# Asymmetric convolution (5x1)
layer_name = "s2_conv1_2"
with tf.name_scope(layer_name):
W = utils.weight_variable([5, 1, ch[1], ch[1]])
b = utils.bias_variable([ch[1]])
conv = utils.conv2d(s2_conv1_1, W, b, 1)
tanh = tf.nn.tanh(conv)
s2_conv1_2 = tf.nn.dropout(tanh, keep_prob)
# Dilated convolution (rate 2)
layer_name = "s2_conv2_1"
with tf.name_scope(layer_name):
W = utils.weight_variable([1, 5, ch[1], ch[1]])
b = utils.bias_variable([ch[1]])
conv = utils.conv2d_dilated(s2_conv1_2, W, b, 2)
tanh = tf.nn.tanh(conv)
s2_conv2_1 = tf.nn.dropout(tanh, keep_prob)
# Asymmetric convolution (1x5)
layer_name = "s2_conv3_1"
with tf.name_scope(layer_name):
W = utils.weight_variable([1, 5, ch[0], ch[1]])
b = utils.bias_variable([ch[1]])
conv = utils.conv2d(s2_conv2_1, W, b, 1)
tanh = tf.nn.tanh(conv)
s2_conv3_1 = tf.nn.dropout(tanh, keep_prob)
# Asymmetric convolution (5x1)
