# Section 2
# Downsampling convolution
layer_name = "s2_conv1"
with tf.name_scope(layer_name):
W = utils.weight_variable([k_size, k_size, ch[0], ch[1]])
b = utils.bias_variable([ch[1]])
conv = utils.conv2d(s1_conv1, W, b, 2)
tanh = tf.nn.tanh(conv)
s2_conv1 = tf.nn.dropout(tanh, keep_prob)
# Regular bottleneck
layer_name = "s2_bn1"
with tf.name_scope(layer_name):
conv = utils.bottleneck(s2_conv1, [k_size, k_size])
tanh = tf.nn.tanh(conv)
s2_bn1 = tf.nn.dropout(tanh, keep_prob)
##############################
# Section 3
# Downsampling convolution
layer_name = "s3_conv1"
with tf.name_scope(layer_name):
W = utils.weight_variable([k_size, k_size, ch[1], ch[2]])
b = utils.bias_variable([ch[2]])
conv = utils.conv2d(s2_bn1, W, b, 2)
tanh = tf.nn.tanh(conv)
s3_conv1 = tf.nn.dropout(tanh, keep_prob)
