def __init__(self,
sess=tf.Session(),
image_size=(32, 32),
z_size=201,
n_iterations=50,
dataset="None",
batch_size=64,
lrate=0.002,
d_size=64):
self.image_size = image_size
self.n_iterations = n_iterations
self.batch_size = batch_size
self.lrate = lrate
self.session = sess
self.base_dim = 512
self.d_size = 32
self.z_size = z_size
self.tau = 1
self.dataset_name = dataset
#self.alpha = tf.constant(1e-6)
self.alpha = tf.constant(0.0)
self.beta = tf.constant(0.0)
self.size = image_size[0]
self.logpath = "log"
self.d_bn0 = ops.BatchNormalization([256], 'd_bn0')
#self.d_bn0 = ops.BatchNormalization([self.d_size], 'd_bn0')
self.d_bn1 = ops.BatchNormalization([self.d_size * 2], 'd_bn1')
self.d_bn2 = ops.BatchNormalization([self.d_size * 4], 'd_bn2')
self.history = {}
self.history["generator"] = []
self.history["discriminator_real"] = []
self.history["discriminator_fake"] = []
with tf.variable_scope('inverse_prgan'):
self.images = tf.placeholder(
tf.float32,
shape=[batch_size, image_size[0], image_size[1], 1],
name='final_image')
self.encodings = tf.placeholder(tf.float32,
shape=[batch_size, self.z_size],
name="encoding")
self.train_flag = tf.placeholder(tf.bool)
with tf.variable_scope('inverse_prgan'):
self.z = self.encode(self.images, self.train_flag)
self.loss = ops.l2(self.encodings, self.z)
self.optimizer = tf.train.AdamOptimizer(1e-5, beta1=0.9).minimize(
self.loss)
self.saver = tf.train.Saver(write_version=tf.train.SaverDef.V2)
def __init__(self,
sess=tf.Session(),
image_size=(32, 32),
z_size=201,
n_iterations=50,
dataset="None",
batch_size=64,
lrate=0.002,
d_size=64):
self.image_size = image_size
self.n_iterations = n_iterations
self.batch_size = batch_size
self.lrate = lrate
self.session = sess
self.base_dim = 512
self.d_size = 256
self.z_size = z_size
self.tau = 1
self.dataset_name = dataset
#self.alpha = tf.constant(1e-6)
self.alpha = tf.constant(0.0)
self.beta = tf.constant(0.0)
self.size = image_size[0]
self.logpath = "log"
self.g_bn0 = ops.BatchNormalization([self.d_size], 'g_bn0')
self.g_bn1 = ops.BatchNormalization([self.d_size / 2], 'g_bn1')
self.g_bn2 = ops.BatchNormalization([self.d_size / 4], 'g_bn2')
self.g_bn3 = ops.BatchNormalization([self.d_size / 8], 'g_bn2')
self.d_bn0 = ops.BatchNormalization([self.d_size], 'd_bn0')
self.d_bn1 = ops.BatchNormalization([self.d_size * 2], 'd_bn1')
self.d_bn2 = ops.BatchNormalization([self.d_size * 4], 'd_bn2')
self.history = {}
self.history["generator"] = []
self.history["discriminator_real"] = []
self.history["discriminator_fake"] = []
with tf.variable_scope('gan'):
self.images = tf.placeholder(
tf.float32,
shape=[batch_size, image_size[0], image_size[1], 1],
name='final_image')
self.z = tf.placeholder(tf.float32,
shape=[batch_size, self.z_size],
name='z')
self.train_flag = tf.placeholder(tf.bool)
self.G = self.generator(self.z, self.train_flag)
with tf.variable_scope('gan'):
self.D_real, self.D_real_logits, self.D_stats_real = self.discriminator(
self.images, self.train_flag)
self.D_fake, self.D_fake_logits, self.D_stats_fake = self.discriminator(
self.G, self.train_flag, reuse=True)
self.D_loss_real = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
self.D_real_logits, tf.ones_like(self.D_real)))
self.D_loss_fake = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
self.D_fake_logits, tf.zeros_like(self.D_fake)))
self.G_loss_classic = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
self.D_fake_logits, tf.ones_like(self.D_fake)))
dr_mean, dr_var = tf.nn.moments(self.D_stats_real, axes=[0])
dl_mean, dl_var = tf.nn.moments(self.D_stats_fake, axes=[0])
self.G_loss = ops.l2(dr_mean, dl_mean)
self.G_loss += ops.l2(dr_var, dl_var)
#print self.G_loss.get_shape()
self.D_loss = self.D_loss_real + self.D_loss_fake
allvars = tf.trainable_variables()
self.D_vars = [v for v in allvars if 'd_' in v.name]
self.G_vars = [v for v in allvars if 'g_' in v.name]
self.D_optim = tf.train.AdamOptimizer(1e-4, beta1=0.5).minimize(
self.D_loss, var_list=self.D_vars)
self.G_optim = tf.train.AdamOptimizer(0.0025, beta1=0.5).minimize(
self.G_loss, var_list=self.G_vars)
self.G_optim_classic = tf.train.AdamOptimizer(
0.0025, beta1=0.5).minimize(self.G_loss_classic,
var_list=self.G_vars)
self.saver = tf.train.Saver(write_version=tf.train.SaverDef.V2)