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 typical_splits(basedir):
"""
Read all lab files.
Find the most common number of segments.
For that number, find the average proportion of each
segment.
Apply it to every song.
"""
# get all lab files
labfiles = get_all_files(basedir, pattern='*.lab')
# find the typical number of segments
hist_nsegs = np.zeros(2000)
for f in labfiles:
startbs, stopbs, labels = read_lab_file(f)
hist_nsegs[len(startbs)] += 1
# most common number of segments
common_nsegs = np.argmax(hist_nsegs)
print 'most common number of segments:', common_nsegs
# for that number, find typical proportion
proportions = np.zeros(common_nsegs)
count_common_segmentation = 0
for f in labfiles:
startbs, stopbs, labels = read_lab_file(f)
if len(startbs) != common_nsegs:
continue
count_common_segmentation += 1
nbeats = stopbs[-1] + 1
diffs = np.diff(startbs + [nbeats])
for k in range(common_nsegs):
proportions[k] += diffs[k] * 1. / nbeats
for k in range(common_nsegs):
proportions[k] *= 1. / count_common_segmentation
# our typical proportions are:
print 'typical proportions:', proportions
# test on all lab files
allprec = 0
allrec = 0
allf = 0
allSo = 0
allSu = 0
for f in labfiles:
startbs, stopbs, labels = read_lab_file(f)
# get ideal fake segmentation
nbeats = stopbs[-1] + 1
nbeats_per_seg = np.zeros(common_nsegs)
for k in range(common_nsegs):
nbeats_per_seg[k] = nbeats * proportions[k]
# adjust so it fits the number of beats
while np.sum(nbeats_per_seg) > nbeats:
nbeats_per_seg[np.argmax(nbeats_per_seg)] -= 1
while np.sum(nbeats_per_seg) < nbeats:
nbeats_per_seg[np.argmax(nbeats_per_seg)] += 1
# create corresponding startbs and stopbs
startbscand = np.zeros(common_nsegs)
startbscand[0] = 0
for k in range(1, common_nsegs):
startbscand[k] = startbscand[k - 1] + nbeats_per_seg[k - 1]
startbscand = list(startbscand)
stopbscand = startbscand[1:] + [nbeats - 1]
# measure
prec, rec, f, So, Su = MEASURES.prec_rec_f_So_Su(
startbs, stopbs, startbscand, stopbscand)
allprec += prec
allrec += rec
allf += f
allSo += So
allSu += Su
# done, average
allprec *= 1. / len(labfiles)
allrec *= 1. / len(labfiles)
allf *= 1. / len(labfiles)
allSo *= 1. / len(labfiles)
allSu *= 1. / len(labfiles)
# print
print 'average prec =', allprec, ', rec =', allrec, ', f =', allf, ', So =', allSo, ', Su =', allSu
cuts = get_cuts(dl,dicts)
for c in cuts:
count_cuts[c] += 1
print 'for longest dict',k,', number of cuts =',len(cuts)
# check possible remaining cuts
if k == longest_p_len:
possible_cuts = set(cuts)
else:
possible_cuts = possible_cuts.intersection(set(cuts))
print 'number of remaining cuts:',len(possible_cuts)
# get measures on segmentation
if labfile != '':
startbref, stopbref, labels = read_lab_file(labfile)
startbcand = np.unique([0] + map(lambda x:x-1,possible_cuts))
stopbcand = np.unique(list(possible_cuts) + [btchroma.shape[1]-1])
prec,rec,fval,So,Su = MEASURES.prec_rec_f_So_Su(startbref,stopbref,
startbcand,stopbcand)
print 'prec =',prec,', rec =',rec,', fval =',fval,', So =',So,', Su =',Su
# iteration done, print
if len(possible_cuts) > 0:
print 'remaining possible cuts:',possible_cuts
# print
import pylab as P
pparams = {'interpolation':'nearest','origin':'lower','cmap':P.cm.gray_r,'aspect':'auto'}
P.subplot(2,1,1)
P.imshow(btchroma,**pparams)
P.subplot(2,1,2)
P.imshow(btchroma_encoded,**pparams)
# add lines
#for c in possible_cuts:
# P.axvline(x=c-.5,ymin=0,ymax=1,color='r')
def typical_splits(basedir):
"""
Read all lab files.
