Deep Generative Modelling

This repository contains learning-oriented PyTorch implementations of basic deep generative models. The code was written as part of self-study to understand the fundamentals of variational autoencoders, generative adversarial networks, and diffusion-based generative models.

Academic Reference

The implementations in this repository were developed after completing the lecture series:

Mathematical Foundations of Generative AI - YouTube Playlist
Instructor: Prof. Prathosh A P
Offered by IIT Madras
Division of Electrical, Electronics, and Computer Science (EECS), IISc Bangalore

The course provided the theoretical foundations for variational inference, adversarial learning, and diffusion-based generative modeling.

Implemented Models

Variational Autoencoder (VAE)

• Encoder–decoder architecture
• Reparameterization trick
• Reconstruction loss with KL divergence regularization

Generative Adversarial Network (GAN)

• Convolutional GAN architecture for MNIST
• Separate generator and discriminator networks trained adversarially
• Binary cross-entropy objective implemented using BCEWithLogitsLoss
• Alternating optimization of generator and discriminator
• Sampling via latent noise vectors

Denoising Diffusion Probabilistic Model (DDPM)

• Forward noising and reverse denoising processes
• Noise prediction objective using MSE loss
• Sampling via the standard DDPM reverse process

Conditional DDPM with Classifier-Free Guidance (DDPM-CFG)

• Class-conditional diffusion model
• Classifier-free guidance implemented via random label dropping during training
• Adjustable guidance scale during sampling

Repository Structure

Deep_Generative_Modelling/
├── models/
│   ├── vae/
│   ├── gan/
│   └── diffusion/
├── train/
├── eval/
├── sample/
├── checkpoints/
├── data/
├── legacy/
├── requirements.txt
└── README.md

Purpose

• To understand generative modeling concepts by implementing them from scratch
• To experiment with VAEs, GANs, and diffusion models in PyTorch
• To analyze training dynamics (loss behavior, adversarial balance, sampling quality)
• To serve as a base for further learning and extensions

Requirements

• Python 3.9+
• PyTorch
• torchvision
• numpy
• matplotlib

Install dependencies using:

pip install -r requirements.txt

Usage

Train models using:

python train/train_vae.py
python train/train_gan.py
python train/train_ddpm.py

Sample from trained models using:

python eval/gan/sample_gan.py
python sample/sample_ddpm.py

Trained model checkpoints are expected in the checkpoints/ directory.