Fake Handwritten digits generation using GAN over MNIST dataset

📌 Project Overview

This project implements a Generative Adversarial Network (GAN) from scratch in Keras & TensorFlow to generate realistic handwritten digits similar to the MNIST dataset.

A GAN consists of two neural networks trained together:

• Generator: Learns to create synthetic digit images from random noise.
• Discriminator: Learns to distinguish between real MNIST images and fake images created by the generator.

Both models compete, and over time the generator produces increasingly realistic digits.

🏗 GAN Architecture used here

Generator	Discriminator

⚙️ Features

• Custom-built Generator and Discriminator networks using TensorFlow/Keras.
• Training loop with tf.GradientTape.
• Model checkpointing to resume or reuse training.
• Visualization of generated digits per epoch.
• Creation of a GIF animation showing generator progress.
• Export of all generated images (epochs 1–100) into a single ZIP archive.

---

📊 Results (Generated Fake Images)

Input	Epoch 1

Epoch 2	Epoch 3

Epoch 9	Epoch 17

Epoch 25	Epoch 50

Epoch 75	Epoch 100

Example (Generated Digits Progress)

---

Real MNIST images:

📈 Possible Improvements

• Implement FID (Fréchet Inception Distance) or Inception Score for evaluation.
• Explore deeper architectures for generator and discriminator.
• Extend GAN training to color image datasets (e.g., CIFAR-10).
• Try Conditional GANs (cGANs) to generate specific digits.