Tidymodels update 2021

Makefile

@@ -6,15 +6,15 @@ all: results/final_model.rds results/accuracy_vs_k.png results/predictor_distrib
 
 # download data
 data/raw/wdbc.feather: src/download_data.py
-	python src/download_data.py --out_type=feather --url=http://mlr.cs.umass.edu/ml/machine-learning-databases/breast-cancer-wisconsin/wdbc.data --out_file=data/raw/wdbc.feather
+	python src/download_data.py --out_type=feather --url=https://archive.ics.uci.edu/ml/machine-learning-databases/breast-cancer-wisconsin/wdbc.data --out_file=data/raw/wdbc.feather 
 
 # pre-process data (e.g., scale and split into train & test)
 data/processed/training.feather data/processed/test.feather scale_factor.rds: src/pre_process_wisc.r data/raw/wdbc.feather
 	Rscript src/pre_process_wisc.r --input=data/raw/wdbc.feather --out_dir=data/processed 
 
 # exploratory data analysis - visualize predictor distributions across classes
 results/predictor_distributions_across_class.png: src/eda_wisc.r data/processed/training.feather
-	Rscript src/eda_wisc.r --train=data/processed/training.feather --out_dir=results
+	Rscript src/eda_wisc.r --train=data/processed/training.feather --out_dir=src/breast_cancer_eda_files/figure-gfm
 
 # tune model (here, find K for k-nn using 30 fold cv with Cohen's Kappa)
 results/final_model.rds results/accuracy_vs_k.png: src/fit_breast_cancer_predict_model.r data/processed/training.feather

src/breast_cancer_test_results.r

@@ -22,11 +22,11 @@ main <- function(test, out_dir) {
 
   # Load and wrangle test data ----------------------------------------------
   test_data <- read_feather(test) 
-  x_test <- test_data %>% 
+  x_test <- test_data |> 
     select(-class, -se_fractal_dimension, -se_smoothness, -se_symmetry, -se_texture)
-  y_test <- test_data %>% 
-    select(class) %>% 
-    mutate(class = as.factor(class)) %>% 
+  y_test <- test_data |> 
+    select(class) |> 
+    mutate(class = as.factor(class))  |>  
     pull()
 
   # Load model and predict --------------------------------------------------

src/eda_wisc.r

@@ -11,9 +11,8 @@ Options:
 --out_dir=<out_dir> Path to directory where the plots should be saved
 " -> doc
 
-library(feather)
+library(arrow)
 library(tidyverse)
-library(caret)
 library(docopt)
 library(ggridges)
 library(ggthemes)
@@ -24,9 +23,9 @@ opt <- docopt(doc)
 main <- function(train, out_dir) {
 
   # visualize predictor distributions by class
-  train_data <- read_feather(train) %>% 
-    gather(key = predictor, value = value, -class) %>% 
-    mutate(predictor = str_replace_all(predictor, "_", " ")) %>% 
+  train_data <- arrow::read_feather(train)  |>  
+    pivot_longer(names_to = "predictor", values_to = "value", -class) |> 
+    mutate(predictor = str_replace_all(predictor, "_", " ")) |> 
     ggplot(aes(x = value, y = class, colour = class, fill = class)) +
     facet_wrap(. ~ predictor, scale = "free", ncol = 4) +
     geom_density_ridges(alpha = 0.8) +

src/fit_breast_cancer_predict_model.r

@@ -24,11 +24,11 @@ main <- function(train, out_dir) {
   # Tune hyperparameters ----------------------------------------------------
 
   train_data <- read_feather(train) 
-  x_train <- train_data %>% 
+  x_train <- train_data |> 
     select(-class, -se_fractal_dimension, -se_smoothness, -se_symmetry, -se_texture)
-  y_train <- train_data %>% 
-    select(class) %>% 
-    mutate(class = as.factor(class)) %>% 
+  y_train <- train_data |> 
+    select(class)  |>  
+    mutate(class = as.factor(class)) |> 
     pull()
   k = data.frame(k = seq(1, 100, by = 2))
   cross_val <- trainControl(method="cv", number = 30)
@@ -37,7 +37,7 @@ main <- function(train, out_dir) {
 
   # Visualize kappa for K's ----------------------------------------------
 
-  kappa_vs_k <- model_cv_30fold$results %>% 
+  kappa_vs_k <- model_cv_30fold$results |> 
     ggplot(aes(x = k, y = Kappa)) +
       geom_point() +
       geom_errorbar(aes(ymin = Kappa - (KappaSD/sqrt(30)), ymax = Kappa + (KappaSD/sqrt(30)))) +

src/pre_process_wisc.r

@@ -11,82 +11,80 @@ Options:
 --out_dir=<out_dir>   Path to directory where the processed data should be written
 " -> doc
 
