def test_ISVTrainInitialize(): # Check that the initialization is consistent and using the rng (cf. issue #118) eps = 1e-10 # UBM GMM ubm = GMMMachine(2,3) ubm.mean_supervector = UBM_MEAN ubm.variance_supervector = UBM_VAR ## ISV ib = ISVBase(ubm, 2) # first round rng = bob.core.random.mt19937(0) it = ISVTrainer(10) #it.rng = rng it.initialize(ib, TRAINING_STATS, rng) u1 = ib.u d1 = ib.d # second round rng = bob.core.random.mt19937(0) #it.rng = rng it.initialize(ib, TRAINING_STATS, rng) u2 = ib.u d2 = ib.d assert numpy.allclose(u1, u2, eps) assert numpy.allclose(d1, d2, eps)
def test_JFATrainInitialize(): # Check that the initialization is consistent and using the rng (cf. issue #118) eps = 1e-10 # UBM GMM ubm = GMMMachine(2,3) ubm.mean_supervector = UBM_MEAN ubm.variance_supervector = UBM_VAR ## JFA jb = JFABase(ubm, 2, 2) # first round rng = bob.core.random.mt19937(0) jt = JFATrainer() #jt.rng = rng jt.initialize(jb, TRAINING_STATS, rng) u1 = jb.u v1 = jb.v d1 = jb.d # second round rng = bob.core.random.mt19937(0) jt.initialize(jb, TRAINING_STATS, rng) u2 = jb.u v2 = jb.v d2 = jb.d assert numpy.allclose(u1, u2, eps) assert numpy.allclose(v1, v2, eps) assert numpy.allclose(d1, d2, eps)
def enroll(self, data):
"""Enrolls a GMM using MAP adaptation given a reference's feature vectors
Returns a GMMMachine tuned from the UBM with MAP on a biometric reference data.
"""
# if input is a list (or SampleBatch) of 2 dimensional arrays, stack them
data = check_data_dim(data, expected_ndim=2)
# Use the array to train a GMM and return it
logger.info("Enrolling with %d feature vectors", data.shape[0])
gmm = GMMMachine(
n_gaussians=self.n_gaussians,
trainer="map",
ubm=copy.deepcopy(self),
convergence_threshold=self.convergence_threshold,
max_fitting_steps=self.enroll_iterations,
random_state=self.random_state,
update_means=self.enroll_update_means,
update_variances=self.enroll_update_variances,
update_weights=self.enroll_update_weights,
mean_var_update_threshold=self.mean_var_update_threshold,
map_relevance_factor=self.enroll_relevance_factor,
map_alpha=self.enroll_alpha,
)
gmm.fit(data)
return gmm
def test_enroll():
# Load the UBM
ubm = GMMMachine.from_hdf5(
pkg_resources.resource_filename("bob.bio.gmm.test",
"data/gmm_ubm.hdf5"))
# Create a GMM object with that UBM
gmm1 = GMM(
number_of_gaussians=2,
enroll_update_means=True,
enroll_update_variances=True,
)
gmm1.ubm = ubm
# Enroll the biometric reference from random features
enroll = utils.random_training_set((20, 45), 5, -5.0, 5.0, seed=seed_value)
biometric_reference = gmm1.enroll(enroll)
assert not biometric_reference.is_similar_to(biometric_reference.ubm)
assert isinstance(biometric_reference, GMMMachine)
reference_file = pkg_resources.resource_filename("bob.bio.gmm.test",
"data/gmm_enrolled.hdf5")
if regenerate_refs:
gmm1.write_biometric_reference(biometric_reference, reference_file)
# Compare to pre-generated file
gmm2 = gmm1.read_biometric_reference(reference_file)
assert biometric_reference.is_similar_to(gmm2)
with tempfile.NamedTemporaryFile(prefix="bob_",
suffix="_bioref.hdf5") as fd:
temp_file = fd.name
gmm1.write_biometric_reference(biometric_reference, temp_file)
assert GMMMachine.from_hdf5(temp_file, ubm).is_similar_to(gmm2)
def test_JFATrainer_updateYandV():
# test the JFATrainer for updating Y and V
v_ref = numpy.array( [0.7228, 0.7892, 0.6475, 0.6080, 0.8631, 0.8416,
1.6512, 1.6068, 0.0500, 0.0101, 0.4325, 0.6719]).reshape((6,2))
y1 = numpy.array([0., 0.])
y2 = numpy.array([0., 0.])
y3 = numpy.array([0.9630, 1.3868])
y4 = numpy.array([0.0426, -0.3721])
y=[y1, y2]
# call the updateY function
ubm = GMMMachine(2,3)
ubm.mean_supervector = UBM_MEAN
ubm.variance_supervector = UBM_VAR
m = JFABase(ubm,2,2)
t = JFATrainer()
t.initialize(m, TRAINING_STATS)
m.u = M_u
m.v = M_v
m.d = M_d
t.__X__ = M_x
t.__Y__ = y
t.__Z__ = M_z
t.e_step_v(m, TRAINING_STATS)
t.m_step_v(m, TRAINING_STATS)
# Expected results(JFA cookbook, matlab)
assert equals(t.__Y__[0], y3, 2e-4)
assert equals(t.__Y__[1], y4, 2e-4)
assert equals(m.v, v_ref, 2e-4)
def test_JFATrainInitialize():
# Check that the initialization is consistent and using the rng (cf. issue #118)
eps = 1e-10
# UBM GMM
ubm = GMMMachine(2, 3)
ubm.means = UBM_MEAN.reshape((2, 3))
ubm.variances = UBM_VAR.reshape((2, 3))
# JFA
it = JFAMachine(2, 2, em_iterations=10, ubm=ubm)
# first round
n_classes = it.estimate_number_of_classes(TRAINING_STATS_y)
it.initialize(TRAINING_STATS_X, TRAINING_STATS_y, n_classes)
u1 = it.U
v1 = it.V
d1 = it.D
# second round
it.initialize(TRAINING_STATS_X, TRAINING_STATS_y, n_classes)
u2 = it.U
v2 = it.V
d2 = it.D
np.testing.assert_allclose(u1, u2, rtol=eps, atol=1e-8)
np.testing.assert_allclose(v1, v2, rtol=eps, atol=1e-8)
np.testing.assert_allclose(d1, d2, rtol=eps, atol=1e-8)
def test_JFATrainer_updateZandD(): # test the JFATrainer for updating Z and D d_ref = numpy.array([0.3110, 1.0138, 0.8297, 1.0382, 0.0095, 0.6320]) z1 = numpy.array([0., 0., 0., 0., 0., 0.]) z2 = numpy.array([0., 0., 0., 0., 0., 0.]) z3_ref = numpy.array([0.3256, 1.8633, 0.6480, 0.8085, -0.0432, 0.2885]) z4_ref = numpy.array([-0.3324, -0.1474, -0.4404, -0.4529, 0.0484, -0.5848]) z=[z1, z2] # call the updateZ function ubm = GMMMachine(2,3) ubm.mean_supervector = UBM_MEAN ubm.variance_supervector = UBM_VAR m = JFABase(ubm,2,2) t = JFATrainer() t.initialize(m, TRAINING_STATS) m.u = M_u m.v = M_v m.d = M_d t.__X__ = M_x t.__Y__ = M_y t.__Z__ = z t.e_step_d(m, TRAINING_STATS) t.m_step_d(m, TRAINING_STATS) # Expected results(JFA cookbook, matlab) assert equals(t.__Z__[0], z3_ref, 2e-4) assert equals(t.__Z__[1], z4_ref, 2e-4) assert equals(m.d, d_ref, 2e-4)
def test_JFATrainer_updateXandU():
# test the JFATrainer for updating X and U
u_ref = numpy.array([
0.6729, 0.3408, 0.0544, 1.0653, 0.5399, 1.3035, 2.4995, 0.4385, 0.1292,
-0.0576, 1.1962, 0.0117
]).reshape((6, 2))
x1 = numpy.array([0., 0., 0., 0.]).reshape((2, 2))
x2 = numpy.array([0., 0., 0., 0.]).reshape((2, 2))
x3 = numpy.array([0.2143, 1.8275, 3.1979, 0.1227]).reshape((2, 2))
x4 = numpy.array([-1.3861, 0.2359, 5.3326, -0.7914]).reshape((2, 2))
x = [x1, x2]
# call the updateX function
ubm = GMMMachine(2, 3)
ubm.mean_supervector = UBM_MEAN
ubm.variance_supervector = UBM_VAR
m = JFABase(ubm, 2, 2)
t = JFATrainer()
t.initialize(m, TRAINING_STATS)
m.u = M_u
m.v = M_v
m.d = M_d
t.__X__ = x
t.__Y__ = M_y
t.__Z__ = M_z
t.e_step_u(m, TRAINING_STATS)
t.m_step_u(m, TRAINING_STATS)
# Expected results(JFA cookbook, matlab)
assert equals(t.__X__[0], x3, 2e-4)
assert equals(t.__X__[1], x4, 2e-4)
assert equals(m.u, u_ref, 2e-4)
def test_JFATrainer_updateXandU():
# test the JFATrainer for updating X and U
u_ref = numpy.array( [0.6729, 0.3408, 0.0544, 1.0653, 0.5399, 1.3035,
2.4995, 0.4385, 0.1292, -0.0576, 1.1962, 0.0117]).reshape((6,2))
x1 = numpy.array([0., 0., 0., 0.]).reshape((2,2))
x2 = numpy.array([0., 0., 0., 0.]).reshape((2,2))
x3 = numpy.array([0.2143, 1.8275, 3.1979, 0.1227]).reshape((2,2))
x4 = numpy.array([-1.3861, 0.2359, 5.3326, -0.7914]).reshape((2,2))
x = [x1, x2]
# call the updateX function
ubm = GMMMachine(2,3)
ubm.mean_supervector = UBM_MEAN
ubm.variance_supervector = UBM_VAR
m = JFABase(ubm,2,2)
t = JFATrainer()
t.initialize(m, TRAINING_STATS)
m.u = M_u
m.v = M_v
m.d = M_d
t.__X__ = x
t.__Y__ = M_y
t.__Z__ = M_z
t.e_step_u(m, TRAINING_STATS)
t.m_step_u(m, TRAINING_STATS)
# Expected results(JFA cookbook, matlab)
assert equals(t.__X__[0], x3, 2e-4)
assert equals(t.__X__[1], x4, 2e-4)
assert equals(m.u, u_ref, 2e-4)
def test_JFATrainInitialize():
# Check that the initialization is consistent and using the rng (cf. issue #118)
eps = 1e-10
# UBM GMM
ubm = GMMMachine(2, 3)
ubm.mean_supervector = UBM_MEAN
ubm.variance_supervector = UBM_VAR
## JFA
jb = JFABase(ubm, 2, 2)
# first round
rng = bob.core.random.mt19937(0)
jt = JFATrainer()
#jt.rng = rng
jt.initialize(jb, TRAINING_STATS, rng)
u1 = jb.u
v1 = jb.v
d1 = jb.d
# second round
rng = bob.core.random.mt19937(0)
jt.initialize(jb, TRAINING_STATS, rng)
u2 = jb.u
v2 = jb.v
d2 = jb.d
assert numpy.allclose(u1, u2, eps)
assert numpy.allclose(v1, v2, eps)
assert numpy.allclose(d1, d2, eps)
def test_JFATrainer_updateZandD():
# test the JFATrainer for updating Z and D
d_ref = numpy.array([0.3110, 1.0138, 0.8297, 1.0382, 0.0095, 0.6320])
z1 = numpy.array([0., 0., 0., 0., 0., 0.])
