def __init__(self, x_dim, y_dim, z_dim, gen_architecture, disc_architecture, aux_architecture, last_layer_activation,
folder="./CWGAN", image_shape=None
):
super(InfoGAN, self).__init__(x_dim, y_dim, z_dim, [gen_architecture, disc_architecture, aux_architecture],
last_layer_activation, folder, image_shape, None)
self._gen_architecture = self._architectures[0]
self._disc_architecture = self._architectures[1]
self._aux_architecture = self._architectures[2]
################# Define architecture
self._gen_architecture.append([logged_dense, {"units": x_dim, "activation": tf.nn.sigmoid, "name": "Output"}])
self._generator = ConditionalGenerator(self._gen_architecture, name="Generator")
self._disc_architecture.append([logged_dense, {"units": 1, "activation": tf.nn.sigmoid, "name": "Output"}])
self._disc = Discriminator(self._disc_architecture, name="Discriminator")
self._aux_architecture.append([logged_dense, {"units": y_dim, "activation": tf.nn.softmax, "name": "Output"}])
self._aux = Encoder(self._aux_architecture, name="Auxiliary")
self._nets = [self._generator, self._disc, self._aux]
################# Connect inputs and networks
self._output_gen = self._generator.generate_net(self._mod_Z_input)
self._output_disc_fake = self._disc.generate_net(self._output_gen)
self._output_disc_real = self._disc.generate_net(self._X_input)
self._output_aux = self._aux.generate_net(self._output_gen)
################# Finalize
self._init_folders()
self._verify_init()
def __init__(self, x_dim, y_dim, z_dim, enc_dim, gen_architecture, enc_architecture, dec_architecture,
last_layer_activation, folder="./CBEGAN", image_shape=None, append_y_at_every_layer=None
):
super(CBEGAN, self).__init__(x_dim, y_dim, z_dim, [gen_architecture, enc_architecture, dec_architecture],
last_layer_activation, folder, image_shape, append_y_at_every_layer)
self._gen_architecture = self._architectures[0]
self._enc_architecture = self._architectures[1]
self._dec_architecture = self._architectures[2]
self._enc_dim = enc_dim
self._kt = 0
self._gamma = 1.
self._lambda = 0.001
################# Define architecture
if len(self._x_dim) == 1:
self._gen_architecture.append([logged_dense, {"units": x_dim, "activation": self._last_layer_activation, "name": "Output"}])
self._dec_architecture.append([logged_dense, {"units": x_dim, "activation": self._last_layer_activation, "name": "Output"}])
else:
self._enc_architecture.append([tf.layers.flatten, {"name": "Flatten"}])
self._gen_architecture[-1][1]["name"] = "Output"
self._enc_architecture.append([logged_dense, {"units": enc_dim, "activation": tf.nn.tanh, "name": "Output"}])
self._generator = ConditionalGenerator(self._gen_architecture, name="Generator")
self._encoder = Encoder(self._enc_architecture, name="Encoder")
self._decoder = Encoder(self._dec_architecture, name="Decoder")
self._nets = [self._generator, self._encoder, self._decoder]
################# Connect inputs and networks
self._output_gen = self._generator.generate_net(self._mod_Z_input,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
with tf.name_scope("InputsEncoder"):
if len(self._x_dim) == 1:
self._input_fake = tf.concat(axis=1, values=[self._output_gen, self._Y_input], name="fake")
self._input_real = tf.concat(axis=1, values=[self._X_input, self._Y_input], name="real")
else:
self._input_fake = image_condition_concat(inputs=self._output_gen, condition=self._Y_input, name="fake")
self._input_real = image_condition_concat(inputs=self._X_input, condition=self._Y_input, name="real")
self._output_encoder_real = self._encoder.generate_net(self._input_real, tf_trainflag=self._is_training)
self._output_encoder_fake = self._encoder.generate_net(self._input_fake, tf_trainflag=self._is_training)
self._output_decoder_real = self._decoder.generate_net(self._output_encoder_real, tf_trainflag=self._is_training)
self._output_decoder_fake = self._decoder.generate_net(self._output_encoder_fake, tf_trainflag=self._is_training)
################# Finalize
self._init_folders()
self._verify_init()
def __init__(self, x_dim, y_dim, z_dim, gen_architecture, critic_architecture, aux_architecture,
last_layer_activation, folder="./CC_CWGAN1", image_shape=None, append_y_at_every_layer=None
):
super(CC_CWGAN1, self).__init__(x_dim, y_dim, z_dim, [gen_architecture, critic_architecture, aux_architecture],
last_layer_activation, folder, image_shape, append_y_at_every_layer)
self._gen_architecture = self._architectures[0]
self._critic_architecture = self._architectures[1]
self._aux_architecture = self._architectures[2]
################# Define architecture
self._gen_architecture.append([logged_dense, {"units": x_dim, "activation": self._last_layer_activation, "name": "Output"}])
self._generator = ConditionalGenerator(self._gen_architecture, name="Generator")
self._critic_architecture.append([logged_dense, {"units": 1, "activation": tf.identity, "name": "Output"}])
self._critic = Critic(self._critic_architecture, name="Critic")
self._aux_architecture.append([logged_dense, {"units": y_dim, "activation": tf.identity, "name": "Output"}])
self._aux = Encoder(self._aux_architecture, name="Auxiliary")
self._nets = [self._generator, self._critic, self._aux]
################# Connect inputs and networks
self._output_gen = self._generator.generate_net(self._mod_Z_input,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
with tf.name_scope("InputsCritic"):
self._input_fake = tf.concat(axis=1, values=[self._output_gen, self._Y_input], name="fake")
self._input_real = tf.concat(axis=1, values=[self._X_input, self._Y_input], name="real")
self._output_critic_real = self._critic.generate_net(self._input_real)
self._output_critic_fake = self._critic.generate_net(self._input_fake)
self._output_aux = self._aux.generate_net(self._output_gen)
self._output_critic_real = self._critic.generate_net(self._input_real,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
self._output_critic_fake = self._critic.generate_net(self._input_fake,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
################# Finalize
self._init_folders()
self._verify_init()
def __init__(self,
x_dim,
y_dim,
z_dim,
dec_architecture,
enc_architecture,
last_layer_activation,
folder="./VAE",
image_shape=None,
append_y_at_every_layer=None):
super(CVAE, self).__init__(x_dim, y_dim, z_dim,
[dec_architecture, enc_architecture],
last_layer_activation, folder, image_shape,
append_y_at_every_layer)
self._gen_architecture = self._architectures[0]
self._enc_architecture = self._architectures[1]
################# Define architecture
if len(self._x_dim) == 1:
self._gen_architecture.append([
logged_dense, {
"units": x_dim,
"activation": self._last_layer_activation,
"name": "Output"
}
])
else:
self._enc_architecture.append(
[tf.layers.flatten, {
"name": "Flatten"
