Skip to content
Snippets Groups Projects

Compare revisions

Changes are shown as if the source revision was being merged into the target revision. Learn more about comparing revisions.

Source

Select target project
No results found

Target

Select target project
  • bob/bob.learn.em
1 result
Show changes
Hide whitespace changes
Inline Side-by-side
Showing
with 1620 additions and 1036 deletions
bob/learn/em/ml_gmm_trainer.cpp
View file @ 61a31463
...@@ -17,7 +17,7 @@ static inline bool f(PyObject* o){return o != 0 && PyObject_IsTrue(o) > 0;} /* ...@@ -17,7 +17,7 @@ static inline bool f(PyObject* o){return o != 0 && PyObject_IsTrue(o) > 0;} /*
static auto ML_GMMTrainer_doc = bob::extension::ClassDoc( static auto ML_GMMTrainer_doc = bob::extension::ClassDoc(
BOB_EXT_MODULE_PREFIX ".ML_GMMTrainer", BOB_EXT_MODULE_PREFIX ".ML_GMMTrainer",
"This class implements the maximum likelihood M-step of the expectation-maximisation algorithm for a GMM Machine." "This class implements the maximum likelihood M-step (:ref:`MLE <mle>`) of the expectation-maximisation algorithm for a GMM Machine."
).add_constructor( ).add_constructor(
bob::extension::FunctionDoc( bob::extension::FunctionDoc(
"__init__", "__init__",
......
bob/learn/em/test/test_kmeans_trainer.py
View file @ 61a31463
...@@ -15,159 +15,179 @@ from bob.io.base.test_utils import datafile ...@@ -15,159 +15,179 @@ from bob.io.base.test_utils import datafile
from bob.learn.em import KMeansMachine, KMeansTrainer from bob.learn.em import KMeansMachine, KMeansTrainer
def equals(x, y, epsilon): def equals(x, y, epsilon):
return (abs(x - y) < epsilon).all() return (abs(x - y) < epsilon).all()
def kmeans_plus_plus(machine, data, seed): def kmeans_plus_plus(machine, data, seed):
"""Python implementation of K-Means++ (initialization)""" """Python implementation of K-Means++ (initialization)"""
n_data = data.shape[0] n_data = data.shape[0]
rng = bob.core.random.mt19937(seed) rng = bob.core.random.mt19937(seed)
u = bob.core.random.uniform('int32', 0, n_data-1) u = bob.core.random.uniform('int32', 0, n_data - 1)
index = u(rng) index = u(rng)
machine.set_mean(0, data[index,:]) machine.set_mean(0, data[index, :])
weights = numpy.zeros(shape=(n_data,), dtype=numpy.float64) weights = numpy.zeros(shape=(n_data,), dtype=numpy.float64)
for m in range(1,machine.dim_c): for m in range(1, machine.dim_c):
for s in range(n_data): for s in range(n_data):
s_cur = data[s,:] s_cur = data[s, :]
w_cur = machine.get_distance_from_mean(s_cur, 0) w_cur = machine.get_distance_from_mean(s_cur, 0)
for i in range(m): for i in range(m):
w_cur = min(machine.get_distance_from_mean(s_cur, i), w_cur) w_cur = min(machine.get_distance_from_mean(s_cur, i), w_cur)
weights[s] = w_cur weights[s] = w_cur
weights *= weights weights *= weights
weights /= numpy.sum(weights) weights /= numpy.sum(weights)
d = bob.core.random.discrete('int32', weights) d = bob.core.random.discrete('int32', weights)
index = d(rng) index = d(rng)
machine.set_mean(m, data[index,:]) machine.set_mean(m, data[index, :])
def NormalizeStdArray(path): def NormalizeStdArray(path):
array = bob.io.base.load(path).astype('float64') array = bob.io.base.load(path).astype('float64')
std = array.std(axis=0) std = array.std(axis=0)
return (array/std, std) return (array / std, std)
def multiplyVectorsByFactors(matrix, vector): def multiplyVectorsByFactors(matrix, vector):
for i in range(0, matrix.shape[0]): for i in range(0, matrix.shape[0]):
for j in range(0, matrix.shape[1]): for j in range(0, matrix.shape[1]):
matrix[i, j] *= vector[j] matrix[i, j] *= vector[j]
def flipRows(array): def flipRows(array):
if len(array.shape) == 2: if len(array.shape) == 2:
return numpy.array([numpy.array(array[1, :]), numpy.array(array[0, :])], 'float64') return numpy.array([numpy.array(array[1, :]), numpy.array(array[0, :])], 'float64')
elif len(array.shape) == 1: elif len(array.shape) == 1:
return numpy.array([array[1], array[0]], 'float64') return numpy.array([array[1], array[0]], 'float64')
else: else:
raise Exception('Input type not supportd by flipRows') raise Exception('Input type not supportd by flipRows')
if hasattr(KMeansTrainer, 'KMEANS_PLUS_PLUS'): if hasattr(KMeansTrainer, 'KMEANS_PLUS_PLUS'):
def test_kmeans_plus_plus(): def test_kmeans_plus_plus():
# Tests the K-Means++ initialization
dim_c = 5
dim_d = 7
n_samples = 150
data = numpy.random.randn(n_samples, dim_d)
seed = 0
# C++ implementation
machine = KMeansMachine(dim_c, dim_d)
trainer = KMeansTrainer()
trainer.rng = bob.core.random.mt19937(seed)
trainer.initialization_method = 'KMEANS_PLUS_PLUS'
trainer.initialize(machine, data)
# Python implementation
py_machine = KMeansMachine(dim_c, dim_d)
kmeans_plus_plus(py_machine, data, seed)
assert equals(machine.means, py_machine.means, 1e-8)
# Tests the K-Means++ initialization
dim_c = 5
dim_d = 7
n_samples = 150
data = numpy.random.randn(n_samples,dim_d)
seed = 0
# C++ implementation def test_kmeans_noduplicate():
# Data/dimensions
dim_c = 2
dim_d = 3
seed = 0
data = numpy.array([[1, 2, 3], [1, 2, 3], [1, 2, 3], [4, 5, 6.]])
