def discriminator(images, layers, spectral, activation, reuse, normalization=None):
net = images
channels = [32, 64, 128, 256, 512, 1024]
if display:
print('Discriminator Information.')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print()
with tf.variable_scope('discriminator', reuse=reuse):
# Padding = 'Same' -> H_new = H_old // Stride
for layer in range(layers):
# Down.
net = convolutional(inputs=net, output_channels=channels[layer], filter_size=5, stride=2, padding='SAME', conv_type='convolutional', spectral=spectral, scope=layer+1)
if normalization is not None: net = normalization(inputs=net, training=True)
net = activation(net)
# Flatten.
net = tf.layers.flatten(inputs=net)
# Dense.
net = dense(inputs=net, out_dim=channels[-1], spectral=spectral, scope=1)
if normalization is not None: net = normalization(inputs=net, training=True)
net = activation(net)
# Dense
logits = dense(inputs=net, out_dim=1, spectral=spectral, scope=2)
output = sigmoid(logits)
print()
return output, logits
def mapping_network(z_input,
z_dim,
layers,
spectral,
activation,
normalization,
init='xavier',
regularizer=None):
if display:
print('MAPPING NETWORK INFORMATION:')
print('Layers: ', layers)
print('Normalization: ', normalization)
print('Activation: ', activation)
print()
with tf.variable_scope('mapping_network'):
net = z_input
for layer in range(layers):
net = normalization(inputs=net,
training=True,
c=None,
spectral=None,
scope=layer)
net = dense(inputs=net,
out_dim=z_dim,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=layer)
w_input = net
return w_input
def discriminator_encoder(enconding, layers, spectral, activation, reuse, init='xavier', regularizer=None, normalization=None, name='dis_encoding'):
net = enconding
channels = [150, 100, 50, 25, 12]
# channels = [200, 150, 100, 50, 24]
if display:
print('DISCRIMINATOR-ENCODER INFORMATION:')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print()
with tf.variable_scope(name, reuse=reuse):
for layer in range(layers):
# Residual Dense layer.
net = residual_block_dense(inputs=net, scope=layer, is_training=True, normalization=normalization, spectral=spectral, activation=activation, init=init, regularizer=regularizer, display=True)
# Dense layer downsample dim.
net = dense(inputs=net, out_dim=channels[layer], spectral=spectral, init=init, regularizer=regularizer, scope=layer)
if normalization is not None: net = normalization(inputs=net, training=True)
net = activation(net)
# Dense
logits_net = dense(inputs=net, out_dim=1, spectral=spectral, init=init, regularizer=regularizer, scope=layer+1)
output = sigmoid(logits_net)
print()
return output, logits_net
def discriminator_resnet_incr(images, layers, spectral, activation, reuse, init='xavier', regularizer=None, normalization=None, attention=None, down='downscale', label=None, label_t='cat', infoGAN=False, c_dim=None, name='discriminator'):
net = images
channels = [32, 64, 128, 256, 512, 1024]
if display:
print('DISCRIMINATOR INFORMATION:')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print('Attention: ', attention)
print()
with tf.variable_scope(name, reuse=reuse):
layer = 0
net = convolutional(inputs=net, output_channels=channels[layer], filter_size=3, stride=1, padding='SAME', conv_type='convolutional', spectral=spectral, init=init, regularizer=regularizer, scope=layer)
for layer in range(layers):
# ResBlock.
net = residual_block(inputs=net, filter_size=3, stride=1, padding='SAME', scope=layer, is_training=True, normalization=normalization, use_bias=True, spectral=spectral, init=init, regularizer=regularizer, activation=activation)
# Attention layer.
if attention is not None and net.shape.as_list()[1]==attention:
net = attention_block(net, spectral=True, init=init, regularizer=regularizer, scope=layers)
# Down.
layer_channel = layer+1
if layer == layers - 1:
layer_channel = -2
net = convolutional(inputs=net, output_channels=channels[layer_channel], filter_size=4, stride=2, padding='SAME', conv_type=down, spectral=spectral, init=init, regularizer=regularizer, scope=layer+1)
if normalization is not None: net = normalization(inputs=net, training=True)
net = activation(net)
# Flatten.
net = tf.layers.flatten(inputs=net)
# Dense.
net = dense(inputs=net, out_dim=channels[-1], spectral=spectral, init=init, regularizer=regularizer, scope=2)
if normalization is not None: net = normalization(inputs=net, training=True)
net = activation(net)
# Dense
logits = dense(inputs=net, out_dim=1, spectral=spectral, init=init, regularizer=regularizer, scope=3)
output = sigmoid(logits)
print()
return output, logits
def generator_resnet_cond(z_input, c_input, image_channels, layers, spectral, activation, reuse, is_train, normalization, up='upscale'):
channels = [32, 64, 128, 256, 512, 1024]
channels = [32, 64, 128, 256, 512, 1024]
reversed_channel = list(reversed(channels[:layers]))
if display:
print('Generator Information.')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
with tf.variable_scope('generator', reuse=reuse):
# Doesn't work ReLU, tried.
# Z Input Shape = (None, 100)
# C Input Shape = (None, 20)
net = tf.concat([z_input, c_input], axis=1)
# Dense.
net = dense(inputs=net, out_dim=1024, spectral=spectral, scope=1)
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Dense.
