def test_set_ground_truth_with_no_ground_truth():
"""set_ground_truth() must raise a PSDSEvalError when GT is None"""
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
psds_eval = PSDSEval()
with pytest.raises(PSDSEvalError,
match="The ground truth cannot be set without data"):
psds_eval.set_ground_truth(None, metadata)
def test_set_ground_truth_with_bad_metadata(bad_data):
"""Setting the ground truth with invalid metadata must raise an error"""
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
psds_eval = PSDSEval()
with pytest.raises(PSDSEvalError, match="The metadata data must be "
"provided in a pandas.DataFrame"):
psds_eval.set_ground_truth(gt, bad_data)
def test_setting_ground_truth_more_than_once():
"""Ensure that an error is raised when the ground truth is set twice"""
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
psds_eval = PSDSEval(metadata=metadata, ground_truth=gt)
with pytest.raises(PSDSEvalError, match="You cannot set the ground truth "
"more than once per evaluation"):
psds_eval.set_ground_truth(gt_t=gt, meta_t=metadata)
def test_set_ground_truth_with_overlapping_events(table_name, raise_error):
"""Gronud truth with overlapping events must raise an error"""
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
gt = pd.read_csv(os.path.join(DATADIR, table_name), sep="\t")
psds_eval = PSDSEval()
if raise_error:
with pytest.raises(
PSDSEvalError,
match="The ground truth dataframe provided has intersecting "
"events/labels for the same class."):
psds_eval.set_ground_truth(gt, metadata)
else:
psds_eval.set_ground_truth(gt, metadata)
assert isinstance(psds_eval.ground_truth, pd.DataFrame) is True
def test_compute_f_score_gt_later(metadata):
"""Test computation is correct when gt is not passed at init time"""
det_t, gt_t = read_gt_and_det()
psds_eval = PSDSEval(dtc_threshold=0.5,
gtc_threshold=0.5,
cttc_threshold=0.3)
psds_eval.set_ground_truth(gt_t, metadata)
f_avg, per_class_f = psds_eval.compute_macro_f_score(det_t)
expected_class_f = [
0.7752161383285303, 0.7421383647798742, 0.548936170212766,
0.44747612551159616, 0.6548881036513545, 0.7663551401869159,
0.9405405405405406, 0.6978021978021978, 0.7102941176470589,
0.8427672955974843
]
assert f_avg == pytest.approx(
0.712641), "The average F-score was incorrect"
for exp_f, class_f in zip(expected_class_f, per_class_f.values()):
assert exp_f == pytest.approx(class_f), "Per-class F-score incorrect"
def test_compute_f_score_gt_later(metadata):
"""Test computation is correct when gt is not passed at init time"""
det_t, gt_t = read_gt_and_det()
psds_eval = PSDSEval(dtc_threshold=0.5,
gtc_threshold=0.5,
cttc_threshold=0.3)
psds_eval.set_ground_truth(gt_t, metadata)
f_avg, per_class_f = psds_eval.compute_macro_f_score(det_t)
expected_class_f = [
0.7752161383285303, 0.7468354430379747, 0.548936170212766,
0.39943342776203966, 0.6548881036513545, 0.7663551401869159,
0.9405405405405406, 0.6978021978021978, 0.7105553512320706,
0.8427672955974843
]
assert f_avg == pytest.approx(0.7083329808351875), \
"The average F-score was incorrect"
for exp_f, class_f in zip(expected_class_f, per_class_f.values()):
assert exp_f == pytest.approx(class_f), "Per-class F-score incorrect"
def test_set_ground_truth_with_no_metadata():
"""set_ground_truth() must raise a PSDSEvalError with None metadata"""
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
psds_eval = PSDSEval()
with pytest.raises(PSDSEvalError, match="Audio metadata is required"):
psds_eval.set_ground_truth(gt, None)
