def transform(self,
X,
indices=None,
sad=None,
save_ivecs=False,
keep_stats=False,
name=None):
"""
Parameters
----------
X : ndarray
Training data [n_samples, n_features]
indices : {Mapping, tuple, list}
in case the data is given by a list of files, `indices`
act as file indicator mapping from
'file_name' -> (start_index_in_X, end_index_in_X)
This mapping can be provided by a dictionary, or list of
tuple.
sad : ndarray
inspired by the "Speech Activity Detection" (SAD) indexing,
this array is indicator of which samples will be taken into
training; the shape should be [n_samples,] or [n_samples, 1]
save_ivecs : bool
if True, save extracted i-vectors to disk at path `ivec_[name]`
if False, return directly the i-vectors without saving
keep_stats : bool
if True, keep the zero and first order statistics.
The first order statistics could consume huge amount
of disk space. Otherwise, they are deleted after training
name : {None, str}
identity of the i-vectors (for re-using in future).
If None, a random name is used
"""
if not self.is_fitted:
raise ValueError(
"Ivector has not been fitted, call Ivector.fit(...) first")
n_files = X.shape[0] if indices is None else len(indices)
if name is None:
name = uuid(length=8)
else:
name = str(name)
# ====== init ====== #
z_path = self.get_z_path(name)
f_path = self.get_f_path(name)
if save_ivecs:
i_path = self.get_i_path(name)
else:
i_path = None
name_path = self.get_name_path(name)
# ====== check exist i-vector file ====== #
if i_path is not None and os.path.exists(i_path):
ivec = MmapArray(path=i_path)
assert ivec.shape[0] == n_files and ivec.shape[1] == self.tv_dim,\
"Need i-vectors for %d files, found exists data at path:'%s' with shape:%s" % \
(n_files, i_path, ivec.shape)
return ivec
# ====== extract Z and F ====== #
if os.path.exists(z_path) and os.path.exists(f_path):
pass
else:
if os.path.exists(z_path):
os.remove(z_path)
if os.path.exists(f_path):
os.remove(f_path)
if os.path.exists(name_path):
os.remove(name_path)
_extract_zero_and_first_stats(X=X,
sad=sad,
indices=indices,
gmm=self.gmm,
z_path=z_path,
f_path=f_path,
name_path=name_path)
Z = MmapArray(path=z_path)
F = MmapArray(path=f_path)
# ====== extract I-vec ====== #
ivec = self.tmat.transform_to_disk(path=i_path,
Z=Z,
F=F,
dtype='float32')
# ====== clean ====== #
Z.close()
F.close()
if not keep_stats:
if os.path.exists(z_path):
os.remove(z_path)
if os.path.exists(f_path):
os.remove(f_path)
else:
print("Zero-order stats saved at:", ctext(z_path, 'cyan'))
print("First-order stats saved at:", ctext(f_path, 'cyan'))
return ivec
def fit(self,
X,
indices=None,
sad=None,
refit_gmm=False,
refit_tmat=False,
extract_ivecs=False,
keep_stats=False):
"""
Parameters
----------
X : ndarray
Training data [n_samples, n_features]
indices : {Mapping, tuple, list}
in case the data is given by a list of files, `indices`
act as file indicator mapping from
'file_name' -> (start_index_in_X, end_index_in_X)
This mapping can be provided by a dictionary, or list of
tuple.
Note: the order provided in indices will be preserved
sad : ndarray
inspired by the "Speech Activity Detection" (SAD) indexing,
this array is indicator of which samples will be taken into
training; the shape should be [n_samples,] or [n_samples, 1]
refit_gmm : bool
if True, re-fit the GMM even though it is fitted,
consequently, the T-matrix will be re-fitted
refit_tmat : bool
if True, re-fit the T-matrix even though it is fitted
extract_ivecs : bool
if True, extract the i-vector for training data
keep_stats : bool
if True, keep the zero and first order statistics.
The first order statistics could consume huge amount
of disk space. Otherwise, they are deleted after training
"""
new_gmm = (not self.gmm.is_fitted or refit_gmm)
# ====== clean error files ====== #
if os.path.exists(self.z_path):
Z = MmapArray(self.z_path)
if Z.shape[0] == 0: # empty file
os.remove(self.z_path)
Z.close()
if os.path.exists(self.f_path):
F = MmapArray(self.f_path)
if F.shape[0] == 0: # empty file
os.remove(self.f_path)
F.close()
if os.path.exists(self.ivec_path):
ivec = MmapArray(self.ivec_path)
if ivec.shape[0] == 0: # empty file
os.remove(self.ivec_path)
ivec.close()
# ====== Training the GMM first ====== #
if new_gmm:
input_data = [X]
if sad is not None:
input_data.append(sad)
if indices is not None:
input_data.append(indices)
self.gmm.fit(input_data)
# ====== some fun, and confusing logics ====== #
# GMM need to be fitted before creating T-matrix model
new_tmat = (not self.tmat.is_fitted or new_gmm or refit_tmat)
# New I-vector is need when:
# - only when `extract_ivecs=True`
# - and new T-matrix is trained but no I-vector is extracted
new_ivec = extract_ivecs and \
(new_tmat or not os.path.exists(self.ivec_path))
# new stats is only needed when
# - GMM is updated
# - training new Tmatrix and the Z and F not exist
# - extracting new I-vector and the Z and F not exist
if not new_gmm and \
(os.path.exists(self.z_path) and os.path.exists(self.f_path)):
new_stats = False
else:
new_stats = new_gmm or new_tmat or new_ivec
# ====== extract the statistics ====== #
if new_stats:
_extract_zero_and_first_stats(X=X,
sad=sad,
indices=indices,
gmm=self.gmm,
z_path=self.z_path,
f_path=self.f_path,
name_path=self.name_path)
# ====== Training the T-matrix and extract i-vector ====== #
if new_tmat or new_ivec:
Z = MmapArray(path=self.z_path)
F = MmapArray(path=self.f_path)
if new_tmat:
self.tmat.fit((Z, F))
if new_ivec:
self.tmat.transform_to_disk(path=self.ivec_path,
Z=Z,
F=F,
dtype='float32',
device='gpu',
override=True)
Z.close()
F.close()
# ====== clean ====== #
if not keep_stats:
if os.path.exists(self.z_path):
os.remove(self.z_path)
if os.path.exists(self.f_path):
os.remove(self.f_path)
return self