net = dense(inputs=net, out_dim=256*7*7, spectral=spectral, scope=2)
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Reshape
net = tf.reshape(tensor=net, shape=(-1, 7, 7, 256), name='reshape')
for layer in range(layers):
# ResBlock.
net = residual_block(inputs=net, filter_size=3, stride=1, padding='SAME', scope=layer, is_training=is_train, spectral=spectral,
activation=activation, normalization=normalization, c_input=c_input)
# Up.
net = convolutional(inputs=net, output_channels=reversed_channel[layer], filter_size=2, stride=2, padding='SAME', conv_type=up, spectral=spectral, scope=layer)
net = normalization(inputs=net, training=is_train, c=c_input, spectral=spectral)
net = activation(net)
logits = convolutional(inputs=net, output_channels=image_channels, filter_size=3, stride=1, padding='SAME', conv_type='convolutional', spectral=spectral, scope='logits')
output = sigmoid(logits)
print()
return output
def generator(z_input, image_channels, layers, spectral, activation, reuse, is_train, normalization):
channels = [32, 64, 128, 256, 512, 1024]
reversed_channel = list(reversed(channels[:layers]))
if display:
print('Generator Information.')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
with tf.variable_scope('generator', reuse=reuse):
# Doesn't work ReLU, tried.
# Dense.
net = dense(inputs=z_input, out_dim=1024, spectral=spectral, scope=1)
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Dense.
net = dense(inputs=net, out_dim=256*7*7, spectral=spectral, scope=2)
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Reshape
net = tf.reshape(tensor=net, shape=(-1, 7, 7, 256), name='reshape')
for layer in range(layers):
# Conv.
net = convolutional(inputs=net, output_channels=reversed_channel[layer], filter_size=2, stride=2, padding='SAME', conv_type='transpose', spectral=spectral, scope=2*(layer+1)-1)
net = normalization(inputs=net, training=is_train)
net = activation(net)
if layer != len(range(layers))-1:
# Conv.
net = convolutional(inputs=net, output_channels=reversed_channel[layer+1], filter_size=5, stride=1, padding='SAME', conv_type='convolutional', spectral=spectral, scope=2*(layer+1))
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Conv.
logits = convolutional(inputs=net, output_channels=image_channels, filter_size=2, stride=2, padding='SAME', conv_type='transpose', spectral=spectral, scope='logits')
output = sigmoid(logits)
print()
return output
def discriminator_resnet_mask_class(images, layers, spectral, activation, reuse, init='xavier', regularizer=None, normalization=None, attention=None, down='downscale', label=None, name='discriminator'):
net = images
# channels = [32, 64, 128, 256, 512, 1024, 2048]
channels = [32, 64, 128, 256, 512, 1024]
if display:
print('DISCRIMINATOR INFORMATION:', name)
print('Total Channels: ', channels)
print('Chosen Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print('Attention: ', attention)
print()
with tf.variable_scope(name, reuse=reuse):
# Discriminator with conditional projection.
batch_size, label_dim = label.shape.as_list()
embedding_size = channels[-1]
for layer in range(layers):
# ResBlock.
net = residual_block(inputs=net, filter_size=3, stride=1, padding='SAME', scope=layer, is_training=True, normalization=normalization, use_bias=True,
spectral=spectral, init=init, regularizer=regularizer, activation=activation)
# Attention layer.
if attention is not None and net.shape.as_list()[1]==attention:
net = attention_block(net, spectral=True, init=init, regularizer=regularizer, scope=layers)
# Down.
net = convolutional(inputs=net, output_channels=channels[layer], filter_size=4, stride=2, padding='SAME', conv_type=down, spectral=spectral, init=init, regularizer=regularizer, scope=layer)
if normalization is not None: net = normalization(inputs=net, training=True)
net = activation(net)
# Feature space extraction
feature_space = tf.layers.max_pooling2d(inputs=net, pool_size=[2, 2], strides=1)
feature_space = tf.layers.flatten(inputs=feature_space)
# Flatten.
net = tf.layers.flatten(inputs=net)
net = dense(inputs=net, out_dim=channels[-1], spectral=spectral, init=init, regularizer=regularizer, scope=1)
net = activation(net)
# Dense Classes
class_logits = dense(inputs=net, out_dim=label_dim, spectral=spectral, init=init, regularizer=regularizer, scope=3)
class_logits = tf.nn.log_softmax(class_logits)
# One encoding for label input
logits = class_logits*label
logits = tf.reduce_sum(logits, axis=-1)
output = sigmoid(logits)
print()
return output, logits, feature_space
def discriminator_resnet_class2(images, layers, spectral, activation, reuse, l_dim, init='xavier', regularizer=None, normalization=None, attention=None, down='downscale', name='discriminator'):
net = images
# channels = [32, 64, 128, 256, 512, 1024, 2048]
channels = [32, 64, 128, 256, 512, 1024]
# New
layers = layers + 1
if display:
print('DISCRIMINATOR INFORMATION:', name)
print('Total Channels: ', channels)
print('Chosen Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print('Attention: ', attention)
print()
with tf.variable_scope(name, reuse=reuse):
for layer in range(layers):
# ResBlock.
net = residual_block(inputs=net, filter_size=3, stride=1, padding='SAME', scope=layer, is_training=True, normalization=normalization, use_bias=True,
spectral=spectral, init=init, regularizer=regularizer, activation=activation)
# Attention layer.
if attention is not None and net.shape.as_list()[1]==attention:
net = attention_block(net, spectral=True, init=init, regularizer=regularizer, scope=layers)
# Down.
net = convolutional(inputs=net, output_channels=channels[layer], filter_size=4, stride=2, padding='SAME', conv_type=down, spectral=spectral, init=init, regularizer=regularizer, scope=layer)
if normalization is not None: net = normalization(inputs=net, training=True)
net = activation(net)
# New
# Flatten.
net = tf.layers.flatten(inputs=net)
