如何将CIFAR数据集转换为与MNIST相同的格式

Question

I'm trying to run a machine learning algorithm on two different datasets. However, the format for the y values are different between the datasets.

from keras.datasets import mnist, cifar10
(x_train, y_train), (x_test, y_test) = mnist.load_data()
print([y_train[i] for i in range(10)])
'''
[5, 0, 4, 1, 9, 2, 1, 3, 1, 4]
'''
(x_train, y_train), (x_test, y_test) = cifar10.load_data()
print([y_train[i] for i in range(10)])
'''
[array([6], dtype=uint8), array([9], dtype=uint8), array([9], dtype=uint8), array([4], dtype=uint8), array([1], dtype=uint8), array([1], dtype=uint8), array([2], dtype=uint8), array([7], dtype=uint8), array([8], dtype=uint8), array([3], dtype=uint8)]
'''

The documentation on Keras says that the format for MNIST is

y_train, y_test: uint8 array of digit labels (integers in range 0-9) with shape (num_samples,)

and that the format for CIFAR is

y_train, y_test: uint8 array of category labels (integers in range 0-9) with shape (num_samples,)

To me these seem like they should be the exact same format. So, I have two questions:

1. How would I tell, from the documentation, that they actually have different formats? (If this is impossible, then just say so.)

2. How can I convert the CIFAR dataset to be in the same format as mnist? (My algorithm currently works on MNIST.)

Answer 1

The issue is that y have slightly different shapes in both datasets, its (60000,) for MNIST, but (50000, 1) in CIFAR10. The extra dimension should not cause any problem, but in any case you can get rid of with with:

y = np.squeeze(y)

Then y's shape will be (50000,) .