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):
clipper, sampler = clipper_ret
sampler.set_data_loader(sample_data_loader)
sampler.keep_memory = False
else:
clipper = clipper_ret
sampler = None
scheduler = get_scheduler(config.scheduler, config)
def callback_before_train(_0, _1, _2):
print(clipper.info())
supervisor = BasicSupervisorMNIST(config, clipper, scheduler, sampler=sampler,
callback_before_train=callback_before_train)
if config.adaptive_rate:
supervisor.put_key("lr", lr)
print(gan_data_loader)
train(config, gan_data_loader, mnist.generator_forward, mnist.discriminator_forward,
gen_optimizer=gen_optimizer,
disc_optimizer=disc_optimizer, accountant=accountant,
supervisor=supervisor, n_samples=n_samples)
disc_optimizer = tf.train.AdamOptimizer(lr, beta1=config.beta1,
beta2=config.beta2)
clipper_ret = get_clipper(config.clipper, config)
if isinstance(clipper_ret, tuple):
clipper, sampler = clipper_ret
sampler.set_data_loader(LSUNLoader(config.sample_dir, block_size=18, max_blocks=256,
num_workers=2, actions=lsun_process_actions()))
sampler.data_loader.start_fetch()
else:
clipper = clipper_ret
sampler = None
scheduler = get_scheduler(config.scheduler, config)
def callback_before_train(_0, _1, _2):
print(clipper.info())
supervisor = BasicSupervisorLSUN(config, clipper, scheduler,
sampler=sampler, callback_before_train=callback_before_train)
if config.adaptive_rate:
supervisor.put_key("lr", lr)
try:
train(config, data_loader, generator_forward, discriminator_forward,
disc_optimizer=disc_optimizer,
gen_optimizer=gen_optimizer, accountant=accountant, supervisor=supervisor)
finally:
data_loader.stop_fetch()
if sampler is not None:
sampler.data_loader.stop_fetch()