Workflowy: łączenie recipes i modeli

Czym jest workflow?

library(workflows)

# Start an empty workflow
wf <- workflow()
print(wf)

# Workflows have two slots: preprocessor and model
# Both must be filled before fitting

add_recipe() i add_model()

library(recipes)
library(parsnip)
library(workflows)

rec <- recipe(price ~ ., data = train) |>
  step_impute_median(all_numeric_predictors()) |>
  step_dummy(all_nominal_predictors()) |>
  step_normalize(all_numeric_predictors())

spec <- linear_reg() |> set_engine('lm')

wf <- workflow() |>
  add_recipe(rec) |>
  add_model(spec)

print(wf)

fit() dla workflow

fitted_wf <- fit(wf, data = housing_train)

# The fitted workflow stores both the prepped recipe
# and the trained model
print(fitted_wf)

predict() dla dopasowanego workflow

# Predictions automatically preprocess the test data
preds <- predict(fitted_wf, new_data = housing_test)
head(preds)

# For classification
preds_prob <- predict(fitted_wf, new_data = test_df, type = 'prob')
preds_class <- predict(fitted_wf, new_data = test_df, type = 'class')

last_fit() — trenowanie na całości i ocena na teście

library(rsample)

split <- initial_split(housing_data, prop = 0.8, strata = price)

# Fit on training, evaluate on test
final_res <- last_fit(wf, split)

# View performance on the held-out test set
collect_metrics(final_res)

extract_fit_parsnip()

fitted_wf <- fit(wf, data = housing_train)

# Extract the parsnip model
parsnip_fit <- extract_fit_parsnip(fitted_wf)

# Now access the underlying model
tidy(parsnip_fit)      # coefficients for lm
vip::vip(parsnip_fit)  # variable importance plot

extract_recipe()

fitted_wf <- fit(wf, data = housing_train)

# Get the prepped recipe
prepped_rec <- extract_recipe(fitted_wf)

# Inspect normalization stats
tidy(prepped_rec, number = 3)  # step_normalize details

# Apply recipe to completely new external data
new_baked <- bake(prepped_rec, new_data = brand_new_df)

augment() dla workflow

fitted_wf <- fit(wf, data = housing_train)

# Append predictions to the test data frame
results <- augment(fitted_wf, new_data = housing_test)

# Now compute metrics directly
library(yardstick)
results |>
  metrics(truth = price, estimate = .pred)

Aktualizowanie workflow

# Original workflow with lm
wf_lm <- workflow() |>
  add_recipe(rec) |>
  add_model(linear_reg() |> set_engine('lm'))

# Swap model to random forest, keep same recipe
wf_rf <- update_model(
  wf_lm,
  rand_forest(trees = 300) |>
    set_engine('ranger') |>
    set_mode('regression')
)

fit_rf <- fit(wf_rf, data = housing_train)

add_formula() a add_recipe()

# Simple baseline — no recipe needed
baseline_wf <- workflow() |>
  add_formula(price ~ sqft + bedrooms + bathrooms) |>
  add_model(linear_reg() |> set_engine('lm'))

baseline_fit <- fit(baseline_wf, data = housing_train)
baseline_preds <- predict(baseline_fit, new_data = housing_test)

# Compare RMSE to recipe-based model
yardstick::rmse_vec(housing_test$price, baseline_preds$.pred)

Zestawy workflow do porównywania

library(workflowsets)

all_workflows <- workflow_set(
  preproc = list(basic = basic_rec, full = full_rec),
  models  = list(
    lm  = linear_reg() |> set_engine('lm'),
    rf  = rand_forest(trees = 200) |> set_engine('ranger') |> set_mode('regression')
  )
)

# Fit all 4 combinations on resamples
results <- workflow_map(all_workflows, 'fit_resamples', resamples = cv_folds)
autoplot(results)

Szybkie sprawdzenie

Podsumowanie workflow

# Canonical tidymodels workflow pattern
split    <- initial_split(data, prop = 0.8)
train    <- training(split)
test     <- testing(split)

rec  <- recipe(y ~ ., data = train) |> step_normalize(all_numeric_predictors())
spec <- rand_forest(trees = 300) |> set_engine('ranger') |> set_mode('regression')

wf   <- workflow() |> add_recipe(rec) |> add_model(spec)

final_res <- last_fit(wf, split)
collect_metrics(final_res)