search

junyanz / CycleGAN

Software that can generate photos from paintings, turn horses into zebras, perform style transfer, and more.
Other
12.34k stars 1.94k forks source link

Discriminator looses (?) #44

Open SavvaI opened 7 years ago

SavvaI commented 7 years ago

I am trying to apply cyclegan to my data, and at the beginning (1-2 epoch) it kinda makes needed transformation (with added noise), but then it just copies the image 2 times. Losses for G and D are near, so it seems that Generator is fooling discriminator with adversarial examples. I haven't find any solution, maybe someone have faced this problem? I have seen quite complex transformations with cyclegan, like face2food and face2doll

junyanz commented 7 years ago

For the complex transformation, I would recommend a small lambda_cycle (e.g. 1 or 2). Could you show some your results as well as loss plots? How big are your G and D losses?

meeio commented 6 years ago

@junyanz how much the lambda_cycle you use for horse2zebra?

rohitsaluja22 commented 6 years ago

Hey @junyanz, this is Rohit. First, thanks for your code. It's really helpful.

I trained your model on around 1.5 k images at A and B sides. I am observing loss as attached image. gan200

Do I expect Discriminators' errors to go to 0.5 ideally? If yes, should I train for more epochs or try changing lambda_cycle? Please suggest, how can I improve..