def build_and_train_models():
# load MNIST dataset
(x_train, y_train), (_, _) = mnist.load_data()
# reshape data for CNN as (28, 28, 1) and normalize
image_size = x_train.shape[1]
x_train = np.reshape(x_train, [-1, image_size, image_size, 1])
x_train = x_train.astype('float32') / 255
# train labels
num_labels = len(np.unique(y_train))
y_train = to_categorical(y_train)
model_name = "acgan_mnist"
# network parameters
latent_size = 100
batch_size = 64
train_steps = 40000
lr = 2e-4
decay = 6e-8
input_shape = (image_size, image_size, 1)
label_shape = (num_labels, )
# build discriminator Model
inputs = Input(shape=input_shape, name='discriminator_input')
# call discriminator builder with 2 outputs, pred source and labels
discriminator = gan.discriminator(inputs, num_labels=num_labels)
# [1] uses Adam, but discriminator converges easily with RMSprop
optimizer = RMSprop(lr=lr, decay=decay)
# 2 loss fuctions: 1) probability image is real
# 2) class label of the image
loss = ['binary_crossentropy', 'categorical_crossentropy']
discriminator.compile(loss=loss, optimizer=optimizer, metrics=['accuracy'])
discriminator.summary()
# build generator model
input_shape = (latent_size, )
inputs = Input(shape=input_shape, name='z_input')
labels = Input(shape=label_shape, name='labels')
# call generator builder with input labels
generator = gan.generator(inputs, image_size, labels=labels)
generator.summary()
# build adversarial model = generator + discriminator
optimizer = RMSprop(lr=lr * 0.5, decay=decay * 0.5)
# freeze the weights of discriminator during adversarial training
discriminator.trainable = False
adversarial = Model([inputs, labels],
discriminator(generator([inputs, labels])),
name=model_name)
# same 2 loss fuctions: 1) probability image is real
# 2) class label of the image
adversarial.compile(loss=loss, optimizer=optimizer, metrics=['accuracy'])
adversarial.summary()
# train discriminator and adversarial networks
models = (generator, discriminator, adversarial)
data = (x_train, y_train)
params = (batch_size, latent_size, train_steps, num_labels, model_name)
train(models, data, params)
def build_and_train_models():
# load MNIST dataset
(x_train, _), (_, _) = mnist.load_data()
# reshape data for CNN as (28, 28, 1) and normalize
image_size = x_train.shape[1]
x_train = np.reshape(x_train, [-1, image_size, image_size, 1])
x_train = x_train.astype('float32') / 255
model_name = "wgan_mnist"
# network parameters
# the latent or z vector is 100-dim
latent_size = 100
# hyper parameters from WGAN paper [2]
n_critic = 5
clip_value = 0.01
batch_size = 64
lr = 5e-5
train_steps = 40000
input_shape = (image_size, image_size, 1)
# build discriminator model
inputs = Input(shape=input_shape, name='discriminator_input')
# WGAN uses linear activation in paper [2]
discriminator = gan.discriminator(inputs, activation='linear')
optimizer = RMSprop(lr=lr)
# WGAN discriminator uses wassertein loss
discriminator.compile(loss=wasserstein_loss,
optimizer=optimizer,
metrics=['accuracy'])
discriminator.summary()
# build generator model
input_shape = (latent_size, )
inputs = Input(shape=input_shape, name='z_input')
generator = gan.generator(inputs, image_size)
generator.summary()
# build adversarial model = generator + discriminator
# freeze the weights of discriminator during adversarial training
discriminator.trainable = False
adversarial = Model(inputs,
discriminator(generator(inputs)),
name=model_name)
adversarial.compile(loss=wasserstein_loss,
optimizer=optimizer,
metrics=['accuracy'])
adversarial.summary()
# train discriminator and adversarial networks
models = (generator, discriminator, adversarial)
params = (batch_size,
latent_size,
n_critic,
clip_value,
train_steps,
model_name)
train(models, x_train, params)
def build_and_train_models():
"""Load the dataset, build LSGAN discriminator,
generator, and adversarial models.
Call the LSGAN train routine.
