def predict(self, x, a=None, training=True, n_samps=5):
""" predict y for given x with q(y|x,a) """
if a is None:
_, _, a = self.sample_pa(x)
py_in = tf.reshape(
tf.concat(
[tf.tile(tf.expand_dims(x, 0), [self.mc_samples, 1, 1]), a],
axis=-1), [-1, self.n_x + self.n_a])
self.W = bnn.sampleCatBNN(self.py_xa, self.n_hid)
preds = dgm.forwardPassCat(py_in,
self.W,
self.n_hid,
self.nonlinearity,
self.bn,
scope='py_xa')
preds = tf.expand_dims(tf.reduce_mean(
tf.reshape(preds, [self.mc_samples, -1, self.n_y]), 0),
axis=0)
for sample in range(n_samps - 1):
self.W = bnn.sampleCatBNN(self.py_xa, self.n_hid)
preds_new = dgm.forwardPassCat(py_in,
self.W,
self.n_hid,
self.nonlinearity,
self.bn,
scope='py_xa')
logits = tf.reduce_mean(tf.reshape(
preds_new, [self.mc_samples, -1, self.n_y]),
axis=0)
preds = tf.concat([preds, tf.expand_dims(logits, 0)], 0)
return tf.reduce_mean(preds, 0)
def predict(self, x, n_samps=5, training=True): """ predict y for given x with p(y|x) """ self.W = bnn.sampleCatBNN(self.py_x, self.n_hid) preds = dgm.forwardPassCat(x, self.W, self.n_hid, self.nonlinearity, self.bn, training=training, scope='py_x') preds = tf.expand_dims(preds, axis=0) for sample in range(n_samps-1): self.W = bnn.sampleCatBNN(self.py_x, self.n_hid) preds_new = dgm.forwardPassCat(x, self.W, self.n_hid, self.nonlinearity, self.bn, training=training, scope='py_x') preds = tf.concat([preds, tf.expand_dims(preds_new, axis=0)], axis=0) return tf.reduce_mean(preds, axis=0)
def predict(self, x, n_iters=150):
""" predict y for given x with p(y|x,z) """
_, _, z = self.sample_z(x)
n_u = tf.shape(x)[0]
x_ = tf.tile(tf.expand_dims(x,0), [self.mc_samples, 1,1])
py_in = tf.reshape(tf.concat([x_, z], axis=-1), [-1,self.n_x+self.n_z])
y_ = dgm.forwardPassCat(py_in, self.py_xz, self.n_hid, self.nonlinearity, self.bn, scope='py_xz')
return tf.reduce_mean(tf.reshape(tf.expand_dims(y_,0), [self.mc_samples, -1, self.n_y]), axis=0)
def predict(self, x, training=True):
""" predict y for given x with q(y|x) """
return dgm.forwardPassCat(x,
self.qy_x,
self.n_hid,
self.nonlinearity,
self.bn,
training=training,
scope='qy_x')
def unlabeled_loss(self, x): n_u = tf.shape(x)[0] qy_l = dgm.forwardPassCat(x, self.qy_x, self.n_hid, self.nonlinearity, self.bn, scope='qy_x') x_r = tf.tile(x, [self.n_y,1]) y_u = tf.reshape(tf.tile(tf.eye(self.n_y), [1, n_u]), [-1, self.n_y]) z_m, z_lv, z = self.sample_z(x_r,y_u) x_ = tf.tile(tf.expand_dims(x_r,0),[self.mc_samples,1,1]) y_ = tf.tile(tf.expand_dims(y_u,0),[self.mc_samples,1,1]) lb_u = tf.transpose(tf.reshape(self.lowerBound(x_, y_, z, z_m, z_lv), [self.n_y, n_u])) lb_u = tf.reduce_sum(qy_l * lb_u, axis=-1) qy_entropy = -tf.reduce_sum(qy_l * tf.log(qy_l + 1e-10), axis=-1) return lb_u + qy_entropy
def encode(self, x, y=None, n_iters=100):
""" encode a new example into z-space (labeled or unlabeled) """
if y is None:
y = tf.one_hot(
tf.argmax(dgm.forwardPassCat(x,
