This project implements a Conditional Generative Adversarial Network (CGAN) for the MNIST handwritten digits dataset, using a simple and effective fully connected (linear) architecture for both the generator and discriminator. The model is implemented in PyTorch Lightning for modularity and ease of training.

Conditional Generation: Generate images of a specific digit (0–9) by conditioning both the generator and discriminator on the label.

Linear (Fully Connected) Architecture: Both networks use only linear layers, making the code simple, fast, and easy to understand.

One-hot Label Conditioning: Labels are encoded as one-hot vectors and concatenated to the noise vector (generator) and image vector (discriminator).

Stable Training: Uses batch normalization, LeakyReLU activations, and the Adam optimizer for robust and stable convergence.

PyTorch Lightning: Modular code structure for easy experimentation and reproducibility.

Quick Training: Trains in minutes on CPU or GPU; 5–20 epochs is usually sufficient for good results.

Simplicity: Minimal code, easy to debug and extend. Ideal for learning and research.

Speed: Trains extremely fast due to low parameter count and small image size (28x28).

Stability: Less prone to mode collapse and instability than deeper or convolutional GANs on MNIST.

Sufficient Capacity: For MNIST, linear models are sufficient to generate high-quality, label-conditional samples. More complex architectures do not yield significant improvements.

Reproducibility: Results are highly reproducible and less sensitive to hyperparameters.

Loss Curves: Discriminator and generator losses converge smoothly.

Sample Quality: Generated digits are visually convincing for all classes.

Advanced Tricks: Techniques like spectral norm, self-attention, or deep CNNs do not significantly improve results on MNIST compared to this linear CGAN.

model.py: Contains the Generator, Discriminator, and GAN LightningModule.

data.py: Loads and preprocesses the MNIST dataset using PyTorch Lightning DataModule.

train.py: Training script (5 epochs by default).

Install dependencies (PyTorch, torchvision, pytorch-lightning).

Run training:

Monitor results: Generated images and loss curves are logged for each epoch.

Educational purposes: Learn GANs and conditional generation with minimal code.

Research baseline: Use as a baseline for MNIST or other simple datasets.

Quick prototyping: Rapidly test new ideas in a stable, reproducible setting.