def test_multi_ops_multiple_times_sequentially():
"""Run 3 times test_full_psds using delete_ops to reset the metric"""
gt = pd.read_csv(join(DATADIR, "baseline_validation_gt.csv"), sep="\t")
metadata = pd.read_csv(join(DATADIR, "baseline_validation_metadata.csv"),
sep="\t")
dets = []
dets.append(
pd.read_csv(join(DATADIR, "baseline_validation_AA_0.005.csv"),
sep="\t"))
for k in range(5):
dets.append(dets[0].sample(4500, random_state=7 * k))
print(dets[k + 1])
psds_eval = PSDSEval(dtc_threshold=0.5,
gtc_threshold=0.5,
cttc_threshold=0.3,
ground_truth=gt,
metadata=metadata)
ref_psds_value = 0.07211376135412327
for k in range(3):
for det_t in dets:
psds_eval.add_operating_point(det_t)
psds = psds_eval.psds(0.0, 0.0, 100)
assert psds.value == pytest.approx(ref_psds_value), \
"PSDS was calculated incorrectly"
psds_eval.clear_all_operating_points()
assert psds_eval.num_operating_points() == 0
assert len(psds_eval.operating_points) == 0
def test_example_1_paper_icassp(metadata):
"""Run PSDSEval on some sample data from the ICASSP paper"""
det = pd.read_csv(join(DATADIR, "test_1.det"), sep="\t")
gt = pd.read_csv(join(DATADIR, "test_1.gt"), sep="\t")
# Record the checksums of the incoming data
gt_hash = pd.util.hash_pandas_object(gt).values
det_hash = pd.util.hash_pandas_object(det).values
psds_eval = PSDSEval(dtc_threshold=0.5,
gtc_threshold=0.5,
cttc_threshold=0.3,
ground_truth=gt,
metadata=metadata)
# matrix (n_class, n_class) last col/row is world (for FP)
exp_counts = np.array([[1, 0, 0, 1], [0, 1, 0, 1], [1, 0, 1, 0],
[0, 0, 0, 0]])
tpr = np.array([1., 1., 1.])
fpr = np.array([12.857143, 12.857143, 0.])
ctr = np.array([[0., 0., 0.], [0., 0., 0.], [720., 0., 0.]])
psds_eval.add_operating_point(det)
assert np.all(psds_eval.operating_points.counts[0] == exp_counts)
np.testing.assert_allclose(psds_eval.operating_points.tpr[0], tpr)
np.testing.assert_allclose(psds_eval.operating_points.fpr[0], fpr)
np.testing.assert_allclose(psds_eval.operating_points.ctr[0], ctr)
psds1 = psds_eval.psds(0.0, 0.0, 100.0)
assert psds1.value == pytest.approx(0.9142857142857143), \
"PSDS value was calculated incorrectly"
# Check that the data has not been messed about with
assert np.all(pd.util.hash_pandas_object(gt).values == gt_hash)
assert np.all(pd.util.hash_pandas_object(det).values == det_hash)
def test_files_from_dcase(metadata):
"""Run PSDSEval on some example data from DCASE"""
det = pd.read_csv(join(DATADIR, "Y23R6_ppquxs_247.000_257000.det"),
sep="\t")
gt = pd.read_csv(join(DATADIR, "Y23R6_ppquxs_247.000_257000.gt"), sep="\t")
# Record the checksums of the incoming data
gt_hash = pd.util.hash_pandas_object(gt).values
det_hash = pd.util.hash_pandas_object(det).values
psds_eval = PSDSEval(dtc_threshold=0.5,
gtc_threshold=0.5,
cttc_threshold=0.3,
ground_truth=gt,
metadata=metadata)
# matrix (n_class, n_class) last col/row is world (for FP)
exp_counts = np.array([[1., 0., 1.], [1., 4., 0.], [0., 0., 0.]])
tpr = np.array([0.25, 1.])
fpr = np.array([12.857143, 0.])
ctr = np.array([[0., 0.], [600.40026684, 0.]])
