def generator(self, z, is_training, reuse=False):
if self.architecture == consts.INFOGAN_ARCH:
return super(AbstractGANWithPenalty,
self).generator(z, is_training, reuse)
elif self.architecture == consts.DCGAN_ARCH:
return dcgan_architecture.generator(z, self.batch_size,
self.output_height,
self.output_width, self.c_dim,
is_training, reuse)
elif self.architecture == consts.RESNET5_ARCH:
assert self.output_height == self.output_width
return resnet_architecture.resnet5_generator(
z,
is_training=is_training,
reuse=reuse,
colors=self.c_dim,
output_shape=self.output_height)
elif self.architecture == consts.RESNET_STL:
return resnet_architecture.resnet_stl_generator(
z, is_training=is_training, reuse=reuse, colors=self.c_dim)
elif self.architecture == consts.RESNET107_ARCH:
return resnet_architecture.resnet107_generator(
z, is_training=is_training, reuse=reuse, colors=self.c_dim)
elif self.architecture == consts.RESNET_CIFAR:
return resnet_architecture.resnet_cifar_generator(
z, is_training=is_training, reuse=reuse, colors=self.c_dim)
elif self.architecture == consts.SNDCGAN_ARCH:
return dcgan_architecture.sn_generator(z, self.batch_size,
self.output_height,
self.output_width,
self.c_dim, is_training,
reuse)
else:
raise NotImplementedError("Architecture %s not implemented." %
self.architecture)
def TestResnet5GeneratorShape(output_shape):
config = tf.ConfigProto(allow_soft_placement=True)
tf.reset_default_graph()
batch_size = 8
z_dim = 64
os = output_shape
with tf.Session(config=config) as sess:
z = tf.random_normal([batch_size, z_dim])
g = resnet_arch.resnet5_generator(
noise=z, is_training=True, reuse=False, colors=3, output_shape=os)
tf.global_variables_initializer().run()
output = sess.run([g])
return [output[0].shape, (batch_size, os, os, 3)]