def _fit_and_plot(constraints, plotting_data):
adjusted_predictor = ThresholdOptimizer(
estimator=ExamplePredictor(scores_ex), constraints=constraints)
adjusted_predictor.fit(plotting_data.X,
plotting_data.y,
sensitive_features=plotting_data.sensitive_features)
fig, (ax) = plt.subplots(1, 1)
plot_threshold_optimizer(adjusted_predictor, ax=ax, show_plot=False)
return fig
def test_no_matplotlib(constraints):
n_samples = 50
n_features = 50
n_sensitive_feature_values = 2
n_classes = 2
threshold_optimizer = ThresholdOptimizer(estimator=FakePredictor(),
constraints=constraints)
threshold_optimizer.fit(X=np.random.random((n_samples, n_features)),
y=np.random.randint(n_classes, size=n_samples),
sensitive_features=np.random.randint(
n_sensitive_feature_values, size=n_samples))
with pytest.raises(RuntimeError) as exc:
plot_threshold_optimizer(threshold_optimizer)
assert str(exc.value) == _MATPLOTLIB_IMPORT_ERROR_MESSAGE
threshold_optimizer = ThresholdOptimizer(
estimator=pipeline,
constraints="demographic_parity",
predict_method="predict_proba",
prefit=False,
)
threshold_optimizer.fit(X_train, y_train, sensitive_features=A_train)
print(threshold_optimizer.predict(X_test, sensitive_features=A_test))
print(
json.dumps(
threshold_optimizer.interpolated_thresholder_.interpolation_dict,
default=str,
indent=4,
))
plot_threshold_optimizer(threshold_optimizer)
# %%
# Similarly, :class:`fairlearn.reductions.ExponentiatedGradient` works with
# pipelines. Since it requires the :code:`sample_weight` parameter of the
# underlying estimator internally we need to provide it with the correct
# way of passing :code:`sample_weight` to just the :code:`"classifier"` step
# using the step name followed by two underscores and :code:`sample_weight`.
exponentiated_gradient = ExponentiatedGradient(
estimator=pipeline,
constraints=DemographicParity(),
sample_weight_name="classifier__sample_weight",
)
exponentiated_gradient.fit(X_train, y_train, sensitive_features=A_train)
print(exponentiated_gradient.predict(X_test))