Project: GAN Generation

Backup of GAN Learning Project (August 2022)

[!NOTE] The project explores various GAN architectures and improvements through iterative versions.

[!IMPORTANT] This project is a personal learning exercise in understanding and implementing different GAN techniques.

This project re-implemented GAN, WGAN and conditional GAN and explored the typical problems that occurred with GAN-based architectures like mode collapse and sensitivity to hyperparameters.

References

• The YouTube video series "Generative Adversarial Networks (GANs)", Aladdin Persson, at https://www.youtube.com/playlist?list=PLhhyoLH6IjfwIp8bZnzX8QR30TRcHO8Va

Setups

Environment:

• Python version: 3.x
• Framework: PyTorch
• Dataset: MNIST
• Runned on Google Colab

Experiments

MNIST Digit Generation

Explored various GAN architectures:

• Standard GAN
• Wasserstein GAN (WGAN)
• Conditional WGAN

Key Experiments

• Experimented with different learning rates
• Observed the phoenomenon of mode collapse and the sensitivity of the GAN architecture to the learning rate
• Understanding the architecture of GAN, improvements made by WGAN, and also the principles of providing class conditions to GANs

Experiment Results

1. Vanilla GAN

   v2 (100 epochs)	v3 (35 epochs *early stopped)	v4 (100 epochs)

2. Wasserstein GAN (WGAN)

   v5 (100 epochs)

3. Conditional Wasserstein GAN (incomplete)