Create a reference / cheat sheet mapping engine-specific parameters to parsnip ones?

[khughitt]

As I'm going through converting algorithm-specific logic to generalized parsnip model specifications, one thing I am thinking that could be useful would be to create one or more reference tables or "cheat sheets" in the parsnip docs describing how to map from implemention-specific parameters (e.g. "eta" in xgboost) to parsnip model parameters (e.g. "learn_rate" in parsnip::boost_tree()).

The simplest form would probably just be a collection of tables (one per model), with columns for each supported engine + a final column corresponding to the parsnip parameter names.

Since this could take a bit of time to add, perhaps if you think this sort of thing would be useful, you could decide on an appropriate location/format, and add placeholders.

Then, myself and other users could help fill it in as we use the models.

[khughitt]

Actually, it looks like the necessary mappings could probably be parsed from the R code (e.g. boost_tree_data.R) in a pretty straight-forward manner. Might be quicker / less error-prone than attempting to manually generate the tables..

[thohan88]

I just opened a similar issue (#238) for the exact same purpose. Based on #235, I learned you can use parsnip::get_model_env to list all arguments. I think it would be useful if the mapping between original and standardized arguments was documented in the function itself and/or in a complete reference table or cheat sheet as suggested, e.g. for boosted trees:

model	parsnip	C5.0	spark	xgboost
boost_tree	learn_rate	NA	step_size	eta
boost_tree	loss_reduction	NA	NA	gamma
boost_tree	min_info_gain	NA	gamma	NA
boost_tree	min_n	minCases	min_instances_per_node	min_child_weight
boost_tree	mtry	NA	feature_subset_strategy	colsample_bytree
boost_tree	sample_size	sample	subsampling_rate	subsample
boost_tree	tree_depth	NA	max_depth	max_depth
boost_tree	trees	trials	max_iter	nrounds

Example code for generating said tables below.

Code

library(tidyverse)
envir <- parsnip::get_model_env()

args <- ls(envir) %>% 
tibble(name = .) %>% 
filter(str_detect(name, "args")) %>% 
mutate(model = str_replace(name, "_args", ""),
       args  = map(name, ~envir[[.x]])) %>% 
unnest(args) %>% 
select(model:original)

Full mapping table

params %>% 
  knitr::kable()

model	engine	parsnip	original
boost_tree	xgboost	tree_depth	max_depth
boost_tree	xgboost	trees	nrounds
boost_tree	xgboost	learn_rate	eta
boost_tree	xgboost	mtry	colsample_bytree
boost_tree	xgboost	min_n	min_child_weight
boost_tree	xgboost	loss_reduction	gamma
boost_tree	xgboost	sample_size	subsample
boost_tree	C5.0	trees	trials
boost_tree	C5.0	min_n	minCases
boost_tree	C5.0	sample_size	sample
boost_tree	spark	tree_depth	max_depth
boost_tree	spark	trees	max_iter
boost_tree	spark	learn_rate	step_size
boost_tree	spark	mtry	feature_subset_strategy
boost_tree	spark	min_n	min_instances_per_node
boost_tree	spark	min_info_gain	gamma
boost_tree	spark	sample_size	subsampling_rate
decision_tree	rpart	tree_depth	maxdepth
decision_tree	rpart	min_n	minsplit
decision_tree	rpart	cost_complexity	cp
decision_tree	C5.0	min_n	minCases
decision_tree	spark	tree_depth	max_depth
decision_tree	spark	min_n	min_instances_per_node
linear_reg	glmnet	penalty	lambda
linear_reg	glmnet	mixture	alpha
linear_reg	spark	penalty	reg_param
linear_reg	spark	mixture	elastic_net_param
logistic_reg	glmnet	penalty	lambda
logistic_reg	glmnet	mixture	alpha
logistic_reg	spark	penalty	reg_param
logistic_reg	spark	mixture	elastic_net_param
logistic_reg	keras	decay	decay
mars	earth	num_terms	nprune
mars	earth	prod_degree	degree
mars	earth	prune_method	pmethod
mlp	keras	hidden_units	hidden_units
mlp	keras	penalty	penalty
mlp	keras	dropout	dropout
mlp	keras	epochs	epochs
mlp	keras	activation	activation
mlp	nnet	hidden_units	size
mlp	nnet	penalty	decay
mlp	nnet	epochs	maxit
multinom_reg	glmnet	penalty	lambda
multinom_reg	glmnet	mixture	alpha
multinom_reg	spark	penalty	reg_param
multinom_reg	spark	mixture	elastic_net_param
multinom_reg	keras	decay	decay
nearest_neighbor	kknn	neighbors	ks
nearest_neighbor	kknn	weight_func	kernel
nearest_neighbor	kknn	dist_power	distance
rand_forest	ranger	mtry	mtry
rand_forest	ranger	trees	num.trees
rand_forest	ranger	min_n	min.node.size
rand_forest	randomForest	mtry	mtry
rand_forest	randomForest	trees	ntree
rand_forest	randomForest	min_n	nodesize
rand_forest	spark	mtry	feature_subset_strategy
rand_forest	spark	trees	num_trees
rand_forest	spark	min_n	min_instances_per_node
surv_reg	flexsurv	dist	dist
surv_reg	survival	dist	dist
svm_poly	kernlab	cost	C
svm_poly	kernlab	degree	degree
svm_poly	kernlab	scale_factor	scale
svm_poly	kernlab	margin	epsilon
svm_rbf	kernlab	cost	C
svm_rbf	kernlab	rbf_sigma	sigma
svm_rbf	kernlab	margin	epsilon