}])
self._gen_architecture[-1][1]["name"] = "Output"
last_layer_mean = [
logged_dense, {
"units": z_dim,
"activation": tf.identity,
"name": "Mean"
}
]
last_layer_std = [
logged_dense, {
"units": z_dim,
"activation": tf.identity,
"name": "Std"
}
]
self._encoder_mean = Encoder(self._enc_architecture +
[last_layer_mean],
name="Encoder")
self._encoder_std = Encoder(self._enc_architecture + [last_layer_std],
name="Encoder")
self._generator = ConditionalDecoder(self._gen_architecture,
name="Generator")
self._nets = [self._generator, self._encoder_mean]
################# Connect inputs and networks
with tf.name_scope("InputsEncoder"):
if len(self._x_dim) == 1:
self._mod_X_input = tf.concat(
axis=1, values=[self._X_input, self._Y_input], name="real")
else:
self._mod_X_input = image_condition_concat(
inputs=self._X_input, condition=self._Y_input, name="real")
self._mean_layer = self._encoder_mean.generate_net(
self._mod_X_input,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
self._std_layer = self._encoder_std.generate_net(
self._mod_X_input,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
self._output_enc = self._mean_layer + tf.exp(
0.5 * self._std_layer) * self._Z_input
with tf.name_scope("InputsGenerator"):
self._dec_input = tf.concat(
axis=1,
values=[self._output_enc, self._Y_input],
name="latent")
self._output_dec = self._generator.generate_net(
self._dec_input,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
self._output_dec_from_encoding = self._generator.generate_net(
self._mod_Z_input,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
################# Finalize
self._init_folders()
self._verify_init()
class CVAE(ConditionalGenerativeModel):
def __init__(self,
x_dim,
y_dim,
z_dim,
dec_architecture,
enc_architecture,
last_layer_activation,
folder="./VAE",
image_shape=None,
append_y_at_every_layer=None):
super(CVAE, self).__init__(x_dim, y_dim, z_dim,
[dec_architecture, enc_architecture],
last_layer_activation, folder, image_shape,
append_y_at_every_layer)
self._gen_architecture = self._architectures[0]
self._enc_architecture = self._architectures[1]
################# Define architecture
if len(self._x_dim) == 1:
self._gen_architecture.append([
logged_dense, {
"units": x_dim,
"activation": self._last_layer_activation,
"name": "Output"
}
])
else:
self._enc_architecture.append(
[tf.layers.flatten, {
"name": "Flatten"
}])
self._gen_architecture[-1][1]["name"] = "Output"
last_layer_mean = [
logged_dense, {
"units": z_dim,
"activation": tf.identity,
"name": "Mean"
}
]
last_layer_std = [
logged_dense, {
"units": z_dim,
"activation": tf.identity,
"name": "Std"
}
]
self._encoder_mean = Encoder(self._enc_architecture +
[last_layer_mean],
name="Encoder")
self._encoder_std = Encoder(self._enc_architecture + [last_layer_std],
name="Encoder")
self._generator = ConditionalDecoder(self._gen_architecture,
name="Generator")
self._nets = [self._generator, self._encoder_mean]
################# Connect inputs and networks
with tf.name_scope("InputsEncoder"):
if len(self._x_dim) == 1:
self._mod_X_input = tf.concat(
axis=1, values=[self._X_input, self._Y_input], name="real")
else:
self._mod_X_input = image_condition_concat(
inputs=self._X_input, condition=self._Y_input, name="real")
self._mean_layer = self._encoder_mean.generate_net(
self._mod_X_input,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
self._std_layer = self._encoder_std.generate_net(
self._mod_X_input,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
self._output_enc = self._mean_layer + tf.exp(
0.5 * self._std_layer) * self._Z_input
with tf.name_scope("InputsGenerator"):
self._dec_input = tf.concat(
axis=1,
values=[self._output_enc, self._Y_input],
name="latent")
self._output_dec = self._generator.generate_net(
self._dec_input,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
self._output_dec_from_encoding = self._generator.generate_net(
self._mod_Z_input,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
################# Finalize
self._init_folders()
self._verify_init()
def compile(self,
logged_images=None,
logged_labels=None,
learning_rate=0.0001,
optimizer=tf.train.AdamOptimizer):
self._define_loss()
with tf.name_scope("Optimizer"):
vae_optimizer = optimizer(learning_rate=learning_rate)
self._vae_optimizer = vae_optimizer.minimize(self._vae_loss,
name="VAE")
self._summarise(logged_images=logged_images,
logged_labels=logged_labels)
def _define_loss(self):
with tf.name_scope("Loss") as scope:
self._data_fidelity_loss = self._X_input * tf.log(
1e-10 + self._output_dec) + (
1 - self._X_input) * tf.log(1e-10 + 1 - self._output_dec)
self._data_fidelity_loss = -tf.reduce_sum(self._data_fidelity_loss,
1)
self._KLdiv = 1 + self._std_layer - tf.square(
self._mean_layer) - tf.exp(self._std_layer)
self._KLdiv = -0.5 * tf.reduce_sum(self._KLdiv, 1)
self._vae_loss = tf.reduce_mean(self._data_fidelity_loss +
self._KLdiv)
tf.summary.scalar("Loss", self._vae_loss)
def train(self,
x_train,
y_train,
x_test=None,
y_test=None,
epochs=100,
batch_size=64,
steps=5,
log_step=3,
gpu_options=None):
self._set_up_training(log_step=log_step, gpu_options=gpu_options)
self._set_up_test_train_sample(x_train, y_train, x_test, y_test)
self._log_results(epoch=0, epoch_time=0)
for epoch in range(epochs):
batch_nr = 0
loss_epoch = 0
start = time.clock()
trained_examples = 0
while trained_examples < len(x_train):
loss_batch = self._optimize(self._trainset, batch_size, steps)
trained_examples += batch_size
loss_epoch += loss_batch
epoch_train_time = (time.clock() - start) / 60
loss_epoch = np.round(loss_epoch, 2)
print("Epoch {}: Loss: {}.".format(epoch, loss_epoch))
if log_step is not None:
self._log(epoch + 1, epoch_train_time)
def _optimize(self, dataset, batch_size, steps):
for i in range(steps):
current_batch_x, current_batch_y = dataset.get_next_batch(
batch_size)
Z_noise = self.sample_noise(n=len(current_batch_x))
_, loss_batch = self._sess.run(
[self._vae_optimizer, self._vae_loss],
feed_dict={
self._X_input: current_batch_x,
self._Y_input: current_batch_y,
self._Z_input: Z_noise,
self._is_training: True
})
return loss_batch
def decode(self, inpt_x, is_encoded):
if not is_encoded:
inpt_x = self._encoder_mean.encode(noise=inpt_x, sess=self._sess)
return self._generator.decode(inpt_x, self._sess)
def __init__(self,
x_dim,
z_dim,
enc_dim,
gen_architecture,
enc_architecture,
dec_architecture,
last_layer_activation,
folder="./BEGAN",
image_shape=None):
super(BEGAN, self).__init__(
x_dim, z_dim,
[gen_architecture, enc_architecture, dec_architecture],
last_layer_activation, folder, image_shape)
self._gen_architecture = self._architectures[0]
self._enc_architecture = self._architectures[1]
self._dec_architecture = self._architectures[2]
self._enc_dim = enc_dim
self._kt = 0
self._gamma = 1.
