From dafa70520f5ad730daf349fa0a36f90639b8431b Mon Sep 17 00:00:00 2001
From: Yannick DAYER <yannick.dayer@idiap.ch>
Date: Tue, 31 Jan 2023 11:20:59 +0100
Subject: [PATCH] [py] Update ISV to accept probe templates.
---
.pre-commit-config.yaml | 8 +--
src/bob/learn/em/factor_analysis.py | 91 +++++++++++++++++------------
src/bob/learn/em/ivector.py | 23 +++++---
3 files changed, 72 insertions(+), 50 deletions(-)
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
index b4d1d1c..6148eb3 100644
--- a/.pre-commit-config.yaml
+++ b/.pre-commit-config.yaml
@@ -2,20 +2,20 @@
# See https://pre-commit.com/hooks.html for more hooks
repos:
- repo: https://github.com/timothycrosley/isort
- rev: 5.10.1
+ rev: 5.12.0
hooks:
- id: isort
args: [--settings-path, "pyproject.toml"]
- repo: https://github.com/psf/black
- rev: 22.3.0
+ rev: 22.12.0
hooks:
- id: black
- repo: https://github.com/pycqa/flake8
- rev: 3.9.2
+ rev: 6.0.0
hooks:
- id: flake8
- repo: https://github.com/pre-commit/pre-commit-hooks
- rev: v4.2.0
+ rev: v4.4.0
hooks:
- id: check-ast
- id: check-case-conflict
diff --git a/src/bob/learn/em/factor_analysis.py b/src/bob/learn/em/factor_analysis.py
index 37c3af6..64ae364 100644
--- a/src/bob/learn/em/factor_analysis.py
+++ b/src/bob/learn/em/factor_analysis.py
@@ -148,7 +148,6 @@ class FactorAnalysisBase(BaseEstimator):
@property
def feature_dimension(self):
"""Get the UBM Dimension"""
-
# TODO: Add this on the GMMMachine class
return self.ubm.means.shape[1]
@@ -159,16 +158,12 @@ class FactorAnalysisBase(BaseEstimator):
@property
def mean_supervector(self):
- """
- Returns the mean supervector
- """
+ """Returns the mean supervector"""
return self.ubm.means.flatten()
@property
def variance_supervector(self):
- """
- Returns the variance supervector
- """
+ """Returns the variance supervector"""
return self.ubm.variances.flatten()
@property
@@ -199,10 +194,7 @@ class FactorAnalysisBase(BaseEstimator):
self._V = np.array(value)
def estimate_number_of_classes(self, y):
- """
- Estimates the number of classes given the labels
- """
-
+ """Estimates the number of classes given the labels"""
return len(unique_labels(y))
def initialize_using_array(self, X):
@@ -279,6 +271,7 @@ class FactorAnalysisBase(BaseEstimator):
D: (n_gaussians*feature_dimension) represents the client offset vector
"""
+
if self.random_state is not None:
np.random.seed(self.random_state)
@@ -319,6 +312,7 @@ class FactorAnalysisBase(BaseEstimator):
(n_classes, n_gaussians) representing the accumulated 0th order statistics
"""
+
# 0th order stats
n_acc = np.zeros((n_classes, self.ubm.n_gaussians), like=X[0].n)
@@ -380,12 +374,11 @@ class FactorAnalysisBase(BaseEstimator):
i: int
Class id to return the statistics for
"""
+
indices = np.where(np.array(y) == i)[0]
return [X[i] for i in indices]
- """
- Estimating U and x
- """
+ # Estimating U and x #
def _compute_id_plus_u_prod_ih(self, x_i, UProd):
"""
@@ -430,6 +423,7 @@ class FactorAnalysisBase(BaseEstimator):
E[y_i] for class `i`
"""
+
f_i = x_i.sum_px
n_i = x_i.n
n_ic = np.repeat(n_i, self.feature_dimension)
@@ -447,9 +441,8 @@ class FactorAnalysisBase(BaseEstimator):
self.mean_supervector + self._D * latent_z_i
)
- """
- # JFA Part (eq 29)
- """
+ # JFA Part (eq 29) #
+
V_dot_v = V @ latent_y_i if latent_y_i is not None else 0
fn_x_ih -= n_ic * V_dot_v if latent_y_i is not None else 0
@@ -545,6 +538,7 @@ class FactorAnalysisBase(BaseEstimator):