z2 = numpy.array([0., 0., 0., 0., 0., 0.])
z3_ref = numpy.array([0.3256, 1.8633, 0.6480, 0.8085, -0.0432, 0.2885])
z4_ref = numpy.array([-0.3324, -0.1474, -0.4404, -0.4529, 0.0484, -0.5848])
z = [z1, z2]
# call the updateZ function
ubm = GMMMachine(2, 3)
ubm.mean_supervector = UBM_MEAN
ubm.variance_supervector = UBM_VAR
m = JFABase(ubm, 2, 2)
t = JFATrainer()
t.initialize(m, TRAINING_STATS)
m.u = M_u
m.v = M_v
m.d = M_d
t.__X__ = M_x
t.__Y__ = M_y
t.__Z__ = z
t.e_step_d(m, TRAINING_STATS)
t.m_step_d(m, TRAINING_STATS)
# Expected results(JFA cookbook, matlab)
assert equals(t.__Z__[0], z3_ref, 2e-4)
assert equals(t.__Z__[1], z4_ref, 2e-4)
assert equals(m.d, d_ref, 2e-4)
def test_JFATrainer_updateYandV():
# test the JFATrainer for updating Y and V
v_ref = numpy.array([
0.7228, 0.7892, 0.6475, 0.6080, 0.8631, 0.8416, 1.6512, 1.6068, 0.0500,
0.0101, 0.4325, 0.6719
]).reshape((6, 2))
y1 = numpy.array([0., 0.])
y2 = numpy.array([0., 0.])
y3 = numpy.array([0.9630, 1.3868])
y4 = numpy.array([0.0426, -0.3721])
y = [y1, y2]
# call the updateY function
ubm = GMMMachine(2, 3)
ubm.mean_supervector = UBM_MEAN
ubm.variance_supervector = UBM_VAR
m = JFABase(ubm, 2, 2)
t = JFATrainer()
t.initialize(m, TRAINING_STATS)
m.u = M_u
m.v = M_v
m.d = M_d
t.__X__ = M_x
t.__Y__ = y
t.__Z__ = M_z
t.e_step_v(m, TRAINING_STATS)
t.m_step_v(m, TRAINING_STATS)
# Expected results(JFA cookbook, matlab)
assert equals(t.__Y__[0], y3, 2e-4)
assert equals(t.__Y__[1], y4, 2e-4)
assert equals(m.v, v_ref, 2e-4)
def test_ISVTrainInitialize():
# Check that the initialization is consistent and using the rng (cf. issue #118)
eps = 1e-10
# UBM GMM
ubm = GMMMachine(2, 3)
ubm.mean_supervector = UBM_MEAN
ubm.variance_supervector = UBM_VAR
## ISV
ib = ISVBase(ubm, 2)
# first round
rng = bob.core.random.mt19937(0)
it = ISVTrainer(10)
#it.rng = rng
it.initialize(ib, TRAINING_STATS, rng)
u1 = ib.u
d1 = ib.d
# second round
rng = bob.core.random.mt19937(0)
#it.rng = rng
it.initialize(ib, TRAINING_STATS, rng)
u2 = ib.u
d2 = ib.d
assert numpy.allclose(u1, u2, eps)
assert numpy.allclose(d1, d2, eps)
def test_ISVBase():
# Creates a UBM
weights = numpy.array([0.4, 0.6], 'float64')
means = numpy.array([[1, 6, 2], [4, 3, 2]], 'float64')
variances = numpy.array([[1, 2, 1], [2, 1, 2]], 'float64')
ubm = GMMMachine(2,3)
ubm.weights = weights
ubm.means = means
ubm.variances = variances
# Creates a ISVBase
U = numpy.array([[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12]], 'float64')
d = numpy.array([0, 1, 0, 1, 0, 1], 'float64')
m = ISVBase(ubm, ru=1)
_,_,ru = m.shape
assert ru == 1
# Checks for correctness
m.resize(2)
m.u = U
m.d = d
n_gaussians,dim,ru = m.shape
supervector_length = m.supervector_length
assert (m.u == U).all()
assert (m.d == d).all()
assert n_gaussians == 2
assert dim == 3
assert supervector_length == 6
assert ru == 2
# Saves and loads
filename = str(tempfile.mkstemp(".hdf5")[1])
m.save(bob.io.base.HDF5File(filename, 'w'))
m_loaded = ISVBase(bob.io.base.HDF5File(filename))
m_loaded.ubm = ubm
assert m == m_loaded
assert (m != m_loaded) is False
assert m.is_similar_to(m_loaded)
# Copy constructor
mc = ISVBase(m)
assert m == mc
# Variant
#mv = ISVBase()
# Checks for correctness
#mv.ubm = ubm
#mv.resize(2)
#mv.u = U
#mv.d = d
#assert (m.u == U).all()
#assert (m.d == d).all()
#ssert m.dim_c == 2
#assert m.dim_d == 3
#assert m.dim_cd == 6
#assert m.dim_ru == 2
# Clean-up
os.unlink(filename)
def test_ml_em():
# Simple GMM test
data = np.array([[1, 2, 2], [2, 1, 2], [7, 8, 9], [7, 7, 8], [7, 9, 7]])
n_gaussians = 2
n_features = data.shape[-1]
machine = GMMMachine(
n_gaussians,
update_means=True,
update_variances=True,
update_weights=True,
)
machine.means = np.repeat([[2], [8]], n_features, 1)
machine.variances = np.ones_like(machine.means)
stats = gmm_module.e_step(
data,
machine,
)
gmm_module.m_step(
[stats],
machine,
)
expected_means = np.array([[1.5, 1.5, 2.0], [7.0, 8.0, 8.0]])
np.testing.assert_almost_equal(machine.means, expected_means)
expected_weights = np.array([2 / 5, 3 / 5])
np.testing.assert_almost_equal(machine.weights, expected_weights)
eps = np.finfo(float).eps
expected_variances = np.array([[1 / 4, 1 / 4, eps], [eps, 2 / 3, 2 / 3]])
np.testing.assert_almost_equal(machine.variances, expected_variances)
def test_ISVTrainInitialize():
# Check that the initialization is consistent and using the rng (cf. issue #118)
eps = 1e-10
# UBM GMM
ubm = GMMMachine(2, 3)
ubm.means = UBM_MEAN.reshape((2, 3))
ubm.variances = UBM_VAR.reshape((2, 3))
# ISV
it = ISVMachine(2, em_iterations=10, ubm=ubm)
# it.rng = rng
n_classes = it.estimate_number_of_classes(TRAINING_STATS_y)
it.initialize(TRAINING_STATS_X, TRAINING_STATS_y, n_classes)
u1 = copy.deepcopy(it.U)
d1 = copy.deepcopy(it.D)
# second round
it.initialize(TRAINING_STATS_X, TRAINING_STATS_y, n_classes)
u2 = it.U
d2 = it.D
np.testing.assert_allclose(u1, u2, rtol=eps, atol=1e-8)
np.testing.assert_allclose(d1, d2, rtol=eps, atol=1e-8)
def enroll(self, data):
"""Enrolls a GMM using MAP adaptation given a reference's feature vectors
Returns a GMMMachine tuned from the UBM with MAP on a biometric reference data.
