Simple playground for learning and visualizing GANs in 2d. After dozens of hours of StyleGAN training, it may be fun to get GANs intuition with fast iterations (30 seconds) for hyperparameters (but I'm not sure that this intuition can be transferred to the “big” GAN models). Inspired by https://poloclub.github.io/ganlab, but maybe someone prefers to play with hyperparameters using code in Colab.
Visualizations
Visualization of training dynamics includes:
real data distribution (black dots)
generated by G "fake" data from fixed noise
decision boundary for D for the entire input space, color coding displays the output probability for D (red - high probability of real data, blue - low)
green arrows for each generated data point - direction of maximization D's output