def test_annotation_trim_no_duration():
# When ann.duration is not set prior to trim should raise warning
ann = jams.Annotation('tag_open')
ann.duration = None
clean_warning_registry()
with warnings.catch_warnings(record=True) as out:
ann_trim = ann.trim(3, 5)
assert len(out) > 0
assert out[0].category is UserWarning
assert 'annotation.duration is not defined' in str(out[0].message).lower()
# When duration is not defined trim should keep all observations in the
# user-specified trim range.
namespace = 'tag_open'
ann = jams.Annotation(namespace)
ann.time = 100
ann.duration = None
ann.append(time=5, duration=2, value='one')
clean_warning_registry()
with warnings.catch_warnings(record=True) as out:
ann_trim = ann.trim(5, 8)
assert len(out) > 0
assert out[0].category is UserWarning
assert 'annotation.duration is not defined' in str(out[0].message).lower()
expected_data = dict(time=[5.0],
duration=[2.0],
value=['one'],
confidence=[None])
expected_ann = jams.Annotation(namespace, data=expected_data, time=5.0,
duration=3.0)
assert ann_trim.data == expected_ann.data
def create_dcase_jam(fid, labelfile, duration=10.0, weak=False):
# Create jam
jam = jams.JAMS()
# Create annotation
ann = jams.Annotation('tag_open')
# duration = sox.file_info.duration(audiofile)
ann.duration = duration
# Get labels from CSV file
fid_ = fid[1:]
labeldf = pd.read_csv(labelfile, header=None, sep='\t')
labeldf.columns = ['filename', 'start_time', 'end_time', 'label']
labeldf = labeldf[labeldf['filename'].str.contains(fid_)]
# assert len(labeldf) > 0
# Add tag for each label
for idx, row in labeldf.iterrows():
if weak:
ann.append(time=0,
duration=duration,
value=row.label,
confidence=1.0)
else:
if row.end_time <= row.start_time:
continue
ann.append(time=row.start_time,
duration=(row.end_time - row.start_time),
value=row.label,
confidence=1.0)
# Fill file metadata
jam.file_metadata.title = fid
jam.file_metadata.release = '1.0'
jam.file_metadata.duration = duration
jam.file_metadata.artist = ''
# Fill annotation metadata
ann.annotation_metadata.version = '1.0'
ann.annotation_metadata.corpus = 'DCASE 2017 Task 4'
ann.annotation_metadata.data_source = 'AudioSet'
ann.annotation_metadata.annotation_tools = 'reference'
# Add annotation to jam
jam.annotations.append(ann)
# Return jam
return jam
def run():
#Capturando os nomes das pastas
names_folders = glob.glob(path_database + '*')
for names_ in names_folders:
#Capturando o caminho completo dos audios
path_audios = glob.glob(names_ + '/*.' + audio_format)
for path_ in path_audios:
#Capturando o nome do arquivo de audio (wav)
match_obj = re.sub(names_, "", path_)
audio_name = re.sub(r'/', "", match_obj)
#Capturando o caminho completo da pasta
fold_path = re.sub(audio_name, "", path_)
os.chdir(fold_path)
#Capturando a duracao do sinal
y, sr = librosa.load(audio_name, sr=None)
duration = librosa.get_duration(y=y, sr=sr)
#Remove os arquivos de audio que possuem duracao inferior a 2.3s
if duration >= 2.3:
#Gerando o arquivo de anotacao jam
jam = jams.JAMS()
#Setando os parametros do arquivo de notacao
jam.file_metadata.duration = duration
ann = jams.Annotation(namespace='beat',
time=0,
duration=jam.file_metadata.duration)
