Hello all,
I executed the program in the GauGAN tutorial and obtained not very impressive results. I’m wondering if anybody obtained better results and if they did anything different than the notebook.
We can notice that the discriminator loss dropped to -33 very fast, but that the generator loss is still decreasing, albeit being still much higher (0.33). Would this be an indicator of the discriminator outperforming the generator too soon and no longer learning? or perhaps even the opposite since although the discriminator loss is lower, it is stagnant?
I’m assuming the loss function take this into account already as it uses Kullback Leibler loss, Feature Loss, and perceptual loss.
Should we tweak the parameters? How to improve the results? Handicap the discriminator? handicap the generator? by slowing the learning rate?
