From c3ad33f3a9acf5b8004ff27938eb178a43ee7a11 Mon Sep 17 00:00:00 2001
From: Amir MOHAMMADI <amir.mohammadi@idiap.ch>
Date: Tue, 12 Apr 2022 14:18:41 +0200
Subject: [PATCH] [factor_analysis] test fit without prior
---
bob/learn/em/factor_analysis.py | 25 +--
bob/learn/em/test/test_factor_analysis.py | 178 ++++++++++++----------
2 files changed, 103 insertions(+), 100 deletions(-)
diff --git a/bob/learn/em/factor_analysis.py b/bob/learn/em/factor_analysis.py
index 3a6b1de..1f2e8c3 100644
--- a/bob/learn/em/factor_analysis.py
+++ b/bob/learn/em/factor_analysis.py
@@ -4,7 +4,6 @@
import logging
-import dask
import numpy as np
from sklearn.base import BaseEstimator
@@ -183,7 +182,7 @@ class FactorAnalysisBase(BaseEstimator):
return len(unique_labels(y))
- def initialize(self, X, y):
+ def initialize(self, X):
"""
Accumulating 0th and 1st order statistics. Trains the UBM if needed.
@@ -211,17 +210,12 @@ class FactorAnalysisBase(BaseEstimator):
# Train the UBM if not already trained
if self.ubm._means is None:
logger.info(f"FA: Training the UBM with {self.ubm}.")
- self.ubm.fit(np.vstack(X)) # GMMMachine.fit takes non-labeled data
-
- logger.info("FA: Projection of training data on the UBM.")
- ubm_projected_X = [dask.delayed(self.ubm.transform(xx)) for xx in X]
+ self.ubm.fit(X) # GMMMachine.fit takes non-labeled data
# Initializing the state matrix
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
@@ -1159,6 +1153,7 @@ class FactorAnalysisBase(BaseEstimator):
return self.score_using_stats(model, self.ubm.transform(data))
def fit(self, X, y):
+ self.initialize(X)
stats = [self.ubm.transform(xx) for xx in X]
return self.fit_using_stats(stats, y)
@@ -1214,20 +1209,6 @@ class ISVMachine(FactorAnalysisBase):
**kwargs,
)
- def initialize(self, X, y):
- """Initializes the ISV parameters and trains a UBM with `X` if needed.
-
- If no UBM has been defined on init, it is trained with a new GMMMachine.
-
- Parameters
- ----------
- X: np.ndarray of shape(n_clients, n_samples, n_features)
- Input data for each client.
- y: np.ndarray of shape(n_clients,)
- Client labels.
- """
- return super().initialize(X, y)
-
def e_step(self, X, y, n_acc, f_acc):
"""
E-step of the EM algorithm
diff --git a/bob/learn/em/test/test_factor_analysis.py b/bob/learn/em/test/test_factor_analysis.py
index f5a85fd..ca1bd2a 100644
--- a/bob/learn/em/test/test_factor_analysis.py
+++ b/bob/learn/em/test/test_factor_analysis.py
@@ -443,7 +443,7 @@ def test_ISVMachine():
np.testing.assert_allclose(score, score_ref, atol=eps)
-def test_ISV_fit():
+def test_ISV_JFA_fit():
np.random.seed(10)
data_class1 = np.random.normal(0, 0.5, (10, 3))
data_class2 = np.random.normal(-0.2, 0.2, (10, 3))
@@ -455,84 +455,106 @@ def test_ISV_fit():
prior_gmm.means = np.vstack(
(np.random.normal(0, 0.5, (1, 3)), np.random.normal(1, 0.5, (1, 3)))
)
-
# All nice and round diagonal covariance
prior_gmm.variances = np.ones((2, 3)) * 0.5
prior_gmm.weights = np.array([0.3, 0.7])
- # Finally doing the ISV training
- isv = ISVMachine(
- 2,
- ubm=prior_gmm,
- relevance_factor=4,
- em_iterations=50,
- )
- isv.fit(data, labels)
-
- # Printing the session offset w.r.t each Gaussian component
- U_ref = [
- [-2.86662863e-02, 4.45865461e-04],
- [-4.51712419e-03, 7.02577809e-05],
- [7.91269855e-02, -1.23071365e-03],
- [3.27129434e-02, -5.08805760e-04],
- [9.17898003e-02, -1.42766668e-03],
- [1.29496881e-01, -2.01414952e-03],
- ]
- # TODO(tiago): The reference used to be the values below but are different now
