Denoising Diffusion Probabilistic Model

This project implements the models proposed in Denoising-Diffusion-Probabilistic-Model (DDPM) paper written by Jonathan Ho, Ajay Jain, and Pieter Abbeel.

The models were ran over the following dataset:

• MNIST

The results of the models can be found in results.pdf.

Below are generated image samples from the DDPM and DDIM methods:

• DDIM (5 steps): Fewer denoising steps, resulting in quicker generation but slightly noisier images.

  

• DDIM (10 steps): Improved clarity over 5 steps while still maintaining fast sampling.

  

• DDIM (20 steps): Better image quality, showing clearer digits with fewer artifacts.

  

• DDIM (50 steps): Nearly matches the DDPM quality, producing very clean and sharp digits.

  

• DDPM Sample after 29 Epochs: A sample generated by the trained DDPM model after 29 epochs of training.

  

Setup

Prerequisites

• Python 3.x
• NumPy
• torch
• torchvision
• matplotlib

You can install the dependencies using:

pip install -r requirements.txt

Running the code

To train the DDPM model on the MNIST dataset:

python main.py

You can customize training using the following command-line arguments:

Argument	Description	Default
--batch_size	Number of images per mini-batch	128
--epochs	Number of training epochs	30
--lr	Initial learning rate for the optimizer	1e-3

Example with Custom Parameters:

python main.py --batch_size 64 --epochs 50 --lr 0.0005

Output Files

• Samples: After every epoch, an image grid of generated MNIST digits is saved in ./samples/ as MNIST_{epoch}.png.
• DDIM Samples: After training, DDIM samples are generated at steps [5, 10, 20, 50] and saved as DDIM_{steps}.png in the same folder.
• Losses: The training loss for each epoch is saved to a file named loss_epochs.pkl.