def test_rand():
n = 5
m = 4
d = 2
atmp = numpy.random.random_sample((n, n))
row_sums = atmp.sum(axis=1)
a = numpy.array(atmp / row_sums[:, numpy.newaxis], dtype=numpy.double)
wtmp = numpy.random.random_sample((n, m))
row_sums = wtmp.sum(axis=1)
w = numpy.array(wtmp / row_sums[:, numpy.newaxis], dtype=numpy.double)
means = numpy.array((0.6 * numpy.random.random_sample((n, m, d)) - 0.3), dtype=numpy.double)
covars = numpy.zeros( (n,m,d,d) )
for i in range(n):
for j in range(m):
for k in range(d):
covars[i][j][k][k] = 1
pitmp = numpy.random.random_sample((n))
pi = numpy.array(pitmp / sum(pitmp), dtype=numpy.double)
gmmhmm = GMHMM(n,m,d,a,means,covars,w,pi,init_type='user',verbose=True)
obs = numpy.array((0.6 * numpy.random.random_sample((40,d)) - 0.3), dtype=numpy.double)
print("Doing Baum-welch")
gmmhmm.train(obs,1000)
print()
print("Pi",gmmhmm.pi)
print("A",gmmhmm.A)
print("weights", gmmhmm.w)
print("means", gmmhmm.means)
print("covars", gmmhmm.covars)
def test_rand():
n = 5
m = 9
d = 2
atmp = numpy.random.random_sample((n, n))
row_sums = atmp.sum(axis=1)
a = numpy.array(atmp / row_sums[:, numpy.newaxis], dtype=numpy.double)
wtmp = numpy.random.random_sample((n, m))
row_sums = wtmp.sum(axis=1)
w = numpy.array(wtmp / row_sums[:, numpy.newaxis], dtype=numpy.double)
means = numpy.array((0.6 * numpy.random.random_sample((n, m, d)) - 0.3), dtype=numpy.double)
covars = numpy.zeros( (n,m,d,d) )
for i in xrange(n):
for j in xrange(m):
for k in xrange(d):
covars[i][j][k][k] = 1
pitmp = numpy.random.random_sample((n))
pi = numpy.array(pitmp / sum(pitmp), dtype=numpy.double)
gmmhmm = GMHMM(n,m,d,a,means,covars,w,pi,init_type='user',verbose=True)
obs = numpy.array((0.6 * numpy.random.random_sample((40,d)) - 0.3), dtype=numpy.double)
print "Doing Baum-welch"
gmmhmm.train(obs,1000)
print
print "Pi",gmmhmm.pi
print "A",gmmhmm.A
print "weights", gmmhmm.w
print "means", gmmhmm.means
print "covars", gmmhmm.covars
def _constructHmmNetwork(self, numStates, ALPHA, withModels ):
'''
top level-function: costruct self.hmmNEtwork that confirms to guyz's code
'''
# sequencePhonemes = sequencePhonemes[0:4]
######## construct transition matrix
#######
if not WITH_DURATIONS:
transMAtrix = self._constructTransMatrixHMMNetwork(self.lyricsWithModels.phonemesNetwork)
# DEBUG
# writeListOfListToTextFile(transMAtrix, None , '/Users/joro/Documents/Phd/UPF/voxforge/myScripts/AlignmentStep/transMatrix')
# construct means, covars, and all the rest params
#########
if numStates == None:
numStates = len(self.lyricsWithModels.statesNetwork)
means, covars, weights, pi = self._constructHMMNetworkParameters(numStates, withModels)
if WITH_DURATIONS:
self.hmmNetwork = GMHMM(numStates,numMixtures,numDimensions,None,means,covars,weights,pi,init_type='user',verbose=True)
self.hmmNetwork.setALPHA(ALPHA)
else:
self.hmmNetwork = GMHMM(numStates,numMixtures,numDimensions,transMAtrix,means,covars,weights,pi,init_type='user',verbose=True)
def __init__(self,
n,
m,
d=1,
A=None,
means=None,
covars=None,
w=None,
pi=None,
min_std=0.01,
init_type='uniform',
precision=numpy.double,
verbose=False):
GMHMM.__init__(self, n, m, d, A, means, covars, w, pi, min_std,
init_type, precision, verbose)
def _constructHmmNetwork(self, numStates, ALPHA, withModels):
'''
top level-function: costruct self.hmmNEtwork that confirms to guyz's code
'''
# sequencePhonemes = sequencePhonemes[0:4]
######## construct transition matrix
#######
if not WITH_DURATIONS:
transMAtrix = self._constructTransMatrixHMMNetwork(
self.lyricsWithModels.phonemesNetwork)
# DEBUG
# writeListOfListToTextFile(transMAtrix, None , '/Users/joro/Documents/Phd/UPF/voxforge/myScripts/AlignmentStep/transMatrix')
# construct means, covars, and all the rest params
#########
if numStates == None:
numStates = len(self.lyricsWithModels.statesNetwork)
means, covars, weights, pi = self._constructHMMNetworkParameters(
numStates, withModels)
if WITH_DURATIONS:
self.hmmNetwork = GMHMM(numStates,
numMixtures,
numDimensions,
None,
means,
covars,
weights,
pi,
init_type='user',
verbose=True)
self.hmmNetwork.setALPHA(ALPHA)
else:
self.hmmNetwork = GMHMM(numStates,
numMixtures,
numDimensions,
transMAtrix,
means,
covars,
weights,
pi,
init_type='user',
verbose=True)
def test_simple():
n = 2
m = 2
d = 2
pi = numpy.array([0.5, 0.5])
A = numpy.ones((n, n), dtype=numpy.double) / float(n)
w = numpy.ones((n, m), dtype=numpy.double)
means = numpy.ones((n, m, d), dtype=numpy.double)
covars = [[numpy.matrix(numpy.eye(d, d)) for j in xrange(m)]
for i in xrange(n)]
w[0][0] = 0.5
w[0][1] = 0.5
w[1][0] = 0.5
w[1][1] = 0.5
means[0][0][0] = 0.5
means[0][0][1] = 0.5
means[0][1][0] = 0.5
means[0][1][1] = 0.5
means[1][0][0] = 0.5
means[1][0][1] = 0.5
means[1][1][0] = 0.5
means[1][1][1] = 0.5
gmmhmm = GMHMM(n,
m,
d,
A,
means,
covars,
w,
pi,
init_type='user',
verbose=True)
obs = numpy.array([[0.3, 0.3], [0.1, 0.1], [0.2, 0.2]])
print "Doing Baum-welch"
gmmhmm.train(obs, 10)
print
print "Pi", gmmhmm.pi
print "A", gmmhmm.A
print "weights", gmmhmm.w
print "means", gmmhmm.means
print "covars", gmmhmm.covars
def __init__(self,
n,
m,
d=1,
A=None,
means=None,
covars=None,
w=None,
pi=None,
min_std=0.01,
init_type='uniform',
precision=numpy.double,
verbose=False):
print(
"Warning: weighted EMs may not converge to local optima, since the log-likelihood function may decrease."