psds_eval.add_operating_point(det)
assert np.all(psds_eval.operating_points.counts[0] == exp_counts)
np.testing.assert_allclose(psds_eval.operating_points.tpr[0], tpr)
np.testing.assert_allclose(psds_eval.operating_points.fpr[0], fpr)
np.testing.assert_allclose(psds_eval.operating_points.ctr[0], ctr)
psds1 = psds_eval.psds(0.0, 0.0, 100.0)
assert psds1.value == pytest.approx(0.6089285714285714), \
"PSDS value was calculated incorrectly"
# Check that the data has not been messed about with
assert np.all(pd.util.hash_pandas_object(gt).values == gt_hash)
assert np.all(pd.util.hash_pandas_object(det).values == det_hash)
def test_empty_det():
"""Run the PSDSEval class with tables that contain no detections"""
gt = pd.DataFrame({
"filename": ["test.wav"],
"onset": [2.4],
"offset": [5.9],
"event_label": ["c1"]
})
det = pd.DataFrame(columns=["filename", "onset", "offset", "event_label"])
metadata = pd.DataFrame({"filename": ["test.wav"], "duration": [10.0]})
# Record the checksums of the incoming data
meta_hash = pd.util.hash_pandas_object(metadata).values
gt_hash = pd.util.hash_pandas_object(gt).values
det_hash = pd.util.hash_pandas_object(det).values
psds_eval = PSDSEval(class_names=['c1'],
metadata=metadata,
ground_truth=gt)
exp_counts = np.array([[0, 0], [0, 0]])
tpr = np.array([0.])
fpr = np.array([0.])
ctr = np.array([[0.]])
psds_eval.add_operating_point(det)
assert np.all(psds_eval.operating_points.counts[0] == exp_counts)
np.testing.assert_allclose(psds_eval.operating_points.tpr[0], tpr)
np.testing.assert_allclose(psds_eval.operating_points.fpr[0], fpr)
np.testing.assert_allclose(psds_eval.operating_points.ctr[0], ctr)
psds1 = psds_eval.psds(0.0, 0.0, 100.0)
assert psds1.value == pytest.approx(0.0), \
"PSDS value was calculated incorrectly"
# Check that the data has not been messed about with
assert np.all(pd.util.hash_pandas_object(gt).values == gt_hash)
assert np.all(pd.util.hash_pandas_object(metadata).values == meta_hash)
assert np.all(pd.util.hash_pandas_object(det).values == det_hash)
def test_example_4(metadata):
"""Run PSDSEval on some sample data and ensure the results are correct"""
det = pd.read_csv(join(DATADIR, "test_4.det"), sep="\t")
gt = pd.read_csv(join(DATADIR, "test_4.gt"), sep="\t")
# Record the checksums of the incoming data
gt_hash = pd.util.hash_pandas_object(gt).values
det_hash = pd.util.hash_pandas_object(det).values
psds_eval = PSDSEval(dtc_threshold=0.5,
gtc_threshold=0.5,
cttc_threshold=0.3,
ground_truth=gt,
metadata=metadata)
# matrix (n_class, n_class) last col/row is world (for FP)
exp_counts = np.array([[2, 0, 0, 1, 0], [0, 0, 0, 1, 3], [0, 1, 1, 0, 0],
[0, 0, 0, 0, 2], [0, 0, 0, 0, 0]])
tpr = np.array([1., 0., 1., 0.])
fpr = np.array([0, 38.57142857, 0, 25.71428571])
ctr = np.array([[0, 0, 0, 87.80487805], [0, 0, 0, 300],
[0, 156.52173913, 0, 0], [0, 0, 0, 0]])
psds_eval.add_operating_point(det)
assert np.all(psds_eval.operating_points.counts[0] == exp_counts)
np.testing.assert_allclose(psds_eval.operating_points.tpr[0], tpr)
np.testing.assert_allclose(psds_eval.operating_points.fpr[0], fpr)
np.testing.assert_allclose(psds_eval.operating_points.ctr[0], ctr)
psds1 = psds_eval.psds(0.0, 0.0, 100.0)
assert psds1.value == pytest.approx(0.5), \
"PSDS value was calculated incorrectly"
# Check that the data has not been messed about with
assert np.all(pd.util.hash_pandas_object(gt).values == gt_hash)
assert np.all(pd.util.hash_pandas_object(det).values == det_hash)
def test_add_operating_point_with_no_metadata():
"""Ensure that add_operating_point raises an error when metadata is none"""
det = pd.read_csv(os.path.join(DATADIR, "test_1.det"), sep="\t")
psds_eval = PSDSEval(metadata=None, ground_truth=None)
with pytest.raises(PSDSEvalError,
match="Ground Truth must be provided "
"before adding the first operating point"):