self._lambda = 0.001
################# Define architecture
if len(self._x_dim) == 1:
self._gen_architecture.append([
logged_dense, {
"units": x_dim,
"activation": self._last_layer_activation,
"name": "Output"
}
])
self._dec_architecture.append([
logged_dense, {
"units": x_dim,
"activation": self._last_layer_activation,
"name": "Output"
}
])
else:
self._enc_architecture.append(
[tf.layers.flatten, {
"name": "Flatten"
}])
self._gen_architecture[-1][1]["name"] = "Output"
self._enc_architecture.append([
logged_dense, {
"units": enc_dim,
"activation": tf.nn.tanh,
"name": "Output"
}
])
self._generator = Generator(self._gen_architecture, name="Generator")
self._encoder = Encoder(self._enc_architecture, name="Encoder")
self._decoder = Encoder(self._dec_architecture, name="Decoder")
self._nets = [self._generator, self._encoder, self._decoder]
################# Connect inputs and networks
self._output_gen = self._generator.generate_net(
self._Z_input, tf_trainflag=self._is_training)
self._output_encoder_real = self._encoder.generate_net(
self._X_input, tf_trainflag=self._is_training)
self._output_encoder_fake = self._encoder.generate_net(
self._output_gen, tf_trainflag=self._is_training)
self._output_decoder_real = self._decoder.generate_net(
self._output_encoder_real, tf_trainflag=self._is_training)
self._output_decoder_fake = self._decoder.generate_net(
self._output_encoder_fake, tf_trainflag=self._is_training)
################# Finalize
self._init_folders()
self._verify_init()
class BEGAN(GenerativeModel):
def __init__(self,
x_dim,
z_dim,
enc_dim,
gen_architecture,
enc_architecture,
dec_architecture,
last_layer_activation,
folder="./BEGAN",
image_shape=None):
super(BEGAN, self).__init__(
x_dim, z_dim,
[gen_architecture, enc_architecture, dec_architecture],
last_layer_activation, folder, image_shape)
self._gen_architecture = self._architectures[0]
self._enc_architecture = self._architectures[1]
self._dec_architecture = self._architectures[2]
self._enc_dim = enc_dim
self._kt = 0
self._gamma = 1.
self._lambda = 0.001
################# Define architecture
if len(self._x_dim) == 1:
self._gen_architecture.append([
logged_dense, {
"units": x_dim,
"activation": self._last_layer_activation,
"name": "Output"
}
])
self._dec_architecture.append([
logged_dense, {
"units": x_dim,
"activation": self._last_layer_activation,
"name": "Output"
}
])
else:
self._enc_architecture.append(
[tf.layers.flatten, {
"name": "Flatten"
}])
self._gen_architecture[-1][1]["name"] = "Output"
self._enc_architecture.append([
logged_dense, {
"units": enc_dim,
"activation": tf.nn.tanh,
"name": "Output"
}
])
self._generator = Generator(self._gen_architecture, name="Generator")
self._encoder = Encoder(self._enc_architecture, name="Encoder")
self._decoder = Encoder(self._dec_architecture, name="Decoder")
self._nets = [self._generator, self._encoder, self._decoder]
################# Connect inputs and networks
self._output_gen = self._generator.generate_net(
self._Z_input, tf_trainflag=self._is_training)
self._output_encoder_real = self._encoder.generate_net(
self._X_input, tf_trainflag=self._is_training)
self._output_encoder_fake = self._encoder.generate_net(
self._output_gen, tf_trainflag=self._is_training)
self._output_decoder_real = self._decoder.generate_net(
self._output_encoder_real, tf_trainflag=self._is_training)
self._output_decoder_fake = self._decoder.generate_net(
self._output_encoder_fake, tf_trainflag=self._is_training)
################# Finalize
self._init_folders()
self._verify_init()
def compile(self,
learning_rate_gen=0.0001,
learning_rate_ae=0.0001,
optimizer=tf.train.AdamOptimizer):
self._define_loss()
with tf.name_scope("Optimizer"):
gen_optimizer = optimizer(learning_rate=learning_rate_gen)
self._gen_optimizer = gen_optimizer.minimize(
self._gen_loss,
var_list=self._get_vars("Generator"),
name="Generator")
ae_optimizer = optimizer(learning_rate=learning_rate_ae)
self._ae_optimizer = ae_optimizer.minimize(
self._ae_loss,
var_list=self._get_vars("Encoder") + self._get_vars("Decoder"),
name="Autoencoder")
self._summarise()
def _define_loss(self):
with tf.name_scope("Loss") as scope:
self._loss_real = tf.reduce_mean(
tf.abs(self._X_input - self._output_decoder_real))
self._loss_fake = tf.reduce_mean(
tf.abs(self._output_gen - self._output_decoder_fake))
self._gen_loss = self._loss_fake
tf.summary.scalar("Generator_loss", self._gen_loss)
self._ae_loss = self._loss_real - self._kt * self._loss_fake
tf.summary.scalar("Autoencoder_loss", self._ae_loss)
self._global_loss = self._loss_real + tf.abs(
self._gamma * self._loss_real - self._loss_fake)
tf.summary.scalar("Global_loss", self._global_loss)
def train(self,
x_train,
x_test=None,
epochs=100,
batch_size=64,
log_step=3,
steps=None,
gpu_options=None):
self._set_up_training(log_step=log_step, gpu_options=gpu_options)
self._set_up_test_train_sample(x_train, x_test)
self._log_results(epoch=0, epoch_time=0)
for epoch in range(epochs):
batch_nr = 0
ae_loss_epoch = 0
gen_loss_epoch = 0
start = time.clock()
trained_examples = 0
while trained_examples < len(x_train):
ae_loss_batch, gen_loss_batch, loss_real_batch, loss_fake_batch = self._optimize(
self._trainset, batch_size)
trained_examples += batch_size
ae_loss_epoch += ae_loss_batch
gen_loss_epoch += gen_loss_batch
self._kt = np.maximum(
np.minimum(
1., self._kt + self._lambda *
(self._gamma * loss_real_batch - loss_fake_batch)), 0.)
epoch_train_time = (time.clock() - start) / 60
ae_loss_epoch = np.round(ae_loss_epoch, 2)
gen_loss_epoch = np.round(gen_loss_epoch, 2)
print("Epoch {}: Autoencoder: {} \n\t\t\tGenerator: {}.".format(
epoch + 1, ae_loss_epoch, gen_loss_epoch))
if self._log_step is not None:
self._log(epoch + 1, epoch_train_time)
def _optimize(self, dataset, batch_size):
current_batch_x = dataset.get_next_batch(batch_size)
Z_noise = self.sample_noise(n=len(current_batch_x))
_, ae_loss_batch, loss_real_batch, loss_fake_batch = self._sess.run(
[
self._ae_optimizer, self._ae_loss, self._loss_real,
self._loss_fake
],
feed_dict={
self._X_input: current_batch_x,
self._Z_input: Z_noise,
self._is_training: True
})
Z_noise = self._generator.sample_noise(n=len(current_batch_x))
_, gen_loss_batch = self._sess.run(
[self._gen_optimizer, self._gen_loss],
feed_dict={
self._Z_input: Z_noise,
self._is_training: True
})
return ae_loss_batch, gen_loss_batch, loss_real_batch, loss_fake_batch
class CC_CWGAN1(ConditionalGenerativeModel):
def __init__(self, x_dim, y_dim, z_dim, gen_architecture, critic_architecture, aux_architecture,
last_layer_activation, folder="./CC_CWGAN1", image_shape=None, append_y_at_every_layer=None
):
