Only simple depthwise convolutions, shorcuts and naive timestep embedding, there you have it! A fully functional denosing diffusion probabilistic model while keeps ultra light weight 4.55MB (the checkpoint has 9.1MB but with ema model double the size).
Install packages
pip install -r requirements.txtStart default setting training
python train_mnist.pyFeel free to tuning training parameters, type python train_mnist.py -h to get help message of arguments.
A neat blog explains how diffusion model works(must read!): https://lilianweng.github.io/posts/2021-07-11-diffusion-models/
The Denoising Diffusion Probabilistic Models paper: https://arxiv.org/pdf/2006.11239.pdf
A pytorch version of DDPM: https://github.com/lucidrains/denoising-diffusion-pytorch