About Project

Generating realistic handwritten digit images requires a model capable of learning complex data distributions from limited training samples. Traditional image generation methods often lack diversity and tend to overfit the dataset.

The key challenge is to train a GAN (Generative Adversarial Network) that can:

• Generate sharp and visually diverse handwritten digits
• Maintain training stability throughout epochs
• Avoid common pitfalls such as mode collapse and vanishing gradients

Problem Statement

Generating realistic handwritten digit images requires a model capable of learning complex data distributions from limited training samples. Traditional image generation methods often lack diversity and tend to overfit the dataset.

The key challenge is to train a GAN (Generative Adversarial Network) that can:

• Generate sharp and visually diverse handwritten digits
• Maintain training stability throughout epochs
• Avoid common pitfalls such as mode collapse and vanishing gradients

Objective

To implement and train a Generative Adversarial Network (GAN) on the MNIST dataset that learns to generate synthetic 28×28 grayscale handwritten digit images.

• Monitor training progression using generated image samples and loss plots
• Perform hyperparameter tuning to identify optimal training configurations
• Justify all architectural and design decisions for both generator and discriminator
• Explore advanced GAN variants such as:
  • DCGAN – Deep Convolutional GAN
  • WGAN – Wasserstein GAN for better stability
  • StyleGAN – For high-quality and stylized image synthesis

Proposed Solution

• Generator: Upsamples random noise into 28×28 grayscale images using Dense and Conv2DTranspose layers for feature expansion and image shaping.
• Discriminator: Classifies real vs. fake images using Conv2D layers with Dropout for regularization and overfitting control.
• Custom GAN Class: Encapsulates the full training loop with a custom train_step() method for adversarial updates of both generator and discriminator.
• LossPlotCallback: Custom Keras callback to visualize and save generated image samples and loss trends after each epoch, aiding in training diagnostics.

Technologies Used

• TensorFlow / Keras – For building and training GAN models
• Matplotlib / PIL – For saving generated images and plotting loss curves

• MNIST – Grayscale handwritten digit dataset (28×28)

• Adam Optimizer – With learning rate and β₁ tuning for stability
• Binary Cross-Entropy – Used as the loss function for both generator and discriminator

• Jupyter Notebook / Python – For implementation and experimentation

Challenges Faced

• Generator instability observed at higher epochs, requiring monitoring and early stopping
• Mode collapse risk was mitigated using dropout in the discriminator and label smoothing (0.9 for real labels)
• Spiking generator loss indicated increasing difficulty in fooling a strong discriminator
• Small image resolution (28×28) limited the expressive power of the generator

Methodology

• Input: 100-dimensional latent vector
• Dense layer: Output shape 7×7×256
• Conv2DTranspose layers:
  • 128 filters → 7×7
  • 64 filters → 14×14
  • 1 filter → 28×28 (final output)
• Activations: LeakyReLU for intermediate layers, Tanh for output
• Output: 28×28 grayscale image

• Input: 28×28×1 image
• Conv2D layers:
  • 64 filters + Dropout
  • 128 filters + Dropout
• Output: Flatten → Dense → Sigmoid (probability real/fake)

• For each batch:
Result / Outcome
🖼️ Visual Output
• Epoch 1: Pure random noise
• Epoch 10: Digit-like shapes begin to emerge
• Epoch 30: Recognizable, sharp handwritten digits
📉 Loss Behavior

Epoch	Discriminator Loss (↓)	Generator Loss (↑)
1	0.45	0.44
10	0.65	0.91
30	0.64	0.95

• Discriminator loss remained relatively stable across epochs
• Generator loss increased — expected behavior as it learns to fool the discriminator
🔁 Hyperparameter Tuning
• Tested configurations:
  • Learning rates: 0.0002, 0.0001
  • β₁ (Adam): 0.5, 0.4
  • Latent dimensions: 100, 128
• Best configuration:
  • Learning rate = 0.0002
  • β₁ = 0.4
  • Latent dimension = 100
• Final loss metrics: Generator = 0.80, Discriminator = 0.63