Flatland

Generative models (GAN, VAE, DDPM, and Transformers) trained in one and two-dimensional data. The point of this is simple: lowering the dimensionality of the training data will aid newcomers in the field to better understand the mechanics of what these networks do: in particular, what they learn. Whenever some work is inspired by previous work, it will be duly credited.

For now, this project will be developed using PyTorch, but if possible/I have both time and patience, I will expand it to other frameworks such as Keras/TensorFlow 2.0, Julia, etc. (this is my way of saying I will accept pull requests for anyone interested in contributing).

TODO:

• Complete 1D GAN code (the following is not necessarily in order and some may even overlap)
  • Make train.py and generate.py code
  • Add manual seed to latents to easily sample from and create interpolations and whatnot (in both scripts)
  • Add command-line arguments (make most things controllable, but with default values); use click for this
  • Add samples of training results to README (distribution plots, videos)
  • Make the neural networks easily editable using config.yml
• Ibidem for VAE, but not as pressing
• Ibidem for DDPM
• Ibidem for Transformer
• Add/update requirements.txt