Files
cleanlab--cleanlab/tests/test_multilabel_classification.py
2026-07-13 12:49:22 +08:00

783 lines
27 KiB
Python

import itertools
import typing
import numpy as np
import pytest
import sklearn
from hypothesis import given, settings
from hypothesis import strategies as st
from hypothesis.extra.numpy import arrays
from hypothesis.strategies import composite
from sklearn.linear_model import LogisticRegression
from sklearn.multiclass import OneVsRestClassifier
from cleanlab import multilabel_classification as ml_classification
from cleanlab.internal import multilabel_scorer as ml_scorer
from cleanlab.internal.multilabel_utils import get_onehot_num_classes, onehot2int, stack_complement
from cleanlab.multilabel_classification import filter
from cleanlab.multilabel_classification.dataset import (
common_multilabel_issues,
multilabel_health_summary,
overall_multilabel_health_score,
rank_classes_by_multilabel_quality,
)
from cleanlab.multilabel_classification.rank import get_label_quality_scores_per_class
@pytest.fixture
def labels():
return np.array(
[
[1, 0, 0],
[0, 1, 0],
[0, 0, 1],
[1, 1, 0],
[1, 0, 1],
[0, 1, 1],
[1, 1, 1],
[0, 0, 0],
[1, 0, 1],
[0, 1, 0],
]
)
@pytest.fixture
def pred_probs_gold(labels):
pred_probs = np.array(
[
[0.203, 0.465, 0.612],
[0.802, 0.596, 0.43],
[0.776, 0.649, 0.391],
[0.201, 0.439, 0.633],
[0.203, 0.443, 0.584],
[0.814, 0.572, 0.332],
[0.201, 0.388, 0.544],
[0.778, 0.646, 0.392],
[0.796, 0.611, 0.387],
[0.199, 0.381, 0.58],
]
)
assert pred_probs.shape == labels.shape
return pred_probs
@pytest.fixture
def pred_probs():
return np.array(
[
[0.9, 0.1, 0.2],
[0.5, 0.6, 0.4],
[0.75, 0.80, 0.85],
[0.9, 0.85, 0.2],
[0.9, 0.1, 0.85],
[0.5, 0.6, 0.85],
[0.9, 0.85, 0.85],
[0.8, 0.4, 0.2],
[0.9, 0.1, 0.85],
[0.15, 0.95, 0.05],
]
)
@pytest.fixture
def pred_probs_multilabel():
return np.array(
[
[0.9, 0.1, 0.0, 0.4, 0.1],
[0.7, 0.8, 0.2, 0.3, 0.1],
[0.9, 0.8, 0.4, 0.2, 0.1],
[0.1, 0.1, 0.8, 0.3, 0.1],
[0.4, 0.5, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.2, 0.1, 0.1],
[0.8, 0.1, 0.2, 0.1, 0.1],
]
)
@pytest.fixture
def labels_multilabel():
return [[0], [0, 1], [0, 1], [2], [0, 2, 3], [], []]
@pytest.fixture
def data_multilabel(num_classes=5):
labels = []
pred_probs = []
for i in range(0, 100):
q = [0.1] * num_classes
pos = i % num_classes
labels.append([pos])
if i > 90:
pos = (pos + 2) % num_classes
q[pos] = 0.9
pred_probs.append(q)
return labels, np.array(pred_probs)
@pytest.fixture
def cv():
return sklearn.model_selection.StratifiedKFold(
n_splits=2,
shuffle=True,
random_state=42,
)
@pytest.fixture
def dummy_features(labels):
np.random.seed(42)
return np.random.rand(labels.shape[0], 2)
def test_public_label_quality_scores(labels, pred_probs):
formatted_labels = onehot2int(labels)
assert isinstance(formatted_labels, list)
scores1 = ml_classification.get_label_quality_scores(formatted_labels, pred_probs)
assert len(scores1) == len(labels)
assert (scores1 >= 0).all() and (scores1 <= 1).all()
scores2 = ml_classification.get_label_quality_scores(
formatted_labels, pred_probs, method="confidence_weighted_entropy"
)
assert not np.isclose(scores1, scores2).all()
scores3 = ml_classification.get_label_quality_scores(
