def decoder(self, inputs, skips, nout):
"""VGG based image decoder.
Args:
inputs: image tensor with size BSxX
skips: skip connections from encoder
nout: number of output channels
Returns:
net: decoded image with size BSx64x64xNout
skips: skip connection after each layer
"""
vgg_layer = common_video.vgg_layer
net = inputs
# d1
net = tfl.conv2d_transpose(net,
512,
kernel_size=4,
padding="VALID",
name="d1_deconv",
activation=None)
net = tfl.batch_normalization(net,
training=self.is_training,
name="d1_bn")
net = tf.nn.leaky_relu(net)
net = common_layers.upscale(net, 2)
# d2
net = tf.concat([net, skips[-1]], axis=3)
net = tfcl.repeat(net, 2, vgg_layer, 512, scope="d2a")
net = tfcl.repeat(net, 1, vgg_layer, 256, scope="d2b")
net = common_layers.upscale(net, 2)
# d3
net = tf.concat([net, skips[-2]], axis=3)
net = tfcl.repeat(net, 2, vgg_layer, 256, scope="d3a")
net = tfcl.repeat(net, 1, vgg_layer, 128, scope="d3b")
net = common_layers.upscale(net, 2)
# d4
net = tf.concat([net, skips[-3]], axis=3)
net = tfcl.repeat(net, 1, vgg_layer, 128, scope="d4a")
net = tfcl.repeat(net, 1, vgg_layer, 64, scope="d4b")
net = common_layers.upscale(net, 2)
# d5
net = tf.concat([net, skips[-4]], axis=3)
net = tfcl.repeat(net, 1, vgg_layer, 64, scope="d5")
# if there are still skip connections left, we have more downscaling to do
for i, s in enumerate(skips[-5::-1]):
net = common_layers.upscale(net, 2)
net = tf.concat([net, s], axis=3)
net = tfcl.repeat(net, 1, vgg_layer, 64, scope="upscale%d" % i)
net = tfl.conv2d_transpose(net,
nout,
kernel_size=3,
padding="SAME",
name="d6_deconv",
activation=tf.sigmoid)
return net
def generator(self, z, reuse=False):
with tf.variable_scope("generator", reuse=reuse):
ch = self.args.g_filters
x = spectral_deconv2d(z, filters=ch, kernel_size=4, stride=1, is_training=self.is_training, padding='VALID',
use_bias=False, scope='deconv2d')
x = batch_norm(x, self.is_training, scope='batch_norm')
x = tf.nn.relu(x)
for i in range(self.layer_num // 2):
if self.args.up_sample:
x = upscale(x, f=2)
x = spectral_conv2d(x, filters=ch // 2, kernel_size=3, stride=1, is_training=self.is_training,
padding='SAME', scope='up_conv2d_' + str(i))
else:
x = spectral_deconv2d(x, filters=ch // 2, kernel_size=4, stride=2, is_training=self.is_training,
use_bias=False, scope='deconv2d_' + str(i))
x = batch_norm(x, self.is_training, scope='batch_norm_' + str(i))
x = tf.nn.relu(x)
ch = ch // 2
# Self Attention
x = attention(x, ch, is_training=self.is_training, scope="attention", reuse=reuse)
for i in range(self.layer_num // 2, self.layer_num):
if self.args.up_sample:
x = upscale(x, f=2)
x = spectral_conv2d(x, filters=ch // 2, kernel_size=3, stride=1, is_training=self.is_training,
padding='SAME', scope='up_conv2d_' + str(i))
else:
x = spectral_deconv2d(x, filters=ch // 2, kernel_size=4, stride=2, is_training=self.is_training,
use_bias=False, scope='deconv2d_' + str(i))
x = batch_norm(x, self.is_training, scope='batch_norm_' + str(i))
x = tf.nn.relu(x)
ch = ch // 2
if self.args.up_sample:
x = upscale(x, f=2)
x = spectral_conv2d(x, filters=self.args.img_size[2], kernel_size=3, stride=1, is_training=self.is_training,
padding='SAME', scope='G_conv_logit')
else:
x = spectral_deconv2d(x, filters=self.args.img_size[2], kernel_size=4, stride=2, is_training=self.is_training,
use_bias=False, scope='G_deconv_logit')
x = tf.nn.tanh(x)
return x
def decoder(self, inputs, skips, nout):
"""VGG based image decoder.
