Simple Diffusion

This repo contains instructions for training the simple diffusion model described in this blog post.

Setup

pip install virtualenv
virtualenv env -p python3
source env/bin/activate
git clone https://github.com/fastai/course22p2.git
cd course22p2

update settings.ini

Change pip_requirements on line 38 in settings.ini to pip_requirements = torch>=1.7 torcheval diffusers einops timm accelerate ipykernel

install packages

pip install -e .
python -m ipykernel install --user --name=env --display-name "simple_diffusion"

Training

Open the simple_diffusion.ipynb notebook, select the simple_diffusion environment and run all cells.

Training simple diffusion on Fashion MNIST for 5 epochs takes ~3 minutes on an A100 with 40GB of VRAM.

Here are some example images generated by simple diffusion after 5 epochs of training.