https://github.com/albarji/toy-diffusion

A toy implementation of a diffusion model for low-dimensional data
https://github.com/albarji/toy-diffusion

Last synced: about 1 month ago
JSON representation

A toy implementation of a diffusion model for low-dimensional data

• Host: GitHub
• URL: https://github.com/albarji/toy-diffusion
• Owner: albarji
• License: mit
• Created: 2023-05-02T14:51:32.000Z (over 1 year ago)
• Default Branch: master
• Last Pushed: 2023-12-06T09:48:09.000Z (about 1 year ago)
• Last Synced: 2023-12-06T11:10:42.488Z (about 1 year ago)
• Language: Jupyter Notebook
• Homepage:
• Size: 5.49 MB
• Stars: 4
• Watchers: 2
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Toy diffusion



        

This is an implementation of a toy diffusion process, able the generate samples from a learned 2-dimensional distribution. It is an alternative implementation of the experiment presented in Sohl-Dickstein et al paper on using diffusion models and deep networks to generate new samples for a given dataset.

A simple pytorch network learns to predict the noise component in a data sample. This is then used in a DDPM sampler to generate new samples from the distribution.

Here is an example of generation of samples for a 2D swiss roll distribution:

![alt text](swissroll_generation.gif "Swiss Roll generation")

## References

* [Sohl-Dickstein et al - Deep Unsupervised Learning using Nonequilibrium Thermodynamics](https://arxiv.org/pdf/1503.03585.pdf)

* [Ho et al - Denoising Diffusion Probabilistic Models](https://arxiv.org/pdf/2006.11239.pdf)

* [Nichol and Dhariwal – Improving Denoising Diffusion Probabilistic Models](https://arxiv.org/pdf/2102.09672.pdf)