From 5ce317d2f234b5eff6a2035253d2a123c1262b1c Mon Sep 17 00:00:00 2001
From: Amir MOHAMMADI <amir.mohammadi@idiap.ch>
Date: Mon, 11 Apr 2022 14:30:29 +0200
Subject: [PATCH] [factor_analysis] Still allow fit and init using gmm stats
---
bob/learn/em/factor_analysis.py | 41 +++++++++++++++++----------
bob/learn/em/test/test_jfa.py | 2 +-
bob/learn/em/test/test_jfa_trainer.py | 22 +++++++-------
3 files changed, 38 insertions(+), 27 deletions(-)
diff --git a/bob/learn/em/factor_analysis.py b/bob/learn/em/factor_analysis.py
index 18cdaa0..184c5b7 100644
--- a/bob/learn/em/factor_analysis.py
+++ b/bob/learn/em/factor_analysis.py
@@ -104,6 +104,9 @@ class FactorAnalysisBase(BaseEstimator):
self.relevance_factor = relevance_factor
+ if ubm is not None:
+ self.create_UVD()
+
@property
def feature_dimension(self):
"""Get the UBM Dimension"""
@@ -216,6 +219,9 @@ class FactorAnalysisBase(BaseEstimator):
if not hasattr(self, "_U") or not hasattr(self, "_D"):
self.create_UVD()
+ self.initialize_using_stats(ubm_projected_X, y)
+
+ def initialize_using_stats(self, ubm_projected_X, y):
# Accumulating 0th and 1st order statistics
# https://gitlab.idiap.ch/bob/bob.learn.em/-/blob/da92d0e5799d018f311f1bf5cdd5a80e19e142ca/bob/learn/em/cpp/ISVTrainer.cpp#L68
# 0th order stats
@@ -814,7 +820,7 @@ class FactorAnalysisBase(BaseEstimator):
np.zeros(
(
self.r_U,
- y.count(y_i),
+ np.sum(y == y_i),
)
)
)
@@ -1192,7 +1198,7 @@ class ISVMachine(FactorAnalysisBase):
ubm=None,
**gmm_kwargs,
):
- super(ISVMachine, self).__init__(
+ super().__init__(
r_U=r_U,
relevance_factor=relevance_factor,
em_iterations=em_iterations,
@@ -1213,7 +1219,7 @@ class ISVMachine(FactorAnalysisBase):
y: np.ndarray of shape(n_clients,)
Client labels.
"""
- return super(ISVMachine, self).initialize(X, y)
+ return super().initialize(X, y)
def e_step(self, X, y, n_acc, f_acc):
"""
@@ -1252,7 +1258,7 @@ class ISVMachine(FactorAnalysisBase):
self.update_U(acc_U_A1, acc_U_A2)
- def fit(self, X, y):
+ def fit_using_stats(self, X, y):
"""
Trains the U matrix (session variability matrix)
@@ -1270,10 +1276,10 @@ class ISVMachine(FactorAnalysisBase):
"""
- y = np.array(y).tolist() if not isinstance(y, list) else y
+ y = np.asarray(y)
# TODO: Point of MAP-REDUCE
- n_acc, f_acc = self.initialize(X, y)
+ n_acc, f_acc = self.initialize_using_stats(X, y)
for i in range(self.em_iterations):
logger.info("U Training: Iteration %d", i)
# TODO: Point of MAP-REDUCE
@@ -1412,20 +1418,25 @@ class JFAMachine(FactorAnalysisBase):
"""
def __init__(
- self, ubm, r_U, r_V, em_iterations=10, relevance_factor=4.0, seed=0
+ self,
+ ubm,
+ r_U,
+ r_V,
+ em_iterations=10,
+ relevance_factor=4.0,
+ seed=0,
+ **kwargs,
):
- super(JFAMachine, self).__init__(
- ubm,
+ super().__init__(
+ ubm=ubm,
r_U=r_U,
r_V=r_V,
relevance_factor=relevance_factor,
em_iterations=em_iterations,
seed=seed,
+ **kwargs,
)
- def initialize(self, X, y):
- return super(JFAMachine, self).initialize(X, y)
-
def e_step_v(self, X, y, n_acc, f_acc):
"""
ISV E-step for the V matrix.
