Old recipe no longer performs the step_pca

[tomazweiss]

The problem

I've found this problem you should probably be aware of:
Back in august 2021 I've created and saved a recipe with several steps, including the step_pca. I have been using the latest recipes package that was available at that time.

rec_obj <- recipe(~ ., data) %>% 
  update_role(username, new_role = "sample ID") %>% 
  step_YeoJohnson(all_predictors()) %>% 
  step_center(all_predictors()) %>% 
  step_scale(all_predictors()) %>% 
  step_pca(all_predictors(), threshold = 0.65)

trained_rec <- prep(rec_obj, training = data)

saveRDS(trained_rec, "models/trained_rec.rds")

Since then, I've been using this saved recipe in scoring process for several months without any issues. Last month I've upgraded to the latest recipes version 0.2.0 and scoring no longer works:

trained_rec <- readRDS("models/trained_rec.rds")

# baked data has the same number of variables as data, all steps except step_pca seem to be executed
baked_data <- bake(trained_rec, new_data = data)

If I print trained_rec , I get the following:

trained_rec

Recipe

Inputs:

      role #variables
 predictor        104
 sample ID          1

Training data contained 29546 data points and no missing data.

Operations:

Yeo-Johnson transformation on ... [trained]
Centering for ... [trained]
Scaling for ... [trained]
No PCA components were extracted from <none> [trained]

PCA step has somehow been lost.

The solution

Downgrading the package version solves my problem:

require(devtools)
install_version("recipes", version = "0.1.17", repos = "http://cran.us.r-project.org")

Reproducible example

Sorry, but I can't (or don't know how to) write one since two versions of the recipe package are involved...

I'm using R version 4.0.4 on Windows.

[juliasilge]

Oh, that is so frustrating to hear, especially because we have invested energy in keeping recipes trained with previous versions working. 😕 We did not believe the breaking changes in recipes 0.2.0 impact step_pca().

When you say:

scoring no longer works

do you get an error? Or is this a silent failure where PCA is not applied?

[tomazweiss]

I get no error or warning.

Here is a reprex (that messes with your installation).

library(tidyverse)
library(janitor)
library(devtools)


data <- iris %>% clean_names() %>% select(-species)


# start with version 0.1.17
# install_version("recipes", version = "0.1.17", repos = "http://cran.us.r-project.org")

library(recipes)

rec_obj <- recipe(~ ., data) %>% 
  step_YeoJohnson(all_predictors()) %>% 
  step_center(all_predictors()) %>% 
  step_scale(all_predictors()) %>% 
  step_pca(all_predictors(), threshold = 0.65)

trained_rec <- prep(rec_obj, training = data)
saveRDS(trained_rec, "trained_rec.rds")

trained_rec

# Recipe
# 
# Inputs:
#   
#   role #variables
# predictor          4
# 
# Training data contained 150 data points and no missing data.
# 
# Operations:
#   
#   Yeo-Johnson transformation on sepal_length, sepal_width, petal_length, petal_width [trained]
# Centering for sepal_length, sepal_width, petal_length, petal_width [trained]
# Scaling for sepal_length, sepal_width, petal_length, petal_width [trained]
# PCA extraction with sepal_length, sepal_width, petal_length, petal_width [trained]



############### restart R



library(tidyverse)
# library(recipes)
library(janitor)
library(devtools)

# upgrade to 0.2.0
install_version("recipes", version = "0.2.0", repos = "http://cran.us.r-project.org")

library(recipes)

packageVersion("recipes")
# ‘0.2.0’


trained_rec <- readRDS("trained_rec.rds")
trained_rec

# Recipe
# 
# Inputs:
#   
#   role #variables
# predictor          4
# 
# Training data contained 150 data points and no missing data.
# 
# Operations:
#   
#   Yeo-Johnson transformation on sepal_length, sepal_width, petal_length, petal_width [trained]
# Centering for sepal_length, sepal_width, petal_length, petal_width [trained]
# Scaling for sepal_length, sepal_width, petal_length, petal_width [trained]
# No PCA components were extracted from <none> [trained]

[EmilHvitfeldt]

I found the cause of the problem. In the effort to support empty selectors #813, a column argument was added to step_pca() to store which columns are being used.

f1773ea#diff-9e3b01309aa3f13e7f0a8acd0fd70758daba477a76aa2062dc04e05b38891879

when using bake.step_pca() the following code is being used to check if PCA should be performed

if (length(object$columns) > 0 && !all(is.na(object$res$rotation))) {

which returns FALSE since the 0.1.17 recipe doesn't have a columns element.

If you have to repair. You can run the following code to the 0.1.17 recipe to make it 0.2.0 compliant. The 4 should be the number step your PCA is at.

pca_columns <- setNames(nm = rownames(trained_rec$steps[[4]]$res$rotation))
trained_rec$steps[[4]]$columns <- pca_columns

See reprex below (only 0.2.0 version)

# In version 0.2.0
library(tidyverse)
library(recipes)
library(janitor)

data <- iris %>% clean_names() %>% select(-species)

packageVersion("recipes")
#> [1] '0.2.0'
# ‘0.2.0’

trained_rec <- readRDS("trained_rec.rds")

# Repair
pca_columns <- setNames(nm = rownames(trained_rec$steps[[4]]$res$rotation))
trained_rec$steps[[4]]$columns <- pca_columns

trained_rec
#> Recipe
#> 
#> Inputs:
#> 
#>       role #variables
#>  predictor          4
#> 
#> Training data contained 150 data points and no missing data.
#> 
#> Operations:
#> 
#> Yeo-Johnson transformation on sepal_length, sepal_width, petal_lengt... [trained]
#> Centering for sepal_length, sepal_width, petal_length, petal_... [trained]
#> Scaling for sepal_length, sepal_width, petal_length, petal_... [trained]
#> PCA extraction with sepal_length, sepal_width, petal_length, petal_w... [trained]

baked_data <- bake(trained_rec, new_data = data)

baked_data
#> # A tibble: 150 × 1
#>      PC1
#>    <dbl>
#>  1 -2.26
#>  2 -2.12
#>  3 -2.43
#>  4 -2.39
#>  5 -2.38
#>  6 -1.99
#>  7 -2.51
#>  8 -2.24
#>  9 -2.47
#> 10 -2.23
#> # … with 140 more rows

^{Created on 2022-03-02 by the reprex package (v2.0.1)}

[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.