is to indicate what number it is in the ordering of
the steps. For example, in r2 there are two steps named
step_epi_lag - the first step where we lag the death rate,
and the second where we lag the case rate. If we want to modify the lags
for the case_rate variable, we would specify the step
number of 2 in which_step.
wf <- wf %>% adjust_epi_recipe(which_step = 2, lag = c(0, 1, 7, 14, 21))
workflows::extract_preprocessor(wf)
#>
#> ── Epi Recipe ──────────────────────────────────────────────────────────────────
#>
#> ── Inputs
#> Number of variables by role
#> raw: 2
#> geo_value: 1
#> time_value: 1
#>
#> ── Operations
#> 1. Lagging: death_rate by 0, 1, 7, 14
#> 2. Lagging: case_rate by 0, 1, 7, 14, 21
#> 3. Leading: death_rate by 14
#> 4. • Removing rows with NA values in: all_predictors()
#> 5. • Removing rows with NA values in: all_outcomes()We could adjust a recipe directly in the same way as we adjust a
recipe in a workflow. The main difference is that we would not input
wf as the first argument to
adjust_epi_recipe() but rather r2.
adjust_epi_recipe(r2, which_step = 2, lag = c(0, 1, 7, 14, 21)) # should be same result as above
#>
#> ── Epi Recipe ──────────────────────────────────────────────────────────────────
#>
#> ── Inputs
#> Number of variables by role
#> raw: 2
#> geo_value: 1
#> time_value: 1
#>
#> ── Operations
#> 1. Lagging: death_rate by 0, 1, 7, 14
#> 2. Lagging: case_rate by 0, 1, 7, 14, 21
#> 3. Leading: death_rate by 7
#> 4. • Removing rows with NA values in: all_predictors()
#> 5. • Removing rows with NA values in: all_outcomes()Note that when we adjust the r2 object directly, we are
not adjusting the recipe in the epi_workflow. That is, if
we modify a step in r2, the change will not automatically
transfer over to wf. We would need to modify the recipe in
wf directly (adjust_epi_recipe() on
wf) or update the recipe in wf with a new
epi_recipe that has undergone the adjustment (using
update_epi_recipe()):
r2 <- adjust_epi_recipe(r2, which_step = 2, lag = 0:21)
workflows::extract_preprocessor(wf)
#>
#> ── Epi Recipe ──────────────────────────────────────────────────────────────────
#>
#> ── Inputs
#> Number of variables by role
#> raw: 2
#> geo_value: 1
#> time_value: 1
#>
#> ── Operations
#> 1. Lagging: death_rate by 0, 1, 7, 14
#> 2. Lagging: case_rate by 0, 1, 7, 14, 21
#> 3. Leading: death_rate by 14
#> 4. • Removing rows with NA values in: all_predictors()
#> 5. • Removing rows with NA values in: all_outcomes()Adjust a single layer of a frosting
Adjusting a layer of a frosting object proceeds in the
same way as adjusting a step in an epi_recipe does. So if
we want to change a single layer in a frosting (that is
either in a standalone object or part of an epi_workflow),
we can use the adjust_frosting() function wherein the layer
to be adjusted is indicated by either its number or name in the
which_layer parameter. In addition, the argument name and
update value must be inputted as ....
Let’s work with the frosting object directly instead of working on it
through the epi_workflow in a simple, illustrative example.
Recall frosting f2 which has the following layers:
f2
#>
#> ── Frosting ────────────────────────────────────────────────────────────────────
#>
#> ── Layers
#> 1. Creating predictions: "<calculated>"
#> 2. Thresholding predictions: .pred to [0, Inf)
#> 3. Adding forecast date: "<calculated>"
#> 4. Adding target date: "<calculated>"Suppose that we decide to change the upper bound of the prediction
threshold to 10 instead of Inf. We can adjust this layer in
frosting object by setting which_layer to the layer number,
3 (which can be found by inspecting f2 or
tidy(f2)):
f2 <- f2 %>% adjust_frosting(which_layer = 2, upper = 10)
f2
#>
#> ── Frosting ────────────────────────────────────────────────────────────────────
#>
#> ── Layers
#> 1. Creating predictions: "<calculated>"
#> 2. Thresholding predictions: .pred to [0, 10]
#> 3. Adding forecast date: "<calculated>"
#> 4. Adding target date: "<calculated>"Alternatively, we may adjust that layer by specifying its full name,
layer_threshold, in which_layer, to achieve
the same result:
f2 %>% adjust_frosting(which_layer = "layer_threshold", upper = 10) # not overwrite f2 because same result
#>
#> ── Frosting ────────────────────────────────────────────────────────────────────
#>
#> ── Layers
#> 1. Creating predictions: "<calculated>"
#> 2. Thresholding predictions: .pred to [0, 10]
#> 3. Adding forecast date: "<calculated>"
#> 4. Adding target date: "<calculated>"On the tidy method to inspect an epi_recipe or a
frosting object
The tidy method, when used on an epi_recipe, will return
a data frame that contains specific overview information about the
recipe including the operation number, the operation class (either
“step” or “check”), the type of method, a boolean value to indicate
whether prep() has been used to estimate the operation, a
boolean value to indicate whether the step is applied when
bake() is called, and the id of the operation.
tidy(r2)
#> # A tibble: 5 × 6
#> number operation type trained skip id
#> <int> <chr> <chr> <lgl> <lgl> <chr>
#> 1 1 step epi_lag FALSE FALSE epi_lag_9QeuR
#> 2 2 step epi_lag FALSE FALSE epi_lag_8HfdV
#> 3 3 step epi_ahead FALSE FALSE epi_ahead_HIyvQ
#> 4 4 step naomit FALSE FALSE naomit_vjF0H
#> 5 5 step naomit FALSE TRUE naomit_gF0fiIn contrast, printing the epi_recipe object shows the
inputs (number and roles of the variables) as well as the ordering and a
brief written summary of the operations:
r2
#>
#> ── Epi Recipe ──────────────────────────────────────────────────────────────────
#>
#> ── Inputs
#> Number of variables by role
#> raw: 2
#> geo_value: 1
#> time_value: 1
#>
#> ── Operations
#> 1. Lagging: death_rate by 0, 1, 7, 14
#> 2. Lagging: case_rate by 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
#> 16,
#> 3. 17, 18, 19...
#> 4. Leading: death_rate by 7
#> 5. • Removing rows with NA values in: all_predictors()
#> 6. • Removing rows with NA values in: all_outcomes()This same general structure persists when we compare the output of a
frosting object to that of its tidy tibble. However, we no longer have
the output specific to a recipe such as the roles in the recipe output
and the trained and skip columns in tidy tibble for it. Thus, the output
of a frosting object and the tidy tibble are simplified in comparison to
those for an epi_recipe.
f
#>
#> ── Frosting ────────────────────────────────────────────────────────────────────
#>
#> ── Layers
#> 1. Creating predictions: "<calculated>"
tidy(f)
#> # A tibble: 1 × 4
#> number operation type id
#> <int> <chr> <chr> <chr>
#> 1 1 layer predict predict_default_qZ3vE