gan_generative_adversial_networks

Fashion Texture Generation using GANs

This project demonstrates the use of a Generative Adversarial Network (GAN) to generate diverse textures for fashion items using the Fashion-MNIST dataset. GANs consist of a Generator and a Discriminator that compete with each other, resulting in realistic generated images. The generator creates new textures, while the discriminator attempts to distinguish between real and generated images.

Table of Contents

1. Project Structure
2. Requirements
3. Dataset
4. Model Architecture
5. Training Process
6. Results
7. Usage
8. Acknowledgements

Project Structure

• GenAI_project.ipynb - Main code for training the GAN.
• generated_images/ - Folder where generated sample images from each epoch will be saved.
• final_generated_images.png - Sample of generated images after the last training epoch.
• README.md - Project documentation.

Requirements

Install the required libraries before running the project:

pip install torch torchvision transformers datasets matplotlib

Dataset

The Fashion-MNIST dataset, hosted on Hugging Face, contains grayscale images of 10 classes of clothing items. For this project:

• Resolution: 28x28 pixels, scaled to range [-1, 1].
• Classes: T-shirt/top, Trouser, Pullover, Dress, Coat, Sandal, Shirt, Sneaker, Bag, Ankle boot.

The dataset is automatically loaded from Hugging Face's dataset repository.

Model Architecture

1. Generator:

   • Takes a random vector (latent vector z) of size 100 as input.
   • Passes through fully connected layers with ReLU activations to produce a 28x28x3 image.
   • Uses Tanh activation for the output, resulting in values between [-1, 1].

2. Discriminator:

   • Takes an image (real or generated) as input.
   • Uses fully connected layers with LeakyReLU activations to classify images as real or fake.
   • Final output is a probability from a Sigmoid activation.

Training Process

1. Loading Data:

   • The dataset is loaded, converted to tensor, and normalized to [-1, 1].
   • DataLoader is set up for batching.

2. Training Loop:

   • Discriminator is trained first to distinguish real images from generated ones, computing a loss for real images and for generated (fake) images.
   • Generator is then trained to improve the realism of generated images, using feedback from the discriminator.
   • Both networks are optimized with Adam optimizers.

3. Logging and Saving:

   • Losses for both generator and discriminator are tracked.
   • Sample images are saved after every 5 epochs.

Results

After training, the GAN generates images similar to Fashion-MNIST textures. These images demonstrate the generator's ability to create diverse fashion patterns.

This will:

1. Load the Fashion-MNIST dataset.
2. Train the GAN model.
3. Save generated samples in generated_images/ after every 5 epochs.
4. Save a final set of generated images as final_generated_images.png.

Acknowledgements

• Hugging Face for providing easy access to datasets and models.
• PyTorch and the deep learning community for extensive GAN tutorials.