def __init__(
self,
block_size = 12, # 1 or two parameters for block size
block_overlap = 11, # 1 or two parameters for block overlap
number_of_dct_coefficients = 45,
normalize_blocks = True,
normalize_dcts = True,
auto_reduce_coefficients = False
):
# call base class constructor
Extractor.__init__(
self,
block_size = block_size,
block_overlap = block_overlap,
number_of_dct_coefficients = number_of_dct_coefficients,
normalize_blocks = normalize_blocks,
normalize_dcts = normalize_dcts,
auto_reduce_coefficients = auto_reduce_coefficients
)
# block parameters
block_size = block_size if isinstance(block_size, (tuple, list)) else (block_size, block_size)
block_overlap = block_overlap if isinstance(block_overlap, (tuple, list)) else (block_overlap, block_overlap)
if block_size[0] < block_overlap[0] or block_size[1] < block_overlap[1]:
raise ValueError("The overlap '%s' is bigger than the block size '%s'. This won't work. Please check your setup!"%(block_overlap, block_size))
if block_size[0] * block_size[1] <= number_of_dct_coefficients:
if auto_reduce_coefficients:
number_of_dct_coefficients = block_size[0] * block_size[1] - 1
else:
raise ValueError("You selected more coefficients %d than your blocks have %d. This won't work. Please check your setup!"%(number_of_dct_coefficients, block_size[0] * block_size[1]))
self.dct_features = bob.ip.base.DCTFeatures(number_of_dct_coefficients, block_size, block_overlap, normalize_blocks, normalize_dcts)
def __init__(
self,
win_length_ms = 20,
win_shift_ms = 10,
n_filters = 24 ,
dct_norm = False,
f_min = 0.0,
f_max = 4000,
delta_win = 2,
mel_scale = True,
with_energy = True,
with_delta = True,
with_delta_delta = True,
n_ceps = 19, # 0-->18
pre_emphasis_coef = 0.95,
features_mask = numpy.arange(0,60),
# Normalization
normalize_flag = True,
**kwargs
):
# call base class constructor with its set of parameters
Extractor.__init__(
self,
win_length_ms = win_length_ms,
win_shift_ms = win_shift_ms,
n_filters = n_filters,
dct_norm = dct_norm,
f_min = f_min,
f_max = f_max,
delta_win = delta_win,
mel_scale = mel_scale,
with_energy = with_energy,
with_delta = with_delta,
with_delta_delta = with_delta_delta,
n_ceps = n_ceps,
pre_emphasis_coef = pre_emphasis_coef,
features_mask = features_mask,
normalize_flag = normalize_flag,
)
# copy parameters
self.win_length_ms = win_length_ms
self.win_shift_ms = win_shift_ms
self.n_filters = n_filters
self.dct_norm = dct_norm
self.f_min = f_min
self.f_max = f_max
self.delta_win = delta_win
self.mel_scale = mel_scale
self.with_energy = with_energy
self.with_delta = with_delta
self.with_delta_delta = with_delta_delta
self.n_ceps = n_ceps
self.pre_emphasis_coef = pre_emphasis_coef
self.features_mask = features_mask
self.normalize_flag = normalize_flag
def __init__(
self,
win_length_ms=20,
win_shift_ms=10,
n_filters=24,
dct_norm=False,
f_min=0.0,
f_max=4000,
delta_win=2,
mel_scale=True,
with_energy=True,
with_delta=True,
with_delta_delta=True,
n_ceps=19, # 0-->18
pre_emphasis_coef=0.95,
features_mask=numpy.arange(0, 60),
# Normalization
normalize_flag=True,
**kwargs):
# call base class constructor with its set of parameters
Extractor.__init__(
