if config.sample_ratio is not None:
kwargs = {}
gan_data_loader = MNISTLoader_aug(expanded_images, expanded_labels,
first=int(party_data_size *100 * (1 - config.sample_ratio)),
seed=config.sample_seed
)
sample_data_loader = MNISTLoader_aug(expanded_images, expanded_labels,
last=int(party_data_size *100 * config.sample_ratio),
seed=config.sample_seed
)
else:
gan_data_loader = MNISTLoader(config.data_dir, include_train=not config.exclude_train,
include_test=not config.exclude_test)
if config.enable_accounting:
accountant = GaussianMomentsAccountant(gan_data_loader.n, config.moment)
if config.log_path:
open(config.log_path, "w").close()
else:
accountant = None
if config.adaptive_rate:
lr = tf.placeholder(tf.float32, shape=())
else:
lr = config.learning_rate
gen_optimizer = tf.train.AdamOptimizer(config.gen_learning_rate, beta1=0.5, beta2=0.9)
disc_optimizer = tf.train.AdamOptimizer(lr, beta1=0.5, beta2=0.9)
clipper_ret = get_clipper(config.clipper, config)
if isinstance(clipper_ret, tuple):
np.random.seed()
if config.enable_accounting:
config.sigma = np.sqrt(2.0 * np.log(1.25 / config.delta)) / config.epsilon
print("Now with new sigma: %.4f" % config.sigma)
if config.image_size == 64:
data_loader = LSUNLoader(config.data_dir, num_workers=4, actions=lsun_process_actions())
data_loader.start_fetch()
generator_forward = d64_resnet_dcgan.generator_forward
discriminator_forward = d64_resnet_dcgan.discriminator_forward
else:
raise NotImplementedError("Unsupported image size %d." % config.image_size)
if config.enable_accounting:
accountant = GaussianMomentsAccountant(data_loader.num_steps(1), config.moment)
if config.log_path:
open(config.log_path, "w").close()
else:
accountant = None
if config.adaptive_rate:
lr = tf.placeholder(tf.float32, shape=())
else:
lr = config.learning_rate
gen_optimizer = tf.train.AdamOptimizer(config.gen_learning_rate, beta1=config.beta1,
beta2=config.beta2)
disc_optimizer = tf.train.AdamOptimizer(lr, beta1=config.beta1,
beta2=config.beta2)