# Defines machine and trainer
machine = KMeansMachine(dim_c, dim_d) machine = KMeansMachine(dim_c, dim_d)
trainer = KMeansTrainer() trainer = KMeansTrainer()
trainer.rng = bob.core.random.mt19937(seed) rng = bob.core.random.mt19937(seed)
trainer.initialization_method = 'KMEANS_PLUS_PLUS' trainer.initialization_method = 'RANDOM_NO_DUPLICATE'
trainer.initialize(machine, data) trainer.initialize(machine, data, rng)
# Makes sure that the two initial mean vectors selected are different
# Python implementation assert equals(machine.get_mean(0), machine.get_mean(1), 1e-8) == False
py_machine = KMeansMachine(dim_c, dim_d)
kmeans_plus_plus(py_machine, data, seed)
assert equals(machine.means, py_machine.means, 1e-8)
def test_kmeans_noduplicate():
# Data/dimensions
dim_c = 2
dim_d = 3
seed = 0
data = numpy.array([[1,2,3],[1,2,3],[1,2,3],[4,5,6.]])
# Defines machine and trainer
machine = KMeansMachine(dim_c, dim_d)
trainer = KMeansTrainer()
rng = bob.core.random.mt19937(seed)
trainer.initialization_method = 'RANDOM_NO_DUPLICATE'
trainer.initialize(machine, data, rng)
# Makes sure that the two initial mean vectors selected are different
assert equals(machine.get_mean(0), machine.get_mean(1), 1e-8) == False
def test_kmeans_a(): def test_kmeans_a():
# Trains a KMeansMachine
# This files contains draws from two 1D Gaussian distributions:
# * 100 samples from N(-10,1)
# * 100 samples from N(10,1)
data = bob.io.base.load(datafile("samplesFrom2G_f64.hdf5", __name__, path="../data/"))
# Trains a KMeansMachine machine = KMeansMachine(2, 1)
# This files contains draws from two 1D Gaussian distributions:
# * 100 samples from N(-10,1)
# * 100 samples from N(10,1)
data = bob.io.base.load(datafile("samplesFrom2G_f64.hdf5", __name__, path="../data/"))
machine = KMeansMachine(2, 1)
trainer = KMeansTrainer()
#trainer.train(machine, data)
bob.learn.em.train(trainer,machine,data)
[variances, weights] = machine.get_variances_and_weights_for_each_cluster(data) trainer = KMeansTrainer()
variances_b = numpy.ndarray(shape=(2,1), dtype=numpy.float64) # trainer.train(machine, data)
weights_b = numpy.ndarray(shape=(2,), dtype=numpy.float64) bob.learn.em.train(trainer, machine, data)
machine.__get_variances_and_weights_for_each_cluster_init__(variances_b, weights_b)
machine.__get_variances_and_weights_for_each_cluster_acc__(data, variances_b, weights_b) [variances, weights] = machine.get_variances_and_weights_for_each_cluster(data)
machine.__get_variances_and_weights_for_each_cluster_fin__(variances_b, weights_b) variances_b = numpy.ndarray(shape=(2, 1), dtype=numpy.float64)
m1 = machine.get_mean(0) weights_b = numpy.ndarray(shape=(2,), dtype=numpy.float64)
m2 = machine.get_mean(1) machine.__get_variances_and_weights_for_each_cluster_init__(variances_b, weights_b)
machine.__get_variances_and_weights_for_each_cluster_acc__(data, variances_b, weights_b)
machine.__get_variances_and_weights_for_each_cluster_fin__(variances_b, weights_b)
m1 = machine.get_mean(0)
m2 = machine.get_mean(1)
## Check means [-10,10] / variances [1,1] / weights [0.5,0.5]
if (m1 < m2):
means = numpy.array(([m1[0], m2[0]]), 'float64')
else:
means = numpy.array(([m2[0], m1[0]]), 'float64')
assert equals(means, numpy.array([-10., 10.]), 2e-1)
assert equals(variances, numpy.array([1., 1.]), 2e-1)
assert equals(weights, numpy.array([0.5, 0.5]), 1e-3)
assert equals(variances, variances_b, 1e-8)
assert equals(weights, weights_b, 1e-8)
## Check means [-10,10] / variances [1,1] / weights [0.5,0.5]
if(m1<m2): means=numpy.array(([m1[0],m2[0]]), 'float64')
else: means=numpy.array(([m2[0],m1[0]]), 'float64')
assert equals(means, numpy.array([-10.,10.]), 2e-1)
assert equals(variances, numpy.array([1.,1.]), 2e-1)
assert equals(weights, numpy.array([0.5,0.5]), 1e-3)
assert equals(variances, variances_b, 1e-8) def test_kmeans_b():
assert equals(weights, weights_b, 1e-8) # Trains a KMeansMachine
(arStd, std) = NormalizeStdArray(datafile("faithful.torch3.hdf5", __name__, path="../data/"))
machine = KMeansMachine(2, 2)
trainer = KMeansTrainer()
# trainer.seed = 1337
bob.learn.em.train(trainer, machine, arStd, convergence_threshold=0.001)
def test_kmeans_b(): [variances, weights] = machine.get_variances_and_weights_for_each_cluster(arStd)
# Trains a KMeansMachine means = numpy.array(machine.means)
(arStd,std) = NormalizeStdArray(datafile("faithful.torch3.hdf5", __name__, path="../data/")) variances = numpy.array(variances)
machine = KMeansMachine(2, 2) multiplyVectorsByFactors(means, std)
multiplyVectorsByFactors(variances, std ** 2)
trainer = KMeansTrainer() gmmWeights = bob.io.base.load(datafile('gmm.init_weights.hdf5', __name__, path="../data/"))
