Denoising Diffusion Probabilistic Models (DDPM) - MNIST Implementation

This repository provides a PyTorch implementation of Denoising Diffusion Probabilistic Models (DDPM), based on the paper:

Denoising Diffusion Probabilistic Models
Ho et al., NeurIPS 2020
📄 Read the paper

This implementation trains a DDPM model on the MNIST dataset, converted from CSV format to image format.

---

🔧 Quickstart

📁 1. Data Preparation

1. Create image folders:

   mkdir -p data/train/images
   mkdir -p data/test/images

2. Download the MNIST CSV files from Kaggle - MNIST in CSV and place them inside the data/ directory:

   data/
   ├── mnist_train.csv
   └── mnist_test.csv
   

3. Extract images using the script:

   python utils/extract_mnist_images.py

4. You should now have:

   data/
   ├── train/
   │   └── images/
   │       ├── 0/
   │       ├── 1/
   │       └── ...
   └── test/
       └── images/
           ├── 0/
           ├── 1/
           └── ...
   

---

🚀 2. Training the Model

Train the DDPM model using:

python train_ddpm.py --config config/default.yaml

This will save model checkpoints to default/ddpm_ckpt.pth.

---

🎨 3. Sampling Images

To generate samples using the trained model:

python sample_ddpm.py --config config/default.yaml

Generated images will be saved in:

default/samples/x0_*.png

---

📦 Installation

git clone https://github.com/momoth12/Denoising-Diffusion-Probabilistic-Models.git
cd Denoising-Diffusion-Probabilistic-Models
pip install -r requirements.txt

---

📁 Project Structure

.
├── config/
│   └── default.yaml
├── dataset/
│   └── mnist_dataset.py
├── models/
│   └── unet_base.py
├── scheduler/
│   └── linear_noise_scheduler.py
├── scripts/
│   ├── train_ddpm.py
│   └── sample_ddpm.py
├── utils/
│   └── extract_mnist_images.py
├── data/

├── config/ │ └── default.yaml ├── dataset/ │ └── mnist_dataset.py ├── models/ │ └── unet_base.py ├── scheduler/ │ └── linear_noise_scheduler.py ├── utils/ │ └── extract_mnist_images.py ├── data/ ├── default/ │ └── ddpm_ckpt.pth (after training) │ └── samples/ ├── train_ddpm.py ├── sample_ddpm.py ├── README.md └── requirements.txt

│   └── ddpm_ckpt.pth (after training)
│   └── samples/
├── README.md
└── requirements.txt

---

📌 Acknowledgements

This repository is a clean reimplementation of the DDPM framework from the original paper by Ho et al. It is designed for educational and research use on small datasets like MNIST.

---

🧠 Future Ideas

• Conditional DDPM on digit labels
• Training on CIFAR-10 or CelebA
• Integration with logging tools like TensorBoard or Weights & Biases