-library(feather)
+
 library(tidyverse)
-library(caret)
+library(tidymodels)
 library(docopt)
+library(arrow) 
 set.seed(2020)
 
 opt <- docopt(doc)
 main <- function(input, out_dir){
   # read data and convert class to factor
-  raw_data <- read_feather(input) 
+  raw_data <- arrow::read_feather(input) 
   colnames(raw_data) <- c("id",
-                         "class",
-                         "mean_radius",
-                         "mean_texture",
-                         "mean_perimeter", 
-                         "mean_area",
-                         "mean_smoothness",
-                         "mean_compactness",
-                         "mean_concavity",
-                         "mean_concave_points",
-                         "mean_symmetry",
-                         "mean_fractal_dimension",
-                         "se_radius",
-                         "se_texture",
-                         "se_perimeter", 
-                         "se_area",
-                         "se_smoothness",
-                         "se_compactness",
-                         "se_concavity",
-                         "se_concave_points",
-                         "se_symmetry",
-                         "se_fractal_dimension",
-                         "max_radius",
-                         "max_texture",
-                         "max_perimeter", 
-                         "max_area",
-                         "max_smoothness",
-                         "max_compactness",
-                         "max_concavity",
-                         "max_concave_points",
-                         "max_symmetry",
-                         "max_fractal_dimension")
-  raw_data <- raw_data %>% 
-    select(-id) %>% 
+                          "class",
+                          "mean_radius",
+                          "mean_texture",
+                          "mean_perimeter", 
+                          "mean_area",
+                          "mean_smoothness",
+                          "mean_compactness",
+                          "mean_concavity",
+                          "mean_concave_points",
+                          "mean_symmetry",
+                          "mean_fractal_dimension",
+                          "se_radius",
+                          "se_texture",
+                          "se_perimeter", 
+                          "se_area",
+                          "se_smoothness",
+                          "se_compactness",
+                          "se_concavity",
+                          "se_concave_points",
+                          "se_symmetry",
+                          "se_fractal_dimension",
+                          "max_radius",
+                          "max_texture",
+                          "max_perimeter", 
+                          "max_area",
+                          "max_smoothness",
+                          "max_compactness",
+                          "max_concavity",
+                          "max_concave_points",
+                          "max_symmetry",
+                          "max_fractal_dimension")
+  raw_data <- raw_data |> 
+    select(-id) |>    
     mutate(class = as.factor(class))
 
-  # split into training and test data sets
-  training_rows <- raw_data %>% 
-    select(class) %>% 
-    pull() %>%
-    createDataPartition(p = 0.75, list = FALSE)
-  training_data <- raw_data %>% slice(training_rows)
-  test_data <- raw_data %>% slice(-training_rows)
+  # split into training and test datasets with rsample package 
+  split <- rsample::initial_split(raw_data, prop = 0.75)
+  training_data <- rsample::training(split)
+  testing_data <- rsample::testing(split)
+
+  # calculate the required statistics into the recipe 
+  class_rec <- recipes::recipe(class ~ ., data = training_data) |>
+    recipes::step_scale(all_predictors()) |>
+    recipes::step_center(all_predictors()) |>
+    recipes::prep()
+
+  # "baking" the recipe: scaling and centering the data
+  training_scaled <- class_rec |> 
+    recipes::bake(new_data = NULL)
 
-  # scale test data using scale factor
-  x_train <- training_data %>% 
-    select(-class) 
-  x_test <- test_data %>% 
-    select(-class)
-  pre_process_scaler <- preProcess(x_train, method = c("center", "scale"))
-  x_train_scaled <- predict(pre_process_scaler, x_train)
-  x_test_scaled <- predict(pre_process_scaler, x_test)
-  training_scaled <- x_train_scaled %>% 
-    mutate(class = training_data %>% select(class) %>% pull())
-  test_scaled <- x_test_scaled %>% 
-    mutate(class = test_data %>% select(class) %>% pull())
+  test_scaled <- class_rec |> 
+    recipes::bake(new_data = testing_data)
 
-  # write scale factor to a file
+  # write the recipe to a file
   try({
     dir.create(out_dir)
   })
-  saveRDS(pre_process_scaler, file = paste0(out_dir, "/scale_factor.rds"))
+  saveRDS(class_rec, file = paste0(out_dir, "/scale_factor.rds"))
 
   # write training and test data to feather files
   write_feather(training_scaled, paste0(out_dir, "/training.feather"))
   write_feather(test_scaled, paste0(out_dir, "/test.feather"))
 }
 
-main(opt[["--input"]], opt[["--out_dir"]])
+main(opt[["--input"]], opt[["--out_dir"]])

src/reticulate_fit_breast_cancer_predict_model.R

@@ -30,16 +30,16 @@ main <- function(train, out_dir) {
   # Load and wrangle data ---------------------------------------------------
 
   # load training data as an R data frame
-  train_data <- read_feather("data/processed/training.feather") %>% 
+  train_data <- read_feather("data/processed/training.feather") |> 
     mutate(class = as.integer(as.factor(class)) )
 
   # create X and Y
-  X_train <- train_data %>% 
+  X_train <- train_data |> 
     select(-class)
-  y_train <- train_data %>% 
-    select(class) %>% 
-    pull(class) %>% 
-    as.factor() %>% 
+  y_train <- train_data  |>  
+    select(class) |> 
+    pull(class) |> 
+    as.factor() |> 
     as.integer()
   y_train <- y_train - 1
 
@@ -77,10 +77,10 @@ main <- function(train, out_dir) {
   ggsave(paste0(out_dir, "/kappa_vs_k.png"), kappa_vs_k, width = 5, height = 3)
 
   # Fit final model ---------------------------------------------------------
-  best_k <- cv_results %>% 
-    filter(cv_mean == max(cv_mean)) %>% 
-    select(k) %>% 
-    pull() %>%
+  best_k <- cv_results |> 
+    filter(cv_mean == max(cv_mean)) |> 
+    select(k) |> 
+    pull() |>
     as.integer()
   knn_final_model <- sklearn_neighbours$KNeighborsClassifier(n_neighbors = best_k)
   knn_final_model_fit <- knn_final_model$fit(X_train, y_train)