formatted_labels, pred_probs, adjust_pred_probs=True
)
assert not np.isclose(scores1, scores3).all()
scores4 = ml_classification.get_label_quality_scores(
formatted_labels,
pred_probs,
method="normalized_margin",
adjust_pred_probs=True,
aggregator_kwargs={"method": "exponential_moving_average"},
)
assert not np.isclose(scores1, scores4).all()
scores5 = ml_classification.get_label_quality_scores(
formatted_labels,
pred_probs,
method="normalized_margin",
adjust_pred_probs=True,
aggregator_kwargs={"method": "softmin"},
)
assert not np.isclose(scores4, scores5).all()
scores6 = ml_classification.get_label_quality_scores(
formatted_labels,
pred_probs,
method="normalized_margin",
adjust_pred_probs=True,
aggregator_kwargs={"method": "softmin", "temperature": 0.002},
)
assert not np.isclose(scores5, scores6).all()
scores7 = ml_classification.get_label_quality_scores(
formatted_labels,
pred_probs,
method="normalized_margin",
adjust_pred_probs=True,
aggregator_kwargs={"method": np.min},
)
assert np.isclose(scores6, scores7, rtol=1e-3).all()
with pytest.raises(ValueError) as e:
_ = ml_classification.get_label_quality_scores(
formatted_labels, pred_probs, method="badchoice"
)
assert "Invalid method name: badchoice" in str(e.value)
with pytest.raises(ValueError) as e:
_ = ml_classification.get_label_quality_scores(
formatted_labels, pred_probs, aggregator_kwargs={"method": "invalid"}
)
assert "Invalid aggregation method specified: 'invalid'" in str(e.value)
class TestAggregator:
"""Test the Aggregator class."""
@pytest.fixture
def base_scores(self):
return np.array([[0.6, 0.3, 0.7, 0.1, 0.9]])
@pytest.mark.parametrize(
"method",
[np.min, np.max, np.mean, np.median, "exponential_moving_average", "softmin"],
ids=lambda x: x.__name__ if callable(x) else str(x),
)
def test_aggregator_callable(self, method):
aggregator = ml_scorer.Aggregator(method=method)
assert callable(aggregator.method), "Aggregator should store a callable method"
assert callable(aggregator), "Aggregator should be callable"
@pytest.mark.parametrize(
"method,expected_score",
[
(np.min, 0.1),
(np.max, 0.9),
(np.mean, 0.52),
(np.median, 0.6),
("exponential_moving_average", 0.436),
("softmin", 0.128),
],
ids=["min", "max", "mean", "median", "exponential_moving_average", "softmin"],
)
def test_aggregator_score(self, base_scores, method, expected_score):
aggregator = ml_scorer.Aggregator(method=method)
scores = aggregator(base_scores)
assert np.isclose(scores, np.array([expected_score]), rtol=1e-3).all()
assert scores.shape == (1,)
def test_invalid_method(self):
with pytest.raises(ValueError) as e:
_ = ml_scorer.Aggregator(method="invalid_method")
assert "Invalid aggregation method specified: 'invalid_method'" in str(
e.value
), "String constructor has limited options"
with pytest.raises(TypeError) as e:
_ = ml_scorer.Aggregator(method=1)
assert "Expected callable method" in str(e.value), "Non-callable methods are not valid"
def test_invalid_score(self, base_scores):
aggregator = ml_scorer.Aggregator(method=np.min)
with pytest.raises(ValueError) as e:
_ = aggregator(base_scores[0])
assert "Expected 2D array" in str(e.value), "Aggregator expects 2D array"
class TestMultilabelScorer:
"""Test the MultilabelScorer class."""