#trainer.seed = 1337 gmmMeans = bob.io.base.load(datafile('gmm.init_means.hdf5', __name__, path="../data/"))
bob.learn.em.train(trainer,machine, arStd, convergence_threshold=0.001) gmmVariances = bob.io.base.load(datafile('gmm.init_variances.hdf5', __name__, path="../data/"))
[variances, weights] = machine.get_variances_and_weights_for_each_cluster(arStd) if (means[0, 0] < means[1, 0]):
means = flipRows(means)
variances = flipRows(variances)
weights = flipRows(weights)
means = numpy.array(machine.means) assert equals(means, gmmMeans, 1e-3)
variances = numpy.array(variances) assert equals(weights, gmmWeights, 1e-3)
assert equals(variances, gmmVariances, 1e-3)
multiplyVectorsByFactors(means, std) # Check that there is no duplicate means during initialization
multiplyVectorsByFactors(variances, std ** 2) machine = KMeansMachine(2, 1)
trainer = KMeansTrainer()
trainer.initialization_method = 'RANDOM_NO_DUPLICATE'
data = numpy.array([[1.], [1.], [1.], [1.], [1.], [1.], [2.], [3.]])
bob.learn.em.train(trainer, machine, data)
assert (numpy.isnan(machine.means).any()) == False
gmmWeights = bob.io.base.load(datafile('gmm.init_weights.hdf5', __name__, path="../data/"))
gmmMeans = bob.io.base.load(datafile('gmm.init_means.hdf5', __name__, path="../data/"))
gmmVariances = bob.io.base.load(datafile('gmm.init_variances.hdf5', __name__, path="../data/"))
if (means[0, 0] < means[1, 0]): def test_trainer_execption():
means = flipRows(means) from nose.tools import assert_raises
variances = flipRows(variances)
weights = flipRows(weights)
assert equals(means, gmmMeans, 1e-3) # Testing Inf
assert equals(weights, gmmWeights, 1e-3) machine = KMeansMachine(2, 2)
assert equals(variances, gmmVariances, 1e-3) data = numpy.array([[1.0, 2.0], [2, 3.], [1, 1.], [2, 5.], [numpy.inf, 1.0]])
trainer = KMeansTrainer()
assert_raises(ValueError, bob.learn.em.train, trainer, machine, data, 10)
# Check that there is no duplicate means during initialization # Testing Nan
machine = KMeansMachine(2, 1) machine = KMeansMachine(2, 2)
trainer = KMeansTrainer() data = numpy.array([[1.0, 2.0], [2, 3.], [1, numpy.nan], [2, 5.], [2.0, 1.0]])
trainer.initialization_method = 'RANDOM_NO_DUPLICATE' trainer = KMeansTrainer()
data = numpy.array([[1.], [1.], [1.], [1.], [1.], [1.], [2.], [3.]]) assert_raises(ValueError, bob.learn.em.train, trainer, machine, data, 10)
bob.learn.em.train(trainer, machine, data)
assert (numpy.isnan(machine.means).any()) == False
bob/learn/em/train.py
View file @ 61a31463
...@@ -7,112 +7,125 @@ ...@@ -7,112 +7,125 @@
import numpy import numpy
import bob.learn.em import bob.learn.em
import logging import logging
logger = logging.getLogger('bob.learn.em') logger = logging.getLogger('bob.learn.em')
def train(trainer, machine, data, max_iterations = 50, convergence_threshold=None, initialize=True, rng=None):
def train(trainer, machine, data, max_iterations=50, convergence_threshold=None, initialize=True, rng=None,
""" check_inputs=True):
Trains a machine given a trainer and the proper data """
Trains a machine given a trainer and the proper data
**Parameters**:
trainer : one of :py:class:`KMeansTrainer`, :py:class:`MAP_GMMTrainer`, :py:class:`ML_GMMTrainer`, :py:class:`ISVTrainer`, :py:class:`IVectorTrainer`, :py:class:`PLDATrainer`, :py:class:`EMPCATrainer` **Parameters**:
A trainer mechanism trainer : one of :py:class:`KMeansTrainer`, :py:class:`MAP_GMMTrainer`, :py:class:`ML_GMMTrainer`, :py:class:`ISVTrainer`, :py:class:`IVectorTrainer`, :py:class:`PLDATrainer`, :py:class:`EMPCATrainer`
machine : one of :py:class:`KMeansMachine`, :py:class:`GMMMachine`, :py:class:`ISVBase`, :py:class:`IVectorMachine`, :py:class:`PLDAMachine`, :py:class:`bob.learn.linear.Machine` A trainer mechanism
A container machine machine : one of :py:class:`KMeansMachine`, :py:class:`GMMMachine`, :py:class:`ISVBase`, :py:class:`IVectorMachine`, :py:class:`PLDAMachine`, :py:class:`bob.learn.linear.Machine`
data : array_like <float, 2D> A container machine
The data to be trained data : array_like <float, 2D>
max_iterations : int The data to be trained
The maximum number of iterations to train a machine max_iterations : int
convergence_threshold : float The maximum number of iterations to train a machine
The convergence threshold to train a machine. If None, the training procedure will stop with the iterations criteria convergence_threshold : float
initialize : bool The convergence threshold to train a machine. If None, the training procedure will stop with the iterations criteria
If True, runs the initialization procedure initialize : bool
rng : :py:class:`bob.core.random.mt19937` If True, runs the initialization procedure