ann.annotation_metadata = jams.AnnotationMetadata(
data_source=data_source)
ann.annotation_metadata = jams.AnnotationMetadata(
validation="")
ann.annotation_metadata.curator = jams.Curator(name=name,
email=email)
jam.annotations.append(ann)
#Salvando o arquivo de notacao para cada audio
jam_name = re.sub(r'\.wav', ".jams", audio_name)
jam.save(jam_name)
else:
os.system('rm ' + path_)
def test_annotation(namespace, tag_data, ann_metadata, ann_sandbox):
ann = jams.Annotation(namespace,
data=tag_data,
annotation_metadata=ann_metadata,
sandbox=ann_sandbox)
assert namespace == ann.namespace
assert dict(ann_metadata) == dict(ann.annotation_metadata)
assert dict(ann_sandbox) == dict(ann.sandbox)
assert len(ann.data) == len(tag_data)
for obs1, obs2 in zip(ann.data, tag_data):
assert obs1._asdict() == obs2
def save_jams(jamsfile, notes, track_duration, orig_filename):
jam = jams.JAMS()
jam.file_metadata.duration = track_duration
jam.file_metadata.title = orig_filename
midi_an = jams.Annotation(namespace='pitch_midi', duration=track_duration)
for n in notes:
midi_an.append(time=n[0], duration=n[1], value=n[2], confidence=0)
jam.annotations.append(midi_an)
jam.save(jamsfile)
def ann_beat():
ann = jams.Annotation(namespace='beat', duration=10)
# for n, i in enumerate(np.arange(0, 10, 0.5)):
# ann.append(time=i, duration=0, value=1 + (n % 4))
# Make up two measures of 4/4, plus two pickup beats
for t, v in [(0, -2), (0.5, -1),
(1, 1), (1.5, 2), (2, 3), (3, 4),
(3.5, 1), (4, 2), (4.5, 3), (5, 4),
(5.5, 1), (6, 2), (6.5, 3), (7, 4)]:
ann.append(time=t, duration=0, value=v)
return ann
def create_tempo_annotation(tempo1=0.0,
tempo2=0.0,
confidence1=1.0,
confidence2=0.0):
tempo = jams.Annotation(namespace='tempo')
tempo.append(time=0.0,
duration='nan',
value=tempo1,
confidence=confidence1)
if tempo2 != 0.0:
tempo.append(time=0.0,
duration='nan',
value=tempo2,
confidence=confidence2)
return tempo
def test_jams_trim_valid():
# For a valid scenario, ensure everything behaves as expected
jam = jams.JAMS()
jam.file_metadata.duration = 15
namespace = 'tag_open'
data = dict(time=[5.0, 5.0, 10.0],
duration=[2.0, 4.0, 4.0],
value=['one', 'two', 'three'],
confidence=[0.9, 0.9, 0.9])
ann = jams.Annotation(namespace, data=data, time=5.0, duration=10.0)
for _ in range(5):
jam.annotations.append(ann)
ann_copy = jams.Annotation(namespace, data=data, time=5.0, duration=10.0)
ann_trim = ann_copy.trim(0, 10, strict=False)
jam_trim = jam.trim(0, 10, strict=False)
for ann in jam_trim.annotations:
assert ann.data == ann_trim.data
assert jam_trim.file_metadata.duration == jam.file_metadata.duration
assert jam_trim.sandbox.trim == [{'start_time': 0, 'end_time': 10}]
# Multiple trims
jam_trim = jam.trim(0, 10).trim(8, 10)
ann_trim = ann_copy.trim(0, 10).trim(8, 10)
for ann in jam_trim.annotations:
assert ann.data == ann_trim.data
assert jam_trim.sandbox.trim == (
[{'start_time': 0, 'end_time': 10}, {'start_time': 8, 'end_time': 10}])
# Make sure file metadata copied over correctly
assert jam_trim.file_metadata == jam.file_metadata
def create_hierarchy(values, offset=0.0, duration=20):
ann = jams.Annotation(namespace='multi_segment')
for level, labels in enumerate(values):
times = np.linspace(offset,
