Transform your images into an MNIST NN Ready Binary:

1. Copy-pasta your jpg and/or png images into one of the class folders, as seen in (e.g. dogs -> 0, cats -> 1, ... giraffes->9)

2. Change the appropriate labels in batches.meta.txt

3. then use the following bash script which processes the images, rescaling all of the png's you placed in the folders the MNIST standard 28x28pixel size

./resize-script.sh

4. lastly, run the following python script to fold all the pics and categories into a single ble binary -- binary will appear as ubyte files ready to tar

python convert-images-to-mnist-format.py

Victory!

You now have the files, and can replace the the conventional data in any standard mnist tutorial with your own data :D

Enjoy!

Tree

example of where files will appear/where-to put png's (ignore placeholder.txt, this is only there so that git could check in their parent folders -- since empty folders can't be committed):

.
├── batches.meta.txt
├── convert-images-to-mnist-format.py
├── LICENSE
├── README.md
├── resize-script.sh
├── test-images
│   ├── 0
│   │   ├── home-cat.png
│   │   └── placeholder.txt
│   ├── 1
│   │   ├── dog-07.png
│   │   └── placeholder.txt
│   ├── 2
│   │   └── placeholder.txt
│   ├── 3
│   │   └── placeholder.txt
│   ├── 4
│   │   └── placeholder.txt
│   ├── 5
│   │   └── placeholder.txt
│   ├── 6
│   │   └── placeholder.txt
│   ├── 7
│   │   └── placeholder.txt
│   ├── 8
│   │   └── placeholder.txt
│   └── 9
│       └── placeholder.txt
└── training-images
    ├── 0
    │   ├── home-cat.png
    │   └── placeholder.txt
    ├── 1
    │   ├── dog-07.png
    │   └── placeholder.txt
    ├── 2
    │   └── placeholder.txt
    ├── 3
    │   └── placeholder.txt
    ├── 4
    │   └── placeholder.txt
    ├── 5
    │   └── placeholder.txt
    ├── 6
    │   └── placeholder.txt
    ├── 7
    │   └── placeholder.txt
    ├── 8
    │   └── placeholder.txt
    └── 9
        └── placeholder.txt