def q_net(x, n_h, n_particles):
with zs.BayesianNet() as proposal:
h1_logits = layers.fully_connected(tf.to_float(x), n_h,
activation_fn=None)
h1 = zs.Bernoulli('h1', h1_logits, n_samples=n_particles,
group_ndims=1, dtype=tf.float32)
h2_logits = layers.fully_connected(h1, n_h, activation_fn=None)
h2 = zs.Bernoulli('h2', h2_logits, group_ndims=1, dtype=tf.float32)
h3_logits = layers.fully_connected(h2, n_h, activation_fn=None)
h3 = zs.Bernoulli('h3', h3_logits, group_ndims=1, dtype=tf.float32)
return proposal
def sbn(observed, n, n_x, n_h, n_particles):
with zs.BayesianNet(observed=observed) as model:
h3_logits = tf.zeros([n, n_h])
h3 = zs.Bernoulli('h3', h3_logits, n_samples=n_particles,
group_ndims=1, dtype=tf.float32)
h2_logits = layers.fully_connected(h3, n_h, activation_fn=None)
h2 = zs.Bernoulli('h2', h2_logits, group_ndims=1, dtype=tf.float32)
h1_logits = layers.fully_connected(h2, n_h, activation_fn=None)
h1 = zs.Bernoulli('h1', h1_logits, group_ndims=1, dtype=tf.float32)
x_logits = layers.fully_connected(h1, n_x, activation_fn=None)
x = zs.Bernoulli('x', x_logits, group_ndims=1)
return model
def q_net(x, n_h, n_particles):
with zs.BayesianNet() as variational:
h1_logits = tf.layers.dense(tf.to_float(x), n_h)
h1 = zs.Bernoulli('h1',
h1_logits,
n_samples=n_particles,
group_ndims=1,
dtype=tf.float32)
h2_logits = tf.layers.dense(h1, n_h)
h2 = zs.Bernoulli('h2', h2_logits, group_ndims=1, dtype=tf.float32)
h3_logits = tf.layers.dense(h2, n_h)
h3 = zs.Bernoulli('h3', h3_logits, group_ndims=1, dtype=tf.float32)
return variational
def sbn(observed, n, n_x, n_h, n_particles):
with zs.BayesianNet(observed=observed) as model:
h3_logits = tf.zeros([n, n_h])
h3 = zs.Bernoulli('h3',
h3_logits,
n_samples=n_particles,
group_ndims=1,
dtype=tf.float32)
h2_logits = tf.layers.dense(h3, n_h)
h2 = zs.Bernoulli('h2', h2_logits, group_ndims=1, dtype=tf.float32)
h1_logits = tf.layers.dense(h2, n_h)
h1 = zs.Bernoulli('h1', h1_logits, group_ndims=1, dtype=tf.float32)
x_logits = tf.layers.dense(h1, n_x)
x = zs.Bernoulli('x', x_logits, group_ndims=1)
return model
def vae_conv(observed, n, n_x, n_z, n_particles, is_training):
with zs.BayesianNet(observed=observed) as model:
normalizer_params = {'is_training': is_training,
'updates_collections': None}
z_mean = tf.zeros([n, n_z])
z = zs.Normal('z', z_mean, std=1., n_samples=n_particles,
group_event_ndims=1)
lx_z = tf.reshape(z, [-1, 1, 1, n_z])
lx_z = layers.conv2d_transpose(
lx_z, 128, kernel_size=3, padding='VALID',
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params)
lx_z = layers.conv2d_transpose(
lx_z, 64, kernel_size=5, padding='VALID',
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params)
lx_z = layers.conv2d_transpose(
lx_z, 32, kernel_size=5, stride=2,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params)
lx_z = layers.conv2d_transpose(
lx_z, 1, kernel_size=5, stride=2,
activation_fn=None)
x_logits = tf.reshape(lx_z, [n_particles, n, -1])
x = zs.Bernoulli('x', x_logits, group_event_ndims=1)
return model
def vae(observed, n, n_x, n_z, n_samples, is_training):
with zs.BayesianNet(observed=observed) as model:
normalizer_params = {'is_training': is_training,
'updates_collections': None}
z_mean = tf.zeros([n, n_z]) # the mean of z is the zero vector
z_std = 1. # the covariance of z is the identity matrix, here a scalar is sufficient because it will be broadcasted to a vector and then used as the diagonal of the covariance matrix in zhusuan
'''
TODO1: sampling z using the Gaussian distribution of zhusuan
> given input
- z_mean, z_std, n_samples
- set group_event_ndims as 1
> e.g.