The Mersenne Twister mt19937 random generator used for the initialization of subspaces/arrays before the EM loop rng : :py:class:`bob.core.random.mt19937`
""" The Mersenne Twister mt19937 random generator used for the initialization of subspaces/arrays before the EM loop
#Initialization check_inputs:
if initialize: Shallow checks in the inputs. Check for inf and NaN
if rng is not None: """
trainer.initialize(machine, data, rng)
else: if check_inputs and type(data) is numpy.ndarray:
trainer.initialize(machine, data)
if numpy.isinf(numpy.sum(data)):
trainer.e_step(machine, data) raise ValueError("Please, check your inputs; numpy.inf detected in `data` ")
average_output = 0
average_output_previous = 0 if numpy.isnan(numpy.sum(data)):
raise ValueError("Please, check your inputs; numpy.nan detected in `data` ")
if hasattr(trainer,"compute_likelihood"):
average_output = trainer.compute_likelihood(machine) # Initialization
if initialize:
for i in range(max_iterations): if rng is not None:
logger.info("Iteration = %d/%d", i, max_iterations) trainer.initialize(machine, data, rng)
average_output_previous = average_output else:
trainer.m_step(machine, data) trainer.initialize(machine, data)
trainer.e_step(machine, data) trainer.e_step(machine, data)
average_output = 0
if hasattr(trainer,"compute_likelihood"): average_output_previous = 0
average_output = trainer.compute_likelihood(machine)
if hasattr(trainer, "compute_likelihood"):
if type(machine) is bob.learn.em.KMeansMachine: average_output = trainer.compute_likelihood(machine)
logger.info("average euclidean distance = %f", average_output)
else: for i in range(max_iterations):
logger.info("log likelihood = %f", average_output) logger.info("Iteration = %d/%d", i, max_iterations)
average_output_previous = average_output
convergence_value = abs((average_output_previous - average_output)/average_output_previous) trainer.m_step(machine, data)
logger.info("convergence value = %f",convergence_value) trainer.e_step(machine, data)
#Terminates if converged (and likelihood computation is set) if hasattr(trainer, "compute_likelihood"):
if convergence_threshold!=None and convergence_value <= convergence_threshold: average_output = trainer.compute_likelihood(machine)
break
if hasattr(trainer,"finalize"): if type(machine) is bob.learn.em.KMeansMachine:
trainer.finalize(machine, data) logger.info("average euclidean distance = %f", average_output)
else:
logger.info("log likelihood = %f", average_output)
convergence_value = abs((average_output_previous - average_output) / average_output_previous)
logger.info("convergence value = %f", convergence_value)
# Terminates if converged (and likelihood computation is set)
if convergence_threshold != None and convergence_value <= convergence_threshold:
break
if hasattr(trainer, "finalize"):
trainer.finalize(machine, data)
def train_jfa(trainer, jfa_base, data, max_iterations=10, initialize=True, rng=None): def train_jfa(trainer, jfa_base, data, max_iterations=10, initialize=True, rng=None):
""" """
Trains a :py:class:`bob.learn.em.JFABase` given a :py:class:`bob.learn.em.JFATrainer` and the proper data Trains a :py:class:`bob.learn.em.JFABase` given a :py:class:`bob.learn.em.JFATrainer` and the proper data
**Parameters**: **Parameters**:
trainer : :py:class:`bob.learn.em.JFATrainer` trainer : :py:class:`bob.learn.em.JFATrainer`
A JFA trainer mechanism A JFA trainer mechanism
jfa_base : :py:class:`bob.learn.em.JFABase` jfa_base : :py:class:`bob.learn.em.JFABase`
A container machine A container machine
data : [[:py:class:`bob.learn.em.GMMStats`]] data : [[:py:class:`bob.learn.em.GMMStats`]]
The data to be trained The data to be trained
max_iterations : int max_iterations : int
The maximum number of iterations to train a machine The maximum number of iterations to train a machine
initialize : bool initialize : bool
If True, runs the initialization procedure If True, runs the initialization procedure
rng : :py:class:`bob.core.random.mt19937` rng : :py:class:`bob.core.random.mt19937`
The Mersenne Twister mt19937 random generator used for the initialization of subspaces/arrays before the EM loops The Mersenne Twister mt19937 random generator used for the initialization of subspaces/arrays before the EM loops
""" """
if initialize: if initialize:
if rng is not None: if rng is not None:
trainer.initialize(jfa_base, data, rng) trainer.initialize(jfa_base, data, rng)
else: else:
trainer.initialize(jfa_base, data) trainer.initialize(jfa_base, data)
#V Subspace # V Subspace
logger.info("V subspace estimation...") logger.info("V subspace estimation...")
