def siplca_btchroma(btchroma,rank=4,win=60,plotiter=0,printiter=10,niter=200,alphaZ=0):
"""
Main method, takes a beat chroma matrix, segments it using Ron's SIPLCA
RETURN
- labels, one number per frame
"""
np.random.seed(123)
labels, W, Z, H, segfun, norm= SEGMENTER.segment_song(btchroma, rank=rank,win=win,
plotiter=plotiter,
printiter=printiter,
niter=niter,alphaZ=alphaZ)
return labels, W, Z, H, segfun, norm
def siplca_btchroma(btchroma,
rank=4,
win=60,
plotiter=0,
printiter=10,
niter=200,
alphaZ=0):
"""
Main method, takes a beat chroma matrix, segments it using Ron's SIPLCA
RETURN
- labels, one number per frame
"""
np.random.seed(123)
labels, W, Z, H, segfun, norm = SEGMENTER.segment_song(btchroma,
rank=rank,
win=win,
plotiter=plotiter,
printiter=printiter,
niter=niter,
alphaZ=alphaZ)
return labels, W, Z, H, segfun, norm
def siplca_method(wavfile,rank=4,win=60,plotiter=10,printiter=10,niter=200,fullspec=False):
"""
Compute beats using Dan's code
Get the fingerprints
Align landmarks with the beats
Run SIPLCA method
Measure errors
"""
# compute beats
print 'compute beats'
x,fs = mlab.wavread(wavfile,nout=2)
x = np.average(x,axis=1)
assert x.shape[0] > 2,'bad signal averaging'
feats,beats = mlab.chrombeatftrs(x,fs,400,1,1,nout=2)
# get the fingerprints
print 'compute landmarks,',beats.shape[1],'beats found'
L,S,T,maxes = LANDMARKS.find_landmarks_from_wav(wavfile)
# LANDMARKS
if not fullspec:
# transform them into per beats features
maxessecs = get_actual_times(maxes)
print 'get features per beat,',len(maxessecs),'landmarks found'
beatfeats = get_fingerprint_feats_per_beat(beats,np.max(maxessecs)+.1,
maxes,maxessecs)
databeat = np.zeros([256,len(beatfeats)])
for bf_idx in range(len(beatfeats)):
bf = beatfeats[bf_idx]
if bf.shape[1] == 0:
continue
for k in range(bf.shape[1]):
databeat[int(bf[1,k]-1),bf_idx] += 1
print 'number of empty rows:',np.shape(np.where(databeat.sum(1)==0))[1],', removed...'
databeat = databeat[np.where(databeat.sum(1)>0)[0],:]
# FULL SPECTROGRAM
else:
# get time for each pos of the spectrogram, then beat for each pos
fakemaxes = np.zeros([2,S.shape[1]])
fakemaxes[0,:] = np.array(range(S.shape[1])).reshape(1,S.shape[1])
times = get_actual_times(fakemaxes)
# fill in databeat
#beats = np.array(beats)[0]
databeat = np.zeros([S.shape[0],beats.shape[1]])
for k in range(S.shape[1]):
t = times[k]
bs = np.where(np.array(beats)[0] > t)[0]
if bs.shape[0] == 0: # last beat
b = databeat.shape[1] - 1
else:
b = max(0,bs[0]-1)
databeat[:,b] += np.exp(S[:,k]) # remove the log for NMF
databeat -= databeat.min()
print 'full spec, max value:',databeat.max(),', shape =',databeat.shape
# launch siplca,
databeat += 1e-16
print 'launch siplca on',wavfile,', databeat.shape=',databeat.shape
np.random.seed(123)
V = databeat.copy()
V/=V.sum()
labels, W, Z, H, segfun, norm= SEGMENTER.segment_song(V, rank=rank,win=win,
plotiter=plotiter,
printiter=printiter,
niter=niter)
#res = SEGMENTER.convert_labels_to_segments(labels, beats[0])
# transform labels output to actuall startbeat and stopbeat
startbeats = [0]
stopbeats = []
currlabel = labels[0]
for k in range(1,len(labels)):
if labels[k] != currlabel:
currlabel = labels[k]
startbeats.append(k)
stopbeats.append(k-1)
stopbeats.append(len(labels)-1)
# get groundtruth
relwavfile = os.path.relpath(wavfile,start=_audio_dir)
labfile = os.path.join(_seglab_dir,relwavfile[:-4]+'.lab')
segstarts = []
fIn = open(labfile,'r')
for line in fIn.readlines():
if line == '' or line.strip() == '':
continue
segstarts.append( float(line.strip().split('\t')[0]) )
fIn.close()
refstartbeats = []
for ss in segstarts: # slow...!
