def join_block(self, x, y, name):
with tf.variable_scope(name):
x_dim = x.get_shape().as_list()[1:3]
scaled = tf.image.resize_images(
y,
x_dim[0],
x_dim[1],
method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
x_norm = get_norm(x)
y_norm = get_norm(scaled)
return tf.concat(3, [x_norm, y_norm], name="concat")
def conv_block(self, bottom, n, S, stride, name, relu=True):
""" conv -> bias -> (norm -> relu | nop) -> out """
with tf.variable_scope(name):
if n > 1 and self.is_training:
pad = n // 2
bottom = tf.pad(bottom,
[[0, 0], [pad, pad], [pad, pad], [0, 0]],
"REFLECT")
C = bottom.get_shape().as_list()[-1]
filt = self.get_conv_filter(n, C, S, name)
if self.is_training:
conv = tf.nn.conv2d(bottom,
filt, [1, stride, stride, 1],
padding='VALID')
else:
conv = tf.nn.conv2d(bottom,
filt, [1, stride, stride, 1],
padding='SAME')
conv_biases = self.get_bias(name, S)
bias = tf.nn.bias_add(conv, conv_biases)
if relu:
norm = get_norm(bias)
return tf.nn.relu(norm)
else:
return bias
def deconv_block(self, bottom, n, S, stride, name):
""" AKA convolutional transpose """
with tf.variable_scope(name):
b, h, w, C = bottom.get_shape().as_list()
filt = self.get_conv_filter(n, S, C, name)
deconv = tf.nn.conv2d_transpose(bottom, filt, [b, h * stride, w * stride, S], [1, stride, stride, 1], padding='SAME')
biases = self.get_bias(name, S)
bias = tf.nn.bias_add(deconv, biases)
norm = get_norm(bias)
return tf.nn.relu(norm)
def get_residual(self, bottom, name, S=3):
with tf.variable_scope(name):
_, _, _, C = bottom.get_shape().as_list()
conv = self.conv_block(bottom, S, C, 1, "%s-a" % name)
return bottom + get_norm(
self.conv_block(conv, S, C, 1, "%s-b" % name, relu=False))