def test_metrics_dict_add_integer() -> None:
"""
Adding a scalar metric where the value is an integer by accident should still store the metric.
"""
m = MetricsDict()
m.add_metric("foo", 1)
assert "foo" in m.values()
assert m.values()["foo"] == [1.0]
def test_delete_hue() -> None:
h1 = "a"
h2 = "b"
a = MetricsDict(hues=[h1, h2])
a.add_metric("foo", 1.0, hue=h1)
a.add_metric("bar", 2.0, hue=h2)
a.delete_hue(h1)
assert a.get_hue_names(include_default=False) == [h2]
assert list(a.enumerate_single_values()) == [(h2, "bar", 2.0)]
def test_delete_metric() -> None:
"""
Deleting a set of metrics from the dictionary.
"""
m = MetricsDict()
m.add_metric(MetricType.LOSS, 1)
assert m.values()[MetricType.LOSS.value] == [1.0]
m.delete_metric(MetricType.LOSS)
assert MetricType.LOSS.value not in m.values()
def test_add_foreground_dice() -> None:
g1 = "Liver"
g2 = "Lung"
ground_truth_ids = [BACKGROUND_CLASS_NAME, g1, g2]
dice = [0.85, 0.75, 0.55]
m = MetricsDict(hues=ground_truth_ids)
for j, ground_truth_id in enumerate(ground_truth_ids):
m.add_metric(MetricType.DICE, dice[j], hue=ground_truth_id)
metrics.add_average_foreground_dice(m)
assert m.get_single_metric(MetricType.DICE) == 0.5 * (dice[1] + dice[2])
def test_metrics_store_mixed_hues() -> None:
"""
Test to make sure metrics dict is able to handle default and non-default hues
"""
m = MetricsDict(hues=["A", "B"])
m.add_metric("foo", 1)
m.add_metric("foo", 1, hue="B")
m.add_metric("bar", 2, hue="A")
assert list(m.enumerate_single_values()) == \
[('A', 'bar', 2), ('B', 'foo', 1), (MetricsDict.DEFAULT_HUE_KEY, 'foo', 1)]
def add_average_foreground_dice(metrics: MetricsDict) -> None:
"""
If the given metrics dictionary contains an entry for Dice score, and only one value for the Dice score per class,
then add an average Dice score for all foreground classes to the metrics dictionary (modified in place).
:param metrics: The object that holds metrics. The average Dice score will be written back into this object.
"""
all_dice = []
for structure_name in metrics.get_hue_names(include_default=False):
if structure_name != BACKGROUND_CLASS_NAME:
all_dice.append(metrics.get_single_metric(MetricType.DICE, hue=structure_name))
metrics.add_metric(MetricType.DICE, np.nanmean(all_dice).item())
def test_classification_metrics_avg() -> None:
hue1 = "H1"
hue2 = "H2"
m = MetricsDict(hues=[hue1, hue2], is_classification_metrics=True)
m.add_metric("foo", 1.0)
m.add_metric("foo", 2.0)
# Perfect predictions for hue1, should give AUC == 1.0
m.add_predictions(["S1", "S2"],
np.array([0.0, 1.0]),
np.array([0.0, 1.0]),
hue=hue1)
expected_hue1_auc = 1.0
# Worst possible predictions for hue2, should give AUC == 0.0
m.add_predictions(["S1", "S2"],
np.array([1.0, 0.0]),
np.array([0.0, 1.0]),
hue=hue2)
expected_hue2_auc = 0.0
averaged = m.average(across_hues=False)
g1_averaged = averaged.values(hue=hue1)
assert MetricType.AREA_UNDER_ROC_CURVE.value in g1_averaged
assert g1_averaged[MetricType.AREA_UNDER_ROC_CURVE.value] == [
expected_hue1_auc
]
assert MetricType.AREA_UNDER_PR_CURVE.value in g1_averaged
assert MetricType.SUBJECT_COUNT.value in g1_averaged
assert g1_averaged[MetricType.SUBJECT_COUNT.value] == [2.0]
default_averaged = averaged.values()
assert default_averaged == {"foo": [1.5]}
can_enumerate = list(averaged.enumerate_single_values())
assert len(can_enumerate) >= 8
assert can_enumerate[0] == (hue1, MetricType.AREA_UNDER_ROC_CURVE.value,
1.0)
assert can_enumerate[-1] == (MetricsDict.DEFAULT_HUE_KEY, "foo", 1.5)
g2_averaged = averaged.values(hue=hue2)
assert MetricType.AREA_UNDER_ROC_CURVE.value in g2_averaged
assert g2_averaged[MetricType.AREA_UNDER_ROC_CURVE.value] == [
expected_hue2_auc
]
averaged_across_hues = m.average(across_hues=True)
assert averaged_across_hues.get_hue_names() == [
MetricsDict.DEFAULT_HUE_KEY
]
assert MetricType.AREA_UNDER_ROC_CURVE.value in averaged_across_hues.values(
)
expected_averaged_auc = 0.5 * (expected_hue1_auc + expected_hue2_auc)
assert averaged_across_hues.values()[
MetricType.AREA_UNDER_ROC_CURVE.value] == [expected_averaged_auc]
def test_metrics_dict_to_string() -> None:
"""
Test to make sure metrics dict is able to be stringified correctly
"""
m = MetricsDict()
m.add_metric("foo", 1.0)
m.add_metric("bar", math.pi)
info_df = pd.DataFrame(columns=MetricsDict.DATAFRAME_COLUMNS)
info_df = info_df.append(
{
MetricsDict.DATAFRAME_COLUMNS[0]: MetricsDict.DEFAULT_HUE_KEY,
MetricsDict.DATAFRAME_COLUMNS[1]: "foo: 1.0000, bar: 3.1416"
},
ignore_index=True)
assert m.to_string() == tabulate_dataframe(info_df)
assert m.to_string(tabulate=False) == info_df.to_string(index=False)
def test_metrics_dict1() -> None:
"""
Test insertion of scalar values into a MetricsDict.