Mapping table by model

args %>% 
split(.$model) %>% 
map(spread, engine, original) %>% 
map(knitr::kable)

$boost_tree

model	parsnip	C5.0	spark	xgboost
boost_tree	learn_rate	NA	step_size	eta
boost_tree	loss_reduction	NA	NA	gamma
boost_tree	min_info_gain	NA	gamma	NA
boost_tree	min_n	minCases	min_instances_per_node	min_child_weight
boost_tree	mtry	NA	feature_subset_strategy	colsample_bytree
boost_tree	sample_size	sample	subsampling_rate	subsample
boost_tree	tree_depth	NA	max_depth	max_depth
boost_tree	trees	trials	max_iter	nrounds

$decision_tree

model	parsnip	C5.0	rpart	spark
decision_tree	cost_complexity	NA	cp	NA
decision_tree	min_n	minCases	minsplit	min_instances_per_node
decision_tree	tree_depth	NA	maxdepth	max_depth

$linear_reg

model	parsnip	glmnet	spark
linear_reg	mixture	alpha	elastic_net_param
linear_reg	penalty	lambda	reg_param

$logistic_reg

model	parsnip	glmnet	keras	spark
logistic_reg	decay	NA	decay	NA
logistic_reg	mixture	alpha	NA	elastic_net_param
logistic_reg	penalty	lambda	NA	reg_param

$mars

model	parsnip	earth
mars	num_terms	nprune
mars	prod_degree	degree
mars	prune_method	pmethod

$mlp

model	parsnip	keras	nnet
mlp	activation	activation	NA
mlp	dropout	dropout	NA
mlp	epochs	epochs	maxit
mlp	hidden_units	hidden_units	size
mlp	penalty	penalty	decay

$multinom_reg

model	parsnip	glmnet	keras	spark
multinom_reg	decay	NA	decay	NA
multinom_reg	mixture	alpha	NA	elastic_net_param
multinom_reg	penalty	lambda	NA	reg_param

$nearest_neighbor

model	parsnip	kknn
nearest_neighbor	dist_power	distance
nearest_neighbor	neighbors	ks
nearest_neighbor	weight_func	kernel

$rand_forest

model	parsnip	randomForest	ranger	spark
rand_forest	min_n	nodesize	min.node.size	min_instances_per_node
rand_forest	mtry	mtry	mtry	feature_subset_strategy
rand_forest	trees	ntree	num.trees	num_trees

$surv_reg

model	parsnip	flexsurv	survival
surv_reg	dist	dist	dist

$svm_poly

model	parsnip	kernlab
svm_poly	cost	C
svm_poly	degree	degree
svm_poly	margin	epsilon
svm_poly	scale_factor	scale

$svm_rbf

model	parsnip	kernlab
svm_rbf	cost	C
svm_rbf	margin	epsilon
svm_rbf	rbf_sigma	sigma

[topepo]

Thanks! I added a subsection to the new Models vignette using your code. Let me know if you want any changes.

[juliasilge]

Just noticed nothing is showing up for stan in the existing tables.

[topepo]

No tuning parameters for those models. The same happens in linear_reg() since there is no lm column.

[github-actions]

This issue 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.