def build_G(self, name, image=None, style_image=None, reuse=False):
with tf.variable_scope('Generator_' + name, reuse=reuse):
if image is None:
self.x_input = tf.placeholder(
tf.float32, [None, self.input_size, self.input_size, 3])
self.y_input = tf.placeholder(
tf.float32, [None, self.input_size, self.input_size, 3])
x = tf.concat(self.x_input, self.y_input, 2)
elif image is not None and style_image is not None:
x = tf.concat([image, style_image], 3)
# filter_size:9, nb_filter: 3->32, step:1
x = ops.conv_layer('first', x, 9, 6, 32, 1)
# filter_size:3, nb_filter: 32->64, step:1
x = ops.conv_layer('dsample_1', x, 3, 32, 64, 2)
# filter_size:3, nb_filter:64->128, step:1
x = ops.conv_layer('dsample_2', x, 3, 64, 128, 2)
for i in range(5): # five residual blocks
name = ('res_block_{}'.format(i))
# filter_size:3, nb_filter:128->128, step:1
x = ops.residual(name, x, 128, 128, 1)
x = ops.upsample_nearest_neighbor('usample_1', x, 3, 128, 64, 2)
x = ops.upsample_nearest_neighbor('usample_2', x, 3, 64, 32, 2)
x = ops.conv_layer('last', x, 9, 32, 3, 1, activation=False)
out = tf.nn.tanh(x) * 150 + 255. / 2
return out
def _resnet(self, input):
with tf.variable_scope(self.name, reuse=self._reuse):
num_filters = [128, 256, 512, 512]
if self._image_size == 256:
num_filters.append(512)
E = input
E = ops.conv_block(E,
64,
'C{}_{}'.format(64, 0),
4,
2,
self._is_train,
self._reuse,
norm=None,
activation='leaky',
bias=True)
for i, n in enumerate(num_filters):
E = ops.residual(E,
n,
'res{}_{}'.format(n, i + 1),
self._is_train,
self._reuse,
norm=self._norm,
bias=True)
E = tf.nn.avg_pool(E, [1, 2, 2, 1], [1, 2, 2, 1], 'SAME')
E = tf.nn.relu(E)
E = tf.nn.avg_pool(E, [1, 8, 8, 1], [1, 8, 8, 1], 'SAME')
E = ops.flatten(E)
mu = ops.mlp(E,
self._latent_dim,
'FC8_mu',
self._is_train,
self._reuse,
norm=None,
activation=None)
log_sigma = ops.mlp(E,
self._latent_dim,
'FC8_sigma',
self._is_train,
self._reuse,
norm=None,
activation=None)
z = mu + tf.random_normal(
shape=tf.shape(self._latent_dim)) * tf.exp(log_sigma)
self._reuse = True
self.var_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
self.name)
return z, mu, log_sigma
def build_G_AdaIN(self,
name,
input_size=256,
image=None,
style_image=None):
with tf.variable_scope('Generator_encoder_' + name):
if image is None:
x_input = tf.placeholder(
tf.float32, [None, self.input_size, self.input_size, 3])
else:
x_input = image
# filter_size:9, nb_filter: 3->32, step:1
x = ops.conv_layer('first', x_input, 9, 3, 32, 1)
# filter_size:3, nb_filter: 32->64, step:1
x = ops.conv_layer('dsample_1', x, 3, 32, 64, 2)
# filter_size:3, nb_filter:64->128, step:1
x = ops.conv_layer('dsample_2', x, 3, 64, 128, 2)
with tf.variable_scope('Generator_encoder_c_' + name, reuse=True):
if image is None:
y_input = tf.placeholder(
tf.float32, [None, self.input_size, self.input_size, 3])
else:
y_input = style_image
# filter_size:9, nb_filter: 3->32, step:1
y = ops.conv_layer('first', y_input, 9, 3, 32, 1)
# filter_size:3, nb_filter: 32->64, step:1
y = ops.conv_layer('dsample_1', x, 3, 32, 64, 2)
# filter_size:3, nb_filter:64->128, step:1
y = ops.conv_layer('dsample_2', x, 3, 64, 128, 2)
with tf.variable_scope('Generator_decoder_' + name):
for i in range(5): # five residual blocks
name = ('res_block_{}'.format(i))
# filter_size:3, nb_filter:128->128, step:1
x = ops.residual(name, x, 128, 128, 1, norm='AdaIN', y=y)
x = ops.upsample_nearest_neighbor('usample_1', x, 3, 128, 64, 2)
x = ops.upsample_nearest_neighbor('usample_2', x, 3, 64, 32, 2)
x = ops.conv_layer('last', x, 9, 32, 3, 1, activation=False)
out = tf.nn.tanh(x) * 150 + 255. / 2
return out
def __call__(self, input):
with tf.variable_scope(self.name, reuse=self._reuse):
G = ops.conv_block(input,
32,
'c7s1-32',
7,
1,
self._is_train,
self._reuse,
self._norm,
self._activation,
pad='REFLECT')
G = ops.conv_block(G, 64, 'd64', 3, 2, self._is_train, self._reuse,
self._norm, self._activation)
G = ops.conv_block(G, 128, 'd128', 3, 2, self._is_train,
self._reuse, self._norm, self._activation)
for i in range(self._num_res_block):
G = ops.residual(G, 128, 'R128_{}'.format(i), self._is_train,
self._reuse, self._norm)
G = ops.deconv_block(G, 64, 'u64', 3, 2, self._is_train,
self._reuse, self._norm, self._activation)
G = ops.deconv_block(G, 32, 'u32', 3, 2, self._is_train,
self._reuse, self._norm, self._activation)
G = ops.conv_block(G,
3,
'c7s1-3',
7,
1,
self._is_train,
self._reuse,
norm=None,
activation='tanh',
pad='REFLECT')
self._reuse = True
self.var_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
self.name)
return G