def _forward_log_det_jacobian(self, x):
# y = sinh((arcsinh(x) + skewness) * tailweight) * multiplier
# Using sinh' = cosh, arcsinh'(x) = 1 / sqrt(x**2 + 1),
# dy/dx
# = cosh((arcsinh(x) + skewness) * tailweight) * tailweight / sqrt(x**2 + 1)
# * multiplier
tailweight = tf.convert_to_tensor(self.tailweight)
return (tfp_math.log_cosh((tf.asinh(x) + self.skewness) * tailweight) -
0.5 * tfp_math.log1psquare(x) + tf.math.log(tailweight) +
tf.math.log(self._output_multiplier(tailweight)))
def _forward_log_det_jacobian(self, x):
# y = sinh((arcsinh(x) + skewness) * tailweight)
# Using sinh' = cosh, arcsinh'(x) = 1 / sqrt(x**2 + 1),
# dy/dx
# = cosh((arcsinh(x) + skewness) * tailweight) * tailweight / sqrt(x**2 + 1)
# This is computed inside the log to avoid catastrophic cancellations
# from cosh((arcsinh(x) + skewness) * tailweight) and sqrt(x**2 + 1).
return (tf.math.log(
tf.cosh((tf.asinh(x) + self.skewness) * self.tailweight)
# TODO(srvasude): Consider using cosh(arcsinh(x)) in cases
# where (arcsinh(x) + skewness) * tailweight ~= arcsinh(x).
/ _sqrtx2p1(x)) + tf.math.log(self.tailweight))
def _inverse_log_det_jacobian(self, y):
# x = sinh(arcsinh(y / multiplier) / tailweight - skewness)
# Using sinh' = cosh, arcsinh'(y) = 1 / sqrt(y**2 + 1),
# dx/dy
# = cosh(arcsinh(y / multiplier) / tailweight - skewness)
# / (tailweight * sqrt((y / multiplier)**2 + 1)) / multiplier
tailweight = tf.convert_to_tensor(self.tailweight)
multiplier = self._output_multiplier(tailweight)
y = y / multiplier
return (tfp_math.log_cosh(tf.asinh(y) / tailweight - self.skewness) -
0.5 * tfp_math.log1psquare(y) - tf.math.log(tailweight) -
tf.math.log(multiplier))
def _inverse_log_det_jacobian(self, y):
# x = sinh(arcsinh(y) / tailweight - skewness)
# Using sinh' = cosh, arcsinh'(y) = 1 / sqrt(y**2 + 1),
# dx/dy
# = cosh(arcsinh(y) / tailweight - skewness)
# / (tailweight * sqrt(y**2 + 1))
# This is computed inside the log to avoid catastrophic cancellations
# from cosh((arcsinh(y) / tailweight) - skewness) and sqrt(x**2 + 1).
return (tf.math.log(
tf.cosh(tf.asinh(y) / self.tailweight - self.skewness)
# TODO(srvasude): Consider using cosh(arcsinh(x)) in cases
# where (arcsinh(x) / tailweight) - skewness ~= arcsinh(x).
/ _sqrtx2p1(y)) - tf.math.log(self.tailweight))
def _inverse_log_det_jacobian(self, y):
# x = sinh(arcsinh(y / multiplier) / tailweight - skewness)
# Using sinh' = cosh, arcsinh'(y) = 1 / sqrt(y**2 + 1),
# dx/dy
# = cosh(arcsinh(y / multiplier) / tailweight - skewness)
# / (tailweight * sqrt((y / multiplier)**2 + 1)) / multiplier
# This is computed inside the log to avoid catastrophic cancellations
# from cosh((arcsinh(y / multiplier) / tailweight)
# - skewness) and sqrt(x**2 + 1).
tailweight = tf.convert_to_tensor(self.tailweight)
multiplier = self._output_multiplier(tailweight)
y = y / multiplier
subtract_term = tf.math.log(multiplier)
return (tf.math.log(
tf.cosh(tf.asinh(y) / tailweight - self.skewness)
