How to use the fairlearn.postprocessing.ThresholdOptimizer function in fairlearn

def test_threshold_optimization_different_input_lengths(X_transform, y_transform,
                                                        sensitive_features_transform,
                                                        constraints):
    n = len(sensitive_features_ex1)
    for permutation in [(0, 1), (1, 0)]:
        with pytest.raises(ValueError, match=DIFFERENT_INPUT_LENGTH_ERROR_MESSAGE
                           .format("X, sensitive_features, and y")):
            X = X_transform(_format_as_list_of_lists(
                sensitive_features_ex1)[:n - permutation[0]])
            y = y_transform(labels_ex[:n - permutation[1]])
            sensitive_features = sensitive_features_transform(sensitive_features_ex1)

            adjusted_predictor = ThresholdOptimizer(unconstrained_predictor=ExamplePredictor(),
                                                    constraints=constraints)
            adjusted_predictor.fit(X, y, sensitive_features=sensitive_features)

    # try providing empty lists in all combinations
    for permutation in [(0, n), (n, 0)]:
        X = X_transform(_format_as_list_of_lists(
            sensitive_features_ex1)[:n - permutation[0]])
        y = y_transform(labels_ex[:n - permutation[1]])
        sensitive_features = sensitive_features_transform(sensitive_features_ex1)

        adjusted_predictor = ThresholdOptimizer(unconstrained_predictor=ExamplePredictor(),
                                                constraints=constraints)
        with pytest.raises(ValueError, match=EMPTY_INPUT_ERROR_MESSAGE):
            adjusted_predictor.fit(X, y, sensitive_features=sensitive_features)

def test_threshold_optimization_demographic_parity_e2e(sensitive_features,
                                                       sensitive_feature_names,
                                                       expected_p0, expected_p1,
                                                       X_transform, y_transform,
                                                       sensitive_features_transform):
    X = X_transform(_format_as_list_of_lists(sensitive_features))
    y = y_transform(labels_ex)
    sensitive_features_ = sensitive_features_transform(sensitive_features)
    adjusted_predictor = ThresholdOptimizer(unconstrained_predictor=ExamplePredictor(),
                                            constraints=DEMOGRAPHIC_PARITY)
    adjusted_predictor.fit(X, y, sensitive_features=sensitive_features_)

    predictions = adjusted_predictor._pmf_predict(X, sensitive_features=sensitive_features_)

    # assert demographic parity
    for sensitive_feature_name in sensitive_feature_names:
        average_probs = np.average(
            predictions[np.array(sensitive_features) == sensitive_feature_name], axis=0)
        assert np.isclose(average_probs[0], expected_p0)
        assert np.isclose(average_probs[1], expected_p1)

def test_predict_different_argument_lengths(sensitive_features, sensitive_feature_names,
                                            X_transform, y_transform,
                                            sensitive_features_transform, constraints):
    X = X_transform(_format_as_list_of_lists(sensitive_features))
    y = y_transform(labels_ex)
    sensitive_features_ = sensitive_features_transform(sensitive_features)
    adjusted_predictor = ThresholdOptimizer(unconstrained_predictor=ExamplePredictor(),
                                            constraints=constraints)
    adjusted_predictor.fit(X, y, sensitive_features=sensitive_features_)

    with pytest.raises(ValueError, match=DIFFERENT_INPUT_LENGTH_ERROR_MESSAGE
                       .format("X and sensitive_features")):
        adjusted_predictor.predict(
            X, sensitive_features=sensitive_features_transform(sensitive_features[:-1]))

    with pytest.raises(ValueError, match=DIFFERENT_INPUT_LENGTH_ERROR_MESSAGE
                       .format("X and sensitive_features")):
        adjusted_predictor.predict(X_transform(_format_as_list_of_lists(sensitive_features))[:-1],
                                   sensitive_features=sensitive_features_)

def test_threshold_optimization_non_binary_labels(X_transform, y_transform,
                                                  sensitive_features_transform, constraints):
    non_binary_labels = copy.deepcopy(labels_ex)
    non_binary_labels[0] = 2