@pytest.fixture
def docs_labels(self):
return np.array([[0, 1, 0], [1, 0, 1]])
@pytest.fixture
def docs_pred_probs(self):
return np.array([[0.1, 0.9, 0.7], [0.4, 0.1, 0.6]])
@pytest.fixture
def default_scorer(self):
return ml_scorer.MultilabelScorer()
@pytest.mark.parametrize(
"base_scorer", [scorer for scorer in ml_scorer.ClassLabelScorer], ids=lambda x: x.name
)
@pytest.mark.parametrize(
"aggregator", [np.min, np.max, np.mean, "exponential_moving_average", "softmin"]
)
@pytest.mark.parametrize("strict", [True, False], ids=["strict", ""])
def test_call(self, base_scorer, aggregator, strict, labels, pred_probs):
scorer = ml_scorer.MultilabelScorer(base_scorer, aggregator, strict=strict)
assert callable(scorer)
test_scores = scorer(labels, pred_probs)
assert isinstance(test_scores, np.ndarray)
assert test_scores.shape == (labels.shape[0],)
# Test base_scorer_kwargs
base_scorer_kwargs = {"adjust_pred_probs": True}
if scorer.base_scorer is not ml_scorer.ClassLabelScorer.CONFIDENCE_WEIGHTED_ENTROPY:
test_scores = scorer(labels, pred_probs, base_scorer_kwargs=base_scorer_kwargs)
assert isinstance(test_scores, np.ndarray)
assert test_scores.shape == (labels.shape[0],)
else:
with pytest.raises(ValueError) as e:
scorer(labels, pred_probs, base_scorer_kwargs=base_scorer_kwargs)
assert "adjust_pred_probs is not currently supported for" in str(e)
@pytest.mark.parametrize(
"base_scorer", [scorer for scorer in ml_scorer.ClassLabelScorer], ids=lambda x: x.name
)
def test_aggregate_kwargs(self, base_scorer):
"""Make sure the instatiated aggregator kwargs can be overridden.
I.e. switching from a forgetting-factor 1.0 to 0.5.
"""
class_label_quality_scores = np.array([[0.9, 0.9, 0.3], [0.4, 0.9, 0.6]])
aggregator = ml_scorer.Aggregator(ml_scorer.exponential_moving_average, alpha=1.0)
scorer = ml_scorer.MultilabelScorer(
base_scorer=base_scorer,
aggregator=aggregator,
)
scores = scorer.aggregate(class_label_quality_scores)
assert np.allclose(scores, np.array([0.3, 0.4]))
# Use different alpha, should change scores
new_scores = scorer.aggregate(class_label_quality_scores, alpha=0.0)
assert np.allclose(new_scores, np.array([0.9, 0.9]))
def test_get_class_label_quality_scores(self, default_scorer, docs_labels, docs_pred_probs):
"""Test the get_class_label_quality_scores method."""
class_label_quality_scores = default_scorer.get_class_label_quality_scores(
docs_labels, docs_pred_probs
)
assert np.allclose(class_label_quality_scores, np.array([[0.9, 0.9, 0.3], [0.4, 0.9, 0.6]]))
@pytest.mark.parametrize(
"method", ["self_confidence", "normalized_margin", "confidence_weighted_entropy"]
)
def test_class_label_scorer_from_str(method):
for m in (method, method.upper()):
scorer = ml_scorer.ClassLabelScorer.from_str(m)
assert callable(scorer)
with pytest.raises(ValueError):
ml_scorer.ClassLabelScorer.from_str(m.replace("_", "-"))
@pytest.fixture
def scorer():
return ml_scorer.MultilabelScorer(
base_scorer=ml_scorer.ClassLabelScorer.SELF_CONFIDENCE,
aggregator=np.min,
)
def test_is_multilabel(labels):
assert ml_scorer._is_multilabel(labels)
assert not ml_scorer._is_multilabel(labels[:, 0])
@pytest.mark.parametrize("class_names", [None, ["Apple", "Cat", "Dog", "Peach", "Bird"]])
def test_common_multilabel_issues(class_names, pred_probs_multilabel, labels_multilabel):
