def test_plot_partial_dependence():
"""
Tests :func:`fatf.vis.feature_influence.plot_partial_dependence` function.
"""
feature_name = 'some feature'
class_index = 1
class_name = 'middle'
figure, axis = fvfi.plot_partial_dependence(FAKE_PD_ARRAY, FAKE_LINESPACE,
class_index, feature_name,
class_name)
assert isinstance(figure, plt.Figure)
p_title, p_x_label, p_x_range, p_y_label, p_y_range = futv.get_plot_data(
axis)
# ...check title
assert p_title == 'Partial Dependence'
# ...check x range
assert np.array_equal(p_x_range, [FAKE_LINESPACE[0], FAKE_LINESPACE[-1]])
# ...check x label
assert p_x_label == feature_name
# ...check y range
assert np.array_equal(p_y_range, [-0.05, 1.05])
# ...check y label
assert p_y_label == '{} class probability'.format(class_name)
# Test the line
assert len(axis.lines) == 1
l_data, l_colour, l_alpha, l_label, l_width = futv.get_line_data(
axis.lines[0])
line_data = np.stack([FAKE_LINESPACE, FAKE_PD_ARRAY[:, class_index]],
axis=1)
assert np.array_equal(l_data, line_data)
assert l_colour == 'lightsalmon'
assert l_alpha == 0.6
assert l_label == 'PD'
assert l_width == 7
# Validate plot legend
legend = [
i for i in axis.get_children()
if isinstance(i, matplotlib.legend.Legend)
]
assert len(legend) == 1
legend_texts = legend[0].get_texts()
assert len(legend_texts) == 1
assert legend_texts[0].get_text() == 'PD'
dp_to_explain_class_index,
class_name=dp_to_explain_class,
feature_name=selected_feature_name)
###############################################################################
# ICE calculation is a base of Partial Dependence (:func:`fatf.transparency.\
# models.feature_influence.partial_dependence_ice`). Then, Partial Dependence
# can be plotted either as an overlay on top of the ICE plot or on its own.
# Calculate Partial Dependence
pd_array = fatf_fi.partial_dependence_ice(ice_array)
# Plot Partial Dependence on its own
pd_plot_clean = fatf_vis_fi.plot_partial_dependence(
pd_array,
ice_linspace,
explanation_class,
class_name=explanation_class_name,
feature_name=selected_feature_name)
# Create a fresh ICE plot and add the Partial Dependence plot on top of it
ice_plot_2 = fatf_vis_fi.plot_individual_conditional_expectation(
ice_array,
ice_linspace,
explanation_class,
class_name=explanation_class_name,
feature_name=selected_feature_name)
ice_plot_2_figure, ice_plot_2_axis = ice_plot_2
pd_plot_overlay = fatf_vis_fi.plot_partial_dependence(
pd_array,
ice_linspace,
explanation_class,
def test_ice_pd_overlay():
"""
Tests overlaying PD plot on top of an ICE plot.
"""
f_name = 'some feature'
c_index = 1
c_name = 'middle'
figure, axis = fvfi.plot_individual_conditional_expectation(
FAKE_ICE_ARRAY, FAKE_LINESPACE, c_index, f_name, c_name)
assert isinstance(figure, plt.Figure)
assert isinstance(axis, plt.Axes)
none, axis = fvfi.plot_partial_dependence(FAKE_PD_ARRAY, FAKE_LINESPACE,
c_index, f_name, c_name, axis)
assert none is None
assert isinstance(axis, plt.Axes)
# Inspect the canvas
p_title, p_x_label, p_x_range, p_y_label, p_y_range = futv.get_plot_data(
axis)
# ...check title
assert p_title == ('Individual Conditional Expectation &\nPartial '
'Dependence')
# ...check x range
assert np.array_equal(p_x_range, [FAKE_LINESPACE[0], FAKE_LINESPACE[-1]])
# ...check x label
assert p_x_label == f_name
# ...check y range
assert np.array_equal(p_y_range, [-0.05, 1.05])
# ...check y label
assert p_y_label == '{} class probability'.format(c_name)
# Check ICE
assert len(axis.collections) == 1
l_data, l_colour, l_alpha, l_label, l_width = futv.get_line_data(
axis.collections[0], is_collection=True)
assert len(l_data) == FAKE_ICE_ARRAY.shape[0]
for i, line_array in enumerate(l_data):
line_data = np.stack([FAKE_LINESPACE, FAKE_ICE_ARRAY[i, :, c_index]],
axis=1)
assert np.array_equal(line_array, line_data)
assert np.isclose(l_colour,
np.array([[0.412, 0.412, 0.412, 0.5]]),
atol=1e-2).all() # dimgray mapping apparently
assert l_alpha == 0.5
assert l_label == 'ICE'
assert l_width == 1.75
# Check PD
assert len(axis.lines) == 1
l_data, l_colour, l_alpha, l_label, l_width = futv.get_line_data(
axis.lines[0])
line_data = np.stack([FAKE_LINESPACE, FAKE_PD_ARRAY[:, c_index]], axis=1)
assert np.array_equal(l_data, line_data)
assert l_colour == 'lightsalmon'
assert l_alpha == 0.6
assert l_label == 'PD'
assert l_width == 7
# Validate plot legend
legend = [
i for i in axis.get_children()
if isinstance(i, matplotlib.legend.Legend)
]
assert len(legend) == 1
legend_texts = legend[0].get_texts()
assert len(legend_texts) == 2
assert legend_texts[0].get_text() == 'PD'
assert legend_texts[1].get_text() == 'ICE'
# Train a model
clf = fatf_models.KNN()
clf.fit(iris_X, iris_y)
# Select a feature to be explained
selected_feature_index = 1
selected_feature_name = iris_feature_names[selected_feature_index]
print('Explaining feature (index: {}): {}.'.format(selected_feature_index,
selected_feature_name))
# Select class for which the explanation will be produced
explanation_class = 2
explanation_class_name = iris_class_names[explanation_class]
print('Explaining class (index: {}): {}.'.format(explanation_class,
explanation_class_name))
# Define the number of samples to be generated (granularity of the explanation)
linspace_samples = 25
# Calculate Partial Dependence
pd_array, pd_linspace = fatf_fi.partial_dependence(
iris_X, clf, selected_feature_index, steps_number=linspace_samples)
# Plot Partial Dependence on its own
pd_plot_clean = fatf_vis_fi.plot_partial_dependence(
pd_array,
pd_linspace,
explanation_class,
class_name=explanation_class_name,
feature_name=selected_feature_name)