def __init__(self):
self.en = encoder(FLAGS.z_dim)
self.de = decoder()
self.di = discriminator()
self.x = tf.placeholder(tf.float32, [None, 28, 28, 1], name='x')
self.real_z = tf.placeholder(tf.float32, [None, FLAGS.z_dim])
self.fake_z = self.en(self.x)
self.rec_x, _ = self.de(self.fake_z, False)
self.gen_x, _ = self.de(self.real_z)
self.g_loss = tf.reduce_mean(self.di(self.fake_z, False))
self.d_loss = tf.reduce_mean(self.di(self.real_z)) - tf.reduce_mean(
self.di(self.fake_z)) + self._dd()
self.a_loss = tf.reduce_mean(tf.square(self.rec_x - self.x))
self.g_optim = tf.train.AdamOptimizer(1e-3).minimize(
self.g_loss, var_list=self.en.vars)
self.d_optim = tf.train.AdamOptimizer(1e-3).minimize(
self.d_loss, var_list=self.di.vars)
self.a_optim = tf.train.AdamOptimizer(1e-3).minimize(
self.a_loss, tf.train.get_or_create_global_step())
self.fit_summary = tf.summary.merge([
tf.summary.scalar('g_loss', self.g_loss),
tf.summary.scalar('d_loss', self.d_loss),
tf.summary.scalar('a_loss', self.a_loss),
tf.summary.histogram('real_z', self.real_z),
tf.summary.histogram('fake_z', self.fake_z),
tf.summary.image('x', self.x, 8),
tf.summary.image('rec_x', self.rec_x, 8)
])
self.gen_summary = tf.summary.merge(
[tf.summary.image('gen_x', self.gen_x, 8)])
def __init__(self):
self.en = encoder()
self.de = decoder()
self.z = tf.placeholder(tf.float32, [None, FLAGS.z_dim], name='z')
self.x = tf.placeholder(tf.float32, [None, 28, 28, 1], name='x')
en_ = self.en(self.x)
self.mu = tf.layers.dense(en_, FLAGS.z_dim)
self.log_var = tf.layers.dense(en_, FLAGS.z_dim)
eps = tf.random_normal(tf.shape(self.mu))
self.z_latent = self.mu + tf.exp(0.5 * self.log_var) * eps
self.rec_x, logits = self.de(self.z_latent, False)
self.gen_x, _ = self.de(self.z)
# self.loss_nll = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=logits, labels=self.x))
self.loss_nll = tf.reduce_mean(tf.square(self.x - self.rec_x))
self.loss_mmd = compute_mmd(self.z, self.z_latent)
self.loss = self.loss_nll + FLAGS.beta * self.loss_mmd
self.optim = tf.train.AdamOptimizer(1e-3).minimize(
self.loss, tf.train.get_or_create_global_step())
self.fit_summary = tf.summary.merge([
tf.summary.scalar('loss', self.loss),
tf.summary.scalar('loss_nll', self.loss_nll),
tf.summary.scalar('loss_mmd', self.loss_mmd),
tf.summary.image('x', self.x, 8),
tf.summary.image('rec_x', self.rec_x, 8),
tf.summary.histogram('z', self.z_latent),
tf.summary.histogram('mu', self.mu),
tf.summary.histogram('logvar', self.log_var)
])
self.gen_summary = tf.summary.merge(
[tf.summary.image('gen_x', self.gen_x, 8)])