psds_eval.add_operating_point(det)
def score_all(hyp_data, ref_data):
# as the unique files are the same, only one unique call is enough
metadata = pd.DataFrame(np.unique(hyp_data['filename']), columns=['filename'])
metadata = metadata.assign(duration=10)
psds_eval = PSDSEval(ground_truth=ref_data, metadata=metadata)
info = {"threshold": 0.5}
psds_eval.add_operating_point(hyp_data, info=info)
return psds_eval.psds(max_efpr=100).value
def test_add_operating_point_with_zero_detections():
"""An error must not be raised when there are no detections"""
det = pd.read_csv(os.path.join(DATADIR, "empty.det"), sep="\t")
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
psds_eval = PSDSEval(metadata=metadata, ground_truth=gt)
psds_eval.add_operating_point(det)
assert psds_eval.num_operating_points() == 1
assert psds_eval.operating_points["id"][0] == \
"e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
def test_that_add_operating_point_added_a_point():
"""Ensure add_operating_point adds an operating point correctly"""
det = pd.read_csv(os.path.join(DATADIR, "test_1.det"), sep="\t")
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
psds_eval = PSDSEval(metadata=metadata, ground_truth=gt)
psds_eval.add_operating_point(det)
assert psds_eval.num_operating_points() == 1
assert psds_eval.operating_points["id"][0] == \
"6f504797195d2df3bae13e416b8bf96ca89ec4e4e4d031dadadd72e382640387"
def test_add_operating_point_with_empty_dataframe():
"""Ensure add_operating_point raises an error when given an
incorrect table"""
det = pd.DataFrame()
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
psds_eval = PSDSEval(metadata=metadata, ground_truth=gt)
with pytest.raises(PSDSEvalError,
match="The data columns need to match the following"):
psds_eval.add_operating_point(det)
def test_that_add_operating_point_added_a_point():
"""Ensure add_operating_point adds an operating point correctly"""
det = pd.read_csv(os.path.join(DATADIR, "test_1.det"), sep="\t")
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
psds_eval = PSDSEval(metadata=metadata, ground_truth=gt)
psds_eval.add_operating_point(det)
assert psds_eval.num_operating_points() == 1
assert psds_eval.operating_points["id"][0] == \
"423089ce6d6554174881f69f9d0e57a8be9f5bc682dfce301462a8753aa6ec5f"
def test_add_operating_point_with_wrong_data_format():
"""Ensure add_operating_point raises an error when the input is not a
pandas table"""
det = pd.read_csv(os.path.join(DATADIR, "test_1.det"), sep="\t").to_numpy()
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
psds_eval = PSDSEval(metadata=metadata, ground_truth=gt)
with pytest.raises(PSDSEvalError,
match="The detection data must be provided "
"in a pandas.DataFrame"):
psds_eval.add_operating_point(det)
def test_two_operating_points_second_has_filtered_out_gtc():
"""Tests a case where the gt coverage df becomes empty for the second op"""
gt = pd.read_csv(join(DATADIR, "test_1.gt"), sep="\t")
metadata = pd.read_csv(join(DATADIR, "test.metadata"), sep="\t")
psds_eval = PSDSEval(1, 1, 1, ground_truth=gt, metadata=metadata)
det = pd.read_csv(join(DATADIR, "test_1.det"), sep="\t")
det2 = pd.read_csv(join(DATADIR, "test_1a.det"), sep="\t")
psds_eval.add_operating_point(det)
psds_eval.add_operating_point(det2)
assert psds_eval.psds(0.0, 0.0, 100.0).value == pytest.approx(0.0), \
"PSDS value was calculated incorrectly"
def test_two_operating_points_one_with_no_detections():
"""Tests a case where the dtc and gtc df's are empty for the second op"""
gt = pd.read_csv(join(DATADIR, "test_1.gt"), sep="\t")
metadata = pd.read_csv(join(DATADIR, "test.metadata"), sep="\t")
psds_eval = PSDSEval(ground_truth=gt, metadata=metadata)
det = pd.read_csv(join(DATADIR, "test_1.det"), sep="\t")
det2 = pd.read_csv(join(DATADIR, "test_4.det"), sep="\t")
