The Pytorch implementation is adapted from openai/guided-diffusion with modifications for classifier-free conditioned generation. The dataset used for training is cifar-10. The platform I use is Ubuntu 20.04.

During the training porcess, I use the file Scheduler.py from zoubohao/DenoisingDiffusionProbabilityModel-ddpm-, which is different from my implementation mainly on one techical detail : How to represent the embeddings of null token of class identifier. According to another academic paper Video Diffusion Models by the same authors, I use the same method as theirs in this detail.

First, you need to do some preparations:

mkdir sample
mkdir model
ln -s absolute/path/to/cifar-10 ./cifar-10-batches-py

make train

NOTICE : hyperparameter settings are in the file train.py

make sample

NOTICE : hyperparameter settings are in the file sample.py