"""
# load MNIST dataset
(x_train, _), (_, _) = mnist.load_data()
# reshape data for CNN as (28, 28, 1) and normalize
image_size = x_train.shape[1]
x_train = np.reshape(x_train, [-1, image_size, image_size, 1])
x_train = x_train.astype('float32') / 255
model_name = "lsgan_mnist"
# network parameters
# the latent or z vector is 100-dim
latent_size = 100
input_shape = (image_size, image_size, 1)
batch_size = 64
lr = 2e-4
decay = 6e-8
train_steps = 40000
# build discriminator model
inputs = Input(shape=input_shape, name='discriminator_input')
discriminator = gan.discriminator(inputs, activation=None)
# [1] uses Adam, but discriminator easily
# converges with RMSprop
optimizer = RMSprop(lr=lr, decay=decay)
# LSGAN uses MSE loss [2]
discriminator.compile(loss='mse',
optimizer=optimizer,
metrics=['accuracy'])
discriminator.summary()
# build generator model
input_shape = (latent_size, )
inputs = Input(shape=input_shape, name='z_input')
generator = gan.generator(inputs, image_size)
generator.summary()
# build adversarial model = generator + discriminator
optimizer = RMSprop(lr=lr * 0.5, decay=decay * 0.5)
# freeze the weights of discriminator
# during adversarial training
discriminator.trainable = False
adversarial = Model(inputs,
discriminator(generator(inputs)),
name=model_name)
# LSGAN uses MSE loss [2]
adversarial.compile(loss='mse', optimizer=optimizer, metrics=['accuracy'])
adversarial.summary()
# train discriminator and adversarial networks
models = (generator, discriminator, adversarial)
params = (batch_size, latent_size, train_steps, model_name)
gan.train(models, x_train, params)
def build_generator(latent_codes, image_size, feature1_dim=256):
"""Build Generator Model sub networks
Two sub networks: 1) Class and noise to feature1
(intermediate feature)
2) feature1 to image
# Arguments
latent_codes (Layers): dicrete code (labels),
noise and feature1 features
image_size (int): Target size of one side
(assuming square image)
feature1_dim (int): feature1 dimensionality
# Returns
gen0, gen1 (Models): Description below
"""
# Latent codes and network parameters
labels, z0, z1, feature1 = latent_codes
# image_resize = image_size // 4
# kernel_size = 5
# layer_filters = [128, 64, 32, 1]
# gen1 inputs
inputs = [labels, z1] # 10 + 50 = 62-dim
x = concatenate(inputs, axis=1)
x = Dense(512, activation='relu')(x)
x = BatchNormalization()(x)
x = Dense(512, activation='relu')(x)
x = BatchNormalization()(x)
fake_feature1 = Dense(feature1_dim, activation='relu')(x)
# gen1: classes and noise (feature2 + z1) to feature1
gen1 = Model(inputs, fake_feature1, name='gen1')
# gen0: feature1 + z0 to feature0 (image)
gen0 = gan.generator(feature1, image_size, codes=z0)
return gen0, gen1
def build_and_train_models(latent_size=100):
# load MNIST dataset
(x_train, y_train), (_, _) = mnist.load_data()
# reshape data for CNN as (28, 28, 1) and normalize
image_size = x_train.shape[1]
x_train = np.reshape(x_train, [-1, image_size, image_size, 1])
x_train = x_train.astype('float32') / 255
# train labels
num_labels = len(np.unique(y_train))
y_train = to_categorical(y_train)
model_name = "infogan_mnist"
# network parameters
batch_size = 64
train_steps = 40000
lr = 2e-4
decay = 6e-8
input_shape = (image_size, image_size, 1)
label_shape = (num_labels, )
code_shape = (1, )
# build discriminator model
inputs = Input(shape=input_shape, name='discriminator_input')
# call discriminator builder with 4 outputs:
# source, label, and 2 codes
discriminator = gan.discriminator(inputs,
num_labels=num_labels,
num_codes=2)
# [1] uses Adam, but discriminator converges easily with RMSprop
optimizer = RMSprop(lr=lr, decay=decay)
# loss functions: 1) probability image is real
# (binary crossentropy)
# 2) categorical cross entropy image label,
# 3) and 4) mutual information loss
loss = [
'binary_crossentropy', 'categorical_crossentropy', mi_loss, mi_loss
]
# lamda or mi_loss weight is 0.5
loss_weights = [1.0, 1.0, 0.5, 0.5]
discriminator.compile(loss=loss,
loss_weights=loss_weights,
optimizer=optimizer,
metrics=['accuracy'])
discriminator.summary()
# build generator model
input_shape = (latent_size, )
inputs = Input(shape=input_shape, name='z_input')
labels = Input(shape=label_shape, name='labels')
code1 = Input(shape=code_shape, name="code1")
code2 = Input(shape=code_shape, name="code2")
# call generator with inputs,
# labels and codes as total inputs to generator
generator = gan.generator(inputs,
image_size,
labels=labels,
codes=[code1, code2])
generator.summary()
# build adversarial model = generator + discriminator
optimizer = RMSprop(lr=lr * 0.5, decay=decay * 0.5)
discriminator.trainable = False
# total inputs = noise code, labels, and codes
inputs = [inputs, labels, code1, code2]
adversarial = Model(inputs,
discriminator(generator(inputs)),
name=model_name)
# same loss as discriminator
adversarial.compile(loss=loss,
loss_weights=loss_weights,
optimizer=optimizer,
metrics=['accuracy'])
adversarial.summary()
# train discriminator and adversarial networks
models = (generator, discriminator, adversarial)
data = (x_train, y_train)
params = (batch_size, latent_size, train_steps, num_labels, model_name)
train(models, data, params)