- x = zs.Bernoulli('x', mu, group_event_ndims=1)
'''
lx_z_1 = layers.fully_connected(
z, 500, normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params) # a mlp layer of 500 hidden units with z as the input
'''
TODO2: add one more mlp layer with 500 hidden units
> given input
- lx_z_1, size of hidden units, normalizer_params
- add batch_norm as the normalizer_fn
> e.g.
see the above line
'''
x_logits = layers.fully_connected(lx_z_2, n_x, activation_fn=None)
x = zs.Bernoulli('x', x_logits, group_event_ndims=1)
return model
def vae(observed, n, x_dim, z_dim, n_particles, is_training):
with zs.BayesianNet(observed=observed) as model:
z_logits = tf.zeros([n, z_dim])
z = zs.Bernoulli('z',
z_logits,
group_ndims=1,
n_samples=n_particles,
dtype=tf.float32)
lx_z = tf.layers.dense(z, 500, use_bias=False)
lx_z = tf.layers.batch_normalization(lx_z, training=is_training)
lx_z = tf.nn.relu(lx_z)
lx_z = tf.layers.dense(lx_z, 500, use_bias=False)
lx_z = tf.layers.batch_normalization(lx_z, training=is_training)
lx_z = tf.nn.relu(lx_z)
x_logits = tf.layers.dense(lx_z, x_dim)
x = zs.Bernoulli('x', x_logits, group_ndims=1)
return model
def vae(observed, n, n_x, n_z):
with zs.BayesianNet(observed=observed) as model:
z_mean = tf.zeros([n, n_z])
z = zs.Normal('z', z_mean, std=1., group_event_ndims=1)
lx_z = layers.fully_connected(z, 500)
lx_z = layers.fully_connected(lx_z, 500)
x_logits = layers.fully_connected(lx_z, n_x, activation_fn=None)
x = zs.Bernoulli('x', x_logits, group_event_ndims=1)
return model, x_logits
def q_net(x, n_h, n_particles, is_training):
with zs.BayesianNet() as variational:
h1_logits = layers.fully_connected(tf.to_float(x),
n_h,
activation_fn=None)
h1 = zs.Bernoulli('h1',
h1_logits,
n_samples=n_particles,
group_event_ndims=1)
h2_logits = layers.fully_connected(tf.to_float(h1),
n_h,
activation_fn=None)
h2 = zs.Bernoulli('h2', h2_logits, group_event_ndims=1)
h3_logits = layers.fully_connected(tf.to_float(h2),
n_h,
activation_fn=None)
h3 = zs.Bernoulli('h3', h3_logits, group_event_ndims=1)
return variational
def vae(observed, x_dim, z_dim, n, n_particles=1):
with zs.BayesianNet(observed=observed) as model:
z_mean = tf.zeros([n, z_dim])
z = zs.Normal('z', z_mean, std=1., group_ndims=1, n_samples=n_particles)
lx_z = tf.layers.dense(z, 500, activation=tf.nn.relu)
lx_z = tf.layers.dense(lx_z, 500, activation=tf.nn.relu)
x_logits = tf.layers.dense(lx_z, x_dim)
x_mean = zs.Implicit("x_mean", tf.sigmoid(x_logits), group_ndims=1)
x = zs.Bernoulli('x', x_logits, group_ndims=1)
return model
def vae(observed, n, n_x, n_z, n_particles, is_training):
with zs.BayesianNet(observed=observed) as model:
normalizer_params = {
'is_training': is_training,
'updates_collections': None
}
z_logits = tf.zeros([n, n_z])
z = zs.Bernoulli('z', z_logits, n_samples=n_particles, group_ndims=1)
lx_z = layers.fully_connected(tf.to_float(z),
500,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params)
lx_z = layers.fully_connected(lx_z,
500,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params)
x_logits = layers.fully_connected(lx_z, n_x, activation_fn=None)
x = zs.Bernoulli('x', x_logits, group_ndims=1)
return model
def vae(observed, x_dim, z_dim, n, y):
with zs.BayesianNet(observed=observed) as model:
z_mean = tf.zeros([n, z_dim])
z = zs.Normal('z', z_mean, std=1., group_ndims=1)
z_plus_y = tf.concat([z, tf.cast(y, tf.float32)], axis=1)
lx_z = tf.layers.dense(z_plus_y, 500, activation=tf.nn.relu)