self,
win_length_ms=win_length_ms,
win_shift_ms=win_shift_ms,
n_filters=n_filters,
dct_norm=dct_norm,
f_min=f_min,
f_max=f_max,
delta_win=delta_win,
mel_scale=mel_scale,
with_energy=with_energy,
with_delta=with_delta,
with_delta_delta=with_delta_delta,
n_ceps=n_ceps,
pre_emphasis_coef=pre_emphasis_coef,
features_mask=features_mask,
normalize_flag=normalize_flag,
)
# copy parameters
self.win_length_ms = win_length_ms
self.win_shift_ms = win_shift_ms
self.n_filters = n_filters
self.dct_norm = dct_norm
self.f_min = f_min
self.f_max = f_max
self.delta_win = delta_win
self.mel_scale = mel_scale
self.with_energy = with_energy
self.with_delta = with_delta
self.with_delta_delta = with_delta_delta
self.n_ceps = n_ceps
self.pre_emphasis_coef = pre_emphasis_coef
self.features_mask = features_mask
self.normalize_flag = normalize_flag
def __init__(self,
features_mask=numpy.arange(0, 60),
normalize_flag=True,
**kwargs):
# call base class constructor with its set of parameters
Extractor.__init__(
self,
features_mask=features_mask,
normalize_flag=normalize_flag,
)
# copy parameters
self.features_mask = features_mask
self.normalize_flag = normalize_flag
def __init__(
self,
features_mask = numpy.arange(0,60),
normalize_flag = True,
**kwargs
):
# call base class constructor with its set of parameters
Extractor.__init__(
self,
features_mask = features_mask,
normalize_flag = normalize_flag,
)
# copy parameters
self.features_mask = features_mask
self.normalize_flag = normalize_flag
def __init__(
self,
split_training_data_by_client=False,
features_mask=numpy.zeros(90), # mask of which features to read
**kwargs):
# call base class constructor with its set of parameters
Preprocessor.__init__(self,
read_original_data=self.read_matlab_files,
**kwargs)
Extractor.__init__(
self,
requires_training=False,
split_training_data_by_client=split_training_data_by_client,
**kwargs)
self.features_mask = features_mask
def __init__(
self,
block_size=12, # 1 or two parameters for block size
block_overlap=11, # 1 or two parameters for block overlap
number_of_dct_coefficients=45,
normalize_blocks=True,
normalize_dcts=True,
auto_reduce_coefficients=False):
# call base class constructor
Extractor.__init__(
self,
block_size=block_size,
block_overlap=block_overlap,
number_of_dct_coefficients=number_of_dct_coefficients,
normalize_blocks=normalize_blocks,
normalize_dcts=normalize_dcts,
auto_reduce_coefficients=auto_reduce_coefficients)
# block parameters
block_size = block_size if isinstance(block_size,
(tuple, list)) else (block_size,
block_size)
block_overlap = block_overlap if isinstance(
block_overlap, (tuple, list)) else (block_overlap, block_overlap)
if block_size[0] < block_overlap[0] or block_size[1] < block_overlap[1]:
raise ValueError(
"The overlap '%s' is bigger than the block size '%s'. This won't work. Please check your setup!"
% (block_overlap, block_size))
if block_size[0] * block_size[1] <= number_of_dct_coefficients:
if auto_reduce_coefficients:
number_of_dct_coefficients = block_size[0] * block_size[1] - 1
else:
raise ValueError(
"You selected more coefficients %d than your blocks have %d. This won't work. Please check your setup!"
