prep, bake, juice三步说的真清楚
Let's walk through what each of these functions does. First, let's define a recipe with a couple of steps. Notice that the data going into the recipe is the training data.
library(recipes)cars_train <- mtcars[1:20,]cars_test <- mtcars[21:32,]cars_rec <- recipe(mpg ~ ., data = cars_train) %>% step_log(disp) %>% step_center(all_predictors())cars_rec#> Data Recipe#> #> Inputs:#> #> role #variables#> outcome 1#> predictor 10#> #> Operations:#> #> Log transformation on disp#> Centering for all_predictors
The preprocessing recipe cars_rec has been defined but no values have been estimated. For example, the log has not been taken for disp, and the mean has not been calculated for predictors so that they can be centered.
The prep() function takes that defined object and computes everything so that the preprocessing steps can be executed. For example, the mean of each predictor is calculated in this example so the predictors can be centered. This is done with the training data.
cars_prep <- prep(cars_rec)cars_prep#> Data Recipe#> #> Inputs:#> #> role #variables#> outcome 1#> predictor 10#> #> Training data contained 20 data points and no missing data.#> #> Operations:#> #> Log transformation on disp [trained]#> Centering for cyl, disp, hp, drat, wt, qsec, vs, am, ... [trained]
Notice that before, with the unprepped recipe, it just said Centering for all_predictors because it had not been evaluated yet. Now it has been evaluated and we know which columns are predictors and what their means are.
The bake() and juice() functions both return data, not a preprocessing recipe object. The bake() function takes a prepped recipe (one that has had all quantities estimated from training data) and applies it to new_data. That new_data could be the training data again...
bake(cars_prep, new_data = cars_train)#> # A tibble: 20 x 11#> cyl disp hp drat wt qsec vs am gear carb mpg#> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>#> 1 -0.2 -0.222 -26.2 0.355 -0.778 -1.98 -0.5 0.7 0.5 1.30 21 #> 2 -0.2 -0.222 -26.2 0.355 -0.523 -1.42 -0.5 0.7 0.5 1.30 21 #> 3 -2.2 -0.615 -43.2 0.305 -1.08 0.169 0.5 0.7 0.5 -1.7 22.8#> 4 -0.2 0.256 -26.2 -0.465 -0.183 0.999 0.5 -0.3 -0.5 -1.7 21.4#> 5 1.8 0.589 38.8 -0.395 0.0415 -1.42 -0.5 -0.3 -0.5 -0.7 18.7#> 6 -0.2 0.119 -31.2 -0.785 0.0615 1.78 0.5 -0.3 -0.5 -1.7 18.1#> 7 1.8 0.589 109. -0.335 0.172 -2.60 -0.5 -0.3 -0.5 1.30 14.3#> 8 -2.2 -0.309 -74.2 0.145 -0.208 1.56 0.5 -0.3 0.5 -0.7 24.4#> 9 -2.2 -0.350 -41.2 0.375 -0.248 4.46 0.5 -0.3 0.5 -0.7 22.8#> 10 -0.2 -0.176 -13.2 0.375 0.0415 -0.141 0.5 -0.3 0.5 1.30 19.2#> 11 -0.2 -0.176 -13.2 0.375 0.0415 0.459 0.5 -0.3 0.5 1.30 17.8#> 12 1.8 0.323 43.8 -0.475 0.672 -1.04 -0.5 -0.3 -0.5 0.300 16.4#> 13 1.8 0.323 43.8 -0.475 0.332 -0.841 -0.5 -0.3 -0.5 0.300 17.3#> 14 1.8 0.323 43.8 -0.475 0.382 -0.441 -0.5 -0.3 -0.5 0.300 15.2#> 15 1.8 0.860 68.8 -0.615 1.85 -0.461 -0.5 -0.3 -0.5 1.30 10.4#> 16 1.8 0.834 78.8 -0.545 2.03 -0.621 -0.5 -0.3 -0.5 1.30 10.4#> 17 1.8 0.790 93.8 -0.315 1.95 -1.02 -0.5 -0.3 -0.5 1.30 14.7#> 18 -2.2 -0.932 -70.2 0.535 -1.20 1.03 0.5 0.7 0.5 -1.7 32.4#> 19 -2.2 -0.970 -84.2 1.38 -1.78 0.079 0.5 0.7 0.5 -0.7 30.4#> 20 -2.2 -1.03 -71.2 0.675 -1.56 1.46 0.5 0.7 0.5 -1.7 33.9
Or it could be the testing data. In this case, the column means from the training data are applied to the testing data, because that is what happens IRL in a modeling workflow. To do otherwise is data leakage.