Accumulated statistics for U_A2(n_gaussians* feature_dimension, r_U)
"""
+
# Inverting A1 over the zero axis
# https://stackoverflow.com/questions/11972102/is-there-a-way-to-efficiently-invert-an-array-of-matrices-with-numpy
inv_A1 = np.linalg.inv(acc_U_A1)
@@ -568,6 +562,7 @@ class FactorAnalysisBase(BaseEstimator):
https://gitlab.idiap.ch/bob/bob.learn.em/-/blob/da92d0e5799d018f311f1bf5cdd5a80e19e142ca/bob/learn/em/cpp/FABaseTrainer.cpp#L325
"""
+
# UProd = (self.ubm.n_gaussians, self.r_U, self.r_U)
Uc = self._U.reshape(
@@ -662,9 +657,7 @@ class FactorAnalysisBase(BaseEstimator):
return acc_U_A1, acc_U_A2
- """
- Estimating D and z
- """
+ # Estimating D and z #
def update_z(self, X, y, latent_x, latent_y, latent_z, n_acc, f_acc):
"""
@@ -878,6 +871,7 @@ class FactorAnalysisBase(BaseEstimator):
latent_x = (n_classes, r_U, n_sessions)
"""
+
kw = dict(like=like) if isinstance(like, dask.array.core.Array) else {}
# x (Eq. 36)
@@ -897,9 +891,7 @@ class FactorAnalysisBase(BaseEstimator):
return latent_x, latent_y, latent_z
- """
- Estimating V and y
- """
+ # Estimating V and y #
def update_y(
self,
@@ -948,6 +940,7 @@ class FactorAnalysisBase(BaseEstimator):
Accumulated 1st order statistics for each class (math:`F_{i}`)
"""
+
# V.T / sigma
VTinvSigma = self._V.T / self.variance_supervector
@@ -1019,7 +1012,7 @@ class FactorAnalysisBase(BaseEstimator):
I = np.eye(self.r_V, self.r_V) # noqa: E741
- # TODO: make the invertion matrix function as a parameter
+ # TODO: make the inversion matrix function as a parameter
return np.linalg.inv(I + (VProd * n_acc_i[:, None, None]).sum(axis=0))
def _compute_vprod(self):
@@ -1176,17 +1169,15 @@ class FactorAnalysisBase(BaseEstimator):
) # Fn_yi = sum_{sessions h}(N_{i,h}*(o_{i,h} - m - D*z_{i})
# Looping over the sessions of a ;ane;
- for session_id in range(len(X_i)):
- n_i = X_i[session_id].n
+ for session_id, x_i_s in enumerate(X_i):
+ n_i = x_i_s[session_id].n
U_dot_x = U @ latent_x_i[:, session_id]
tmp_CD = np.repeat(n_i, self.feature_dimension)
fn_y_i -= tmp_CD * U_dot_x
return fn_y_i
- """
- Scoring
- """
+ # Scoring #
def estimate_x(self, X):
@@ -1212,6 +1203,7 @@ class FactorAnalysisBase(BaseEstimator):
X_i: list of :py:class:`bob.learn.em.GMMStats`
List of statistics for a class
"""
+
I = np.eye(self.r_U, self.r_U) # noqa: E741
Uc = self._U.reshape(
@@ -1319,10 +1311,11 @@ class FactorAnalysisBase(BaseEstimator):
z
"""
+
return self.enroll([self.ubm.acc_stats(X)])
def _prepare_dask_input(self, X, y):
- """Perpare the input for the fit method"""
+ """Prepare the input for the fit method"""
logger.info(
"Rechunking bag of stats to delayed list of stats per class. If your worker runs "
"out of memory in this training step, you have to use workers with more memory."
@@ -1406,9 +1399,7 @@ class ISVMachine(FactorAnalysisBase):
)
def e_step(self, X, y, n_samples_per_class, n_acc, f_acc):
- """
- E-step of the EM algorithm
- """
+ """E-step of the EM algorithm."""
# self.initialize_XYZ(y)
UProd = self._compute_uprod()
_, _, latent_z = self.initialize_XYZ(
@@ -1452,6 +1443,7 @@ class ISVMachine(FactorAnalysisBase):
Accumulated statistics for U_A2(n_gaussians* feature_dimension, r_U)
"""
+
acc_U_A1 = [acc[0] for acc in acc_U_A1_acc_U_A2_list]
acc_U_A2 = [acc[1] for acc in acc_U_A1_acc_U_A2_list]
@@ -1476,6 +1468,7 @@ class ISVMachine(FactorAnalysisBase):
Returns self.
