def __create_decoder(self, encoder_results, lastGlobalNet, output_channels):
layer_specs = [
(self.ngf * 8, 0.5), # decoder_8: [batch, 1, 1, ngf * 8] => [batch, 2, 2, ngf * 8]
(self.ngf * 8, 0.5), # decoder_7: [batch, 2, 2, ngf * 8 ] => [batch, 4, 4, ngf * 8]
(self.ngf * 8, 0.5), # decoder_6: [batch, 4, 4, ngf * 8 ] => [batch, 8, 8, ngf * 8] #Dropout was 0.5 until here
(self.ngf * 8, 0.0), # decoder_5: [batch, 8, 8, ngf * 8 ] => [batch, 16, 16, ngf * 8]
(self.ngf * 4, 0.0), # decoder_4: [batch, 16, 16, ngf * 8 ] => [batch, 32, 32, ngf * 4]
(self.ngf * 2, 0.0), # decoder_3: [batch, 32, 32, ngf * 4] => [batch, 64, 64, ngf * 2]
(self.ngf, 0.0), # decoder_2: [batch, 64, 64, ngf * 2] => [batch, 128, 128, ngf]
]
decoder_results = []
num_encoder_layers = len(encoder_results)
for decoder_layer, (out_channels, dropout) in enumerate(layer_specs):
skip_layer = num_encoder_layers - decoder_layer - 1
with tf.variable_scope("decoder_%d" % (skip_layer + 1)):
if decoder_layer == 0:
# first decoder layer doesn't have skip connections
# since it is directly connected to the skip_layer
input = encoder_results[-1]
else:
input = tf.concat([decoder_results[-1], encoder_results[skip_layer]], axis=3)
rectified = tfHelpers.lrelu(input, 0.2)
# [batch, in_height, in_width, in_channels] => [batch, in_height*2, in_width*2, out_channels]
output = tfHelpers.deconv(rectified, out_channels)
output, mean, variance = tfHelpers.instancenorm(output)
output, lastGlobalNet = self._addSecondaryNetBlock(output, mean, lastGlobalNet, out_channels, out_channels, num_encoder_layers + len(decoder_results))
if dropout > 0.0:
output = tf.nn.dropout(output, keep_prob=1 - dropout)
decoder_results.append(output)
# decoder_1: [batch, 128, 128, ngf * 2] => [batch, 256, 256, output_channels]
with tf.variable_scope("decoder_1"):
input = tf.concat([decoder_results[-1], encoder_results[0]], axis=3)
rectified = tfHelpers.lrelu(input, 0.2)
deconved = tfHelpers.deconv(rectified, output_channels)
#should we normalize it ?
deconved, lastGlobalNet = self._addSecondaryNetBlock(deconved, None, lastGlobalNet, output_channels, output_channels, num_encoder_layers + len(decoder_results), True)
#output = tf.tanh(deconved)
decoder_results.append(deconved)
return decoder_results[-1], lastGlobalNet
def create_generator(self,
generator_inputs,
generator_outputs_channels,
reuse_bool=True):
with tf.variable_scope("generator", reuse=reuse_bool) as scope:
#Print the shape to check we are inputting a tensor with a reasonable shape
print("generator_inputs :" + str(generator_inputs.get_shape()))
print("generator_outputs_channels :" +
str(generator_outputs_channels))
layers = []
#Input here should be [batch, 256,256,3]
inputMean, inputVariance = tf.nn.moments(generator_inputs,
axes=[1, 2],
keep_dims=False)
globalNetworkInput = inputMean
globalNetworkOutputs = []
with tf.variable_scope("globalNetwork_fc_1"):
globalNetwork_fc_1 = tfHelpers.fullyConnected(
globalNetworkInput, self.ngf * 2, True,
"globalNetworkLayer" + str(len(globalNetworkOutputs) + 1))
globalNetworkOutputs.append(tf.nn.selu(globalNetwork_fc_1))
#encoder_1: [batch, 256, 256, in_channels] => [batch, 128, 128, ngf]
with tf.variable_scope("encoder_1"):
#Convolution with stride 2 and kernel size 4x4.