self.qy_x,
self.n_hid,
self.nonlinearity,
self.bn,
scope='qy_x'),
axis=1), self.n_y)
_, _, z = self.sample_z(x, y)
return z
def predictq(self, x, a=None, training=True):
""" predict y for given x with q(y|x,a) """
if a is None:
_, _, a = self.sample_qa(x)
qy_in = tf.reshape(
tf.concat(
[tf.tile(tf.expand_dims(x, 0), [self.mc_samples, 1, 1]), a],
axis=-1), [-1, self.n_x + self.n_a])
predictions = dgm.forwardPassCat(qy_in, self.qy_xa, self.n_hid,
self.nonlinearity, self.bn, training,
'qy_xa')
return tf.reduce_mean(tf.reshape(predictions,
[self.mc_samples, -1, self.n_y]),
axis=0)
def unlabeled_loss(self, x): """ compute the unlabeled loss """ z_m, z_lv, z = self.sample_z(x) x_r, z_r = tf.tile(tf.tile(tf.expand_dims(x,0), [self.mc_samples,1,1]), [1,self.n_y,1]), tf.tile(z, [1,self.n_y,1]) z_mr, z_lvr = tf.tile(tf.expand_dims(z_m,0), [1,self.n_y,1]), tf.tile(tf.expand_dims(z_lv,0), [1,self.n_y,1]) y_u = tf.reshape(tf.tile(tf.eye(self.n_y), [1, tf.shape(self.x_u)[0]]), [-1, self.n_y]) y_u, n_u = tf.tile(tf.expand_dims(y_u,0), [self.mc_samples,1,1]), tf.shape(x)[0] lb_u = tf.transpose(tf.reshape(self.lowerBound(x_r, y_u, z_r, z_mr, z_lvr), [self.n_y, n_u])) qy_in = tf.reshape(tf.concat([tf.tile(tf.expand_dims(x,0),[self.mc_samples, 1,1]), z], axis=-1), [-1, self.n_x+self.n_z]) qy_l = dgm.forwardPassCat(qy_in, self.qy_xz, self.n_hid, self.nonlinearity, self.bn, scope='qy_xz') qy_l = tf.reduce_mean(tf.reshape(qy_l, [self.mc_samples, -1, self.n_y]), axis=0) lb_u = tf.reduce_sum(qy_l * lb_u, axis=-1) qy_entropy = -tf.reduce_sum(qy_l * tf.log(qy_l + 1e-10), axis=-1) return lb_u + qy_entropy
def build_model(self):
""" Define model components and variables """
self.create_placeholders()
self.initialize_networks()
## model variables and relations ##
# infernce #
self.y_ = dgm.forwardPassCatLogits(self.x,
self.qy_x,
self.n_hid,
self.nonlinearity,
self.bn,
scope='qy_x',
reuse=False)
self.qz_in = tf.concat([self.x, self.y], axis=-1)
self.qz_mean, self.qz_lv, self.z_ = dgm.samplePassGauss(
self.qz_in,
self.qz_xy,
self.n_hid,
self.nonlinearity,
self.bn,
scope='qz_xy',
reuse=False)
# generative #
self.z_prior = tf.random_normal([100, self.n_z])
if self.x_dist == 'Gaussian':
self.px_mean, self.px_lv, self.x_ = dgm.samplePassGauss(
self.z_prior,
self.px_z,
self.n_hid,
self.nonlinearity,
self.bn,
scope='px_z',
reuse=False)
elif self.x_dist == 'Bernoulli':
self.x_ = dgm.forwardPassBernoulli(self.z_prior,
self.px_z,
self.n_hid,
self.nonlinearity,
self.bn,
scope='px_z',
reuse=False)
self.py_in = tf.concat([self.x_, self.z_prior], axis=-1)
self.py_ = dgm.forwardPassCat(self.py_in,
self.py_xz,
self.n_hid,
self.nonlinearity,
self.bn,
scope='py_xz',
reuse=False)
self.predictions = self.predict(self.x)
def predict(self, x, n_iters=150):
""" predict y for given x with p(y|x,z) """
y_ = dgm.forwardPassCat(x,
self.qy_x,
self.n_hid,
self.nonlinearity,
self.bn,
scope='qy_x')
yq = y_
