Post by Frank Mollard

Training GANs is super easy!!! 😏 💡 In a recent project, I explored several variants of Generative Adversarial Networks (GANs), with a particular focus on Wasserstein GANs (WGANs). One of the key advantages of WGANs is their ability to maintain stable convergence even when the critic (the WGAN counterpart of the traditional discriminator) becomes significantly stronger than the generator. In many classical GAN setups, convergence between the discriminator (💔) and generator eventually breaks down because both components must remain very carefully balanced⚖️. In this experiment, documented in the accompanying kaggle-notebook, I investigated whether convergence could be re-stimulated through controlled shocks to the learning rate. I also experimented with perturbations in weights and dropout parameters, although these approaches did not produce meaningful improvements. Interestingly, the results show that adjusting the learning rate can indeed allow the system to resume convergence 🎢... https://lnkd.in/d7ts5MkD For the project, I trained several generators and built a small interactive dashboard that allows users to blend different classes. If you have ever wondered what a hybrid between a human and a dog might look like, the interface lets you explore exactly that: https://lnkd.in/dZZsUGFP A fun experiment - and a small illustration of how delicate yet fascinating the dynamics of adversarial training can be. Image settings: (Generator128_norm.keras, latent space: ~N(0, 1.4), ) #kaggle #datascience #cnn #innovation #deeplearning #python #tensorflow #gan #ai #genai #computervision