# Dense.
feature_space = dense(inputs=net, out_dim=channels[-1], spectral=spectral, init=init, regularizer=regularizer, scope=2)
net = activation(feature_space)
# Dense
logits = dense(inputs=net, out_dim=1, spectral=spectral, init=init, regularizer=regularizer, scope=3)
output = sigmoid(logits)
net = dense(inputs=net, out_dim=channels[-1], spectral=spectral, init=init, regularizer=regularizer, scope=4)
net = activation(net)
# Dense Classes
class_logits = dense(inputs=net, out_dim=l_dim, spectral=spectral, init=init, regularizer=regularizer, scope=5)
print()
return output, logits, feature_space, class_logits
def generator_decoder_resnet(z_input, image_channels, layers, spectral, activation, reuse, is_train, normalization, attention=None, up='upscale'):
channels = [32, 64, 128, 256, 512, 1024]
reversed_channel = list(reversed(channels[:layers]))
if display:
print('GENERATOR-DECODER INFORMATION:')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print('Attention H/W: ', attention)
print()
with tf.variable_scope('generator_decoder', reuse=reuse):
# Doesn't work ReLU, tried.
# Dense.
net = dense(inputs=z_input, out_dim=1024, spectral=spectral, scope=1)
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Dense.
net = dense(inputs=net, out_dim=256*7*7, spectral=spectral, scope=2)
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Reshape
net = tf.reshape(tensor=net, shape=(-1, 7, 7, 256), name='reshape')
for layer in range(layers):
# ResBlock.
net = residual_block(inputs=net, filter_size=3, stride=1, padding='SAME', scope=layer, is_training=is_train, spectral=spectral, activation=activation, normalization=normalization)
# Attention layer.
if attention is not None and net.shape.as_list()[1]==attention:
net = attention_block(net, spectral=True, scope=layers)
# if (vae_dim/2.) == net.shape.as_list()[1]:
# lr_logs2_xi_z = convolutional(inputs=net, output_channels=reversed_channel[layer], filter_size=2, stride=2, padding='SAME', conv_type=up, spectral=spectral, scope='lr_logs2_xi_z')
# if (vae_dim/2.) == net.shape.as_list()[1]:
# scope = 'lr_mean_xi_z'
# else:
# scope = layer
# Up.
net = convolutional(inputs=net, output_channels=reversed_channel[layer], filter_size=2, stride=2, padding='SAME', conv_type=up, spectral=spectral, scope=layer)
net = normalization(inputs=net, training=is_train)
net = activation(net)
# if vae_dim == net.shape.as_list()[1]:
# lr_mean_xi_z = sigmoid(net)
# Final outputs
logits = convolutional(inputs=net, output_channels=image_channels, filter_size=3, stride=1, padding='SAME', conv_type='convolutional', spectral=spectral, scope='mean_xi_z')
mean_xi_z = sigmoid(logits)
# Final outputs
logs2_xi_z = convolutional(inputs=net, output_channels=image_channels, filter_size=3, stride=1, padding='SAME', conv_type='convolutional', spectral=spectral, scope='logs2_xi_z')
print()
# return output, lr_mean_xi_z, lr_logs2_xi_z
return mean_xi_z, logs2_xi_z
def generator_resnet_cond(z_input,
c_input,
image_channels,
layers,
spectral,
activation,
reuse,
is_train,
normalization,
up='upscale'):
channels = [32, 64, 128, 256, 512, 1024]
channels = [32, 64, 128, 256, 512, 1024]
reversed_channel = list(reversed(channels[:layers]))
if display:
print('Generator Information.')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
with tf.variable_scope('generator', reuse=reuse):
# Doesn't work ReLU, tried.
# Z Input Shape = (None, 100)
# C Input Shape = (None, 20)
net = tf.concat([z_input, c_input], axis=1)
# Dense.
net = dense(inputs=net, out_dim=1024, spectral=spectral, scope=1)
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Dense.
net = dense(inputs=net,
out_dim=256 * 7 * 7,
spectral=spectral,
scope=2)
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Reshape
net = tf.reshape(tensor=net, shape=(-1, 7, 7, 256), name='reshape')
for layer in range(layers):
# ResBlock.
net = residual_block(inputs=net,
filter_size=3,
stride=1,
padding='SAME',
scope=layer,
is_training=is_train,
spectral=spectral,
activation=activation,
normalization=normalization,
c_input=c_input)
# Up.
net = convolutional(inputs=net,
output_channels=reversed_channel[layer],
filter_size=2,
stride=2,
padding='SAME',
conv_type=up,
spectral=spectral,
scope=layer)
net = normalization(inputs=net,
training=is_train,
c=c_input,
spectral=spectral)
net = activation(net)
logits = convolutional(inputs=net,
output_channels=image_channels,
filter_size=3,
stride=1,
padding='SAME',
conv_type='convolutional',
spectral=spectral,
scope='logits')
output = sigmoid(logits)
print()
return output
def encoder_resnet(images,
z_dim,
layers,
spectral,
activation,
reuse,
init='xavier',
regularizer=None,
normalization=None,
attention=None,
down='downscale',
name='encoder'):
net = images
channels = [32, 64, 128, 256, 512, 1024]
if display:
print('ENCODER INFORMATION:', name)
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print('Attention: ', attention)
print()
with tf.variable_scope(name, reuse=reuse):
for layer in range(layers + 1):
# ResBlock.
net = residual_block(inputs=net,
filter_size=3,
stride=1,
padding='SAME',
scope=layer,
is_training=True,
normalization=normalization,
use_bias=True,
spectral=spectral,
init=init,
regularizer=regularizer,
activation=activation)
# Attention layer.
if attention is not None and net.shape.as_list()[1] == attention:
net = attention_block(net,
spectral=True,
init=init,
regularizer=regularizer,
scope=layers)
