Starting Simple - MNIST Generator

First, generative AI I’m working on is MNIST generator. MNIST is a dataset of hand written digits.

Two things that were attractive about this dataset was that all the images had 28×28 pixels and they were grayscale.

Took around 3-4 days to code a python script that would train VQVAE model and Auto-Regressive model on top of it. Used 4×4 pixel patches and tokenized into codes using VQVAE (had to use very small commitment loss term).

And voila, it was able to create random numbers at the end without any issues.

Next, I will be moving to Moving MNIST dataset and see if I can create a model that can create digits bouncing around.