def train(self):
# set phase
self._enter_('train')
# get data pipeline
data, info, _ = loads(self.config)
c1_real, p1_real, c2_real, p2_real = tf.unstack(data, axis=1)
label, cond = tf.unstack(info, axis=1)
# encode image to a vector
c1_mu, c1_sigma, feat_c1 = self._encoder(c1_real)
p2_mu, p2_sigma, feat_p2 = self._encoder(p2_real, True)
# children to parent
c1_z = c1_mu + c1_sigma * tf.random_normal(tf.shape(c1_mu))
c1_z = self._generator(c1_z, cond)
c1_fake = tf.clip_by_value(c1_z, 1e-8, 1 - 1e-8)
# parent to children
p2_z = p2_mu + p2_sigma * tf.random_normal(tf.shape(p2_mu))
p2_z = self._generator(p2_z, cond, True)
p2_fake = tf.clip_by_value(p2_z, 1e-8, 1 - 1e-8)
# loss for genertor
E1_loss = self._loss_vae(c1_real, c1_fake, c1_mu, c1_sigma)
E2_loss = self._loss_vae(p2_real, p2_fake, p2_mu, p2_sigma)
E_loss = E1_loss + E2_loss
loss = E_loss
# # allocate two optimizer
global_step = tf.train.create_global_step()
var_e = variables.select_vars('encoder')
var_g = variables.select_vars('generator')
op1 = updater.default(self.config, loss, global_step, var_e, 0)
op2 = updater.default(self.config, loss, None, var_g, 1)
train_op = tf.group(op1, op2)
# update at the same time
saver = tf.train.Saver(var_list=variables.all())
# hooks
self.add_hook(self.snapshot.init())
self.add_hook(self.summary.init())
self.add_hook(
context.Running_Hook(config=self.config.log,
step=global_step,
keys=['E'],
values=[E_loss],
func_test=self.test,
func_val=None))
with context.DefaultSession(self.hooks) as sess:
self.snapshot.restore(sess, saver)
while not sess.should_stop():
sess.run(train_op)
def train(self):
"""
"""
# set phase
self._enter_('train')
# get data pipeline
data, label, _ = loads(self.config)
# generate fake images
logit_G, net_G = self._generator(label)
# discriminate fake images
logit_F, net_F, out_F = self._discriminator(logit_G,
label,
reuse=False)
# discriminate real images
logit_R, net_R, out_R = self._discriminator(data, label, reuse=True)
# classify fake
c_F = self._classifier(out_F, reuse=False)
# classify real
c_R = self._classifier(out_R, reuse=True)
# loss
D_loss, G_loss, C_loss = self._loss(label, logit_F, logit_R, c_F, c_R)
# t_vars = tf.trainable_variables()
d_vars = variables.select_vars('discriminator')
g_vars = variables.select_vars('generator')
c_vars = variables.select_vars('classifier')
# update
step = tf.train.create_global_step()
d_op = updater.default(self.config, D_loss, step, d_vars, 0)
g_op = updater.default(self.config, G_loss, step, g_vars, 0)
c_op = updater.default(self.config, C_loss, step, c_vars, 0)
# assemble
train_op = [[d_op, g_op, c_op]]
saver = tf.train.Saver(variables.all())
# hooks
self.add_hook(self.snapshot.init())
self.add_hook(self.summary.init())
self.add_hook(
context.Running_Hook(config=self.config.log,
step=step,
keys=['D_loss', 'G_loss', 'C_loss'],
values=[D_loss, G_loss, C_loss],
func_test=self.test,
func_val=None))
# monitor session
with context.DefaultSession(self.hooks) as sess:
self.snapshot.restore(sess, saver)
while not sess.should_stop():
sess.run(train_op)
def train(self):
""" """
# set phase
self._enter_('train')
# get data pipeline
data, info, _ = loads(self.config)
c1_real, p1_real, c2_real, p2_real = tf.unstack(data, axis=1)
label, cond = tf.unstack(info, axis=1)
# encode image to a vector
c1_mu, c1_sigma, feat_c1 = self._encoder(c1_real)
p2_mu, p2_sigma, feat_p2 = self._encoder(p2_real, True)
# children to parent
with tf.variable_scope('net1'):
c1_z = c1_mu + c1_sigma * tf.random_normal(tf.shape(c1_mu))
c1_z = self._generator(c1_z, cond)
c1_fake = tf.clip_by_value(c1_z, 1e-8, 1 - 1e-8)
# parent to children
with tf.variable_scope('net2'):
p2_z = p2_mu + p2_sigma * tf.random_normal(tf.shape(p2_mu))
p2_z = self._generator(p2_z, cond)
p2_fake = tf.clip_by_value(p2_z, 1e-8, 1 - 1e-8)
# discriminator
D_c1_fake = self._discriminator(c1_fake, cond, reuse=False)
D_p1_real = self._discriminator(p1_real, cond, reuse=True)
D_p2_fake = self._discriminator(p2_fake, cond, reuse=True)
D_c2_real = self._discriminator(c2_real, cond, reuse=True)
# loss for encoder
R_loss, _ = self._loss_metric(feat_c1, feat_p2, label)
# loss for genertor
E1_loss = self._loss_vae(p1_real, c1_fake, c1_mu, c1_sigma)
E2_loss = self._loss_vae(c2_real, p2_fake, p2_mu, p2_sigma)
E_loss = E1_loss + E2_loss
# loss for discriminator
D1_loss, G1_loss = self._loss_gan(D_c1_fake, D_p1_real)
D2_loss, G2_loss = self._loss_gan(D_p2_fake, D_c2_real)
D_loss = D1_loss + D2_loss
G_loss = G1_loss + G2_loss
loss = E_loss + D_loss + G_loss + R_loss
# # allocate two optimizer
global_step = tf.train.create_global_step()
var_e = variables.select_vars('encoder')
var_g = variables.select_vars('generator')
var_d = variables.select_vars('discriminator')
op1 = updater.default(self.config, loss, global_step, var_e, 0)
op2 = updater.default(self.config, loss, None, var_g, 1)
op3 = updater.default(self.config, loss, None, var_d, 0)
train_op = tf.group(op1, op2, op3)
# update at the same time
saver = tf.train.Saver(var_list=variables.all())
# hooks
self.add_hook(self.snapshot.init())
self.add_hook(self.summary.init())
self.add_hook(
context.Running_Hook(config=self.config.log,
step=global_step,
keys=['E', 'D', 'G', 'R'],
values=[E_loss, D_loss, G_loss, R_loss],
func_test=self.test,
func_val=None))
with context.DefaultSession(self.hooks) as sess:
self.snapshot.restore(sess, saver)
while not sess.should_stop():
sess.run(train_op)