df = common_multilabel_issues(
labels=labels_multilabel, pred_probs=pred_probs_multilabel, class_names=class_names
)
expected_issue_probabilities = [
0.14285714285714285,
0.14285714285714285,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
]
assert len(df) == 10
assert np.isclose(np.array(expected_issue_probabilities), df["Issue Probability"]).all()
if class_names:
expected_res = [
"Apple",
"Dog",
"Apple",
"Cat",
"Cat",
"Dog",
"Peach",
"Peach",
"Bird",
"Bird",
]
assert list(df["Class Name"]) == expected_res
else:
assert "Class Name" not in df.columns
def test_multilabel_find_label_issues(data_multilabel):
labels, pred_probs = data_multilabel
issues = filter.find_label_issues(
labels=labels, pred_probs=pred_probs, return_indices_ranked_by="self_confidence"
)
issues_lm = filter.find_label_issues(
labels, pred_probs, low_memory=True, return_indices_ranked_by="self_confidence"
)
intersection = len(list(set(issues).intersection(set(issues_lm))))
union = len(set(issues)) + len(set(issues_lm)) - intersection
assert float(intersection) / union > 0.95
# Check with return_indices_ranked_by=None
issues_mask = filter.find_label_issues(labels=labels, pred_probs=pred_probs)
issues_lm_mask = filter.find_label_issues(labels, pred_probs, low_memory=True)
issues_from_mask = np.where(issues_mask)[0]
issues_lm_from_mask = np.where(issues_lm_mask)[0]
intersection = len(list(set(issues_from_mask).intersection(set(issues_lm_from_mask))))
union = len(set(issues_from_mask)) + len(set(issues_lm_from_mask)) - intersection
assert float(intersection) / union > 0.95
# Check with low_memory=True, unused parameters rank_by_kwargs and n_jobs
rank_by_kwargs = {"adjust_pred_probs": None}
issues_lm2 = filter.find_label_issues(
labels,
pred_probs,
low_memory=True,
return_indices_ranked_by="self_confidence",
rank_by_kwargs=rank_by_kwargs,
n_jobs=1,
)
np.testing.assert_array_equal(issues_lm2, issues_lm)
@pytest.mark.parametrize("min_examples_per_class", [10, 90])
def test_multilabel_min_examples_per_class(data_multilabel, min_examples_per_class):
labels, pred_probs = data_multilabel
issues = filter.find_label_issues(
labels=labels, pred_probs=pred_probs, min_examples_per_class=min_examples_per_class
)
if min_examples_per_class == 10:
assert sum(issues) == 9
else:
assert sum(issues) == 0
@pytest.mark.parametrize("num_to_remove_per_class", [None, [1, 1, 0, 0, 2], [1, 1, 0, 0, 1]])
def test_multilabel_num_to_remove_per_class(data_multilabel, num_to_remove_per_class):
labels, pred_probs = data_multilabel
issues = filter.find_label_issues(
labels=labels, pred_probs=pred_probs, num_to_remove_per_class=num_to_remove_per_class
)
num_issues = sum(issues)
if num_to_remove_per_class is None:
assert num_issues == 9
else:
assert num_issues == sum(num_to_remove_per_class)
@pytest.mark.parametrize("class_names", [None, ["Apple", "Cat", "Dog", "Peach", "Bird"]])
def test_rank_classes_by_multilabel_quality(pred_probs_multilabel, labels_multilabel, class_names):
df_ranked = rank_classes_by_multilabel_quality(
pred_probs=pred_probs_multilabel, labels=labels_multilabel, class_names=class_names
)
expected_Label_Issues = [1, 0, 0, 0, 0]
expected_Label_Noise = [0.14285714285714285, 0.0, 0.0, 0.0, 0.0]
expected_Label_Quality_Score = [0.8571428571428572, 1.0, 1.0, 1.0, 1.0]
expected_Inverse_Label_Issues = [0, 1, 0, 0, 0]
