FshnGan Project

Overview

This project implements a Generative Adversarial Network (GAN) using TensorFlow to generate synthetic images resembling the Fashion MNIST dataset. The GAN consists of a generator and a discriminator trained in an adversarial manner to produce realistic fashion images.

Project Structure

The project is organized as follows:

• 1. Import Dependencies and Data: Install necessary libraries, import TensorFlow, and load the Fashion MNIST dataset.

• 2. Visualize Data and Build Dataset: Visualize a few samples from the dataset, perform data transformations, and build a pipeline for training.

• 3. Build Neural Network:

  • 3.1 Import Modelling Components: Import essential components for building the generator and discriminator.
  • 3.2 Build Generator: Define the architecture of the generator model.
  • 3.3 Build Discriminator: Define the architecture of the discriminator model.

• 4. Construct Training Loop:

  • 4.1 Setup Losses and Optimizers: Define loss functions and optimizers.
  • 4.2 Build Subclassed Model: Subclass a model for training the GAN.
  • 4.3 Build Callback: Implement a callback to monitor and save generated images during training.
  • 4.3 Train: Train the GAN model using the Fashion MNIST dataset.
  • 4.4 Review Performance: Plot and review the training loss curves.

• 5. Test Out the Generator:

  • 5.1 Generate Images: Generate sample images using the trained generator.
  • 5.2 Save the Model: Save the generator and discriminator models.

Usage

1. Ensure you have the required dependencies installed by running:

   pip install tensorflow tensorflow-gpu matplotlib tensorflow-datasets ipywidgets

2. Execute the provided code in a Jupyter Notebook or Python environment.

3. Review the generated images during training in the 'images' directory.

4. Customize hyperparameters, model architectures, and other settings as needed.

Results

The trained GAN produces synthetic fashion images that closely resemble the Fashion MNIST dataset. Check the 'images' directory for visualizations.

Notes

• This implementation assumes a TensorFlow environment and requires GPU support for optimal training performance.

• Experiment with hyperparameters, model architectures, and training duration for better results.

Author

Ruchit

License

This project is licensed under the MIT License.