lx_z = tf.layers.dense(lx_z, 500, activation=tf.nn.relu)
x_logits = tf.layers.dense(lx_z, x_dim)
x_mean = zs.Implicit("x_mean", tf.sigmoid(x_logits), group_ndims=1)
x = zs.Bernoulli('x', x_logits, group_ndims=1)
return model, x_mean
def sbn(observed, n, n_x, n_h, n_particles, is_training):
with zs.BayesianNet(observed=observed) as model:
h3_logits = tf.zeros([n, n_h])
h3 = zs.Bernoulli('h3',
h3_logits,
n_samples=n_particles,
group_event_ndims=1)
h2_logits = layers.fully_connected(tf.to_float(h3),
n_h,
activation_fn=None)
h2 = zs.Bernoulli('h2', h2_logits, group_event_ndims=1)
h1_logits = layers.fully_connected(tf.to_float(h2),
n_h,
activation_fn=None)
h1 = zs.Bernoulli('h1', h1_logits, group_event_ndims=1)
x_logits = layers.fully_connected(tf.to_float(h1),
n_x,
activation_fn=None)
x = zs.Bernoulli('x', x_logits, group_event_ndims=1)
return model
def p_net(observed, n, x_dim, z_dim):
'''
Decoder: p(x|z)
'''
with zs.BayesianNet(observed=observed) as model:
z_mean = tf.zeros([n, z_dim])
z = zs.Normal('z', z_mean, std=1., group_ndims=1)
lx_z = layers.fully_connected(z, 500)
lx_z = layers.fully_connected(lx_z, 500)
x_logits = layers.fully_connected(lx_z, x_dim, activation_fn=None)
x = zs.Bernoulli('x', x_logits, group_ndims=1)
return model
def M2(observed, n, n_x, n_y, n_z, n_particles):
with zs.BayesianNet(observed=observed) as model:
z_mean = tf.zeros([n, n_z])
z = zs.Normal('z', z_mean, std=1., n_samples=n_particles,
group_ndims=1)
y_logits = tf.zeros([n, n_y])
y = zs.OnehotCategorical('y', y_logits, n_samples=n_particles)
lx_zy = layers.fully_connected(tf.concat([z, tf.to_float(y)], 2), 500)
lx_zy = layers.fully_connected(lx_zy, 500)
x_logits = layers.fully_connected(lx_zy, n_x, activation_fn=None)
x = zs.Bernoulli('x', x_logits, group_ndims=1)
return model
def vae(observed, n, x_dim, z_dim, n_particles):
with zs.BayesianNet(observed=observed) as model:
z_mean = tf.zeros([n, z_dim])
z = zs.Normal('z',
z_mean,
std=1.,
group_ndims=1,
n_samples=n_particles)
lx_z = layers.fully_connected(z, 500)
lx_z = layers.fully_connected(lx_z, 500)
x_logits = layers.fully_connected(lx_z, x_dim, activation_fn=None)
x = zs.Bernoulli('x', x_logits, group_ndims=1)
return model, x_logits
def wae(observed, n, n_x, n_z, n_particles):
with zs.BayesianNet(observed=observed) as model:
z_mean = tf.zeros([n, n_z])
z = zs.Normal('z',
z_mean,
std=1.,
group_ndims=1,
n_samples=n_particles)
lx_z = tf.layers.dense(z, units=256, activation=tf.nn.relu)
lx_z = tf.layers.dense(lx_z, units=256, activation=tf.nn.relu)
x = tf.layers.dense(lx_z, units=784, activation=None)
zs.Bernoulli('x', x, group_ndims=1, dtype=tf.int32)
return model, z, tf.sigmoid(x)
def vae(observed, n, n_x, n_z, n_particles):
with zs.BayesianNet(observed=observed) as model:
z_mean = tf.zeros([n, n_z])
z = zs.Normal('z',
z_mean,
std=1.,
n_samples=n_particles,
group_ndims=1)
lx_z = tf.layers.dense(z, 500, activation=tf.nn.relu)
lx_z = tf.layers.dense(lx_z, 500, activation=tf.nn.relu)
x_logits = tf.layers.dense(lx_z, n_x)
x = zs.Bernoulli('x', x_logits, group_ndims=1)
return model
def q_net(x, z_dim, n_particles, is_training):
with zs.BayesianNet() as variational:
lz_x = tf.layers.dense(tf.to_float(x), 500, use_bias=False)
lz_x = tf.layers.batch_normalization(lz_x, training=is_training)
lz_x = tf.nn.relu(lz_x)
lz_x = tf.layers.dense(lz_x, 500, use_bias=False)
lz_x = tf.layers.batch_normalization(lz_x, training=is_training)
lz_x = tf.nn.relu(lz_x)
z_logits = tf.layers.dense(lz_x, z_dim)
