def build(self):
if self.is_built:
return
self.is_built = True
gen_factory = self.create_generator()
dis_factory = self.create_discriminator()
smoothing = 0.9 if self.options.label_smoothing else 1
seed = seed = self.options.seed
kernel = self.options.kernel_size
self.input_rgb = tf.placeholder(tf.float32,
shape=(None, None, None, 3),
name='input_rgb')
self.input_rgb_prev = tf.placeholder(tf.float32,
shape=(None, None, None, 3),
name='input_rgb_prev')
self.input_gray = tf.image.rgb_to_grayscale(self.input_rgb)
self.input_color = preprocess(self.input_rgb,
colorspace_in=COLORSPACE_RGB,
colorspace_out=self.options.color_space)
self.input_color_prev = preprocess(
self.input_rgb_prev,
colorspace_in=COLORSPACE_RGB,
colorspace_out=self.options.color_space)
gen = gen_factory.create(
tf.concat([self.input_gray, self.input_color_prev], 3), kernel,
seed)
dis_real = dis_factory.create(
tf.concat([self.input_color, self.input_color_prev], 3), kernel,
seed)
dis_fake = dis_factory.create(tf.concat([gen, self.input_color_prev],
3),
kernel,
seed,
reuse_variables=True)
gen_ce = tf.nn.sigmoid_cross_entropy_with_logits(
logits=dis_fake, labels=tf.ones_like(dis_fake))
dis_real_ce = tf.nn.sigmoid_cross_entropy_with_logits(
logits=dis_real, labels=tf.ones_like(dis_real) * smoothing)
dis_fake_ce = tf.nn.sigmoid_cross_entropy_with_logits(
logits=dis_fake, labels=tf.zeros_like(dis_fake))
self.dis_loss_real = tf.reduce_mean(dis_real_ce)
self.dis_loss_fake = tf.reduce_mean(dis_fake_ce)
self.dis_loss = tf.reduce_mean(dis_real_ce + dis_fake_ce)
self.gen_loss_gan = tf.reduce_mean(gen_ce)
self.gen_loss_l1 = tf.reduce_mean(
tf.abs(self.input_color - gen)) * self.options.l1_weight
#self.gen_loss_l1 = tf.reduce_mean(tf.abs(self.input_gray - tf.image.rgb_to_grayscale(gen))) * self.options.l1_weight
self.gen_loss = self.gen_loss_l1 #self.gen_loss_gan + self.gen_loss_l1
self.sampler = gen_factory.create(tf.concat(
[self.input_gray, self.input_color_prev], 3),
kernel,
seed,
reuse_variables=True)
self.accuracy = pixelwise_accuracy(self.input_color, gen,
self.options.color_space,
self.options.acc_thresh)
self.learning_rate = tf.constant(self.options.lr)
# learning rate decay
if self.options.lr_decay_rate > 0:
self.learning_rate = tf.maximum(
1e-8,
tf.train.exponential_decay(
learning_rate=self.options.lr,
global_step=self.global_step,
decay_steps=self.options.lr_decay_steps,
decay_rate=self.options.lr_decay_rate))
# generator optimizaer
self.gen_train = tf.train.AdamOptimizer(
learning_rate=self.learning_rate,
beta1=self.options.beta1).minimize(self.gen_loss,
var_list=gen_factory.var_list)
# discriminator optimizaer
self.dis_train = tf.train.AdamOptimizer(
learning_rate=self.learning_rate,
beta1=self.options.beta1).minimize(self.dis_loss,
var_list=dis_factory.var_list,
global_step=self.global_step)
self.saver = tf.train.Saver()
def build(self):
if self.is_built:
return
self.is_built = True
gen_factory = self.create_generator()
dis_factory = self.create_discriminator()
smoothing = 0.9 if self.options.label_smoothing else 1
seed = self.options.seed
kernel = 4
# model input placeholder: RGB imaege
self.input_rgb = tf.placeholder(tf.float32,
shape=(None, None, None, 3),
name='input_rgb')
# model input after preprocessing: LAB image
self.input_color = preprocess(self.input_rgb,
colorspace_in=COLORSPACE_RGB,
colorspace_out=self.options.color_space)
# test mode: model input is a graycale placeholder
if self.options.mode == 1:
self.input_gray = tf.placeholder(tf.float32,
shape=(None, None, None, 1),
name='input_gray')
# train/turing-test we extract grayscale image from color image
else:
self.input_gray = tf.placeholder(tf.float32,
shape=(None, None, None, 1),
name='input_gray')
gen = gen_factory.create(self.input_gray, kernel, seed)
dis_real = dis_factory.create(
tf.concat([self.input_gray, self.input_color], 3), kernel, seed)
dis_fake = dis_factory.create(tf.concat([self.input_gray, gen], 3),
kernel,
seed,
reuse_variables=True)
gen_ce = tf.nn.sigmoid_cross_entropy_with_logits(
logits=dis_fake, labels=tf.ones_like(dis_fake))
dis_real_ce = tf.nn.sigmoid_cross_entropy_with_logits(
logits=dis_real, labels=tf.ones_like(dis_real) * smoothing)
dis_fake_ce = tf.nn.sigmoid_cross_entropy_with_logits(
logits=dis_fake, labels=tf.zeros_like(dis_fake))
self.dis_loss_real = tf.reduce_mean(dis_real_ce)
self.dis_loss_fake = tf.reduce_mean(dis_fake_ce)
self.dis_loss = tf.reduce_mean(dis_real_ce + dis_fake_ce)
self.gen_loss_gan = tf.reduce_mean(gen_ce)
self.gen_loss_l1 = tf.reduce_mean(
tf.abs(self.input_color - gen)) * self.options.l1_weight
self.gen_loss = self.gen_loss_gan + self.gen_loss_l1
self.sampler = tf.identity(gen_factory.create(self.input_gray,
kernel,
seed,
reuse_variables=True),
name='output')
self.accuracy = pixelwise_accuracy(self.input_color, gen,
self.options.color_space,
self.options.acc_thresh)
self.learning_rate = tf.constant(self.options.lr)
# learning rate decay
if self.options.lr_decay and self.options.lr_decay_rate > 0:
self.learning_rate = tf.maximum(
1e-6,
tf.train.exponential_decay(
learning_rate=self.options.lr,
global_step=self.global_step,
decay_steps=self.options.lr_decay_steps,
decay_rate=self.options.lr_decay_rate))
# generator optimizaer
# 学习者的优化器
self.gen_train = tf.train.AdamOptimizer(
learning_rate=self.learning_rate,
beta1=self.options.beta1).minimize(self.gen_loss,
var_list=gen_factory.var_list)
# discriminator optimizaer
# 打分者的优化器
self.dis_train = tf.train.AdamOptimizer(
learning_rate=self.learning_rate / 10,
beta1=self.options.beta1).minimize(self.dis_loss,
var_list=dis_factory.var_list,
global_step=self.global_step)
self.saver = tf.train.Saver()