DDPM from Scratch

This project implements a Denoising Diffusion Probabilistic Model (DDPM) inspired by Stable Diffusion, trained on the MNIST dataset. The model uses a UNet architecture for image generation and denoising.

Project Overview

The DDPM model is designed to generate new images by reversing a gradual noising process. This project includes the following components:

1. Linear Noise Scheduler: Manages the noise scheduling for the diffusion process.
2. MNIST Dataset: Custom dataset class for loading and preprocessing MNIST images.
3. UNet Model: A UNet-based architecture for image denoising and generation.
4. Training and Sampling Scripts: Scripts to train the model and generate new images.

Project Structure

• linear_noise_scheduler.py: Implements the linear noise scheduler for the diffusion process.
• mnist.py: Custom dataset class for loading MNIST images.
• mnist_downloader.py: Script to download and preprocess MNIST data.
• sample.py: Script to generate new images using the trained model.
• train.py: Script to train the DDPM model.
• unet.py: Implements the UNet architecture.
• unet_test.py: Tests for the UNet model.
• config/default.yaml: Configuration file for training and sampling.

Getting Started

Prerequisites

• Python 3.x
• PyTorch
• NumPy
• Pillow
• tqdm
• yaml

Installation

1. Clone the repository:

   git clone https://github.com/yourusername/ddpm-from-scratch.git
   cd ddpm-from-scratch

2. Install the required dependencies:

   pip install -r requirements.txt

Data Preparation

Download and preprocess the MNIST dataset:

python mnist_downloader.py

Training

Train the DDPM model using the following command:

python train.py --config config/default.yaml --val True

Sampling

Generate new images using the trained model:

python sample.py --config config/default.yaml --from_val False

Blog Post

For an in-depth explanation of the mathematics behind DDPM and this implementation, check out the blog post: DDPM from Scratch.

Acknowledgements

This project is inspired by the work on Stable Diffusion and various implementations of DDPM. Special thanks to the authors of the original papers and the open-source community for their contributions.

Feel free to explore the code and experiment with different configurations and if you want to, feel free to send a PR 😊 Happy coding! 🚀