def clustering(args):
''' run clustering on a single k'''
print "Feature Analysis/Clustering Mode: single k"
feature_holder = featurevector.feature_holder(filename=FEATURE_VECTOR_FILENAME)
sones_holder = featurevector.feature_holder(filename=SONE_VECTOR_FILENAME)
k = args.k
print feature_holder
mfccs = feature_holder.get_feature('mfcc')
print sones_holder
sones = sones_holder.get_feature('sones')
centroids, distortion = Get_Best_Centroids(k, 1)
print "Distortion for this run: %0.3f" % (distortion)
classes,dist = kmeans.scipy_vq(mfccs, centroids)
# Get the inter class dist matrix
inter_class_dist_matrix = mir_utils.GetSquareDistanceMatrix(centroids)
eventBeginnings = feature_holder.get_event_start_indecies()
# write audio if given -w
if args.plot_segments:
PlotWaveformWClasses(k, feature_holder,classes)
if args.write_audio_results:
WriteAudioFromClasses(k, feature_holder, classes)
plot.plot(mfccs, sones, eventBeginnings, centroids, inter_class_dist_matrix, classes)
def clustering(args):
''' run clustering on a single k'''
print "Feature Analysis/Clustering Mode: single k"
feature_holder = featurevector.feature_holder(
filename=FEATURE_VECTOR_FILENAME)
sones_holder = featurevector.feature_holder(filename=SONE_VECTOR_FILENAME)
k = args.k
print feature_holder
mfccs = feature_holder.get_feature('mfcc')
print sones_holder
sones = sones_holder.get_feature('sones')
centroids, distortion = Get_Best_Centroids(k, 1)
print "Distortion for this run: %0.3f" % (distortion)
classes, dist = kmeans.scipy_vq(mfccs, centroids)
# Get the inter class dist matrix
inter_class_dist_matrix = mir_utils.GetSquareDistanceMatrix(centroids)
eventBeginnings = feature_holder.get_event_start_indecies()
# write audio if given -w
if args.plot_segments:
PlotWaveformWClasses(k, feature_holder, classes)
if args.write_audio_results:
WriteAudioFromClasses(k, feature_holder, classes)
plot.plot(mfccs, sones, eventBeginnings, centroids,
inter_class_dist_matrix, classes)
def Get_Best_Centroids(k, iterations):
print "Feature Analysis/Clustering Mode - feature selection from multiple k's"
feature_holder = featurevector.feature_holder(
filename=FEATURE_VECTOR_FILENAME)
mfccs = feature_holder.get_feature('mfcc')
j_measures = np.zeros(iterations)
max = 0
bestCentroids = 0
bestDistortion = 0
for i in range(iterations):
centroids, distortion = kmeans.scipy_kmeans(mfccs, k)
classes, dist = kmeans.scipy_vq(mfccs, centroids)
j_measures[i] = calcJ(mfccs, classes, centroids, k)
if j_measures[i] > max:
max = j_measures[i]
bestCentroids = centroids
bestDistortion = distortion
return bestCentroids, bestDistortion
def feature_selection(args):
''' run clustering on a range of k's'''
print "Feature Analysis/Clustering Mode - feature selection from multiple k's"
feature_holder = featurevector.feature_holder(
filename=FEATURE_VECTOR_FILENAME)
kMin = args.k_min
kMax = args.k_max
kHop = args.k_hop
mfccs = feature_holder.get_feature('mfcc')
nmfcc = len(mfccs)
print "N MFCCS:", nmfcc
results = []
for k in range(kMin, kMax, kHop):
print "Running k-Means with k=%d" % (k)
if k >= nmfcc:
print "WARNING! k is greater than the number of samples!"
