zackory/keras-mnist-gan

Simple Generative Adversarial Networks for MNIST data with Keras.

keras-mnist-gan's Issues

requirements.txt

please if possible add requirements.txt with all dependencies. right now I'm trying with Keras 2 and Keras 1 - both cases throw errors

Problem with the model.trainable

Hi, Erickson. I found that there may be a problem in your code, you used "discriminator.trainable = False" to freeze the discriminator model, but I found that the model must be recompiled to freeze it .

Why is binary_crossentropy loss used??

According to the paper, Shouldn't loss of discriminator be log(D(x))+ log(1-D(G(z))
And generator be log(1-D(G(z))
so why the loss in both cases is xlog(x) + (1-x)(log(1-x) i.e binary_crossentropy

If they are equivalent then can you show how??

Shows this error when compiling the code

15 from keras import initializers
16
---> 17 K.set_image_dim_ordering('th')
18
19 # The dimensionality has been set at 100 for consistency with other GAN implementations.

AttributeError: module 'keras.backend' has no attribute 'set_image_dim_ordering'