offset + duration,
num=len(labels),
endpoint=False)
durations = list(np.diff(times))
durations.append(duration + offset - times[-1])
for t, d, v in zip(times, durations, labels):
ann.append(time=t, duration=d, value=dict(label=v, level=level))
return ann
def test_annotation_to_samples(confidence):
ann = jams.Annotation('tag_open')
ann.append(time=0, duration=0.5, value='one', confidence=0.1)
ann.append(time=0.25, duration=0.5, value='two', confidence=0.2)
ann.append(time=0.75, duration=0.5, value='three', confidence=0.3)
ann.append(time=1.5, duration=0.5, value='four', confidence=0.4)
values = ann.to_samples([0.2, 0.4, 0.75, 1.25, 1.75, 1.4], confidence=confidence)
if confidence:
values, confs = values
assert confs == [[0.1], [0.1, 0.2], [0.2, 0.3], [0.3], [0.4], []]
assert values == [['one'], ['one', 'two'], ['two', 'three'], ['three'], ['four'], []]
def serializeTrack(path, track: Track, features=[{"namespace": "beat", "data_source": "Madmom", 'feature': "beats"}]):
"""
Serialize a track in jams format
"""
jam = jams.JAMS()
jam.file_metadata.duration = track.getDuration()
for feature in features:
annotation = jams.Annotation(namespace=feature["namespace"])
annotation.annotation_metadata = jams.AnnotationMetadata(data_source=feature["data_source"])
for t in track.getFeature(feature["feature"]):
annotation.append(time=t, duration=0.0)
jam.annotations.append(annotation)
jam.save(path)
def test_annotation_array_data():
data = dict(time=[0.0, 1.0],
duration=[0.5, 0.5],
value=['one', 'two'],
confidence=[0.9, 0.9])
ann = jams.Annotation('tag_open', data=data)
arr = jams.AnnotationArray(annotations=[ann, ann])
eq_(len(arr), 2)
arr.append(ann)
eq_(len(arr), 3)
for t_ann in arr:
assert ann.data.equals(t_ann.data)
def __test(namespace, data, amd, sandbox):
ann = jams.Annotation(namespace,
data=data,
annotation_metadata=amd,
sandbox=sandbox)
eq_(namespace, ann.namespace)
if amd is not None:
eq_(dict(amd), dict(ann.annotation_metadata))
if sandbox is not None:
eq_(dict(sandbox), dict(ann.sandbox))
if data is not None:
assert ann.data.equals(jams.JamsFrame.from_dict(data))
def test_annotation_append():
data = [
dict(time=0, duration=0.5, value='one', confidence=0.9),
dict(time=1.0, duration=0.5, value='two', confidence=0.9)
]
namespace = 'tag_open'
ann = jams.Annotation(namespace, data=data)
update = dict(time=2.0, duration=1.0, value='three', confidence=0.8)
ann.append(**update)
assert ann.data[-1]._asdict() == update
def test_annotation_array_serialize():
data = dict(time=[0.0, 1.0],
duration=[0.5, 0.5],
value=['one', 'two'],
confidence=[0.9, 0.9])
namespace = 'tag_open'
ann = jams.Annotation(namespace, data=data)
arr = jams.AnnotationArray(annotations=[ann, ann])
arr_js = arr.__json__
arr2 = jams.AnnotationArray(annotations=arr_js)
eq_(arr, arr2)
def modernize_chord_annotation(annot):
# fields = ['time', 'duration', 'value', 'confidence']
data = []
for item in annot['data']:
t0 = item['start']['value']
t1 = item['end']['value']
label = item['label']['value']
record = dict(time=t0,
duration=t1 - t0,
value=str(label),
confidence=1.0)
data.append(record)
return jams.Annotation(namespace='chord',
data=data,
sandbox=annot.get('sandbox', dict()),
annotation_metadata=annot.get(
