def readGroundTruth(fileList, gtExt = '.gtruth', pitchExt = '.pitchSilIntrpPP', tonicExt = '.tonicFine'):
"""
Returns the start and end time of ground truth phrases in seconds
along with the label representing the annotated name of the phrase
"""
lines = open(fileList,'r').readlines()
for ii, line in enumerate(lines):
gt_filename = line.strip() + gtExt
gt_file = np.loadtxt(gt_filename)
pitch, time, Hop = BO.readPitchFile(line.strip() + pitchExt)
tonic = np.loadtxt(line.strip() + tonicExt)
pcents = BO.PitchHz2Cents(pitch, tonic)
segment = []
time_stamps = []
count = 4
st, en = [0.0]*count, [0.0]*count
for ii in range(count):
s, e = gt_file[ii][0], gt_file[ii][1]
st[ii], en[ii] = s, e
start_ind = find_ind(time, s)
end_ind = find_ind(time, e)
#print start_ind, end_ind
time_stamp = np.arange(start_ind, end_ind)
pitch_vals = pcents[start_ind:end_ind]
time_stamps.append(time_stamp)
segment.append(pitch_vals)
return segment, time_stamps, st, en, Hop, pcents
def plotFoundMatches(fileList, pitchExt = '.pitchSilIntrpPP', tonicExt = '.tonicFine'):
"""
"""
song_str, st_seg, en_seg, spl_arr = readFullTransFile(fileList, fullTransExt = '.fullTrans')
centroids = readCentroids(centroids_file = 'centroids.npy')
aligned = getAlignment(fileList)
lines = open(fileList,'r').readlines()
for ii, line in enumerate(lines):
pitch, time, Hop = BO.readPitchFile(line.strip() + pitchExt)
tonic = np.loadtxt(line.strip() + tonicExt)
pcents = BO.PitchHz2Cents(pitch, tonic)
#recons = get_quantized_ts(st_seg/Hop, en_seg/Hop, song_str, centroids, pcents)
for ii in range(len(aligned)):
matches = aligned[ii]
print "Query index:", ii+1
count = 0
for s, e in matches:
st = st_seg[s]
en = en_seg[e]
#print st, en
contour = pcents[st/Hop:en/Hop]
plt.plot(np.arange(len(contour))*Hop, contour)
plt.ylim((-300,1100))
#plt.show()
count += 1
print "# motifs found: ", count
def plotFoundMatches(fileList,
pitchExt='.pitchSilIntrpPP',
tonicExt='.tonicFine'):
"""
"""
song_str, st_seg, en_seg, spl_arr = readFullTransFile(
fileList, fullTransExt='.fullTrans')
centroids = readCentroids(centroids_file='centroids.npy')
aligned = getAlignment(fileList)
lines = open(fileList, 'r').readlines()
for ii, line in enumerate(lines):
pitch, time, Hop = BO.readPitchFile(line.strip() + pitchExt)
tonic = np.loadtxt(line.strip() + tonicExt)
pcents = BO.PitchHz2Cents(pitch, tonic)
#recons = get_quantized_ts(st_seg/Hop, en_seg/Hop, song_str, centroids, pcents)
for ii in range(len(aligned)):
matches = aligned[ii]
print "Query index:", ii + 1
count = 0
for s, e in matches:
st = st_seg[s]
en = en_seg[e]
#print st, en
contour = pcents[st / Hop:en / Hop]
plt.plot(np.arange(len(contour)) * Hop, contour)
plt.ylim((-300, 1100))
#plt.show()
count += 1
print "# motifs found: ", count
def get_transients(fileList,
map_file,
segExt='.seg',
pitchExt='.pitchSilIntrpPP',
tonicExt='.tonicFine'):
"""
"""
map_data = pickle.load(open(map_file, 'r'))
lines = open(fileList, 'r').readlines()
ids_data = []
cnt = 0
for ii, line in enumerate(lines):
print line.strip()
seg_filename = line.strip() + segExt
seg_file = np.loadtxt(seg_filename)
pitch, time, Hop = BO.readPitchFile(line.strip() + pitchExt)
tonic = np.loadtxt(line.strip() + tonicExt)
pcents = BO.PitchHz2Cents(pitch, tonic)
#print pitch
for jj in range(seg_file.shape[0]):
if seg_file[jj][2] == -20000:
ids_data.append(map_data['file_row_to_id_map'][ii][jj])
start_time = seg_file[jj][0]
end_time = seg_file[jj][1]
#trans_id = map_data[][ii][jj]
start_ind = find_ind(time, start_time)
end_ind = find_ind(time, end_time)
segment = pitch[start_ind:end_ind]
#print len(segment)
if len(segment) >= 60:
segment_norm = polyfit_shapes_norm(segment)
if cnt == 0:
aggregate = np.array([segment_norm])
else:
aggregate = np.vstack((aggregate, segment_norm))
cnt += 1
print aggregate.shape
#plt.show()