psds_eval.add_operating_point(det)
psds_eval.add_operating_point(det2)
assert psds_eval.psds(0.0, 0.0, 100.0).value == \
pytest.approx(0.9142857142857143), \
"PSDS value was calculated incorrectly"
def compute_psds_from_operating_points(list_predictions,
groundtruth_df,
meta_df,
dtc_threshold=0.5,
gtc_threshold=0.5,
cttc_threshold=0.3):
psds = PSDSEval(dtc_threshold,
gtc_threshold,
cttc_threshold,
ground_truth=groundtruth_df,
metadata=meta_df)
for prediction_df in list_predictions:
psds.add_operating_point(prediction_df)
return psds
def test_add_operating_point_with_info_using_column_names():
"""Check for non-permitted keys in the info"""
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
det1 = pd.read_csv(os.path.join(DATADIR, "test_1.det"), sep="\t")
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
info1 = {"counts": 0, "threshold1": 1}
psds_eval = PSDSEval(dtc_threshold=0.5, gtc_threshold=0.5,
cttc_threshold=0.3, ground_truth=gt,
metadata=metadata)
with pytest.raises(PSDSEvalError,
match="the 'info' cannot contain the keys 'id', "
"'counts', 'tpr', 'fpr' or 'ctr'"):
psds_eval.add_operating_point(det1, info=info1)
def test_unknown_class_constraint_check():
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
det1 = pd.read_csv(os.path.join(DATADIR, "test_1.det"), sep="\t")
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
info1 = {"name": "test_1", "threshold1": 1}
psds_eval = PSDSEval(dtc_threshold=0.5, gtc_threshold=0.5,
cttc_threshold=0.3, ground_truth=gt,
metadata=metadata)
psds_eval.add_operating_point(det1, info=info1)
constraints = pd.DataFrame([
{"class_name": "class1", "constraint": "tpr", "value": 1.}])
with pytest.raises(PSDSEvalError,
match="Unknown class: class1"):
psds_eval.select_operating_points_per_class(constraints,
alpha_ct=1., beta=1.)
def test_delete_ops():
"""Perform deletion of ops"""
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
det = pd.read_csv(os.path.join(DATADIR, "test_1.det"), sep="\t")
det_2 = pd.read_csv(os.path.join(DATADIR, "test_1a.det"), sep="\t")
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
psds_eval = PSDSEval(dtc_threshold=0.5, gtc_threshold=0.5,
cttc_threshold=0.3, ground_truth=gt,
metadata=metadata)
assert psds_eval.operating_points.empty
psds_eval.add_operating_point(det)
psds_eval.add_operating_point(det_2)
assert psds_eval.num_operating_points() == 2
psds_eval.clear_all_operating_points()
assert psds_eval.operating_points.empty
def test_add_operating_points_with_overlapping_events(table_name, raise_error):
"""Detections with overlapping events must raise an error"""
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
det = pd.read_csv(os.path.join(DATADIR, table_name), sep="\t")
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
psds_eval = PSDSEval(dtc_threshold=0.5,
gtc_threshold=0.5,
cttc_threshold=0.3,
ground_truth=gt,
metadata=metadata)
if raise_error:
with pytest.raises(
PSDSEvalError,
match="The detection dataframe provided has intersecting "
"events/labels for the same class."):
psds_eval.add_operating_point(det)
else:
psds_eval.add_operating_point(det)
assert psds_eval.num_operating_points() == 1
def test_impossible_constraint_check():
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
det1 = pd.read_csv(os.path.join(DATADIR, "test_1.det"), sep="\t")
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
info1 = {"name": "test_1", "threshold1": 1}
psds_eval = PSDSEval(dtc_threshold=0.5, gtc_threshold=0.5,
cttc_threshold=0.3, ground_truth=gt,
metadata=metadata)
psds_eval.add_operating_point(det1, info=info1)
constraints = pd.DataFrame([
{"class_name": "c2", "constraint": "fpr", "value": 11.},
{"class_name": "c1", "constraint": "tpr", "value": 1.1}])
chosen_op_points = \
psds_eval.select_operating_points_per_class(constraints, alpha_ct=1.,
beta=1.)