)
print()
GMHMM.__init__(self, n, m, d, A, means, covars, w, pi, min_std,
init_type, precision, verbose) #@UndefinedVariable
Linear.__init__(self)
def test_simple():
n = 2
m = 2
d = 2
pi = numpy.array([0.5, 0.5])
A = numpy.ones((n,n),dtype=numpy.double)/float(n)
w = numpy.ones((n,m),dtype=numpy.double)
means = numpy.ones((n,m,d),dtype=numpy.double)
covars = [[ numpy.matrix(numpy.eye(d,d)) for j in xrange(m)] for i in xrange(n)]
w[0][0] = 0.5
w[0][1] = 0.5
w[1][0] = 0.5
w[1][1] = 0.5
means[0][0][0] = 0.5
means[0][0][1] = 0.5
means[0][1][0] = 0.5
means[0][1][1] = 0.5
means[1][0][0] = 0.5
means[1][0][1] = 0.5
means[1][1][0] = 0.5
means[1][1][1] = 0.5
gmmhmm = GMHMM(n,m,d,A,means,covars,w,pi,init_type='user',verbose=True)
obs = numpy.array([ [0.3,0.3], [0.1,0.1], [0.2,0.2]])
print "Doing Baum-welch"
gmmhmm.train(obs,10)
print
print "Pi",gmmhmm.pi
print "A",gmmhmm.A
print "weights", gmmhmm.w
print "means", gmmhmm.means
print "covars", gmmhmm.covars
def trainModel(self, obs):
pi = numpy.array([0.2, 0.2, 0.2, 0.2, 0.2])
A = numpy.ones((self.n, self.n), dtype=numpy.double) / float(self.n)
w = numpy.ones((self.n, self.m), dtype=numpy.double)
means = numpy.ones((self.n, self.m, self.d), dtype=numpy.double)
covars = [[
numpy.matrix(numpy.eye(self.d, self.d)) for j in xrange(self.m)
] for i in xrange(self.n)]
n_iter = 20
'''w[0][0] = 0.5
w[0][1] = 0.5
w[1][0] = 0.5
w[1][1] = 0.5
means[0][0][0] = 0.5
means[0][0][1] = 0.5
means[0][1][0] = 0.5
means[0][1][1] = 0.5
means[1][0][0] = 0.5
means[1][0][1] = 0.5
means[1][1][0] = 0.5
means[1][1][1] = 0.5 '''
gmmhmm = GMHMM(self.n,
self.m,
self.d,
A,
means,
covars,
w,
pi,
init_type='user',
verbose=True)
print "Doing Baum-welch"
#gmmhmm.train(obs,10)
if len(obs.shape) == 2:
gmmhmm.train(obs)
return self
elif len(obs.shape) == 3:
count = obs.shape[0]
for n in range(count):
gmmhmm.train(obs[n, :, :])
return self
# w[1][0] = 0.5
# w[1][1] = 0.5
# btmp = np.random.random_sample((N, M))
# row_sums = btmp.sum(axis=1)
# B = btmp / row_sums[:, np.newaxis]
# print("Initial B: ", B)
print("HMM node started... Call a service")
#print("--- Check rows sum up to 1 ---")
HMM_model = GMHMM(N,
M,
D,
A,
means,
covars,
w,
pi,
init_type='user',
verbose=True)
def get_training_set_srv(req):
print("Start training requested")
print("Run a bag...")
global train_hmm
train_hmm = True
return
def start_training_srv(req):
model.transmat_prior = model.transmat_
model.fit(obs)
pi = model.startprob_
A = model.transmat_
w = numpy.ones((n, m), dtype=numpy.double)
hmm_means = numpy.ones((n, m, d), dtype=numpy.double)
hmm_means[0][0] = model.means_[0]
hmm_means[1][0] = model.means_[1]
hmm_means[2][0] = model.means_[2]
hmm_covars = numpy.array([[ numpy.matrix(numpy.eye(d,d)) for j in xrange(m)] for i in xrange(n)])
hmm_covars[0][0] = model.covars_[0]
hmm_covars[1][0] = model.covars_[1]
hmm_covars[2][0] = model.covars_[2]
gmmhmm = GMHMM(n,m,d,A,hmm_means,hmm_covars,w,pi,init_type='user',verbose=False)
# hidden_state = model.predict(obs)
hidden_state = gmmhmm.decode(obs)
mean_sequence = [None] * len(obs)
var_sequence = [None] * len(obs)
for i in range(len(obs)):
mean_sequence[i] = model.means_[hidden_state[i]]
var_sequence[i] = numpy.diag(model.covars_[hidden_state[i]])
means.append(mean_sequence)
vars.append(var_sequence)
hiddens.append(hidden_state)
except:
means.append(None)
w = numpy.ones((n, m), dtype=numpy.double)
hmm_means = numpy.ones((n, m, d), dtype=numpy.double)
hmm_means[0][0] = model.means_[0]
hmm_means[1][0] = model.means_[1]
hmm_means[2][0] = model.means_[2]
hmm_covars = numpy.array(
[[numpy.matrix(numpy.eye(d, d)) for j in xrange(m)]
for i in xrange(n)])
hmm_covars[0][0] = model.covars_[0]
hmm_covars[1][0] = model.covars_[1]
hmm_covars[2][0] = model.covars_[2]