"""
+
if isinstance(X, dask.bag.Bag):
X, y = self._prepare_dask_input(X, y)
@@ -1535,6 +1528,7 @@ class ISVMachine(FactorAnalysisBase):
z
"""
+
iterations = self.enroll_iterations
# We have only one class for enrollment
y = list(np.zeros(len(X), dtype=np.int32))
@@ -1584,6 +1578,7 @@ class ISVMachine(FactorAnalysisBase):
z
"""
+
return self.enroll([self.ubm.acc_stats(X)])
def score(self, latent_z, data):
@@ -1596,7 +1591,7 @@ class ISVMachine(FactorAnalysisBase):
Latent representation of the client (E[z_i])
data : list of :py:class:`bob.learn.em.GMMStats`
- List of statistics to be scored
+ List of statistics of one probe template to be scored
Returns
-------
@@ -1604,7 +1599,14 @@ class ISVMachine(FactorAnalysisBase):
The linear scored
"""
- x = self.estimate_x(data)
+
+ template_data = data[0]
+
+ if len(data) > 1:
+ for d in data[1:]:
+ template_data += d
+
+ x = self.estimate_x(template_data)
Ux = self._U @ x
# TODO: I don't know why this is not the enrolled model
@@ -1615,10 +1617,10 @@ class ISVMachine(FactorAnalysisBase):
return linear_scoring(
z.reshape((self.ubm.n_gaussians, self.feature_dimension)),
self.ubm,
- data,
+ template_data,
Ux.reshape((self.ubm.n_gaussians, self.feature_dimension)),
frame_length_normalization=True,
- )[0]
+ )[0][0]
class JFAMachine(FactorAnalysisBase):
@@ -1759,6 +1761,7 @@ class JFAMachine(FactorAnalysisBase):
Accumulated statistics for V_A2(n_gaussians* feature_dimension, r_V)
"""
+
acc_V_A1 = [acc[0] for acc in acc_V_A1_acc_V_A2_list]
acc_V_A2 = [acc[1] for acc in acc_V_A1_acc_V_A2_list]
@@ -1807,6 +1810,7 @@ class JFAMachine(FactorAnalysisBase):
E[y]
"""
+
VProd = self._compute_vprod()
latent_x, latent_y, latent_z = self.initialize_XYZ(
@@ -1856,6 +1860,7 @@ class JFAMachine(FactorAnalysisBase):
Accumulated statistics for U_A2(n_gaussians* feature_dimension, r_U)
"""
+
# self.initialize_XYZ(y)
UProd = self._compute_uprod()
latent_x, _, latent_z = self.initialize_XYZ(
@@ -1897,6 +1902,7 @@ class JFAMachine(FactorAnalysisBase):
Accumulated statistics for V_A2(n_gaussians* feature_dimension, r_V)
"""
+
acc_U_A1 = [acc[0] for acc in acc_U_A1_acc_U_A2_list]
acc_U_A2 = [acc[1] for acc in acc_U_A1_acc_U_A2_list]
@@ -1995,6 +2001,7 @@ class JFAMachine(FactorAnalysisBase):
Accumulated statistics for D_A2(n_gaussians* feature_dimension, )
"""
+
_, _, latent_z = self.initialize_XYZ(
n_samples_per_class=n_samples_per_class
)
@@ -2030,6 +2037,7 @@ class JFAMachine(FactorAnalysisBase):
Accumulated statistics for D_A2(n_gaussians* feature_dimension, )
"""
+
acc_D_A1 = [acc[0] for acc in acc_D_A1_acc_D_A2_list]
acc_D_A2 = [acc[1] for acc in acc_D_A1_acc_D_A2_list]
@@ -2055,6 +2063,7 @@ class JFAMachine(FactorAnalysisBase):
z, y latent variables
"""
+
iterations = self.enroll_iterations
# We have only one class for enrollment
y = list(np.zeros(len(X), dtype=np.int32))
@@ -2118,6 +2127,7 @@ class JFAMachine(FactorAnalysisBase):
Returns self.
"""
+
if isinstance(X, dask.bag.Bag):
X, y = self._prepare_dask_input(X, y)
@@ -2240,9 +2250,16 @@ class JFAMachine(FactorAnalysisBase):
The linear scored
"""
+
latent_y = model[0]
latent_z = model[1]
+ template_data = data[0]
+
+ if len(data > 1):
+ for d in data[1:]:
+ template_data += d
+
x = self.estimate_x(data)
Ux = self._U @ x
diff --git a/src/bob/learn/em/ivector.py b/src/bob/learn/em/ivector.py
index 927e51e..3b0d5b3 100644
--- a/src/bob/learn/em/ivector.py
+++ b/src/bob/learn/em/ivector.py
@@ -287,8 +287,13 @@ class IVectorMachine(BaseEstimator):
)
self.sigma = copy.deepcopy(self.ubm.variances)
+ logger.info("Training I-Vector...")
for step in range(self.max_iterations):
+ logger.info(
+ f"IVector step {step+1:{len(str(self.max_iterations))}d}/{self.max_iterations}."
+ )
if chunky:
+ # Compute the IVectorStats of each chunk
stats = [
dask.delayed(e_step)(
machine=self,
@@ -300,27 +305,27 @@ class IVectorMachine(BaseEstimator):
# Workaround to prevent memory issues at compute with too many chunks.
# This adds pairs of stats together instead of sending all the stats to
# one worker.
- while (l := len(stats)) > 1:
+ while (length := len(stats)) > 1:
last = stats[-1]
stats = [
- dask.delayed(operator.add)(stats[i], stats[l // 2 + i])
- for i in range(l // 2)
+ dask.delayed(operator.add)(
+ stats[i], stats[length // 2 + i]
+ )
+ for i in range(length // 2)
]
- if l % 2 != 0:
+ if length % 2 != 0:
stats.append(last)
-
stats_sum = stats[0]
+
+ # Update the machine parameters with the aggregated stats
new_machine = dask.compute(
dask.delayed(m_step)(self, stats_sum)
)[0]
for attr in ["T", "sigma"]:
setattr(self, attr, getattr(new_machine, attr))
- else:
+ else: # Working directly on numpy array, not dask.Bags
stats = e_step(machine=self, data=X)
_ = m_step(self, stats)
- logger.info(
- f"IVector step {step+1:{len(str(self.max_iterations))}d}/{self.max_iterations}."
- )
logger.info(f"Reached {step+1} steps.")
return self
--
GitLab