"""
for feature in data:
self._check_feature(feature)
# if input is a list (or SampleBatch) of 2 dimensional arrays, stack them
if data[0].ndim == 2:
data = np.vstack(data)
# Use the array to train a GMM and return it
logger.info("Enrolling with %d feature vectors", data.shape[0])
gmm = GMMMachine(
n_gaussians=self.number_of_gaussians,
trainer="map",
ubm=copy.deepcopy(self.ubm),
convergence_threshold=self.training_threshold,
max_fitting_steps=self.gmm_enroll_iterations,
random_state=self.rng,
update_means=self.enroll_update_means,
update_variances=self.enroll_update_variances,
update_weights=self.enroll_update_weights,
mean_var_update_threshold=self.variance_threshold,
map_relevance_factor=self.enroll_relevance_factor,
map_alpha=self.enroll_alpha,
)
gmm.fit(data)
return gmm
def _create_ubm_prior(means):
# Creating a fake prior with 2 gaussians
prior_gmm = GMMMachine(2)
prior_gmm.means = means.copy()
# All nice and round diagonal covariance
prior_gmm.variances = np.ones((2, 3)) * 0.5
prior_gmm.weights = np.array([0.3, 0.7])
return prior_gmm
def test_ISVTrainAndEnrol():
# Train and enroll an 'ISVMachine'
eps = 1e-10
d_ref = numpy.array([
0.39601136, 0.07348469, 0.47712682, 0.44738127, 0.43179856, 0.45086029
], 'float64')
u_ref = numpy.array([[0.855125642430777, 0.563104284748032],
[-0.325497865404680, 1.923598985291687],
[0.511575659503837, 1.964288663083095],
[9.330165761678115, 1.073623827995043],
[0.511099245664012, 0.278551249248978],
[5.065578541930268, 0.509565618051587]], 'float64')
z_ref = numpy.array([
-0.079315777443826, 0.092702428248543, -0.342488761656616,
-0.059922635809136, 0.133539981073604, 0.213118695516570
], 'float64')
# Calls the train function
ubm = GMMMachine(2, 3)
ubm.mean_supervector = UBM_MEAN
ubm.variance_supervector = UBM_VAR
mb = ISVBase(ubm, 2)
t = ISVTrainer(4.)
t.initialize(mb, TRAINING_STATS)
mb.u = M_u
for i in range(10):
t.e_step(mb, TRAINING_STATS)
t.m_step(mb)
assert numpy.allclose(mb.d, d_ref, eps)
assert numpy.allclose(mb.u, u_ref, eps)
# Calls the enroll function
m = ISVMachine(mb)
Ne = numpy.array([0.1579, 0.9245, 0.1323, 0.2458]).reshape((2, 2))
Fe = numpy.array([
0.1579, 0.1925, 0.3242, 0.1234, 0.2354, 0.2734, 0.2514, 0.5874, 0.3345,
0.2463, 0.4789, 0.5236
]).reshape((6, 2))
gse1 = GMMStats(2, 3)
gse1.n = Ne[:, 0]
gse1.sum_px = Fe[:, 0].reshape(2, 3)
gse2 = GMMStats(2, 3)
gse2.n = Ne[:, 1]
gse2.sum_px = Fe[:, 1].reshape(2, 3)
gse = [gse1, gse2]
t.enroll(m, gse, 5)
assert numpy.allclose(m.z, z_ref, eps)
#Testing exceptions
nose.tools.assert_raises(RuntimeError, t.initialize, mb, [1, 2, 2])
nose.tools.assert_raises(RuntimeError, t.initialize, mb, [[1, 2, 2]])
nose.tools.assert_raises(RuntimeError, t.e_step, mb, [1, 2, 2])
nose.tools.assert_raises(RuntimeError, t.e_step, mb, [[1, 2, 2]])
nose.tools.assert_raises(RuntimeError, t.enroll, m, [[1, 2, 2]], 5)
def loadGMM():
gmm = GMMMachine(2, 2)
gmm.weights = bob.io.base.load(datafile('gmm.init_weights.hdf5', __name__, path="../data/"))
gmm.means = bob.io.base.load(datafile('gmm.init_means.hdf5', __name__, path="../data/"))
gmm.variances = bob.io.base.load(datafile('gmm.init_variances.hdf5', __name__, path="../data/"))
#gmm.variance_thresholds = numpy.array([[0.001, 0.001],[0.001, 0.001]], 'float64')
return gmm
def test_gmm_MAP_3():
# Train a GMMMachine with MAP_GMMTrainer; compares to old reference
ar = load_array(resource_filename("bob.learn.em", "data/dataforMAP.hdf5"))
# Initialize GMMMachine
n_gaussians = 5
prior_gmm = GMMMachine(n_gaussians)
prior_gmm.means = load_array(
resource_filename("bob.learn.em", "data/meansAfterML.hdf5"))
prior_gmm.variances = load_array(
resource_filename("bob.learn.em", "data/variancesAfterML.hdf5"))
prior_gmm.weights = load_array(
resource_filename("bob.learn.em", "data/weightsAfterML.hdf5"))
threshold = 0.001
prior_gmm.variance_thresholds = threshold
# Initialize MAP Trainer
prior = 0.001
accuracy = 0.00001
gmm = GMMMachine(
n_gaussians,
trainer="map",
ubm=prior_gmm,
convergence_threshold=prior,
max_fitting_steps=1,
update_means=True,
update_variances=False,
update_weights=False,
mean_var_update_threshold=accuracy,
map_relevance_factor=None,
)
gmm.variance_thresholds = threshold
# Test results
# Load torch3vision reference
meansMAP_ref = load_array(
resource_filename("bob.learn.em", "data/meansAfterMAP.hdf5"))
variancesMAP_ref = load_array(
resource_filename("bob.learn.em", "data/variancesAfterMAP.hdf5"))
weightsMAP_ref = load_array(
resource_filename("bob.learn.em", "data/weightsAfterMAP.hdf5"))
for transform in (to_numpy, to_dask_array):
ar = transform(ar)
# Train
gmm = gmm.fit(ar)
# Compare to current results
# Gaps are quite large. This might be explained by the fact that there is no
# adaptation of a given Gaussian in torch3 when the corresponding responsibilities
# are below the responsibilities threshold
np.testing.assert_allclose(gmm.means, meansMAP_ref, atol=2e-1)
np.testing.assert_allclose(gmm.variances, variancesMAP_ref, atol=1e-4)
np.testing.assert_allclose(gmm.weights, weightsMAP_ref, atol=1e-4)
def loadGMM():
gmm = GMMMachine(n_gaussians=2)
gmm.weights = load_array(
resource_filename("bob.learn.em", "data/gmm.init_weights.hdf5"))
gmm.means = load_array(
resource_filename("bob.learn.em", "data/gmm.init_means.hdf5"))
gmm.variances = load_array(
resource_filename("bob.learn.em", "data/gmm.init_variances.hdf5"))
return gmm
def loadGMM():
gmm = GMMMachine(2, 2)
gmm.weights = bob.io.base.load(
datafile('gmm.init_weights.hdf5', __name__, path="../data/"))
gmm.means = bob.io.base.load(
datafile('gmm.init_means.hdf5', __name__, path="../data/"))
gmm.variances = bob.io.base.load(
datafile('gmm.init_variances.hdf5', __name__, path="../data/"))
#gmm.variance_thresholds = numpy.array([[0.001, 0.001],[0.001, 0.001]], 'float64')
return gmm
def test_GMMMachine_3():
# Test a GMMMachine (log-likelihood computation)
data = bob.io.base.load(datafile('data.hdf5', __name__, path="../data/"))
gmm = GMMMachine(2, 50)
gmm.weights = bob.io.base.load(datafile('weights.hdf5', __name__, path="../data/"))
gmm.means = bob.io.base.load(datafile('means.hdf5', __name__, path="../data/"))
gmm.variances = bob.io.base.load(datafile('variances.hdf5', __name__, path="../data/"))
# Compare the log-likelihood with the one obtained using Chris Matlab
# implementation
matlab_ll_ref = -2.361583051672024e+02
assert abs(gmm(data) - matlab_ll_ref) < 1e-10
def test_gmm_MAP_2():
# Train a GMMMachine with MAP_GMMTrainer and compare with matlab reference
data = bob.io.base.load(datafile('data.hdf5', __name__, path="../data/"))
data = data.reshape((1, data.shape[0])) # make a 2D array out of it
means = bob.io.base.load(datafile('means.hdf5', __name__, path="../data/"))
variances = bob.io.base.load(datafile('variances.hdf5', __name__, path="../data/"))
weights = bob.io.base.load(datafile('weights.hdf5', __name__, path="../data/"))
gmm = GMMMachine(2,50)
gmm.means = means
gmm.variances = variances
gmm.weights = weights
map_adapt = MAP_GMMTrainer(update_means=True, update_variances=False, update_weights=False, mean_var_update_responsibilities_threshold=0.,prior_gmm=gmm, relevance_factor=4.)