for i in range(max_iterations): for i in range(max_iterations):
logger.info("Iteration = %d/%d", i, max_iterations) logger.info("Iteration = %d/%d", i, max_iterations)
trainer.e_step_v(jfa_base, data) trainer.e_step_v(jfa_base, data)
trainer.m_step_v(jfa_base, data) trainer.m_step_v(jfa_base, data)
trainer.finalize_v(jfa_base, data) trainer.finalize_v(jfa_base, data)
#U subspace # U subspace
logger.info("U subspace estimation...") logger.info("U subspace estimation...")
for i in range(max_iterations): for i in range(max_iterations):
logger.info("Iteration = %d/%d", i, max_iterations) logger.info("Iteration = %d/%d", i, max_iterations)
trainer.e_step_u(jfa_base, data) trainer.e_step_u(jfa_base, data)
trainer.m_step_u(jfa_base, data) trainer.m_step_u(jfa_base, data)
trainer.finalize_u(jfa_base, data) trainer.finalize_u(jfa_base, data)
# D subspace # D subspace
logger.info("D subspace estimation...") logger.info("D subspace estimation...")
for i in range(max_iterations): for i in range(max_iterations):
logger.info("Iteration = %d/%d", i, max_iterations) logger.info("Iteration = %d/%d", i, max_iterations)
trainer.e_step_d(jfa_base, data) trainer.e_step_d(jfa_base, data)
trainer.m_step_d(jfa_base, data) trainer.m_step_d(jfa_base, data)
trainer.finalize_d(jfa_base, data) trainer.finalize_d(jfa_base, data)
bob/learn/em/ztnorm.cpp
View file @ 61a31463
...@@ -12,7 +12,7 @@ ...@@ -12,7 +12,7 @@
/*** zt_norm ***/ /*** zt_norm ***/
bob::extension::FunctionDoc zt_norm = bob::extension::FunctionDoc( bob::extension::FunctionDoc zt_norm = bob::extension::FunctionDoc(
"ztnorm", "ztnorm",
"Normalise raw scores with ZT-Norm." "Normalise raw scores with :ref:`ZT-Norm <ztnorm>`."
"Assume that znorm and tnorm have no common subject id.", "Assume that znorm and tnorm have no common subject id.",
0, 0,
true true
...@@ -72,7 +72,7 @@ PyObject* PyBobLearnEM_ztNorm(PyObject*, PyObject* args, PyObject* kwargs) { ...@@ -72,7 +72,7 @@ PyObject* PyBobLearnEM_ztNorm(PyObject*, PyObject* args, PyObject* kwargs) {
/*** t_norm ***/ /*** t_norm ***/
bob::extension::FunctionDoc t_norm = bob::extension::FunctionDoc( bob::extension::FunctionDoc t_norm = bob::extension::FunctionDoc(
"tnorm", "tnorm",
"Normalise raw scores with T-Norm", "Normalise raw scores with :ref:`T-Norm <tnorm>`",
0, 0,
true true
) )
...@@ -109,7 +109,7 @@ PyObject* PyBobLearnEM_tNorm(PyObject*, PyObject* args, PyObject* kwargs) { ...@@ -109,7 +109,7 @@ PyObject* PyBobLearnEM_tNorm(PyObject*, PyObject* args, PyObject* kwargs) {
/*** z_norm ***/ /*** z_norm ***/
bob::extension::FunctionDoc z_norm = bob::extension::FunctionDoc( bob::extension::FunctionDoc z_norm = bob::extension::FunctionDoc(
"znorm", "znorm",
"Normalise raw scores with Z-Norm", "Normalise raw scores with :ref:`Z-Norm <znorm>`",
0, 0,
true true
) )
......
bootstrap-buildout.py deleted 100644 → 0
View file @ 68cfd39a
##############################################################################
#
# Copyright (c) 2006 Zope Foundation and Contributors.
# All Rights Reserved.
#
# This software is subject to the provisions of the Zope Public License,
# Version 2.1 (ZPL). A copy of the ZPL should accompany this distribution.
# THIS SOFTWARE IS PROVIDED "AS IS" AND ANY AND ALL EXPRESS OR IMPLIED
# WARRANTIES ARE DISCLAIMED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
# WARRANTIES OF TITLE, MERCHANTABILITY, AGAINST INFRINGEMENT, AND FITNESS
# FOR A PARTICULAR PURPOSE.
#
##############################################################################
"""Bootstrap a buildout-based project
Simply run this script in a directory containing a buildout.cfg.
The script accepts buildout command-line options, so you can
use the -c option to specify an alternate configuration file.
"""
import os
import shutil
import sys
import tempfile
from optparse import OptionParser
__version__ = '2015-07-01'
# See zc.buildout's changelog if this version is up to date.
tmpeggs = tempfile.mkdtemp(prefix='bootstrap-')
usage = '''\
[DESIRED PYTHON FOR BUILDOUT] bootstrap.py [options]
Bootstraps a buildout-based project.
Simply run this script in a directory containing a buildout.cfg, using the
Python that you want bin/buildout to use.
Note that by using --find-links to point to local resources, you can keep
this script from going over the network.
