Feature selection on MNIST data

set.seed(123)

library(celavi)
library(dplyr)
library(klassets)

use_small_data <- TRUE

data <- klassets::mnist_train
data <- dplyr::mutate(data, label = factor(label))

test <- klassets::mnist_test
test <- dplyr::mutate(test, label = factor(label))

iter  <- 20
frac <- .25

if(use_small_data){
  
  data <- dplyr::sample_n(data, 10000)
  data <- dplyr::select(data, c(1, sample(2:785, 500)))  
  test <- dplyr::sample_n(test, 1000)
  
  iter <- 20
  frac <- 1/10
  
} 

x <- feature_selection(
  ranger::ranger,
  data = data,
  test = test,
  response = "label",
  # stat = function(x) quantile(x, .25),
  stat = function(x) quantile(x, .75),
  iterations = iter,
  sample_frac = frac,
  predict_function = function(object, newdata){ranger:::predict.ranger(object, data = newdata)$predictions},
  parallel = FALSE,
  max.depth = 15
)
#> ℹ Using 1 - accuracy as loss function.
#> ℹ Fitting 1st model using 500 predictor variables.
#> 
#> ── Round #1 ──
#> 
#> ℹ Using `dplyr::sample_frac` as sampler.
#> ℹ Removing 4 variables. Fitting new model with 496 variables.
#> 
#> ── Round #2 ──
#> 
#> ℹ Using `dplyr::sample_frac` as sampler.
#> ℹ Removing 9 variables. Fitting new model with 487 variables.
#> 
#> ── Round #3 ──
#> 
#> ℹ Using `dplyr::sample_frac` as sampler.
#> ℹ Removing 108 variables. Fitting new model with 379 variables.
#> 
#> ── Round #4 ──
#> 
#> ℹ Using `dplyr::sample_frac` as sampler.
#> ℹ Removing 5 variables. Fitting new model with 374 variables.
#> 
#> ── Round #5 ──
#> 
#> ℹ Using `dplyr::sample_frac` as sampler.
#> ℹ Removing 18 variables. Fitting new model with 356 variables.
#> 
#> ── Round #6 ──
#> 
#> ℹ Using `dplyr::sample_frac` as sampler.

x
#> # A tibble: 6 × 5
#>   round mean_value values     n_variables variables  
#>   <dbl>      <dbl> <list>           <int> <list>     
#> 1     1     0.0425 <dbl [20]>         500 <chr [500]>
#> 2     2     0.0455 <dbl [20]>         496 <chr [496]>
#> 3     3     0.053  <dbl [20]>         487 <chr [487]>
#> 4     4     0.046  <dbl [20]>         379 <chr [379]>
#> 5     5     0.0505 <dbl [20]>         374 <chr [374]>
#> 6     6     0.0485 <dbl [20]>         356 <chr [356]>

plot(x)