DCGAN for MNIST Data Generation

Overview

This repository implements a Deep Convolutional Generative Adversarial Network (DCGAN) to generate synthetic handwritten digit images similar to those in the MNIST dataset.

Project Description

The goal is to train a DCGAN that learns the distribution of the MNIST dataset and produces realistic handwritten digits. The implementation consists of two main neural networks:

• Generator: Converts a random noise vector into a 28×28 grayscale image.
• Discriminator: Distinguishes between real images from the MNIST dataset and images generated by the Generator.

The adversarial training process allows both models to improve over time—ultimately resulting in a generator capable of producing plausible digits.

Implementation Details

• Language & Framework: Python 3.12 with TensorFlow 2.x (using Keras API)

• Training Parameters:

  • Epochs: 200
  • Loss Function: Binary Cross-Entropy is used for both generator and discriminator losses
  • Optimizers: Adam optimizer with customized learning rates and decay schedules

• Data: Utilized MNIST dataset. MNIST is a dataset of 70,000 28×28 grayscale images of handwritten digits, split into 60,000 training and 10,000 test samples. It is widely used as a benchmark for image classification and deep learning experiments due to its simplicity and standardized format. MNIST DATA

• Generator

    def build_generator():
        model = tf.keras.Sequential([
            layers.Dense(7*7*256, use_bias=False, input_shape=(100,)),  # Input: Noise vector
            layers.Reshape((7, 7, 256)),  # Reshape to small feature map
            layers.BatchNormalization(),
            layers.LeakyReLU(alpha=0.2),

            layers.Conv2DTranspose(128, (4,4), strides=(2,2), padding='same', use_bias=False),
            layers.BatchNormalization(),
            layers.LeakyReLU(alpha=0.2),

            layers.Conv2DTranspose(64, (4,4), strides=(2,2), padding='same', use_bias=False),
            layers.BatchNormalization(),
            layers.LeakyReLU(alpha=0.2),
            
            layers.Conv2DTranspose(1, (4,4), strides=1, padding="same", activation="tanh")
        ])

• Descriminator

    def build_discriminator():
    base_model = keras.Sequential([
        layers.Conv2D(6, kernel_size=5, strides=1, padding="valid", input_shape=(28, 28, 1)),
        layers.LeakyReLU(0.2),
        layers.AveragePooling2D(pool_size=(2, 2)),

        layers.Conv2D(16, kernel_size=5, strides=1, padding="valid"),
        layers.LeakyReLU(0.2),
        layers.AveragePooling2D(pool_size=(2, 2)),

        layers.Flatten(),
        layers.Dense(120, activation="relu"),
        layers.Dense(84, activation="relu"),
        layers.Dense(1),  # Output probability (Real or Fake)
    ])

Results

Final Epoch Grid	Training Gif

Loss Function Over 200 Epochs

The loss plot below depicts the training progress, showing the generator and discriminator losses throughout 200 epochs.

Reference

• Deep Convolutional Generative Adversarial Network