for k in range(beats.shape[1]-1):
if beats[0,k] <= ss and beats[0,k+1] > ss:
refstartbeats.append(k)
break
if ss > beats[0,-1]:
refstartbeats.append(beats.shape[1]-1)
elif ss < beats[0,0]:
refstartbeats.append(0)
refstartbeats = list(np.unique(refstartbeats))
refstopbeats = list(np.array(refstartbeats[1:]) - 1) + [beats.shape[1]-1]
# measure error
prec,rec,f,So,Su = MEASURES.prec_rec_f_So_Su(refstartbeats,
refstopbeats,
startbeats,
stopbeats)
print 'prec =',prec,', rec =',rec,', f =',f,', So =',So,', Su =',Su
return prec,rec,f,So,Su
def siplca_method(wavfile,
rank=4,
win=60,
plotiter=10,
printiter=10,
niter=200,
fullspec=False):
"""
Compute beats using Dan's code
Get the fingerprints
Align landmarks with the beats
Run SIPLCA method
Measure errors
"""
# compute beats
print 'compute beats'
x, fs = mlab.wavread(wavfile, nout=2)
x = np.average(x, axis=1)
assert x.shape[0] > 2, 'bad signal averaging'
feats, beats = mlab.chrombeatftrs(x, fs, 400, 1, 1, nout=2)
# get the fingerprints
print 'compute landmarks,', beats.shape[1], 'beats found'
L, S, T, maxes = LANDMARKS.find_landmarks_from_wav(wavfile)
# LANDMARKS
if not fullspec:
# transform them into per beats features
maxessecs = get_actual_times(maxes)
print 'get features per beat,', len(maxessecs), 'landmarks found'
beatfeats = get_fingerprint_feats_per_beat(beats,
np.max(maxessecs) + .1,
maxes, maxessecs)
databeat = np.zeros([256, len(beatfeats)])
for bf_idx in range(len(beatfeats)):
bf = beatfeats[bf_idx]
if bf.shape[1] == 0:
continue
for k in range(bf.shape[1]):
databeat[int(bf[1, k] - 1), bf_idx] += 1
print 'number of empty rows:', np.shape(
np.where(databeat.sum(1) == 0))[1], ', removed...'
databeat = databeat[np.where(databeat.sum(1) > 0)[0], :]
# FULL SPECTROGRAM
else:
# get time for each pos of the spectrogram, then beat for each pos
fakemaxes = np.zeros([2, S.shape[1]])
fakemaxes[0, :] = np.array(range(S.shape[1])).reshape(1, S.shape[1])
times = get_actual_times(fakemaxes)
# fill in databeat
#beats = np.array(beats)[0]
databeat = np.zeros([S.shape[0], beats.shape[1]])
for k in range(S.shape[1]):
t = times[k]
bs = np.where(np.array(beats)[0] > t)[0]
if bs.shape[0] == 0: # last beat
b = databeat.shape[1] - 1
else:
b = max(0, bs[0] - 1)
databeat[:, b] += np.exp(S[:, k]) # remove the log for NMF
databeat -= databeat.min()
print 'full spec, max value:', databeat.max(
), ', shape =', databeat.shape
# launch siplca,
databeat += 1e-16
print 'launch siplca on', wavfile, ', databeat.shape=', databeat.shape
np.random.seed(123)
V = databeat.copy()
V /= V.sum()
labels, W, Z, H, segfun, norm = SEGMENTER.segment_song(V,
rank=rank,
win=win,
plotiter=plotiter,
printiter=printiter,
niter=niter)
#res = SEGMENTER.convert_labels_to_segments(labels, beats[0])
# transform labels output to actuall startbeat and stopbeat
startbeats = [0]
stopbeats = []
currlabel = labels[0]
for k in range(1, len(labels)):
if labels[k] != currlabel:
currlabel = labels[k]
startbeats.append(k)
stopbeats.append(k - 1)
stopbeats.append(len(labels) - 1)
# get groundtruth
relwavfile = os.path.relpath(wavfile, start=_audio_dir)
labfile = os.path.join(_seglab_dir, relwavfile[:-4] + '.lab')
segstarts = []
fIn = open(labfile, 'r')
for line in fIn.readlines():
if line == '' or line.strip() == '':
continue
segstarts.append(float(line.strip().split('\t')[0]))
fIn.close()
refstartbeats = []
for ss in segstarts: # slow...!
for k in range(beats.shape[1] - 1):
if beats[0, k] <= ss and beats[0, k + 1] > ss:
refstartbeats.append(k)
break
if ss > beats[0, -1]:
refstartbeats.append(beats.shape[1] - 1)
elif ss < beats[0, 0]:
refstartbeats.append(0)
refstartbeats = list(np.unique(refstartbeats))
refstopbeats = list(np.array(refstartbeats[1:]) - 1) + [beats.shape[1] - 1]
# measure error
prec, rec, f, So, Su = MEASURES.prec_rec_f_So_Su(refstartbeats,
refstopbeats, startbeats,
stopbeats)
print 'prec =', prec, ', rec =', rec, ', f =', f, ', So =', So, ', Su =', Su
return prec, rec, f, So, Su
