[go: up one dir, main page]

Skip to content

Kei18/tiny-tiny-diffusion

Repository files navigation

tiny-tiny-diffusion

PyTorch

Tiny-tiny PyTorch implementation of diffusion probabilistic models [1] for self-study, based on the amazing tiny-diffusion repository. All this repo, tiny-tiny-diffusion, can do is reconstruct a 2D dinosaur 🦖 from random noise (but it's cool, right?)

Thanks to the simplicity and limited capabilities, the training can be done in a few minutes with my laptop without GPUs. The inference speed is negligible 🚅.

setup

You need Poetry. After cloning this repo, run the following to complete the setup:

poetry install

demo

The demo is available in examples.ipynb. You can start Jupyter Lab with the following command:

poetry run jupyter lab

learning new data

Actually, this repo can also draw a cat 🐈 by learning new data:

poetry run python ddpm/train.py dataset.csv_file=assets/shape/cat.csv

The trained model will be saved in learning_results.

The training script uses Hydra to specify hyperparameters. A base configuration can be found in ddpm/train_conf/example.yaml.

background

  • Okay, so you are interested in diffusion models. Yay! But I am not an ML guy. Indeed, the original paper [1] was quite difficult to understand without preliminary knowledge.
  • After skimming some blog posts, I came across [2]; I recommend this paper as a starting point.
  • Okay, I was able to grasp the basics. But it is always good to code it for a better understanding. So I searched through some code repositories and found tiny-diffusion. Quite impressive! I decided to mimic it for leaning.
  • Although the code in this repository has been rewritten from scratch for self-study, I have heavily referenced tiny-diffusion. Therefore, I don't strongly claim authorship. Please respect the original repository.

notes

  • Please remember that I am NOT an ML person. Maybe I made some mistakes. I am still learning!
  • The shape data in assets/shape was retrieved from Data Morph.
  • As a minimalist, the repo omits positional encodings; it works
  • Setup auto formatting when committing with pre-commit:
poetry run pre-commit install

references

  1. Ho, J., Jain, A., & Abbeel, P. Denoising diffusion probabilistic models. NeurIPS. 2020.
  2. Luo, C. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970. 2022.