gmm_adapted = GMMMachine(2,50)
gmm_adapted.means = means
gmm_adapted.variances = variances
gmm_adapted.weights = weights
#map_adapt.max_iterations = 1
#map_adapt.train(gmm_adapted, data)
bob.learn.em.train(map_adapt, gmm_adapted, data, max_iterations = 1)
new_means = bob.io.base.load(datafile('new_adapted_mean.hdf5', __name__, path="../data/"))
# print new_means[0,:]
# print gmm_adapted.means[:,0]
# Compare to matlab reference
assert equals(new_means[0,:], gmm_adapted.means[:,0], 1e-4)
assert equals(new_means[1,:], gmm_adapted.means[:,1], 1e-4)
def load_model(self, ubm_file):
"""Loads the projector (UBM) from a file."""
hdf5file = HDF5File(ubm_file, "r")
logger.debug("Loading model from file '%s'", ubm_file)
# Read the UBM
self.ubm = GMMMachine.from_hdf5(hdf5file)
self.ubm.variance_thresholds = self.variance_threshold
def test_score():
gmm1 = GMM(number_of_gaussians=2)
gmm1.load_model(
pkg_resources.resource_filename("bob.bio.gmm.test",
"data/gmm_ubm.hdf5"))
biometric_reference = GMMMachine.from_hdf5(
pkg_resources.resource_filename("bob.bio.gmm.test",
"data/gmm_enrolled.hdf5"),
ubm=gmm1.ubm,
)
probe = GMMStats.from_hdf5(
pkg_resources.resource_filename("bob.bio.gmm.test",
"data/gmm_projected.hdf5"))
probe_data = utils.random_array((20, 45), -5.0, 5.0, seed=seed_value)
reference_score = 0.6509
numpy.testing.assert_almost_equal(gmm1.score(biometric_reference, probe),
reference_score,
decimal=5)
multi_refs = gmm1.score_multiple_biometric_references(
[biometric_reference, biometric_reference, biometric_reference], probe)
assert multi_refs.shape == (3, 1), multi_refs.shape
numpy.testing.assert_almost_equal(multi_refs, reference_score, decimal=5)
# With not projected data
numpy.testing.assert_almost_equal(gmm1.score(biometric_reference,
probe_data),
reference_score,
decimal=5)
def test_GMMMachine_3():
# Test a GMMMachine (log-likelihood computation)
data = bob.io.base.load(datafile('data.hdf5', __name__, path="../data/"))
gmm = GMMMachine(2, 50)
gmm.weights = bob.io.base.load(
datafile('weights.hdf5', __name__, path="../data/"))
gmm.means = bob.io.base.load(
datafile('means.hdf5', __name__, path="../data/"))
gmm.variances = bob.io.base.load(
datafile('variances.hdf5', __name__, path="../data/"))
# Compare the log-likelihood with the one obtained using Chris Matlab
# implementation
matlab_ll_ref = -2.361583051672024e+02
assert abs(gmm(data) - matlab_ll_ref) < 1e-10
def test_training():
"""Tests the generation of the UBM."""
# Set a small training iteration count
gmm1 = GMM(
number_of_gaussians=2,
kmeans_training_iterations=5,
ubm_training_iterations=5,
init_seed=seed_value,
kmeans_oversampling_factor=2,
)
train_data = utils.random_training_set((100, 45),
count=5,
minimum=-5.0,
maximum=5.0)
train_data = numpy.vstack(train_data)
# Train the UBM (projector)
gmm1.fit(train_data)
# Test saving and loading of projector
with tempfile.NamedTemporaryFile(prefix="bob_",
suffix="_model.hdf5") as fd:
temp_file = fd.name
gmm1.save_model(temp_file)
reference_file = pkg_resources.resource_filename(
"bob.bio.gmm.test", "data/gmm_ubm.hdf5")
if regenerate_refs:
gmm1.save_model(reference_file)
gmm2 = GMM(number_of_gaussians=2)
gmm2.load_model(temp_file)
ubm_reference = GMMMachine.from_hdf5(reference_file)
assert gmm2.ubm.is_similar_to(ubm_reference)
def test_gmm_ML_1():
"""Trains a GMMMachine with ML_GMMTrainer"""
ar = load_array(
resource_filename("bob.learn.em", "data/faithful.torch3_f64.hdf5"))
gmm_ref = GMMMachine.from_hdf5(
HDF5File(resource_filename("bob.learn.em", "data/gmm_ML.hdf5"), "r"))
for transform in (to_numpy, to_dask_array):
ar = transform(ar)
gmm = loadGMM()
# test rng handling
gmm.convergence_threshold = 0.001
gmm.update_means = True
gmm.update_variances = True
gmm.update_weights = True
gmm.random_state = np.random.RandomState(seed=12345)
gmm = gmm.fit(ar)
gmm = loadGMM()
gmm.convergence_threshold = 0.001
gmm.update_means = True
gmm.update_variances = True
gmm.update_weights = True
# Generate reference
# gmm.save(HDF5File(resource_filename("bob.learn.em", "data/gmm_ML.hdf5"), "w"))
gmm = gmm.fit(ar)
assert_gmm_equal(gmm, gmm_ref)
def test_likelihood():
data = np.array([[1, 1, 1], [-1, 0, 0], [0, 0, 1], [2, 2, 2]])
n_gaussians = 3
machine = GMMMachine(n_gaussians)
machine.means = np.repeat([[0], [1], [-1]], 3, 1)
machine.variances = np.ones_like(machine.means)
for transform in (to_numpy, to_dask_array):
data = transform(data)
log_likelihood = machine.log_likelihood(data)
expected_ll = np.array([
-3.6519900964986527,
-3.83151883210222,
-3.83151883210222,
-5.344374066745753,
])
np.testing.assert_almost_equal(log_likelihood, expected_ll)
def test_gmm_MAP_1():
# Train a GMMMachine with MAP_GMMTrainer
ar = bob.io.base.load(
datafile('faithful.torch3_f64.hdf5', __name__, path="../data/"))
# test with rng
rng = bob.core.random.mt19937(12345)
gmm = GMMMachine(
bob.io.base.HDF5File(datafile("gmm_ML.hdf5", __name__,
path="../data/")))
gmmprior = GMMMachine(
bob.io.base.HDF5File(datafile("gmm_ML.hdf5", __name__,
path="../data/")))
map_gmmtrainer = MAP_GMMTrainer(update_means=True,
update_variances=False,
update_weights=False,
prior_gmm=gmmprior,
relevance_factor=4.)
bob.learn.em.train(map_gmmtrainer, gmm, ar, rng=rng)
gmm = GMMMachine(
bob.io.base.HDF5File(datafile("gmm_ML.hdf5", __name__,
path="../data/")))
gmmprior = GMMMachine(
bob.io.base.HDF5File(datafile("gmm_ML.hdf5", __name__,
path="../data/")))
map_gmmtrainer = MAP_GMMTrainer(update_means=True,
update_variances=False,
update_weights=False,
prior_gmm=gmmprior,
relevance_factor=4.)
#map_gmmtrainer.train(gmm, ar)
bob.learn.em.train(map_gmmtrainer, gmm, ar)
gmm_ref = GMMMachine(
bob.io.base.HDF5File(
datafile('gmm_MAP.hdf5', __name__, path="../data/")))
assert (equals(gmm.means, gmm_ref.means, 1e-3)
and equals(gmm.variances, gmm_ref.variances, 1e-3)
and equals(gmm.weights, gmm_ref.weights, 1e-3))
def test_gmm_test():
# Tests a GMMMachine by computing scores against a model and compare to
# an old reference
ar = bob.io.base.load(
datafile('dataforMAP.hdf5', __name__, path="../data/"))
# Initialize GMMMachine
n_gaussians = 5
n_inputs = 45
gmm = GMMMachine(n_gaussians, n_inputs)
gmm.means = bob.io.base.load(
datafile('meansAfterML.hdf5', __name__, path="../data/"))
gmm.variances = bob.io.base.load(
datafile('variancesAfterML.hdf5', __name__, path="../data/"))
gmm.weights = bob.io.base.load(
datafile('weightsAfterML.hdf5', __name__, path="../data/"))
threshold = 0.001
gmm.set_variance_thresholds(threshold)
# Test against the model
score_mean_ref = -1.50379e+06
score = 0.
for v in ar:
score += gmm(v)
score /= len(ar)
# Compare current results to torch3vision
assert abs(score - score_mean_ref) / score_mean_ref < 1e-4
def test_ISVTrainAndEnrol():
# Train and enroll an 'ISVMachine'
eps = 1e-10
d_ref = numpy.array([0.39601136, 0.07348469, 0.47712682, 0.44738127, 0.43179856, 0.45086029], 'float64')
u_ref = numpy.array([[0.855125642430777, 0.563104284748032], [-0.325497865404680, 1.923598985291687], [0.511575659503837, 1.964288663083095], [9.330165761678115, 1.073623827995043], [0.511099245664012, 0.278551249248978], [5.065578541930268, 0.509565618051587]], 'float64')
z_ref = numpy.array([-0.079315777443826, 0.092702428248543, -0.342488761656616, -0.059922635809136 , 0.133539981073604, 0.213118695516570], 'float64')
# Calls the train function
ubm = GMMMachine(2,3)
ubm.mean_supervector = UBM_MEAN
ubm.variance_supervector = UBM_VAR
mb = ISVBase(ubm,2)
t = ISVTrainer(4.)