'annotation_metadata', dict()))
def test_jamsframe_add_observation_fail():
@raises(jams.ParameterError)
def __test(ann, time, duration, value, confidence):
ann.data.add_observation(time=time,
duration=duration,
value=value,
confidence=confidence)
ann = jams.Annotation(namespace='tag_open')
yield __test, ann, None, None, 'foo', 1
yield __test, ann, 0.0, None, 'foo', 1
yield __test, ann, None, 1.0, 'foo', 1
yield __test, ann, -1, -1, 'foo', 1
yield __test, ann, 0.0, -1, 'foo', 1
yield __test, ann, -1, 1.0, 'foo', 1
def inverse(self, encoded, duration=None):
'''Inverse static tag transformation'''
ann = jams.Annotation(namespace=self.namespace, duration=duration)
if np.isrealobj(encoded):
detected = (encoded >= 0.5)
else:
detected = encoded
for vd in self.encoder.inverse_transform(np.atleast_2d(detected))[0]:
vid = np.flatnonzero(self.encoder.transform(np.atleast_2d(vd)))
ann.append(time=0,
duration=duration,
value=vd,
confidence=encoded[vid])
return ann
def test_pitch_hz_to_midi():
ann = jams.Annotation(namespace='pitch_hz')
ann.append(time=0, duration=1, value=440.0, confidence=0.5)
ann2 = jams.convert(ann, 'pitch_midi')
ann.validate()
ann2.validate()
# Check the namespace
eq_(ann2.namespace, 'pitch_midi')
# midi 69 = 440.0 Hz
eq_(ann2.data.value.loc[0], 69)
# Check all else is equal
pdt.assert_series_equal(ann.data.time, ann2.data.time)
pdt.assert_series_equal(ann.data.duration, ann2.data.duration)
pdt.assert_series_equal(ann.data.confidence, ann2.data.confidence)
def test_annotation_data_frame():
namespace = 'tag_open'
data = dict(time=[5.0, 5.0, 10.0],
duration=[2.0, 4.0, 4.0],
value=['one', 'two', 'three'],
confidence=[0.9, 0.9, 0.9])
ann = jams.Annotation(namespace, data=data, time=5.0, duration=10.0)
df = ann.to_dataframe()
assert list(df.columns) == ['time', 'duration', 'value', 'confidence']
for i, row in df.iterrows():
assert row.time == data['time'][i]
assert row.duration == data['duration'][i]
assert row.value == data['value'][i]
assert row.confidence == data['confidence'][i]
def test_annotation_append():
data = dict(time=[0.0, 1.0],
duration=[0.5, 0.5],
value=['one', 'two'],
confidence=[0.9, 0.9])
namespace = 'tag_open'
ann = jams.Annotation(namespace, data=data)
update = dict(time=2.0, duration=1.0, value='three', confidence=0.8)
ann.append(**update)
jf = jams.JamsFrame.from_dict(data)
jf.add_observation(**update)
assert ann.data.equals(jf)
def test_jams_add(tag_data):
fn = 'tests/fixtures/valid.jams'
# The original jam
jam_orig = jams.load(fn)
jam = jams.load(fn)
# Make a new jam with the same metadata and different data
jam2 = jams.load(fn)
ann = jams.Annotation('tag_open', data=tag_data)
jam2.annotations = jams.AnnotationArray(annotations=[ann])
# Add the two
jam.add(jam2)
assert len(jam.annotations) == 3
assert jam.annotations[:-1] == jam_orig.annotations
assert jam.annotations[-1] == jam2.annotations[0]
def tags_to_jams(tags, duration=0, namespace="tag_open", description=None):
"""Convert tags annotations into jams format.
Args:
tags (annotations.Tags): tags annotation object
namespace (str): the jams-compatible tag namespace
description (str): annotation description
Returns:
jams.Annotation: jams annotation object.
