def discriminator(self, x, is_training, reuse=False):
if self.architecture == consts.INFOGAN_ARCH:
return super(AbstractGANWithPenalty,
self).discriminator(x, is_training, reuse)
elif self.architecture == consts.DCGAN_ARCH:
return dcgan_architecture.discriminator(
x, self.batch_size, is_training,
self.discriminator_normalization, reuse)
elif self.architecture == consts.RESNET5_ARCH:
return resnet_architecture.resnet5_discriminator(
x, is_training, self.discriminator_normalization, reuse)
elif self.architecture == consts.RESNET107_ARCH:
return resnet_architecture.resnet107_discriminator(
x, is_training, self.discriminator_normalization, reuse)
elif self.architecture == consts.RESNET_CIFAR:
return resnet_architecture.resnet_cifar_discriminator(
x, is_training, self.discriminator_normalization, reuse)
elif self.architecture == consts.RESNET_STL:
return resnet_architecture.resnet_stl_discriminator(
x, is_training, self.discriminator_normalization, reuse)
elif self.architecture == consts.SNDCGAN_ARCH:
assert self.discriminator_normalization in [
consts.SPECTRAL_NORM, consts.NO_NORMALIZATION
]
return dcgan_architecture.sn_discriminator(
x,
self.batch_size,
reuse,
use_sn=self.discriminator_normalization ==
consts.SPECTRAL_NORM)
else:
raise NotImplementedError("Architecture %s not implemented." %
self.architecture)
def testResnet5DiscriminatorRuns(self):
config = tf.ConfigProto(allow_soft_placement=True)
tf.reset_default_graph()
batch_size = 8
with tf.Session(config=config) as sess:
images = tf.random_normal([batch_size, 128, 128, 3])
out, _, _ = resnet_arch.resnet5_discriminator(
images, is_training=True, discriminator_normalization="spectral_norm",
reuse=False)
tf.global_variables_initializer().run()
output = sess.run([out])
self.assertEquals(output[0].shape, (batch_size, 1))