output = tfHelpers.conv(generator_inputs, self.ngf, stride=2)
layers.append(output)
#Default ngf is 64
layer_specs = [
self.ngf *
2, # encoder_2: [batch, 128, 128, ngf] => [batch, 64, 64, ngf * 2]
self.ngf *
4, # encoder_3: [batch, 64, 64, ngf * 2] => [batch, 32, 32, ngf * 4]
self.ngf *
8, # encoder_4: [batch, 32, 32, ngf * 4] => [batch, 16, 16, ngf * 8]
self.ngf *
8, # encoder_5: [batch, 16, 16, ngf * 8] => [batch, 8, 8, ngf * 8]
self.ngf *
8, # encoder_6: [batch, 8, 8, ngf * 8] => [batch, 4, 4, ngf * 8]
self.ngf *
8, # encoder_7: [batch, 4, 4, ngf * 8] => [batch, 2, 2, ngf * 8]
#self.ngf * 8, # encoder_8: [batch, 2, 2, ngf * 8] => [batch, 1, 1, ngf * 8]
]
for layerCount, out_channels in enumerate(layer_specs):
with tf.variable_scope("encoder_%d" % (len(layers) + 1)):
#We use a leaky relu instead of a relu to let a bit more expressivity to the network.
rectified = tfHelpers.lrelu(layers[-1], 0.2)
# [batch, in_height, in_width, in_channels] => [batch, in_height/2, in_width/2, out_channels]
convolved = tfHelpers.conv(rectified,
out_channels,
stride=2)
#here mean and variance will be [batch, 1, 1, out_channels] and we run an instance normalization
outputs, mean, variance = tfHelpers.instancenorm(convolved)
#Get the last value in the global feature secondary network and transform it to be added to the current Unet layer output.
outputs = outputs + tfHelpers.GlobalToGenerator(
globalNetworkOutputs[-1], out_channels)
with tf.variable_scope("globalNetwork_fc_%d" %
(len(globalNetworkOutputs) + 1)):
#Prepare the input to the next global feature secondary network step and run it.
nextGlobalInput = tf.concat([
tf.expand_dims(tf.expand_dims(
globalNetworkOutputs[-1], axis=1),
axis=1), mean
],
axis=-1)
globalNetwork_fc = ""
if layerCount + 1 < len(layer_specs) - 1:
globalNetwork_fc = tfHelpers.fullyConnected(
nextGlobalInput, layer_specs[layerCount + 1],
True, "globalNetworkLayer" +
str(len(globalNetworkOutputs) + 1))
else:
globalNetwork_fc = tfHelpers.fullyConnected(
nextGlobalInput, layer_specs[layerCount], True,
"globalNetworkLayer" +
str(len(globalNetworkOutputs) + 1))
#We use selu as we are in a fully connected network and it has auto normalization properties.
globalNetworkOutputs.append(
tf.nn.selu(globalNetwork_fc))
layers.append(outputs)
with tf.variable_scope("encoder_8"):
#The last encoder is mostly similar to previous layers except that we don't normalize the output.
