Overview
This project uses Generative Adversarial Networks (GANs) to generate realistic images of human faces. The primary objective was to build and train a GAN model that could synthesize facial images convincingly using the CelebA dataset - a large-scale collection of celebrity faces.
Method
- Used the CelebA dataset with 49,736 images loaded via PyTorch's Dataset tool
- Each image was resized to 3x128x128 pixels to optimize training time
- GAN architecture with a generator and discriminator network, both initialized with specific weights
- Trained using the Adam optimizer with Binary Cross Entropy loss over 350 epochs (~4 minutes per epoch)
- Used Python and PyTorch libraries for visualization and loss tracking
Outcomes
- Successfully generated realistic facial images after 350 training epochs
- Loss-vs-iterations plots for both generator and discriminator provided insights into training dynamics
- Demonstrated the practical application of DCGANs for high-quality image synthesis
Media