# For training data
#------------------
#np.save('transientIds',np.array(ids_data))
#np.save('transientShapes',aggregate)
# For unknown data
#-----------------
np.save('transientIds_eval', np.array(ids_data))
np.save('transientShapes_eval', aggregate)
def get_transients(fileList, map_file, segExt = '.seg', pitchExt = '.pitchSilIntrpPP', tonicExt = '.tonicFine'):
"""
"""
map_data = pickle.load(open(map_file,'r'))
lines = open(fileList,'r').readlines()
ids_data = []
cnt = 0
for ii, line in enumerate(lines):
print line.strip()
seg_filename = line.strip() + segExt
seg_file = np.loadtxt(seg_filename)
pitch, time, Hop = BO.readPitchFile(line.strip() + pitchExt)
tonic = np.loadtxt(line.strip() + tonicExt)
pcents = BO.PitchHz2Cents(pitch, tonic)
#print pitch
for jj in range(seg_file.shape[0]):
if seg_file[jj][2] == -20000:
ids_data.append(map_data['file_row_to_id_map'][ii][jj])
start_time = seg_file[jj][0]
end_time = seg_file[jj][1]
#trans_id = map_data[][ii][jj]
start_ind = find_ind(time, start_time)
end_ind = find_ind(time, end_time)
segment = pitch[start_ind:end_ind]
#print len(segment)
if len(segment) >= 60:
segment_norm = polyfit_shapes_norm(segment)
if cnt == 0:
aggregate = np.array([segment_norm])
else:
aggregate = np.vstack((aggregate, segment_norm))
cnt += 1
print aggregate.shape
#plt.show()
# For training data
#------------------
#np.save('transientIds',np.array(ids_data))
#np.save('transientShapes',aggregate)
# For unknown data
#-----------------
np.save('transientIds_eval',np.array(ids_data))
np.save('transientShapes_eval',aggregate)
def generateLinearDataset(self, root_dir, output_dir, pitchExt, tonicExt, downsampleFactor, min_nyas_dur=-1):
pitch=np.array([])
timeInfo=np.array([])
fileInfo={}
filenames = BP.GetFileNamesInDir(root_dir,pitchExt)
for filename in filenames:
fname, ext = os.path.splitext(filename)
#reading pitch and tonic data
pitchData,timeStamps,pHop = BPO.readPitchFile(fname+pitchExt)
tonic = np.loadtxt(open(fname+tonicExt,"r"))
pCents = BPO.PitchHz2Cents(pitchData, tonic)
#some preprocessing
#removing flat regions
if (min_nyas_dur>0):
msObj = MS.nyasSegmentation()
msObj.ComputeNyasCandidates(pitchData, tonic.tolist(), pHop)
msObj.FilterNyasCandidates(min_nyas_duration=min_nyas_dur)
for swar in msObj.nyasInfo.keys():
for seg in msObj.nyasInfo[swar]:
pCents[seg[0]:seg[1]]=-5000
#downsampling
factor=downsampleFactor
pCents, pHop, timeStamps = BPO.downsamplesPitchData(pCents,pHop,timeStamps, factor)
#removing silence regions
ind_silence = np.where(pCents<-4000)[0] ###Please correct this silence condition once log eps is used
pCents = np.delete(pCents,ind_silence)
timeStamps = np.delete(timeStamps,ind_silence)
#accumulating
pitch = np.append(pitch, pCents)
timeInfo = np.append(timeInfo, timeStamps)
fileInfo[filename]= [timeInfo.size-timeStamps.size, timeInfo.size]
np.savetxt(output_dir+'/'+'AggPitch.txt', pitch, fmt='%.2f')
np.savetxt(output_dir+'/'+'AggTime.txt', timeInfo, fmt='%.2f')
stream = file(output_dir+'/'+'fileInfo.yaml','w')
yaml.dump(fileInfo, stream)
def get_quantized_ts_onlySteadyNotes(fileList, pitchExt='.pitch'):
"""
"""
lines = open(fileList, 'r').readlines()
for ii, line in enumerate(lines):
pitch, time, Hop = BO.readPitchFile(line.strip() + pitchExt)
#tonic = np.loadtxt(line.strip() + tonicExt)
#pcents = BO.PitchHz2Cents(pitch, tonic)
song_str, st_seg, en_seg, spl_arr = readFullTransFile(
fileList, fullTransExt='.fullTrans')
st, en = st_seg / Hop, en_seg / Hop
qts = np.array([None] * (max(en)))
for ii in range(len(st)):
if song_str[ii] != -100000 and song_str[ii] % 100 == 0:
qts[st[ii]:en[ii]] = song_str[ii]
#print st[ii], en[ii], song_str[ii]
#plt.plot(qts,linewidth=2)
#plt.ylim((-500,1700))
#plt.show()
#segment, time_stamps, st, en, Hop, pcents = readGroundTruth(fileList)
#for ii in range(len(segment)):