y_ = tf.one_hot(tf.argmax(y_, axis=1), self.n_y)
y_samps = tf.expand_dims(y_, axis=2)
for i in range(n_iters):
_, _, z = self.sample_z(x, y_)
z = tf.reshape(z, [-1, self.n_z])
py_in = tf.concat([x, z], axis=-1)
y_ = dgm.forwardPassCat(py_in,
self.py_xz,
self.n_hid,
self.nonlinearity,
self.bn,
scope='py_xz')
y_samps = tf.concat([y_samps, tf.expand_dims(y_, axis=2)], axis=2)
y_ = tf.one_hot(tf.argmax(y_, axis=1), self.n_y)
return tf.reduce_mean(y_samps, axis=2)
def encode(self, x, y=None, n_iters=100):
""" encode a new example into z-space (labeled or unlabeled) """
_, _, a = dgm.samplePassGauss(x, self.qa_x, self.n_hid,
self.nonlinearity, self.bn, True, 'qa_x')
if y is None:
h = tf.concat([x, a], axis=1)
y = tf.one_hot(
tf.argmax(dgm.forwardPassCat(h, self.qy_xa, self.n_hid,
self.nonlinearity, self.bn, True,
'qa_x'),
axis=1), self.n_y)
h = tf.concat([x, y, a], axis=1)
_, _, z = dgm.samplePassGauss(h, self.qz_xya, self.n_hid,
self.nonlinearity, self.bn, True,
'qz_xya')
return z
def unlabeled_loss(self, x):
""" compute the unlabeled loss """
qy_l = dgm.forwardPassCat(x,
self.qy_x,
self.n_hid,
self.nonlinearity,
self.bn,
scope='qy_x')
x_r = tf.tile(x, [self.n_y, 1])
y_u = tf.reshape(tf.tile(tf.eye(self.n_y),
[1, tf.shape(self.x_u)[0]]), [-1, self.n_y])
n_u = tf.shape(x)[0]
lb_u = tf.transpose(
tf.reshape(self.labeled_loss(x_r, y_u), [self.n_y, n_u]))
lb_u = tf.reduce_sum(qy_l * lb_u, axis=-1)
qy_entropy = -tf.reduce_sum(qy_l * tf.log(qy_l + 1e-10), axis=-1)
return lb_u + qy_entropy
def build_model(self):
""" Define model components and variables """
self.create_placeholders()
self.initialize_networks()
## model variables and relations ##
# inference #
self.y_ = dgm.forwardPassCatLogits(self.x, self.qy_x, self.n_hid, self.nonlinearity, self.bn, scope='qy_x', reuse=False)
self.qz_in = tf.concat([self.x, self.y], axis=-1)
self.qz_mean, self.qz_lv, self.z_ = dgm.samplePassGauss(self.qz_in, self.qz_xy, self.n_hid, self.nonlinearity, self.bn, scope='qz_xy', reuse=False)
# generative #
self.z_prior = tf.random_normal([tf.shape(self.y)[0], self.n_z])
self.px_in = tf.concat([self.y, self.z_prior], axis=-1)
if self.x_dist == 'Gaussian':
self.px_mean, self.px_lv, self.x_ = dgm.samplePassGauss(self.px_in, self.px_yz, self.n_hid, self.nonlinearity, self.bn, scope='px_yz', reuse=False)
self.x_ = tf.reshape(self.x_, [-1, self.n_x])
elif self.x_dist == 'Bernoulli':
self.x_ = dgm.forwardPassBernoulli(self.px_in, self.px_yz, self.n_hid, self.nonlinearity, self.bn, scope='px_yz', reuse=False)
self.W = bnn.sampleCatBNN(self.py_x, self.n_hid)
self.py = dgm.forwardPassCat(self.x_, self.W, self.n_hid, self.nonlinearity, self.bn, scope='py_x', reuse=False)
self.predictions = self.predict(self.x, n_samps=50, training=False)
def predictq(self, x): """ predict y for given x with q(y|x) """ return dgm.forwardPassCat(x, self.qy_x, self.n_hid, self.nonlinearity, self.bn, scope='qy_x')