super(CC_CWGAN1, self).__init__(x_dim, y_dim, z_dim, [gen_architecture, critic_architecture, aux_architecture],
last_layer_activation, folder, image_shape, append_y_at_every_layer)
self._gen_architecture = self._architectures[0]
self._critic_architecture = self._architectures[1]
self._aux_architecture = self._architectures[2]
################# Define architecture
self._gen_architecture.append([logged_dense, {"units": x_dim, "activation": self._last_layer_activation, "name": "Output"}])
self._generator = ConditionalGenerator(self._gen_architecture, name="Generator")
self._critic_architecture.append([logged_dense, {"units": 1, "activation": tf.identity, "name": "Output"}])
self._critic = Critic(self._critic_architecture, name="Critic")
self._aux_architecture.append([logged_dense, {"units": y_dim, "activation": tf.identity, "name": "Output"}])
self._aux = Encoder(self._aux_architecture, name="Auxiliary")
self._nets = [self._generator, self._critic, self._aux]
################# Connect inputs and networks
self._output_gen = self._generator.generate_net(self._mod_Z_input,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
with tf.name_scope("InputsCritic"):
self._input_fake = tf.concat(axis=1, values=[self._output_gen, self._Y_input], name="fake")
self._input_real = tf.concat(axis=1, values=[self._X_input, self._Y_input], name="real")
self._output_critic_real = self._critic.generate_net(self._input_real)
self._output_critic_fake = self._critic.generate_net(self._input_fake)
self._output_aux = self._aux.generate_net(self._output_gen)
self._output_critic_real = self._critic.generate_net(self._input_real,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
self._output_critic_fake = self._critic.generate_net(self._input_fake,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
################# Finalize
self._init_folders()
self._verify_init()
def compile(self, logged_images=None, logged_labels=None, learning_rate_gen=0.0001, learning_rate_critic=0.0001,
learning_rate_aux=0.0001, optimizer=tf.train.RMSPropOptimizer):
self._define_loss()
with tf.name_scope("Optimizer"):
gen_optimizer = optimizer(learning_rate=learning_rate_gen)
self._gen_optimizer = gen_optimizer.minimize(self._gen_loss, var_list=self._get_vars(scope="Generator"), name="Generator")
critic_optimizer = optimizer(learning_rate=learning_rate_critic)
self._critic_optimizer = critic_optimizer.minimize(self._critic_loss, var_list=self._get_vars(scope="Critic"), name="Critic")
aux_optimizer = optimizer(learning_rate=learning_rate_aux)
self._aux_optimizer = aux_optimizer.minimize(self._aux_loss, var_list=self._get_vars(scope="Generator")+self._get_vars(scope="Auxiliary"), name="Auxiliary")
self._clip_critic_param = [p.assign(tf.clip_by_value(p, -0.01, 0.01)) for p in self._get_vars("Critic")]
self._summarise(logged_images=logged_images, logged_labels=logged_labels)
def _define_loss(self):
with tf.name_scope("Loss") as scope:
self._gen_loss = -tf.reduce_mean(self._output_critic_fake)
tf.summary.scalar("Generator_loss", self._gen_loss)
self._critic_loss = -(tf.reduce_mean(self._output_critic_real) - tf.reduce_mean(self._output_critic_fake))
tf.summary.scalar("Critic_loss", self._critic_loss)
self._aux_loss = tf.reduce_mean(tf.abs(self._Y_input - self._output_aux))
tf.summary.scalar("Auxiliary_loss", self._aux_loss)
def train(self, x_train, y_train, x_test=None, y_test=None, epochs=100, batch_size=64, gen_steps=1, critic_steps=5, log_step=3):
self._set_up_training(log_step=log_step)
self._set_up_test_train_sample(x_train, y_train, x_test, y_test)
self._log_results(epoch=0, epoch_time=0)
for epoch in range(epochs):
batch_nr = 0
critic_loss_epoch = 0
gen_loss_epoch = 0
aux_loss_epoch = 0
start = time.clock()
trained_examples = 0
while trained_examples < len(x_train):
critic_loss_batch, gen_loss_batch, aux_loss_batch = self._optimize(self._trainset, batch_size, critic_steps, gen_steps)
trained_examples += batch_size
critic_loss_epoch += critic_loss_batch
gen_loss_epoch += gen_loss_batch
aux_loss_epoch += aux_loss_batch
epoch_train_time = (time.clock() - start)/60
critic_loss_epoch = np.round(critic_loss_epoch, 2)
gen_loss_epoch = np.round(gen_loss_epoch, 2)
print("Epoch {}: Critic: {} \n\t\t\tGenerator: {}\n\t\t\tAuxiliary: {}.".format(epoch, critic_loss_epoch, gen_loss_epoch, aux_loss_epoch))
if log_step is not None:
self._log(epoch+1, epoch_train_time)
def _optimize(self, dataset, batch_size, critic_steps, gen_steps):
for i in range(critic_steps):
current_batch_x, current_batch_y = dataset.get_next_batch(batch_size)
Z_noise = self.sample_noise(n=len(current_batch_x))
_, critic_loss_batch, clipping_D = self._sess.run([
self._critic_optimizer, self._critic_loss, self._clip_critic_param
],
feed_dict={self._X_input: current_batch_x, self._Y_input: current_batch_y,
self._Z_input: Z_noise, self._is_training: True
})
for _ in range(gen_steps):
Z_noise = self._generator.sample_noise(n=len(current_batch_x))
_, gen_loss_batch = self._sess.run([self._gen_optimizer, self._gen_loss],
feed_dict={self._Z_input: Z_noise, self._Y_input: current_batch_y,
self._is_training: True})
_, aux_loss_batch = self._sess.run([self._aux_optimizer, self._aux_loss],
feed_dict={self._Z_input: Z_noise, self._Y_input: current_batch_y,
self._is_training: True})
return critic_loss_batch, gen_loss_batch, aux_loss_batch
def predict(self, inpt_x, inpt_y, is_encoded):
if is_encoded:
inpt_x = self._sess.run(self._mod_Z_input, feed_dict={self._Z_input: inpt_x, self._Y_input_oneHot: inpt_y})
inpt_x = self._generator.generate_samples(noise=inpt_x, sess=self._sess)
inpt = self._sess.run(self._input_real, feed_dict={self._X_input: inpt_x, self._Y_input: inpt_y})
return self._critic.predict(inpt, self._sess)
def __init__(self,
x_dim,
y_dim,
z_dim,
gen_architecture,
disc_architecture,
enc_architecture,
last_layer_activation,
folder="./Results/CBiGAN_log",
image_shape=None,
append_y_at_every_layer=None):
super(CBiGAN, self).__init__(
x_dim, y_dim, z_dim,
[gen_architecture, disc_architecture, enc_architecture],
last_layer_activation, folder, image_shape,
append_y_at_every_layer)
self._gen_architecture = self._architectures[0]
self._disc_architecture = self._architectures[1]
self._enc_architecture = self._architectures[2]
################# Define architecture
if len(self._x_dim) == 1:
self._gen_architecture.append([
logged_dense, {
"units": x_dim,
"activation": self._last_layer_activation,
"name": "Output"
}
])
else:
self._disc_architecture.append(
[tf.layers.flatten, {
"name": "Flatten"
}])
self._enc_architecture.append(
[tf.layers.flatten, {
"name": "Flatten"
}])
self._gen_architecture[-1][1]["name"] = "Output"