rectified = tfHelpers.lrelu(layers[-1], 0.2)
# [batch, in_height, in_width, in_channels] => [batch, in_height/2, in_width/2, out_channels]
convolvedNoGlobal = tfHelpers.conv(rectified,
self.ngf * 8,
stride=2)
convolved = convolvedNoGlobal + tfHelpers.GlobalToGenerator(
globalNetworkOutputs[-1], self.ngf * 8)
with tf.variable_scope("globalNetwork_fc_%d" %
(len(globalNetworkOutputs) + 1)):
mean, variance = tf.nn.moments(convolvedNoGlobal,
axes=[1, 2],
keep_dims=True)
nextGlobalInput = tf.concat([
tf.expand_dims(tf.expand_dims(globalNetworkOutputs[-1],
axis=1),
axis=1), mean
],
axis=-1)
globalNetwork_fc = tfHelpers.fullyConnected(
nextGlobalInput, self.ngf * 8, True,
"globalNetworkLayer" +
str(len(globalNetworkOutputs) + 1))
globalNetworkOutputs.append(tf.nn.selu(globalNetwork_fc))
layers.append(convolved)
#default nfg = 64
layer_specs = [
(
self.ngf * 8, 0.5
), # decoder_8: [batch, 1, 1, ngf * 8] => [batch, 2, 2, ngf * 8 * 2]
(
self.ngf * 8, 0.5
), # decoder_7: [batch, 2, 2, ngf * 8 * 2] => [batch, 4, 4, ngf * 8 * 2]
(
self.ngf * 8, 0.5
), # decoder_6: [batch, 4, 4, ngf * 8 * 2] => [batch, 8, 8, ngf * 8 * 2]
(
self.ngf * 8, 0.0
), # decoder_5: [batch, 8, 8, ngf * 8 * 2] => [batch, 16, 16, ngf * 8 * 2]
(
self.ngf * 4, 0.0
), # decoder_4: [batch, 16, 16, ngf * 8 * 2] => [batch, 32, 32, ngf * 4 * 2]
(
self.ngf * 2, 0.0
), # decoder_3: [batch, 32, 32, ngf * 4 * 2] => [batch, 64, 64, ngf * 2 * 2]
(
self.ngf, 0.0
), # decoder_2: [batch, 64, 64, ngf * 2 * 2] => [batch, 128, 128, ngf * 2]
]
#Start the decoder here
num_encoder_layers = len(layers)
for decoder_layer, (out_channels,
dropout) in enumerate(layer_specs):
skip_layer = num_encoder_layers - decoder_layer - 1
#Evaluate which layer from the encoder has to be contatenated for the skip connection
with tf.variable_scope("decoder_%d" % (skip_layer + 1)):
if decoder_layer == 0:
# first decoder layer doesn't have skip connections
# since it is directly connected to the skip_layer
input = layers[-1]
else:
input = tf.concat([layers[-1], layers[skip_layer]],
axis=3)
#Leaky relu some more (same reason as in the encoder)
rectified = tfHelpers.lrelu(input, 0.2)
# [batch, in_height, in_width, in_channels] => [batch, in_height*2, in_width*2, out_channels]
#The deconvolution has stride 1 and shape 4x4. Theorically, it should be shape 3x3 to avoid any effects on the image borders, but it doesn't seem to have such a strong effect.
output = tfHelpers.deconv(rectified, out_channels)
#Instance norm and global feature secondary network similar to the decoder.
output, mean, variance = tfHelpers.instancenorm(output)
output = output + tfHelpers.GlobalToGenerator(
globalNetworkOutputs[-1], out_channels)
with tf.variable_scope("globalNetwork_fc_%d" %
(len(globalNetworkOutputs) + 1)):
nextGlobalInput = tf.concat([
tf.expand_dims(tf.expand_dims(
globalNetworkOutputs[-1], axis=1),
axis=1), mean
],
axis=-1)
globalNetwork_fc = tfHelpers.fullyConnected(
nextGlobalInput, out_channels, True,
"globalNetworkLayer" +
str(len(globalNetworkOutputs) + 1))
globalNetworkOutputs.append(
tf.nn.selu(globalNetwork_fc))
if dropout > 0.0:
#We use dropout as described in the pix2pix paper.