z = zs.Bernoulli('z',
z_logits,
group_ndims=1,
n_samples=n_particles,
dtype=tf.float32)
return variational
def cvae(observed, x_dim, y_dim, z_dim, n, n_particles=1):
with zs.BayesianNet(observed=observed) as model:
y = zs.Empirical('y', tf.int32, (n, y_dim))
z_mean = tf.zeros([n, z_dim])
z = zs.Normal('z',
z_mean,
std=1.,
group_ndims=1,
n_samples=n_particles)
z = tf.to_float(z[0])
yz = tf.concat([tf.to_float(y), z], axis=1)
lx_yz = tf.layers.dense(tf.to_float(yz), 500, activation=tf.nn.relu)
lx_yz = tf.layers.dense(lx_yz, 500, activation=tf.nn.relu)
x_logits = tf.layers.dense(lx_yz, x_dim)
x_mean = zs.Implicit('x_mean', tf.sigmoid(x_logits), group_ndims=1)
x = zs.Bernoulli('x', logits=x_logits, group_ndims=1)
return model
def vae(observed, n, n_x, n_z, n_particles):
with zs.BayesianNet(observed=observed) as model:
z_mean = tf.zeros([n, n_z])
z = zs.Normal('z', z_mean, std=1., group_ndims=1,
n_samples=n_particles)
lx_z = layers.fully_connected(z, 7 * 7 * ngf * 2)
lx_z = tf.reshape(lx_z, [-1, 7, 7, ngf * 2])
lx_z = deconv_res_block(lx_z, [7, 7, ngf * 2])
lx_z = deconv_res_block(lx_z, [14, 14, ngf * 2], resize=True)
lx_z = deconv_res_block(lx_z, [14, 14, ngf * 2])
lx_z = deconv_res_block(lx_z, [28, 28, ngf], resize=True)
lx_z = deconv_res_block(lx_z, [28, 28, ngf])
lx_z = conv2d_transpose(lx_z, [28, 28, 1], kernel_size=(3, 3),
stride=(1, 1), activation_fn=None)
x_logits = tf.reshape(lx_z, [n_particles, -1, n_x])
x = zs.Bernoulli('x', x_logits, group_ndims=1)
return model, x_logits
def q_net(x, n_z, n_particles, is_training):
with zs.BayesianNet() as variational:
normalizer_params = {
'is_training': is_training,
'updates_collections': None
}
lz_x = layers.fully_connected(tf.to_float(x),
500,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params)
lz_x = layers.fully_connected(lz_x,
500,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params)
z_logits = layers.fully_connected(lz_x, n_z, activation_fn=None)
z = zs.Bernoulli('z', z_logits, n_samples=n_particles, group_ndims=1)
return variational
def vae(observed, x_dim, z_dim, n, n_particles=1):
with zs.BayesianNet(observed=observed) as model:
y = zs.OnehotCategorical('y',
logits=tf.ones([n, 10]),
group_ndims=1,
n_samples=n_particles)
y = tf.to_float(y)
y = tf.reshape(y, (1, n, 10))
z_mean = tf.zeros([n, z_dim])
z = zs.Normal('z',
z_mean,
std=1.,
group_ndims=1,
n_samples=n_particles)
z = tf.concat([z, y], 2)
lx_z = tf.layers.dense(z, 500, activation=tf.nn.relu)
lx_z = tf.layers.dense(lx_z, 500, activation=tf.nn.relu)
x_logits = tf.layers.dense(lx_z, x_dim)
x_mean = zs.Implicit("x_mean", tf.sigmoid(x_logits), group_ndims=1)
x = zs.Bernoulli('x', x_logits, group_ndims=1)
return model
def vae(observed, n, n_x, n_h, n_z):
with zs.BayesianNet(observed=observed) as model:
pi = tf.get_variable(name='pi',
dtype=tf.float32,
initializer=tf.truncated_normal([1, n_z],
stddev=0.1))
zpi = tf.tile(pi, [n, 1])
# zpi = tf.zeros([n, n_z])
z = zs.OnehotDiscrete('z', zpi, group_ndims=0)
h_mean = layers.fully_connected(tf.to_float(z),
n_h,
activation_fn=None)
h_logstd = layers.fully_connected(tf.to_float(z),
n_h,
activation_fn=None)
h = zs.Normal('h', h_mean, logstd=h_logstd, group_ndims=1)
lx_h = layers.fully_connected(h, 512)
lx_h = layers.fully_connected(lx_h, 512)
x_logits = layers.fully_connected(lx_h, n_x, activation_fn=None)
x = zs.Bernoulli('x', x_logits, group_ndims=1)
return model, x_logits, pi
def vae(observed, n, y, is_training):
with zs.BayesianNet(observed=observed) as decoder:
normalizer_params = {'is_training': is_training,
'updates_collections': None}
z_mean = tf.zeros([n, n_z])
z = zs.Normal('z', z_mean, std=1., n_samples=1, group_event_ndims=1)
z = tf.reshape(z,[-1 ,n_z])
yb = tf.reshape(y, [-1, 1, 1, n_code])
lx_z = layers.fully_connected(tf.concat([z, y], 1), 1024, activation_fn=tf.nn.relu,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params,
weights_initializer=init_ops.RandomNormal(stddev=0.02)
)
lx_z = layers.fully_connected(tf.concat([lx_z, y], 1), ngf * 8 * 4 * 4, activation_fn=tf.nn.relu,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params,
weights_initializer=init_ops.RandomNormal(stddev=0.02)
)
lx_z = tf.reshape(lx_z, [-1, 4, 4, ngf * 8])
# assert tf.shape(lx_z)[0] == n
lx_z = layers.conv2d_transpose(conv_cond_concat(lx_z, yb), ngf * 4, 5, stride=2, activation_fn=tf.nn.relu,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params,
weights_initializer=init_ops.RandomNormal(stddev=0.02))
lx_z = layers.conv2d_transpose(conv_cond_concat(lx_z, yb), ngf * 2, 5, stride=2, activation_fn=tf.nn.relu,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params,
weights_initializer=init_ops.RandomNormal(stddev=0.02))
lx_z = layers.conv2d_transpose(conv_cond_concat(lx_z, yb), ngf * 1, 5, stride=2, activation_fn=tf.nn.relu,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params,
weights_initializer=init_ops.RandomNormal(stddev=0.02))
lx_z = layers.conv2d_transpose(conv_cond_concat(lx_z, yb), 1, 5, stride=2, activation_fn=None,
weights_initializer=init_ops.RandomNormal(stddev=0.02))
x_logits = tf.reshape(lx_z, [1, -1, n_x])
x = zs.Bernoulli('x', x_logits, group_event_ndims=1)
return decoder, x_logits
def vae_conv(observed, n, x_dim, z_dim, n_particles, nf=16):
with zs.BayesianNet(observed=observed) as model:
z_mean = tf.zeros([n, z_dim])
z = zs.Normal("z",
z_mean,
std=1.,
group_ndims=1,
n_samples=n_particles)
lx_z = tf.layers.dense(z, 7 * 7 * nf * 2, activation=tf.nn.relu)
lx_z = tf.reshape(lx_z, [-1, 7, 7, nf * 2])
lx_z = deconv_resnet_block(lx_z, [7, 7, nf * 2])
lx_z = deconv_resnet_block(lx_z, [14, 14, nf * 2], resize=True)
lx_z = deconv_resnet_block(lx_z, [14, 14, nf * 2])
lx_z = deconv_resnet_block(lx_z, [28, 28, nf], resize=True)
lx_z = deconv_resnet_block(lx_z, [28, 28, nf])
lx_z = conv2d_transpose(lx_z, [28, 28, 1],
kernel_size=(3, 3),
stride=(1, 1),
activation_fn=None)
x_logits = tf.reshape(lx_z, [n_particles, -1, x_dim])
x = zs.Bernoulli("x", x_logits, group_ndims=1)
return model, x_logits
def vae(observed, batch_size, n_x, n_h, n_z, n_particles):
with zs.BayesianNet(observed=observed) as model:
log_pi = tf.get_variable('log_pi',
n_z,
initializer=tf.zeros_initializer())
mu = tf.get_variable('mu', [n_h, n_z],
initializer=tf.random_uniform_initializer(-1, 1))
log_sigma = tf.get_variable('log_sigma', [n_h, n_z],
initializer=tf.random_normal_initializer(
0, 0.1))
n_log_pi = tf.tile(tf.expand_dims(log_pi, 0),
[batch_size, 1]) # (batch_size, n_z)
z = zs.OnehotCategorical(
'z', n_log_pi, n_samples=n_particles,
group_event_ndims=0) # (n_particles, batch_size, n_z)
z_tensor = tf.reshape(z.tensor, [-1, n_z])
h_mean = tf.matmul(tf.to_float(z_tensor), tf.transpose(
mu)) # (n_particles x batch_size, n_z) OneHot val_shape [n_z]
h_logstd = tf.matmul(tf.to_float(z_tensor), tf.transpose(log_sigma))
h_mean = tf.reshape(
h_mean, [-1, batch_size, n_h]) # (n_particles, batch_size, n_h)
h_logstd = tf.reshape(h_logstd, [-1, batch_size, n_h])