'''
parser = OptionParser(usage=usage)
parser.add_option("--version",
action="store_true", default=False,
help=("Return bootstrap.py version."))
parser.add_option("-t", "--accept-buildout-test-releases",
dest='accept_buildout_test_releases',
action="store_true", default=False,
help=("Normally, if you do not specify a --version, the "
"bootstrap script and buildout gets the newest "
"*final* versions of zc.buildout and its recipes and "
"extensions for you. If you use this flag, "
"bootstrap and buildout will get the newest releases "
"even if they are alphas or betas."))
parser.add_option("-c", "--config-file",
help=("Specify the path to the buildout configuration "
"file to be used."))
parser.add_option("-f", "--find-links",
help=("Specify a URL to search for buildout releases"))
parser.add_option("--allow-site-packages",
action="store_true", default=False,
help=("Let bootstrap.py use existing site packages"))
parser.add_option("--buildout-version",
help="Use a specific zc.buildout version")
parser.add_option("--setuptools-version",
help="Use a specific setuptools version")
parser.add_option("--setuptools-to-dir",
help=("Allow for re-use of existing directory of "
"setuptools versions"))
options, args = parser.parse_args()
if options.version:
print("bootstrap.py version %s" % __version__)
sys.exit(0)
######################################################################
# load/install setuptools
try:
from urllib.request import urlopen
except ImportError:
from urllib2 import urlopen
ez = {}
if os.path.exists('ez_setup.py'):
exec(open('ez_setup.py').read(), ez)
else:
exec(urlopen('https://bootstrap.pypa.io/ez_setup.py').read(), ez)
if not options.allow_site_packages:
# ez_setup imports site, which adds site packages
# this will remove them from the path to ensure that incompatible versions
# of setuptools are not in the path
import site
# inside a virtualenv, there is no 'getsitepackages'.
# We can't remove these reliably
if hasattr(site, 'getsitepackages'):
for sitepackage_path in site.getsitepackages():
# Strip all site-packages directories from sys.path that
# are not sys.prefix; this is because on Windows
# sys.prefix is a site-package directory.
if sitepackage_path != sys.prefix:
sys.path[:] = [x for x in sys.path
if sitepackage_path not in x]
setup_args = dict(to_dir=tmpeggs, download_delay=0)
if options.setuptools_version is not None:
setup_args['version'] = options.setuptools_version
if options.setuptools_to_dir is not None:
setup_args['to_dir'] = options.setuptools_to_dir
ez['use_setuptools'](**setup_args)
import setuptools
import pkg_resources
# This does not (always?) update the default working set. We will
# do it.
for path in sys.path:
if path not in pkg_resources.working_set.entries:
pkg_resources.working_set.add_entry(path)
######################################################################
# Install buildout
ws = pkg_resources.working_set
setuptools_path = ws.find(
pkg_resources.Requirement.parse('setuptools')).location
# Fix sys.path here as easy_install.pth added before PYTHONPATH
cmd = [sys.executable, '-c',
'import sys; sys.path[0:0] = [%r]; ' % setuptools_path +
'from setuptools.command.easy_install import main; main()',
'-mZqNxd', tmpeggs]
find_links = os.environ.get(
'bootstrap-testing-find-links',
options.find_links or
('http://downloads.buildout.org/'
if options.accept_buildout_test_releases else None)
)
if find_links:
cmd.extend(['-f', find_links])
requirement = 'zc.buildout'
version = options.buildout_version
if version is None and not options.accept_buildout_test_releases:
# Figure out the most recent final version of zc.buildout.
import setuptools.package_index
_final_parts = '*final-', '*final'
def _final_version(parsed_version):
try:
return not parsed_version.is_prerelease
except AttributeError:
# Older setuptools
for part in parsed_version:
if (part[:1] == '*') and (part not in _final_parts):
return False
return True
index = setuptools.package_index.PackageIndex(
search_path=[setuptools_path])
if find_links:
index.add_find_links((find_links,))
req = pkg_resources.Requirement.parse(requirement)
if index.obtain(req) is not None:
best = []
bestv = None
for dist in index[req.project_name]:
distv = dist.parsed_version
if _final_version(distv):
if bestv is None or distv > bestv:
best = [dist]
bestv = distv
elif distv == bestv:
best.append(dist)
if best:
best.sort()
version = best[-1].version
if version:
requirement = '=='.join((requirement, version))
cmd.append(requirement)
import subprocess
if subprocess.call(cmd) != 0:
raise Exception(
"Failed to execute command:\n%s" % repr(cmd)[1:-1])
######################################################################
# Import and run buildout
ws.add_entry(tmpeggs)
ws.require(requirement)
import zc.buildout.buildout
if not [a for a in args if '=' not in a]:
args.append('bootstrap')
# if -c was provided, we push it back into args for buildout' main function
if options.config_file is not None:
args[0:0] = ['-c', options.config_file]
zc.buildout.buildout.main(args)
shutil.rmtree(tmpeggs)
doc/conf.py 100644 → 100755
View file @ 61a31463
...@@ -25,6 +25,7 @@ extensions = [ ...@@ -25,6 +25,7 @@ extensions = [
'sphinx.ext.intersphinx', 'sphinx.ext.intersphinx',
'sphinx.ext.napoleon', 'sphinx.ext.napoleon',
'sphinx.ext.viewcode', 'sphinx.ext.viewcode',
'matplotlib.sphinxext.plot_directive'
] ]
import sphinx import sphinx
...@@ -231,7 +232,6 @@ autodoc_member_order = 'bysource' ...@@ -231,7 +232,6 @@ autodoc_member_order = 'bysource'
autodoc_default_flags = [ autodoc_default_flags = [
'members', 'members',
'undoc-members', 'undoc-members',
'inherited-members',
'show-inheritance', 'show-inheritance',
] ]
......
doc/guide.rst
View file @ 61a31463
.. vim: set fileencoding=utf-8 : .. vim: set fileencoding=utf-8 :
.. Laurent El Shafey <Laurent.El-Shafey@idiap.ch>
.. Wed Mar 14 12:31:35 2012 +0100
..