# Down.
net = convolutional(inputs=net,
output_channels=channels[layer],
filter_size=4,
stride=2,
padding='SAME',
conv_type=down,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=layer)
if normalization is not None:
net = normalization(inputs=net, training=True, scope=layer)
net = activation(net)
# Flatten.
net = tf.layers.flatten(inputs=net)
# Dense.
net = dense(inputs=net,
out_dim=channels[-1],
spectral=spectral,
init=init,
regularizer=regularizer,
scope=1)
if normalization is not None:
net = normalization(inputs=net, training=True)
net = activation(net)
# Dense
w_latent = dense(inputs=net,
out_dim=z_dim,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=2)
print()
return w_latent
def encoder_resnet_incr(images,
z_dim,
layers,
spectral,
activation,
reuse,
is_train,
init='xavier',
regularizer=None,
normalization=None,
attention=None,
stack_layers=False,
concat_img=False,
down='downscale',
name='encoder'):
out_stack_layers = list()
net = images
channels = [32, 64, 128, 256, 512, 1024]
if display:
print('ENCODER INFORMATION:')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print('Attention: ', attention)
print()
_, height, width, _ = images.shape.as_list()
with tf.variable_scope(name, reuse=reuse):
layer = 0
net = convolutional(inputs=net,
output_channels=channels[layer],
filter_size=3,
stride=1,
padding='SAME',
conv_type='convolutional',
spectral=spectral,
init=init,
regularizer=regularizer,
scope=layer)
for layer in range(layers):
# ResBlock.
net = residual_block(inputs=net,
filter_size=3,
stride=1,
padding='SAME',
scope=layer,
is_training=is_train,
normalization=normalization,
use_bias=True,
spectral=spectral,
init=init,
regularizer=regularizer,
activation=activation)
if concat_img and layer != 0:
down_sample = tf.image.resize_images(
images=images,
size=(int(height / (2**layer)), int(width / (2**layer))),
method=tf.image.ResizeMethod.BILINEAR,
align_corners=False)
print('down_sample', down_sample.shape)
print('net', net.shape)
net = tf.concat([net, down_sample], axis=-1)
print('net', net.shape)
# Attention layer.
if attention is not None and net.shape.as_list()[1] == attention:
net = attention_block(net,
spectral=True,
init=init,
regularizer=regularizer,
scope=layers)
if stack_layers:
print('Adding layer output to stack layer output.')
out_stack_layers.append(net)
# Down.
layer_channel = layer + 1
if layer == layers - 1:
layer_channel = -2
net = convolutional(inputs=net,
output_channels=channels[layer_channel],
filter_size=4,
stride=2,
padding='SAME',
conv_type=down,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=layer + 1)
if normalization is not None:
net = normalization(inputs=net, training=is_train)
net = activation(net)
if stack_layers:
print('Adding layer output to stack layer output.')
out_stack_layers.append(net)
if concat_img and layer != 0:
down_sample = tf.image.resize_images(
images=images,
size=(int(height / (2**(layer + 1))),
int(width / (2**(layer + 1)))),
method=tf.image.ResizeMethod.BILINEAR,
align_corners=False)
print('down_sample', down_sample.shape)
print('net', net.shape)
net = tf.concat([net, down_sample], axis=-1)
print('net', net.shape)
# Flatten.
net = tf.layers.flatten(inputs=net)
# shape = int(np.product(net.shape.as_list()[1:3])/2)
# # # Dense.
# net = dense(inputs=net, out_dim=shape, spectral=spectral, init=init, regularizer=regularizer, scope=1)
# if normalization is not None: net = normalization(inputs=net, training=True)
# net = activation(net)
# Dense.
net = dense(inputs=net,
out_dim=channels[-1],
spectral=spectral,
init=init,
regularizer=regularizer,
scope=2)
if normalization is not None:
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Dense
w_latent = dense(inputs=net,
out_dim=z_dim,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=3)
print()
if stack_layers:
return w_latent, out_stack_layers
return w_latent
def generator_resnet(z_input,
image_channels,
layers,
spectral,
activation,
reuse,
is_train,
normalization,
init='xavier',
noise_input_f=False,
regularizer=None,
cond_label=None,
attention=None,
up='upscale',
bigGAN=False):
channels = [32, 64, 128, 256, 512, 1024]
reversed_channel = list(reversed(channels[:layers]))
# Question here: combine z dims for upscale and the conv after, or make them independent.
if bigGAN:
z_dim = z_input.shape.as_list()[-1]
blocks = 2 + layers
block_dims = math.floor(z_dim / blocks)
remainder = z_dim - block_dims * blocks
if remainder == 0:
z_sets = [block_dims] * (blocks + 1)
else:
z_sets = [block_dims] * blocks + [remainder]
z_splits = tf.split(z_input, num_or_size_splits=z_sets, axis=-1)
if display:
print('GENERATOR INFORMATION:')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print('Attention H/W: ', attention)
print()
with tf.variable_scope('generator', reuse=reuse):
if bigGAN:
z_input_block = z_splits[0]
label = z_splits[1]
else:
z_input_block = z_input
label = z_input
if cond_label is not None:
if 'training_gate' in cond_label.name:
label = cond_label
else:
label = tf.concat([cond_label, label], axis=-1)
# Dense.
net = dense(inputs=z_input_block,
out_dim=1024,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=1)
net = normalization(inputs=net,
training=is_train,
c=label,
spectral=spectral,
scope='dense_1')
net = activation(net)
if bigGAN: label = z_splits[2]
else: label = z_input
if cond_label is not None:
if 'training_gate' in cond_label.name:
label = cond_label
else:
label = tf.concat([cond_label, label], axis=-1)