gmmhmm = GMHMM(n,
m,
d,
A,
hmm_means,
hmm_covars,
w,
pi,
init_type='user',
verbose=False)
# hidden_state = model.predict(obs)
hidden_state = gmmhmm.decode(obs)
mean_sequence = [None] * len(obs)
var_sequence = [None] * len(obs)
for i in range(len(obs)):
mean_sequence[i] = model.means_[hidden_state[i]]
var_sequence[i] = numpy.diag(model.covars_[hidden_state[i]])
means.append(mean_sequence)
class Decoder(object):
'''
holds structures used in decoding and decoding result
'''
def __init__(self,
lyricsWithModels,
ALPHA,
numStates=None,
withModels=True):
'''
Constructor
'''
self.lyricsWithModels = lyricsWithModels
'''
of class HMM
'''
self.hmmNetwork = []
self._constructHmmNetwork(numStates, float(ALPHA), withModels)
self.hmmNetwork.logger.setLevel(loggingLevel)
# Path class object
self.path = None
def decodeAudio(self, observationFeatures, usePersistentFiles,
URI_recording_noExt, listDurations):
''' decode path for given exatrcted features for audio
HERE is decided which decoding scheme: with or without duration (based on WITH_DURATION parameter)
'''
if self.lyricsWithModels.ONLY_MIDDLE_STATE:
URI_bmap = URI_recording_noExt + '.bmap_onlyMiddleState'
else:
URI_bmap = URI_recording_noExt + '.bmap'
self.hmmNetwork.setPersitentFiles(usePersistentFiles, URI_bmap)
# double check that features are in same dimension as model
if observationFeatures.shape[1] != numDimensions:
sys.exit(
"dimension of feature vector should be {} but is {} ".format(
numDimensions, observationFeatures.shape[1]))
# observationFeatures = observationFeatures[0:100,:]
if WITH_DURATIONS:
# transMatrix for 0 state which is silence
transMatrix = self.lyricsWithModels.phonemesNetwork[
0].getTransMatrix()
self.hmmNetwork.setWaitProbSilState(transMatrix[2, 2])
self.hmmNetwork.setDurForStates(listDurations)
# if os.path.exists(PATH_CHI) and os.path.exists(PATH_PSI):
# chiBackPointer = numpy.loadtxt(PATH_CHI)
# psiBackPointer = numpy.loadtxt(PATH_PSI)
#
# else:
# standard viterbi forced alignment
if not WITH_DURATIONS:
path_, psi, delta = self.hmmNetwork._viterbiForced(
observationFeatures)
self.path = Path(None, None)
self.path.setPatRaw(path_)
# duration-HMM
else:
chiBackPointer, psiBackPointer = self.hmmNetwork._viterbiForcedDur(
observationFeatures)
# writeListOfListToTextFile(chiBackPointer, None , PATH_CHI)
# writeListOfListToTextFile(psiBackPointer, None , PATH_PSI)
self.path = Path(chiBackPointer, psiBackPointer)
print "\n"
# DEBUG
# self.path.printDurations()
# writeListToTextFile(self.path.pathRaw, None , '/tmp/path')
def _constructHmmNetwork(self, numStates, ALPHA, withModels):
'''
top level-function: costruct self.hmmNEtwork that confirms to guyz's code
'''
# sequencePhonemes = sequencePhonemes[0:4]
######## construct transition matrix
#######
if not WITH_DURATIONS:
transMAtrix = self._constructTransMatrixHMMNetwork(
self.lyricsWithModels.phonemesNetwork)
# DEBUG
# writeListOfListToTextFile(transMAtrix, None , '/Users/joro/Documents/Phd/UPF/voxforge/myScripts/AlignmentStep/transMatrix')
# construct means, covars, and all the rest params
#########
if numStates == None:
numStates = len(self.lyricsWithModels.statesNetwork)
means, covars, weights, pi = self._constructHMMNetworkParameters(
numStates, withModels)
if WITH_DURATIONS:
self.hmmNetwork = GMHMM(numStates,
numMixtures,
numDimensions,
None,
means,
covars,
weights,
pi,
init_type='user',
verbose=True)
self.hmmNetwork.setALPHA(ALPHA)
else:
self.hmmNetwork = GMHMM(numStates,
numMixtures,
numDimensions,
transMAtrix,
means,
covars,
weights,
pi,
init_type='user',
verbose=True)
def _constructTransMatrixHMMNetwork(self, sequencePhonemes):
'''
tranform other htkModel params to format of gyuz's hmm class
take from sequencePhonemes' attached htk models the transprobs.