Find the most common number of segments.
For that number, find the average proportion of each
segment.
Apply it to every song.
"""
# get all lab files
labfiles = get_all_files(basedir, pattern="*.lab")
# find the typical number of segments
hist_nsegs = np.zeros(2000)
for f in labfiles:
startbs, stopbs, labels = read_lab_file(f)
hist_nsegs[len(startbs)] += 1
# most common number of segments
common_nsegs = np.argmax(hist_nsegs)
print "most common number of segments:", common_nsegs
# for that number, find typical proportion
proportions = np.zeros(common_nsegs)
count_common_segmentation = 0
for f in labfiles:
startbs, stopbs, labels = read_lab_file(f)
if len(startbs) != common_nsegs:
continue
count_common_segmentation += 1
nbeats = stopbs[-1] + 1
diffs = np.diff(startbs + [nbeats])
for k in range(common_nsegs):
proportions[k] += diffs[k] * 1.0 / nbeats
for k in range(common_nsegs):
proportions[k] *= 1.0 / count_common_segmentation
# our typical proportions are:
print "typical proportions:", proportions
# test on all lab files
allprec = 0
allrec = 0
allf = 0
allSo = 0
allSu = 0
for f in labfiles:
startbs, stopbs, labels = read_lab_file(f)
# get ideal fake segmentation
nbeats = stopbs[-1] + 1
nbeats_per_seg = np.zeros(common_nsegs)
for k in range(common_nsegs):
nbeats_per_seg[k] = nbeats * proportions[k]
# adjust so it fits the number of beats
while np.sum(nbeats_per_seg) > nbeats:
nbeats_per_seg[np.argmax(nbeats_per_seg)] -= 1
while np.sum(nbeats_per_seg) < nbeats:
nbeats_per_seg[np.argmax(nbeats_per_seg)] += 1
# create corresponding startbs and stopbs
startbscand = np.zeros(common_nsegs)
startbscand[0] = 0
for k in range(1, common_nsegs):
startbscand[k] = startbscand[k - 1] + nbeats_per_seg[k - 1]
startbscand = list(startbscand)
stopbscand = startbscand[1:] + [nbeats - 1]
# measure
prec, rec, f, So, Su = MEASURES.prec_rec_f_So_Su(startbs, stopbs, startbscand, stopbscand)
allprec += prec
allrec += rec
allf += f
allSo += So
allSu += Su
# done, average
allprec *= 1.0 / len(labfiles)
allrec *= 1.0 / len(labfiles)
allf *= 1.0 / len(labfiles)
allSo *= 1.0 / len(labfiles)
allSu *= 1.0 / len(labfiles)
# print
print "average prec =", allprec, ", rec =", allrec, ", f =", allf, ", So =", allSo, ", Su =", allSu
for c in cuts:
count_cuts[c] += 1
print 'for longest dict', k, ', number of cuts =', len(cuts)
# check possible remaining cuts
if k == longest_p_len:
possible_cuts = set(cuts)
else:
possible_cuts = possible_cuts.intersection(set(cuts))
print 'number of remaining cuts:', len(possible_cuts)
# get measures on segmentation
if labfile != '':
startbref, stopbref, labels = read_lab_file(labfile)
startbcand = np.unique([0] + map(lambda x: x - 1, possible_cuts))
stopbcand = np.unique(
list(possible_cuts) + [btchroma.shape[1] - 1])
prec, rec, fval, So, Su = MEASURES.prec_rec_f_So_Su(
startbref, stopbref, startbcand, stopbcand)
print 'prec =', prec, ', rec =', rec, ', fval =', fval, ', So =', So, ', Su =', Su
# iteration done, print
if len(possible_cuts) > 0:
print 'remaining possible cuts:', possible_cuts
# print
import pylab as P
pparams = {
'interpolation': 'nearest',
'origin': 'lower',
'cmap': P.cm.gray_r,
'aspect': 'auto'
}
P.subplot(2, 1, 1)
P.imshow(btchroma, **pparams)
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