t.initialize(mb, TRAINING_STATS)
mb.u = M_u
for i in range(10):
t.e_step(mb, TRAINING_STATS)
t.m_step(mb)
assert numpy.allclose(mb.d, d_ref, eps)
assert numpy.allclose(mb.u, u_ref, eps)
# Calls the enroll function
m = ISVMachine(mb)
Ne = numpy.array([0.1579, 0.9245, 0.1323, 0.2458]).reshape((2,2))
Fe = numpy.array([0.1579, 0.1925, 0.3242, 0.1234, 0.2354, 0.2734, 0.2514, 0.5874, 0.3345, 0.2463, 0.4789, 0.5236]).reshape((6,2))
gse1 = GMMStats(2,3)
gse1.n = Ne[:,0]
gse1.sum_px = Fe[:,0].reshape(2,3)
gse2 = GMMStats(2,3)
gse2.n = Ne[:,1]
gse2.sum_px = Fe[:,1].reshape(2,3)
gse = [gse1, gse2]
t.enroll(m, gse, 5)
assert numpy.allclose(m.z, z_ref, eps)
#Testing exceptions
nose.tools.assert_raises(RuntimeError, t.initialize, mb, [1,2,2])
nose.tools.assert_raises(RuntimeError, t.initialize, mb, [[1,2,2]])
nose.tools.assert_raises(RuntimeError, t.e_step, mb, [1,2,2])
nose.tools.assert_raises(RuntimeError, t.e_step, mb, [[1,2,2]])
nose.tools.assert_raises(RuntimeError, t.enroll, m, [[1,2,2]],5)
def test_gmm_test():
# Tests a GMMMachine by computing scores against a model and comparing to a reference
ar = load_array(resource_filename("bob.learn.em", "data/dataforMAP.hdf5"))
# Initialize GMMMachine
n_gaussians = 5
gmm = GMMMachine(n_gaussians)
gmm.means = load_array(
resource_filename("bob.learn.em", "data/meansAfterML.hdf5"))
gmm.variances = load_array(
resource_filename("bob.learn.em", "data/variancesAfterML.hdf5"))
gmm.weights = load_array(
resource_filename("bob.learn.em", "data/weightsAfterML.hdf5"))
threshold = 0.001
gmm.variance_thresholds = threshold
# Test against the model
score_mean_ref = -1.50379e06
for transform in (to_numpy, to_dask_array):
ar = transform(ar)
score = gmm.log_likelihood(ar).sum()
score /= len(ar)
# Compare current results to torch3vision
assert abs(score - score_mean_ref) / score_mean_ref < 1e-4
def convert_gmm(gmm):
"""
Converts rasr_cache.MixtureFile to bob.em.learn.GMMachine
:param gmm: (MixtureFile)
:return: (GMMachine)
"""
ubm = GMMMachine(gmm.nMeans, gmm.dim)
tmp_m = np.ndarray((gmm.nMeans, gmm.dim))
tmp_c = np.ndarray((gmm.nCovs, gmm.dim))
for i in range(gmm.nMeans):
tmp_m[i, :] = np.array(gmm.getMeanByIdx(i))
tmp_c[i, :] = np.array(
gmm.getCovByIdx(0)
) # TODO figure out to generate same number of covariances as means
ubm.means = tmp_m
ubm.variances = tmp_c
return ubm
def test_gmm_MAP_1():
# Train a GMMMachine with MAP_GMMTrainer
ar = load_array(
resource_filename("bob.learn.em", "data/faithful.torch3_f64.hdf5"))
# test with rng
gmmprior = GMMMachine.from_hdf5(
HDF5File(resource_filename("bob.learn.em", "data/gmm_ML.hdf5"), "r"))
gmm = GMMMachine.from_hdf5(
HDF5File(resource_filename("bob.learn.em", "data/gmm_ML.hdf5"), "r"),
ubm=gmmprior,
)
gmm.update_means = True
gmm.update_variances = False
gmm.update_weights = False
rng = np.random.RandomState(seed=12345)
gmm.random_state = rng
gmm = gmm.fit(ar)
gmmprior = GMMMachine.from_hdf5(
HDF5File(resource_filename("bob.learn.em", "data/gmm_ML.hdf5"), "r"))
gmm = GMMMachine.from_hdf5(
HDF5File(resource_filename("bob.learn.em", "data/gmm_ML.hdf5"), "r"),
ubm=gmmprior,
)
gmm.update_means = True
gmm.update_variances = False
gmm.update_weights = False
# Generate reference
# gmm.save(HDF5File(resource_filename("bob.learn.em", "data/gmm_MAP.hdf5"), "w"))
gmm_ref = GMMMachine.from_hdf5(
HDF5File(resource_filename("bob.learn.em", "data/gmm_MAP.hdf5"), "r"))
for transform in (to_numpy, to_dask_array):
ar = transform(ar)
gmm = gmm.fit(ar)
np.testing.assert_almost_equal(gmm.means, gmm_ref.means, decimal=3)
np.testing.assert_almost_equal(gmm.variances,
gmm_ref.variances,
decimal=3)
np.testing.assert_almost_equal(gmm.weights, gmm_ref.weights, decimal=3)
def test_GMMMachine_4():
import numpy
numpy.random.seed(3) # FIXING A SEED
data = numpy.random.rand(100,50) #Doesn't matter if it is ramdom. The average of 1D array (in python) MUST output the same result for the 2D array (in C++)
gmm = GMMMachine(2, 50)
gmm.weights = bob.io.base.load(datafile('weights.hdf5', __name__, path="../data/"))
gmm.means = bob.io.base.load(datafile('means.hdf5', __name__, path="../data/"))
gmm.variances = bob.io.base.load(datafile('variances.hdf5', __name__, path="../data/"))
ll = 0
for i in range(data.shape[0]):
ll += gmm(data[i,:])
ll /= data.shape[0]
assert ll==gmm(data)
def test_gmm_kmeans_plusplus_init():
n_gaussians = 3
machine = GMMMachine(
n_gaussians,
k_means_trainer=KMeansMachine(n_clusters=n_gaussians,
init_method="k-means++"),
)
data = np.array([[1.5, 1], [1, 1.5], [-1, 0.5], [-1.5, 0], [2, 2],
[2.5, 2.5]])
for transform in (to_numpy, to_dask_array):
data = transform(data)
machine = machine.fit(data)
expected_means = np.array([[2.25, 2.25], [-1.25, 0.25], [1.25, 1.25]])
expected_variances = np.array([[1 / 16, 1 / 16], [1 / 16, 1 / 16],
[1 / 16, 1 / 16]])
np.testing.assert_almost_equal(machine.means,
expected_means,
decimal=3)
np.testing.assert_almost_equal(machine.variances, expected_variances)
def test_gmm_MAP_2():
# Train a GMMMachine with MAP_GMMTrainer and compare with matlab reference
data = bob.io.base.load(datafile('data.hdf5', __name__, path="../data/"))
data = data.reshape((1, data.shape[0])) # make a 2D array out of it
means = bob.io.base.load(datafile('means.hdf5', __name__, path="../data/"))
variances = bob.io.base.load(
datafile('variances.hdf5', __name__, path="../data/"))
weights = bob.io.base.load(
datafile('weights.hdf5', __name__, path="../data/"))
gmm = GMMMachine(2, 50)
gmm.means = means
gmm.variances = variances
gmm.weights = weights
map_adapt = MAP_GMMTrainer(update_means=True,
update_variances=False,
update_weights=False,
mean_var_update_responsibilities_threshold=0.,
prior_gmm=gmm,
relevance_factor=4.)