expected_Inverse_Label_Noise = [0.0, 0.14285714285714285, 0.0, 0.0, 0.0]
assert list(df_ranked["Label Issues"]) == expected_Label_Issues
assert np.isclose(np.array(expected_Label_Noise), df_ranked["Label Noise"]).all()
assert np.isclose(
np.array(expected_Label_Quality_Score), df_ranked["Label Quality Score"]
).all()
assert list(df_ranked["Inverse Label Issues"]) == expected_Inverse_Label_Issues
assert np.isclose(
np.array(expected_Inverse_Label_Noise), df_ranked["Inverse Label Noise"]
).all()
if class_names:
expected_res = [
"Dog",
"Apple",
"Cat",
"Peach",
"Bird",
]
assert list(df_ranked["Class Name"]) == expected_res
else:
assert "Class Name" not in df_ranked.columns
def test_overall_multilabel_health_score(data_multilabel):
labels, pred_probs = data_multilabel
overall_label_health_score = overall_multilabel_health_score(
pred_probs=pred_probs, labels=labels
)
assert np.isclose(overall_label_health_score, 0.91)
def test_get_class_label_quality_scores():
pred_probs = np.array(
[
[0.9, 0.1, 0.0, 0.4, 0.1],
[0.7, 0.8, 0.2, 0.3, 0.1],
[0.9, 0.8, 0.4, 0.2, 0.1],
[0.1, 0.1, 0.8, 0.3, 0.1],
[0.4, 0.5, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.2, 0.1, 0.1],
[0.8, 0.1, 0.2, 0.1, 0.1],
]
)
labels = [[0], [0, 1], [0, 1], [2], [0, 2, 3], [], []]
scores = get_label_quality_scores_per_class(pred_probs=pred_probs, labels=labels)
expected_res = [
[0.9, 0.9, 1.0, 0.6, 0.9],
[0.7, 0.8, 0.8, 0.7, 0.9],
[0.9, 0.8, 0.6, 0.8, 0.9],
[0.9, 0.9, 0.8, 0.7, 0.9],
[0.4, 0.5, 0.1, 0.1, 0.9],
[0.9, 0.9, 0.8, 0.9, 0.9],
[0.2, 0.9, 0.8, 0.9, 0.9],
]
assert np.isclose(scores, np.array(expected_res)).all()
def test_health_summary_multilabel(pred_probs_multilabel, labels_multilabel):
health_summary_multilabel = multilabel_health_summary(
pred_probs=pred_probs_multilabel, labels=labels_multilabel
)
expected_keys = [
"classes_by_multilabel_quality",
"common_multilabel_issues",
"overall_multilabel_health_score",
]
assert sorted(health_summary_multilabel.keys()) == expected_keys
@pytest.mark.parametrize(
"input",
[
[[0], [1, 2], [0, 2]],
[["a", "b"], ["b"]],
np.array([[[0, 1], [0, 1]], [[1, 1], [0, 0]]]),
1,
],
ids=["lists of ids", "lists of strings", "3d array", "scalar"],
)
def test_is_multilabel_is_false(input):
assert not ml_scorer._is_multilabel(input)
def test_stack_complement():
# Toy example
pred_probs_class = np.array([0.1, 0.9, 0.3, 0.8])
pred_probs_extended = stack_complement(pred_probs_class)
pred_probs_expected = np.array(
[
[0.9, 0.1],
[0.1, 0.9],
[0.7, 0.3],
[0.2, 0.8],
]
)
assert np.isclose(pred_probs_extended, pred_probs_expected).all()
# Check preservation of probabilities
pred_probs_class = np.random.rand(100)
pred_probs_extended = stack_complement(pred_probs_class)
assert np.sum(pred_probs_extended, axis=1).all() == 1
@pytest.mark.parametrize(
"pred_probs_test",
(None, "pred_probs"),
ids=["Without probabilities", "With probabilities"],
)
def test_get_onehot_num_classes(labels, pred_probs_test, request):
pred_probs_test = (
request.getfixturevalue(pred_probs_test)
if isinstance(pred_probs_test, str)
else pred_probs_test
)
labels_list = [np.nonzero(x)[0].tolist() for x in labels]
_, num_classes = get_onehot_num_classes(labels_list, pred_probs_test)
assert num_classes == 3
def test_get_label_quality_scores_output(labels, pred_probs, scorer):