output = tf.nn.dropout(output, keep_prob=1 - dropout)
layers.append(output)
# decoder_1: [batch, 128, 128, ngf * 2] => [batch, 256, 256, generator_outputs_channels]
with tf.variable_scope("decoder_1"):
input = tf.concat([layers[-1], layers[0]], axis=3)
rectified = tfHelpers.lrelu(input, 0.2)
output = tfHelpers.deconv(rectified,
generator_outputs_channels)
lastGlobalNet = tfHelpers.GlobalToGenerator(
globalNetworkOutputs[-1], generator_outputs_channels)
output = output + lastGlobalNet
#output = tf.tanh(output)
layers.append(output)
return layers[-1], lastGlobalNet
def create_generator(self,
generator_inputs,
generator_outputs_channels,
materialEncoded,
reuse_bool=True):
with tf.variable_scope("generator", reuse=reuse_bool) as scope:
layers = []
#Input here should be [batch, 256,256,3]
inputMean, inputVariance = tf.nn.moments(generator_inputs,
axes=[1, 2],
keep_dims=False)
globalNetworkInput = inputMean
globalNetworkOutputs = []
with tf.variable_scope("globalNetwork_fc_1"):
globalNetwork_fc_1 = tfHelpers.fullyConnected(
globalNetworkInput, self.ngf * 2, True,
"globalNetworkLayer" + str(len(globalNetworkOutputs) + 1))
globalNetworkOutputs.append(tf.nn.selu(globalNetwork_fc_1))
#encoder_1: [batch, 256, 256, in_channels] => [batch, 128, 128, ngf]
with tf.variable_scope("encoder_1"):
output = tfHelpers.conv(generator_inputs, self.ngf, stride=2)
layers.append(output)
#Default ngf is 64
layer_specs = [
self.ngf *
2, # encoder_2: [batch, 128, 128, ngf] => [batch, 64, 64, ngf * 2]
self.ngf *
4, # encoder_3: [batch, 64, 64, ngf * 2] => [batch, 32, 32, ngf * 4]
self.ngf *
8, # encoder_4: [batch, 32, 32, ngf * 4] => [batch, 16, 16, ngf * 8]
self.ngf *
8, # encoder_5: [batch, 16, 16, ngf * 8] => [batch, 8, 8, ngf * 8]
self.ngf *
8, # encoder_6: [batch, 8, 8, ngf * 8] => [batch, 4, 4, ngf * 8]
self.ngf *
8, # encoder_7: [batch, 4, 4, ngf * 8] => [batch, 2, 2, ngf * 8]
#self.ngf * 8, # encoder_8: [batch, 2, 2, ngf * 8] => [batch, 1, 1, ngf * 8]
]
for layerCount, out_channels in enumerate(layer_specs):
with tf.variable_scope("encoder_%d" % (len(layers) + 1)):
rectified = tfHelpers.lrelu(layers[-1], 0.2)
# [batch, in_height, in_width, in_channels] => [batch, in_height/2, in_width/2, out_channels]
convolved = tfHelpers.conv(rectified,
out_channels,
stride=2)
#here mean and variance will be [batch, 1, 1, out_channels]
outputs, mean, variance = tfHelpers.instancenorm(convolved)
outputs = outputs + self.GlobalToGenerator(
globalNetworkOutputs[-1], out_channels)
with tf.variable_scope("globalNetwork_fc_%d" %
(len(globalNetworkOutputs) + 1)):
nextGlobalInput = tf.concat([
tf.expand_dims(tf.expand_dims(
globalNetworkOutputs[-1], axis=1),
axis=1), mean
],
axis=-1)
globalNetwork_fc = ""
if layerCount + 1 < len(layer_specs) - 1:
globalNetwork_fc = tfHelpers.fullyConnected(
nextGlobalInput, layer_specs[layerCount + 1],
True, "globalNetworkLayer" +
str(len(globalNetworkOutputs) + 1))
else:
globalNetwork_fc = tfHelpers.fullyConnected(
nextGlobalInput, layer_specs[layerCount], True,
"globalNetworkLayer" +
str(len(globalNetworkOutputs) + 1))
globalNetworkOutputs.append(
tf.nn.selu(globalNetwork_fc))
layers.append(outputs)
with tf.variable_scope("encoder_8"):