self._disc_architecture.append([
logged_dense, {
"units": 1,
"activation": tf.sigmoid,
"name": "Output"
}
])
self._enc_architecture.append([
logged_dense, {
"units": z_dim,
"activation": tf.identity,
"name": "Output"
}
])
self._generator = ConditionalGenerator(self._gen_architecture,
name="Generator")
self._discriminator = Discriminator(self._disc_architecture,
name="Discriminator")
self._encoder = Encoder(self._enc_architecture, name="Encoder")
self._nets = [self._generator, self._discriminator, self._encoder]
################# Connect inputs and networks
with tf.name_scope("InputsEncoder"):
if len(self._x_dim) == 1:
self._mod_X_input = tf.concat(
axis=1,
values=[self._X_input, self._Y_input],
name="modified_x")
else:
self._mod_X_input = image_condition_concat(
images=self._X_input, condition=self._Y_input, name="fake")
self._output_gen = self._generator.generate_net(
self._mod_Z_input,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
self._output_enc = self._encoder.generate_net(
self._mod_X_input,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
with tf.name_scope("InputsCritic"):
self._mod_output_enc = tf.concat(
axis=1,
values=[self._output_enc, self._Y_input],
name="modified_encoder")
if len(self._x_dim) == 1:
self._disc_input_fake = tf.concat(
axis=1,
values=[self._output_gen, self._mod_Z_input],
name="fake")
self._disc_input_real = tf.concat(
axis=1,
values=[self._X_input, self._mod_output_enc],
name="real")
else:
self._disc_input_fake = image_condition_concat(
inputs=self._output_gen,
condition=self._mod_Z_input,
name="fake")
self._disc_input_real = image_condition_concat(
inputs=self._X_input,
condition=self._mod_output_enc,
name="real")
self._output_disc_fake = self._discriminator.generate_net(
self._disc_input_fake,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
self._output_disc_real = self._discriminator.generate_net(
self._disc_input_real,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
################# Finalize
self._init_folders()
self._verify_init()
class InfoGAN(ConditionalGenerativeModel):
def __init__(self, x_dim, y_dim, z_dim, gen_architecture, disc_architecture, aux_architecture, last_layer_activation,
folder="./CWGAN", image_shape=None
):
super(InfoGAN, self).__init__(x_dim, y_dim, z_dim, [gen_architecture, disc_architecture, aux_architecture],
last_layer_activation, folder, image_shape, None)
self._gen_architecture = self._architectures[0]
self._disc_architecture = self._architectures[1]
self._aux_architecture = self._architectures[2]
################# Define architecture
self._gen_architecture.append([logged_dense, {"units": x_dim, "activation": tf.nn.sigmoid, "name": "Output"}])
self._generator = ConditionalGenerator(self._gen_architecture, name="Generator")
self._disc_architecture.append([logged_dense, {"units": 1, "activation": tf.nn.sigmoid, "name": "Output"}])
self._disc = Discriminator(self._disc_architecture, name="Discriminator")
self._aux_architecture.append([logged_dense, {"units": y_dim, "activation": tf.nn.softmax, "name": "Output"}])
self._aux = Encoder(self._aux_architecture, name="Auxiliary")
self._nets = [self._generator, self._disc, self._aux]
################# Connect inputs and networks
self._output_gen = self._generator.generate_net(self._mod_Z_input)
self._output_disc_fake = self._disc.generate_net(self._output_gen)
self._output_disc_real = self._disc.generate_net(self._X_input)
self._output_aux = self._aux.generate_net(self._output_gen)
################# Finalize
self._init_folders()
self._verify_init()
def compile(self, learning_rate=0.0003, learning_rate_gen=None, learning_rate_disc=None, learning_rate_aux=None, optimizer=tf.train.RMSPropOptimizer):
self._define_loss()
if learning_rate_gen is None:
learning_rate_gen = learning_rate
if learning_rate_disc is None:
learning_rate_disc = learning_rate
if learning_rate_aux is None:
learning_rate_aux = learning_rate
with tf.name_scope("Optimizer"):
gen_optimizer = optimizer(learning_rate=learning_rate_gen)
self._gen_optimizer = gen_optimizer.minimize(self._gen_loss, var_list=self._get_vars(scope="Generator"), name="Generator")
disc_optimizer = optimizer(learning_rate=learning_rate_disc)
self._disc_optimizer = disc_optimizer.minimize(self._disc_loss, var_list=self._get_vars(scope="Discriminator"), name="Discriminator")
aux_optimizer = optimizer(learning_rate=learning_rate_aux)
self._aux_optimizer = aux_optimizer.minimize(self._aux_loss, var_list=self._get_vars("Auxiliary")+self._get_vars("Generator"), name="Auxiliary")
self._summarise(logged_labels=np.identity(self._y_dim))
def _define_loss(self):
with tf.name_scope("Loss") as scope:
self._gen_loss = -tf.reduce_mean(tf.log(self._output_disc_fake+0.00001))
tf.summary.scalar("Generator_loss", self._gen_loss)
self._disc_loss = -tf.reduce_mean(tf.log(self._output_disc_real+0.00001) + tf.log(1.0-self._output_disc_fake+0.00001))
tf.summary.scalar("Discriminator_loss", self._disc_loss)
self._aux_loss = -tf.reduce_mean(-tf.reduce_sum(tf.log(self._output_aux+0.00001)*self._Y_input, 1))
tf.summary.scalar("Auxiliary_loss", self._aux_loss)
def train(self, x_train, epochs=100, batch_size=64, gen_steps=1, disc_steps=1, log_step=3):
self._set_up_training(log_step=log_step)
self._trainset = Dataset(x_train)
nr_test_samples = 5000 if len(x_train) >=5000 else len(x_train)
self._x_test = self._trainset.sample(nr_test_samples)
self._y_test = self.sample_condition(n=len(self._x_test))
self._z_test = self._generator.sample_noise(n=nr_test_samples)
for epoch in range(epochs):
batch_nr = 0
disc_loss_epoch = 0
gen_loss_epoch = 0
aux_loss_epoch = 0
start = time.clock()
trained_examples = 0
while trained_examples < len(x_train):
disc_loss_batch, gen_loss_batch, aux_loss_batch = self._optimize(self._trainset, batch_size, disc_steps, gen_steps)
trained_examples += batch_size
disc_loss_epoch += disc_loss_batch
gen_loss_epoch += gen_loss_batch
aux_loss_epoch += aux_loss_batch
epoch_train_time = (time.clock() - start)/60
disc_loss_epoch = np.round(disc_loss_epoch, 2)
gen_loss_epoch = np.round(gen_loss_epoch, 2)
aux_loss_epoch = np.round(aux_loss_epoch, 2)
print("Epoch {}: Discriminator: {}\n\t\t\tGenerator: {}\n\t\t\tAuxiliary: {}.".format(epoch, disc_loss_epoch, gen_loss_epoch, aux_loss_epoch))
if log_step is not None:
self._log(epoch, epoch_train_time)
def _optimize(self, dataset, batch_size, disc_steps, gen_steps):
for i in range(disc_steps):
current_batch_x = dataset.get_next_batch(batch_size)
Z_noise = self.sample_noise(n=len(current_batch_x))
C_noise = self.sample_condition(n=len(current_batch_x))
latent_space = np.concatenate((Z_noise, C_noise), axis=1)
