Rakibul HaqueOnto the next

Watch a GAN train, live

A real generative adversarial network, training in your browser. Both networks are tiny multilayer perceptrons with hand-written backprop and Adam (no libraries). The generator maps 2-D noise to points; the discriminator tries to tell generated points from real ones sampled off a ring. Press play and watch the amber cloud collapse onto the blue ring as the two networks fight to a standstill. This is the same adversarial idea behind my face-generation GAN and SRGAN, stripped down to 2-D so you can actually see it happen.

iter 0 · D loss - · G loss - · equilibrium ≈ 0.69

← All experiments