Privileged-GAN introduces a novel way to train classical GAN models, ensuring consistent and stable results. This framework combines two GANs and an autoencoder and has been trained on the MNIST dataset.
Privileged-GAN is an unsupervised learning framework based on adversarial training. It consists of:
- Two GANs: A parent GAN and a child GAN.
- An autoencoder: Functions as the generator for the child GAN.
The architecture for Version 1 is shown below:
Details:
- The parent GAN starts with a generator initialized with random noise.
- The child GAN's generator is an autoencoder that processes outputs from the parent GAN, making it a "Privileged-GAN."
- Both GANs follow the classical GAN training approach: the parent GAN is trained first, and the child GAN is trained afterward, learning from the parent's successes.
- With ANN:
- With CNN:
The architecture for Version 2 is shown below:
Improvements in Version 2:
- Similar structure to Version 1.
- The child GAN learns from both the successes and failures of the parent GAN, further improving its training dynamics.
- With ANN:
- With CNN:
- Stable Training: Incorporates a privileged learning approach for smoother convergence.
- Modular Framework: Combines GANs and autoencoders for flexible training pipelines.
- Scalability: Demonstrated success on MNIST, adaptable to other datasets.
Clone the repository and follow the provided documentation to train your own Privileged-GAN models.
git clone https://github.com/basantbhandari/Privilaged-GAN.git
cd Privileged-GAN