Skip to content

Repository for Skoltech Deep Learning Final Project. Reproduction of Star-Shaped Denoising Diffusion Probabilistic Models pre-print https://arxiv.org/pdf/2302.05259.pdf

Notifications You must be signed in to change notification settings

SKholkin/SS_DDPM

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

20 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Star-Shaped DDPM Implementation

Repository for Skoltech Deep Learning Final Project. Reproduction of Star-Shaped Denoising Diffusion Probabilistic Models (Okhotin et. al https://arxiv.org/pdf/2302.05259.pdf) in PyTorch.

Motivation

From paper, For data distributed on manifolds, bounded volumes, or with other features, the injection of Gaussian noise can be unnatural, breaking the data structure. Following this observation, authors propose to use non-Gaussian noising scheme for constrained manifolds (e.g sphere), such as Mises-Fisher, Dirichlet, Beta etc.

Generating Dataset samples in Offline Regime

Since in SSDDPM forward process can not be implemented in closed form (as for example in casual DDPM), it is more efficient to prepare in offline. In order to obtain such dataset, run (you can change hyper parameters in file)

python generate_dataset.py

Training SSDDPM

To run training of UNet model after obtaining offline dataset, run

python 

Notebooks

Some research code & dirty notebooks can be found in playground folder. UNet model is located in models.py and has nothing special in it.

About

Repository for Skoltech Deep Learning Final Project. Reproduction of Star-Shaped Denoising Diffusion Probabilistic Models pre-print https://arxiv.org/pdf/2302.05259.pdf

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published