#plt.subplot(2,2,ii+1)
#plt.plot(time_stamps[ii]*Hop, segment[ii])
#plt.plot(time_stamps[ii]*Hop, qts[time_stamps[ii]], 'r', linewidth=3)
#plt.ylim((-300,1100))
#plt.show()
stylized_pitch = np.array(qts, dtype=np.float)
lines = open(fileList, 'r').readlines()
for ii, line in enumerate(lines):
filename = line.strip()
steadyContourFilename = filename + '.steadyPitch'
fid = open(steadyContourFilename, 'w')
for ii in range(len(stylized_pitch))[::2]:
#print ii, (Hop*ii), stylized_pitch[ii]
fid.write("%f\t%f" % ((Hop * ii), stylized_pitch[ii]))
fid.write('\n')
fid.close()
return qts
def get_quantized_ts_onlySteadyNotes(fileList, pitchExt = '.pitch'):
"""
"""
lines = open(fileList,'r').readlines()
for ii, line in enumerate(lines):
pitch, time, Hop = BO.readPitchFile(line.strip() + pitchExt)
#tonic = np.loadtxt(line.strip() + tonicExt)
#pcents = BO.PitchHz2Cents(pitch, tonic)
song_str, st_seg, en_seg, spl_arr = readFullTransFile(fileList, fullTransExt = '.fullTrans')
st, en = st_seg/Hop, en_seg/Hop
qts = np.array([None]*(max(en)))
for ii in range(len(st)):
if song_str[ii] != -100000 and song_str[ii]%100 == 0:
qts[st[ii]:en[ii]] = song_str[ii]
#print st[ii], en[ii], song_str[ii]
#plt.plot(qts,linewidth=2)
#plt.ylim((-500,1700))
#plt.show()
#segment, time_stamps, st, en, Hop, pcents = readGroundTruth(fileList)
#for ii in range(len(segment)):
#plt.subplot(2,2,ii+1)
#plt.plot(time_stamps[ii]*Hop, segment[ii])
#plt.plot(time_stamps[ii]*Hop, qts[time_stamps[ii]], 'r', linewidth=3)
#plt.ylim((-300,1100))
#plt.show()
stylized_pitch = np.array(qts, dtype=np.float)
lines = open(fileList,'r').readlines()
for ii, line in enumerate(lines):
filename = line.strip()
steadyContourFilename = filename + '.steadyPitch'
fid = open(steadyContourFilename,'w')
for ii in range(len(stylized_pitch))[::2]:
#print ii, (Hop*ii), stylized_pitch[ii]
fid.write("%f\t%f"%((Hop*ii),stylized_pitch[ii]))
fid.write('\n')
fid.close()
return qts
def generateSubsequenceDataset(self, pitchFile, tonicFile, outputCandidateFile):
#reading pitch data
tonic = np.loadtxt(open(tonicFile,"r"))
pitchData,timeData,pHop = BPO.readPitchFile(pitchFile)
pCents = BPO.PitchHz2Cents(pitchData, tonic)
#preprocessing pitch data
factor=3
pCents, pHop, timeData = BPO.downsamplesPitchData(pCents,pHop,timeData, factor)
if self.params.keys()[0]=='slidingWindow':
segments = self.slidingWindowCandidates(pCents, pHop, self.params['slidingWindow']['windowLength'],pHop*3)
dataset = []
[dataset.append(pCents[segment[0]:segment[1]]) for segment in segments]
np.savetxt(outputCandidateFile,np.array(dataset),fmt='%.2f')
def batchProc(root_dir,
audioExt='.mp3',
pitchExt='.pitchSilIntrpPP',
tonicExt='.tonicFine'):
filenames = BP.GetFileNamesInDir(root_dir, '.mp3')
segObj = seg.melodySegmentation()
#fig = plt.figure(figsize=(15,10), dpi=80)
for filename in filenames[:]:
print "Processing file %s" % filename
#======================
## This is done for all
#======================
fname, ext = os.path.splitext(filename)
pitch, time, Hop = BO.readPitchFile(fname + pitchExt)
tonic = np.loadtxt(fname + tonicExt)
pcents = BO.PitchHz2Cents(pitch, tonic)
pdata = (time, pcents, Hop)
## Extract Breath Phrases
#------------------------
breathPhrases = findBreathPhrases(segObj, fname, pcents, Hop)
## Histogram processing to extract note locations
#------------------------------------------------
svaraSemitone, ignoreNotes = findValidSvaras(pitch, tonic)
#print svaraSemitone, ignoreNotes
print "Notes being ignored are: %s" % ignoreNotes
## Read valid region for evolution
#---------------------------------
#endTime = readValidVistarRegion(fname)
## Svara transcription
#---------------------
transcription = transcribePitch(fname, pdata, ignoreNotes)
#print transcription
print "-------\nDone !!\n-------"
'''
