def px_graph(z, y):
reuse = len(tf.get_collection(tf.GraphKeys.VARIABLES, scope='px')) > 0
# -- p(z)
with tf.variable_scope('pz'):
zm = Dense(y, 64, 'zm', reuse=reuse)
zv = Dense(y, 64, 'zv', tf.nn.softplus, reuse=reuse)
# -- p(x)
with tf.variable_scope('px'):
h1 = Dense(z, 512, 'layer1', tf.nn.relu, reuse=reuse)
h2 = Dense(h1, 512, 'layer2', tf.nn.relu, reuse=reuse)
px_logit = Dense(h2, 784, 'logit', reuse=reuse)
return zm, zv, px_logit
tf.reset_default_graph()
x = Placeholder((None, 784), 'x')
# binarize data and create a y "placeholder"
with tf.name_scope('x_binarized'):
xb = tf.cast(tf.greater(x, tf.random_uniform(tf.shape(x), 0, 1)),
tf.float32)
with tf.name_scope('y_'):
y_ = tf.fill(tf.pack([tf.shape(x)[0], k]), 0.0)
# propose distribution over y
qy_logit, qy = qy_graph(xb, k)
# for each proposed y, infer z and reconstruct x
z, zm, zv, zm_prior, zv_prior, px_logit = [[None] * k for i in xrange(6)]
for i in xrange(k):
with tf.name_scope('graphs/hot_at{:d}'.format(i)):
zm = Dense(y, 64, 'zm', reuse=reuse)
zv = Dense(y, 64, 'zv', tf.nn.softplus, reuse=reuse)
# -- p(x)
with tf.variable_scope('px'):
with tf.name_scope('layer1'):
zy = zm + tf.sqrt(zv) * z
h1 = custom_layer(zy, reuse)
h2 = Dense(h1, 512, 'layer2', tf.nn.relu, reuse=reuse)
h3 = Dense(h2, 512, 'layer3', tf.nn.relu, reuse=reuse)
h4 = Dense(h3, 512, 'layer4', tf.nn.relu, reuse=reuse)
px_logit = Dense(h2, 784, 'logit', reuse=reuse)
return px_logit
tf.reset_default_graph()
x = Placeholder((None, 784), name='x')
# binarize data and create a y "placeholder"
with tf.name_scope('x_binarized'):
xb = tf.cast(tf.greater(x, tf.random_uniform(tf.shape(x), 0, 1)),
tf.float32)
with tf.name_scope('y_'):
y_ = tf.fill(tf.stack([tf.shape(x)[0], 10]), 0.0)
# propose distribution over y
qy_logit, qy = qy_graph(xb)
# for each proposed y, infer z and reconstruct x
z, zm, zv, px_logit = [[None] * 10 for i in xrange(4)]
for i in xrange(10):
with tf.name_scope('graphs/hot_at{:d}'.format(i)):