# Dense.
net = dense(inputs=net,
out_dim=256 * 7 * 7,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=2)
net = normalization(inputs=net,
training=is_train,
c=label,
spectral=spectral,
scope='dense_2')
net = activation(net)
# Reshape
net = tf.reshape(tensor=net, shape=(-1, 7, 7, 256), name='reshape')
for layer in range(layers):
if bigGAN: label = z_splits[3 + layer]
else: label = z_input
if cond_label is not None:
if 'training_gate' in cond_label.name:
label = cond_label
else:
label = tf.concat([cond_label, label], axis=-1)
# ResBlock.
net = residual_block(inputs=net,
filter_size=3,
stride=1,
padding='SAME',
scope=layer,
is_training=is_train,
spectral=spectral,
init=init,
regularizer=regularizer,
noise_input_f=noise_input_f,
activation=activation,
normalization=normalization,
cond_label=label)
# Attention layer.
if attention is not None and net.shape.as_list()[1] == attention:
net = attention_block(net,
spectral=True,
init=init,
regularizer=regularizer,
scope=layer)
# ResBlock.
# net = residual_block(inputs=net, filter_size=3, stride=1, padding='SAME', scope=layer+layers, is_training=is_train, spectral=spectral, init=init, regularizer=regularizer, noise_input_f=noise_input_f,
# activation=activation, normalization=normalization, cond_label=label)
# Up.
net = convolutional(inputs=net,
output_channels=reversed_channel[layer],
filter_size=2,
stride=2,
padding='SAME',
conv_type=up,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=layer)
if noise_input_f:
net = noise_input(inputs=net, scope=layer)
# net = normalization(inputs=net, training=is_train, c=label, spectral=spectral, scope=layer+layers)
net = normalization(inputs=net,
training=is_train,
c=label,
spectral=spectral,
scope=layer)
net = activation(net)
logits = convolutional(inputs=net,
output_channels=image_channels,
filter_size=3,
stride=1,
padding='SAME',
conv_type='convolutional',
spectral=spectral,
init=init,
regularizer=regularizer,
scope='logits')
output = sigmoid(logits)
print()
return output
def encoder_resnet_instnorm(images,
latent_dim,
layers,
spectral,
activation,
reuse,
is_train,
init='xavier',
regularizer=None,
normalization=instance_norm,
attention=None,
down='downscale',
name='encoder'):
net = images
channels = [32, 64, 128, 256, 512, 1024]
if display:
print('ENCODER INFORMATION:')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print('Attention: ', attention)
print()
_, height, width, _ = images.shape.as_list()
with tf.variable_scope(name, reuse=reuse):
layer = 0
net = convolutional(inputs=net,
output_channels=channels[layer],
filter_size=3,
stride=1,
padding='SAME',
conv_type='convolutional',
spectral=spectral,
init=init,
regularizer=regularizer,
scope=layer)
# Style extraction.
styles = style_extract(inputs=net,
latent_dim=latent_dim,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=layer)
if normalization is not None:
net = normalization(inputs=net, training=is_train)
net = activation(net)
for layer in range(layers):
# ResBlock.
net, style = residual_block(inputs=net,
filter_size=3,
stride=1,
padding='SAME',
scope=layer,
style_extract_f=True,
latent_dim=latent_dim,
is_training=is_train,
normalization=normalization,
use_bias=True,
spectral=spectral,
init=init,
regularizer=regularizer,
activation=activation)
styles += style
# Attention layer.
if attention is not None and net.shape.as_list()[1] == attention:
net = attention_block(net,
spectral=True,
init=init,
regularizer=regularizer,
scope=layers)
# Down.
layer_channel = layer + 1
if layer == layers - 1:
layer_channel = -2
net = convolutional(inputs=net,
output_channels=channels[layer_channel],
filter_size=4,
stride=2,
padding='SAME',
conv_type=down,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=layer + 1)
# Style extraction.
style = style_extract(inputs=net,
latent_dim=latent_dim,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=layer + 1)
styles += style
if normalization is not None:
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Flatten.
net = tf.layers.flatten(inputs=net)
# Dense.
net = dense(inputs=net,
out_dim=channels[-1],
spectral=spectral,
init=init,
regularizer=regularizer,
scope=2)
net = activation(net)
# Dense
style = dense(inputs=net,
out_dim=latent_dim,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=3)
styles += style
print()
return styles
def generator_resnet_style(z_input,
image_channels,
layers,
spectral,
activation,
reuse,
is_train,
normalization,
init='xavier',
noise_input_f=False,
regularizer=None,
cond_label=None,
attention=None,
up='upscale'):
channels = [32, 64, 128, 256, 512, 1024]
reversed_channel = list(reversed(channels[:layers]))
if display:
print('GENERATOR INFORMATION:')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print('Attention H/W: ', attention)
print()
with tf.variable_scope('generator', reuse=reuse):
z_input_block = z_input[:, :, 0]
label = z_input[:, :, 0]
# Dense.
net = dense(inputs=z_input_block,
out_dim=1024,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=1)
net = normalization(inputs=net,
training=is_train,
c=label,
spectral=spectral,
scope='dense_1')
net = activation(net)
# Dense.
net = dense(inputs=net,
out_dim=256 * 7 * 7,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=2)
net = normalization(inputs=net,
training=is_train,
c=label,
spectral=spectral,
scope='dense_2')
net = activation(net)
# Reshape
net = tf.reshape(tensor=net, shape=(-1, 7, 7, 256), name='reshape')
for layer in range(layers):
label = z_input[:, :, layer]
# ResBlock.
net = residual_block(inputs=net,
filter_size=3,
stride=1,
padding='SAME',
scope=layer,
is_training=is_train,
spectral=spectral,
init=init,
regularizer=regularizer,
noise_input_f=noise_input_f,
activation=activation,
normalization=normalization,
cond_label=label)
print('Z input:', layer)
# Attention layer.
if attention is not None and net.shape.as_list()[1] == attention:
net = attention_block(net,
spectral=True,
init=init,
regularizer=regularizer,
scope=layer)