def batchProc(fileList, centroids_file = 'centroids.npy', audioExt = '.mp3', pitchExt = '.pitchSilIntrpPP', tonicExt = '.tonicFine', fullTransExt = '.fullTrans', gtExt = '.gtruth'):
"""
"""
centroids = readCentroids(centroids_file)
lines = open(fileList,'r').readlines()
for ii, line in enumerate(lines):
filename = line.strip()
print "Processing file: %s" %filename
# Read pitch data
#----------------
pitch, time, Hop = BO.readPitchFile(filename + pitchExt)
tonic = np.loadtxt(line.strip() + tonicExt)
pcents = BO.PitchHz2Cents(pitch, tonic)
# Read transcription
#-------------------
song_str, st_seg, en_seg = readFullTransFile(filename, fullTransExt = fullTransExt)
# Read ground truth
#------------------
st_gt, en_gt, str_gt = visualizeGroundTruth(filename, pcents, time, Hop, song_str, st_seg, en_seg)
# Get query strings
#------------------
note_sym = getQuertString(str_gt)
# Get song string
#----------------
search_str = getSearchString(song_str)
# Get aligned contour indices by SW
#----------------------------------
aligned = getAlignment(song_str, note_sym)
# Get contour segments
#---------------------
plotFoundMatches(aligned, st_seg, en_seg, pcents, Hop)
print "-------\nDone !!\n-------"
def readGroundTruth(fileList,
gtExt='.gtruth',
pitchExt='.pitchSilIntrpPP',
tonicExt='.tonicFine'):
"""
Returns the start and end time of ground truth phrases in seconds
along with the label representing the annotated name of the phrase
"""
lines = open(fileList, 'r').readlines()
for ii, line in enumerate(lines):
gt_filename = line.strip() + gtExt
gt_file = np.loadtxt(gt_filename)
pitch, time, Hop = BO.readPitchFile(line.strip() + pitchExt)
tonic = np.loadtxt(line.strip() + tonicExt)
pcents = BO.PitchHz2Cents(pitch, tonic)
segment = []
time_stamps = []
count = 4
st, en = [0.0] * count, [0.0] * count
for ii in range(count):
s, e = gt_file[ii][0], gt_file[ii][1]
st[ii], en[ii] = s, e
start_ind = find_ind(time, s)
end_ind = find_ind(time, e)
#print start_ind, end_ind
time_stamp = np.arange(start_ind, end_ind)
pitch_vals = pcents[start_ind:end_ind]
time_stamps.append(time_stamp)
segment.append(pitch_vals)
return segment, time_stamps, st, en, Hop, pcents
def generateSubsequenceDataset(self, root_dir, output_dir, pitchExt, tonicExt, downsampleFactor, windowLength, combineData = 0, writeBinary = 1, meanNormalize = 0, flatnessThreshold=0.8, binsPOctave=120, fixPointData=1):
if combineData:
timeInfo=np.array([])
fileInfo={}
#obtaining all the files in the directory
filenames = BP.GetFileNamesInDir(root_dir,pitchExt)
# iterating over each file
for kk, filename in enumerate(filenames):
#separate file name from extension
fname, ext = os.path.splitext(filename)
audiofile = fname.split('/')[-1]
#if data is not to be combined, then output goes into individual directories (dirname = filename)
if not combineData:
out_dir = output_dir + '/' + audiofile
#creating directory if doesn't exist
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
#reading pitch and tonic data
pitchData,timeStamps,pHop = BPO.readPitchFile(fname+pitchExt)
tonic = np.loadtxt(open(fname+tonicExt,"r"))
#Convert the pitch values into cents, Note that we add a offset of 1 octave to make everything positive, assuming that pitch wont go below one octave to tonic
pCents=np.round(binsPOctave*np.log2((eps+pitchData)/tonic)).astype(np.int) + binsPOctave
#downsampling pitch data
factor=downsampleFactor
pCents, pHop, timeStamps = BPO.downsamplesPitchData(pCents,pHop,timeStamps, factor)
#removing silence regions from the pitch sequence
ind_silence = np.where(pCents<0)[0] ###Please correct this silence condition once log eps is used
pCents = np.delete(pCents,ind_silence)
timeStamps = np.delete(timeStamps,ind_silence)
#computing all the indices which have non flat region of pitch (binsPOctave is an input to decide threshold, thats it!!)