centroids, distortion = kmeans.scipy_kmeans(mfccs, k)
classes, dist = kmeans.scipy_vq(mfccs, centroids)
J0 = calcJ(mfccs, classes, centroids, k)
results.append((k, distortion, dist, J0))
plot.plot_feature_selection(kMin, kMax, kHop, results)
def setUp(self):
self.nMFCC = 20
self.vector_template = [('sones', 0, 1),
('mfcc', 1, self.nMFCC)]
self.fvh = fv.feature_holder(self.vector_template)
self.check_nfeats = sum([z for x,y,z in self.vector_template])
def Get_Best_Centroids(k, iterations):
feature_holder = featurevector.feature_holder(
filename=FEATURE_VECTOR_FILENAME)
mfccs = feature_holder.get_feature('mfcc')
j_measures = np.zeros(iterations)
max = 0
bestCentroids = 0
bestDistortion = 0
for i in range(iterations):
centroids, distortion = kmeans.scipy_kmeans(mfccs, k)
classes, dist = kmeans.scipy_vq(mfccs, centroids)
j_measures[i] = calcJ(mfccs, classes, centroids, k)
if j_measures[i] > max:
max = j_measures[i]
bestCentroids = centroids
bestDistortion = distortion
'''
plt.close()
fig, (ax1) = plt.subplots(1)
ax1.plot(j_measures)
ax1.set_title("J measures over multiple iterations of k")
ax1.set_xlabel("iterations")
ax1.set_ylabel("J-measure values")
plt.show()
'''
return bestCentroids, bestDistortion
def Get_Best_Centroids(k, iterations):
feature_holder = featurevector.feature_holder(filename=FEATURE_VECTOR_FILENAME)
mfccs = feature_holder.get_feature('mfcc')
j_measures = np.zeros(iterations)
max = 0;
bestCentroids = 0
bestDistortion = 0
for i in range(iterations):
centroids, distortion = kmeans.scipy_kmeans(mfccs, k)
classes, dist = kmeans.scipy_vq(mfccs, centroids)
j_measures[i] = calcJ(mfccs, classes, centroids, k)
if j_measures[i] > max:
max = j_measures[i]
bestCentroids = centroids
bestDistortion = distortion
'''
plt.close()
fig, (ax1) = plt.subplots(1)
ax1.plot(j_measures)
ax1.set_title("J measures over multiple iterations of k")
ax1.set_xlabel("iterations")
ax1.set_ylabel("J-measure values")
plt.show()
'''
return bestCentroids, bestDistortion
def feature_selection(args):
''' run clustering on a range of k's'''
print "Feature Analysis/Clustering Mode - feature selection from multiple k's"
feature_holder = featurevector.feature_holder(filename=FEATURE_VECTOR_FILENAME)
kMin = args.k_min
kMax = args.k_max
kHop = args.k_hop
mfccs = feature_holder.get_feature('mfcc')
nmfcc = len(mfccs)
print "N MFCCS:", nmfcc
results = []
for k in range(kMin, kMax, kHop):
print "Running k-Means with k=%d" % (k)
if k >= nmfcc:
print "WARNING! k is greater than the number of samples!"
centroids, distortion = Get_Best_Centroids(k,20)
classes, dist = kmeans.scipy_vq(mfccs, centroids)
J0 = calcJ(mfccs, classes, centroids, k)
results.append( (k, distortion, dist, J0) )
plot.plot_feature_selection(kMin, kMax, kHop, results)
def Get_Best_Centroids(k, iterations):
print "Feature Analysis/Clustering Mode - feature selection from multiple k's"
feature_holder = featurevector.feature_holder(filename=FEATURE_VECTOR_FILENAME)
mfccs = feature_holder.get_feature('mfcc')
j_measures = np.zeros(iterations)
max = 0;
bestCentroids = 0
bestDistortion = 0
for i in range(iterations):
centroids, distortion = kmeans.scipy_kmeans(mfccs, k)
classes, dist = kmeans.scipy_vq(mfccs, centroids)
j_measures[i] = calcJ(mfccs, classes, centroids, k)
if j_measures[i] > max:
max = j_measures[i]
bestCentroids = centroids
bestDistortion = distortion
return bestCentroids, bestDistortion
def test_loadsave(self):
# Starts cleared
self.assertEquals(self.fvh.get_nvectors(), 0)
# Add a new vector at 3
newVect = np.random.rand(self.check_nfeats)
self.fvh.add_vector(newVect, 3)
self.assertEquals(self.fvh.get_nvectors(), 4)
self.assertTrue((self.fvh.vector[3] == newVect).all())
feature = 'mfcc'
label = (10, 15)
newFeats = np.random.randn(15, self.nMFCC)
self.fvh.add_feature(feature, newFeats, timelabel=label)
self.assertEquals(self.fvh.get_nvectors(), 19)
storedFeatures1 = self.fvh.get_feature(feature)
self.assertTrue( (self.fvh.vector[:, 1:] == storedFeatures1).all() )
storedFeatures2 = self.fvh.get_feature_by_label(label, feature)