assert np.isnan(chosen_op_points.TPR[0]), \
"NaN value is not returned for 0, 0 operating point"
assert np.isnan(chosen_op_points.TPR[1]), \
"NaN value is not returned for non-existing operating point"
def test_add_same_operating_point_with_different_info():
"""Check the use of conflicting info for the same operating point"""
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
det1 = pd.read_csv(os.path.join(DATADIR, "test_1.det"), sep="\t")
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
info1 = {"name": "test_1", "threshold1": 1}
info2 = {"name": "test_1_2", "threshold2": 0}
psds_eval = PSDSEval(dtc_threshold=0.5, gtc_threshold=0.5,
cttc_threshold=0.3, ground_truth=gt,
metadata=metadata)
psds_eval.add_operating_point(det1, info=info1)
psds_eval.add_operating_point(det1, info=info2)
assert psds_eval.num_operating_points() == 1
assert psds_eval.operating_points.name[0] == "test_1", \
"The info name is not correctly reported."
assert psds_eval.operating_points.threshold1[0] == 1, \
"The info threshold1 is not correctly reported."
assert "threshold2" not in psds_eval.operating_points.columns, \
"The info of ignored operating point modified the operating " \
"points table."
def test_full_dcase_validset():
"""Run PSDSEval on all the example data from DCASE"""
det = pd.read_csv(join(DATADIR, "baseline_validation_AA_0.005.csv"),
sep="\t")
gt = pd.read_csv(join(DATADIR, "baseline_validation_gt.csv"), sep="\t")
metadata = pd.read_csv(join(DATADIR, "baseline_validation_metadata.csv"),
sep="\t")
# Record the checksums of the incoming data
meta_hash = pd.util.hash_pandas_object(metadata).values
gt_hash = pd.util.hash_pandas_object(gt).values
det_hash = pd.util.hash_pandas_object(det).values
psds_eval = PSDSEval(dtc_threshold=0.5,
gtc_threshold=0.5,
cttc_threshold=0.3,
ground_truth=gt,
metadata=metadata)
# matrix (n_class, n_class) last col/row is world (for FP)
exp_counts = np.array(
[[269, 9, 63, 41, 120, 13, 7, 18, 128, 2, 302],
[5, 59, 4, 45, 29, 31, 35, 46, 86, 58,
416], [54, 17, 129, 19, 105, 13, 14, 16, 82, 20, 585],
[37, 43, 8, 164, 56, 9, 63, 63, 87, 7, 1100],
[45, 10, 79, 73, 278, 7, 24, 51, 154, 22, 1480],
[14, 22, 11, 24, 30, 41, 51, 26, 62, 43, 386],
[3, 20, 12, 136, 96, 35, 87, 103, 97, 27, 840],
[8, 41, 13, 119, 93, 48, 135, 127, 185, 32, 662],
[89, 120, 74, 493, 825, 203, 403, 187, 966, 89, 1340],
[0, 83, 1, 12, 58, 27, 46, 46, 120, 67, 390],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]])
psds_eval.add_operating_point(det)
assert np.all(psds_eval.operating_points.counts[0] == exp_counts)
psds1 = psds_eval.psds(0.0, 0.0, 100.0)
# Check that all the psds metrics match
assert psds1.value == pytest.approx(0.0044306914546640595), \
"PSDS value was calculated incorrectly"
# Check that the data has not been messed about with
assert np.all(pd.util.hash_pandas_object(gt).values == gt_hash)
assert np.all(pd.util.hash_pandas_object(metadata).values == meta_hash)
assert np.all(pd.util.hash_pandas_object(det).values == det_hash)
def test_retrieve_desired_operating_point():
"""Check if operating points can be found with requested constraints"""
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
det1 = pd.read_csv(os.path.join(DATADIR, "test_1.det"), sep="\t")
det2 = pd.read_csv(os.path.join(DATADIR, "test_2.det"), sep="\t")
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
info1 = {"name": "test_1", "threshold1": 1}
info2 = {"name": "test_2", "threshold2": 0}
psds_eval = PSDSEval(dtc_threshold=0.5, gtc_threshold=0.5,
cttc_threshold=0.3, ground_truth=gt,
metadata=metadata)
psds_eval.add_operating_point(det1, info=info1)
psds_eval.add_operating_point(det2, info=info2)
constraints = pd.DataFrame([
{"class_name": "c1", "constraint": "tpr", "value": 1.},
{"class_name": "c1", "constraint": "tpr", "value": 0.8},
{"class_name": "c2", "constraint": "fpr", "value": 13.},
{"class_name": "c3", "constraint": "efpr", "value": 240.},
{"class_name": "c3", "constraint": "efpr", "value": 26.},
{"class_name": "c1", "constraint": "fscore", "value": np.nan}])
chosen_op_points = \
psds_eval.select_operating_points_per_class(constraints, alpha_ct=1.,
beta=1.)
