def _anomaly_log_prob(self, targets, prediction_ops):
prediction = prediction_ops["mean"]
if self._anomaly_distribution == AnomalyMixtureARModel.GAUSSIAN_ANOMALY:
anomaly_variance = prediction_ops["anomaly_params"]
anomaly_sigma = math_ops.sqrt(
gen_math_ops.maximum(anomaly_variance, 1e-5))
log_prob = math_utils.normal_log_prob(targets, anomaly_sigma, prediction)
else:
assert self._anomaly_distribution == AnomalyMixtureARModel.CAUCHY_ANOMALY
anomaly_scale = prediction_ops["anomaly_params"]
log_prob = math_utils.cauchy_log_prob(targets, anomaly_scale, prediction)
return log_prob
def _anomaly_log_prob(self, targets, prediction_ops):
prediction = prediction_ops["mean"]
if self._anomaly_distribution == AnomalyMixtureARModel.GAUSSIAN_ANOMALY:
anomaly_variance = prediction_ops["anomaly_params"]
anomaly_sigma = math_ops.sqrt(
gen_math_ops.maximum(anomaly_variance, 1e-5))
log_prob = math_utils.normal_log_prob(targets, anomaly_sigma, prediction)
else:
assert self._anomaly_distribution == AnomalyMixtureARModel.CAUCHY_ANOMALY
anomaly_scale = prediction_ops["anomaly_params"]
log_prob = math_utils.cauchy_log_prob(targets, anomaly_scale, prediction)
return log_prob
def loss_op(self, targets, prediction_ops):
"""Create loss_op."""
prediction = prediction_ops["mean"]
if self.loss == ARModel.NORMAL_LIKELIHOOD_LOSS:
covariance = prediction_ops["covariance"]
sigma = math_ops.sqrt(gen_math_ops.maximum(covariance, 1e-5))
loss_op = -math_ops.reduce_sum(
math_utils.normal_log_prob(targets, sigma, prediction))
else:
assert self.loss == ARModel.SQUARED_LOSS, self.loss
loss_op = math_ops.reduce_sum(math_ops.square(prediction - targets))
loss_op /= math_ops.cast(
math_ops.reduce_prod(array_ops.shape(targets)), loss_op.dtype)
return loss_op
def loss_op(self, targets, prediction_ops):
"""Create loss_op."""
prediction = prediction_ops["mean"]
if self.loss == ARModel.NORMAL_LIKELIHOOD_LOSS:
covariance = prediction_ops["covariance"]
sigma = math_ops.sqrt(gen_math_ops.maximum(covariance, 1e-5))
loss_op = -math_ops.reduce_sum(
math_utils.normal_log_prob(targets, sigma, prediction))
else:
assert self.loss == ARModel.SQUARED_LOSS, self.loss
loss_op = math_ops.reduce_sum(
math_ops.squared_difference(prediction, targets))
loss_op /= math_ops.cast(
math_ops.reduce_prod(array_ops.shape(targets)), loss_op.dtype)
return loss_op
def loss_op(self, targets, prediction_ops):
"""Create loss_op."""
prediction = prediction_ops["mean"]
covariance = prediction_ops["covariance"]
# Normal data log probability.
sigma = math_ops.sqrt(gen_math_ops.maximum(covariance, 1e-5))
log_prob1 = math_utils.normal_log_prob(targets, sigma, prediction)
log_prob1 += math_ops.log(1 - self._anomaly_prior_probability)
# Anomaly log probability.
log_prob2 = self._anomaly_log_prob(targets, prediction_ops)
log_prob2 += math_ops.log(self._anomaly_prior_probability)
# We need to compute log(exp(log_prob1) + exp(log_prob2). For numerical
# stability, we rewrite the expression as below.
p1 = gen_math_ops.minimum(log_prob1, log_prob2)
p2 = gen_math_ops.maximum(log_prob1, log_prob2)
mixed_log_prob = p2 + math_ops.log(1 + gen_math_ops.exp(p1 - p2))
loss_op = -math_ops.reduce_sum(mixed_log_prob)
loss_op /= math_ops.cast(
math_ops.reduce_prod(array_ops.shape(targets)), self.dtype)
return loss_op
def loss_op(self, targets, prediction_ops):
"""Create loss_op."""
prediction = prediction_ops["mean"]
covariance = prediction_ops["covariance"]
# Normal data log probability.
sigma = math_ops.sqrt(gen_math_ops.maximum(covariance, 1e-5))
log_prob1 = math_utils.normal_log_prob(targets, sigma, prediction)
log_prob1 += math_ops.log(1 - self._anomaly_prior_probability)
# Anomaly log probability.
log_prob2 = self._anomaly_log_prob(targets, prediction_ops)
log_prob2 += math_ops.log(self._anomaly_prior_probability)
# We need to compute log(exp(log_prob1) + exp(log_prob2). For numerical
# stability, we rewrite the expression as below.
p1 = gen_math_ops.minimum(log_prob1, log_prob2)
p2 = gen_math_ops.maximum(log_prob1, log_prob2)
mixed_log_prob = p2 + math_ops.log(1 + gen_math_ops.exp(p1 - p2))
loss_op = -math_ops.reduce_sum(mixed_log_prob)
loss_op /= math_ops.cast(
math_ops.reduce_prod(array_ops.shape(targets)), self.dtype)
return loss_op