如何更改 CIFAR-10 数据集标签？

Question

I want to change some labels of dataset CIFAR-10. That mean, I want to train my networks with some wrongs labels in this dataset.

After change the dataset labels, I need it in some format like:

dataset_cifar_with_some_wrongs_labels = (amount of pictures, label, image)

image.shape=(32,32,3)

I know I can download it as:

https://www.cs.toronto.edu/~kriz/cifar.html

Someone know how can i do it?

Answer 1

You can use keras internal cifar10 dataset. I have implemented a noise mechanism via permutation but there are probably many other choices.

import keras
import numpy as np
(x_train, y_train), (x_test, y_test) = keras.datasets.cifar10.load_data()
n = 1000
y_train_ = y_train.copy()
idx_to_shuffle = np.random.choice(y_train.shape[0], n)
y_train[idx_to_shuffle] = np.random.permutation(y_train[idx_to_shuffle])