    X = X_transform(_format_as_list_of_lists(sensitive_features_ex1))
    y = y_transform(non_binary_labels)
    sensitive_features = sensitive_features_transform(sensitive_features_ex1)

    adjusted_predictor = ThresholdOptimizer(unconstrained_predictor=ExamplePredictor(),
                                            constraints=constraints)

    with pytest.raises(ValueError, match=NON_BINARY_LABELS_ERROR_MESSAGE):
        adjusted_predictor.fit(X, y, sensitive_features=sensitive_features)

def test_predict_multiple_sensitive_features_columns_error(
        sensitive_features, sensitive_feature_names, X_transform, y_transform, constraints):
    X = X_transform(_format_as_list_of_lists(sensitive_features))
    y = y_transform(labels_ex)
    sensitive_features_ = pd.DataFrame({"A1": sensitive_features, "A2": sensitive_features})
    adjusted_predictor = ThresholdOptimizer(unconstrained_predictor=ExamplePredictor(),
                                            constraints=constraints)
    adjusted_predictor.fit(X, y, sensitive_features=sensitive_features_)

    with pytest.raises(ValueError,
                       match=MULTIPLE_DATA_COLUMNS_ERROR_MESSAGE.format("sensitive_features")):
        adjusted_predictor.predict(X, sensitive_features=sensitive_features_)

def test_threshold_optimization_degenerate_labels(X_transform, y_transform,
                                                  sensitive_features_transform, constraints):
    X = X_transform(_format_as_list_of_lists(sensitive_features_ex1))
    y = y_transform(degenerate_labels_ex)
    sensitive_features = sensitive_features_transform(sensitive_features_ex1)

    adjusted_predictor = ThresholdOptimizer(unconstrained_predictor=ExamplePredictor(),
                                            constraints=constraints)

    with pytest.raises(ValueError, match=DEGENERATE_LABELS_ERROR_MESSAGE.format('A')):
        adjusted_predictor.fit(X, y, sensitive_features=sensitive_features)

def test_threshold_optimization_equalized_odds_e2e(
        sensitive_features, sensitive_feature_names, expected_positive_p0, expected_positive_p1,
        expected_negative_p0, expected_negative_p1, X_transform, y_transform,
        sensitive_features_transform):
    X = X_transform(_format_as_list_of_lists(sensitive_features))
    y = y_transform(labels_ex)
    sensitive_features_ = sensitive_features_transform(sensitive_features)
    adjusted_predictor = ThresholdOptimizer(unconstrained_predictor=ExamplePredictor(),
                                            constraints=EQUALIZED_ODDS)
    adjusted_predictor.fit(X, y, sensitive_features=sensitive_features_)

    predictions = adjusted_predictor._pmf_predict(X, sensitive_features=sensitive_features_)

    # assert equalized odds
    for a in sensitive_feature_names:
        positive_indices = (np.array(sensitive_features) == a) * \
            (np.array(labels_ex) == 1)
        negative_indices = (np.array(sensitive_features) == a) * \
            (np.array(labels_ex) == 0)
        average_probs_positive_indices = np.average(
            predictions[positive_indices], axis=0)
        average_probs_negative_indices = np.average(
            predictions[negative_indices], axis=0)
        assert np.isclose(

def test_not_predictor(constraints):
    with pytest.raises(ValueError, match=MISSING_PREDICT_ERROR_MESSAGE):
        ThresholdOptimizer(unconstrained_predictor=ExampleNotPredictor(),
                           constraints=constraints)

def test_predict_before_fit_error(X_transform, sensitive_features_transform, predict_method_name,
                                  constraints):
    X = X_transform(_format_as_list_of_lists(sensitive_features_ex1))
    sensitive_features = sensitive_features_transform(sensitive_features_ex1)
    adjusted_predictor = ThresholdOptimizer(unconstrained_predictor=ExamplePredictor(),
                                            constraints=constraints)

    with pytest.raises(ValueError, match=PREDICT_BEFORE_FIT_ERROR_MESSAGE):
        getattr(adjusted_predictor, predict_method_name)(X, sensitive_features=sensitive_features)

def test_both_predictor_and_estimator_error(constraints):
    with pytest.raises(ValueError, match=EITHER_PREDICTOR_OR_ESTIMATOR_ERROR_MESSAGE):
        ThresholdOptimizer(unconstrained_predictor=ExamplePredictor(),
                           estimator=ExampleEstimator(),
                           constraints=constraints)