def _compute_Ky(self):
"""
Computes the train covariance matrix.
"""
self.K = self._compute_covariance_matrix(self.X_train, self.X_train, self.sigma_l_tf, self.sigma_f_tf, name='K')\
+ self._compute_covariance_matrix(self.locs_train, self.locs_train, self.sigma_l_tf_loc, self.sigma_f_tf_loc, name='K_loc')
# Add diagonal for noisy observations
if self.sigma_n is not None:
self.Ky = tf.add(self.K,
tf.mul(tf.square(self.sigma_n_tf),
tf_utils.eye(self.n_tf)),
name='Ky')
else:
self.Ky = tf.identity(self.K, name='Ky')
def _compute_det(self):
"""
Computes determinant terms in lml.
"""
if self.method == 'pseudo':
self.det = tf.mul(-0.5, self.Ky_logdet, name='det')
elif self.method == 'inverse':
self.det = tf.mul(-0.5,
tf.log(tf.matrix_determinant(self.Ky)),
name='det')
else:
# Compute log determinant using previous determinant constant and
# Cholesky decomposition
self.mask_Ky = tf_utils.eye(tf.shape(self.Ky)[0], name='mask_Ky')
self.L_Ky = tf.cholesky(self.Ky)
self.det = tf.neg(tf.reduce_sum(
self.mask_Ky *
tf.log(tf.abs(0.5 * (self.L_Ky + tf.transpose(self.L_Ky)))),
keep_dims=True),
name='det')
def _compute_compound_loss(self):
"""
Computes semi-supervised loss that is a weighted sum of the negative
log marginal likelihood and the sum of test variances.
"""
# Compute sum of test variances
self.sum_test_variances = tf.reduce_sum(tf.mul(
tf_utils.eye(tf.shape(self.y_test_cov)[0]), self.y_test_cov),
name='sum_test_variances')
# Compute weighted compound loss for semi-supervised training
self.lml_component = tf.div(tf.neg(self.lml),
tf.to_float(tf.shape(self.X_train)[0]),
name='lml_component')
self.sum_test_variances_component = tf.mul(
(self.alpha / (tf.to_float(tf.shape(self.X_test)[0]))),
self.sum_test_variances,
name='sum_test_variances_component')
self.semisup_loss = tf.add(self.lml_component,
self.sum_test_variances_component,
name='semisup_loss')