def get_v(n):
ret = tf.get_variable(n + '_unused',
[param.batch_size, param.rnn_size],
trainable=False,
initializer=tf.constant_initializer())
ret = symbolic_functions.shapeless_placeholder(ret, 0, name=n)
return ret
def _build_graph(self, inputs):
image_pos, y = inputs
image_pos = tf.expand_dims(image_pos * 2.0 - 1, -1)
y = tf.one_hot(y, 10, name='label_onehot')
z = tf.random_uniform([BATCH, 100], -1, 1, name='z_train')
z = symbf.shapeless_placeholder(z, [0], name='z')
with argscope([Conv2D, Deconv2D, FullyConnected],
W_init=tf.truncated_normal_initializer(stddev=0.02)):
with tf.variable_scope('gen'):
image_gen = self.generator(z, y)
tf.summary.image('gen', image_gen, 30)
with tf.variable_scope('discrim'):
vecpos = self.discriminator(image_pos, y)
vecneg = self.discriminator(image_gen, y)
self.build_losses(vecpos, vecneg)
self.collect_variables()
def _build_graph(self, inputs):
image_pos, y = inputs
image_pos = tf.expand_dims(image_pos * 2.0 - 1, -1)
y = tf.one_hot(y, 10, name='label_onehot')
z = tf.random_uniform([BATCH, 100], -1, 1, name='z_train')
z = symbf.shapeless_placeholder(z, [0], name='z')
with argscope([Conv2D, Deconv2D, FullyConnected],
W_init=tf.truncated_normal_initializer(stddev=0.02)):
with tf.variable_scope('gen'):
image_gen = self.generator(z, y)
tf.summary.image('gen', image_gen, 30)
with tf.variable_scope('discrim'):
vecpos = self.discriminator(image_pos, y)
vecneg = self.discriminator(image_gen, y)
self.build_losses(vecpos, vecneg)
self.collect_variables()
def _build_graph(self, inputs):
real_sample = inputs[0]
real_sample = tf.expand_dims(real_sample, -1)
# sample the latent code:
zc = symbf.shapeless_placeholder(sample_prior(BATCH), 0, name='z_code')
z_noise = symbf.shapeless_placeholder(tf.random_uniform(
[BATCH, NOISE_DIM], -1, 1),
0,
name='z_noise')
z = tf.concat([zc, z_noise], 1, name='z')
with argscope([Conv2D, Deconv2D, FullyConnected],
W_init=tf.truncated_normal_initializer(stddev=0.02)):
with tf.variable_scope('gen'):
fake_sample = self.generator(z)
fake_sample_viz = tf.cast((fake_sample) * 255.0,
tf.uint8,
name='viz')
tf.summary.image('gen', fake_sample_viz, max_outputs=30)
# may need to investigate how bn stats should be updated across two discrim
with tf.variable_scope('discrim'):
real_pred, _ = self.discriminator(real_sample)
fake_pred, dist_param = self.discriminator(fake_sample)
"""
Mutual information between x (i.e. zc in this case) and some
information s (the generated samples in this case):
I(x;s) = H(x) - H(x|s)
= H(x) + E[\log P(x|s)]
The distribution from which zc is sampled, in this case, is set to a fixed prior already.
So the first term is a constant.
For the second term, we can maximize its variational lower bound:
E_{x \sim P(x|s)}[\log Q(x|s)]
where Q(x|s) is a proposal distribution to approximate P(x|s).
Here, Q(x|s) is assumed to be a distribution which shares the form
of P, and whose parameters are predicted by the discriminator network.