'''
# just for initialization totalNumPhonemes
totalNumStates = 0
for phoneme in sequencePhonemes:
currNumStates = phoneme.htkModel.tmat.numStates - 2
totalNumStates += currNumStates
transMAtrix = numpy.zeros((totalNumStates, totalNumStates),
dtype=numpy.double)
counterOverallStateNum = 0
for phoneme in sequencePhonemes:
currNumStates = phoneme.htkModel.tmat.numStates - 2
# disregard 1st and last states from transMat because they are the non-emitting states
currTransMat = phoneme.getTransMatrix()
transMAtrix[counterOverallStateNum:counterOverallStateNum +
currNumStates,
counterOverallStateNum:counterOverallStateNum +
currNumStates] = currTransMat[1:-1, 1:-1]
# transition probability to next state
# TODO: here multiply by [0,1] matrix next state. check if it exists
nextStateTransition = 1
if (counterOverallStateNum + currNumStates) < transMAtrix.shape[1]:
val = currTransMat[-2, -1] * nextStateTransition
transMAtrix[counterOverallStateNum + currNumStates - 1,
counterOverallStateNum + currNumStates] = val
# increment in final trans matrix
counterOverallStateNum += currNumStates
return transMAtrix
def _constructHMMNetworkParameters(self,
numStates,
withModels=True,
sequenceStates=None):
'''
tranform other htkModel params to format of gyuz's hmm class
'''
means = numpy.empty((numStates, numMixtures, numDimensions))
# init covars
covars = [[
numpy.matrix(numpy.eye(numDimensions, numDimensions))
for j in xrange(numMixtures)
] for i in xrange(numStates)]
weights = numpy.ones((numStates, numMixtures), dtype=numpy.double)
# start probs :
pi = numpy.zeros((numStates), dtype=numpy.double)
# avoid log(0)
pi.fill(sys.float_info.min)
# allow to start only at first state
pi[0] = 1
# equal prob. for states to start
# pi = numpy.ones( (numStates)) *(1.0/numStates)
if not withModels:
return None, None, None, pi
sequenceStates = self.lyricsWithModels.statesNetwork
if sequenceStates == None:
sys.exit('no state sequence')
for i in range(len(sequenceStates)):
state = sequenceStates[i]
for (numMixture, weight, mixture) in state.mixtures:
weights[i, numMixture - 1] = weight
means[i, numMixture - 1, :] = mixture.mean.vector
variance_ = mixture.var.vector
for k in range(len(variance_)):
covars[i][numMixture - 1][k, k] = variance_[k]
return means, covars, weights, pi
def path2ResultWordList(self):
'''
makes sense of path indices : maps numbers to states and phonemes, uses self.lyricsWithModels.statesNetwork and self.lyricsWithModels.listWords)
to be called after decoding
'''
# indices in pathRaw
self.path._path2stateIndices()
#sanity check
numStates = len(self.lyricsWithModels.statesNetwork)
numdecodedStates = len(self.path.indicesStateStarts)
if numStates != numdecodedStates:
logging.warn(
"detected path has {} states, but stateNetwork transcript has {} states"
.format(numdecodedStates, numStates))
# num misssed states in the beginning of the path
howManyMissedStates = numStates - numdecodedStates
# WORKAROUND: assume missed states start at time 0
for i in range(howManyMissedStates):
self.path.indicesStateStarts[:0] = [0]
dummy = 0
detectedWordList = expandlyrics2Words(
self.lyricsWithModels, self.path, dummy,
_constructTimeStampsForWordDetected)
return detectedWordList
class Decoder(object):
'''
holds structures used in decoding and decoding result
