Take a sample of the MNIST dataset

I am working with the MNIST dataset and performing different classification methods on it, but my runtimes are ridiculous, so I am looking for a way to maybe use an a portion of the training part of the set, but keep the test portion at 10K. I have tried a number of different options but nothing is working.

I need to take a sample either from the entire set, or lower the training x and y from 60000 to maybe 20000.

My current code:


library(keras)

mnist <- dataset_mnist()

train_images <- mnist$train$x 
train_labels <- mnist$train$y 
test_images <- mnist$test$x   
test_labels <- mnist$test$y

I have tried to use the sample() function and other types of splits to no avail.

Solution

In the following example I'm downloading MNIST myself and loading it through reticulate / numpy. Shouldn't make much difference. When you want to get a sample with sample(), you usually take a sample of indices you'll use for subsetting. To get a balanced sample, you might want to draw a specific number or proportion from each label group:

library(reticulate)
library(dplyr)

# Download MNIST dataset as numpy npz, 
# load through reticulate, build something along the lines of keras::dataset_mnist() output
np <- import("numpy")
mnist_npz <- curl::curl_download("https://storage.googleapis.com/tensorflow/tf-keras-datasets/mnist.npz", "mnist.npz")
mnist_np <- np$load(mnist_npz)

mnist_lst <- list(
  train = list(
    x = mnist_np[["x_train"]],
    y = mnist_np[["y_train"]]
  ),
  test = list(
    x = mnist_np[["x_test"]],
    y = mnist_np[["y_test"]]
  )
)

train_images <- mnist_lst$train$x 
train_labels <- mnist_lst$train$y 
test_images  <- mnist_lst$test$x   
test_labels  <- mnist_lst$test$y 

# sample row indices, 
# 100 per class to keep the dataset balanced
sample_idx <- 
  train_labels |>
  tibble(y = _) |>
  tibble::rowid_to_column("idx") |>
  slice_sample(n = 100, by = y ) |>
  arrange(idx) |>
  pull(idx)

# use sample_idx for subsetting
train_images_sample <- train_images[sample_idx,,] 
train_labels_sample <- train_labels[sample_idx]

str(train_images_sample)
#>  int [1:1000, 1:28, 1:28] 0 0 0 0 0 0 0 0 0 0 ...
str(train_labels_sample)
#>  int [1:1000(1d)] 9 7 5 6 8 7 7 5 2 9 ...

# original label distribution
table(train_labels)
#> train_labels
#>    0    1    2    3    4    5    6    7    8    9 
#> 5923 6742 5958 6131 5842 5421 5918 6265 5851 5949

# sample distribution
table(train_labels_sample)
#> train_labels_sample
#>   0   1   2   3   4   5   6   7   8   9 
#> 100 100 100 100 100 100 100 100 100 100

^{Created on 2024-03-29 with reprex v2.1.0}