bake(cars_prep, new_data = cars_test)#> # A tibble: 12 x 11#> cyl disp hp drat wt qsec vs am gear carb mpg#> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>#> 1 -2.2 -0.509 -39.2 0.155 -0.933 1.57 0.5 -0.3 -0.5 -1.7 21.5#> 2 1.8 0.465 13.8 -0.785 0.122 -1.57 -0.5 -0.3 -0.5 -0.7 15.5#> 3 1.8 0.420 13.8 -0.395 0.0366 -1.14 -0.5 -0.3 -0.5 -0.7 15.2#> 4 1.8 0.561 109. 0.185 0.442 -3.03 -0.5 -0.3 -0.5 1.30 13.3#> 5 1.8 0.694 38.8 -0.465 0.447 -1.39 -0.5 -0.3 -0.5 -0.7 19.2#> 6 -2.2 -0.928 -70.2 0.535 -1.46 0.459 0.5 0.7 0.5 -1.7 27.3#> 7 -2.2 -0.507 -45.2 0.885 -1.26 -1.74 -0.5 0.7 1.5 -0.7 26 #> 8 -2.2 -0.742 -23.2 0.225 -1.89 -1.54 0.5 0.7 1.5 -0.7 30.4#> 9 1.8 0.564 128. 0.675 -0.228 -3.94 -0.5 0.7 1.5 1.30 15.8#> 10 -0.2 -0.320 38.8 0.075 -0.628 -2.94 -0.5 0.7 1.5 3.3 19.7#> 11 1.8 0.410 199. -0.00500 0.172 -3.84 -0.5 0.7 1.5 5.3 15 #> 12 -2.2 -0.501 -27.2 0.565 -0.618 0.159 0.5 0.7 0.5 -0.7 21.4
The juice() function is a nice little shortcut. Because the prepped recipe was estimated from the training data, you can process the training data only from it. Picture yourself squeezing the prepped recipe to get the training data back out that you used to estimate the preprocessing parameters with to start with.
juice(cars_prep)#> # A tibble: 20 x 11#> cyl disp hp drat wt qsec vs am gear carb mpg#> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>#> 1 -0.2 -0.222 -26.2 0.355 -0.778 -1.98 -0.5 0.7 0.5 1.30 21 #> 2 -0.2 -0.222 -26.2 0.355 -0.523 -1.42 -0.5 0.7 0.5 1.30 21 #> 3 -2.2 -0.615 -43.2 0.305 -1.08 0.169 0.5 0.7 0.5 -1.7 22.8#> 4 -0.2 0.256 -26.2 -0.465 -0.183 0.999 0.5 -0.3 -0.5 -1.7 21.4#> 5 1.8 0.589 38.8 -0.395 0.0415 -1.42 -0.5 -0.3 -0.5 -0.7 18.7#> 6 -0.2 0.119 -31.2 -0.785 0.0615 1.78 0.5 -0.3 -0.5 -1.7 18.1#> 7 1.8 0.589 109. -0.335 0.172 -2.60 -0.5 -0.3 -0.5 1.30 14.3#> 8 -2.2 -0.309 -74.2 0.145 -0.208 1.56 0.5 -0.3 0.5 -0.7 24.4#> 9 -2.2 -0.350 -41.2 0.375 -0.248 4.46 0.5 -0.3 0.5 -0.7 22.8#> 10 -0.2 -0.176 -13.2 0.375 0.0415 -0.141 0.5 -0.3 0.5 1.30 19.2#> 11 -0.2 -0.176 -13.2 0.375 0.0415 0.459 0.5 -0.3 0.5 1.30 17.8#> 12 1.8 0.323 43.8 -0.475 0.672 -1.04 -0.5 -0.3 -0.5 0.300 16.4#> 13 1.8 0.323 43.8 -0.475 0.332 -0.841 -0.5 -0.3 -0.5 0.300 17.3#> 14 1.8 0.323 43.8 -0.475 0.382 -0.441 -0.5 -0.3 -0.5 0.300 15.2#> 15 1.8 0.860 68.8 -0.615 1.85 -0.461 -0.5 -0.3 -0.5 1.30 10.4#> 16 1.8 0.834 78.8 -0.545 2.03 -0.621 -0.5 -0.3 -0.5 1.30 10.4#> 17 1.8 0.790 93.8 -0.315 1.95 -1.02 -0.5 -0.3 -0.5 1.30 14.7#> 18 -2.2 -0.932 -70.2 0.535 -1.20 1.03 0.5 0.7 0.5 -1.7 32.4#> 19 -2.2 -0.970 -84.2 1.38 -1.78 0.079 0.5 0.7 0.5 -0.7 30.4#> 20 -2.2 -1.03 -71.2 0.675 -1.56 1.46 0.5 0.7 0.5 -1.7 33.9
Created on 2020-06-04 by the reprex package (v0.3.0)