@@ -1769,7 +1780,7 @@ class JFAMachine(FactorAnalysisBase):
"""
return self.enroll([self.ubm.transform(X)], iterations)
- def fit(self, X, y):
+ def fit_using_stats(self, X, y):
"""
Trains the U matrix (session variability matrix)
@@ -1795,10 +1806,10 @@ class JFAMachine(FactorAnalysisBase):
):
self.create_UVD()
- y = np.array(y).tolist() if not isinstance(y, list) else y
+ y = np.asarray(y)
# TODO: Point of MAP-REDUCE
- n_acc, f_acc = self.initialize(X, y)
+ n_acc, f_acc = self.initialize_using_stats(X, y)
# Updating V
for i in range(self.em_iterations):
diff --git a/bob/learn/em/test/test_jfa.py b/bob/learn/em/test/test_jfa.py
index 8e20b4a..dc15b71 100644
--- a/bob/learn/em/test/test_jfa.py
+++ b/bob/learn/em/test/test_jfa.py
@@ -65,7 +65,7 @@ def test_ISVMachine():
ubm.variances = np.array([[1, 2, 1], [2, 1, 2]], "float64")
# Creates a ISVMachine
- isv_machine = ISVMachine(ubm, r_U=2, em_iterations=10)
+ isv_machine = ISVMachine(ubm=ubm, r_U=2, em_iterations=10)
isv_machine.U = np.array(
[[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12]], "float64"
)
diff --git a/bob/learn/em/test/test_jfa_trainer.py b/bob/learn/em/test/test_jfa_trainer.py
index cc3d6ad..e38beec 100644
--- a/bob/learn/em/test/test_jfa_trainer.py
+++ b/bob/learn/em/test/test_jfa_trainer.py
@@ -126,7 +126,7 @@ def test_JFATrainAndEnrol():
it.U = copy.deepcopy(M_u)
it.V = copy.deepcopy(M_v)
it.D = copy.deepcopy(M_d)
- it.fit(TRAINING_STATS_X, TRAINING_STATS_y)
+ it.fit_using_stats(TRAINING_STATS_X, TRAINING_STATS_y)
v_ref = np.array(
[
@@ -225,7 +225,7 @@ def test_JFATrainAndEnrolWithNumpy():
it.U = copy.deepcopy(M_u)
it.V = copy.deepcopy(M_v)
it.D = copy.deepcopy(M_d)
- it.fit(TRAINING_STATS_X, TRAINING_STATS_y)
+ it.fit_using_stats(TRAINING_STATS_X, TRAINING_STATS_y)
v_ref = np.array(
[
@@ -337,14 +337,14 @@ def test_ISVTrainAndEnrol():
ubm.variances = UBM_VAR.reshape((2, 3))
it = ISVMachine(
- ubm,
+ ubm=ubm,
r_U=2,
relevance_factor=4.0,
em_iterations=10,
)
it.U = copy.deepcopy(M_u)
- it = it.fit(TRAINING_STATS_X, TRAINING_STATS_y)
+ it = it.fit_using_stats(TRAINING_STATS_X, TRAINING_STATS_y)
np.testing.assert_allclose(it.D, d_ref, rtol=eps, atol=1e-8)
np.testing.assert_allclose(it.U, u_ref, rtol=eps, atol=1e-8)
@@ -417,14 +417,14 @@ def test_ISVTrainAndEnrolWithNumpy():
ubm.variances = UBM_VAR.reshape((2, 3))
it = ISVMachine(
- ubm,
+ ubm=ubm,
r_U=2,
relevance_factor=4.0,
em_iterations=10,
)
it.U = copy.deepcopy(M_u)
- it = it.fit(TRAINING_STATS_X, TRAINING_STATS_y)
+ it = it.fit_using_stats(TRAINING_STATS_X, TRAINING_STATS_y)
np.testing.assert_allclose(it.D, d_ref, rtol=eps, atol=1e-8)
np.testing.assert_allclose(it.U, u_ref, rtol=eps, atol=1e-8)
@@ -466,13 +466,13 @@ def test_JFATrainInitialize():
it = JFAMachine(ubm, 2, 2, em_iterations=10)
# first round
- it.initialize(TRAINING_STATS_X, TRAINING_STATS_y)
+ it.initialize_using_stats(TRAINING_STATS_X, TRAINING_STATS_y)
u1 = it.U
v1 = it.V
d1 = it.D
# second round
- it.initialize(TRAINING_STATS_X, TRAINING_STATS_y)
+ it.initialize_using_stats(TRAINING_STATS_X, TRAINING_STATS_y)
u2 = it.U
v2 = it.V
d2 = it.D
@@ -493,15 +493,15 @@ def test_ISVTrainInitialize():
ubm.variances = UBM_VAR.reshape((2, 3))
# ISV
- it = ISVMachine(ubm, 2, em_iterations=10)
+ it = ISVMachine(2, em_iterations=10, ubm=ubm)
# it.rng = rng
- it.initialize(TRAINING_STATS_X, TRAINING_STATS_y)
+ it.initialize_using_stats(TRAINING_STATS_X, TRAINING_STATS_y)
u1 = copy.deepcopy(it.U)
d1 = copy.deepcopy(it.D)
# second round
- it.initialize(TRAINING_STATS_X, TRAINING_STATS_y)
+ it.initialize_using_stats(TRAINING_STATS_X, TRAINING_STATS_y)
u2 = it.U
d2 = it.D
--
GitLab