Implementation for Denoising Diffusion Probabilistic Models. The goal is to implement the models from scratch including the U-Net architecture and experiment with other architecture understanding what is happening in the Diffusion Models [and hopefully try to explain it to you as well !:)]
I have used 2 datasets for the experimentation :
- Mnist : Hand written digits
- Cifar-10 : Images belonging to 10 classes In order to replicate, please download them from the above links. Unzip the files and store them in the data dfirectory in the folder data/mnist/ and data/cifar-10/
There are two models :
- SimplifiedModel : This is a simple model comprising of CNN and without attention or the usual standard UNet model that is followed. It however does have a class embedding and a 2D positional embedding. It also supports generating conditional distribution given the class label. You can find the model architecture SimplifiedModel.py
- TO DO (Some complicated modelling)
These with are with conditional modelling i.e. passing the label that gets embeded similar (almost quite identical to positional embeding)
