MNIST generation using GAN and DCGAN

• Binary classifier (real vs fake images)
• Input: Flattened MNIST images (784 pixels)
• Architecture:
  • 4 fully-connected layers (1024 → 512 → 256 → 1)
  • ReLU activations with Dropout (0.3) for regularization
  • Sigmoid output for probability scoring
• Purpose: Distinguishes between real and generated images

• Image generator from random noise
• Input: 100-dimensional random vector (latent space)
• Architecture:
  • 4 fully-connected layers (256 → 512 → 1024 → 784)
  • ReLU activations in hidden layers
  • Tanh output (values between -1 and 1)
• Output: Generated images reshaped to 28×28 pixels (MNIST format)

• Optimizer: Adam (learning rate: 0.0001)
• Loss Function: Binary Cross Entropy (BCELoss)
• Epochs: 50

• Input: 1×28×28 images
• Architecture: Convolutional classifier
  • Layers:
    • 4 convolutional layers with stride 2
    • Batch normalization + LeakyReLU (0.2)
    • No activation on final output (uses BCEWithLogits)
• Feature Maps: 28 → 56 → 112 → 1
• Output: Real/fake classification score

• Input: 100×1×1 random noise vector
• Architecture: Convolutional transpose (deconvolution) network
  • Layers:
    • 4 transposed convolutional layers
    • Batch normalization + ReLU activation
    • Tanh output (pixel values between -1 and 1)
• Feature Maps: 64 → 128 → 256 → 1
• Output: 1×28×28 generated images

• Weight Initialization: Normal distribution (mean 0.0, std 0.02)
• Optimizer: Adam (beta1=0.5)
• Loss: BCEWithLogits
• Gradient Tracking: Monitors generator gradients (mean & std)std)
• Flexible Loss: Supports both saturating and non-saturating GAN objectives