def processFlat(self):
"""Main process.
Returns
-------
est_idxs : np.array(N)
Estimated indeces the segment boundaries in frame indeces.
est_labels : np.array(N-1)
Estimated labels for the segments.
"""
# Preprocess to obtain features (array(n_frames, n_features))
F = self._preprocess()
F = librosa.util.normalize(F, axis=0)
F = librosa.feature.stack_memory(F.T).T
# F = pre.normalize(F, axis=0)
ideal_t = vmo.find_threshold(F, dim=F.shape[1])
oracle = vmo.build_oracle(F,
flag='a',
threshold=ideal_t[0][1],
dim=F.shape[1])
my_bounds, my_labels = segmentation(
oracle,
method=self.config['method'],
connectivity=self.config['connectivity'])
# Post process estimations
est_idxs, est_labels = self._postprocess(my_bounds, my_labels[:-1])
assert est_idxs[0] == 0 and est_idxs[-1] == F.shape[0] - 1
# We're done!
return est_idxs, est_labels
def vmo_routine(feature):
ideal_t = vmo.find_threshold(feature, dim=feature.shape[1])
oracle = vmo.build_oracle(feature,
flag='a',
threshold=ideal_t[0][1],
dim=feature.shape[1])
return oracle
def _self_sim_vmo(feature, d, m='lrs'):
if d == 'cosine' or 'correlation':
r = (0., 1., 0.02)
elif d == 'euclidean':
r = (0., 30., .5)
elif d == 'sqeuclidean':
r = (0., 800., 16.)
ideal_t = vmo.find_threshold(feature, r=r, flag='a', dfunc=d, dim=feature.shape[1])
oracle = vmo.build_oracle(feature,
flag='a',
threshold=ideal_t[0][1],
dfunc=d, dim=feature.shape[1])
return vse.create_selfsim(oracle, method=m)
def generate(self, audio_path, sr=44100, hop_length=512):
"""Segmentation of an audio recording using the Constant Q Transform
(CQT)
Args:
audio_path (str): A string representing the path of the audio file
sr (int): Sampling rate (default: 44100)
hop_length (int): Number of samples between successive CQT columns
(default: 512)
"""
# Extract feature
y, sr = librosa.load(audio_path, sr=sr)
cqt = librosa.core.cqt(y, sr=sr, hop_length=hop_length)
beat = range(1, len(y), 30) # create synthetic beat of 33 frames
cqt_sync = librosa.feature.sync(cqt, beat, aggregate=np.median)
cqt_stack = librosa.feature.stack_memory(cqt_sync, n_steps=3)
# Find ideal threshold
r = (0, 1.1, 0.1)
threshold = vmo.find_threshold(cqt_stack.T,
r=r,
dim=cqt_stack.shape[0])
ideal_t = threshold[0][1]
# Build Oracle
cqt_vmo = vmo.build_oracle(cqt_stack.T,
flag='a',
threshold=ideal_t,
dim=cqt_stack.shape[0])
# Symbol spectral
method = 'symbol_spectral'
est_boundaries, est_labels = van.segmentation(cqt_vmo,
method=method,
connectivity='lrs')
est_intervals = zip(est_boundaries[:-1], est_boundaries[1:])
ulabel, invind = np.unique(est_labels, return_inverse=True)
# Setting attributes
self._oracle = cqt_vmo
self._segmentation = [
Section(l, Region(i1, i2))
for l, (i1, i2) in zip(invind, est_intervals)
]
def vmo_routine(feature):
ideal_t = vmo.find_threshold(feature, dim=feature.shape[1])
oracle = vmo.build_oracle(feature, flag='a', threshold=ideal_t[0][1], dim=feature.shape[1])
return oracle
for i, p in enumerate(audio_test['pattern']):
for _p in p:
start = _p[0] - audio_test['info'][2][0]
end = _p[1] - audio_test['info'][2][0]
len_list.append(end - start)
ground[i][start:end + 1] = i + 1
min_len = int(
min(len_list) * len(subbeat_mat[ind]) /
(audio_test['info'][2][1] - audio_test['info'][2][0])) + 1
min_len_list.append(min_len)
print min_len_list
start_time = time.time()
for ind in range(5):
chroma_frames = feature_mat[ind].transpose()
r = (0.0, 1.0, 0.01)
ideal_v_inv = vmo.find_threshold(chroma_frames,
r=r,
flag='a',
dfunc='other',
dfunc_handle=trnspose_inv,
VERBOSE=False)
oracle_inv = vmo.build_oracle(chroma_frames,
flag='a',
threshold=ideal_v_inv[0][1],
feature='chroma',
dfunc='other',
dfunc_handle=trnspose_inv)
pattern = van.find_repeated_patterns(oracle_inv, lower=5)
print str(time.time() - start_time)
tempo, beats = librosa.beat.beat_track(y=y, sr=sr, hop_length=hop_size)
### Sub-Beat-Synchronous Chromagram
subbeats = []
for bs, be in zip(beats[:-1], beats[1:]):
