An object with class "model_spec" is a container for information about a model that will be fit.
The main elements of the object are:
args: A vector of the main arguments for the model. The
names of these arguments may be different from their
counterparts n the underlying model function. For example, for a
glmnet model, the argument name for the amount of the penalty
is called "penalty" instead of "lambda" to make it more general
and usable across different types of models (and to not be
specific to a particular model function). The elements of
varying(). If left to their defaults (
arguments will use the underlying model functions default value.
As discussed below, the arguments in
args are captured as
quosures and are not immediately executed.
...: Optional model-function-specific
parameters. As with
args, these will be quosures and can be
mode: The type of model, such as "regression" or
"classification". Other modes will be added once the package
adds more functionality.
method: This is a slot that is filled in later by the
model's constructor function. It generally contains lists of
information that are used to create the fit and prediction code
as well as required packages and similar data.
engine: This character string declares exactly what
software will be used. It can be a package name or a technology
This class and structure is the basis for how parsnip stores model objects prior to seeing the data.
An important detail to understand when creating model specifications is that they are intended to be functionally independent of the data. While it is true that some tuning parameters are data dependent, the model specification does not interact with the data at all.
For example, most R functions immediately evaluate their
arguments. For example, when calling
mean(dat_vec), the object
dat_vec is immediately evaluated inside of the function.
parsnip model functions do not do this. For example, using
rand_forest(mtry = ncol(mtcars) - 1)
does not execute
ncol(mtcars) - 1 when creating the specification.
This can be seen in the output:
> rand_forest(mtry = ncol(mtcars) - 1) Random Forest Model Specification (unknown) Main Arguments: mtry = ncol(mtcars) - 1
The model functions save the argument expressions and their
associated environments (a.k.a. a quosure) to be evaluated later
called with the actual data.
The consequence of this strategy is that any data required to get the parameter values must be available when the model is fit. The two main ways that this can fail is if:
The data have been modified between the creation of the model specification and when the model fit function is invoked.
If the model specification is saved and loaded into a new session where those same data objects do not exist.
The best way to avoid these issues is to not reference any data
objects in the global environment but to use data descriptors
.cols(). Another way of writing the previous
rand_forest(mtry = .cols() - 1)
This is not dependent on any specific data object and is evaluated immediately before the model fitting process begins.
One less advantageous approach to solving this issue is to use quasiquotation. This would insert the actual R object into the model specification and might be the best idea when the data object is small. For example, using
rand_forest(mtry = ncol(!!mtcars) - 1)
would work (and be reproducible between sessions) but embeds
the entire mtcars data set into the
> rand_forest(mtry = ncol(!!mtcars) - 1) Random Forest Model Specification (unknown) Main Arguments: mtry = ncol(structure(list(Sepal.Length = c(5.1, 4.9, 4.7, 4.6, 5, <snip>
However, if there were an object with the number of columns in it, this wouldn't be too bad:
> mtry_val <- ncol(mtcars) - 1 > mtry_val  10 > rand_forest(mtry = !!mtry_val) Random Forest Model Specification (unknown) Main Arguments: mtry = 10
More information on quosures and quasiquotation can be found at https://tidyeval.tidyverse.org.