Args:
inputs: image tensor with size BSxX
skips: skip connections from encoder
nout: number of output channels
Returns:
net: decoded image with size BSx64x64xNout
skips: skip connection after each layer
"""
vgg_layer = self.vgg_layer
net = inputs
# d1
net = slim.conv2d_transpose(net, 512, kernel_size=4, padding="VALID",
scope="d1_deconv", activation_fn=None)
net = slim.batch_norm(net, scope="d1_bn")
net = tf.nn.leaky_relu(net)
net = common_layers.upscale(net, 2)
# d2
net = tf.concat([net, skips[3]], axis=3)
net = slim.repeat(net, 2, vgg_layer, 512, scope="d2a")
net = slim.repeat(net, 1, vgg_layer, 256, scope="d2b")
net = common_layers.upscale(net, 2)
# d3
net = tf.concat([net, skips[2]], axis=3)
net = slim.repeat(net, 2, vgg_layer, 256, scope="d3a")
net = slim.repeat(net, 1, vgg_layer, 128, scope="d3b")
net = common_layers.upscale(net, 2)
# d4
net = tf.concat([net, skips[1]], axis=3)
net = slim.repeat(net, 1, vgg_layer, 128, scope="d4a")
net = slim.repeat(net, 1, vgg_layer, 64, scope="d4b")
net = common_layers.upscale(net, 2)
# d5
net = tf.concat([net, skips[0]], axis=3)
net = slim.repeat(net, 1, vgg_layer, 64, scope="d5")
net = slim.conv2d_transpose(net, nout, kernel_size=3, padding="SAME",
scope="d6_deconv", activation_fn=tf.sigmoid)
return net
def decoder(self, inputs, skips, nout):
"""VGG based image decoder.
Args:
inputs: image tensor with size BSxX
skips: skip connections from encoder
nout: number of output channels
Returns:
net: decoded image with size BSx64x64xNout
skips: skip connection after each layer
"""
vgg_layer = self.vgg_layer
net = inputs
# d1
net = slim.conv2d_transpose(net, 512, kernel_size=4, padding="VALID",
scope="d1_deconv", activation_fn=None)
net = slim.batch_norm(net, scope="d1_bn")
net = tf.nn.leaky_relu(net)
net = common_layers.upscale(net, 2)
# d2
net = tf.concat([net, skips[3]], axis=3)
net = slim.repeat(net, 2, vgg_layer, 512, scope="d2a")
net = slim.repeat(net, 1, vgg_layer, 256, scope="d2b")
net = common_layers.upscale(net, 2)
# d3
net = tf.concat([net, skips[2]], axis=3)
net = slim.repeat(net, 2, vgg_layer, 256, scope="d3a")
net = slim.repeat(net, 1, vgg_layer, 128, scope="d3b")
net = common_layers.upscale(net, 2)
# d4
net = tf.concat([net, skips[1]], axis=3)
net = slim.repeat(net, 1, vgg_layer, 128, scope="d4a")
net = slim.repeat(net, 1, vgg_layer, 64, scope="d4b")
net = common_layers.upscale(net, 2)
# d5
net = tf.concat([net, skips[0]], axis=3)
net = slim.repeat(net, 1, vgg_layer, 64, scope="d5")
net = slim.conv2d_transpose(net, nout, kernel_size=3, padding="SAME",
scope="d6_deconv", activation_fn=tf.sigmoid)
return net
def decoder(self, inputs, nout, skips=None, has_batchnorm=True):
"""VGG based image decoder.
Args:
inputs: image tensor with size BSxX
nout: number of output channels
skips: optional skip connections from encoder
has_batchnorm: variable to use or not use batch normalization
Returns:
net: decoded image with size BSx64x64xNout
skips: skip connection after each layer
"""
vgg_layer = common_video.vgg_layer
net = inputs
# d1
net = tfl.conv2d_transpose(net,
512,
kernel_size=4,
padding="VALID",
name="d1_deconv",
activation=tf.nn.relu)
if has_batchnorm:
net = tfl.batch_normalization(net,
training=self.is_training,
name="d1_bn")
net = tf.nn.relu(net)
net = common_layers.upscale(net, 2)
# d2
if skips is not None:
net = tf.concat([net, skips[-1]], axis=3)
net = tfcl.repeat(net,
2,
vgg_layer,
512,
scope="d2a",
is_training=self.is_training,
activation=tf.nn.relu,
has_batchnorm=has_batchnorm)
net = tfcl.repeat(net,
1,
vgg_layer,
256,
scope="d2b",
is_training=self.is_training,
activation=tf.nn.relu,
has_batchnorm=has_batchnorm)
net = common_layers.upscale(net, 2)
# d3
if skips is not None:
net = tf.concat([net, skips[-2]], axis=3)
net = tfcl.repeat(net,
2,
vgg_layer,
256,
scope="d3a",
