def model_r(self, g):
with tf.variable_scope('r'):
r_b1_alpha = tf.contrib.layers.conv2d(g,
32,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=tf.sigmoid)
r_b1_beta = tf.contrib.layers.conv2d(g,
32,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
r_b1_con = tf.concat([self.images_mos, self.images_r], 3)
r_b1_conv_low1 = tf.contrib.layers.conv2d(r_b1_con,
32,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
r_b1_transform = r_b1_alpha * r_b1_conv_low1 + r_b1_beta
r_b1_conv_low2 = tf.contrib.layers.conv2d(r_b1_transform,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME')
r_b1_conv1_1 = tf.contrib.layers.conv2d(r_b1_conv_low2,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME')
r_b1_conv1_2 = tf.contrib.layers.conv2d(r_b1_conv1_1,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
rate=2,
biases_initializer=None)
r_b1_conv1_3 = tf.contrib.layers.conv2d(r_b1_conv1_2,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
rate=2,
biases_initializer=None)
r_b1_conv1_5 = tf.contrib.layers.conv2d(r_b1_conv1_3,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
r_b1_conv1_add = r_b1_conv_low2 - r_b1_conv1_5
r_b2_conv_high1 = tf.contrib.layers.conv2d(r_b1_conv1_add,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME')
r_b2_conv_high2 = tf.contrib.layers.conv2d(r_b2_conv_high1,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
r_b2_conv_add = r_b2_conv_high2 + r_b1_conv1_add
r_b2_residual = tf.contrib.layers.conv2d(r_b2_conv_add,
1,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
r_out = r_b2_residual + self.images_r
r_lp = laplacian_loss(r_b1_conv1_add)
self.vars_r = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
'r')
return r_lp, r_out
def model_b(self, g):
with tf.variable_scope('b'):
b_b1_alpha = tf.contrib.layers.conv2d(g,
32,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=tf.sigmoid)
b_b1_beta = tf.contrib.layers.conv2d(g,
32,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
b_b1_con = tf.concat([self.images_mos, self.images_b], 3)
b_b1_conv_low1 = tf.contrib.layers.conv2d(b_b1_con,
32,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
b_b1_transform = b_b1_alpha * b_b1_conv_low1 + b_b1_beta
b_b1_conv_low2 = tf.contrib.layers.conv2d(b_b1_transform,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME')
b_b1_conv1_1 = tf.contrib.layers.conv2d(b_b1_conv_low2,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME')
b_b1_conv1_2 = tf.contrib.layers.conv2d(b_b1_conv1_1,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
rate=2,
biases_initializer=None)
b_b1_conv1_3 = tf.contrib.layers.conv2d(b_b1_conv1_2,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
rate=2,
biases_initializer=None)
b_b1_conv1_5 = tf.contrib.layers.conv2d(b_b1_conv1_3,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
b_b1_conv1_add = b_b1_conv_low2 - b_b1_conv1_5
b_b2_conv_high1 = tf.contrib.layers.conv2d(b_b1_conv1_add,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME')
b_b2_conv_high2 = tf.contrib.layers.conv2d(b_b2_conv_high1,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
b_b2_conv_add = b_b2_conv_high2 + b_b1_conv1_add
b_b2_residual = tf.contrib.layers.conv2d(b_b2_conv_add,
1,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
b_out = b_b2_residual + self.images_b
b_lp = laplacian_loss(b_b1_conv1_add)
self.vars_b = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
'b')
return b_lp, b_out
def model_g(self):
with tf.variable_scope('g'):
g_conv_low1 = tf.contrib.layers.conv2d(self.images_mos,
32,
kernel_size=(3, 3),
stride=1,
padding='SAME')
g_conv_low2 = tf.contrib.layers.conv2d(g_conv_low1,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
g_b1_conv1_1 = tf.contrib.layers.conv2d(g_conv_low2,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME')
g_b1_conv1_2 = tf.contrib.layers.conv2d(g_b1_conv1_1,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
rate=2,
biases_initializer=None)
g_b1_conv1_3 = tf.contrib.layers.conv2d(g_b1_conv1_2,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
rate=2,
biases_initializer=None)
g_b1_conv1_4 = tf.contrib.layers.conv2d(g_b1_conv1_3,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
g_b1_conv1_add = g_conv_low2 - 0.1 * g_b1_conv1_4
g_b1_conv_5 = tf.contrib.layers.conv2d(g_b1_conv1_add,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME')
g_b1_conv_6 = tf.contrib.layers.conv2d(g_b1_conv_5,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
g_b1_conv_add2 = g_b1_conv_6 + 0.1 * g_b1_conv1_add
g_b2_conv1_1 = tf.contrib.layers.conv2d(g_b1_conv_add2,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME')
g_b2_conv1_2 = tf.contrib.layers.conv2d(g_b2_conv1_1,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
rate=2,
biases_initializer=None)
g_b2_conv1_3 = tf.contrib.layers.conv2d(g_b2_conv1_2,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
rate=2,
biases_initializer=None)
g_b2_conv1_4 = tf.contrib.layers.conv2d(g_b2_conv1_3,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
g_b2_conv1_add = g_b1_conv_add2 - 0.1 * g_b2_conv1_4
g_b2_conv_5 = tf.contrib.layers.conv2d(g_b2_conv1_add,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME')
g_b2_conv_6 = tf.contrib.layers.conv2d(g_b2_conv_5,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
g_b2_conv_add2 = g_b2_conv_6 + 0.1 * g_b2_conv1_add
g_b3_conv1_1 = tf.contrib.layers.conv2d(g_b2_conv_add2,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME')
g_b3_conv1_2 = tf.contrib.layers.conv2d(g_b3_conv1_1,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
rate=2,
biases_initializer=None)
g_b3_conv1_3 = tf.contrib.layers.conv2d(g_b3_conv1_2,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
rate=2,
biases_initializer=None)
g_b3_conv1_4 = tf.contrib.layers.conv2d(g_b3_conv1_3,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
g_b3_conv1_add = g_b2_conv_add2 - 0.1 * g_b3_conv1_4
g_b3_conv_5 = tf.contrib.layers.conv2d(g_b3_conv1_add,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME')
g_b3_conv_6 = tf.contrib.layers.conv2d(g_b3_conv_5,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
g_b3_conv_add2 = g_b3_conv_6 + 0.1 * g_b3_conv1_add
g_b3_total = tf.concat(
[g_b1_conv_add2, g_b2_conv_add2, g_b3_conv_add2], 3)
g_b3_red = tf.contrib.layers.conv2d(g_b3_total,
64,
kernel_size=(1, 1),
stride=1,
padding='SAME')
g_conv_high1 = tf.contrib.layers.conv2d(g_b3_red,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME')
g_conv_high2 = tf.contrib.layers.conv2d(g_conv_high1,
64,
kernel_size=(3, 3),
stride=1,
padding='SAME')
g_residual = tf.contrib.layers.conv2d(g_conv_high2,
1,
kernel_size=(3, 3),
stride=1,
padding='SAME',
activation_fn=None)
g_out = g_residual + self.images_mos
g_lp = (laplacian_loss(g_b1_conv1_add) +
laplacian_loss(g_b2_conv1_add) +
laplacian_loss(g_b3_conv1_add)) / 3
self.vars_g = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
'g')
return g_lp, g_out