# ResBlock.
# net = residual_block(inputs=net, filter_size=3, stride=1, padding='SAME', scope=layer+layers, is_training=is_train, spectral=spectral, init=init, regularizer=regularizer, noise_input_f=noise_input_f, activation=activation, normalization=normalization, cond_label=label)
label = z_input[:, :, layer + 1]
# Up.
net = convolutional(inputs=net,
output_channels=reversed_channel[layer],
filter_size=2,
stride=2,
padding='SAME',
conv_type=up,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=layer)
if noise_input_f:
net = noise_input(inputs=net, scope=layer)
net = normalization(inputs=net,
training=is_train,
c=label,
spectral=spectral,
scope=layer)
net = activation(net)
print('Z input:', layer + 1)
logits = convolutional(inputs=net,
output_channels=image_channels,
filter_size=3,
stride=1,
padding='SAME',
conv_type='convolutional',
spectral=spectral,
init=init,
regularizer=regularizer,
scope='logits')
output = sigmoid(logits)
print()
return output
def encoder_resnet(images,
z_dim,
layers,
spectral,
activation,
reuse,
normalization=None,
is_train=None,
attention=None,
down='downscale'):
net = images
channels = [32, 64, 128, 256, 512, 1024]
channels = [64, 128, 256, 512, 1024]
if display:
print('ENCODER INFORMATION:')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print('Attention: ', attention)
print()
with tf.variable_scope('encoder', reuse=reuse):
for layer in range(layers):
# ResBlock.
# if vae_dim == net.shape.as_list()[1]:
# scope = 'vae_out'
# else:
# scope = layer
net = residual_block(inputs=net,
filter_size=3,
stride=1,
padding='SAME',
scope=layer,
is_training=is_train,
normalization=normalization,
use_bias=True,
spectral=spectral,
activation=activation)
# Attention layer.
if attention is not None and net.shape.as_list()[1] == attention:
net = attention_block(net, spectral=True, scope=layers)
# if vae_dim == net.shape.as_list()[1]:
# vae_out = sigmoid(net)
# Down.
net = convolutional(inputs=net,
output_channels=channels[layer],
filter_size=4,
stride=2,
padding='SAME',
conv_type=down,
spectral=spectral,
scope=layer)
if normalization is not None:
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Flatten.
net = tf.layers.flatten(inputs=net)
# Dense.
net = dense(inputs=net,
out_dim=channels[-1],
spectral=spectral,
scope=1)
if normalization is not None:
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Dense.
mean_z_xi = dense(inputs=net,
out_dim=z_dim,
spectral=spectral,
scope='mean_z_xi')
logs2_z_xi = dense(inputs=net,
out_dim=z_dim,
spectral=spectral,
scope='logs2_z_xi')
print()
# return mean_z_xi, logs2_z_xi, vae_out
return mean_z_xi, logs2_z_xi
def discriminator_resnet(images,
layers,
spectral,
activation,
reuse,
init='xavier',
regularizer=None,
normalization=None,
attention=None,
down='downscale',
label=None,
label_t='cat',
infoGAN=False,
c_dim=None):
net = images
channels = [32, 64, 128, 256, 512, 1024]
if display:
print('DISCRIMINATOR INFORMATION:')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print('Attention: ', attention)
print()
with tf.variable_scope('discriminator', reuse=reuse):
# Discriminator with conditional projection.
if label is not None:
batch_size, label_dim = label.shape.as_list()
embedding_size = channels[-1]
# Categorical Embedding.
print(label_t)
if label_t == 'cat':
# emb = embedding(shape=(label_dim, embedding_size), init=init, regularizer=regularizer, power_iterations=1)
emb = embedding(shape=(label_dim, embedding_size),
init=init,
power_iterations=1)
index = tf.argmax(label, axis=-1)
label_emb = tf.nn.embedding_lookup(emb, index)
# Linear conditioning, using NN to produce embedding.
else:
inter_dim = int((label_dim + net.shape.as_list()[-1]) / 2)
net_label = dense(inputs=label,
out_dim=inter_dim,
spectral=spectral,
init='xavier',
regularizer=None,
scope='label_nn_1')
if normalization is not None:
net_label = normalization(inputs=net_label, training=True)
net_label = activation(net_label)
label_emb = dense(inputs=net_label,
out_dim=embedding_size,
spectral=spectral,
init='xavier',
regularizer=None,
scope='label_nn_2')
for layer in range(layers):
# ResBlock.
net = residual_block(inputs=net,
filter_size=3,
stride=1,
padding='SAME',
scope=layer,
is_training=True,
normalization=normalization,
use_bias=True,
spectral=spectral,
init=init,
regularizer=regularizer,
activation=activation)
# Attention layer.
if attention is not None and net.shape.as_list()[1] == attention:
net = attention_block(net,
spectral=True,
init=init,
regularizer=regularizer,
scope=layers)
# Down.
net = convolutional(inputs=net,
output_channels=channels[layer],
filter_size=4,
stride=2,
padding='SAME',
conv_type=down,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=layer)
if normalization is not None:
net = normalization(inputs=net, training=True)
net = activation(net)
# Flatten.
net = tf.layers.flatten(inputs=net)