- # U_ref = [
- # [-0.01, -0.027],
- # [-0.002, -0.004],
- # [0.028, 0.074],
- # [0.012, 0.03],
- # [0.033, 0.085],
- # [0.046, 0.12],
- # ]
- np.testing.assert_allclose(isv.U, U_ref, atol=1e-7)
-
-
-def test_JFA_fit():
- np.random.seed(10)
- data_class1 = np.random.normal(0, 0.5, (10, 3))
- data_class2 = np.random.normal(-0.2, 0.2, (10, 3))
- data = np.concatenate([data_class1, data_class2], axis=0)
- labels = [0] * 10 + [1] * 10
-
- # Creating a fake prior with 2 gaussians
- prior_gmm = GMMMachine(2)
- prior_gmm.means = np.vstack(
- (np.random.normal(0, 0.5, (1, 3)), np.random.normal(1, 0.5, (1, 3)))
- )
-
- # All nice and round diagonal covariance
- prior_gmm.variances = np.ones((2, 3)) * 0.5
- prior_gmm.weights = np.array([0.3, 0.7])
-
- # Finally doing the JFA training
- jfa = JFAMachine(
- 2,
- 2,
- ubm=prior_gmm,
- relevance_factor=4,
- em_iterations=50,
- )
- jfa.fit(data, labels)
-
- # Printing the session offset w.r.t each Gaussian component
- V_ref = [
- [-0.00459188, 0.00463761],
- [-0.06622346, 0.06688288],
- [0.41800691, -0.4221692],
- [0.40218688, -0.40619164],
- [0.61849675, -0.6246554],
- [0.57576069, -0.5814938],
- ]
- # TODO(tiago): The reference used to be the values below but are different now
- # V_ref = [
- # [0.003, -0.006],
- # [0.041, -0.084],
- # [-0.261, 0.53],
- # [-0.252, 0.51],
- # [-0.387, 0.785],
- # [-0.36, 0.73],
- # ]
- np.testing.assert_allclose(jfa.V, V_ref, atol=1e-7)
+ for prior, machine_type, ref in [
+ (
+ None,
+ "isv",
+ [
+ [0.02619036, 0.07607595],
+ [-0.02570657, -0.07451667],
+ [-0.0430513, -0.12514552],
+ [-0.09729266, -0.28582205],
+ [-0.01035388, -0.03041718],
+ [0.0733034, 0.21534741],
+ ],
+ ),
+ (
+ prior_gmm,
+ "isv",
+ [
+ [-0.02361267, 0.0157274],
+ [-0.00372588, 0.00248165],
+ [0.06517179, -0.04340818],
+ [0.02694231, -0.01794513],
+ [0.07560949, -0.05036029],
+ [0.10668997, -0.07106169],
+ ],
+ # TODO(tiago): The reference used to be the values below but are different now
+ # [
+ # [-0.01, -0.027],
+ # [-0.002, -0.004],
+ # [0.028, 0.074],
+ # [0.012, 0.03],
+ # [0.033, 0.085],
+ # [0.046, 0.12],
+ # ]
+ ),
+ (
+ None,
+ "jfa",
+ [
+ [-1.72285693e-01, 1.47171193e-01],
+ [-1.08402014e-01, 9.25999920e-02],
+ [1.55349449e-02, -1.32703786e-02],
+ [2.13389657e-04, -1.82283334e-04],
+ [1.84127661e-05, -1.57286929e-05],
+ [-1.90492196e-04, 1.62723691e-04],
+ ],
+ ),
+ (
+ prior_gmm,
+ "jfa",
+ [
+ [6.54547662e-03, 1.98699266e-04],
+ [9.48510389e-02, 2.87936736e-03],
+ [-5.98879972e-01, -1.81800375e-02],
+ [-5.76350228e-01, -1.74961082e-02],
+ [-8.86302168e-01, -2.69052355e-02],
+ [-8.25011907e-01, -2.50446636e-02],
+ ],
+ # TODO(tiago): The reference used to be the values below but are different now
+ # [[ 0.003 -0.006]
+ # [ 0.041 -0.084]
+ # [-0.261 0.53 ]
+ # [-0.252 0.51 ]
+ # [-0.387 0.785]
+ # [-0.36 0.73 ]]
+ ),
+ ]:
+ ref = np.asarray(ref)
+ ubm_kwargs = dict(n_gaussians=2) if prior is None else None
+
+ # Doing the training
+ if machine_type == "isv":
+ machine = ISVMachine(
+ 2,
+ ubm=prior,
+ relevance_factor=4,
+ em_iterations=50,
+ ubm_kwargs=ubm_kwargs,
+ seed=10,
+ )
+ test_attr = "U"
+ else:
+ machine = JFAMachine(
+ 2,
+ 2,
+ ubm=prior,
+ relevance_factor=4,
+ em_iterations=50,
+ ubm_kwargs=ubm_kwargs,
+ seed=10,
+ )
+ test_attr = "V"
+ machine.fit(data, labels)
+ arr = getattr(machine, test_attr)
+ np.testing.assert_allclose(
+ arr,
+ ref,
+ atol=1e-7,
+ err_msg=f"Test failed with prior={prior} and machine_type={machine_type}",
+ )
--
GitLab