From Denoising Diffusions to Denoising Markov Models

Running experiments

For the g-and-k distribution example, run python main_score_sde.py experiment=conditional dataset.num_quantiles=250 dataset.num_samples=250

For the MNIST inpainting example, go into the discrete_ctmc directory and run python dist_train.py conditional_mnist

For the ImageNet super-resolution example, go into the discrete_ctmc directory and run python train.py conditional_imagenet

For the SO3 example, run python main.py experiment=so3 dataset.K=16

For the pose estimation example, run python main.py experiment=symsol

References

This codebase is largely based on the existing works of [1] and [2], both also developed at Oxford in the Department of Statistics. It also uses a modified version of geomstats and haikumodels.

[1] Valentin De Bortoli, Emile Mathieu, Michael Hutchinson, James Thornton, Yee Whye Teh, Arnaud Doucet, Riemannian Score-Based Generative Modeling, NeurIPS 2022.

[2] Andrew Campbell, Joe Benton, Valentin De Bortoli, Tom Rainforth, George Deligiannidis, Arnaud Doucet, A Continuous Time Framework for Discrete Denoising Models