% (number_of_dct_coefficients,
block_size[0] * block_size[1]))
self.dct_features = bob.ip.base.DCTFeatures(number_of_dct_coefficients,
block_size, block_overlap,
normalize_blocks,
normalize_dcts)
def __init__(
self,
# Gabor parameters
gabor_directions = 8,
gabor_scales = 5,
gabor_sigma = 2. * math.pi,
gabor_maximum_frequency = math.pi / 2.,
gabor_frequency_step = math.sqrt(.5),
gabor_power_of_k = 0,
gabor_dc_free = True,
# what kind of information to extract
normalize_gabor_jets = True,
# setup of the aligned grid
eyes = None, # if set, the grid setup will be aligned to the eye positions {'leye' : LEFT_EYE_POS, 'reye' : RIGHT_EYE_POS},
nodes_between_eyes = 4,
nodes_along_eyes = 2,
nodes_above_eyes = 3,
nodes_below_eyes = 7,
# setup of static grid
node_distance = None, # one or two integral values
first_node = None, # one or two integral values, or None -> automatically determined
):
# call base class constructor
Extractor.__init__(
self,
gabor_directions = gabor_directions,
gabor_scales = gabor_scales,
gabor_sigma = gabor_sigma,
gabor_maximum_frequency = gabor_maximum_frequency,
gabor_frequency_step = gabor_frequency_step,
gabor_power_of_k = gabor_power_of_k,
gabor_dc_free = gabor_dc_free,
normalize_gabor_jets = normalize_gabor_jets,
eyes = eyes,
nodes_between_eyes = nodes_between_eyes,
nodes_along_eyes = nodes_along_eyes,
nodes_above_eyes = nodes_above_eyes,
nodes_below_eyes = nodes_below_eyes,
node_distance = node_distance,
first_node = first_node
)
# create Gabor wavelet transform class
self.gwt = bob.ip.gabor.Transform(
number_of_scales = gabor_scales,
number_of_directions = gabor_directions,
sigma = gabor_sigma,
k_max = gabor_maximum_frequency,
k_fac = gabor_frequency_step,
power_of_k = gabor_power_of_k,
dc_free = gabor_dc_free
)
# create graph extractor
if eyes is not None:
self._aligned_graph = bob.ip.gabor.Graph(
righteye = [int(e) for e in eyes['reye']],
lefteye = [int(e) for e in eyes['leye']],
between = int(nodes_between_eyes),
along = int(nodes_along_eyes),
above = int(nodes_above_eyes),
below = int(nodes_below_eyes)
)
else:
if node_distance is None:
raise ValueError("Please specify either 'eyes' or the grid parameters 'node_distance' (and 'first_node')!")
self._aligned_graph = None
self._last_image_resolution = None
self.first_node = first_node
self.node_distance = node_distance
if isinstance(self.node_distance, (int, float)):
self.node_distance = (int(self.node_distance), int(self.node_distance))
self.normalize_jets = normalize_gabor_jets
self.trafo_image = None
def __init__(
self,
# Block setup
block_size, # one or two parameters for block size
block_overlap = 0, # one or two parameters for block overlap
# Gabor parameters
gabor_directions = 8,
gabor_scales = 5,
gabor_sigma = 2. * math.pi,
gabor_maximum_frequency = math.pi / 2.,
gabor_frequency_step = math.sqrt(.5),
gabor_power_of_k = 0,
gabor_dc_free = True,
use_gabor_phases = False,
# LBP parameters
lbp_radius = 2,
lbp_neighbor_count = 8,
lbp_uniform = True,
lbp_circular = True,
lbp_rotation_invariant = False,
lbp_compare_to_average = False,
lbp_add_average = False,
# histogram options
sparse_histogram = False,
split_histogram = None
):
# call base class constructor
Extractor.__init__(
self,
block_size = block_size,
block_overlap = block_overlap,
gabor_directions = gabor_directions,
gabor_scales = gabor_scales,
gabor_sigma = gabor_sigma,
gabor_maximum_frequency = gabor_maximum_frequency,
gabor_frequency_step = gabor_frequency_step,
gabor_power_of_k = gabor_power_of_k,
gabor_dc_free = gabor_dc_free,
use_gabor_phases = use_gabor_phases,
lbp_radius = lbp_radius,
lbp_neighbor_count = lbp_neighbor_count,
lbp_uniform = lbp_uniform,
lbp_circular = lbp_circular,
lbp_rotation_invariant = lbp_rotation_invariant,
lbp_compare_to_average = lbp_compare_to_average,
lbp_add_average = lbp_add_average,
sparse_histogram = sparse_histogram,
split_histogram = split_histogram
)
# block parameters
self.block_size = block_size if isinstance(block_size, (tuple, list)) else (block_size, block_size)
self.block_overlap = block_overlap if isinstance(block_overlap, (tuple, list)) else (block_overlap, block_overlap)
if self.block_size[0] < self.block_overlap[0] or self.block_size[1] < self.block_overlap[1]:
raise ValueError("The overlap is bigger than the block size. This won't work. Please check your setup!")
