This repository contains an implementation of a Relativistic Average Least Squares GAN (RaLSGAN), a variation of Generative Adversarial Networks (GANs) designed to improve stability and realism in image generation. The model is trained to generate high-quality images using adversarial training techniques.
The dataset consists of a collection of images used for training the GAN. The images undergo preprocessing to enhance the training process, including resizing, normalization, and conversion to tensors.
- Images are loaded using
PILandtorchvision.transforms. - Resized to a fixed dimension suitable for the model.
- Normalized to a range of [-1,1] to improve GAN stability.
- Converted into PyTorch tensors for training.
RaLSGAN consists of:
- Generator: A deep neural network that learns to generate realistic images from random noise.
- Discriminator: A classifier that distinguishes between real and generated images using the relativistic average loss, which compares real and fake samples in a more stable manner.
- Loss Function: The least squares loss function is used to improve training dynamics and reduce mode collapse.
The architecture is implemented using PyTorch.
- The model is trained for 100 epochs on a dataset of 2000 images.
- Uses the Adam optimizer for both Generator and Discriminator.
- The training loop alternates between updating the Generator and Discriminator using relativistic average loss.
- Training progress is tracked using loss curves and generated images.
- Generated images are saved and visually inspected to assess quality.
- Loss curves for both the Generator and Discriminator are plotted to evaluate model stability.
- Model outputs are compared at different epochs to observe improvements.
- The model successfully generates high-quality images with realistic features.
- The relativistic approach leads to better convergence and image diversity compared to standard GANs.
- The training process demonstrates stable loss trends, minimizing mode collapse.