如何随机获取 numpy 数组的一定数量的元素，每个 class 至少有一个元素？

Question

I have a dataset of 400 images, 10 images of 40 different people. There are 2 NumPy arrays, "olivetti_faces" contains the images (400x64x64), and "olivetti_faces_target" contains the classes of those images (400), one class for each person. So "olivetti_faces" is of the form: array([<img1>, <img2>, ..., <img400>]) where <img> is a 64x64 array of numbers, and "olivetti_faces_target" is of the form: array([0, 0, ..., 39]) .

You can access the dataset here . You can load them after downloading as follows:

import numpy as np
data=np.load("olivetti_faces.npy")
target=np.load("olivetti_faces_target.npy")

I would like to randomly choose 100 of the images, with at least one image of each of the 40 people. How can I achieve this in NumPy?

So far I could randomly get 100 images using the following code:

n = 100 # number of images to retrieve
rand_indeces = np.random.choice(data.shape[0], n, replace=False)
data_random = data[rand_indeces]
target_random = target_random[rand_indeces]

But it does not guarantee that at least one image of each of the 40 classes is included in data_random .

Answer 1

As suggested in my comment, first pick a random index from each class. Then choose random indexes from the entire array. This will guarantee that each class has an entry in the final result.

Since each class has 10 elements, you can loop through classes 0-39 and pick a value 0-9.

Try this code:

import numpy as np
import random

data=np.load("olivetti_faces.npy")
target=np.load("olivetti_faces_target.npy")

# target is groups of 10, so select random index in each block
for i in range(40):  # class 0-39
   rndindex.append(i*10 + random.randint(0,9)) # one per class
   
for i in range(60):  # up to 100
   idx = rndindex[0]
   while idx in rndindex:  # prevent duplicates
       idx = random.randint(0,399)  # other indexes can be anywhere
   rndindex.append(idx)

rand_indeces = []  # np array objects
for idx in rndindex:
   rand_indeces.append(data[idx])

print(rndindex)
#print(rand_indeces)

Output (note that the first 40 fall within blocks of 10)

[9, 17, 23, 31, 41, 52, 60, 72, 83, 95, 
 100, 119, 121, 136, 140, 150, 166, 175, 188, 198, 
 209, 211, 221, 238, 243, 250, 261, 276, 289, 290, 
 306, 315, 325, 333, 344, 351, 368, 376, 382, 391, 
 62, 296, 327, 241, 393, 215, 64, 59, 185, 286, 
 162, 163, 364, 309, 220, 273, 32, 214, 217, 182, 
 172, 98, 19, 358, 92, 322, 68, 399, 226, 285, 
 103, 155, 249, 1, 75, 303, 311, 125, 339, 106, 
 127, 94, 101, 113, 35, 20, 189, 199, 128, 30, 
 131, 317, 337, 156, 340, 99, 397, 385, 384, 193]

Answer 2

One way to do this, as @xrisk suggested, would be randomly getting 100 images until the condition is satisfied, as follows:

n = 100 # number of images to retrieve
n_unique = len(np.unique(target)) # number of classes
containsAllCats = False
while not containsAllCats:
    rand_indeces = np.random.choice(data.shape[0], n, replace=False)
    data_random = data[rand_indeces]
    target_random = target[rand_indeces]
    containsAllCats = len(np.unique(target_random)) == n_unique

However, it seems to be rather inefficient, most of the time requiring several iterations.