def __init__(self,
gan_type='gan',
net_type='cnn',
optim_type='adam',
batch_size=64,
noise_dim=50,
learning_rate=2e-4,
optim_num=0.5,
clip_num=0.03,
critic_iter=5,
plot_iter=5,
verbose=True):
self.noise_dim = noise_dim
self.class_num = 10
self.clip_num = None if clip_num == 0 else clip_num
self.lr = learning_rate
self.optim_num = optim_num
self.critic_iter = critic_iter
self.plot_iter = plot_iter
self.verbose = verbose
super(GAN, self).__init__(batch_size, gan_type + "_" + net_type)
self.gan_type = gan_type
self.optim_type = optim_type
self.data = datamanager_mnist(train_ratio=1.0,
fold_k=None,
norm=True,
expand_dim=True,
seed=23333)
sample_data = datamanager_mnist(train_ratio=1.0,
fold_k=None,
norm=True,
expand_dim=True,
seed=23333)
self.sample_data = sample_data(self.batch_size,
var_list=["data", "labels"])
if net_type == 'cnn':
self.generator = Generator_CNN('cnn_generator')
self.discriminator = Discriminator_CNN(class_num=self.class_num,
name='cnn_discriminator')
elif net_type == 'mlp':
self.generator = Generator_MLP('mlp_generator')
self.discriminator = Discriminator_MLP(class_num=self.class_num,
name='mlp_discriminator')
self.build_placeholder()
self.build_gan()
self.build_optimizer(optim_type)
self.build_summary()
self.build_sess()
self.build_dirs()
def __init__(self,
optim_type='adam',
net_type='cnn',
batch_size=64,
noise_dim=50,
learning_rate=2e-4,
optim_num=0.5,
critic_iter=1,
plot_iter=5,
verbose=True):
self.noise_dim = noise_dim
self.class_num = 10
self.lr = learning_rate
self.optim_num = optim_num
self.critic_iter = critic_iter
self.plot_iter = plot_iter
self.verbose = verbose
super(InfoGAN, self).__init__(batch_size, "infogan_"+net_type)
# code
self.len_discrete_code = self.class_num # categorical distribution (i.e. label)
self.len_continuous_code = 2 # gaussian distribution (e.g. rotation, thickness)
self.optim_type = optim_type
self.SUPERVISED = True
self.data = datamanager_mnist(train_ratio=1.0, fold_k=None, norm=True, expand_dim=True, seed=23333)
sample_data = datamanager_mnist(train_ratio=1.0, fold_k=None, norm=True, expand_dim=True, seed=23333)
self.sample_data = sample_data(self.batch_size, var_list=["data", "labels"])
if net_type == 'cnn':
self.generator = Generator_CNN(name='cnn_generator')
self.discriminator = Discriminator_CNN(name='cnn_discriminator')
elif net_type == 'mlp':
self.generator = Generator_MLP(name='mlp_generator')
self.discriminator = Discriminator_CNN(name='discriminator')
self.classifier = Classifier_MLP(self.len_discrete_code + self.len_continuous_code, name='classifier')
self.build_placeholder()
self.build_gan()
self.build_optimizer(optim_type)
self.build_summary()
self.build_sess()
self.build_dirs()
def __init__(self, batch_size, version='daegan'):
super(DAEGAN, self).__init__(batch_size, version)
self.critic_iter = 3
self.data_A = datamanager_mnist(datapath=mnist_path, train_ratio=0.8, fold_k=None, expand_dim=True, norm=True, seed=0)
self.data_B = datamanager_mnist(datapath=colorbackground_mnist_path, train_ratio=0.8, fold_k=None, expand_dim=False, norm=True, seed=1)
self.sample_A = self.data_A(self.batch_size, phase='test', var_list=['data', 'labels'])
self.sample_B = self.data_B(self.batch_size, phase='test', var_list=['data', 'labels'])
self.autoencoder_A = CNN_Generator(output_dim=1, name='ae_A')
self.autoencoder_B = CNN_Generator(output_dim=3, name='ae_B')
self.Latent_classifier = MLP_Classifier(output_dim=10, layers=[128]*2, name='latent_c')
self.Latent_discriminator = CNN_Latent_discriminator(name='latent_d')
self.build_placeholder()
self.build_network()
self.build_optimizer()
self.build_summary()
self.build_sess()
self.build_dirs()
def __init__(self, batch_size, version="AE"):
super(AutoEncoder, self).__init__(batch_size, version)
self.data = datamanager_mnist(train_ratio=0.8,
fold_k=None,
expand_dim=True,
norm=True)
self.sample_data = self.data(self.batch_size,
phase='test',
var_list=["data", "labels"])
self.emb_dim = 10
self.encoder = CNN_Encoder(output_dim=self.emb_dim, sn=False)
self.decoder = CNN_Decoder(sn=False)
self.build_placeholder()
self.build_network()
self.build_optimizer()
self.build_sess()
self.build_dirs()
],
feed_dict=feed_dict)
print self.version + " Epoch [%3d/%3d] Iter [%3d/%3d] Da=%.3f Db=%.3f Ga=%.3f Gb=%.3f R=%.3f" % (
epoch, epoches, idx, batches_per_epoch, da, db, ga, gb,
r)
self.writer.add_summary(sum_str, cnt)
if epoch % 20 == 0:
self.sample(epoch)
self.sample(epoch)
self.saver.save(self.sess,
os.path.join(self.model_dir, 'model.ckpt'),
global_step=cnt)
data_A = datamanager_mnist(datapath=mnist_path,
train_ratio=0.8,
fold_k=None,
expand_dim=True,
norm=True,
seed=0)
data_B = datamanager_mnist(datapath=colorbackground_mnist_path,
train_ratio=0.8,
fold_k=None,
expand_dim=False,
norm=True,
seed=1)
cyclegan = CycleGAN(data_A, data_B, 64, gan_type='', version='cyclegan')
cyclegan.train(100)
cyclegan.plot_loss()