# Gabor wavelet transform class
self.gwt = bob.ip.gabor.Transform(
number_of_scales = gabor_scales,
number_of_directions = gabor_directions,
sigma = gabor_sigma,
k_max = gabor_maximum_frequency,
k_fac = gabor_frequency_step,
power_of_k = gabor_power_of_k,
dc_free = gabor_dc_free
)
self.trafo_image = None
self.use_phases = use_gabor_phases
self.lbp = bob.ip.base.LBP(
neighbors = lbp_neighbor_count,
radius = float(lbp_radius),
circular = lbp_circular,
to_average = lbp_compare_to_average,
add_average_bit = lbp_add_average,
uniform = lbp_uniform,
rotation_invariant = lbp_rotation_invariant,
border_handling = 'wrap'
)
self.split = split_histogram
self.sparse = sparse_histogram
if self.sparse and self.split:
raise ValueError("Sparse histograms cannot be split! Check your setup!")
def __init__(
self,
win_length_ms=20., # 20 ms
win_shift_ms=10., # 10 ms
n_filters=40,
f_min=0.0, # 0 Hz
f_max=8000, # 8 KHz - this is an important value. Normally it should be half of the sampling frequency
pre_emphasis_coef=1.0,
mel_scale=True,
rect_filter=False,
inverse_filter=False,
delta_win=2,
n_ceps=19, # 0-->18,
dct_norm=False,
ssfc_features=False,
scfc_features=False,
scmc_features=False,
with_delta=True,
with_delta_delta=True,
with_energy=False,
normalize_spectrum=False,
keep_only_deltas=True,
log_filter=True,
energy_filter=False,
vad_filter="no_filter", # we do apply any trim filter by default
normalize_feature_vector=False,
**kwargs):
# call base class constructor with its set of parameters
Extractor.__init__(self,
requires_training=False,
split_training_data_by_client=False,
**kwargs)
# copy parameters
self.win_length_ms = win_length_ms
self.win_shift_ms = win_shift_ms
self.n_filters = n_filters
self.f_min = f_min
self.f_max = f_max
self.pre_emphasis_coef = pre_emphasis_coef
self.mel_scale = mel_scale
self.rect_filter = rect_filter
self.inverse_filter = inverse_filter
self.delta_win = delta_win
self.n_ceps = n_ceps
self.dct_norm = dct_norm
self.ssfc_features = ssfc_features
self.scfc_features = scfc_features
self.scmc_features = scmc_features
self.with_delta = with_delta
self.with_delta_delta = with_delta_delta
self.with_energy = with_energy
self.normalize_spectrum = normalize_spectrum
self.keep_only_deltas = keep_only_deltas
self.log_filter = log_filter
self.energy_filter = energy_filter
self.vad_filter = vad_filter
self.normalize_feature_vector = normalize_feature_vector
# compute the size of the feature vector
self.features_len = self.n_ceps
if self.with_delta:
self.features_len += self.n_ceps
if self.with_delta_delta:
self.features_len += self.n_ceps
def __init__(self, subspace_dimension): # We have to register that this function will need a training step Extractor.__init__(self, requires_training = True, subspace_dimension = subspace_dimension) self.subspace_dimension = subspace_dimension
def __init__(self, **kwargs): Extractor.__init__(self, requires_training=True) self.model = False
def __init__(self, **kwargs):
Extractor.__init__(self, requires_training=True)
self.model = False
def __init__(
self,
win_length_ms=20., # 20 ms
win_shift_ms=10., # 10 ms
n_filters=40,
f_min=0.0, # 0 Hz
f_max=8000, # 8 KHz - this is an important value. Normally it should be half of the sampling frequency
pre_emphasis_coef=1.0,
mel_scale=False,
rect_filter=False,
inverse_filter=False,
delta_win=2,
n_ceps=19, # 0-->18,
dct_norm=False,
d= 1,
p= 3,
k= 7,
with_delta=True,
with_delta_delta=False,
with_energy=False,
normalize_spectrum=False,
keep_only_deltas=True,
log_filter=True,
energy_filter=False,
vad_filter="no_filter", # we do apply any trim filter by default
normalize_feature_vector = False,
**kwargs
):
# call base class constructor with its set of parameters
Extractor.__init__(
self,
requires_training=False, split_training_data_by_client=False,
**kwargs
)
# copy parameters
self.win_length_ms = win_length_ms
self.win_shift_ms = win_shift_ms
self.n_filters = n_filters
self.f_min = f_min
self.f_max = f_max
self.pre_emphasis_coef = pre_emphasis_coef
self.mel_scale = mel_scale
self.rect_filter = rect_filter
