Fashion-MNIST converted to png
This repo contains the following:
-
Simple script to convert MNIST-like datasets to png images.
-
Fashion-MNIST converted to png format stored in the
data.zipfile.
I was needed Fashion-MNIST dataset for cool simplified example of DL framework usage, and I want to make it as close to usual Kaggle-like dataset workflow as possible.
Sometimes you just don't want all this parsing stuff.
I was about to fork one of repos or gists (1, 2, 3) which convert original MNIST into png images. But some of them don't have licenses, some have too restricted as for me. That's why I've written this script under MIT license.
Unzip it and you good to go:
unzip data.zip
The structure of the data directory after unzip is following:
data
├── test
│ ├── 0
│ ├── 1
│ ├── 2
│ ├── 3
│ ├── 4
│ ├── 5
│ ├── 6
│ ├── 7
│ ├── 8
│ └── 9
└── train
├── 0
├── 1
├── 2
├── 3
├── 4
├── 5
├── 6
├── 7
├── 8
└── 9
Images in those directories named independently from directory structure, so if you merge all those folders in one - none of images will be overwritten.
Names are from 0.png to 59999.png for train folder and from 0.png
to 9999.png for test.
In case if you want to convert MNIST or other MNIST-similar dataset like
Fashion-MNIST - you can use mnist_to_png.py script.
There are lots of hard coded stuff, but it's easy to tweak if you need.
But it should work as is if you put your MNIST files into original_data folder in
root of this repo:
original_data
├── train-images-idx3-ubyte.gz
├── train-labels-idx1-ubyte.gz
├── t10k-images-idx3-ubyte.gz
└── t10k-labels-idx1-ubyte.gz
mnist_to_png.pyThen just run:
python3 mnist_to_png.pyAnd it will create data directory with all the images stored according to
structure above.
Feel free to ask for help in issues.
I've tried to minimize requirements for this script:
- Python 3+
- pypng
You can install them by:
pip install -r requirements.txtFashion-MNIST: Copyright © 2017 Zalando SE, https://tech.zalando.com