eriklindernoren/Keras-GAN

CycleGAN Buffer

wave-transmitter opened this issue · 0 comments

Hello there,

according to the paper of CycleGAN:

we update the discriminators using a history of generated images rather than the
ones produced by the latest generators. We keep an image
buffer that stores the 50 previously created images.

As far as I understand this is not implemented in the current model. Any tips on how to implemented? How crucial is it to use that kind of buffer in training?