# Dense.
net = dense(inputs=net,
out_dim=channels[-1],
spectral=spectral,
init=init,
regularizer=regularizer,
scope=1)
if normalization is not None:
net = normalization(inputs=net, training=True)
# net = activation(net)
# Dense
logits = dense(inputs=net,
out_dim=1,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=2)
if label is not None:
inner_prod = tf.reduce_sum(net * label_emb, axis=-1, keepdims=True)
output = sigmoid(logits + inner_prod)
else:
output = sigmoid(logits)
if infoGAN:
mean_c_x = dense(inputs=net,
out_dim=c_dim,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=3)
logs2_c_x = dense(inputs=net,
out_dim=c_dim,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=4)
return output, logits, mean_c_x, logs2_c_x
print()
return output, logits
def generator_resnet_style(w_input,
image_channels,
layers,
spectral,
activation,
reuse,
is_train,
normalization,
init='xavier',
noise_input_f=False,
regularizer=None,
cond_label=None,
attention=None,
stack_layers=False,
up='upscale',
name='generator'):
out_stack_layers = list()
channels = [32, 64, 128, 256, 512, 1024]
i_pixel = 7
reversed_channel = list(reversed(channels[:layers]))
if display:
print('GENERATOR INFORMATION:', name)
print('Total Channels: ', channels)
print('Chosen Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print('Attention H/W: ', attention)
print()
with tf.variable_scope(name, reuse=reuse):
w_input_block = w_input[:, :, 0]
# Dense.
label = w_input[:, :, 0]
net = dense(inputs=w_input_block,
out_dim=1024,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=1)
net = normalization(inputs=net,
training=is_train,
c=label,
spectral=spectral,
scope='dense_1')
net = activation(net)
# Dense.
net = dense(inputs=net,
out_dim=256 * i_pixel * i_pixel,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=2)
net = normalization(inputs=net,
training=is_train,
c=label,
spectral=spectral,
scope='dense_2')
net = activation(net)
# Reshape
# net = tf.reshape(tensor=net, shape=(-1, i_pixel, i_pixel, 1024), name='reshape')
net = tf.reshape(tensor=net,
shape=(-1, i_pixel, i_pixel, 256),
name='reshape')
# Loop for convolutional layers.
for layer in range(layers):
# ResBlock.
label = w_input[:, :, layer]
net = residual_block(inputs=net,
filter_size=3,
stride=1,
padding='SAME',
scope=layer,
is_training=is_train,
spectral=spectral,
init=init,
regularizer=regularizer,
noise_input_f=noise_input_f,
activation=activation,
normalization=normalization,
cond_label=label)
# Attention layer.
if attention is not None and net.shape.as_list()[1] == attention:
net = attention_block(net,
spectral=True,
init=init,
regularizer=regularizer,
scope=layers)
if stack_layers:
print('Adding layer output to stack layer output.')
out_stack_layers.append(net)
# Convolutional Up.
label = w_input[:, :, layer + 1]
net = convolutional(inputs=net,
output_channels=reversed_channel[layer],
filter_size=2,
stride=2,
padding='SAME',
conv_type=up,
spectral=spectral,
init=init,
regularizer=regularizer,
scope=layer)
if noise_input_f: net = noise_input(inputs=net, scope=layer)
net = normalization(inputs=net,
training=is_train,
c=label,
spectral=spectral,
scope=layer)
net = activation(net)
# net = residual_block(inputs=net, filter_size=3, stride=1, padding='SAME', scope=layer+1, is_training=is_train, spectral=spectral, init=init, regularizer=regularizer, noise_input_f=noise_input_f, activation=activation, normalization=normalization, cond_label=label)
if stack_layers:
print('Adding layer output to stack layer output.')
out_stack_layers.append(net)
logits = convolutional(inputs=net,
output_channels=image_channels,
filter_size=3,
stride=1,
padding='SAME',
conv_type='convolutional',
spectral=spectral,
init=init,
regularizer=regularizer,
scope='logits')
logits = normalization(inputs=logits,
training=is_train,
c=label,
spectral=spectral,
scope='logits_norm')
output = sigmoid(logits)
print()
if stack_layers:
return output, out_stack_layers
return output
def generator_decoder_resnet(z_input,
image_channels,
layers,
spectral,
activation,
reuse,
is_train,
normalization,
attention=None,
up='upscale'):
channels = [32, 64, 128, 256, 512, 1024]
reversed_channel = list(reversed(channels[:layers]))
if display:
print('GENERATOR-DECODER INFORMATION:')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print('Attention H/W: ', attention)
print()
with tf.variable_scope('generator_decoder', reuse=reuse):
# Doesn't work ReLU, tried.
# Dense.
net = dense(inputs=z_input, out_dim=1024, spectral=spectral, scope=1)
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Dense.
net = dense(inputs=net,
out_dim=256 * 7 * 7,
spectral=spectral,
scope=2)
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Reshape
net = tf.reshape(tensor=net, shape=(-1, 7, 7, 256), name='reshape')
for layer in range(layers):
# ResBlock.
net = residual_block(inputs=net,
filter_size=3,
stride=1,
padding='SAME',
scope=layer,
is_training=is_train,
spectral=spectral,
activation=activation,
normalization=normalization)
# Attention layer.
if attention is not None and net.shape.as_list()[1] == attention:
net = attention_block(net, spectral=True, scope=layers)
# if (vae_dim/2.) == net.shape.as_list()[1]:
# lr_logs2_xi_z = convolutional(inputs=net, output_channels=reversed_channel[layer], filter_size=2, stride=2, padding='SAME', conv_type=up, spectral=spectral, scope='lr_logs2_xi_z')
