This is keep-it-simple-and-stupid realization of Score-Based Generative Modeling through Stochastic Differential Equations. The whole model is written in pure PyTorch and made as self-explanatory as possible.

This realization contains a basic convolutional U-Net-like score approximation model and predictor-corrector as a sampler. The whole code is based on a different parts of the mentioned repo. Some major improvements like EMA of weights are implemented, leading to reproducing nearly SotA results on CIFAR10, while using tutorial-like architecture.

Firstly, build docker file with

docker build -t score_sde .

Then you can specify your wandb key in run.yaml and run training with crafting

crafting run.yaml