I didn't properly manage all the python dependencies, but for sure you will need:

computer
python
numpy
matplotlib
tensorflow      (with GPU support if you don't want to witness heat death of universe)
pandas          (for visualizations, python's built in csv reader sucks)
Pillow          (python package, not irl, but it also will come in handy whilst waiting)
pygame          (if you want to see how it does with your digits)
python-mnist
mnist image binaries

Place *-ubyte decompressed files into mnist folder in the root of repo.

You can get these files here: click. To unzip use gunzip <name> or GUI like 7zip or WinRAR.

I don't want to spam github with over 2GBs of raw model weights, if you really want them, open a issue and I'll contact you. If you want to train those models yourself, just type python train_<name>.py into terminal and sit back or use the highly sophisticated script train.sh and go to sleep.

With my GTX960 I got about 30 secs per epoch, with 2070S, about half of that. 30s * 150 epochs * 5 models ~= 6hrs

I've made a small little teensy-weensy sketch, that you can use to see how's your model doing trying to guess your hand-writing.

Start it with: python interactive.py <path_to_model_file>, eg. python interactive.py output/own/model/150. For now (and probably ever) it only supports the "own" model due to special preprocessing of mobile and efficient networks inputs.

Use keys ~ through 0 (above keys, not on the keypad) to pick a response layer (~ is used for the background, and rest of the keys correspond to the class), and keys q through p to set the brush size. Backspace clears the draw area, Return predicts, x inserts single generated image the same as for training into drawing area and Escape quits.

Don't be the cause of AI uprising and stay safe.