"""
with tf.name_scope("mutual_information"):
with tf.name_scope('prior_entropy'):
cat, uni = get_distributions(DIST_PRIOR_PARAM[:NUM_CLASS],
DIST_PRIOR_PARAM[NUM_CLASS:])
ents = [
cat.entropy(name='cat_entropy'),
tf.reduce_sum(uni.entropy(), name='uni_entropy')
]
entropy = tf.add_n(ents, name='total_entropy')
# Note that the entropy of prior is a constant. The paper mentioned it but didn't use it.
with tf.name_scope('conditional_entropy'):
cond_ents = entropy_from_samples(zc, dist_param)
cond_entropy = tf.add_n(cond_ents, name="total_entropy")
MI = tf.subtract(entropy, cond_entropy, name='mutual_information')
summary.add_moving_summary(entropy, cond_entropy, MI, *cond_ents)
# default GAN objective
self.build_losses(real_pred, fake_pred)
# subtract mutual information for latent factors (we want to maximize them)
self.g_loss = tf.subtract(self.g_loss, MI, name='total_g_loss')
self.d_loss = tf.subtract(self.d_loss, MI, name='total_d_loss')
summary.add_moving_summary(self.g_loss, self.d_loss)
# distinguish between variables of generator and discriminator updates
self.collect_variables()
def _build_graph(self, inputs):
real_sample = inputs[0]
real_sample = tf.expand_dims(real_sample * 2.0 - 1, -1)
# latent space is cat(10) x uni(1) x uni(1) x noise(NOISE_DIM)
self.factors = ProductDistribution("factors", [
CategoricalDistribution("cat", 10),
GaussianWithUniformSample("uni_a", 1),
GaussianWithUniformSample("uni_b", 1)
])
# prior: the assumption how the factors are presented in the dataset
prior = tf.constant([0.1] * 10 + [0, 0],
tf.float32, [12],
name='prior')
batch_prior = tf.tile(tf.expand_dims(prior, 0), [BATCH, 1],
name='batch_prior')
# sample the latent code:
zc = symbf.shapeless_placeholder(self.factors.sample(BATCH, prior),
0,
name='z_code')
z_noise = symbf.shapeless_placeholder(tf.random_uniform(
[BATCH, NOISE_DIM], -1, 1),
0,
name='z_noise')
z = tf.concat([zc, z_noise], 1, name='z')
with argscope([Conv2D, Deconv2D, FullyConnected],
W_init=tf.truncated_normal_initializer(stddev=0.02)):
with tf.variable_scope('gen'):
fake_sample = self.generator(z)
fake_sample_viz = tf.cast((fake_sample + 1) * 128.0,
tf.uint8,
name='viz')
tf.summary.image('gen', fake_sample_viz, max_outputs=30)
# TODO investigate how bn stats should be updated across two discrim
with tf.variable_scope('discrim'):
real_pred, _ = self.discriminator(real_sample)
with tf.variable_scope('discrim', reuse=True):
fake_pred, dist_param = self.discriminator(fake_sample)
# post-process output vector from discriminator to become valid
# distribution parameters
encoder_activation = self.factors.encoder_activation(dist_param)
"""
Mutual information between x (i.e. zc in this case) and some
information s (the generated samples in this case):
I(x;s) = H(x) - H(x|s)
= H(x) + E[\log P(x|s)]
The distribution from which zc is sampled, in this case, is set to a fixed prior already.
For the second term, we can maximize its variational lower bound:
E_{x \sim P(x|s)}[\log Q(x|s)]
where Q(x|s) is a proposal distribution to approximate P(x|s).
Here, Q(x|s) is assumed to be a distribution which shares the form
of self.factors, and whose parameters are predicted by the discriminator network.
"""
with tf.name_scope("mutual_information"):
ents = self.factors.entropy(zc, batch_prior)
entropy = tf.add_n(ents, name='total_entropy')