.. Copyright (C) 2011-2014 Idiap Research Institute, Martigny, Switzerland
.. testsetup:: * .. testsetup:: *
...@@ -21,671 +17,570 @@ ...@@ -21,671 +17,570 @@
User guide User guide
============ ============
This section includes the machine/trainer guides for learning techniques The EM algorithm is an iterative method that estimates parameters for
available in this package. statistical models, where the model depends on unobserved latent variables. The
EM iteration alternates between performing an expectation (E) step, which
creates a function for the expectation of the log-likelihood evaluated using
the current estimate for the parameters, and a maximization (M) step, which
computes parameters maximizing the expected log-likelihood found on the E step.
These parameter-estimates are then used to determine the distribution of the
latent variables in the next E step [8]_.
Machines *Machines* and *trainers* are the core components of Bob's machine learning
-------- packages. *Machines* represent statistical models or other functions defined by
parameters that can be learned by *trainers* or manually set. Below you will
find machine/trainer guides for learning techniques available in this package.
Machines are one of the core components of |project|. They represent
statistical models or other functions defined by parameters that can be learnt
or set by using Trainers.
K-means machines K-Means
================ -------
.. _kmeans:
`k-means <http://en.wikipedia.org/wiki/K-means_clustering>`_ is a clustering **k-means** [7]_ is a clustering method which aims to partition a set of
method which aims to partition a set of observations into :math:`k` clusters. :math:`N` observations into
The `training` procedure is described further below. Here, we explain only how :math:`C` clusters with equal variance minimizing the following cost function
to use the resulting machine. For the sake of example, we create a new :math:`J = \sum_{i=0}^{N} \min_{\mu_j \in C} ||x_i - \mu_j||`, where
:py:class:`bob.learn.em.KMeansMachine` as follows: :math:`\mu` is a given mean (also called centroid) and
:math:`x_i` is an observation.
.. doctest:: This implementation has two stopping criteria. The first one is when the
:options: +NORMALIZE_WHITESPACE maximum number of iterations is reached; the second one is when the difference
between :math:`Js` of successive iterations are lower than a convergence
threshold.
>>> machine = bob.learn.em.KMeansMachine(2,3) # Two clusters with a feature dimensionality of 3 In this implementation, the training consists in the definition of the
>>> machine.means = numpy.array([[1,0,0],[0,0,1]], 'float64') # Defines the two clusters statistical model, called machine, (:py:class:`bob.learn.em.KMeansMachine`) and
this statistical model is learned via a trainer
(:py:class:`bob.learn.em.KMeansTrainer`).
Then, given some input data, it is possible to determine to which cluster the Follow bellow an snippet on how to train a KMeans using Bob_.
data is the closest as well as the min distance.
.. doctest:: .. doctest::
:options: +NORMALIZE_WHITESPACE :options: +NORMALIZE_WHITESPACE
>>> sample = numpy.array([2,1,-2], 'float64') >>> import bob.learn.em
>>> print(machine.get_closest_mean(sample)) # Returns the index of the closest mean and the distance to it at the power of 2 >>> import numpy
(0, 6.0) >>> data = numpy.array(
... [[3,-3,100],
... [4,-4,98],
Gaussian machines ... [3.5,-3.5,99],
================= ... [-7,7,-100],
... [-5,5,-101]], dtype='float64')
The :py:class:`bob.learn.em.Gaussian` represents a `multivariate diagonal >>> # Create a kmeans m with k=2 clusters with a dimensionality equal to 3
Gaussian (or normal) distribution >>> kmeans_machine = bob.learn.em.KMeansMachine(2, 3)
<http://en.wikipedia.org/wiki/Multivariate_normal_distribution>`_. In this >>> kmeans_trainer = bob.learn.em.KMeansTrainer()
context, a *diagonal* Gaussian refers to the covariance matrix of the >>> max_iterations = 200
distribution being diagonal. When the covariance matrix is diagonal, each >>> convergence_threshold = 1e-5
variable in the distribution is independent of the others. >>> # Train the KMeansMachine
>>> bob.learn.em.train(kmeans_trainer, kmeans_machine, data,
Objects of this class are normally used as building blocks for more complex ... max_iterations=max_iterations,
:py:class:`bob.learn.em.GMMMachine` or GMM objects, but can also be used ... convergence_threshold=convergence_threshold)
individually. Here is how to create one multivariate diagonal Gaussian >>> print(kmeans_machine.means)
distribution: [[ -6. 6. -100.5]
[ 3.5 -3.5 99. ]]
.. doctest::
>>> g = bob.learn.em.Gaussian(2) #bi-variate diagonal normal distribution
>>> g.mean = numpy.array([0.3, 0.7], 'float64')
>>> g.mean
array([ 0.3, 0.7])
>>> g.variance = numpy.array([0.2, 0.1], 'float64')
>>> g.variance
array([ 0.2, 0.1])
Once the :py:class:`bob.learn.em.Gaussian` has been set, you can use it to Bellow follow an intuition (source code + plot) of a kmeans training using the
estimate the log-likelihood of an input feature vector with a matching number Iris flower `dataset <https://en.wikipedia.org/wiki/Iris_flower_data_set>`_.
of dimensions:
.. doctest:: .. plot:: plot/plot_kmeans.py
:include-source: False
>>> log_likelihood = g(numpy.array([0.4, 0.4], 'float64'))
As with other machines you can save and re-load machines of this type using
:py:meth:`bob.learn.em.Gaussian.save` and the class constructor
respectively.
Gaussian mixture models Gaussian mixture models
======================= -----------------------
The :py:class:`bob.learn.em.GMMMachine` represents a Gaussian `mixture model
<http://en.wikipedia.org/wiki/Mixture_model>`_ (GMM), which consists of a
mixture of weighted :py:class:`bob.learn.em.Gaussian`\s.