"""
ann = jams.Annotation(namespace=namespace)
ann.annotation_metadata = jams.AnnotationMetadata(data_source="soundata")
for t, c in zip(tags.labels, tags.confidence):
ann.append(time=0.0, duration=duration, value=t, confidence=c)
if description is not None:
ann.sandbox = jams.Sandbox(name=description)
return ann
def test_task_key_present(SR, HOP_LENGTH, SPARSE):
# Create jams with key annotation
jam = jams.JAMS(file_metadata=dict(duration=12.0))
ann = jams.Annotation(namespace='key_mode')
ann.append(time=0, duration=2.0, value='C:minor')
ann.append(time=2, duration=2.0, value='N')
ann.append(time=4, duration=2.0, value='Eb')
ann.append(time=8, duration=2.0, value='D:major')
ann.append(time=10, duration=2.0, value='D:lydian')
jam.annotations.append(ann)
trans = pumpp.task.KeyTransformer(name='key',
sr=SR,
hop_length=HOP_LENGTH,
sparse=SPARSE)
output = trans.transform(jam)
# Make sure we have the mask
assert np.all(
output['key/_valid'] == [0, 12 * trans.sr // trans.hop_length])
# Ideal vectors:
# pcp = Cmin, N, Eb, N, D, D_lyd
pcp_true = np.array([
_encode_key_str('C:minor', SPARSE)[0],
_encode_key_str('N', SPARSE)[0],
_encode_key_str('Eb:major', SPARSE)[0],
_encode_key_str('N', SPARSE)[0],
_encode_key_str('D', SPARSE)[0],
_encode_key_str('D:lydian', SPARSE)[0]
])
assert np.all(output['key/pitch_profile'] == np.repeat(
pcp_true, (SR * 2 // HOP_LENGTH), axis=0))
for key in trans.fields:
assert shape_match(output[key].shape[1:], trans.fields[key].shape)
assert type_match(output[key].dtype, trans.fields[key].dtype)
def test_task_beat_present(SR, HOP_LENGTH):
# Construct a jam
jam = jams.JAMS(file_metadata=dict(duration=4.0))
ann = jams.Annotation(namespace='beat')
ann.append(time=0, duration=0.0, value=1)
ann.append(time=1, duration=0.0, value=2)
ann.append(time=2, duration=0.0, value=3)
ann.append(time=3, duration=0.0, value=1)
jam.annotations.append(ann)
trans = pumpp.task.BeatTransformer(name='beat')
output = trans.transform(jam)
# Make sure we have the masks
assert np.all(
output['beat/_valid'] == [0, 4 * trans.sr // trans.hop_length])
assert output['beat/mask_downbeat']
# The first channel measures beats
# The second channel measures downbeats
assert output['beat/beat'].shape == (1, 4 * (SR // HOP_LENGTH), 1)
assert output['beat/downbeat'].shape == (1, 4 * (SR // HOP_LENGTH), 1)
# Ideal vectors:
# a beat every second (two samples)
# a downbeat every three seconds (6 samples)
beat_true = np.asarray([[1, 1, 1, 1]]).T
downbeat_true = np.asarray([[1, 0, 0, 1]]).T
assert np.all(output['beat/beat'][0, ::(SR // HOP_LENGTH)] == beat_true)
assert np.all(output['beat/downbeat'][0, ::(SR //
HOP_LENGTH)] == downbeat_true)
for key in trans.fields:
assert shape_match(output[key].shape[1:], trans.fields[key].shape)
assert type_match(output[key].dtype, trans.fields[key].dtype)
def inverse(self, encoded, downbeat=None, duration=None):
'''Inverse transformation for beats and optional downbeats'''
ann = jams.Annotation(namespace=self.namespace, duration=duration)
beat_times = np.asarray([
t for t, _ in self.decode_events(encoded,
transition=self.beat_transition,
p_init=self.beat_p_init,
p_state=self.beat_p_state) if _
])
beat_frames = time_to_frames(beat_times,
sr=self.sr,
hop_length=self.hop_length)
if downbeat is not None:
downbeat_times = set([
t
for t, _ in self.decode_events(downbeat,
transition=self.down_transition,
p_init=self.down_p_init,
p_state=self.down_p_state) if _
])
pickup_beats = len(
[t for t in beat_times if t < min(downbeat_times)])
else:
downbeat_times = set()
pickup_beats = 0
value = -pickup_beats - 1
for beat_t, beat_f in zip(beat_times, beat_frames):
if beat_t in downbeat_times:
value = 1
else:
value += 1
confidence = encoded[beat_f]
ann.append(time=beat_t,
duration=0,
value=value,
confidence=confidence)
return ann
def inverse(self, encoded, duration=None):
'''Inverse transformation'''
ann = jams.Annotation(self.namespace, duration=duration)
for start, end, value in self.decode_intervals(
encoded,
duration=duration,
multi=False,
sparse=self.sparse,
transition=self.transition,
p_init=self.p_init,
p_state=self.p_state):
# Map start:end to frames
f_start, f_end = time_to_frames([start, end],
sr=self.sr,
hop_length=self.hop_length)
# Reverse the index
if self.sparse:
# Compute the confidence
if encoded.shape[1] == 1:
# This case is for full-confidence prediction (just the index)
confidence = 1.
