page-david / python-mnist

Simple MNIST data parser written in Python

python-mnist

Simple MNIST and EMNIST data parser written in pure Python.

MNIST is a database of handwritten digits available on http://yann.lecun.com/exdb/mnist/. EMNIST is an extended MNIST database https://www.nist.gov/itl/iad/image-group/emnist-dataset.

Requirements

• Python 2 or Python 3

Usage

• git clone https://github.com/sorki/python-mnist
• cd python-mnist

• Get MNIST data:

  ./get_data.sh

• Check preview with:

  PYTHONPATH=. ./bin/mnist_preview

Installation

Get the package from PyPi:

pip install python-mnist

or install with setup.py:

python setup.py install

Code sample:

from mnist import MNIST
mndata = MNIST('./dir_with_mnist_data_files')
images, labels = mndata.load_training()

To enable loading of gzip-ed files use:

mndata.gz = True

Library tries to load files named t10k-images-idx3-ubyte train-labels-idx1-ubyte train-images-idx3-ubyte and t10k-labels-idx1-ubyte. If loading throws an exception check if these names match.

EMNIST

• Get EMNIST data:

  ./get_emnist_data.sh

• Check preview with:

  PYTHONPATH=. ./bin/emnist_preview

To use EMNIST datasets you need to call:

mndata.select_emnist('digits')

Where digits is one of the available EMNIST datasets. You can choose from

• balanced
• byclass
• bymerge
• digits
• letters
• mnist

EMNIST loader uses gziped files by default, this can be disabled by by setting:

mndata.gz = False

You also need to unpack EMNIST files as get_emnist_data.sh script won't do it for you. EMNIST loader also needs to mirror and rotate images so it is a bit slower (If this is an issue for you, you should repack the data to avoid mirroring and rotation on each load).

Notes

This package doesn't use numpy by design as when I've tried to find a working implementation all of them were based on some archaic version of numpy and none of them worked. This loads data files with struct.unpack instead.

Example

$ PYTHONPATH=. ./bin/mnist_preview
Showing num: 3

............................
............................
............................
............................
............................
............................
.............@@@@@..........
..........@@@@@@@@@@........
.......@@@@@@......@@.......
.......@@@........@@@.......
.................@@.........
................@@@.........
...............@@@@@........
.............@@@............
.............@.......@......
.....................@......
.....................@@.....
....................@@......
...................@@@......
.................@@@@.......
................@@@@........
....@........@@@@@..........
....@@@@@@@@@@@@............
......@@@@@@................
............................
............................
............................
............................