# returned tensor of log_prob() has shape (... + )batch_shape[:-group_event_ndims] = (n_particles, batch_size).
h = zs.Normal(
'h', h_mean, h_logstd, group_event_ndims=1
) # Multivariate Normal. val_shape []. see zhusuan Basic Concepts.
lx_h = layers.fully_connected(h, 500)
lx_h = layers.fully_connected(lx_h, 500)
x_logits = layers.fully_connected(
lx_h, n_x, activation_fn=None) # (n_particles, batch_size, n_x)
x = zs.Bernoulli(
'x', x_logits, group_event_ndims=1
) # (n_particles, batch_size, n_x) n_x=784 pixel as one event
return model, x_logits, z.tensor
def vae(observed, n, n_x, n_z, n_k, tau, n_particles, relaxed=False):
with zs.BayesianNet(observed=observed) as model:
z_stacked_logits = tf.zeros([n, n_z, n_k])
if relaxed:
z = zs.ExpConcrete('z',
tau,
z_stacked_logits,
n_samples=n_particles,
group_event_ndims=1)
z = tf.exp(tf.reshape(z, [n_particles, n, n_z * n_k]))
else:
z = zs.OnehotCategorical('z',
z_stacked_logits,
dtype=tf.float32,
n_samples=n_particles,
group_event_ndims=1)
z = tf.reshape(z, [n_particles, n, n_z * n_k])
lx_z = layers.fully_connected(z, 200, activation_fn=tf.tanh)
lx_z = layers.fully_connected(lx_z, 200, activation_fn=tf.tanh)
x_logits = layers.fully_connected(lx_z, n_x, activation_fn=None)
x = zs.Bernoulli('x', x_logits, group_event_ndims=1)
return model
def VLAE(observed, n, is_training):
with zs.BayesianNet(observed=observed) as decoder:
normalizer_params = {
'is_training': is_training,
'updates_collections': None
}
z_char_mean = tf.zeros([n, char_dim])
z_char = zs.Normal('z_char',
z_char_mean,
std=1.,
n_samples=1,
group_event_ndims=1)
z_char = tf.reshape(z_char, [-1, char_dim])
z_font_mean = tf.zeros([n, font_dim])
z_font = zs.Normal('z_font',
z_font_mean,
std=1.,
n_samples=1,
group_event_ndims=1)
z_font = tf.reshape(z_font, [-1, font_dim])
latent2 = z_char
latent1 = z_font
ladder2 = layers.fully_connected(
latent2,
ngf * 8 * 4 * 4,
activation_fn=tf.nn.relu,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params,
weights_initializer=init_ops.RandomNormal(stddev=0.02))
ladder2 = tf.reshape(ladder2, [-1, 4, 4, ngf * 8])
ladder2 = layers.conv2d_transpose(
ladder2,
ngf * 8,
4,
stride=1,
activation_fn=tf.nn.relu,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params,
weights_initializer=init_ops.RandomNormal(stddev=0.02))
ladder2 = layers.conv2d_transpose(
ladder2,
ngf * 4,
4,
stride=2,
activation_fn=tf.nn.relu,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params,
weights_initializer=init_ops.RandomNormal(stddev=0.02))
inference1 = layers.conv2d_transpose(
ladder2,
ngf * 4,
4,
stride=1,
activation_fn=tf.nn.relu,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params,
weights_initializer=init_ops.RandomNormal(stddev=0.02))
ladder1 = layers.fully_connected(
latent1,
ngf * 4 * 8 * 8,
activation_fn=tf.nn.relu,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params,
weights_initializer=init_ops.RandomNormal(stddev=0.02))
ladder1 = tf.reshape(ladder1, [-1, 8, 8, ngf * 4])
ladder1 = tf.concat([ladder1, inference1], 3)
