GAN with keras (cifar10)

[yim....@gmail.com]

Hi everyone!
I'm new in GAN and I'm stuck for cifar10..
MNIST worked fine with this : https://github.com/jacobgil/keras-dcgan/blob/master/dcgan.py

Now, I'm trying to adapt this code for cifar-10
I already readapted the generative model to get 3x32x32 in output.
But, the images I got in output are all gray or all black. And I don't really understand from where it come from.

How would you do ?
Or maybe, do you have a functionnal test examples ?

Thanks!

[Benjamin Striner]

Hello everybody!

Building GANs in keras usually means building your own loops or your own training functions. I abstracted that out in a library that should

make GANs easier to build. It basically lets you wrap several Keras models into a single multiplayer model that can be trained with model.fit. Works for GANs, BiGANs, AAEs, etc.

https://github.com/bstriner/keras_adversarial

https://github.com/bstriner/keras_adversarial/blob/master/examples/cifar10_utils.py

Just put up the CIFAR10 example, so that should be a good start. 

GANs are wild beasts, so don't be surprised if you don't get beautiful images at first, a few small changes to the network can make some radical differences.

Cheers,

Ben

[scs...@gmail.com]

If you discriminator is not adjusted for convolution 3d (images in color - 3 channels rgb), then your generator will never learn to output a color image.

[bice...@gmail.com]

Did you get it working, none of the repos on github has trained cifar10 well as far as my experiments go.

Any help appreciated.