self.assertTrue( (newFeats == storedFeatures2).all() )
file = 'features_test.npy'
self.fvh.save(file)
self.assertTrue(os.path.exists(file))
fvh2 = fv.feature_holder(self.vector_template)
fvh2.load(file)
self.assertEquals(self.fvh.vector_name_map, fvh2.vector_name_map)
self.assertEquals(self.fvh.vector_index_map, fvh2.vector_index_map)
self.assertEquals(self.fvh.vector_length, fvh2.vector_length)
self.assertTrue( (self.fvh.vector == fvh2.vector).all() )
fvh3 = fv.feature_holder(filename=file)
self.assertEquals(self.fvh.vector_name_map, fvh3.vector_name_map)
self.assertEquals(self.fvh.vector_index_map, fvh3.vector_index_map)
self.assertEquals(self.fvh.vector_length, fvh3.vector_length)
self.assertTrue( (self.fvh.vector == fvh3.vector).all() )
self.assertEquals(self.fvh.get_nvectors(), 19)
storedFeatures1 = self.fvh.get_feature(feature)
self.assertTrue( (self.fvh.vector[:, 1:] == storedFeatures1).all() )
storedFeatures2 = self.fvh.get_feature_by_label(label, feature)
self.assertTrue( (newFeats == storedFeatures2).all() )
def test_loadsave(self):
# Starts cleared
self.assertEquals(self.fvh.get_nvectors(), 0)
# Add a new vector at 3
newVect = np.random.rand(self.check_nfeats)
self.fvh.add_vector(newVect, 3)
self.assertEquals(self.fvh.get_nvectors(), 4)
self.assertTrue((self.fvh.vector[3] == newVect).all())
feature = 'mfcc'
label = (10, 15)
newFeats = np.random.randn(15, self.nMFCC)
self.fvh.add_feature(feature, newFeats, timelabel=label)
self.assertEquals(self.fvh.get_nvectors(), 19)
storedFeatures1 = self.fvh.get_feature(feature)
self.assertTrue((self.fvh.vector[:, 1:] == storedFeatures1).all())
storedFeatures2 = self.fvh.get_feature_by_label(label, feature)
self.assertTrue((newFeats == storedFeatures2).all())
file = 'features_test.npy'
self.fvh.save(file)
self.assertTrue(os.path.exists(file))
fvh2 = fv.feature_holder(self.vector_template)
fvh2.load(file)
self.assertEquals(self.fvh.vector_name_map, fvh2.vector_name_map)
self.assertEquals(self.fvh.vector_index_map, fvh2.vector_index_map)
self.assertEquals(self.fvh.vector_length, fvh2.vector_length)
self.assertTrue((self.fvh.vector == fvh2.vector).all())
fvh3 = fv.feature_holder(filename=file)
self.assertEquals(self.fvh.vector_name_map, fvh3.vector_name_map)
self.assertEquals(self.fvh.vector_index_map, fvh3.vector_index_map)
self.assertEquals(self.fvh.vector_length, fvh3.vector_length)
self.assertTrue((self.fvh.vector == fvh3.vector).all())
self.assertEquals(self.fvh.get_nvectors(), 19)
storedFeatures1 = self.fvh.get_feature(feature)
self.assertTrue((self.fvh.vector[:, 1:] == storedFeatures1).all())
storedFeatures2 = self.fvh.get_feature_by_label(label, feature)
self.assertTrue((newFeats == storedFeatures2).all())
def getfeatures(args):
''' write the extracted features from the input audio file into numpy files for reading in the other steps. '''
debug = args.debug
filepath = args.audiofile
chunk_len = args.audio_seg_length
afm = af.audiofile_manager(filepath, chunk_len)
# FFT Parameters
fs = afm.afReader.samplerate()
N = 2048
hopDenom = 2
hopSize = N / hopDenom
zp = 0
winfunc = np.hamming
fftParams = fftparams.FFTParams(fs, N, hopDenom, zp, winfunc)
# MFCC Paramters
nFilters = 40
nDCTCoefs = 20
minFreq = 50
maxFreq = 8000
nIndexSkip = 2
seglen = 1
mfccParams = fftparams.MFCCParams(nFilters, nDCTCoefs, minFreq, maxFreq,
nIndexSkip)
# Feature Vector parameters
# Template : ('name', order index, length)
vector_template = [('sones', 0, 1), ('mfcc', 1, nDCTCoefs - nIndexSkip)]
sone_template = [('sones', 0, 1)]
# Initialize the feature vector holder
feature_holder = featurevector.feature_holder(vector_template, filepath)
sone_holder = featurevector.feature_holder(sone_template, filepath)
envelopeHolder = []
audioHolder = []
sonesHolder = []
maxEnvelope = 0
count = 0
print "Feature Extraction Mode\n"
print datetime.now()
# For each chunk of audio
while afm.HasMoreData():
count += 1
audioChunk, chunkIndex = afm.GetNextSegment()
if debug:
print "Read %d sample chunk of audio (%0.2fs)" % (
len(audioChunk), len(audioChunk) / fs)
