search

tatsy / keras-generative

Deep generative networks, coded with Keras.
122 stars 28 forks source link

cvaegan diverge when dataset is cifar10 #1

Closed xzqjack closed 3 years ago

xzqjack commented 7 years ago

Hi, your "keras-generative" implementation is really fascinating. When training the cvaegan.py on cifar10 with your code and hyper-setting parameters, the generator diverge and the generated image are noise -like. I haven't try celebA dataset. Have you generated real-like image with cvaegan.py? Is there any setting i should have pay attention to? I look forward your reply. Thanks in advance.

xtarx commented 6 years ago

Same here when trained using ali.py ,but the other-way around the discriminator diverge

tatsy commented 6 years ago

Talking about the dataset cifar10, I think it is a little difficult to generate real-like image because the variation of the data is quite wide. So, maybe neither the algorithm CVAE-GAN nor ALI will not generate real-like images, while I have not tested them.