assert chosen_op_points.name[0] == "test_1", \
"Correct operating point is not chosen for tpr criteria with equality"
assert chosen_op_points.name[1] == "test_1", \
"Correct operating point is not chosen for tpr criteria with " \
"inequality"
assert chosen_op_points.name[2] == "test_1", \
"Correct operating point is not chosen for fpr criteria with " \
"inequality"
assert chosen_op_points.name[3] == "test_1", \
"Correct operating point is not chosen for efpr criteria with " \
"equality"
assert chosen_op_points.name[4] == "test_1", \
"Correct operating point is not chosen for efpr criteria with " \
"inequality"
assert chosen_op_points.name[5] == "test_1", \
"Correct operating point is not chosen for fscore criteria"
assert chosen_op_points.Fscore[5] == pytest.approx(2./3.), \
"Correct operating point is not chosen for fscore criteria"
def test_full_psds():
"""Run a full example of the PSDSEval and test the result"""
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
det = pd.read_csv(os.path.join(DATADIR, "test_1.det"), sep="\t")
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
psds_eval = PSDSEval(dtc_threshold=0.5,
gtc_threshold=0.5,
cttc_threshold=0.3,
ground_truth=gt,
metadata=metadata)
# matrix (n_class, n_class) last col/row is world (for FP)
exp_counts = np.array([[1, 0, 0, 1], [0, 1, 0, 1], [1, 0, 1, 0],
[0, 0, 0, 0]])
psds_eval.add_operating_point(det)
assert np.all(psds_eval.operating_points.counts[0] == exp_counts), \
"Expected counts do not match"
psds = psds_eval.psds(0.0, 0.0, 100.0)
assert psds.value == pytest.approx(0.9142857142857143), \
"PSDS was calculated incorrectly"
def psds_results(predictions, gtruth_df, gtruth_durations):
try:
dtc_threshold = 0.5
gtc_threshold = 0.5
cttc_threshold = 0.3
# Instantiate PSDSEval
psds = PSDSEval(dtc_threshold,
gtc_threshold,
cttc_threshold,
ground_truth=gtruth_df,
metadata=gtruth_durations)
psds.add_operating_point(predictions)
psds_score = psds.psds(alpha_ct=0, alpha_st=0, max_efpr=100)
print(f"\nPSD-Score (0, 0, 100): {psds_score.value:.5f}")
psds_score = psds.psds(alpha_ct=1, alpha_st=0, max_efpr=100)
print(f"\nPSD-Score (1, 0, 100): {psds_score.value:.5f}")
psds_score = psds.psds(alpha_ct=0, alpha_st=1, max_efpr=100)
print(f"\nPSD-Score (0, 1, 100): {psds_score.value:.5f}")
except psds_eval.psds.PSDSEvalError as e:
logger.error("psds did not work ....")
logger.error(e)
def test_example_2_paper_icassp(metadata):
"""Run PSDSEval on some sample data from the ICASSP paper"""
det = pd.read_csv(join(DATADIR, "test_2.det"), sep="\t")
gt = pd.read_csv(join(DATADIR, "test_2.gt"), sep="\t")
# Record the checksums of the incoming data
gt_hash = pd.util.hash_pandas_object(gt).values
det_hash = pd.util.hash_pandas_object(det).values
psds_eval = PSDSEval(dtc_threshold=0.5,
gtc_threshold=0.5,
cttc_threshold=0.3,
ground_truth=gt,
metadata=metadata)
exp_counts = np.array([[0, 0, 1, 1], [1, 0, 1, 0], [0, 0, 1, 1],
[0, 0, 0, 0]])
psds_eval.add_operating_point(det)