"""
m = MetricsDict()
assert m.get_hue_names() == [MetricsDict.DEFAULT_HUE_KEY]
name = "foo"
v1 = 2.7
v2 = 3.14
m.add_metric(name, v1)
m.add_metric(name, v2)
assert m.values()[name] == [v1, v2]
with pytest.raises(ValueError) as ex:
# noinspection PyTypeChecker
m.add_metric(name, [1.0]) # type: ignore
assert "Expected the metric to be a scalar" in str(ex)
assert m.skip_nan_when_averaging[name] is False
v3 = 3.0
name2 = "bar"
m.add_metric(name2, v3, skip_nan_when_averaging=True)
assert m.skip_nan_when_averaging[name2] is True
# Expected average: Metric "foo" averages over two values v1 and v2. For "bar", we only inserted one value anyhow
average = m.average()
mean_v1_v2 = mean([v1, v2])
assert average.values() == {name: [mean_v1_v2], name2: [v3]}
num_entries = m.num_entries()
assert num_entries == {name: 2, name2: 1}
def test_metrics_dict_average_metrics_averaging() -> None:
"""
Test if averaging metrics avoid NaN as expected.
"""
m = MetricsDict()
metric1 = "foo"
v1 = 1.0
m.add_metric(metric1, v1)
m.add_metric(metric1, np.nan, skip_nan_when_averaging=True)
metric2 = "bar"
v2 = 2.0
m.add_metric(metric2, v2)
m.add_metric(metric2, np.nan, skip_nan_when_averaging=False)
average = m.average()
assert average.values()[metric1] == [v1]
assert np.isnan(average.values()[metric2])
def test_metrics_dict_flatten(hues: Optional[List[str]]) -> None:
m = MetricsDict(hues=hues)
_hues = hues or [MetricsDict.DEFAULT_HUE_KEY] * 2
m.add_metric("foo", 1.0, hue=_hues[0])
m.add_metric("foo", 2.0, hue=_hues[1])
m.add_metric("bar", 3.0, hue=_hues[0])
m.add_metric("bar", 4.0, hue=_hues[1])
if hues is None:
average = m.average(across_hues=True)
# We should be able to flatten out all the singleton values that the `average` operation returns
all_values = list(average.enumerate_single_values())
assert all_values == [(MetricsDict.DEFAULT_HUE_KEY, "foo", 1.5),
(MetricsDict.DEFAULT_HUE_KEY, "bar", 3.5)]
# When trying to flatten off a dictionary that has two values, this should fail:
with pytest.raises(ValueError) as ex:
list(m.enumerate_single_values())
assert "only hold 1 item" in str(ex)
else:
average = m.average(across_hues=False)
all_values = list(average.enumerate_single_values())
assert all_values == [('A', 'foo', 1.0), ('A', 'bar', 3.0),
('B', 'foo', 2.0), ('B', 'bar', 4.0)]
def test_get_single_metric() -> None:
h1 = "a"
m = MetricsDict(hues=[h1])
m1, v1 = ("foo", 1.0)
m2, v2 = (MetricType.LOSS, 2.0)
m.add_metric(m1, v1, hue=h1)
m.add_metric(m2, v2)
assert m.get_single_metric(m1, h1) == v1
assert m.get_single_metric(m2) == v2
with pytest.raises(KeyError) as ex1:
m.get_single_metric(m1, "no such hue")
assert "no such hue" in str(ex1)
with pytest.raises(KeyError) as ex2:
m.get_single_metric("no such metric", h1)
assert "no such metric" in str(ex2)
m.add_metric(m2, v2)
with pytest.raises(ValueError) as ex3:
m.get_single_metric(m2)
assert "Expected a single entry" in str(ex3)