b = utils.bias_variable([ch[3]])
conv = utils.conv2d(s4_conv3, W, b, 1)
tanh = tf.nn.tanh(conv)
s4_conv4 = tf.nn.dropout(tanh, keep_prob)
##############################
# Section 5
# Upsampling convolution (with skip connection from section 2)
layer_name = "s5_conv1"
with tf.name_scope(layer_name):
W = utils.weight_variable([k_size, k_size, ch[4], ch[3]])
b = utils.bias_variable([ch[4]])
conv = utils.conv2d_transpose(s4_conv4, W, b, tf.shape(s2_conv1), 2)
tanh = tf.nn.tanh(conv)
s5_conv1 = tf.nn.dropout(tanh, keep_prob)
# Asymmetric convolution
layer_name = "s5_conv2"
with tf.name_scope(layer_name):
W = utils.weight_variable([k_size, 1, ch[4], ch[4]])
b = utils.bias_variable([ch[4]])
conv = utils.conv2d(s5_conv1, W, b, 1)
tanh = tf.nn.tanh(conv)
s5_conv2 = tf.nn.dropout(tanh, keep_prob)
# Asymmetric convolution
layer_name = "s5_conv3"
b = utils.bias_variable([ch[2]])
conv = utils.conv2d(s3_conv1, W, b, 1)
tanh = tf.nn.tanh(conv)
s3_conv2 = tf.nn.dropout(tanh, keep_prob)
##############################
# Section 4
# Upsampling convolution (with skip connection from section 2)
layer_name = "s4_conv1"
with tf.name_scope(layer_name):
W = utils.weight_variable([k_size, k_size, ch[3], ch[2]])
b = utils.bias_variable([ch[3]])
conv = utils.conv2d_transpose(s3_conv2, W, b, tf.shape(s1_conv1), 2)
tanh = tf.nn.tanh(conv)
s4_conv1 = tf.nn.dropout(tanh, keep_prob)
# Convolution
layer_name = "s4_conv2"
with tf.name_scope(layer_name):
W = utils.weight_variable([k_size, k_size, ch[3], ch[3]])
b = utils.bias_variable([ch[3]])
conv = utils.conv2d(s4_conv1, W, b, 1)
tanh = tf.nn.tanh(conv)
s4_conv2 = tf.nn.dropout(tanh, keep_prob)
##############################
# Section 5
