%20p1.jpeg){kind=link}
%20p2.jpeg){kind=link}
This is a lightly edited version of the following implementation of a Denoising Diffusion Probabilistic Model (DDPM):
https://www.youtube.com/watch?v=S_il77Ttrmg&list=WL&index=4
See math.ipynb.
A UNet is used to model the mapping from noisy image to denoised image (see model.py).
I overfitted the diffusion model to a single training image (racoon.jpg), and checked that the model generates exactly this image,
when presented with random noise (racoon.ipynb).
I then trained the model on CIFAR10 using a single-GPU g4dn.xlarge EC2 instance for 160 epochs over a period of about 6 hours.
Some statistics from the training run can be seen at: https://api.wandb.ai/links/peter-thomas-mchale/76bqm8a8
The model is about 240Mb in size.
Finally, I used the trained model to generate new images,
some of which look realistic (generate-from-cifar10.ipynb).
Most of the images generated from the CIFAR10 model are not realistic. There are two opposite explanations:
- The labels were not one-hot encoded (and masked out), as they were in the conditional diffusion model described here. It's possible that, without masking the label, the model can over-rely upon the labels and not learn how to generate images at all.
- The conditional diffusion model we trained above is actually ignoring the labels, which could be corrected using classifier or classifier-free guidance, as described in Luo 2022.
See also https://github.com/petermchale/minDiffusion