gmm_adapted = GMMMachine(2, 50)
gmm_adapted.means = means
gmm_adapted.variances = variances
gmm_adapted.weights = weights
#map_adapt.max_iterations = 1
#map_adapt.train(gmm_adapted, data)
bob.learn.em.train(map_adapt, gmm_adapted, data, max_iterations=1)
new_means = bob.io.base.load(
datafile('new_adapted_mean.hdf5', __name__, path="../data/"))
# print new_means[0,:]
# print gmm_adapted.means[:,0]
# Compare to matlab reference
assert equals(new_means[0, :], gmm_adapted.means[:, 0], 1e-4)
assert equals(new_means[1, :], gmm_adapted.means[:, 1], 1e-4)
def test_gmm_MAP_3():
# Train a GMMMachine with MAP_GMMTrainer; compares to old reference
ar = bob.io.base.load(datafile('dataforMAP.hdf5', __name__, path="../data/"))
# Initialize GMMMachine
n_gaussians = 5
n_inputs = 45
prior_gmm = GMMMachine(n_gaussians, n_inputs)
prior_gmm.means = bob.io.base.load(datafile('meansAfterML.hdf5', __name__, path="../data/"))
prior_gmm.variances = bob.io.base.load(datafile('variancesAfterML.hdf5', __name__, path="../data/"))
prior_gmm.weights = bob.io.base.load(datafile('weightsAfterML.hdf5', __name__, path="../data/"))
threshold = 0.001
prior_gmm.set_variance_thresholds(threshold)
# Initialize MAP Trainer
relevance_factor = 0.1
prior = 0.001
max_iter_gmm = 1
accuracy = 0.00001
map_factor = 0.5
map_gmmtrainer = MAP_GMMTrainer(prior_gmm, alpha=map_factor, update_means=True, update_variances=False, update_weights=False, mean_var_update_responsibilities_threshold=accuracy)
#map_gmmtrainer.max_iterations = max_iter_gmm
#map_gmmtrainer.convergence_threshold = accuracy
gmm = GMMMachine(n_gaussians, n_inputs)
gmm.set_variance_thresholds(threshold)
# Train
#map_gmmtrainer.train(gmm, ar)
bob.learn.em.train(map_gmmtrainer, gmm, ar, max_iterations = max_iter_gmm, convergence_threshold=prior)
# Test results
# Load torch3vision reference
meansMAP_ref = bob.io.base.load(datafile('meansAfterMAP.hdf5', __name__, path="../data/"))
variancesMAP_ref = bob.io.base.load(datafile('variancesAfterMAP.hdf5', __name__, path="../data/"))
weightsMAP_ref = bob.io.base.load(datafile('weightsAfterMAP.hdf5', __name__, path="../data/"))
# Compare to current results
# Gaps are quite large. This might be explained by the fact that there is no
# adaptation of a given Gaussian in torch3 when the corresponding responsibilities
# are below the responsibilities threshold
assert equals(gmm.means, meansMAP_ref, 2e-1)
assert equals(gmm.variances, variancesMAP_ref, 1e-4)
assert equals(gmm.weights, weightsMAP_ref, 1e-4)
def test_gmm_test():
# Tests a GMMMachine by computing scores against a model and compare to
# an old reference
ar = bob.io.base.load(datafile('dataforMAP.hdf5', __name__, path="../data/"))
# Initialize GMMMachine
n_gaussians = 5
n_inputs = 45
gmm = GMMMachine(n_gaussians, n_inputs)
gmm.means = bob.io.base.load(datafile('meansAfterML.hdf5', __name__, path="../data/"))
gmm.variances = bob.io.base.load(datafile('variancesAfterML.hdf5', __name__, path="../data/"))
gmm.weights = bob.io.base.load(datafile('weightsAfterML.hdf5', __name__, path="../data/"))
threshold = 0.001
gmm.set_variance_thresholds(threshold)
# Test against the model
score_mean_ref = -1.50379e+06
score = 0.
for v in ar: score += gmm(v)
score /= len(ar)
# Compare current results to torch3vision
assert abs(score-score_mean_ref)/score_mean_ref<1e-4
def test_GMMMachine_2():
# Test a GMMMachine (statistics)
arrayset = bob.io.base.load(datafile("faithful.torch3_f64.hdf5", __name__, path="../data/"))
gmm = GMMMachine(2, 2)
gmm.weights = numpy.array([0.5, 0.5], 'float64')
gmm.means = numpy.array([[3, 70], [4, 72]], 'float64')
gmm.variances = numpy.array([[1, 10], [2, 5]], 'float64')
gmm.variance_thresholds = numpy.array([[0, 0], [0, 0]], 'float64')
stats = GMMStats(2, 2)
gmm.acc_statistics(arrayset, stats)
stats_ref = GMMStats(bob.io.base.HDF5File(datafile("stats.hdf5",__name__, path="../data/")))
assert stats.t == stats_ref.t
assert numpy.allclose(stats.n, stats_ref.n, atol=1e-10)
#assert numpy.array_equal(stats.sumPx, stats_ref.sumPx)
#Note AA: precision error above
assert numpy.allclose(stats.sum_px, stats_ref.sum_px, atol=1e-10)
assert numpy.allclose(stats.sum_pxx, stats_ref.sum_pxx, atol=1e-10)
def test_gmm_ML_2():
# Trains a GMMMachine with ML_GMMTrainer; compares to an old reference
ar = bob.io.base.load(datafile('dataNormalized.hdf5', __name__, path="../data/"))
# Initialize GMMMachine
gmm = GMMMachine(5, 45)
gmm.means = bob.io.base.load(datafile('meansAfterKMeans.hdf5', __name__, path="../data/")).astype('float64')
gmm.variances = bob.io.base.load(datafile('variancesAfterKMeans.hdf5', __name__, path="../data/")).astype('float64')
gmm.weights = numpy.exp(bob.io.base.load(datafile('weightsAfterKMeans.hdf5', __name__, path="../data/")).astype('float64'))
threshold = 0.001
gmm.set_variance_thresholds(threshold)
# Initialize ML Trainer
prior = 0.001
max_iter_gmm = 25
accuracy = 0.00001
ml_gmmtrainer = ML_GMMTrainer(True, True, True, prior)
# Run ML
#ml_gmmtrainer.train(gmm, ar)
bob.learn.em.train(ml_gmmtrainer, gmm, ar, max_iterations = max_iter_gmm, convergence_threshold=accuracy)
# Test results
# Load torch3vision reference
meansML_ref = bob.io.base.load(datafile('meansAfterML.hdf5', __name__, path="../data/"))
variancesML_ref = bob.io.base.load(datafile('variancesAfterML.hdf5', __name__, path="../data/"))
weightsML_ref = bob.io.base.load(datafile('weightsAfterML.hdf5', __name__, path="../data/"))
# Compare to current results
assert equals(gmm.means, meansML_ref, 3e-3)
assert equals(gmm.variances, variancesML_ref, 3e-3)
assert equals(gmm.weights, weightsML_ref, 1e-4)
def test_JFAMachine():
# Creates a UBM
weights = numpy.array([0.4, 0.6], 'float64')
means = numpy.array([[1, 6, 2], [4, 3, 2]], 'float64')
variances = numpy.array([[1, 2, 1], [2, 1, 2]], 'float64')
ubm = GMMMachine(2,3)
ubm.weights = weights
ubm.means = means
ubm.variances = variances
# Creates a JFABase
U = numpy.array([[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12]], 'float64')
V = numpy.array([[6, 5], [4, 3], [2, 1], [1, 2], [3, 4], [5, 6]], 'float64')
d = numpy.array([0, 1, 0, 1, 0, 1], 'float64')
base = JFABase(ubm,2,2)
base.u = U
base.v = V
base.d = d
# Creates a JFAMachine
y = numpy.array([1,2], 'float64')
z = numpy.array([3,4,1,2,0,1], 'float64')
m = JFAMachine(base)
m.y = y
m.z = z
n_gaussians,dim,ru,rv = m.shape
supervector_length = m.supervector_length
assert n_gaussians == 2
assert dim == 3
assert supervector_length == 6
assert ru == 2
assert rv == 2
assert (m.y == y).all()
assert (m.z == z).all()
# Saves and loads
filename = str(tempfile.mkstemp(".hdf5")[1])
m.save(bob.io.base.HDF5File(filename, 'w'))
m_loaded = JFAMachine(bob.io.base.HDF5File(filename))
m_loaded.jfa_base = base
assert m == m_loaded
assert (m != m_loaded) is False
assert m.is_similar_to(m_loaded)
# Copy constructor
mc = JFAMachine(m)
assert m == mc
# Variant
#mv = JFAMachine()
# Checks for correctness
#mv.jfa_base = base
#m.y = y
#m.z = z
#assert m.dim_c == 2
#assert m.dim_d == 3
#assert m.dim_cd == 6
#assert m.dim_ru == 2
#assert m.dim_rv == 2
#assert (m.y == y).all()
#assert (m.z == z).all()
# Defines GMMStats
gs = GMMStats(2,3)
log_likelihood = -3.