assert np.all(psds_eval.operating_points.counts[0] == exp_counts)
psds1 = psds_eval.psds(0.0, 0.0, 100.0)
assert psds1.value == pytest.approx(0.29047619047619044), \
"PSDS value was calculated incorrectly"
# Check that the data has not been messed about with
assert np.all(pd.util.hash_pandas_object(gt).values == gt_hash)
assert np.all(pd.util.hash_pandas_object(det).values == det_hash)
def test_adding_shuffled_operating_points():
"""Avoid the addition of the same operating point after shuffling"""
det = pd.read_csv(os.path.join(DATADIR, "test_1.det"), sep="\t")
metadata = pd.read_csv(os.path.join(DATADIR, "test.metadata"), sep="\t")
gt = pd.read_csv(os.path.join(DATADIR, "test_1.gt"), sep="\t")
psds_eval = PSDSEval(metadata=metadata, ground_truth=gt)
psds_eval.add_operating_point(det)
det_shuffled = det.copy(deep=True)
det_shuffled = det_shuffled.sample(frac=1.).reset_index(drop=True)
psds_eval.add_operating_point(det_shuffled)
det_shuffled2 = det.copy(deep=True)
det_shuffled2 = det_shuffled2[["onset", "event_label", "offset",
"filename"]]
psds_eval.add_operating_point(det_shuffled2)
assert psds_eval.num_operating_points() == 1
assert psds_eval.operating_points["id"][0] == \
"423089ce6d6554174881f69f9d0e57a8be9f5bc682dfce301462a8753aa6ec5f"
def test_full_dcase_validset():
"""Run PSDSEval on all the example data from DCASE"""
det = pd.read_csv(join(DATADIR, "baseline_validation_AA_0.005.csv"),
sep="\t")
gt = pd.read_csv(join(DATADIR, "baseline_validation_gt.csv"), sep="\t")
metadata = pd.read_csv(join(DATADIR, "baseline_validation_metadata.csv"),
sep="\t")
# Record the checksums of the incoming data
meta_hash = pd.util.hash_pandas_object(metadata).values
gt_hash = pd.util.hash_pandas_object(gt).values
det_hash = pd.util.hash_pandas_object(det).values
psds_eval = PSDSEval(dtc_threshold=0.5,
gtc_threshold=0.5,
cttc_threshold=0.3,
ground_truth=gt,
metadata=metadata)
# matrix (n_class, n_class): axis 0 = gt, axis 1 = det
exp_counts = np.array([
[269, 5, 54, 37, 45, 14, 3, 8, 89, 0, 302],
[9, 59, 17, 43, 10, 22, 20, 41, 120, 83, 416],
[63, 4, 129, 8, 79, 11, 12, 13, 74, 1, 585],
[41, 45, 19, 141, 73, 24, 139, 119, 495, 12, 1103],
[120, 29, 105, 56, 278, 30, 96, 93, 825, 58, 1480],
[13, 31, 13, 9, 7, 41, 35, 48, 203, 27, 386],
[7, 35, 14, 63, 24, 51, 87, 135, 403, 46, 840],
[18, 46, 16, 63, 51, 26, 103, 127, 187, 46, 662],
[128, 86, 82, 87, 154, 62, 97, 185, 966, 120, 1340],
[2, 58, 20, 7, 22, 43, 27, 32, 89, 67, 390],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
])
tpr = np.array([
0.64047619, 0.62105263, 0.37829912, 0.25044405, 0.4877193, 0.63076923,
0.92553191, 0.53586498, 0.55105533, 0.72826087
])
fpr = np.array([
93.08219178, 128.21917808, 180.30821918, 339.96575342, 456.16438356,
118.97260274, 258.90410959, 204.04109589, 413.01369863, 120.20547945
])
ctr = np.array([[
np.nan, 36.63950594, 412.06956854, 378.4295175, 201.60236547,
100.2921207, 13.91555321, 23.16073227, 122.18868771, 0.
