tune attempts to evaluate the candidate models in the
shortest amount of time. This vignette is a summary of those
approaches.
Sub-model speed-ups
For some types of models, such as boosted models or regularized
models, the number of models that are actually fit can be far
less than the number of models evaluated. For example, suppose a boosted
tree is fit with 1000 trees. Many boosting implementations let the user
make predictions for any number of trees less than what was originally
fit (1000 in this example). This “sub-model trick” can greatly speed up
the training time for many models (e.g. see this
example in the caret documentation).
In order to know what models allow this, the parsnip
package contains a multi_predict() function that enables
this feature. Printing the S3 methods for it lists the possible
models:
library(tidymodels)
methods("multi_predict")## [1] multi_predict._C5.0* multi_predict._earth*
## [3] multi_predict._elnet* multi_predict._glmnetfit*
## [5] multi_predict._lognet* multi_predict._multnet*
## [7] multi_predict._torch_mlp* multi_predict._train.kknn*
## [9] multi_predict._xgb.Booster* multi_predict.default*
## see '?methods' for accessing help and source code
# There are arguments for the parameter(s) that can create multiple predictions.
# For xgboost, `trees` are cheap to evaluate:
parsnip:::multi_predict._xgb.Booster %>%
formals() %>%
names()## [1] "object" "new_data" "type" "trees" "..."The same feature does not exist for recipes though.
Expensive pre-processing
When tuning a recipe and a model, it makes sense to avoid recreating the recipe for each model.
For example, suppose that Isomap multi-dimensional scaling is used to pre-process the data prior to tuning a K-nearest neighbor regression:
data(Chicago)
iso_rec <-
recipe(ridership ~ ., data = Chicago) %>%
step_dummy(all_nominal()) %>%
step_isomap(all_predictors(), num_terms = tune())## 3 packages (dimRed, RSpectra, and RANN) are needed for this step but are
## not installed.## To install run: `install.packages(c("dimRed", "RSpectra", "RANN"))`
knn_mod <-
nearest_neighbor(neighbors = tune(), weight_func = tune()) %>%
set_engine("kknn") %>%
set_mode("regression")With the following grid:
grid <-
parameters(num_terms(c(1, 9)), neighbors(), weight_func()) %>%
grid_regular(levels = c(5, 10, 7)) %>%
arrange(num_terms, neighbors, weight_func)
grid ## # A tibble: 350 × 3
## num_terms neighbors weight_func
## <int> <int> <chr>
## 1 1 1 biweight
## 2 1 1 cos
## 3 1 1 epanechnikov
## 4 1 1 inv
## 5 1 1 rectangular
## 6 1 1 triangular
## 7 1 1 triweight
## 8 1 2 biweight
## 9 1 2 cos
## 10 1 2 epanechnikov
## # ℹ 340 more rowsTo evaluate these 350 candidate models, we would have to compute the same recipe 70 times per resample. Since Isomap is expensive, this is really inefficient.
tune_grid() determines when this occurs and fits all 70
candidate models for each unique configuration of the recipe. In
essence, it nests the model parameters inside the unique parameters of
the recipe:
## # A tibble: 5 × 2
## num_terms data
## <int> <list>
## 1 1 <tibble [70 × 2]>
## 2 3 <tibble [70 × 2]>
## 3 5 <tibble [70 × 2]>
## 4 7 <tibble [70 × 2]>
## 5 9 <tibble [70 × 2]>When parallel_over = "resamples", the default, only 5
recipes are prepared and, within each, all of the appropriate models are
fit from the same recipe. In this example, once the recipe with
num_terms = 1 is created, the model parameters are
iteratively tuned:
## # A tibble: 70 × 2
## neighbors weight_func
## <int> <chr>
## 1 1 biweight
## 2 1 cos
## 3 1 epanechnikov
## 4 1 inv
## 5 1 rectangular
## 6 1 triangular
## 7 1 triweight
## 8 2 biweight
## 9 2 cos
## 10 2 epanechnikov
## # ℹ 60 more rowsThe same will be true for post-processing parameters being tuned. For each unique set of recipe and model parameters, the post-processing parameters will be evaluated without unnecessary re-fitting.
Also, when using a model formula, the model matrix is only created once per resample.
Parallel processing
tune allows users, when possible, to use multiple cores
or separate machines fit models. The package is currently able to
parallelize over either the resampling loop of grid search (via
parallel_over = "resamples" in control_grid(),
the default) or both the resampling and preprocessing loops (via
parallel_over = "everything"). When
parallel_over = "everything", an outer parallel loop will
iterate over resamples and an inner parallel loop will iterate over all
unique combinations of preprocessor and model tuning parameters for that
specific resample. This will result in the preprocessor being
re-processed multiple times, but can be faster if that preprocessing is
extremely fast.
tune supports parallel processing using the future framework. future
supports a variety of technologies to share the computations and the
choice of technology is determined by the chosen plan. To run
tuning code in parallel, just provide a plan() and tune
will take care of the rest. For example:
foreach (legacy)
Before the 1.2.0 release, tune supported parallelism using the
foreach framework. Support has been deprecated for backends registered
for foreach and will be fully removed in an upcoming release. foreach
can use a variety of technologies to share the computations and the
choice of technology is determined by which parallel backend
package is chosen (aka the do{technology} packages). For
example, the doMC package uses the forking mechanism on
Unix-like systems to split the computations across multiple cores. As of
this writing, the backend packages are doMC,
doMPI, doParallel, doRedis,
doSNOW, and doAzureParallel (GitHub only).
Registering a parallel backend is also somewhat dependent of the
package. For doParallel, one could use:
all_cores <- parallel::detectCores(logical = FALSE)
library(doParallel)
cl <- makePSOCKcluster(all_cores)
registerDoParallel(cl)To transition from foreach to future, remove lines
loading any foreach backend packages library(do*) as well
as lines registering that technology with registerDo*(),
and add the following lines:
Switch out multisession for another strategy if
desired.
In practice
One downside to parallel processing is that the different
technologies handle inputs and outputs differently. For example,
multicore forking tends to carry the loaded packages and
objects into the worker processes. Others do not. To make sure that the
correct packages are loaded (but not attached) in the workers is to use
the pkg option in control_grid().
Some helpful advice to avoid errors in parallel processing is to not use variables in the global environment. These may not be found when the code is run inside of a worker process. For example:
num_pcs <- 3
recipe(mpg ~ ., data = mtcars) %>%
# Bad since num_pcs might not be found by a worker process
step_pca(all_predictors(), num_comp = num_pcs)
recipe(mpg ~ ., data = mtcars) %>%
# Good since the value is injected into the object
step_pca(all_predictors(), num_comp = !!num_pcs)This issue is likely to occur if dplyr::one_of() is used
as a sector.
Also note that almost all of the logging provided by
tune_grid() will not be seen when running in parallel.
Again, this is dependent on the backend package and technology being
used.