self.inverse_filter = inverse_filter
self.delta_win = delta_win
self.n_ceps = n_ceps
self.dct_norm = dct_norm
self.d = d
self.p = p
self.k = k
self.with_delta = with_delta
self.with_delta_delta = with_delta_delta
self.with_energy = with_energy
self.normalize_spectrum = normalize_spectrum
self.keep_only_deltas = keep_only_deltas
self.log_filter = log_filter
self.energy_filter = energy_filter
self.vad_filter = vad_filter
self.normalize_feature_vector = normalize_feature_vector
# compute the size of the feature vector
self.features_len = self.n_ceps
if self.with_delta:
self.features_len += self.n_ceps
if self.with_delta_delta:
self.features_len += self.n_ceps
def normalize_features(self, features):
mean = numpy.mean(features, axis=0)
std = numpy.std(features, axis=0)
return numpy.divide(features-mean, std)
def compute_ceps(self, rate, data):
ceps = bob.ap.Ceps(rate, self.win_length_ms, self.win_shift_ms, self.n_filters, self.n_ceps, self.f_min,
self.f_max, self.delta_win, self.pre_emphasis_coef)
ceps.dct_norm = self.dct_norm
ceps.mel_scale = self.mel_scale
ceps.rect_filter = self.rect_filter
ceps.inverse_filter = self.inverse_filter
ceps.with_energy = self.with_energy
ceps.with_delta = self.with_delta
ceps.with_delta_delta = self.with_delta_delta
ceps.normalize_spectrum = self.normalize_spectrum
ceps.log_filter = self.log_filter
ceps.energy_filter = self.energy_filter
cepstral_features = ceps(data)
if self.keep_only_deltas:
cepstral_features = cepstral_features[:, self.n_ceps:]
return cepstral_features
def shifted_delta_cepstral(cep, d, p, k):
y = numpy.r_[numpy.resize(cep[0, :], (d, cep.shape[1])), cep, numpy.resize(cep[-1, :], (k * 3 + d, cep.shape[1]))]
delta = self.with_delta
sdc = numpy.empty((cep.shape[0], cep.shape[1] * k))
idx = numpy.zeros(delta.shape[0], dtype='bool')
for ii in range(k):
idx[d + ii * p] = True
for ff in range(len(cep)):
sdc[ff, :] = delta[idx, :].reshape(1, -1)
idx = numpy.roll(idx, 1)
return numpy.hstack((cep, sdc))
def _call_(self, input_data):
ceps = self._call_(input_data)
return shifted_delta_cepstral(ceps,d,p,k)
def __init__(self, subspace_dimension):
# We have to register that this function will need a training step
Extractor.__init__(self, requires_training=True, subspace_dimension=subspace_dimension)
self.subspace_dimension = subspace_dimension
def __init__(
self,
# Block setup
block_size, # one or two parameters for block size
block_overlap=0, # one or two parameters for block overlap
# Gabor parameters
gabor_directions=8,
gabor_scales=5,
gabor_sigma=2. * math.pi,
gabor_maximum_frequency=math.pi / 2.,
gabor_frequency_step=math.sqrt(.5),
gabor_power_of_k=0,
gabor_dc_free=True,
use_gabor_phases=False,
# LBP parameters
lbp_radius=2,
lbp_neighbor_count=8,
lbp_uniform=True,
lbp_circular=True,
lbp_rotation_invariant=False,
lbp_compare_to_average=False,
lbp_add_average=False,
# histogram options
sparse_histogram=False,
split_histogram=None):
# call base class constructor
Extractor.__init__(self,
block_size=block_size,
block_overlap=block_overlap,
gabor_directions=gabor_directions,
gabor_scales=gabor_scales,
gabor_sigma=gabor_sigma,
gabor_maximum_frequency=gabor_maximum_frequency,
gabor_frequency_step=gabor_frequency_step,
gabor_power_of_k=gabor_power_of_k,
gabor_dc_free=gabor_dc_free,
use_gabor_phases=use_gabor_phases,
lbp_radius=lbp_radius,
lbp_neighbor_count=lbp_neighbor_count,
lbp_uniform=lbp_uniform,
lbp_circular=lbp_circular,
lbp_rotation_invariant=lbp_rotation_invariant,
lbp_compare_to_average=lbp_compare_to_average,
lbp_add_average=lbp_add_average,
sparse_histogram=sparse_histogram,
split_histogram=split_histogram)
# block parameters
self.block_size = block_size if isinstance(block_size,
(tuple,
list)) else (block_size,
block_size)
self.block_overlap = block_overlap if isinstance(
block_overlap, (tuple, list)) else (block_overlap, block_overlap)
if self.block_size[0] < self.block_overlap[0] or self.block_size[
1] < self.block_overlap[1]:
raise ValueError(
"The overlap is bigger than the block size. This won't work. Please check your setup!"
