I’ve recently been having this weird issue where my discriminator losses tend to 0 quite quickly, while my generator losses tend to 1 (see photo). My understanding of GAN’s is that the optimal solution should approach the Nash equilibrium (around 0.5 for both generator and discriminator). My GAN loss objective function is a least squares formulation, as per the options default.
I've tried lambda-identity 0.5 and 0.1, and the same phenomenon occurs.
Hi,
I’ve recently been having this weird issue where my discriminator losses tend to 0 quite quickly, while my generator losses tend to 1 (see photo). My understanding of GAN’s is that the optimal solution should approach the Nash equilibrium (around 0.5 for both generator and discriminator). My GAN loss objective function is a least squares formulation, as per the options default.
I've tried lambda-identity 0.5 and 0.1, and the same phenomenon occurs.