def __init__(self, num_phonemes, num_features, var_diag_interval,
var_offdiag_interval):
self.num_features = num_features
self.num_phonemes = num_phonemes
self._labels = tuple("label%d" % i for i in range(num_phonemes))
# Randomly choose num_phonemes vectors
# KJB - need something more sophisticated here to ensure minimum
# distance between phonemes
self._targets = tuple(tuple(random() for i in range(num_features))
for j in range(num_phonemes))
def buildOneCovarMatrix():
uni = (uniform(var_offdiag_interval[0], var_offdiag_interval[1])
for i in xrange(num_features**2))
mat = numpy.fromiter(uni, 'float64').reshape(num_features, num_features)
# Make symetric
lower_tri = numpy.tril(mat) # lower triangular
mat = lower_tri + lower_tri.transpose()
# Now clobber diagonal with other values
for i in range(num_features):
mat[i,i] = uniform(var_diag_interval[0], var_diag_interval[1])
return mat
# Build covar matrices
covars = tuple(buildOneCovarMatrix() for i in range(num_phonemes))
# Create a dictionary from _labels to SimpleGaussianModels
self._models = dict()
for i,label in enumerate(self._labels):
m = SimpleGaussianModel(num_features, SimpleGaussianModel.FULL_COVARIANCE)
m.set_model(self._targets[i], covars[i])
self._models[label] = m
def __init__(self, primary_classifier, variant_labels):
self._variant_labels = variant_labels
self._variant_models = {}
for label in variant_labels:
model = SimpleGaussianModel()
# KJB - this guy shouldn't have to know so much about models -
# maybe do some kind of cloning from models in the classifier?
model.set_model(primary_classifier.num_features, 'diagonal')
self._variant_models[label] = model
self._primary = primary_classifier
self._variant_classifier = None
def __init__(self, primary_classifier, variant_labels):
self._variant_labels = variant_labels
self._variant_models = {}
for label in variant_labels:
model = SimpleGaussianModel()
# KJB - this guy shouldn't have to know so much about models -
# maybe do some kind of cloning from models in the classifier?
model.set_model(primary_classifier.num_features, 'diagonal')
self._variant_models[label] = model
self._primary = primary_classifier
self._variant_classifier = None
def test2a(num_obs, num_passes):
dimension = 2
# Data generator setup
target_means = (1,1)
target_vars = (0.1,0.1)
generator = SimpleGaussianModel(dimension, SimpleGaussianModel.DIAGONAL_COVARIANCE)
generator.set_model(target_means, target_vars)
SimpleGaussianModel.seed(0)
data = [generator.sample() for i in xrange(num_obs)]
# Gmm setup
num_mixtures = 2
gmm0 = make_gmm_diag(dimension, num_mixtures)
gmm1 = make_gmm_diag(dimension, num_mixtures)
mm = GmmMgr((gmm1,))
# Hmm setup
# A transition probability matrix with a p ~= 1 self-loop for the real state.
# The entry state feeds into the real state with p=1. We use p ~= 1 for the
# second self loop since we need *some* probability of finishing.
trans = array(((0.0, 1.0, 0.0),
(0.0, 0.999999999999, 0.000000000001),
(0.0, 0.0, 0.0)))
hmm0 = Hmm(1, log_domain=True)
hmm0.build_model(mm, (0,), 1, 1, trans)
print hmm0.to_string(True) + '\n'
print gmm0
print '\n\n'
# Try some adaptation
for p in xrange(num_passes):
mm.set_adaptation_state("INITIALIZING")
mm.clear_all_accumulators()
hmm0.begin_adapt("STANDALONE")
mm.set_adaptation_state("ACCUMULATING")
hmm0.adapt_one_sequence(data)
mm.set_adaptation_state("APPLYING")
hmm0.end_adapt()
mm.apply_all_accumulators()
mm.set_adaptation_state("NOT_ADAPTING")
really_print = False
with DebugPrint("gaussian", "gaussian_pt", "gaussian_gmm_score") if really_print else DebugPrint():
gmm0.adapt(data, max_iters = num_passes)
print hmm0.to_string(True) + '\n'
print gmm0
def test1(num_obs, num_passes):
dimension = 2
# Data generator setup
target_means = (1,1)
target_vars = (0.1,0.1)
generator = SimpleGaussianModel(dimension, SimpleGaussianModel.DIAGONAL_COVARIANCE)
generator.set_model(target_means, target_vars)
SimpleGaussianModel.seed(0)
GaussianMixtureModel.seed(0)
# Gmm setup
num_mixtures = 2
gmm0 = make_gmm(dimension, num_mixtures)
gmm1 = make_gmm(dimension, num_mixtures)
mm = GmmMgr((gmm1,))
# Hmm setup
hmm0 = Hmm(1, log_domain=True)