_, disc_loss_batch = self._sess.run([self._disc_optimizer, self._disc_loss],
feed_dict={self._X_input: current_batch_x, self._Y_input: C_noise, self._Z_input: Z_noise
})
for _ in range(gen_steps):
Z_noise = self.sample_noise(n=len(current_batch_x))
C_noise = self.sample_condition(n=len(current_batch_x))
latent_space = np.concatenate((Z_noise, C_noise), axis=1)
_, gen_loss_batch = self._sess.run([self._gen_optimizer, self._gen_loss], feed_dict={self._Y_input: C_noise, self._Z_input: Z_noise})
_, aux_loss_batch = self._sess.run([self._aux_optimizer, self._aux_loss], feed_dict={self._Y_input: C_noise, self._Z_input: Z_noise})
return disc_loss_batch, gen_loss_batch, aux_loss_batch
def sample_condition(self, n):
return np.random.multinomial(1, self._y_dim*[1/self._y_dim], size=n)
def predict(self, inpt_x, is_encoded, inpt_c=None):
if is_encoded:
if inpt_c is None:
raise ValueError("If input is encoded, the conditional input is also needed")
else:
input_x = np.concatenate((inpt_x, inpt_c), axis=1)
inpt_x = self._generator.generate_samples(noise=inpt_x, sess=self._sess)
return self._disc.predict(inpt_x, self._sess)
class CBiGAN(ConditionalGenerativeModel):
def __init__(self,
x_dim,
y_dim,
z_dim,
gen_architecture,
disc_architecture,
enc_architecture,
last_layer_activation,
folder="./Results/CBiGAN_log",
image_shape=None,
append_y_at_every_layer=None):
super(CBiGAN, self).__init__(
x_dim, y_dim, z_dim,
[gen_architecture, disc_architecture, enc_architecture],
last_layer_activation, folder, image_shape,
append_y_at_every_layer)
self._gen_architecture = self._architectures[0]
self._disc_architecture = self._architectures[1]
self._enc_architecture = self._architectures[2]
################# Define architecture
if len(self._x_dim) == 1:
self._gen_architecture.append([
logged_dense, {
"units": x_dim,
"activation": self._last_layer_activation,
"name": "Output"
}
])
else:
self._disc_architecture.append(
[tf.layers.flatten, {
"name": "Flatten"
}])
self._enc_architecture.append(
[tf.layers.flatten, {
"name": "Flatten"
}])
self._gen_architecture[-1][1]["name"] = "Output"
self._disc_architecture.append([
logged_dense, {
"units": 1,
"activation": tf.sigmoid,
"name": "Output"
}
])
self._enc_architecture.append([
logged_dense, {
"units": z_dim,
"activation": tf.identity,
"name": "Output"
}
])
self._generator = ConditionalGenerator(self._gen_architecture,
name="Generator")
self._discriminator = Discriminator(self._disc_architecture,
name="Discriminator")
self._encoder = Encoder(self._enc_architecture, name="Encoder")
self._nets = [self._generator, self._discriminator, self._encoder]
################# Connect inputs and networks
with tf.name_scope("InputsEncoder"):
if len(self._x_dim) == 1:
self._mod_X_input = tf.concat(
axis=1,
values=[self._X_input, self._Y_input],
name="modified_x")
else:
self._mod_X_input = image_condition_concat(
images=self._X_input, condition=self._Y_input, name="fake")
self._output_gen = self._generator.generate_net(
self._mod_Z_input,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
self._output_enc = self._encoder.generate_net(
self._mod_X_input,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
with tf.name_scope("InputsCritic"):
self._mod_output_enc = tf.concat(
axis=1,
values=[self._output_enc, self._Y_input],
name="modified_encoder")
if len(self._x_dim) == 1:
self._disc_input_fake = tf.concat(
axis=1,
values=[self._output_gen, self._mod_Z_input],
name="fake")
self._disc_input_real = tf.concat(
axis=1,
values=[self._X_input, self._mod_output_enc],
name="real")
else:
self._disc_input_fake = image_condition_concat(
inputs=self._output_gen,
condition=self._mod_Z_input,
name="fake")
self._disc_input_real = image_condition_concat(
inputs=self._X_input,
condition=self._mod_output_enc,
name="real")
self._output_disc_fake = self._discriminator.generate_net(
self._disc_input_fake,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
self._output_disc_real = self._discriminator.generate_net(
self._disc_input_real,
append_elements_at_every_layer=self._append_at_every_layer,
tf_trainflag=self._is_training)
################# Finalize
self._init_folders()
self._verify_init()
def compile(self,
logged_images=None,
logged_labels=None,
learning_rate=0.0003,
learning_rate_gen=None,
learning_rate_disc=None,
optimizer=tf.train.AdamOptimizer):
self._define_loss()
if learning_rate_gen is None:
learning_rate_gen = learning_rate
if learning_rate_disc is None:
learning_rate_disc = learning_rate
with tf.name_scope("Optimizer"):
gen_optimizer = optimizer(learning_rate=learning_rate_gen)
self._gen_optimizer = gen_optimizer.minimize(
self._gen_loss,
var_list=self._get_vars(scope="Generator") +
self._get_vars(scope="Encoder"),
name="Generator")
disc_optimizer = optimizer(learning_rate=learning_rate_disc)
self._disc_optimizer = disc_optimizer.minimize(
self._disc_loss,
var_list=self._get_vars(scope="Discriminator"),
name="Discriminator")
self._summarise(logged_images=logged_images,
logged_labels=logged_labels)
def _define_loss(self):
with tf.name_scope("Loss") as scope:
self._gen_loss = -tf.reduce_mean(
tf.log(self._output_disc_fake + 0.00001) +
tf.log(1.0 - self._output_disc_real + 0.00001))
tf.summary.scalar("Generator_loss", self._gen_loss)
self._disc_loss = -tf.reduce_mean(
tf.log(self._output_disc_real + 0.00001) +
tf.log(1.0 - self._output_disc_fake + 0.00001))
tf.summary.scalar("Discriminator_loss", self._disc_loss)
def train(self,
x_train,
y_train,
x_test=None,
y_test=None,
epochs=100,
batch_size=64,
gen_steps=1,
disc_steps=1,
steps=None,
log_step=3,
gpu_options=None):
if steps is not None:
gen_steps = 1
disc_steps = steps
self._set_up_training(log_step=log_step, gpu_options=gpu_options)
self._set_up_test_train_sample(x_train, y_train, x_test, y_test)
self._log_results(epoch=0, epoch_time=0)
for epoch in range(epochs):
batch_nr = 0
disc_loss_epoch = 0
gen_loss_epoch = 0
start = time.clock()
trained_examples = 0
while trained_examples < len(x_train):
disc_loss_batch, gen_loss_batch = self._optimize(
self._trainset, batch_size, disc_steps, gen_steps)
trained_examples += batch_size
disc_loss_epoch += disc_loss_batch
gen_loss_epoch += gen_loss_batch
epoch_train_time = (time.clock() - start) / 60
disc_loss_epoch = np.round(disc_loss_epoch, 2)
gen_loss_epoch = np.round(gen_loss_epoch, 2)
acc_real = self.get_accuracy(inpt=self._x_test,
inpt_y=self._y_test,
labels=np.ones(len(self._x_test)),
is_encoded=False)
acc_fake = self.get_accuracy(inpt=self._z_test,
inpt_y=self._y_test,
labels=np.zeros(len(self._z_test)),
is_encoded=True)
print(
"Epoch {}: Discriminator: {} ({})\n\t\t\tGenerator: {} ({}).".