],
[
39.38051054, np.nan, 129.72560491, 439.79646629,
44.80052566, 157.60190396, 92.77035472, 118.69875286,
164.74879241, 382.88155531
],
[
275.66357376, 29.31160475, np.nan, 81.82259838,
353.92415271, 78.80095198, 55.66221283, 37.63618993,
101.59508866, 4.61303079
],
[
179.40010356, 329.75555343, 144.98744078, np.nan,
327.04383731, 171.92934977, 644.75396529, 344.51589244,
679.58876871, 55.35636944
],
[
525.07347382, 212.50913443, 801.24638326, 572.75818865,
np.nan, 214.91168722, 445.29770265, 269.24351258,
1132.64794785, 267.55578564
],
[
56.88295966, 227.16493681, 99.20193317, 92.05042318,
31.36036796, np.nan, 162.34812076, 138.96439359,
278.7000405, 124.55183125
],
[
30.62928597, 256.47654156, 106.8328511, 644.35296223,
107.52126158, 365.34986827, np.nan, 390.83735697,
553.28136119, 212.1994162
],
[
78.76102107, 337.08345462, 122.09468697, 644.35296223,
228.48268086, 186.25679559, 477.7673268, np.nan,
256.73353485, 212.1994162
],
[
560.07837208, 630.19950211, 625.73527074, 889.82075737,
689.92809516, 444.15082025, 449.93622038, 535.59193363,
np.nan, 553.56369442
],
[
8.75122456, 425.01826886, 152.61835872, 71.59477358,
98.56115645, 308.04008501, 125.23997887, 92.64292906,
122.18868771, np.nan
]])
psds_eval.add_operating_point(det)
assert np.all(psds_eval.operating_points.counts[0] == exp_counts)
np.testing.assert_allclose(psds_eval.operating_points.tpr[0], tpr)
np.testing.assert_allclose(psds_eval.operating_points.fpr[0], fpr)
np.testing.assert_allclose(psds_eval.operating_points.ctr[0], ctr)
psds1 = psds_eval.psds(0.0, 0.0, 100.0)
# Check that all the psds metrics match
assert psds1.value == pytest.approx(0.0044306914546640595), \
"PSDS value was calculated incorrectly"
# Check that the data has not been messed about with
assert np.all(pd.util.hash_pandas_object(gt).values == gt_hash)
assert np.all(pd.util.hash_pandas_object(metadata).values == meta_hash)
assert np.all(pd.util.hash_pandas_object(det).values == det_hash)
data_dir = os.path.join(os.path.dirname(__file__), "data")
ground_truth_csv = os.path.join(data_dir, "dcase2019t4_gt.csv")
metadata_csv = os.path.join(data_dir, "dcase2019t4_meta.csv")
gt_table = pd.read_csv(ground_truth_csv, sep="\t")
meta_table = pd.read_csv(metadata_csv, sep="\t")
# Instantiate PSDSEval
psds_eval = PSDSEval(dtc_threshold, gtc_threshold, cttc_threshold,
ground_truth=gt_table, metadata=meta_table)
# Add the operating points, with the attached information
for i, th in enumerate(np.arange(0.1, 1.1, 0.1)):
csv_file = os.path.join(data_dir, f"baseline_{th:.1f}.csv")
det_t = pd.read_csv(os.path.join(csv_file), sep="\t")
info = {"name": f"Op {i + 1}", "threshold": th}
psds_eval.add_operating_point(det_t, info=info)
print(f"\rOperating point {i+1} added", end=" ")
# Calculate the PSD-Score
psds = psds_eval.psds(alpha_ct, alpha_st, max_efpr)
print(f"\nPSD-Score: {psds.value:.5f}")
# Plot the PSD-ROC
plot_psd_roc(psds)
# Plot per class tpr vs fpr/efpr/ctr
tpr_vs_fpr, _, tpr_vs_efpr = psds_eval.psd_roc_curves(alpha_ct=1.)
plot_per_class_psd_roc(tpr_vs_fpr, psds_eval.class_names,
title="Per-class TPR-vs-FPR PSDROC",
xlabel="FPR")
plot_per_class_psd_roc(tpr_vs_efpr, psds_eval.class_names,
gtc_threshold = 0.5
cttc_threshold = 0.3
alpha_ct = 0.0
alpha_st = 0.0
max_efpr = 100
# Load metadata and ground truth tables
data_dir = os.path.join(os.path.dirname(__file__), "data")
ground_truth_csv = os.path.join(data_dir, "dcase2019t4_gt.csv")
metadata_csv = os.path.join(data_dir, "dcase2019t4_meta.csv")
gt_table = pd.read_csv(ground_truth_csv, sep="\t")
meta_table = pd.read_csv(metadata_csv, sep="\t")
# Instantiate PSDSEval
psds_eval = PSDSEval(dtc_threshold, gtc_threshold, cttc_threshold,
ground_truth=gt_table, metadata=meta_table)
# Add the operating points
for i, th in enumerate(np.arange(0.1, 1.1, 0.1)):
csv_file = os.path.join(data_dir, f"baseline_{th:.1f}.csv")
det_t = pd.read_csv(os.path.join(csv_file), sep="\t")
psds_eval.add_operating_point(det_t)
print(f"\rOperating point {i+1} added", end=" ")
# Calculate the PSD-Score
psds = psds_eval.psds(alpha_ct, alpha_st, max_efpr)
print(f"\nPSD-Score: {psds.value:.5f}")
# Plot the PSD-ROC
plot_psd_roc(psds)
