def build(self):
"""Build the model."""
self.global_step = tf.Variable(0, trainable=False, name='global_step')
# Create placeholders
self.z = tf.placeholder(
tf.float32,
(self.config['batch_size'], self.config['net_g']['z_dim']), 'z')
data_shape = (self.config['batch_size'], self.config['num_bar'],
self.config['num_timestep'], self.config['num_pitch'],
self.config['num_track'])
# print('[*]', data_shape)
self.x = tf.placeholder(tf.bool, data_shape, 'x')
self.x_ = tf.cast(self.x, tf.float32, 'x_')
# Components
self.G = Generator(self.z, self.config, name='G')
self.test_round = self.G.tensor_out > 0.5
self.test_bernoulli = self.G.tensor_out > tf.random_uniform(data_shape)
self.D_fake = Discriminator(self.G.tensor_out, self.config, name='D')
self.D_real = Discriminator(self.x_, self.config, name='D', reuse=True)
self.components = (self.G, self.D_fake)
# Losses
self.g_loss, self.d_loss = self.get_adversarial_loss()
# Optimizers
with tf.variable_scope('Optimizer'):
self.g_optimizer = self.get_optimizer()
self.g_step = self.g_optimizer.minimize(self.g_loss,
self.global_step,
self.G.vars)
self.d_optimizer = self.get_optimizer()
self.d_step = self.d_optimizer.minimize(self.d_loss,
self.global_step,
self.D_fake.vars)
# Apply weight clipping
if self.config['gan']['type'] == 'wgan':
with tf.control_dependencies([self.d_step]):
self.d_step = tf.group(*(tf.assign(
var,
tf.clip_by_value(var,
-self.config['gan']['clip_value'],
self.config['gan']['clip_value']))
for var in self.D_fake.vars))
# Metrics
self.metrics = Metrics(self.config)
# Saver
self.saver = tf.train.Saver()
# Print and save model information
self.print_statistics()
self.save_statistics()
self.print_summary()
self.save_summary()
def build(self):
"""Build the model."""
# Create global step variable
self.global_step = tf.Variable(0, trainable=False, name='global_step')
# Get tensors from the pretrained model
self.z = tf.placeholder(
tf.float32,
(self.config['batch_size'], self.config['net_g']['z_dim']), 'z')
self.x = tf.placeholder(tf.bool, data_shape[self.resolution], 'x')
self.x_ = tf.cast(self.x, tf.float32, 'x_')
self.G = Generator(self.z, self.resolution, self.config, name='G')
self.test_round = self.G.tensor_out > 0.5
self.test_bernoulli = self.G.tensor_out > tf.random_uniform(
data_shape[self.resolution])
self.slope_tensor = tf.Variable(1.0)
self.R = Refiner(self.pretrained.G.tensor_out,
self.config,
slope_tensor=self.slope_tensor,
name='R')
self.D_fake = Discriminator(self.R.tensor_out, self.config, name='D')
self.D_real = Discriminator(self.x_, self.config, name='D', reuse=True)
self.components = (self.G, self.R, self.D_fake)
#self.R = Refiner()
# Slope tensor for applying slope annealing trick to stochastic neurons
# self.slope_tensor = tf.Variable(1.0)
# self.D_real = self.pretrained.D_real
# with tf.variable_scope(self.pretrained.scope, reuse=True):
# self.D_fake = Discriminator(self.G.tensor_out, self.config,
# name='D')
# self.components = (self.pretrained.G, self.G, self.D_fake)
# Losses
self.g_loss, self.d_loss = self.get_adversarial_loss()
# Optimizers
with tf.variable_scope('Optimizer'):
self.g_optimizer = self.get_optimizer()
if self.config['joint_training']:
self.g_step = self.g_optimizer.minimize(
self.g_loss, self.global_step, (self.R.vars + self.G.vars))
else:
self.g_step = self.g_optimizer.minimize(
self.g_loss, self.global_step, self.R.vars)
self.d_optimizer = self.get_optimizer()
self.d_step = self.d_optimizer.minimize(self.d_loss,
self.global_step,
self.D_fake.vars)
# Apply weight clipping
if self.config['gan']['type'] == 'wgan':
with tf.control_dependencies([self.d_step]):
self.d_step = tf.group(*(tf.assign(
var,
tf.clip_by_value(var,
-self.config['gan']['clip_value'],
self.config['gan']['clip_value']))
for var in self.D_fake.vars))
# Optimizers
with tf.variable_scope('Optimizer'):
self.g_optimizer = self.get_optimizer()
self.g_step = self.g_optimizer.minimize(self.g_loss,
self.global_step,
self.G.vars)
self.d_optimizer = self.get_optimizer()
self.d_step = self.d_optimizer.minimize(self.d_loss,
self.global_step,
self.D_fake.vars)
# Apply weight clipping
if self.config['gan']['type'] == 'wgan':
with tf.control_dependencies([self.d_step]):
self.d_step = tf.group(*(tf.assign(
var,
tf.clip_by_value(var,
-self.config['gan']['clip_value'],
self.config['gan']['clip_value']))
for var in self.D_fake.vars))
# Metrics
self.metrics = Metrics(self.config)
# Saver
self.saver = tf.train.Saver()
# Print and save model information
self.print_statistics()
self.save_statistics()
self.print_summary()
self.save_summary()
def build(self):
"""Build the model."""
# Create global step variable
self.global_step = tf.Variable(0, trainable=False, name='global_step')
# Get tensors from the pretrained model
self.z = self.pretrained.z
self.x = self.pretrained.x
self.x_ = self.pretrained.x_
# Slope tensor for applying slope annealing trick to stochastic neurons
self.slope_tensor = tf.Variable(1.0)
# Components
self.G = Refiner(self.pretrained.G.tensor_out,
self.config,
slope_tensor=self.slope_tensor,
name='R')
self.D_real = self.pretrained.D_real
with tf.variable_scope(self.pretrained.scope, reuse=True):
self.D_fake = Discriminator(self.G.tensor_out,
self.config,
name='D')
self.components = (self.pretrained.G, self.G, self.D_fake)
# Losses
self.g_loss, self.d_loss = self.get_adversarial_loss(
self.pretrained.scope)
# Optimizers
with tf.variable_scope('Optimizer'):
self.g_optimizer = self.get_optimizer()
if self.config['joint_training']:
self.g_step = self.g_optimizer.minimize(
self.g_loss, self.global_step,
(self.G.vars + self.pretrained.G.vars))
else:
self.g_step = self.g_optimizer.minimize(
self.g_loss, self.global_step, self.G.vars)
self.d_optimizer = self.get_optimizer()
self.d_step = self.d_optimizer.minimize(self.d_loss,
self.global_step,
self.D_fake.vars)
# Apply weight clipping
if self.config['gan']['type'] == 'wgan':
with tf.control_dependencies([self.d_step]):
self.d_step = tf.group(*(tf.assign(
var,
tf.clip_by_value(var,
-self.config['gan']['clip_value'],
self.config['gan']['clip_value']))
for var in self.D_fake.vars))
# Metrics
self.metrics = Metrics(self.config)
# Saver
self.saver = tf.train.Saver()
# Print and save model information
self.print_statistics()
self.save_statistics()
self.print_summary()
self.save_summary()