format(epoch, disc_loss_epoch, acc_real, gen_loss_epoch,
acc_fake))
if log_step is not None:
self._log(epoch + 1, epoch_train_time)
def _optimize(self, dataset, batch_size, disc_steps, gen_steps):
for i in range(disc_steps):
current_batch_x, current_batch_y = dataset.get_next_batch(
batch_size)
Z_noise = self.sample_noise(n=len(current_batch_x))
_, disc_loss_batch = self._sess.run(
[self._disc_optimizer, self._disc_loss],
feed_dict={
self._X_input: current_batch_x,
self._Y_input: current_batch_y,
self._Z_input: Z_noise,
self._is_training: True
})
for _ in range(gen_steps):
Z_noise = self.sample_noise(n=len(current_batch_x))
_, gen_loss_batch = self._sess.run(
[self._gen_optimizer, self._gen_loss],
feed_dict={
self._X_input: current_batch_x,
self._Y_input: current_batch_y,
self._Z_input: Z_noise,
self._is_training: True
})
return disc_loss_batch, gen_loss_batch
def get_accuracy(self, inpt, inpt_y, labels, is_encoded):
if not is_encoded:
inpt = self._sess.run(self._disc_input_real,
feed_dict={
self._X_input: inpt,
self._Y_input: inpt_y,
self._is_training: False
})
else:
inpt = self._sess.run(self._disc_input_fake,
feed_dict={
self._Z_input: inpt,
self._Y_input: inpt_y,
self._is_training: False
})
return self._discriminator.get_accuracy(inpt, labels, self._sess)
def predict(self, inpt_x, inpt_y, is_encoded):
if not is_encoded:
inpt_x = self._encoder.encode(inpt=inpt_x, sess=self._sess)
inpt_x = self._sess.run(self._mod_X_input,
feed_dict={
self._X_input: inpt_x,
self._Y_input: inpt_y,
self._is_training: False
})
return self._discriminator.predict(inpt_x, self._sess)
def generate_image_from_noise(self, n):
noise = self.sample_noise(n=n)
return self.generate_samples(noise, self._sess)
def __init__(self,
x_dim,
z_dim,
gen_architecture,
disc_architecture,
enc_architecture,
last_layer_activation,
folder="./WGAN",
image_shape=None):
super(BiGAN, self).__init__(
x_dim, z_dim,
[gen_architecture, disc_architecture, enc_architecture],
last_layer_activation, folder, image_shape)
self._gen_architecture = self._architectures[0]
self._disc_architecture = self._architectures[1]
self._enc_architecture = self._architectures[2]
################# Define architecture
self._gen_architecture.append([
logged_dense, {
"units": x_dim,
"activation": self._last_layer_activation,
"name": "Output"
}
])
self._generator = Generator(self._gen_architecture, name="Generator")
self._disc_architecture.append([
logged_dense, {
"units": 1,
"activation": tf.nn.sigmoid,
"name": "Output"
}
])
self._discriminator = Discriminator(self._disc_architecture,
name="Discriminator")
self._enc_architecture.append([
logged_dense, {
"units": z_dim,
"activation": tf.identity,
"name": "Output"
}
])
self._encoder = Encoder(self._enc_architecture, name="Encoder")
self._nets = [self._generator, self._discriminator, self._encoder]
################# Connect inputs and networks
self._output_gen = self._generator.generate_net(self._Z_input)
self._output_enc = self._encoder.generate_net(self._X_input)
with tf.name_scope("InputsDiscriminator"):
self._disc_input_fake = tf.concat(
axis=1, values=[self._output_gen, self._Z_input])
self._disc_input_real = tf.concat(
axis=1, values=[self._X_input, self._output_enc])
self._output_disc_fake = self._discriminator.generate_net(
self._disc_input_fake)
self._output_disc_real = self._discriminator.generate_net(
self._disc_input_real)
################# Finalize
self._init_folders()
self._verify_init()
class BiGAN(GenerativeModel):
def __init__(self,
x_dim,
z_dim,
gen_architecture,
disc_architecture,
enc_architecture,
last_layer_activation,
folder="./WGAN",
image_shape=None):
super(BiGAN, self).__init__(
x_dim, z_dim,
[gen_architecture, disc_architecture, enc_architecture],
last_layer_activation, folder, image_shape)
self._gen_architecture = self._architectures[0]
self._disc_architecture = self._architectures[1]
self._enc_architecture = self._architectures[2]
################# Define architecture
self._gen_architecture.append([
logged_dense, {
"units": x_dim,
"activation": self._last_layer_activation,
"name": "Output"
}
])
self._generator = Generator(self._gen_architecture, name="Generator")
self._disc_architecture.append([
logged_dense, {
"units": 1,
"activation": tf.nn.sigmoid,
"name": "Output"
}
])
self._discriminator = Discriminator(self._disc_architecture,
name="Discriminator")
self._enc_architecture.append([
logged_dense, {
"units": z_dim,
"activation": tf.identity,
"name": "Output"
}
])
self._encoder = Encoder(self._enc_architecture, name="Encoder")
self._nets = [self._generator, self._discriminator, self._encoder]
################# Connect inputs and networks
self._output_gen = self._generator.generate_net(self._Z_input)
self._output_enc = self._encoder.generate_net(self._X_input)
with tf.name_scope("InputsDiscriminator"):
self._disc_input_fake = tf.concat(
axis=1, values=[self._output_gen, self._Z_input])
self._disc_input_real = tf.concat(
axis=1, values=[self._X_input, self._output_enc])
self._output_disc_fake = self._discriminator.generate_net(
self._disc_input_fake)
self._output_disc_real = self._discriminator.generate_net(
self._disc_input_real)
################# Finalize
self._init_folders()
self._verify_init()
def compile(self,
learning_rate=0.0003,
learning_rate_gen=None,
learning_rate_disc=None,
optimizer=tf.train.AdamOptimizer):
self._define_loss()
if learning_rate_gen is None:
learning_rate_gen = learning_rate
if learning_rate_disc is None:
learning_rate_disc = learning_rate
with tf.name_scope("Optimizer"):
gen_optimizer = optimizer(learning_rate=learning_rate_gen)
self._gen_optimizer = gen_optimizer.minimize(
self._gen_loss,
var_list=self._get_vars(scope="Generator") +
self._get_vars(scope="Encoder"),
name="Generator")
disc_optimizer = optimizer(learning_rate=learning_rate_disc)
self._disc_optimizer = disc_optimizer.minimize(
self._disc_loss,
var_list=self._get_vars(scope="Discriminator"),
name="Discriminator")
self._summarise()
def _define_loss(self):
with tf.name_scope("Loss") as scope:
self._gen_loss = -tf.reduce_mean(
tf.log(self._output_disc_fake + 0.00001) +
tf.log(1.0 - self._output_disc_real + 0.00001))
tf.summary.scalar("Generator_loss", self._gen_loss)
self._disc_loss = -tf.reduce_mean(
tf.log(self._output_disc_real + 0.00001) +
tf.log(1.0 - self._output_disc_fake + 0.00001))
tf.summary.scalar("Discriminator_loss", self._disc_loss)
def train(self,
x_train,
x_test=None,
epochs=100,
batch_size=64,
gen_steps=1,
disc_steps=1,
log_step=3):
self._set_up_training(log_step=log_step)
self._set_up_test_train_sample(x_train, x_test)
for epoch in range(epochs):
batch_nr = 0
disc_loss_epoch = 0
gen_loss_epoch = 0
start = time.clock()
trained_examples = 0
while trained_examples < len(x_train):
disc_loss_batch, gen_loss_batch = self._optimize(
self._trainset, batch_size, disc_steps, gen_steps)
trained_examples += batch_size
disc_loss_epoch += disc_loss_batch
gen_loss_epoch += gen_loss_batch
epoch_train_time = (time.clock() - start) / 60
disc_loss_epoch = np.round(disc_loss_epoch, 2)
gen_loss_epoch = np.round(gen_loss_epoch, 2)
acc_real = self.get_accuracy(inpt_x=self._x_test,
labels=np.ones(len(self._x_test)))
acc_fake = np.round(
100 - self.get_accuracy(inpt_x=self._z_test,
labels=np.zeros(len(self._z_test))), 2)
print(
"Epoch {}: Discriminator: {} ({})\n\t\t\tGenerator: {} ({}).".
