Abstract by Michael Brodie
The recent CoachGAN algorithm provides an efficient, inference-time output enhancement method for generative models. We introduce a new model health metric called the Discriminator Conditioning Number (DCN). DCN extracts the maximum and minimum eigenvalues of the squared Jacobian of the discriminator for a random input sample. The quotient of these values serves as a compact but informative number that indicates the stability of CoachGAN if used with the current discriminator. Using state-of-the-art GAN models and a variety of datasets, we use FID and IS metrics to quantitatively validate that DCN is an accurate measure of CoachGAN effectiveness ahead of time. We also introduce a fine-tuning algorithm that improves the DCN and leads to higher FID and IS scores. We demonstrate these improvements across a variety of tasks, including image superresolution, image inpainting, image-to-image translation, conditional image generation, and unconditional image generation.