modelblueprint: a model-agnostic container for ML model lifecycles

Arguments

model: A fitted model object. Any class implementing predict().
train, test, holdout: Datasets as data.frame. Default NULL.
pre_process_fun: function(df) -> df. Pre-processing pipeline.
feat_eng_fun: function(df) -> df. Feature engineering pipeline.
post_process_fun: function(preds, df_raw) -> numeric. Post-processing.
x_original_inputs: [character] Original input feature names.
x_names: [character] Engineered feature names.
y_name: [character(1)] Target variable name.
yhat_name: [character(1)] Prediction column name.
expo_name: [character(1)] Exposure column name.
expo_val: [numeric(1)] Reference exposure value.
expo_0_rep: [numeric(1)] Replacement for zero-exposure rows.
offset_name: [character(1)] Offset variable name.
offset_value: [numeric(1)] Offset value.
model_display_name: [character(1)] Human-readable label.
deploy_notes: [character(1)] Deployment notes.

Construction

Create a blueprint by passing a fitted model plus, optionally, its data splits and metadata:

modelblueprint(
  model,
  train = NULL, test = NULL, holdout = NULL,
  pre_process_fun  = function(df) df,
  feat_eng_fun     = function(df) df,
  post_process_fun = function(preds, df_raw) preds,
  x_original_inputs = NA_character_, x_names = NA_character_,
  y_name = NA_character_, yhat_name = NA_character_,
  expo_name = "exposure", expo_val = 1, expo_0_rep = 0.1,
  offset_name = NA_character_, offset_value = NA_real_,
  model_display_name = NA_character_, deploy_notes = NA_character_
)

Only model is required; every other property has a sensible default. See the argument list and the example below for details.

Examples

# Wrap a fitted model together with its training data and metadata.
mb <- modelblueprint(
  model              = lm(mpg ~ wt + hp, data = mtcars),
  train              = mtcars,
  y_name             = "mpg",
  x_original_inputs  = c("wt", "hp"),
  model_display_name = "lm_mpg"
)

mb                       # print method shows a structured summary
#> ============================================================ 
#> modelblueprint
#> ============================================================ 
#>   Model:        lm
#>   Display name: lm_mpg
#>   Target:       mpg
#>   Exposure:     exposure (val = 1)
#>   Features:     2 original / 0 engineered
#> ------------------------------------------------------------ 
#>   Train rows:   32
#>   Test rows:    <not set>
#>   Holdout rows: <not set>
#> ============================================================ 
predict(mb, head(mtcars))
#> [1] 23.57233 22.58348 25.27582 21.26502 18.32727 20.47382