# TODO(srvasude): Consider using cosh(arcsinh(x)) in cases
# where (arcsinh(x) / tailweight) - skewness ~= arcsinh(x).
/ _sqrtx2p1(y)) - tf.math.log(tailweight)) - subtract_term
def _inverse(self, y):
tailweight = tf.convert_to_tensor(self.tailweight)
multiplier = self._output_multiplier(tailweight)
return tf.sinh(tf.asinh(y / multiplier) / tailweight - self.skewness)
def _forward(self, x):
tailweight = tf.convert_to_tensor(self.tailweight)
multiplier = self._output_multiplier(tailweight)
bijector_output = tf.sinh((tf.asinh(x) + self.skewness) * tailweight)
return bijector_output * multiplier
def _output_multiplier(self, tailweight):
return self._scale_number / tf.sinh(
tf.asinh(self._scale_number) * tailweight)
def call(self, inputs):
if (not isinstance(inputs, random_variable.RandomVariable)
and not isinstance(self.kernel, random_variable.RandomVariable)
and not isinstance(self.bias, random_variable.RandomVariable)):
return super(DenseDVI, self).call(inputs)
self.call_weights()
inputs_mean, inputs_variance, inputs_covariance = get_moments(inputs)
kernel_mean, kernel_variance, _ = get_moments(self.kernel)
if self.use_bias:
bias_mean, _, bias_covariance = get_moments(self.bias)
# E[outputs] = E[inputs] * E[kernel] + E[bias]
mean = tf.tensordot(inputs_mean, kernel_mean, [[-1], [0]])
if self.use_bias:
mean = tf.nn.bias_add(mean, bias_mean)
# Cov = E[inputs**2] Cov(kernel) + E[W]^T Cov(inputs) E[W] + Cov(bias)
# For first term, assume Cov(kernel) = 0 on off-diagonals so we only
# compute diagonal term.
covariance_diag = tf.tensordot(inputs_variance + inputs_mean**2,
kernel_variance, [[-1], [0]])
# Compute quadratic form E[W]^T Cov E[W] from right-to-left. First is
# [..., features, features], [features, units] -> [..., features, units].
cov_w = tf.tensordot(inputs_covariance, kernel_mean, [[-1], [0]])
# Next is [..., features, units], [features, units] -> [..., units, units].
w_cov_w = tf.tensordot(cov_w, kernel_mean, [[-2], [0]])
covariance = w_cov_w
if self.use_bias:
covariance += bias_covariance
covariance = tf.linalg.set_diag(
covariance,
tf.linalg.diag_part(covariance) + covariance_diag)
if self.activation in (tf.keras.activations.relu, tf.nn.relu):
# Compute activation's moments with variable names from Wu et al. (2018).
variance = tf.linalg.diag_part(covariance)
scale = tf.sqrt(variance)
mu = mean / (scale + tf.keras.backend.epsilon())
mean = scale * soft_relu(mu)
pairwise_variances = (tf.expand_dims(variance, -1) *
tf.expand_dims(variance, -2)
) # [..., units, units]
rho = covariance / tf.sqrt(pairwise_variances +
tf.keras.backend.epsilon())
rho = tf.clip_by_value(rho,
-1. / (1. + tf.keras.backend.epsilon()),
1. / (1. + tf.keras.backend.epsilon()))
s = covariance / (rho + tf.keras.backend.epsilon())
mu1 = tf.expand_dims(mu, -1) # [..., units, 1]
mu2 = tf.linalg.matrix_transpose(mu1) # [..., 1, units]
a = (soft_relu(mu1) * soft_relu(mu2) +
rho * tfp.distributions.Normal(0., 1.).cdf(mu1) *
tfp.distributions.Normal(0., 1.).cdf(mu2))
gh = tf.asinh(rho)
bar_rho = tf.sqrt(1. - rho**2)
gr = gh + rho / (1. + bar_rho)
# Include numerically stable versions of gr and rho when multiplying or
# dividing them. The sign of gr*rho and rho/gr is always positive.
safe_gr = tf.abs(gr) + 0.5 * tf.keras.backend.epsilon()
safe_rho = tf.abs(rho) + tf.keras.backend.epsilon()
exp_negative_q = gr / (
2. * math.pi) * tf.exp(-safe_rho / (2. * safe_gr *
(1 + bar_rho)) +
(gh - rho) /
(safe_gr * safe_rho) * mu1 * mu2)
covariance = s * (a + exp_negative_q)
elif self.activation not in (tf.keras.activations.linear, None):
raise NotImplementedError(
'Activation is {}. Deterministic variational '
'inference is only available if activation is '
'ReLU or None.'.format(self.activation))
return generated_random_variables.MultivariateNormalFullCovariance(
mean, covariance)
def _inverse(self, y):
return tf.sinh(tf.asinh(y) / self.tailweight - self.skewness)
def _forward(self, x):
return tf.sinh((tf.asinh(x) + self.skewness) * self.tailweight)
def _inverse(self, y):
return tf.asinh(y)