Denoising Diffusion Probabilistic Model from Scratch

This repository implements a fast, yet simple version of the Denoising Diffusion Probabilistic Models paper for the image generation task. Denoising score matching technique is used to allow the network for rapid estimation of data distribution’s gradient. Moreover, the Langevin sampling method is also performed to generate images from their corresponding true data distribution. This implementation provides unconditional along with conditional (Classifier-free Diffusion Guidance (CFG) and Exponential Moving Average (EMA)) sampling approaches.

Requirements

• PyTorch
• torchvision
• numpy
• PIL
• matplotlib
• logging
• tqdm

Usage

Data

The unconditional model is trained on Landscape Pictures dataset and the conditional model is trained on CIFAR-10 dataset.

Training and Testing

• To see the implementation of conditional and unconditional sampling methods, check DDPM.py.
• All the network architectures for EMA, UNet, etc. can be found in the models.py file.
• To train DDPMs for either of the sampling approaches (conditional and unconditional) and generate the images, run DDPM.py.
• All hyperparameters to control the training and testing phases of the model are provided in the given DDPM.py file.

Results

The images generated by both conditional and unconditional models can be seen below:-

Training Dataset	Sampling Type	Generated Images
Landscape Pictures	unconditional
CIFAR-10	conditional (on Dog and Deer classes)