'''
def __init__(self, lyricsWithModels, ALPHA, numStates=None, withModels=True):
'''
Constructor
'''
self.lyricsWithModels = lyricsWithModels
'''
of class HMM
'''
self.hmmNetwork = []
self._constructHmmNetwork(numStates, float(ALPHA), withModels)
self.hmmNetwork.logger.setLevel(loggingLevel)
# Path class object
self.path = None
def decodeAudio( self, observationFeatures, usePersistentFiles, URI_recording_noExt, listDurations):
''' decode path for given exatrcted features for audio
HERE is decided which decoding scheme: with or without duration (based on WITH_DURATION parameter)
'''
if self.lyricsWithModels.ONLY_MIDDLE_STATE:
URI_bmap = URI_recording_noExt + '.bmap_onlyMiddleState'
else:
URI_bmap = URI_recording_noExt + '.bmap'
self.hmmNetwork.setPersitentFiles( usePersistentFiles, URI_bmap )
# double check that features are in same dimension as model
if observationFeatures.shape[1] != numDimensions:
sys.exit("dimension of feature vector should be {} but is {} ".format(numDimensions, observationFeatures.shape[1]) )
# observationFeatures = observationFeatures[0:100,:]
if WITH_DURATIONS:
# transMatrix for 0 state which is silence
transMatrix = self.lyricsWithModels.phonemesNetwork[0].getTransMatrix()
self.hmmNetwork.setWaitProbSilState(transMatrix[2,2])
self.hmmNetwork.setDurForStates(listDurations)
# if os.path.exists(PATH_CHI) and os.path.exists(PATH_PSI):
# chiBackPointer = numpy.loadtxt(PATH_CHI)
# psiBackPointer = numpy.loadtxt(PATH_PSI)
#
# else:
# standard viterbi forced alignment
if not WITH_DURATIONS:
path_, psi, delta = self.hmmNetwork._viterbiForced(observationFeatures)
self.path = Path(None, None)
self.path.setPatRaw(path_)
# duration-HMM
else:
chiBackPointer, psiBackPointer = self.hmmNetwork._viterbiForcedDur(observationFeatures)
# writeListOfListToTextFile(chiBackPointer, None , PATH_CHI)
# writeListOfListToTextFile(psiBackPointer, None , PATH_PSI)
self.path = Path(chiBackPointer, psiBackPointer)
print "\n"
# DEBUG
# self.path.printDurations()
# writeListToTextFile(self.path.pathRaw, None , '/tmp/path')
def _constructHmmNetwork(self, numStates, ALPHA, withModels ):
'''
top level-function: costruct self.hmmNEtwork that confirms to guyz's code
'''
# sequencePhonemes = sequencePhonemes[0:4]
######## construct transition matrix
#######
if not WITH_DURATIONS:
transMAtrix = self._constructTransMatrixHMMNetwork(self.lyricsWithModels.phonemesNetwork)
# DEBUG
# writeListOfListToTextFile(transMAtrix, None , '/Users/joro/Documents/Phd/UPF/voxforge/myScripts/AlignmentStep/transMatrix')
# construct means, covars, and all the rest params
#########
if numStates == None:
numStates = len(self.lyricsWithModels.statesNetwork)
means, covars, weights, pi = self._constructHMMNetworkParameters(numStates, withModels)
if WITH_DURATIONS:
self.hmmNetwork = GMHMM(numStates,numMixtures,numDimensions,None,means,covars,weights,pi,init_type='user',verbose=True)
self.hmmNetwork.setALPHA(ALPHA)
else:
self.hmmNetwork = GMHMM(numStates,numMixtures,numDimensions,transMAtrix,means,covars,weights,pi,init_type='user',verbose=True)
def _constructTransMatrixHMMNetwork(self, sequencePhonemes):
'''
tranform other htkModel params to format of gyuz's hmm class
take from sequencePhonemes' attached htk models the transprobs.
