Add defaults to engine specific params in model docs

[juliasilge]

This PR closes #211.

What I have so far here is my initial idea of how to add the default values for each engine specific parameter into the tables in the docs. These values are only needed for devtools::document() so I am thinking let's make little internal functions via paste0() rather than get_model_env() and add all these values via some set_... function and keep them around.

What are your thoughts?

Current results for the one model implemented right now:

library(parsnip)
convert_args("decision_tree")
#> Joining, by = c("model", "engine", "parsnip", "original")

parsnip	rpart	C5.0	spark
tree_depth	maxdepth (30)	NA	max_depth (5)
min_n	minsplit (20)	minCases (2)	min_instances_per_node (1)
cost_complexity	cp (0.01)	NA	NA

^{Created on 2020-05-27 by the reprex package (v0.3.0)}

[apreshill]

@juliasilge I believe this also would close #235? 🎉

[juliasilge]

For now I'm only adding these to the tables in the docs, with functions that are only run during devtools::document(). So no user-facing functions quite yet.

Users could look at the docs and get the values but we haven't moved quite all the way to a dynamic lookup.

[juliasilge]

We're going to have to make a call on what to do with the rand_forest() table, for mtry and min_n.

For ranger's mtry:

Number of variables to possibly split at in each node. Default is the (rounded down) square root
of the number variables. Alternatively, a single argument function returning an integer, given the 
number of independent variables.

For ranger's min.node.size:

Minimal node size. Default 1 for classification, 5 for regression, 3 for survival, and 10 for 
probability.

Neither of these are available via formals().

For randomForest, the function argument themselves are defined as:

mtry=if (!is.null(y) && !is.factor(y))
             max(floor(ncol(x)/3), 1) else floor(sqrt(ncol(x)))

nodesize = if (!is.null(y) && !is.factor(y)) 5 else 1

I think I'm going to put some of these in manually.

[topepo]

Yes. I would add them as expressions to your data structure. Maybe something like "1C|5R|3S" with a legend on the bottom?

[juliasilge]

The issues around rand_forest() and SVM make me think #235 is going to be quite difficult.

[juliasilge]

The SVM options are less than ideal. They are hard-coded mostly, and I dug around in the liquidSVM code and came up with... this:

> convert_args("svm_rbf")

parsnip	kernlab	liquidSVM
cost	C (1)	lambdas (varies)
rbf_sigma	sigma (1)	gammas (varies)
margin	epsilon (0.1)	NA

[juliasilge]

I updated the sigma value for kernlab because although rbfdot() has a default of 1, during training the value used depends on the data.

[github-actions]

This pull request has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.