def decoder(self, net, reuse=None, training=True):
"""Builds a decoder on top of net.
Args:
net: Input to the decoder (output of encoder)
reuse: Whether to reuse already defined variables
training: Whether in train or eval mode.
Returns:
Decoded image with 3 channels.
"""
f_dims = DEFAULT_FILTER_DIMS
with tf.variable_scope('decoder', reuse=reuse):
with slim.arg_scope(
toon_net_argscope(padding='SAME', training=training)):
for l in range(0, self.num_layers - 1):
net = up_conv2d(net,
num_outputs=f_dims[self.num_layers - l -
2],
scope='deconv_{}'.format(l))
net = up_conv2d(net,
num_outputs=32,
scope='deconv_{}'.format(self.num_layers))
net = slim.conv2d(net,
num_outputs=3,
scope='deconv_{}'.format(self.num_layers +
1),
stride=1,
activation_fn=tf.nn.tanh,
normalizer_fn=None)
return net
def generator(self, net, drop_mask, reuse=None, training=True):
"""Builds a generator with the given inputs. Noise is induced in all convolutional layers.
Args:
net: Input to the generator (i.e. cartooned image and/or edge-map)
reuse: Whether to reuse already defined variables
training: Whether in train or eval mode.
Returns:
Encoding of the input.
"""
f_dims = DEFAULT_FILTER_DIMS
res_dim = DEFAULT_FILTER_DIMS[self.num_layers - 1]
with tf.variable_scope('generator', reuse=reuse):
with slim.arg_scope(toon_net_argscope(padding='SAME', training=training)):
net_in = net
for i in range(self.num_res_layers):
net = res_block_bottleneck(net, res_dim, res_dim / 4, noise_channels=32, scope='res_{}'.format(i))
net = net_in + (1.0-drop_mask)*net
for l in range(0, self.num_layers - 1):
net = up_conv2d(net, num_outputs=f_dims[self.num_layers - l - 2], scope='deconv_{}'.format(l))
net = tf.image.resize_images(net, (self.im_shape[0], self.im_shape[1]),
tf.image.ResizeMethod.NEAREST_NEIGHBOR)
net = slim.conv2d(net, num_outputs=32, scope='deconv_{}'.format(self.num_layers), stride=1)
net = slim.conv2d(net, num_outputs=3, scope='deconv_{}'.format(self.num_layers + 1), stride=1,
activation_fn=tf.nn.tanh, normalizer_fn=None)
return net