ladder1 = layers.conv2d_transpose(
ladder1,
ngf * 2,
4,
stride=2,
activation_fn=tf.nn.relu,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params,
weights_initializer=init_ops.RandomNormal(stddev=0.02))
ladder1 = layers.conv2d_transpose(
ladder1,
ngf * 2,
4,
stride=1,
activation_fn=tf.nn.relu,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params,
weights_initializer=init_ops.RandomNormal(stddev=0.02))
inference0 = layers.conv2d_transpose(
ladder1,
ngf,
4,
stride=2,
activation_fn=tf.nn.relu,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params,
weights_initializer=init_ops.RandomNormal(stddev=0.02))
# ladder0 = layers.fully_connected(latent0, ngf * 1 * 32 * 32, activation_fn=tf.nn.relu,
# normalizer_fn=layers.batch_norm,
# normalizer_params=normalizer_params,
# weights_initializer=init_ops.RandomNormal(stddev=0.02))
# ladder0 = tf.reshape(ladder0, [-1, 32, 32, ngf * 1])
# ladder0 = tf.concat([ladder0, inference0], 3)
ladder0 = layers.conv2d_transpose(
inference0,
ngf,
4,
stride=1,
activation_fn=tf.nn.relu,
normalizer_fn=layers.batch_norm,
normalizer_params=normalizer_params,
weights_initializer=init_ops.RandomNormal(stddev=0.02))
x_logits = layers.conv2d_transpose(
ladder0,
1,
4,
stride=2,
activation_fn=tf.identity,
weights_initializer=init_ops.RandomNormal(stddev=0.02))
x_logits = tf.reshape(x_logits, [1, -1, n_x])
x = zs.Bernoulli('x', x_logits, n_samples=1, group_event_ndims=1)
return decoder, x_logits
def vae(observed,
n,
n_particles,
is_training,
dim_h=40,
dim_z=10,
dim_x=784):
'''decoder: z-->h-->x
n: batch_size
dim_z: K = 10
dim_x: 784
dim_h: D = 40
'''
with zs.BayesianNet(observed=observed) as model:
normalizer_params = {
'is_training': is_training,
'updates_collections': None
}
pai = tf.get_variable('pai',
shape=[dim_z],
dtype=tf.float32,
trainable=True,
initializer=tf.constant_initializer(1.0))
n_pai = tf.tile(tf.expand_dims(pai, 0), [n, 1])
z = zs.OnehotCategorical('z',
logits=n_pai,
dtype=tf.float32,
n_samples=n_particles)
mu = tf.get_variable('mu',
shape=[dim_z, dim_h],
dtype=tf.float32,
initializer=tf.random_uniform_initializer(
-1, 1))
log_sigma = tf.get_variable(
'log_sigma',
shape=[dim_z, dim_h],
dtype=tf.float32,
initializer=tf.random_uniform_initializer(-3, -2))
h_mean = tf.reshape(
tf.matmul(tf.reshape(z, [-1, dim_z]), mu),
[n_particles, -1, dim_h]) # [n_particles, None, dim_x]
h_logstd = tf.reshape(
tf.matmul(tf.reshape(z, [-1, dim_z]), log_sigma),
[n_particles, -1, dim_h])
h = zs.Normal(
'h',
mean=h_mean,
logstd=h_logstd,
#n_samples=n_particles,
group_event_ndims=1)
lx_h = layers.fully_connected(
h,
512,
# normalizer_fn=layers.batch_norm,
# normalizer_params=normalizer_params
)
lx_h = layers.fully_connected(
lx_h,
512,
# normalizer_fn=layers.batch_norm,
# normalizer_params=normalizer_params
)
x_logits = layers.fully_connected(
lx_h, dim_x, activation_fn=None) # the log odds of being 1
x = zs.Bernoulli(
'x',
x_logits,
#n_samples=n_particles,
group_event_ndims=1)
return model, x_logits, h, z.tensor