Fashion-GAN: Generating Diverse Fashion Images with Deep Convolutional Generative Adversarial Networks

Overview

This project harnesses the power of Generative Adversarial Networks (GANs) to try to produce conventional and novel images of various clothing and apparel items. It leverages the Fashion-MNIST dataset, a collection of 60,000 grayscale images (28x28 pixels) spanning 10 fashion categories, commonly used for image classification tasks. This project explores the potential of GANs to generate creative fashion designs, offering a unique twist on traditional classification approaches.

Dataset and Model Architecture

Dataset:

Fashion-MNIST:
    60,000 28x28 grayscale images
    10 fashion categories (e.g., shoes, jeans, etc.)
    Available within TensorFlow Keras API datasets

Model Architecture:

Generator:
    7-layer Deep Convolutional GAN (DCGAN) architecture
    Alternating feed-forward and convolutional layers
    Leaky ReLU activation for feed-forward layers
Discriminator:
    6-layer DCGAN architecture
    Alternating feed-forward and convolutional layers

Key Features:

GAN-Based Image Generation: Trains a GAN to produce realistic fashion images.
Fashion-MNIST Exploration: Leverages a dataset commonly used for classification in a generative context.
Leaky ReLU Experimentation: Explores Leaky ReLU's effectiveness in model performance.

^{For in-depth analysis of hyperparameter tuning experiments and future research directions, refer to the comprehensive report}