else:
confidence = np.mean(encoded[f_start:f_end + 1, value])
value_dec = self.encoder.inverse_transform(value)
else:
confidence = np.mean(encoded[f_start:f_end + 1,
np.argmax(value)])
value_dec = self.encoder.inverse_transform(
np.atleast_2d(value))
for vd in value_dec:
ann.append(time=start,
duration=end - start,
value=vd,
confidence=float(confidence))
return ann
def test_task_beat_nometer(SR, HOP_LENGTH):
# Construct a jam
jam = jams.JAMS(file_metadata=dict(duration=4.0))
ann = jams.Annotation(namespace='beat')
ann.append(time=0, duration=0.0)
ann.append(time=1, duration=0.0)
ann.append(time=2, duration=0.0)
ann.append(time=3, duration=0.0)
jam.annotations.append(ann)
# One second = one frame
trans = pumpp.task.BeatTransformer(name='beat')
output = trans.transform(jam)
# Make sure we have the mask
assert np.all(
output['beat/_valid'] == [0, 4 * trans.sr // trans.hop_length])
assert not output['beat/mask_downbeat']
# Check the shape: 4 seconds at 2 samples per second
assert output['beat/beat'].shape == (1, 4 * (SR // HOP_LENGTH), 1)
assert output['beat/downbeat'].shape == (1, 4 * (SR // HOP_LENGTH), 1)
# Ideal vectors:
# a beat every second (two samples)
# no downbeats
beat_true = np.asarray([1, 1, 1, 1])
downbeat_true = np.asarray([0, 0, 0, 0])
assert np.all(output['beat/beat'][0, ::(SR // HOP_LENGTH)] == beat_true)
assert np.all(output['beat/downbeat'][0, ::(SR //
HOP_LENGTH)] == downbeat_true)
for key in trans.fields:
assert shape_match(output[key].shape[1:], trans.fields[key].shape)
assert type_match(output[key].dtype, trans.fields[key].dtype)
def test_pitch_midi_to_hz():
ann = jams.Annotation(namespace='pitch_midi')
ann.append(time=0, duration=1, value=69, confidence=0.5)
ann2 = jams.convert(ann, 'pitch_hz')
ann.validate()
ann2.validate()
# Check the namespace
assert ann2.namespace == 'pitch_hz'
# midi 69 = 440.0 Hz
assert ann2.data[0].value == 440.0
# Check all else is equal
assert len(ann.data) == len(ann2.data)
for obs1, obs2 in zip(ann.data, ann2.data):
assert obs1.time == obs2.time
assert obs1.duration == obs2.duration
assert obs1.confidence == obs2.confidence
def create_annotation(values,
namespace='beat',
offset=0.0,
duration=1,
confidence=1):
ann = jams.Annotation(namespace=namespace)
time = np.arange(offset, offset + len(values))
if np.isscalar(duration):
time = time * duration
duration = [duration] * len(time)
if np.isscalar(confidence):
confidence = [confidence] * len(time)
for t, d, v, c in zip(time, duration, values, confidence):
ann.append(time=t, duration=d, value=v, confidence=c)
return ann
