numpyvision: Vision datasets as numpy arrays

numpyvision is a drop-in replacement for torchvision.datasets with an easy access to MNIST and other MNIST-like datasets (FashionMNIST, KMNIST, EMNIST) in your numpy code.

numpyvision replicates the functionality of torchvision.datasets.mnist without the need to download dozens of dependencies - numpyvision has only one dependency: numpy.

Usage

Each dataset stores train/test images as numpy arrays of shape (n_samples, img_height, img_width) and train/test labels as numpy arrays of shape (n_samples,).

MNIST example:

>>> from numpyvision.datasets import MNIST
>>> mnist = MNIST(train=True)
>>> type(mnist.data)
<class 'numpy.ndarray'>
>>> mnist.data.dtype
dtype('uint8')
>>> mnist.data.min()
0
>>> mnist.data.max()
255
>>> mnist.data.shape
(60000, 28, 28)
>>> mnist.targets.shape
(60000,)
>>> mnist.classes[:3]
['0 - zero', '1 - one', '2 - two']

FashionMNIST example:

from numpyvision.datasets import FashionMNIST
import matplotlib.pyplot as plt

fmnist = FashionMNIST()
img, label = fmnist[0]
plt.imshow(img, cmap='gray')
plt.title(fmnist.classes[label])
plt.axis('off')
plt.show()

EMNIST example

from numpyvision.datasets import EMNIST
import matplotlib.pyplot as plt

letters = EMNIST('letters')
plt.imshow(
    letters.data[:256]
        .reshape(16, 16, 28, 28)
        .swapaxes(1, 2)
        .reshape(16 * 28, -1),
    cmap='gray')
plt.axis('off')
plt.show()

Installation

Install numpyvision from PyPi:

pip install numpyvision

or from source:

pip install -U git+https://github.com/pczarnik/numpyvision

The only requirements for numpyvision are numpy>=1.22 and python>=3.9.

If you want to have progress bars while downloading datasets, install with

pip install numpyvision[tqdm]

Acknowledgments

The main inspirations for numpyvision were mnist and torchvision.datasets.mnist.