def encoder(x, y, phase, scope='enc', reuse=None, internal_update=False):
with tf.variable_scope(scope, reuse=reuse):
with arg_scope([conv2d, dense], bn=True, phase=phase, activation=leaky_relu), \
arg_scope([batch_norm], internal_update=internal_update):
# Ignore y
x = conv2d(x, 64, 3, 2)
x = conv2d(x, 128, 3, 2)
x = conv2d(x, 256, 3, 2)
x = dense(x, 1024)
# Autoregression (4 steps)
ms = []
vs = []
zs = [x]
for i in xrange(5):
h = tf.concat(zs, axis=-1)
h = dense(h, 100)
m = dense(h, 20, activation=None)
v = dense(h, 20, activation=tf.nn.softplus) + 1e-5
z = gaussian_sample(m, v)
ms += [m]
vs += [v]
zs += [z]
m = tf.concat(ms, 1)
v = tf.concat(vs, 1)
z = tf.concat(zs[1:], 1)
return z, (m, v)
def qz_graph(x, y, n_z):
reuse = len(tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='qz')) > 0
# -- q(z)
with tf.variable_scope('qz'):
xy = tf.concat((x, y), 1, name='xy/concat')
h1 = tf.contrib.layers.fully_connected(xy,
4,
scope='layer1',
activation_fn=tf.nn.relu,
reuse=reuse)
zm = tf.contrib.layers.fully_connected(h1,
n_z,
scope='zm',
activation_fn=None,
reuse=reuse)
zv = tf.contrib.layers.fully_connected(h1,
n_z,
scope='zv',
activation_fn=tf.nn.softplus,
reuse=reuse)
z = gaussian_sample(zm, zv, 'z')
# Used to feed into z when sampling
z = tf.identity(z, name='z_sample')
return z, zm, zv
def encoder(x, y, phase, scope='enc', reuse=None, internal_update=False):
with tf.variable_scope(scope, reuse=reuse):
with arg_scope([conv2d, dense], bn=True, phase=phase, activation=leaky_relu), \
arg_scope([batch_norm], internal_update=internal_update):
# Ignore y
x = conv2d(x, 64, 3, 2)
x = conv2d(x, 128, 3, 2)
x = conv2d(x, 256, 3, 2)
x = dense(x, 1024)
m = dense(x, 100, activation=None)
v = dense(x, 100, activation=tf.nn.softplus) + 1e-5
z = gaussian_sample(m, v)
return z, (m, v)
def z_graph(zm, zv):
with tf.variable_scope('z'):
z = gaussian_sample(zm, zv, 'z')
# Used to feed into z when sampling
z = tf.identity(z, name='z_sample')
return z