'''
# just for initialization totalNumPhonemes
totalNumStates = 0
for phoneme in sequencePhonemes:
currNumStates = phoneme.htkModel.tmat.numStates - 2
totalNumStates += currNumStates
transMAtrix = numpy.zeros((totalNumStates, totalNumStates), dtype=numpy.double)
counterOverallStateNum = 0
for phoneme in sequencePhonemes:
currNumStates = phoneme.htkModel.tmat.numStates - 2
# disregard 1st and last states from transMat because they are the non-emitting states
currTransMat = phoneme.getTransMatrix()
transMAtrix[counterOverallStateNum : counterOverallStateNum + currNumStates, counterOverallStateNum : counterOverallStateNum + currNumStates ] = currTransMat[1:-1,1:-1]
# transition probability to next state
# TODO: here multiply by [0,1] matrix next state. check if it exists
nextStateTransition = 1
if (counterOverallStateNum + currNumStates) < transMAtrix.shape[1]:
val = currTransMat[-2,-1] * nextStateTransition
transMAtrix[counterOverallStateNum + currNumStates -1, counterOverallStateNum + currNumStates] = val
# increment in final trans matrix
counterOverallStateNum +=currNumStates
return transMAtrix
def _constructHMMNetworkParameters(self, numStates, withModels=True, sequenceStates=None):
'''
tranform other htkModel params to format of gyuz's hmm class
'''
means = numpy.empty((numStates, numMixtures, numDimensions))
# init covars
covars = [[ numpy.matrix(numpy.eye(numDimensions,numDimensions)) for j in xrange(numMixtures)] for i in xrange(numStates)]
weights = numpy.ones((numStates,numMixtures),dtype=numpy.double)
# start probs :
pi = numpy.zeros((numStates), dtype=numpy.double)
# avoid log(0)
pi.fill(sys.float_info.min)
# allow to start only at first state
pi[0] = 1
# equal prob. for states to start
# pi = numpy.ones( (numStates)) *(1.0/numStates)
if not withModels:
return None, None, None, pi
sequenceStates = self.lyricsWithModels.statesNetwork
if sequenceStates==None:
sys.exit('no state sequence')
for i in range(len(sequenceStates) ):
state = sequenceStates[i]
for (numMixture, weight, mixture) in state.mixtures:
weights[i,numMixture-1] = weight
means[i,numMixture-1,:] = mixture.mean.vector
variance_ = mixture.var.vector
for k in range(len( variance_) ):
covars[i][numMixture-1][k,k] = variance_[k]
return means, covars, weights, pi
def path2ResultWordList(self):
'''
makes sense of path indices : maps numbers to states and phonemes, uses self.lyricsWithModels.statesNetwork and self.lyricsWithModels.listWords)
to be called after decoding
'''
# indices in pathRaw
self.path._path2stateIndices()
#sanity check
numStates = len(self.lyricsWithModels.statesNetwork)
numdecodedStates = len(self.path.indicesStateStarts)
if numStates != numdecodedStates:
logging.warn("detected path has {} states, but stateNetwork transcript has {} states".format( numdecodedStates, numStates ) )
# num misssed states in the beginning of the path
howManyMissedStates = numStates - numdecodedStates
# WORKAROUND: assume missed states start at time 0
for i in range(howManyMissedStates):
self.path.indicesStateStarts[:0] = [0]
dummy= 0
detectedWordList = expandlyrics2Words (self.lyricsWithModels, self.path, dummy, _constructTimeStampsForWordDetected)
return detectedWordList