subbeats.extend(np.linspace(bs, be, num=2, endpoint=False).astype("int").tolist())
subbeats.extend(np.linspace(beats[-1], C.shape[1], num=2, endpoint=False).astype("int").tolist())
C_sync = librosa.feature.sync(C, subbeats, aggregate=np.median)
subbeats.append(C.shape[1])
feature = np.log(C_sync + np.finfo(float).eps)
feature = pre.normalize(feature, axis=0)
### Create VMO
chroma_frames = feature.transpose()
ideal_v_inv = vmo.find_threshold(
chroma_frames, r=r, flag="a", dfunc="other", dfunc_handle=trnspose_inv, dim=chroma_frames.shape[1]
)
oracle_inv = vmo.build_oracle(
chroma_frames,
flag="a",
threshold=ideal_v_inv[0][1],
feature="chroma",
dfunc="other",
dfunc_handle=trnspose_inv,
dim=chroma_frames.shape[1],
)
### Gather Ground Truth from Dataset
ground = np.zeros((len(audio_test["pattern"]), audio_test["info"][2][1] - audio_test["info"][2][0]))
len_list = []
C = librosa.feature.chromagram(y=y, sr=sr, n_fft=fft_size, hop_length=hop_size)
tempo, beats = librosa.beat.beat_track(y=y, sr=sr, hop_length=hop_size)
### Sub-Beat-Synchronous Chromagram
subbeats = []
for bs, be in zip(beats[:-1],beats[1:]):
subbeats.extend(np.linspace(bs, be, num=2, endpoint = False).astype('int').tolist())
subbeats.extend(np.linspace(beats[-1], C.shape[1], num=2, endpoint=False).astype('int').tolist())
C_sync = librosa.feature.sync(C, subbeats, aggregate=np.median)
subbeats.append(C.shape[1])
feature = np.log(C_sync+np.finfo(float).eps)
feature = pre.normalize(feature, axis=0)
### Create VMO
chroma_frames = feature.transpose()
ideal_v_inv = vmo.find_threshold(chroma_frames, r=r, flag='a',
dfunc='other', dfunc_handle=trnspose_inv, dim=chroma_frames.shape[1])
oracle_inv= vmo.build_oracle(chroma_frames, flag='a',
threshold=ideal_v_inv[0][1],
feature='chroma', dfunc='other', dfunc_handle=trnspose_inv,
dim=chroma_frames.shape[1])
### Gather Ground Truth from Dataset
ground = np.zeros((len(audio_test['pattern']), audio_test['info'][2][1]-audio_test['info'][2][0]))
len_list = []
for i,p in enumerate(audio_test['pattern']):
for _p in p:
start = _p[0] - audio_test['info'][2][0]
end = _p[1] - audio_test['info'][2][0]
len_list.append(end-start)
ground[i][start:end+1] = i+1
plt.imshow(chroma, interpolation='nearest', aspect='auto')
plt.title('Chroma', fontsize=18)
plt.xlabel('Analysis Frame', fontsize=14)
plt.ylabel('Chroma', fontsize=14)
plt.tight_layout()
# <markdowncell>
# ### VMO - Variable Markov Oracle
# <codecell>
r = (0., 0.8, 0.02)
ideal_t = vmo.find_threshold(features,
r=r,
flag='a',
dfunc='euclidean',
dim=features.shape[1])
x_a = [i[1] for i in ideal_t[1]]
y_a = [i[0] for i in ideal_t[1]]
plt.figure()
plt.plot(x_a, y_a, linewidth=2)
plt.title('IR vs. Threshold Value', fontsize=18)
plt.grid(b='on')
plt.xlabel('Threshold', fontsize=14)
plt.ylabel('IR', fontsize=14)
# <markdowncell>
# Build the best oracle by choosing the ideal threshold (one that gives most informative oracle).
min_len_list = []
for ind in range(5):
audio_test = audio_list[ind]
ground = np.zeros((len(audio_test['pattern']), audio_test['info'][2][1]-audio_test['info'][2][0]))
len_list = []
for i,p in enumerate(audio_test['pattern']):
for _p in p:
start = _p[0] - audio_test['info'][2][0]
end = _p[1] - audio_test['info'][2][0]
len_list.append(end-start)
ground[i][start:end+1] = i+1
min_len = int(min(len_list)*len(subbeat_mat[ind])/(audio_test['info'][2][1]-audio_test['info'][2][0]))+1
min_len_list.append(min_len)
print min_len_list
start_time = time.time()
for ind in range(5):
chroma_frames = feature_mat[ind].transpose()
r = (0.0, 1.0, 0.01)
ideal_v_inv = vmo.find_threshold(chroma_frames, r = r,flag = 'a', dfunc = 'other', dfunc_handle = trnspose_inv, VERBOSE = False)
oracle_inv= vmo.build_oracle(chroma_frames, flag = 'a',
threshold = ideal_v_inv[0][1],
feature = 'chroma', dfunc = 'other', dfunc_handle = trnspose_inv)
pattern = van.find_repeated_patterns(oracle_inv, lower = 5)
print str(time.time()-start_time)