Automated Ranger training for a grid of specifications and datasets. …

…Separated export and validation of predictions from training. Switched .RDS files handling to readr package and bz2 algorithm. Partly finished Ranger predictions validation code.

dataset_prep.R

@@ -30,13 +30,13 @@ data2 <- data2 %>%
 
 data2 %>% filter_vars_by_iv(significance_thres = 0.01) %>%
   initial_split(prop = 0.75) %>%
-  saveRDS("data/split_raw.RDS")
+  write_rds("data/split_raw.RDS", compress = "gz2")
 
 data3 <- data2 %>% 
   factorize(bin_methods = "tree") %>% 
   as_tibble() %>% 
   filter_vars_by_iv(significance_thres = 0.01)
 
-dataset_split <- data3 %>% initial_split(prop = 0.75) %>% saveRDS("data/split.RDS")
+dataset_split <- data3 %>% initial_split(prop = 0.75) %>% write_rds("data/split.RDS", compress = "gz2")
 
 rm(list = ls())

funs_valid.R

@@ -0,0 +1,16 @@
+#### validation and visualisation automating functions
+
+get_all_metrics <- function(pred_dfs, spec = 1L) {
+  f <- function(df) metrics(df, Exited, .pred_class, .pred_No)
+  model1_metrics <- f(pred_dfs[[1]][[spec]])
+
+  all_metrics <- do.call(
+    function(...) bind_cols(model1_metrics, ...), 
+    pred_dfs %>% select(-1) %>% map(~ f(.x[[spec]])$.estimate)) %>%
+    rename(model_1 = .estimate) %>% 
+    return
+}
+
+exportable_conf_matrix <- function(df) {
+  conf_mat(df, Exited, .pred_class)$table
+}

rand_forest.R

@@ -10,52 +10,35 @@ rm(list = ls())
 
 # source("dataset_prep.R")
 
-dataset_split1 <- readRDS("data/split.RDS")
-dataset_split2 <- readRDS("data/split_raw.RDS")
-
-df_train1 <- dataset_split1 %>% training()
-df_test1 <- dataset_split1 %>% testing()
-df_train2 <- dataset_split2 %>% training()
-df_test2 <- dataset_split2 %>% testing()
-
-ranger_model_specs <- rand_forest("classification", 2, 1000, 5) %>% 
-  # set_engine("ranger", num.threads = 8, replace = F, sample.fraction = 0.8, importance = "impurity") %>%
-  set_engine("ranger", num.threads = 8, replace = F, sample.fraction = 0.8, importance = "permutation", local.importance = T)
-
-ranger_model_1 <- ranger_model_specs %>% fit(Exited ~ ., data = df_train1)
-
-ranger_model_2 <- ranger_model_specs %>% fit(Exited ~ ., data = df_train2)
+dataset_splits <- list(
+  read_rds("data/split.RDS"),
+  read_rds("data/split_raw.RDS")
+  )
+
+testing_sets <- dataset_splits %>% map(~ .x %>% testing())
+
+models_specs <- list(
+  rand_forest("classification", 2, 1000, 5) %>% 
+    # set_engine("ranger", num.threads = 8, replace = F, sample.fraction = 0.8, importance = "impurity") %>%
+    set_engine("ranger", num.threads = 8, replace = F, sample.fraction = 0.8, importance = "permutation", local.importance = T)
+  )
+
+spec_names <- str_c("model_", 1:length(dataset_splits))
+fitted_models <- dataset_splits %>% 
+  map(~ .x %>% training()) %>% 
+  map2_dfc(spec_names, function(df, col_name) {
+    tibble(!!col_name := models_specs %>% map(~ .x %>%  fit(Exited ~ ., data = df)))
+  })
 
-df_pred1 <- ranger_model_1 %>% 
-  predict(df_test1) %>% 
-  bind_cols(df_test1)
-
-df_pred2 <- ranger_model_2 %>% 
-  predict(df_test2) %>% 
-  bind_cols(df_test2)
-
-df_pred1 %>% metrics(Exited, .pred_class)
-df_pred2 %>% metrics(Exited, .pred_class)
-
-df_pred_probs1 <- ranger_model_1 %>% 
-  predict(df_test1, type = "prob") %>% 
-  bind_cols(df_test1)
-
-df_pred_probs2 <- ranger_model_2 %>% 
-  predict(df_test2, type = "prob") %>% 
-  bind_cols(df_test2)
-
-df_pred_probs1 %>% roc_auc(Exited, .pred_No)
-df_pred_probs2 %>% roc_auc(Exited, .pred_No)
-
-df_pred_probs1 %>% roc_curve(Exited, .pred_No) %>% autoplot()
-df_pred_probs2 %>% roc_curve(Exited, .pred_No) %>% autoplot()
-
-vi(ranger_model_1)
-vi(ranger_model_2)
-
-vip(ranger_model_1)
-vip(ranger_model_2)
-
-
-
+pred_dfs <- list(fitted_models, testing_sets, spec_names) %>% pmap_dfc(function(models_by_spec, df, spec_name) {
+  tibble(!!spec_name := 
+           models_by_spec %>% map(function(model) {
+             df %>% bind_cols(
+               model %>% predict(df),
+               model %>% predict(df, type = "prob")
+             )
+           }))
+})
+
+fitted_models %>% write_rds("data/fitted_models.RDS", compress = "bz2")
+pred_dfs %>% write_rds("data/predictions.RDS", compress = "bz2")

validation_rf.R

@@ -0,0 +1,29 @@
+#### walidation
+
+library(tidyverse)
+library(stringr)
+library(tidymodels)
+library(ranger)
+library(vip)
+library(knitr)
+library(kableExtra)
+
+rm(list = ls())
+
+fitted_models <- read_rds("data/fitted_models.RDS")
+pred_dfs <- read_rds("data/predictions.RDS")
+
+source("funs_valid.R")
+
+all_metrics <- get_all_metrics(pred_dfs)
+
+pred_dfs[[1]][[1]] %>% exportable_conf_matrix %>% kable(format = "markdown")
+
+# df_pred_probs1 %>% roc_curve(Exited, .pred_No) %>% autoplot()
+# df_pred_probs2 %>% roc_curve(Exited, .pred_No) %>% autoplot()
+# 
+# vi(ranger_model_1)
+# vi(ranger_model_2)
+# 
+# vip(ranger_model_1)
+# vip(ranger_model_2)