T = 1
n = numpy.array([0.4, 0.6], 'float64')
sumpx = numpy.array([[1., 2., 3.], [4., 5., 6.]], 'float64')
sumpxx = numpy.array([[10., 20., 30.], [40., 50., 60.]], 'float64')
gs.log_likelihood = log_likelihood
gs.t = T
gs.n = n
gs.sum_px = sumpx
gs.sum_pxx = sumpxx
# Forward GMMStats and check estimated value of the x speaker factor
eps = 1e-10
x_ref = numpy.array([0.291042849767692, 0.310273618998444], 'float64')
score_ref = -2.111577181208289
score = m.log_likelihood(gs)
assert numpy.allclose(m.x, x_ref, eps)
assert abs(score_ref-score) < eps
# x and Ux
x = numpy.ndarray((2,), numpy.float64)
m.estimate_x(gs, x)
n_gaussians, dim,_,_ = m.shape
x_py = estimate_x(n_gaussians, dim, ubm.mean_supervector, ubm.variance_supervector, U, n, sumpx)
assert numpy.allclose(x, x_py, eps)
ux = numpy.ndarray((6,), numpy.float64)
m.estimate_ux(gs, ux)
n_gaussians, dim,_,_ = m.shape
ux_py = estimate_ux(n_gaussians, dim, ubm.mean_supervector, ubm.variance_supervector, U, n, sumpx)
assert numpy.allclose(ux, ux_py, eps)
assert numpy.allclose(m.x, x, eps)
score = m.forward_ux(gs, ux)
assert abs(score_ref-score) < eps
# Clean-up
os.unlink(filename)
def test_JFATrainAndEnrol():
# Train and enroll a JFAMachine
# Calls the train function
ubm = GMMMachine(2,3)
ubm.mean_supervector = UBM_MEAN
ubm.variance_supervector = UBM_VAR
mb = JFABase(ubm, 2, 2)
t = JFATrainer()
t.initialize(mb, TRAINING_STATS)
mb.u = M_u
mb.v = M_v
mb.d = M_d
bob.learn.em.train_jfa(t,mb, TRAINING_STATS, initialize=False)
v_ref = numpy.array([[0.245364911936476, 0.978133261775424], [0.769646805052223, 0.940070736856596], [0.310779202800089, 1.456332053893072],
[0.184760934399551, 2.265139705602147], [0.701987784039800, 0.081632150899400], [0.074344030229297, 1.090248340917255]], 'float64')
u_ref = numpy.array([[0.049424652628448, 0.060480486336896], [0.178104127464007, 1.884873813495153], [1.204011484266777, 2.281351307871720],
[7.278512126426286, -0.390966087173334], [-0.084424326581145, -0.081725474934414], [4.042143689831097, -0.262576386580701]], 'float64')
d_ref = numpy.array([9.648467e-18, 2.63720683155e-12, 2.11822157653706e-10, 9.1047243e-17, 1.41163442535567e-10, 3.30581e-19], 'float64')
eps = 1e-10
assert numpy.allclose(mb.v, v_ref, eps)
assert numpy.allclose(mb.u, u_ref, eps)
assert numpy.allclose(mb.d, d_ref, eps)
# Calls the enroll function
m = JFAMachine(mb)
Ne = numpy.array([0.1579, 0.9245, 0.1323, 0.2458]).reshape((2,2))
Fe = numpy.array([0.1579, 0.1925, 0.3242, 0.1234, 0.2354, 0.2734, 0.2514, 0.5874, 0.3345, 0.2463, 0.4789, 0.5236]).reshape((6,2))
gse1 = GMMStats(2,3)
gse1.n = Ne[:,0]
gse1.sum_px = Fe[:,0].reshape(2,3)
gse2 = GMMStats(2,3)
gse2.n = Ne[:,1]
gse2.sum_px = Fe[:,1].reshape(2,3)
gse = [gse1, gse2]
t.enroll(m, gse, 5)
y_ref = numpy.array([0.555991469319657, 0.002773650670010], 'float64')
z_ref = numpy.array([8.2228e-20, 3.15216909492e-13, -1.48616735364395e-10, 1.0625905e-17, 3.7150503117895e-11, 1.71104e-19], 'float64')
assert numpy.allclose(m.y, y_ref, eps)
assert numpy.allclose(m.z, z_ref, eps)
#Testing exceptions
nose.tools.assert_raises(RuntimeError, t.initialize, mb, [1,2,2])
nose.tools.assert_raises(RuntimeError, t.initialize, mb, [[1,2,2]])
nose.tools.assert_raises(RuntimeError, t.e_step_u, mb, [1,2,2])
nose.tools.assert_raises(RuntimeError, t.e_step_u, mb, [[1,2,2]])
nose.tools.assert_raises(RuntimeError, t.m_step_u, mb, [1,2,2])
nose.tools.assert_raises(RuntimeError, t.m_step_u, mb, [[1,2,2]])
nose.tools.assert_raises(RuntimeError, t.e_step_v, mb, [1,2,2])
nose.tools.assert_raises(RuntimeError, t.e_step_v, mb, [[1,2,2]])
nose.tools.assert_raises(RuntimeError, t.m_step_v, mb, [1,2,2])
nose.tools.assert_raises(RuntimeError, t.m_step_v, mb, [[1,2,2]])
nose.tools.assert_raises(RuntimeError, t.e_step_d, mb, [1,2,2])
nose.tools.assert_raises(RuntimeError, t.e_step_d, mb, [[1,2,2]])
nose.tools.assert_raises(RuntimeError, t.m_step_d, mb, [1,2,2])
nose.tools.assert_raises(RuntimeError, t.m_step_d, mb, [[1,2,2]])
nose.tools.assert_raises(RuntimeError, t.enroll, m, [[1,2,2]],5)
def test_LinearScoring(): ubm = GMMMachine(2, 2) ubm.weights = numpy.array([0.5, 0.5], 'float64') ubm.means = numpy.array([[3, 70], [4, 72]], 'float64') ubm.variances = numpy.array([[1, 10], [2, 5]], 'float64') ubm.variance_thresholds = numpy.array([[0, 0], [0, 0]], 'float64') model1 = GMMMachine(2, 2) model1.weights = numpy.array([0.5, 0.5], 'float64') model1.means = numpy.array([[1, 2], [3, 4]], 'float64') model1.variances = numpy.array([[9, 10], [11, 12]], 'float64') model1.variance_thresholds = numpy.array([[0, 0], [0, 0]], 'float64') model2 = GMMMachine(2, 2) model2.weights = numpy.array([0.5, 0.5], 'float64') model2.means = numpy.array([[5, 6], [7, 8]], 'float64') model2.variances = numpy.array([[13, 14], [15, 16]], 'float64') model2.variance_thresholds = numpy.array([[0, 0], [0, 0]], 'float64') stats1 = GMMStats(2, 2) stats1.sum_px = numpy.array([[1, 2], [3, 4]], 'float64') stats1.n = numpy.array([1, 2], 'float64') stats1.t = 1+2 stats2 = GMMStats(2, 2) stats2.sum_px = numpy.array([[5, 6], [7, 8]], 'float64') stats2.n = numpy.array([3, 4], 'float64') stats2.t = 3+4 stats3 = GMMStats(2, 2) stats3.sum_px = numpy.array([[5, 6], [7, 3]], 'float64') stats3.n = numpy.array([3, 4], 'float64') stats3.t = 3+4 test_channeloffset = [numpy.array([9, 8, 7, 6], 'float64'), numpy.array([5, 4, 3, 2], 'float64'), numpy.array([1, 0, 1, 2], 'float64')] # Reference scores (from Idiap internal matlab implementation) ref_scores_00 = numpy.array([[2372.9, 5207.7, 5275.7], [2215.7, 4868.1, 4932.1]], 'float64') ref_scores_01 = numpy.array( [[790.9666666666667, 743.9571428571428, 753.6714285714285], [738.5666666666667, 695.4428571428572, 704.5857142857144]], 'float64') ref_scores_10 = numpy.array([[2615.5, 5434.1, 5392.5], [2381.5, 4999.3, 5022.5]], 'float64') ref_scores_11 = numpy.array([[871.8333333333332, 776.3000000000001, 770.3571428571427], [793.8333333333333, 714.1857142857143, 717.5000000000000]], 'float64') # 1/ Use GMMMachines # 1/a/ Without test_channelOffset, without frame-length normalisation scores = linear_scoring([model1, model2], ubm, [stats1, stats2, stats3]) assert (abs(scores - ref_scores_00) < 1e-7).all() # 1/b/ Without test_channelOffset, with frame-length normalisation scores = linear_scoring([model1, model2], ubm, [stats1, stats2, stats3], [], True) assert (abs(scores - ref_scores_01) < 1e-7).all() #scores = linear_scoring([model1, model2], ubm, [stats1, stats2, stats3], (), True) #assert (abs(scores - ref_scores_01) < 1e-7).all() #scores = linear_scoring([model1, model2], ubm, [stats1, stats2, stats3], None, True) #assert (abs(scores - ref_scores_01) < 1e-7).all() # 1/c/ With test_channelOffset, without frame-length normalisation scores = linear_scoring([model1, model2], ubm, [stats1, stats2, stats3], test_channeloffset) assert (abs(scores - ref_scores_10) < 1e-7).all() # 1/d/ With test_channelOffset, with frame-length normalisation scores = linear_scoring([model1, model2], ubm, [stats1, stats2, stats3], test_channeloffset, True) assert (abs(scores - ref_scores_11) < 1e-7).all() # 2/ Use mean/variance supervectors # 2/a/ Without test_channelOffset, without frame-length normalisation scores = linear_scoring([model1.mean_supervector, model2.mean_supervector], ubm.mean_supervector, ubm.variance_supervector, [stats1, stats2, stats3]) assert (abs(scores - ref_scores_00) < 1e-7).all() # 2/b/ Without test_channelOffset, with frame-length normalisation scores = linear_scoring([model1.mean_supervector, model2.mean_supervector], ubm.mean_supervector, ubm.variance_supervector, [stats1, stats2, stats3], [], True) assert (abs(scores - ref_scores_01) < 1e-7).all() # 2/c/ With test_channelOffset, without frame-length normalisation scores = linear_scoring([model1.mean_supervector, model2.mean_supervector], ubm.mean_supervector, ubm.variance_supervector, [stats1, stats2, stats3], test_channeloffset) assert (abs(scores - ref_scores_10) < 1e-7).all() # 2/d/ With test_channelOffset, with frame-length normalisation scores = linear_scoring([model1.mean_supervector, model2.mean_supervector], ubm.mean_supervector, ubm.variance_supervector, [stats1, stats2, stats3], test_channeloffset, True) assert (abs(scores - ref_scores_11) < 1e-7).all() # 3/ Using single model/sample # 3/a/ without frame-length normalisation score = linear_scoring(model1.mean_supervector, ubm.mean_supervector, ubm.variance_supervector, stats1, test_channeloffset[0]) assert abs(score - ref_scores_10[0,0]) < 1e-7 score = linear_scoring(model1.mean_supervector, ubm.mean_supervector, ubm.variance_supervector, stats2, test_channeloffset[1]) assert abs(score - ref_scores_10[0,1]) < 1e-7 score = linear_scoring(model1.mean_supervector, ubm.mean_supervector, ubm.variance_supervector, stats3, test_channeloffset[2]) assert abs(score - ref_scores_10[0,2]) < 1e-7 score = linear_scoring(model2.mean_supervector, ubm.mean_supervector, ubm.variance_supervector, stats1, test_channeloffset[0]) assert abs(score - ref_scores_10[1,0]) < 1e-7 score = linear_scoring(model2.mean_supervector, ubm.mean_supervector, ubm.variance_supervector, stats2, test_channeloffset[1]) assert abs(score - ref_scores_10[1,1]) < 1e-7 score = linear_scoring(model2.mean_supervector, ubm.mean_supervector, ubm.variance_supervector, stats3, test_channeloffset[2]) assert abs(score - ref_scores_10[1,2]) < 1e-7 # 3/b/ without frame-length normalisation score = linear_scoring(model1.mean_supervector, ubm.mean_supervector, ubm.variance_supervector, stats1, test_channeloffset[0], True) assert abs(score - ref_scores_11[0,0]) < 1e-7 score = linear_scoring(model1.mean_supervector, ubm.mean_supervector, ubm.variance_supervector, stats2, test_channeloffset[1], True) assert abs(score - ref_scores_11[0,1]) < 1e-7 score = linear_scoring(model1.mean_supervector, ubm.mean_supervector, ubm.variance_supervector, stats3, test_channeloffset[2], True) assert abs(score - ref_scores_11[0,2]) < 1e-7 score = linear_scoring(model2.mean_supervector, ubm.mean_supervector, ubm.variance_supervector, stats1, test_channeloffset[0], True) assert abs(score - ref_scores_11[1,0]) < 1e-7 score = linear_scoring(model2.mean_supervector, ubm.mean_supervector, ubm.variance_supervector, stats2, test_channeloffset[1], True) assert abs(score - ref_scores_11[1,1]) < 1e-7 score = linear_scoring(model2.mean_supervector, ubm.mean_supervector, ubm.variance_supervector, stats3, test_channeloffset[2], True) assert abs(score - ref_scores_11[1,2]) < 1e-7
def test_GMMMachine_1(): # Test a GMMMachine basic features weights = numpy.array([0.5, 0.5], 'float64') weights2 = numpy.array([0.6, 0.4], 'float64') means = numpy.array([[3, 70, 0], [4, 72, 0]], 'float64') means2 = numpy.array([[3, 7, 0], [4, 72, 0]], 'float64') variances = numpy.array([[1, 10, 1], [2, 5, 2]], 'float64') variances2 = numpy.array([[10, 10, 1], [2, 5, 2]], 'float64') varianceThresholds = numpy.array([[0, 0, 0], [0, 0, 0]], 'float64') varianceThresholds2 = numpy.array([[0.0005, 0.0005, 0.0005], [0, 0, 0]], 'float64') # Initializes a GMMMachine gmm = GMMMachine(2,3) # Sets the weights, means, variances and varianceThresholds and # Checks correctness gmm.weights = weights gmm.means = means gmm.variances = variances gmm.variance_thresholds = varianceThresholds assert gmm.shape == (2,3) assert (gmm.weights == weights).all() assert (gmm.means == means).all() assert (gmm.variances == variances).all() assert (gmm.variance_thresholds == varianceThresholds).all() # Checks supervector-like accesses assert (gmm.mean_supervector == means.reshape(means.size)).all() assert (gmm.variance_supervector == variances.reshape(variances.size)).all() newMeans = numpy.array([[3, 70, 2], [4, 72, 2]], 'float64') newVariances = numpy.array([[1, 1, 1], [2, 2, 2]], 'float64') # Checks particular varianceThresholds-related methods varianceThresholds1D = numpy.array([0.3, 1, 0.5], 'float64') gmm.set_variance_thresholds(varianceThresholds1D) assert (gmm.variance_thresholds[0,:] == varianceThresholds1D).all() assert (gmm.variance_thresholds[1,:] == varianceThresholds1D).all() gmm.set_variance_thresholds(0.005) assert (gmm.variance_thresholds == 0.005).all() # Checks Gaussians access gmm.means = newMeans gmm.variances = newVariances assert (gmm.get_gaussian(0).mean == newMeans[0,:]).all() assert (gmm.get_gaussian(1).mean == newMeans[1,:]).all() assert (gmm.get_gaussian(0).variance == newVariances[0,:]).all() assert (gmm.get_gaussian(1).variance == newVariances[1,:]).all() # Checks resize gmm.resize(4,5) assert gmm.shape == (4,5) # Checks comparison gmm2 = GMMMachine(gmm) gmm3 = GMMMachine(2,3) gmm3.weights = weights2 gmm3.means = means gmm3.variances = variances #gmm3.varianceThresholds = varianceThresholds gmm4 = GMMMachine(2,3) gmm4.weights = weights gmm4.means = means2 gmm4.variances = variances #gmm4.varianceThresholds = varianceThresholds gmm5 = GMMMachine(2,3) gmm5.weights = weights gmm5.means = means gmm5.variances = variances2 #gmm5.varianceThresholds = varianceThresholds gmm6 = GMMMachine(2,3) gmm6.weights = weights gmm6.means = means gmm6.variances = variances #gmm6.varianceThresholds = varianceThresholds2 assert gmm == gmm2 assert (gmm != gmm2) is False assert gmm.is_similar_to(gmm2) assert gmm != gmm3 assert (gmm == gmm3) is False assert gmm.is_similar_to(gmm3) is False assert gmm != gmm4 assert (gmm == gmm4) is False assert gmm.is_similar_to(gmm4) is False assert gmm != gmm5 assert (gmm == gmm5) is False assert gmm.is_similar_to(gmm5) is False assert gmm != gmm6 assert (gmm == gmm6) is False assert gmm.is_similar_to(gmm6) is False
def test_ISVMachine():
# Creates a UBM
weights = numpy.array([0.4, 0.6], 'float64')
means = numpy.array([[1, 6, 2], [4, 3, 2]], 'float64')
variances = numpy.array([[1, 2, 1], [2, 1, 2]], 'float64')
ubm = GMMMachine(2,3)
ubm.weights = weights
ubm.means = means
ubm.variances = variances
# Creates a ISVBaseMachine
U = numpy.array([[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12]], 'float64')
#V = numpy.array([[0], [0], [0], [0], [0], [0]], 'float64')
d = numpy.array([0, 1, 0, 1, 0, 1], 'float64')
base = ISVBase(ubm,2)
base.u = U
#base.v = V
base.d = d
# Creates a JFAMachine
z = numpy.array([3,4,1,2,0,1], 'float64')
m = ISVMachine(base)
m.z = z
n_gaussians,dim,ru = m.shape
supervector_length = m.supervector_length
assert n_gaussians == 2
assert dim == 3
assert supervector_length == 6
assert ru == 2
assert (m.z == z).all()
# Saves and loads
filename = str(tempfile.mkstemp(".hdf5")[1])
m.save(bob.io.base.HDF5File(filename, 'w'))
m_loaded = ISVMachine(bob.io.base.HDF5File(filename))
m_loaded.isv_base = base
assert m == m_loaded
assert (m != m_loaded) is False
assert m.is_similar_to(m_loaded)
# Copy constructor
mc = ISVMachine(m)
assert m == mc
# Variant
mv = ISVMachine(base)
# Checks for correctness
#mv.isv_base = base
m.z = z
n_gaussians,dim,ru = m.shape
supervector_length = m.supervector_length
assert n_gaussians == 2
assert dim == 3
assert supervector_length == 6
assert ru == 2
assert (m.z == z).all()
# Defines GMMStats
gs = GMMStats(2,3)
log_likelihood = -3.
T = 1
n = numpy.array([0.4, 0.6], 'float64')
sumpx = numpy.array([[1., 2., 3.], [4., 5., 6.]], 'float64')
sumpxx = numpy.array([[10., 20., 30.], [40., 50., 60.]], 'float64')
gs.log_likelihood = log_likelihood
gs.t = T
gs.n = n
gs.sum_px = sumpx
gs.sum_pxx = sumpxx
# Forward GMMStats and check estimated value of the x speaker factor
eps = 1e-10
x_ref = numpy.array([0.291042849767692, 0.310273618998444], 'float64')
score_ref = -3.280498193082100
score = m(gs)
assert numpy.allclose(m.x, x_ref, eps)
assert abs(score_ref-score) < eps
# Check using alternate forward() method
supervector_length = m.supervector_length
Ux = numpy.ndarray(shape=(supervector_length,), dtype=numpy.float64)
m.estimate_ux(gs, Ux)
score = m.forward_ux(gs, Ux)
assert abs(score_ref-score) < eps
# x and Ux
x = numpy.ndarray((2,), numpy.float64)
m.estimate_x(gs, x)
n_gaussians,dim,_ = m.shape
x_py = estimate_x(n_gaussians, dim, ubm.mean_supervector, ubm.variance_supervector, U, n, sumpx)
assert numpy.allclose(x, x_py, eps)
ux = numpy.ndarray((6,), numpy.float64)
m.estimate_ux(gs, ux)
n_gaussians,dim,_ = m.shape
ux_py = estimate_ux(n_gaussians, dim, ubm.mean_supervector, ubm.variance_supervector, U, n, sumpx)
assert numpy.allclose(ux, ux_py, eps)
assert numpy.allclose(m.x, x, eps)
score = m.forward_ux(gs, ux)
assert abs(score_ref-score) < eps
# Clean-up
os.unlink(filename)