)
# Gabor wavelet transform class
self.gwt = bob.ip.gabor.Transform(
number_of_scales=gabor_scales,
number_of_directions=gabor_directions,
sigma=gabor_sigma,
k_max=gabor_maximum_frequency,
k_fac=gabor_frequency_step,
power_of_k=gabor_power_of_k,
dc_free=gabor_dc_free)
self.trafo_image = None
self.use_phases = use_gabor_phases
self.lbp = bob.ip.base.LBP(neighbors=lbp_neighbor_count,
radius=float(lbp_radius),
circular=lbp_circular,
to_average=lbp_compare_to_average,
add_average_bit=lbp_add_average,
uniform=lbp_uniform,
rotation_invariant=lbp_rotation_invariant,
border_handling='wrap')
self.split = split_histogram
self.sparse = sparse_histogram
if self.sparse and self.split:
raise ValueError(
"Sparse histograms cannot be split! Check your setup!")
def __init__(
self,
win_length_ms=20., # 20 ms
win_shift_ms=10., # 10 ms
n_filters=40,
f_min=0.0, # 0 Hz
f_max=8000, # 8 KHz - this is an important value. Normally it should be half of the sampling frequency
pre_emphasis_coef=1.0,
mel_scale=True,
rect_filter=False,
inverse_filter=False,
delta_win=2,
n_ceps=19, # 0-->18,
dct_norm=False,
ssfc_features=False,
scfc_features=False,
scmc_features=False,
with_delta=True,
with_delta_delta=True,
with_energy=False,
normalize_spectrum=False,
keep_only_deltas=True,
log_filter=True,
energy_filter=False,
vad_filter="no_filter", # we do apply any trim filter by default
normalize_feature_vector = False,
**kwargs
):
# call base class constructor with its set of parameters
Extractor.__init__(
self,
requires_training=False, split_training_data_by_client=False,
**kwargs
)
# copy parameters
self.win_length_ms = win_length_ms
self.win_shift_ms = win_shift_ms
self.n_filters = n_filters
self.f_min = f_min
self.f_max = f_max
self.pre_emphasis_coef = pre_emphasis_coef
self.mel_scale = mel_scale
self.rect_filter = rect_filter
self.inverse_filter = inverse_filter
self.delta_win = delta_win
self.n_ceps = n_ceps
self.dct_norm = dct_norm
self.ssfc_features = ssfc_features
self.scfc_features = scfc_features
self.scmc_features = scmc_features
self.with_delta = with_delta
self.with_delta_delta = with_delta_delta
self.with_energy = with_energy
self.normalize_spectrum = normalize_spectrum
self.keep_only_deltas = keep_only_deltas
self.log_filter = log_filter
self.energy_filter = energy_filter
self.vad_filter = vad_filter
self.normalize_feature_vector = normalize_feature_vector
# compute the size of the feature vector
self.features_len = self.n_ceps
if self.with_delta:
self.features_len += self.n_ceps
if self.with_delta_delta:
self.features_len += self.n_ceps