is_training=self.is_training,
activation=tf.nn.relu,
has_batchnorm=has_batchnorm)
net = tfcl.repeat(net,
1,
vgg_layer,
128,
scope="d3b",
is_training=self.is_training,
activation=tf.nn.relu,
has_batchnorm=has_batchnorm)
net = common_layers.upscale(net, 2)
# d4
if skips is not None:
net = tf.concat([net, skips[-3]], axis=3)
net = tfcl.repeat(net,
1,
vgg_layer,
128,
scope="d4a",
is_training=self.is_training,
activation=tf.nn.relu,
has_batchnorm=has_batchnorm)
net = tfcl.repeat(net,
1,
vgg_layer,
64,
scope="d4b",
is_training=self.is_training,
activation=tf.nn.relu,
has_batchnorm=has_batchnorm)
net = common_layers.upscale(net, 2)
# d5
if skips is not None:
net = tf.concat([net, skips[-4]], axis=3)
net = tfcl.repeat(net,
1,
vgg_layer,
64,
scope="d5",
is_training=self.is_training,
activation=tf.nn.relu,
has_batchnorm=has_batchnorm)
# if there are still skip connections left, we have more upscaling to do
if skips is not None:
for i, s in enumerate(skips[-5::-1]):
net = common_layers.upscale(net, 2)
net = tf.concat([net, s], axis=3)
net = tfcl.repeat(net,
1,
vgg_layer,
64,
scope="upscale%d" % i,
is_training=self.is_training,
activation=tf.nn.relu,
has_batchnorm=has_batchnorm)
net = tfl.conv2d_transpose(net,
nout,
kernel_size=3,
padding="SAME",
name="d6_deconv",
activation=None)
return net
def decoder(self, inputs, nout, skips=None, has_batchnorm=True):
"""VGG based image decoder.
Args:
inputs: image tensor with size BSxX
nout: number of output channels
skips: optional skip connections from encoder
has_batchnorm: variable to use or not use batch normalization
Returns:
net: decoded image with size BSx64x64xNout
skips: skip connection after each layer
"""
vgg_layer = common_video.vgg_layer
net = inputs
# d1
net = tfl.conv2d_transpose(net, 512, kernel_size=4, padding="VALID",
name="d1_deconv", activation=tf.nn.relu)
if has_batchnorm:
net = tfl.batch_normalization(
net, training=self.is_training, name="d1_bn")
net = tf.nn.relu(net)
net = common_layers.upscale(net, 2)
# d2
if skips is not None:
net = tf.concat([net, skips[-1]], axis=3)
net = tfcl.repeat(net, 2, vgg_layer, 512, scope="d2a",
is_training=self.is_training,
activation=tf.nn.relu, has_batchnorm=has_batchnorm)
net = tfcl.repeat(net, 1, vgg_layer, 256, scope="d2b",
is_training=self.is_training,
activation=tf.nn.relu, has_batchnorm=has_batchnorm)
net = common_layers.upscale(net, 2)
# d3
if skips is not None:
net = tf.concat([net, skips[-2]], axis=3)
net = tfcl.repeat(net, 2, vgg_layer, 256, scope="d3a",
is_training=self.is_training,
activation=tf.nn.relu, has_batchnorm=has_batchnorm)
net = tfcl.repeat(net, 1, vgg_layer, 128, scope="d3b",
is_training=self.is_training,
activation=tf.nn.relu, has_batchnorm=has_batchnorm)
net = common_layers.upscale(net, 2)
# d4
if skips is not None:
net = tf.concat([net, skips[-3]], axis=3)
net = tfcl.repeat(net, 1, vgg_layer, 128, scope="d4a",
is_training=self.is_training,
activation=tf.nn.relu, has_batchnorm=has_batchnorm)
net = tfcl.repeat(net, 1, vgg_layer, 64, scope="d4b",
is_training=self.is_training,
activation=tf.nn.relu, has_batchnorm=has_batchnorm)
net = common_layers.upscale(net, 2)
# d5
if skips is not None:
net = tf.concat([net, skips[-4]], axis=3)
net = tfcl.repeat(net, 1, vgg_layer, 64, scope="d5",
is_training=self.is_training,
activation=tf.nn.relu, has_batchnorm=has_batchnorm)
# if there are still skip connections left, we have more upscaling to do
if skips is not None:
for i, s in enumerate(skips[-5::-1]):
net = common_layers.upscale(net, 2)
net = tf.concat([net, s], axis=3)
net = tfcl.repeat(net, 1, vgg_layer, 64, scope="upscale%d" % i,
is_training=self.is_training,
activation=tf.nn.relu, has_batchnorm=has_batchnorm)
net = tfl.conv2d_transpose(net, nout, kernel_size=3, padding="SAME",
name="d6_deconv", activation=None)
return net