search

eriklindernoren / PyTorch-GAN

PyTorch implementations of Generative Adversarial Networks.
MIT License
16.39k stars 4.07k forks source link

Question : training the discriminator #135

Closed pfeatherstone closed 3 years ago

pfeatherstone commented 3 years ago

@eriklindernoren when training the discriminator, can you concatenate the real images and fake images along the batch dimension, same for the labels, then shuffle the batch dimensions before passing through the discriminator network. Or do you have to infer the real images then the fake images separately?

pfeatherstone commented 3 years ago

Apparently that can cause the discriminator to converge too quickly and you don't reach equilibrium. So probs not a good idea. I think i've answered my own question. I'll keep this open in case someone comes up with some other good arguments.

andylida commented 2 years ago

Hi there! I'm now devoting to a similar problem. If one concat them and there are batchnorm in D, It's easy to fail to function as a traditional GAN. However ,for some task that don't need batchnorm, this doesn't have any side effect. And now I just change all BN to GN.

lanlanwei commented 2 years ago

您好，您的来信我已收到，感谢来信，祝天天好心情。

baicaiPCX commented 2 years ago

您好，您的邮件小皮已经收到啦！！！