Keras-GAN-Animeface-Character

Keras-GAN-Animeface-Character copied to clipboard

The minibatch discrimination proposed here indicates that a bias term is added before concatenating the kernels to the main input. As I'm not 100% sure which implementation is the correct one, I would suggest to check and improve your code accordingly, with an update in the build method

self.b = self.add_weight(shape=(self.nb_kernels,),
                         initializer=keras.initializers.zeros(),
                         name='bias',
                         regularizer=self.W_regularizer,
                         trainable=True,
                         constraint=self.W_constraint)

and in the call method

minibatch_features += self.b

Thank you for the proposal. I might add this as well: https://github.com/atiyo/unbounded_gan