# if (vae_dim/2.) == net.shape.as_list()[1]:
# scope = 'lr_mean_xi_z'
# else:
# scope = layer
# Up.
net = convolutional(inputs=net,
output_channels=reversed_channel[layer],
filter_size=2,
stride=2,
padding='SAME',
conv_type=up,
spectral=spectral,
scope=layer)
net = normalization(inputs=net, training=is_train)
net = activation(net)
# if vae_dim == net.shape.as_list()[1]:
# lr_mean_xi_z = sigmoid(net)
# Final outputs
logits = convolutional(inputs=net,
output_channels=image_channels,
filter_size=3,
stride=1,
padding='SAME',
conv_type='convolutional',
spectral=spectral,
scope='mean_xi_z')
mean_xi_z = sigmoid(logits)
# Final outputs
logs2_xi_z = convolutional(inputs=net,
output_channels=image_channels,
filter_size=3,
stride=1,
padding='SAME',
conv_type='convolutional',
spectral=spectral,
scope='logs2_xi_z')
print()
# return output, lr_mean_xi_z, lr_logs2_xi_z
return mean_xi_z, logs2_xi_z
def generator_resnet(z_input, image_channels, layers, spectral, activation, reuse, is_train, normalization, init='xavier', regularizer=None, cond_label=None, attention=None, up='upscale', bigGAN=False):
channels = [32, 64, 128, 256, 512, 1024]
reversed_channel = list(reversed(channels[:layers]))
# Question here: combine z dims for upscale and the conv after, or make them independent.
if bigGAN:
z_dim = z_input.shape.as_list()[-1]
blocks = 2 + layers
block_dims = math.floor(z_dim/blocks)
remainder = z_dim - block_dims*blocks
if remainder == 0:
z_sets = [block_dims]*(blocks + 1)
else:
z_sets = [block_dims]*blocks + [remainder]
z_splits = tf.split(z_input, num_or_size_splits=z_sets, axis=-1)
if display:
print('GENERATOR INFORMATION:')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
print('Attention H/W: ', attention)
print()
with tf.variable_scope('generator', reuse=reuse):
if bigGAN:
z_input_block = z_splits[0]
label = z_splits[1]
else:
z_input_block = z_input
label = z_input
if cond_label is not None: label = tf.concat([cond_label, label], axis=-1)
# Dense.
net = dense(inputs=z_input_block, out_dim=1024, spectral=spectral, init=init, regularizer=regularizer, scope=1)
# net = batch_norm(inputs=net, training=is_train)
# Introducing instance norm into dense layers.
net = normalization(inputs=net, training=is_train, c=label, spectral=spectral, scope='dense_1')
net = activation(net)
if bigGAN: label = z_splits[2]
else: label = z_input
if cond_label is not None: label = tf.concat([cond_label, label], axis=-1)
# Dense.
net = dense(inputs=net, out_dim=256*7*7, spectral=spectral, init=init, regularizer=regularizer, scope=2)
# net = batch_norm(inputs=net, training=is_train)
# Introducing instance norm into dense layers.
net = normalization(inputs=net, training=is_train, c=label, spectral=spectral, scope='dense_2')
net = activation(net)
# Reshape
net = tf.reshape(tensor=net, shape=(-1, 7, 7, 256), name='reshape')
for layer in range(layers):
if bigGAN: label = z_splits[3+layer]
else: label = z_input
if cond_label is not None: label = tf.concat([cond_label, label], axis=-1)
# ResBlock.
net = residual_block(inputs=net, filter_size=3, stride=1, padding='SAME', scope=layer, is_training=is_train, spectral=spectral, init=init, regularizer=regularizer,
activation=activation, normalization=normalization, cond_label=label)
# Attention layer.
if attention is not None and net.shape.as_list()[1]==attention:
net = attention_block(net, spectral=True, init=init, regularizer=regularizer, scope=layers)
# Up.
net = convolutional(inputs=net, output_channels=reversed_channel[layer], filter_size=2, stride=2, padding='SAME', conv_type=up, spectral=spectral, init=init, regularizer=regularizer, scope=layer)
net = normalization(inputs=net, training=is_train, c=label, spectral=spectral, scope=layer)
net = activation(net)
logits = convolutional(inputs=net, output_channels=image_channels, filter_size=3, stride=1, padding='SAME', conv_type='convolutional', spectral=spectral, init=init, regularizer=regularizer, scope='logits')
output = sigmoid(logits)
print()
return output
def generator(z_input, image_channels, layers, spectral, activation, reuse,
is_train, normalization):
channels = [32, 64, 128, 256, 512, 1024]
reversed_channel = list(reversed(channels[:layers]))
if display:
print('Generator Information.')
print('Channels: ', channels[:layers])
print('Normalization: ', normalization)
print('Activation: ', activation)
with tf.variable_scope('generator', reuse=reuse):
# Doesn't work ReLU, tried.
# Dense.
net = dense(inputs=z_input, out_dim=1024, spectral=spectral, scope=1)
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Dense.
net = dense(inputs=net,
out_dim=256 * 7 * 7,
spectral=spectral,
scope=2)
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Reshape
net = tf.reshape(tensor=net, shape=(-1, 7, 7, 256), name='reshape')
for layer in range(layers):
# Conv.
net = convolutional(inputs=net,
output_channels=reversed_channel[layer],
filter_size=2,
stride=2,
padding='SAME',
conv_type='transpose',
spectral=spectral,
scope=2 * (layer + 1) - 1)
net = normalization(inputs=net, training=is_train)
net = activation(net)
if layer != len(range(layers)) - 1:
# Conv.
net = convolutional(inputs=net,
output_channels=reversed_channel[layer +
1],
filter_size=5,
stride=1,
padding='SAME',
conv_type='convolutional',
spectral=spectral,
scope=2 * (layer + 1))
net = normalization(inputs=net, training=is_train)
net = activation(net)
# Conv.
logits = convolutional(inputs=net,
output_channels=image_channels,
filter_size=2,
stride=2,
padding='SAME',
conv_type='transpose',
spectral=spectral,
scope='logits')
output = sigmoid(logits)
print()
return output