format(epoch, disc_loss_epoch, acc_real, gen_loss_epoch,
acc_fake))
if log_step is not None:
self._log(epoch, epoch_train_time)
def _optimize(self, dataset, batch_size, disc_steps, gen_steps):
for i in range(disc_steps):
current_batch_x = dataset.get_next_batch(batch_size)
Z_noise = self.sample_noise(n=len(current_batch_x))
_, disc_loss_batch = self._sess.run(
[self._disc_optimizer, self._disc_loss],
feed_dict={
self._X_input: current_batch_x,
self._Z_input: Z_noise
})
for _ in range(gen_steps):
Z_noise = self.sample_noise(n=len(current_batch_x))
_, gen_loss_batch = self._sess.run(
[self._gen_optimizer, self._gen_loss],
feed_dict={
self._X_input: current_batch_x,
self._Z_input: Z_noise
})
return disc_loss_batch, gen_loss_batch
def get_accuracy(self, inpt_x, labels):
if labels[0] == 0:
inpt_image = self._sess.run(self._output_gen,
feed_dict={self._Z_input: inpt_x})
inpt_x = np.concatenate((inpt_image, inpt_x), axis=1)
else:
inpt_encoding = self._encoder.encode(inpt=inpt_x, sess=self._sess)
inpt_x = np.concatenate((inpt_x, inpt_encoding), axis=1)
return self._discriminator.get_accuracy(inpt_x, labels, self._sess)
def predict(self, inpt_x, is_encoded):
if is_encoded:
inpt_image = self._generator.generate_samples(noise=inpt_x,
sess=self._sess)
inpt_x = np.concatenate((inpt_image, inpt_x), axis=1)
else:
inpt_encoding = self._generator.generate_samples(noise=inpt_x,
sess=self._sess)
inpt_x = np.concatenate((inpt_x, inpt_encoding), axis=1)
return self._discriminator.predict(inpt_x, self._sess)
def generate_image_from_noise(self, n):
noise = self.sample_noise(n=n)
return self.generate_samples(noise, self._sess)
def generate_image_from_image(self, inpt_x):
encoding = self._encoder.encode(inpt=inpt_x, sess=self._sess)
return self._generator.generate_samples(encoding, self._sess)
def __init__(self,
x_dim,
z_dim,
dec_architecture,
enc_architecture,
last_layer_activation,
folder="./VAE",
image_shape=None):
super(VAE, self).__init__(x_dim, z_dim,
[dec_architecture, enc_architecture],
last_layer_activation, folder, image_shape)
self._gen_architecture = self._architectures[0]
self._enc_architecture = self._architectures[1]
################# Define architecture
last_layer_mean = [
logged_dense, {
"units": z_dim,
"activation": tf.identity,
"name": "Mean"
}
]
self._encoder_mean = Encoder(self._enc_architecture +
[last_layer_mean],
name="Encoder")
last_layer_std = [
logged_dense, {
"units": z_dim,
"activation": tf.identity,
"name": "Std"
}
]
self._encoder_std = Encoder(self._enc_architecture + [last_layer_std],
name="Encoder")
self._gen_architecture.append([
logged_dense, {
"units": x_dim,
"activation": self._last_layer_activation,
"name": "Output"
}
])
self._decoder = Decoder(self._gen_architecture, name="Generator")
self._nets = [self._decoder, self._encoder_mean]
################# Connect inputs and networks
self._mean_layer = self._encoder_mean.generate_net(self._X_input)
self._std_layer = self._encoder_std.generate_net(self._X_input)
self._output_enc_with_noise = self._mean_layer + tf.exp(
0.5 * self._std_layer) * self._Z_input
self._output_dec = self._decoder.generate_net(
self._output_enc_with_noise)
self._output_dec_from_encoding = self._decoder.generate_net(
self._Z_input)
################# Finalize
self._init_folders()
self._verify_init()
class VAE(GenerativeModel):
def __init__(self,
x_dim,
z_dim,
dec_architecture,
enc_architecture,
last_layer_activation,
folder="./VAE",
image_shape=None):
super(VAE, self).__init__(x_dim, z_dim,
[dec_architecture, enc_architecture],
last_layer_activation, folder, image_shape)
self._gen_architecture = self._architectures[0]
self._enc_architecture = self._architectures[1]
################# Define architecture
last_layer_mean = [
logged_dense, {
"units": z_dim,
"activation": tf.identity,
"name": "Mean"
}
]
self._encoder_mean = Encoder(self._enc_architecture +
[last_layer_mean],
name="Encoder")
last_layer_std = [
logged_dense, {
"units": z_dim,
"activation": tf.identity,
"name": "Std"
}
]
self._encoder_std = Encoder(self._enc_architecture + [last_layer_std],
name="Encoder")
self._gen_architecture.append([
logged_dense, {
"units": x_dim,
"activation": self._last_layer_activation,
"name": "Output"
}
])
self._decoder = Decoder(self._gen_architecture, name="Generator")
self._nets = [self._decoder, self._encoder_mean]
################# Connect inputs and networks
self._mean_layer = self._encoder_mean.generate_net(self._X_input)
self._std_layer = self._encoder_std.generate_net(self._X_input)
self._output_enc_with_noise = self._mean_layer + tf.exp(
0.5 * self._std_layer) * self._Z_input
self._output_dec = self._decoder.generate_net(
self._output_enc_with_noise)
self._output_dec_from_encoding = self._decoder.generate_net(
self._Z_input)
################# Finalize
self._init_folders()
self._verify_init()
def compile(self, learning_rate=0.0001, optimizer=tf.train.AdamOptimizer):
self._define_loss()
with tf.name_scope("Optimizer"):
vae_optimizer = optimizer(learning_rate=learning_rate)
self._vae_optimizer = vae_optimizer.minimize(self._vae_loss,
name="VAE")
self._summarise()
def _define_loss(self):
with tf.name_scope("Loss") as scope:
self._data_fidelity_loss = self._X_input * tf.log(
1e-10 + self._output_dec) + (
1 - self._X_input) * tf.log(1e-10 + 1 - self._output_dec)
self._data_fidelity_loss = -tf.reduce_sum(self._data_fidelity_loss,
1)
self._KLdiv = 0.5 * (tf.square(self._mean_layer) +
tf.exp(self._std_layer) - self._std_layer - 1)
self._KLdiv = tf.reduce_sum(self._KLdiv, 1)
self._vae_loss = tf.reduce_mean(self._data_fidelity_loss +
self._KLdiv)
tf.summary.scalar("Loss", self._vae_loss)
def train(self,
x_train,
x_test,
epochs=100,
batch_size=64,
steps=5,
log_step=3):
self._set_up_training(log_step=log_step)
self._set_up_test_train_sample(x_train, x_test)
for epoch in range(epochs):
batch_nr = 0
loss_epoch = 0
start = time.clock()
trained_examples = 0
while trained_examples < len(x_train):
loss_batch = self._optimize(self._trainset, batch_size, steps)
trained_examples += batch_size
loss_epoch += loss_batch
epoch_train_time = (time.clock() - start) / 60
loss_epoch = np.round(loss_epoch, 2)
print("Epoch {}: Loss: {}.".format(epoch, loss_epoch))
if log_step is not None:
self._log(epoch, epoch_train_time)
def _optimize(self, dataset, batch_size, steps):
for i in range(steps):
current_batch_x = dataset.get_next_batch(batch_size)
Z_noise = self._decoder.sample_noise(n=len(current_batch_x))
_, loss_batch = self._sess.run(
[self._vae_optimizer, self._vae_loss],
feed_dict={
self._X_input: current_batch_x,
self._Z_input: Z_noise
})
return loss_batch
def decode(self, inpt_x, is_encoded):
if not is_encoded:
inpt_x = self._encoder_mean.encode(noise=inpt_x, sess=self._sess)
return self._decoder.decode(inpt_x, self._sess)
def generate_image_from_noise(self, n):
noise = self.sample_noise(n=n)
return self._decoder.generate_samples(noise, self._sess)
def generate_image_from_image(self, inpt_x):
encoding = self._encoder.encode(inpt=inpt_x, sess=self._sess)
return self._decoder.generate_samples(encoding, self._sess)