Generative Adversarial Networks (GAN) – Fashion MNIST

• Vanilla GAN (Fully Connected Networks)
• Deep Convolutional GAN (DCGAN)

• Image quality
• Training stability
• Effect of architecture on performance

Dataset

• Fashion-MNIST
• 28×28 grayscale images of clothing items
• Dataset automatically downloaded using PyTorch

Preprocessing

• Converted images to tensors
• Normalized pixel values to range [-1, 1]

Model Architectures

1. Vanilla GAN

• Generator
  • Fully connected layers
  • ReLU activations
• Discriminator
  • Fully connected layers
  • LeakyReLU activations

2. DCGAN

• Generator
  • Transposed Convolution layers
  • Batch Normalization
  • ReLU activations
• Discriminator
  • Convolution layers
  • Batch Normalization
  • LeakyReLU activations

Training Process

1. Train discriminator on:
   • Real images (label = 1)
   • Fake images (label = 0)
2. Train generator to:
   • Fool discriminator (label = 1)
3. Loss function:
   • Binary Cross Entropy (BCE)
4. Optimizer:
   • Adam optimizer (lr = 0.0002)

Observations

• DCGAN produces sharper and more realistic images
• Vanilla GAN generates blurry outputs
• Training is unstable in early epochs

Loss Behavior

• Generator and discriminator losses oscillate
• DCGAN shows relatively stable convergence