# Check that the function returns a dictionary with the correct keys.
scores = ml_scorer.get_label_quality_scores(labels, pred_probs, method=scorer)
assert isinstance(scores, np.ndarray)
assert scores.shape == (labels.shape[0],)
assert np.all(scores >= 0) and np.all(scores <= 1)
assert np.all(np.isfinite(scores))
@pytest.mark.parametrize(
"given_labels,expected",
[
(
"labels",
np.full((3, 2), 0.5),
),
(np.array([[0, 1], [0, 0], [1, 1]]), np.array([[2 / 3, 1 / 3], [1 / 3, 2 / 3]])),
(np.array([[0, 1], [0, 0], [0, 1], [0, 1]]), np.array([[4 / 4, 0 / 4], [1 / 4, 3 / 4]])),
(
np.array([[0, 1, 0, 0, 0, 0, 0, 0, 0]]),
np.array([[1, 0] if i != 1 else [0, 1] for i in range(9)]),
),
],
ids=[
"default",
"Missing class assignment configuration",
"Missing class",
"Handle more than 8 classes",
],
)
def test_multilabel_py(given_labels, expected, request):
given_labels = (
request.getfixturevalue(given_labels) if isinstance(given_labels, str) else given_labels
)
py = ml_scorer.multilabel_py(given_labels)
assert isinstance(py, np.ndarray)
assert py.shape == (given_labels.shape[1], 2)
assert np.isclose(py, expected).all()
@pytest.mark.parametrize("K", [2, 3, 4], ids=["K=2", "K=3", "K=4"])
def test_get_split_generator(cv, K):
all_configurations = np.array(list(itertools.product([0, 1], repeat=K)))
given_labels = np.repeat(all_configurations, 2, axis=0)
split_generator = ml_scorer._get_split_generator(given_labels, cv)
assert isinstance(split_generator, typing.Generator)
train, test = next(split_generator)
for split in (train, test):
assert isinstance(split, np.ndarray)
assert np.isin(split, np.arange(given_labels.shape[0])).all()
# Test that the label distribution is relatively equal among the splits.
train_labels, test_labels = given_labels[train], given_labels[test]
_, train_counts = np.unique(train_labels, axis=0, return_counts=True)
_, test_counts = np.unique(test_labels, axis=0, return_counts=True)
# cv.get_n_splits() is 2, so we expect 1/2 of the labels in each split.
assert np.all(train_counts == 1)
assert np.all(test_counts == 1)
# Test split_generator with rare/missing multilabel configurations
@pytest.mark.parametrize("K", [2, 3, 4], ids=["K=2", "K=3", "K=4"])
def test_get_split_generator_rare_configurations(cv, K):
all_configurations = np.array(list(itertools.product([0, 1], repeat=K)))
given_labels = np.repeat(all_configurations, 2, axis=0)
# Remove one configuration
given_labels = given_labels[~np.all(given_labels == all_configurations[0], axis=1)]
split_generator = ml_scorer._get_split_generator(given_labels, cv)
train, test = next(split_generator)
train_labels, test_labels = given_labels[train], given_labels[test]
# Test that the label distribution is relatively equal among the splits.
_, train_counts = np.unique(train_labels, axis=0, return_counts=True)
_, test_counts = np.unique(test_labels, axis=0, return_counts=True)
# cv.get_n_splits() is 2, so we expect 1/2 of the labels in each split.
assert np.all(train_counts == 1)
assert np.all(test_counts == 1)
assert len(train_counts) == len(test_counts) == len(all_configurations) - 1
# Remove one instance from labels
given_labels = given_labels[1:, :]
split_generator = ml_scorer._get_split_generator(given_labels, cv)
train, test = next(split_generator)
train_labels, test_labels = given_labels[train], given_labels[test]
# Test that the label distribution is relatively equal among the splits.
_, train_counts = np.unique(train_labels, axis=0, return_counts=True)
_, test_counts = np.unique(test_labels, axis=0, return_counts=True)
# cv.get_n_splits() is 2, so we expect 1/2 of the labels in each split,
# except for the class with one fewer instances.
assert len(train_counts) != len(test_counts)
def test_get_cross_validated_multilabel_pred_probs(dummy_features, labels, cv, pred_probs_gold):
clf = OneVsRestClassifier(LogisticRegression(random_state=0))
pred_probs = ml_scorer.get_cross_validated_multilabel_pred_probs(
dummy_features,
labels,
clf=clf,
cv=cv,
)
assert isinstance(pred_probs, np.ndarray)
assert pred_probs.shape == labels.shape
assert np.all(pred_probs >= 0) and np.all(pred_probs <= 1)
assert np.all(np.isfinite(pred_probs))
# Gold master test - Ensure output is consistent
assert dummy_features.shape == (10, 2)
assert np.allclose(pred_probs, pred_probs_gold, atol=5e-4)
class TestExponentialMovingAverage:
"""Test the ml_scorer.expontential_moving_average function."""