# Note that dropping this term has no effect because the entropy
# of prior is a constant. The paper mentioned it but didn't use it.
# Adding this term may make the curve less stable because the
# entropy estimated from the samples is not the true value.
cond_ents = self.factors.entropy(zc, encoder_activation)
cond_entropy = tf.add_n(cond_ents,
name="total_conditional_entropy")
MI = tf.subtract(entropy, cond_entropy, name='mutual_information')
summary.add_moving_summary(entropy, cond_entropy, MI, *ents)
# default GAN objective
self.build_losses(real_pred, fake_pred)
# subtract mutual information for latent factores (we want to maximize them)
self.g_loss = tf.subtract(self.g_loss, MI, name='total_g_loss')
self.d_loss = tf.subtract(self.d_loss, MI, name='total_d_loss')
summary.add_moving_summary(self.g_loss, self.d_loss)
# distinguish between variables of generator and discriminator updates
self.collect_variables()
def _build_graph(self, inputs):
real_sample = inputs[0]
real_sample = tf.expand_dims(real_sample, -1)
# latent space is cat(10) x uni(1) x uni(1) x noise(NOISE_DIM)
self.factors = ProductDistribution("factors", [CategoricalDistribution("cat", 10),
GaussianWithUniformSample("uni_a", 1),
GaussianWithUniformSample("uni_b", 1)])
# prior: the assumption how the factors are presented in the dataset
prior = tf.constant([0.1] * 10 + [0, 0], tf.float32, [12], name='prior')
batch_prior = tf.tile(tf.expand_dims(prior, 0), [BATCH, 1], name='batch_prior')
# sample the latent code:
zc = symbf.shapeless_placeholder(
self.factors.sample(BATCH, prior), 0, name='z_code')
z_noise = symbf.shapeless_placeholder(
tf.random_uniform([BATCH, NOISE_DIM], -1, 1), 0, name='z_noise')
z = tf.concat([zc, z_noise], 1, name='z')
with argscope([Conv2D, Deconv2D, FullyConnected],
W_init=tf.truncated_normal_initializer(stddev=0.02)):
with tf.variable_scope('gen'):
fake_sample = self.generator(z)
fake_sample_viz = tf.cast((fake_sample) * 255.0, tf.uint8, name='viz')
tf.summary.image('gen', fake_sample_viz, max_outputs=30)
# may need to investigate how bn stats should be updated across two discrim
with tf.variable_scope('discrim'):
real_pred, _ = self.discriminator(real_sample)
with tf.variable_scope('discrim', reuse=True):
fake_pred, dist_param = self.discriminator(fake_sample)
"""
Mutual information between x (i.e. zc in this case) and some
information s (the generated samples in this case):
I(x;s) = H(x) - H(x|s)
= H(x) + E[\log P(x|s)]
The distribution from which zc is sampled, in this case, is set to a fixed prior already.
For the second term, we can maximize its variational lower bound:
E_{x \sim P(x|s)}[\log Q(x|s)]
where Q(x|s) is a proposal distribution to approximate P(x|s).
Here, Q(x|s) is assumed to be a distribution which shares the form
of self.factors, and whose parameters are predicted by the discriminator network.
"""
with tf.name_scope("mutual_information"):
ents = self.factors.entropy(zc, batch_prior)
entropy = tf.add_n(ents, name='total_entropy')
# Note that dropping this term has no effect because the entropy
# of prior is a constant. The paper mentioned it but didn't use it.
# Adding this term may make the curve less stable because the
# entropy estimated from the samples is not the true value.
# post-process output vector from discriminator to obtain valid distribution parameters
encoder_activation = self.factors.encoder_activation(dist_param)
cond_ents = self.factors.entropy(zc, encoder_activation)
cond_entropy = tf.add_n(cond_ents, name="total_conditional_entropy")
MI = tf.subtract(entropy, cond_entropy, name='mutual_information')
summary.add_moving_summary(entropy, cond_entropy, MI, *ents)
# default GAN objective
self.build_losses(real_pred, fake_pred)
# subtract mutual information for latent factors (we want to maximize them)
self.g_loss = tf.subtract(self.g_loss, MI, name='total_g_loss')
self.d_loss = tf.subtract(self.d_loss, MI, name='total_d_loss')
summary.add_moving_summary(self.g_loss, self.d_loss)
# distinguish between variables of generator and discriminator updates
self.collect_variables()