rectified = tfHelpers.lrelu(layers[-1], 0.2)
# [batch, in_height, in_width, in_channels] => [batch, in_height/2, in_width/2, out_channels]
convolved = tfHelpers.conv(rectified, self.ngf * 8, stride=2)
convolved = convolved + self.GlobalToGenerator(
globalNetworkOutputs[-1], self.ngf * 8)
with tf.variable_scope("globalNetwork_fc_%d" %
(len(globalNetworkOutputs) + 1)):
mean, variance = tf.nn.moments(convolved,
axes=[1, 2],
keep_dims=True)
nextGlobalInput = tf.concat([
tf.expand_dims(tf.expand_dims(globalNetworkOutputs[-1],
axis=1),
axis=1), mean
],
axis=-1)
globalNetwork_fc = tfHelpers.fullyConnected(
nextGlobalInput, self.ngf * 8, True,
"globalNetworkLayer" +
str(len(globalNetworkOutputs) + 1))
globalNetworkOutputs.append(tf.nn.selu(globalNetwork_fc))
layers.append(convolved)
#default nfg = 64
layer_specs = [
(
self.ngf * 8, 0.5
), # decoder_8: [batch, 1, 1, ngf * 8] => [batch, 2, 2, ngf * 8 * 2]
(
self.ngf * 8, 0.5
), # decoder_7: [batch, 2, 2, ngf * 8 * 2] => [batch, 4, 4, ngf * 8 * 2]
(
self.ngf * 8, 0.5
), # decoder_6: [batch, 4, 4, ngf * 8 * 2] => [batch, 8, 8, ngf * 8 * 2]
(
self.ngf * 8, 0.0
), # decoder_5: [batch, 8, 8, ngf * 8 * 2] => [batch, 16, 16, ngf * 8 * 2]
(
self.ngf * 4, 0.0
), # decoder_4: [batch, 16, 16, ngf * 8 * 2] => [batch, 32, 32, ngf * 4 * 2]
(
self.ngf * 2, 0.0
), # decoder_3: [batch, 32, 32, ngf * 4 * 2] => [batch, 64, 64, ngf * 2 * 2]
(
self.ngf, 0.0
), # decoder_2: [batch, 64, 64, ngf * 2 * 2] => [batch, 128, 128, ngf * 2]
]
num_encoder_layers = len(layers)
for decoder_layer, (out_channels,
dropout) in enumerate(layer_specs):
skip_layer = num_encoder_layers - decoder_layer - 1
with tf.variable_scope("decoder_%d" % (skip_layer + 1)):
if decoder_layer == 0:
# first decoder layer doesn't have skip connections
# since it is directly connected to the skip_layer
input = layers[-1]
else:
input = tf.concat([layers[-1], layers[skip_layer]],
axis=3)
rectified = tfHelpers.lrelu(input, 0.2)
# [batch, in_height, in_width, in_channels] => [batch, in_height*2, in_width*2, out_channels]
output = tfHelpers.deconv(rectified, out_channels)
output, mean, variance = tfHelpers.instancenorm(output)
output = output + self.GlobalToGenerator(
globalNetworkOutputs[-1], out_channels)
with tf.variable_scope("globalNetwork_fc_%d" %
(len(globalNetworkOutputs) + 1)):
nextGlobalInput = tf.concat([
tf.expand_dims(tf.expand_dims(
globalNetworkOutputs[-1], axis=1),
axis=1), mean
],
axis=-1)
globalNetwork_fc = tfHelpers.fullyConnected(
nextGlobalInput, out_channels, True,
"globalNetworkLayer" +
str(len(globalNetworkOutputs) + 1))
globalNetworkOutputs.append(
tf.nn.selu(globalNetwork_fc))
if dropout > 0.0:
output = tf.nn.dropout(output, rate=dropout)
layers.append(output)
# decoder_1: [batch, 128, 128, ngf * 2] => [batch, 256, 256, generator_outputs_channels]
with tf.variable_scope("decoder_1"):
input = tf.concat([layers[-1], layers[0]], axis=3)
rectified = tfHelpers.lrelu(input, 0.2)
output = tfHelpers.deconv(rectified,
generator_outputs_channels)
lastGlobalNet = self.GlobalToGenerator(
globalNetworkOutputs[-1], generator_outputs_channels)
output = output + lastGlobalNet
output = tf.tanh(output)
layers.append(output)
return layers[-1], lastGlobalNet