@pytest.mark.parametrize("alpha", [0.5, None])
def test_valid_alpha(self, alpha):
# Test valid alpha values
for x, expected_ema in zip(
[
np.ones(5).reshape(1, -1),
np.array([[0.1, 0.2, 0.3]]),
np.array([x / 10 for x in range(1, 7)]).reshape(2, 3),
],
[1, 0.175, np.array([0.175, 0.475])],
):
ema = ml_scorer.exponential_moving_average(x, alpha=alpha)
assert np.allclose(ema, expected_ema, atol=1e-4)
@pytest.mark.parametrize(
"alpha,expected_ema",
[[0, 0.3], [1, 0.1]],
ids=["alpha=0", "alpha=1"],
)
def test_alpha_boundary(self, alpha, expected_ema):
# alpha = 0(1) should return the largest(smallest) value
X = np.array([[0.1, 0.2, 0.3]])
ema = ml_scorer.exponential_moving_average(X, alpha=alpha)
assert np.allclose(ema, expected_ema, atol=1e-4)
def test_invalid_alpha(self):
# Test that the exponential moving average raises an error
# when alpha is not in the interval [0, 1].
partial_error_msg = r"alpha must be in the interval \[0, 1\]"
for alpha in [-0.5, 1.5]:
with pytest.raises(ValueError, match=partial_error_msg):
ml_scorer.exponential_moving_average(np.ones(5).reshape(1, -1), alpha=alpha)
def flip_labels(label, flip_prob):
"""Flips binary labels with a given probability."""
rand_flip = np.random.choice(
[0, 1], size=label.shape, replace=True, p=[1 - flip_prob, flip_prob]
)
return np.abs(label - rand_flip).astype(int)
@composite
def cleanlab_data_strategy(draw):
num_classes = draw(st.integers(min_value=2, max_value=3))
num_samples = draw(st.integers(min_value=10, max_value=50))
# Generate true labels as one-hot encoded vectors for multi-label
true_labels = draw(
arrays(dtype=np.int8, shape=(num_samples, num_classes), elements=st.integers(0, 1))
)
# Generate noise matrix for multi-label and flip those values
flip_prob = 0.2
noisy_labels = flip_labels(true_labels, flip_prob)
# Multilabel find_issues raises a ValueError if all values are the same
# To avoid that we flip the first two values if all values are equal.
for i in range(noisy_labels.shape[1]):
if np.all(noisy_labels[:, i] == noisy_labels[0, i]):
noisy_labels[:2, i] = 1 - noisy_labels[:2, i]
# Generate predicted probabilities for each class for each sample
pred_probs = draw(
arrays(
dtype=np.float32,
shape=(num_samples, num_classes),
elements=st.floats(min_value=0, max_value=1, width=32), # Specify width here
)
)
for i in range(num_samples):
for j in range(num_classes):
if draw(st.floats(min_value=0, max_value=1)) < 0.1:
# Set some probability values to exactly 0.5
pred_probs[i][j] = 0.5
return true_labels, noisy_labels, np.array(pred_probs)
class TestMultiLabel:
@given(cleanlab_data_strategy())
@settings(deadline=20000)
def test_find_label_issues(self, data):
true_labels, noisy_labels, pred_probs = data
noisy_labels_list = onehot2int(noisy_labels)
is_issue = filter.find_label_issues(
labels=noisy_labels_list, pred_probs=np.array(pred_probs), n_jobs=1
)
threshold = 0.5
predicted_labels = (pred_probs >= threshold).astype(int)
# Check if predicted labels are the same as noisy labels for each example
labels_match = np.all(predicted_labels == noisy_labels, axis=1)
# For any example flagged as having an issue, there should be at least one label mismatch
assert not np.any(
is_issue & labels_match
), "Examples with issues must have at least one label mismatch."