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.
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.
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
To run training of UNet model after obtaining offline dataset, run
python
Some research code & dirty notebooks can be found in playground folder. UNet model is located in models.py and has nothing special in it.