# Get Events
eventTimes, envelope = GetEvents(audioChunk, fftParams, debug)
if maxEnvelope < envelope.max():
maxEnvelope = envelope.max()
if debug: print "EVENTTIMES:", eventTimes
envelopeHolder.append(envelope)
eventTimesSamps = np.asarray(np.multiply(eventTimes, fs), dtype=int)
# Get event audio segments
eventSegments = GetEventAudioSegments(eventTimesSamps, audioChunk,
debug)
#get sones
eventSegmentSones = GetEventSones(eventSegments, fftParams, debug)
# Get the MFCCs for each segment / event
eventSegmentMFCCs = GetEventMFCCs(eventSegments, fftParams, mfccParams,
debug)
# Time-average for each segment / event
averagedEventSegmentMFCCs, averagedEventSegmentSones = AverageEventFeatures(
eventSegmentMFCCs, eventSegmentSones, seglen, fftParams, debug)
# Store these vectors in the feature_holder, labelled with their time
StoreFeatureVector(feature_holder, sone_holder,
averagedEventSegmentMFCCs,
averagedEventSegmentSones, chunkIndex, chunk_len,
eventTimes, debug)
# Write features to disk
print datetime.now()
fileSize = feature_holder.save(FEATURE_VECTOR_FILENAME)
print "Wrote", fileSize, "bytes to disk. (%s)" % (FEATURE_VECTOR_FILENAME)
fileSize = sone_holder.save(SONE_VECTOR_FILENAME)
print "Wrote", fileSize, "bytes to disk. (%s)" % (SONE_VECTOR_FILENAME)
def setUp(self):
self.nMFCC = 20
self.vector_template = [('sones', 0, 1), ('mfcc', 1, self.nMFCC)]
self.fvh = fv.feature_holder(self.vector_template)
self.check_nfeats = sum([z for x, y, z in self.vector_template])
def getfeatures(args):
''' write the extracted features from the input audio file into numpy files for reading in the other steps. '''
debug = args.debug
filepath = args.audiofile
chunk_len = args.audio_seg_length
afm = af.audiofile_manager(filepath, chunk_len)
# FFT Parameters
fs = afm.afReader.samplerate()
N = 2048
hopDenom = 2
hopSize = N/hopDenom
zp = 0
winfunc=np.hamming
fftParams = fftparams.FFTParams(fs, N, hopDenom, zp, winfunc)
# MFCC Paramters
nFilters = 40
nDCTCoefs = 20
minFreq = 50
maxFreq = 8000
nIndexSkip = 2
seglen = 1
mfccParams = fftparams.MFCCParams(nFilters, nDCTCoefs, minFreq, maxFreq, nIndexSkip)
# Feature Vector parameters
# Template : ('name', order index, length)
vector_template = [('sones', 0, 1),
('mfcc', 1, nDCTCoefs - nIndexSkip)]
sone_template = [('sones', 0, 1)]
# Initialize the feature vector holder
feature_holder = featurevector.feature_holder(vector_template, filepath)
sone_holder = featurevector.feature_holder(sone_template, filepath)
envelopeHolder = []
audioHolder = []
sonesHolder = []
maxEnvelope = 0;
count =0
print "Feature Extraction Mode\n"
print datetime.now()
# For each chunk of audio
while afm.HasMoreData():
count +=1
audioChunk, chunkIndex = afm.GetNextSegment()
if debug: print "Read %d sample chunk of audio (%0.2fs)" % (len(audioChunk), len(audioChunk) / fs)
# Get Events
eventTimes, envelope = GetEvents(audioChunk, fftParams, debug)
if maxEnvelope < envelope.max():
maxEnvelope = envelope.max()
if debug: print "EVENTTIMES:", eventTimes
envelopeHolder.append(envelope)
eventTimesSamps = np.asarray(np.multiply(eventTimes,fs),dtype=int)
# Get event audio segments
eventSegments = GetEventAudioSegments(eventTimesSamps, audioChunk, debug)
#get sones
eventSegmentSones = GetEventSones(eventSegments, fftParams, debug)
# Get the MFCCs for each segment / event
eventSegmentMFCCs = GetEventMFCCs(eventSegments, fftParams, mfccParams, debug)
# Time-average for each segment / event
averagedEventSegmentMFCCs, averagedEventSegmentSones = AverageEventFeatures(eventSegmentMFCCs, eventSegmentSones, seglen, fftParams, debug)
# Store these vectors in the feature_holder, labelled with their time
StoreFeatureVector(feature_holder, sone_holder, averagedEventSegmentMFCCs, averagedEventSegmentSones, chunkIndex, chunk_len, eventTimes, debug)
# Write features to disk
print datetime.now()
fileSize = feature_holder.save(FEATURE_VECTOR_FILENAME)
print "Wrote", fileSize, "bytes to disk. (%s)" % (FEATURE_VECTOR_FILENAME)
fileSize = sone_holder.save(SONE_VECTOR_FILENAME)
print "Wrote", fileSize, "bytes to disk. (%s)" % (SONE_VECTOR_FILENAME)