nonFlatIndexes = self.nonFlatIndexes(pCents, pHop,binsPOctave)
#making an array which will be 1 for nonflat and 0 for flat regions
flatNonflat = np.zeros(pCents.shape[0])
flatNonflat[nonFlatIndexes]=1
#given the windowLength or motigLength compute how many samples do they correspond to
windowSamples = int(np.round(windowLength/pHop))
#to create a matlab buffer like thing, we do the following
row = np.array([np.arange(windowSamples)])
col = np.array([np.arange(pCents.size-windowSamples)])
col = np.transpose(col)
col2 = copy.deepcopy(col)
col2[:]=1
ind = row*col2 + col
#so after creating matrix where each row is a subsequence, look which subsequence to reject based on flatness threshold
mtx = flatNonflat[ind]
mean_array = np.mean(mtx,axis=1)
ind_Invalid = np.where(mean_array<flatnessThreshold)[0]
ind = np.delete(ind,[ind_Invalid],axis=0)
#finally obtain pitch matrix and timestamp array
mtx = pCents[ind]
timeStamps = timeStamps[ind[:,0]]
###TODO impement mean normalizatoin
### Adding a crucial check!!
if timeStamps.shape[0] != mtx.shape[0]:
print filename
if combineData:#keep on appending the data to combine it
if kk==0:
pitch = copy.deepcopy(mtx)
timeInfo = copy.deepcopy(timeStamps)
else:
pitch = np.append(pitch, mtx,axis=0)
timeInfo = np.append(timeInfo, timeStamps)
fileInfo[filename]= [timeInfo.size-timeStamps.size, timeInfo.size]
else:#just dump for each file
if writeBinary:
if fixPointData:
mtx.astype(np.uint32).tofile(out_dir+'/'+ audiofile +'.pitchSubDBbin')
else:
mtx.astype(np.float).tofile(out_dir+'/'+ audiofile +'.pitchSubDBbin')
timeStamps.astype(np.float).tofile(out_dir+'/'+ audiofile +'.timeSubDBbin')
else:
np.savetxt(out_dir+'/'+ audiofile +'.pitchSubDBtxt', mtx , fmt='%d')
np.savetxt(out_dir+'/'+ audiofile +'.timeSubDBtxt', timeStamps, fmt='%.3f')
if combineData:
#another crucial check at each step
if pitch.shape[0] != timeInfo.shape[0]:
print filename
if combineData:
if writeBinary:
if fixPointData:
pitch.astype(np.uint32).tofile(output_dir+'/'+'AggPitch.bin')
else:
pitch.astype(np.float).tofile(output_dir+'/'+'AggPitch.bin')
timeInfo.astype(np.float).tofile(output_dir+'/'+'AggTime.bin')
else:
np.savetxt(output_dir+'/'+'AggPitch.txt', pitch , fmt='%d')
np.savetxt(output_dir+'/'+'AggTime.txt', timeInfo, fmt='%.3f')
stream = file(output_dir+'/'+'fileInfo.yaml','w')
yaml.dump(fileInfo, stream)
print "Total number of time series : " + str(pitch.shape[0])
print "Length of single Sub sequence : " + str(pitch.shape[1])
