def test_true_positive_rate(self):
"""Checks that `true_positive_rate` calculates the right quantity."""
# For the penalty, the default loss is hinge.
expected_penalty_numerator = np.sum(
np.maximum(0.0, 1.0 + self._penalty_predictions) *
(self._penalty_labels > 0.0) * self._penalty_weights *
self._penalty_predicate)
expected_penalty_denominator = np.sum(
(self._penalty_labels > 0.0) * self._penalty_weights *
self._penalty_predicate)
expected_penalty_value = (expected_penalty_numerator /
expected_penalty_denominator)
# For the constraint, the default loss is zero-one.
expected_constraint_numerator = np.sum(
(0.5 * (1.0 + np.sign(self._constraint_predictions))) *
(self._constraint_labels > 0.0) * self._constraint_weights *
self._constraint_predicate)
expected_constraint_denominator = np.sum(
(self._constraint_labels > 0.0) * self._constraint_weights *
self._constraint_predicate)
expected_constraint_value = (expected_constraint_numerator /
expected_constraint_denominator)
actual_expression = binary_rates.true_positive_rate(
self._split_context)
self._check_rates(expected_penalty_value, expected_constraint_value,
actual_expression)
def problem_fn(logits, labels, features, weight_column=None):
# Returns a `RateMinimizationProblem` minimizing error rate subject to
# recall >= recall_target.
del features
del weight_column
context = subsettable_context.rate_context(logits, labels)
objective = binary_rates.error_rate(context)
constraints = [
binary_rates.true_positive_rate(context) >= recall_target
]
return rate_minimization_problem.RateMinimizationProblem(
objective, constraints)