# A transition probability matrix with a p=1/2 exit for the real state.
# The entry state feeds into the real state with p=1.
trans = array(((0.0, 1.0, 0.0),
(0.0, 0.5, 0.5),
(0.0, 0.0, 0.0)))
hmm0.build_model(mm, (0,), 1, 1, trans)
print hmm0.to_string(True)
print gmm0
# Try some adaptation. Note that we are feeding the entire data set as one stream
# to the Hmm adaption call.
data = [generator.sample() for i in xrange(num_obs)]
for p in xrange(num_passes):
mm.set_adaptation_state("INITIALIZING")
mm.clear_all_accumulators()
hmm0.begin_adapt("STANDALONE")
mm.set_adaptation_state("ACCUMULATING")
hmm0.adapt_one_sequence(data)
mm.set_adaptation_state("APPLYING")
hmm0.end_adapt()
mm.apply_all_accumulators()
mm.set_adaptation_state("NOT_ADAPTING")
gmm0.adapt(data, max_iters = num_passes)
print hmm0.to_string(True)
print gmm0
def read_smm(stream, params, _):
covariance_type, num_components, dimension = params
v, smm_means = stream.read_array("means", rtype=float, dim=1, shape=(dimension,))
if covariance_type is GaussianMixtureModel.FULL_COVARIANCE:
var_dim = 2
var_shape = (dimension, dimension)
else:
assert covariance_type is GaussianMixtureModel.DIAGONAL_COVARIANCE
var_dim = 1
var_shape = (dimension,)
v, smm_vars = stream.read_array("vars", rtype=float, dim=var_dim, shape=var_shape)
ret = SimpleGaussianModel(dimension, covariance_type)
ret.set_model(smm_means, smm_vars)
return ret
def test0(num_obs, num_passes):
dimension = 2
# Data generator setup
target_means = (1,1)
target_vars = (0.1,0.1)
generator = SimpleGaussianModel(dimension, SimpleGaussianModel.DIAGONAL_COVARIANCE)
generator.set_model(target_means, target_vars)
SimpleGaussianModel.seed(0)
GaussianMixtureModel.seed(0)
mm = GmmMgr(dimension)
# Hmm setup
hmm0 = Hmm(0, log_domain=True)
# A transition probability matrix with no real state.
# The entry state feeds into the exit state with p=1.
trans = array(((0.0, 1.0),
(0.0, 0.0)))
hmm0.build_model(mm, (), 1, 1, trans)
print hmm0.to_string(True)
# Try some adaptation. Note that we are feeding the entire data set as one stream
# to the Hmm adaption call.
data = [generator.sample() for i in xrange(num_obs)]
for p in xrange(num_passes):
mm.set_adaptation_state("INITIALIZING")
mm.clear_all_accumulators()
hmm0.begin_adapt("STANDALONE")
mm.set_adaptation_state("ACCUMULATING")
with DebugPrint("hmm_gxfs", "hmm_aos") if False else DebugPrint():
hmm0.adapt_one_sequence(data)
mm.set_adaptation_state("APPLYING")
hmm0.end_adapt()
mm.apply_all_accumulators()
mm.set_adaptation_state("NOT_ADAPTING")
print hmm0.to_string(True)
def read_smm(stream, params, _):
covariance_type, num_components, dimension = params
v, smm_means = stream.read_array("means",
rtype=float,
dim=1,
shape=(dimension, ))
if covariance_type is GaussianMixtureModel.FULL_COVARIANCE:
var_dim = 2
var_shape = (dimension, dimension)
else:
assert (covariance_type is GaussianMixtureModel.DIAGONAL_COVARIANCE)
var_dim = 1
var_shape = (dimension, )
v, smm_vars = stream.read_array("vars",
rtype=float,
dim=var_dim,
shape=var_shape)
ret = SimpleGaussianModel(dimension, covariance_type)
ret.set_model(smm_means, smm_vars)
return ret
ncomps = 7, 3 nfeatures = numceps # mfcc, with c0: gathered from about 60,000 frames of 'train' data from Ken and Hugh skyping true_means_prime = N.array((-5.5522, -1.0517, 1.0240, 3.1609, 0.5628, -122.5789), dtype=N.float32) true_vars_prime = N.array((26.2238, 26.4064, 59.7242, 35.1180, 47.2471, 3967.2240), dtype=N.float32) false_means_prime = N.array((-9.4634, -5.3991, -4.2773, 1.7494, -0.0822, -228.6211), dtype=N.float32) false_vars_prime = N.array((3.0097, 6.0277, 8.3711, 10.7198, 13.4285, 456.7074), dtype=N.float32) # mfcc, no c0: 20,000 frames of Hugh talking true_means_prime = N.array((-4.8087, 3.9863, -0.5217, 1.3076, 0.7514, -4.6497), dtype=N.float32) true_vars_prime = N.array((26.8496, 32.6631, 32.3662, 24.2963, 36.2244, 34.1555), dtype=N.float32) false_means_prime = N.array((-6.8806, -1.3424, -3.8147, 0.4520, 0.7129, -3.1560), dtype=N.float32) false_vars_prime = N.array((2.7468, 6.2286, 7.4355, 10.1530, 13.3865, 15.9309), dtype=N.float32) true_prime = SimpleGaussianModel(nfeatures, GaussianModelBase.DIAGONAL_COVARIANCE) true_prime.set_model(true_means_prime, true_vars_prime) false_prime = SimpleGaussianModel(nfeatures, GaussianModelBase.DIAGONAL_COVARIANCE) false_prime.set_model(false_means_prime, false_vars_prime) primer = (true_prime, false_prime) GaussianMixtureModel.seed(0) gmm_mgr0 = GmmMgr(ncomps, nfeatures, GaussianModelBase.DIAGONAL_COVARIANCE, primer) gmm_mgr1 = GmmMgr(ncomps, nfeatures, GaussianModelBase.DIAGONAL_COVARIANCE, primer) classify0 = AdaptingGmmClassifier(gmm_mgr0, izip(labels, count())) classify1 = AdaptingGmmClassifier(gmm_mgr1, izip(labels, count())) classify0.set_relevance(333) classify1.set_relevance(333) classify0.set_num_em_iterations(2) classify1.set_num_em_iterations(2) classifier0 = AdaptingGmmClassProcessor(classify0)