.. doctest::
>>> gmm = bob.learn.em.GMMMachine(2,3) # Mixture of two diagonal Gaussian of dimension 3
By default, the diagonal Gaussian distributions of the GMM are initialized with
zero mean and unit variance, and the weights are identical. This can be updated
using the :py:attr:`bob.learn.em.GMMMachine.means`,
:py:attr:`bob.learn.em.GMMMachine.variances` or
:py:attr:`bob.learn.em.GMMMachine.weights`.
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> gmm.weights = numpy.array([0.4, 0.6], 'float64')
>>> gmm.means = numpy.array([[1, 6, 2], [4, 3, 2]], 'float64')
>>> gmm.variances = numpy.array([[1, 2, 1], [2, 1, 2]], 'float64')
>>> gmm.means
array([[ 1., 6., 2.],
[ 4., 3., 2.]])
Once the :py:class:`bob.learn.em.GMMMachine` has been set, you can use it to
estimate the log-likelihood of an input feature vector with a matching number
of dimensions:
.. doctest::
>>> log_likelihood = gmm(numpy.array([5.1, 4.7, -4.9], 'float64'))
As with other machines you can save and re-load machines of this type using
:py:meth:`bob.learn.em.GMMMachine.save` and the class constructor respectively.
Gaussian mixture models Statistics
==================================
The :py:class:`bob.learn.em.GMMStats` is a container for the sufficient
statistics of a GMM distribution.
Given a GMM, the sufficient statistics of a sample can be computed as
follows:
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> gs = bob.learn.em.GMMStats(2,3)
>>> sample = numpy.array([0.5, 4.5, 1.5])
>>> gmm.acc_statistics(sample, gs)
>>> print(gs) # doctest: +SKIP
Then, the sufficient statistics can be accessed (or set as below), by
considering the following attributes.
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> gs = bob.learn.em.GMMStats(2,3)
>>> log_likelihood = -3. # log-likelihood of the accumulated samples
>>> T = 1 # Number of samples used to accumulate statistics
>>> n = numpy.array([0.4, 0.6], 'float64') # zeroth order stats
>>> sumpx = numpy.array([[1., 2., 3.], [4., 5., 6.]], 'float64') # first order stats
>>> sumpxx = numpy.array([[10., 20., 30.], [40., 50., 60.]], 'float64') # second order stats
>>> gs.log_likelihood = log_likelihood
>>> gs.t = T
>>> gs.n = n
>>> gs.sum_px = sumpx
>>> gs.sum_pxx = sumpxx
Joint Factor Analysis
=====================
Joint Factor Analysis (JFA) [1]_ [2]_ is a session variability modelling
technique built on top of the Gaussian mixture modelling approach. It utilises
a within-class subspace :math:`U`, a between-class subspace :math:`V`, and a
subspace for the residuals :math:`D` to capture and suppress a significant
portion of between-class variation.
An instance of :py:class:`bob.learn.em.JFABase` carries information about
the matrices :math:`U`, :math:`V` and :math:`D`, which can be shared between
several classes. In contrast, after the enrollment phase, an instance of
:py:class:`bob.learn.em.JFAMachine` carries class-specific information about
the latent variables :math:`y` and :math:`z`.
An instance of :py:class:`bob.learn.em.JFABase` can be initialized as
follows, given an existing GMM:
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> jfa_base = bob.learn.em.JFABase(gmm,2,2) # dimensions of U and V are both equal to 2
>>> U = numpy.array([[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12]], 'float64')
>>> V = numpy.array([[6, 5], [4, 3], [2, 1], [1, 2], [3, 4], [5, 6]], 'float64')
>>> d = numpy.array([0, 1, 0, 1, 0, 1], 'float64')
>>> jfa_base.u = U
>>> jfa_base.v = V
>>> jfa_base.d = d
Next, this :py:class:`bob.learn.em.JFABase` can be shared by several
instances of :py:class:`bob.learn.em.JFAMachine`, the initialization being
as follows:
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> m = bob.learn.em.JFAMachine(jfa_base)
>>> m.y = numpy.array([1,2], 'float64')
>>> m.z = numpy.array([3,4,1,2,0,1], 'float64')
Once the :py:class:`bob.learn.em.JFAMachine` has been configured for a
specific class, the log-likelihood (score) that an input sample belongs to the
enrolled class, can be estimated, by first computing the GMM sufficient
statistics of this input sample, and then calling the
:py:meth:`bob.learn.em.JFAMachine.log_likelihood` on the sufficient statistics.
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> gs = bob.learn.em.GMMStats(2,3)
>>> gmm.acc_statistics(sample, gs)
>>> score = m(gs)
As with other machines you can save and re-load machines of this type using
:py:meth:`bob.learn.em.JFAMachine.save` and the class constructor
respectively.
Inter-Session Variability
=========================
Similarly to Joint Factor Analysis, Inter-Session Variability (ISV) modelling
[3]_ [2]_ is another session variability modelling technique built on top of
the Gaussian mixture modelling approach. It utilises a within-class subspace
:math:`U` and a subspace for the residuals :math:`D` to capture and suppress a
significant portion of between-class variation. The main difference compared to
JFA is the absence of the between-class subspace :math:`V`.
Similarly to JFA, an instance of :py:class:`bob.learn.em.JFABase` carries
information about the matrices :math:`U` and :math:`D`, which can be shared
between several classes, whereas an instance of
:py:class:`bob.learn.em.JFAMachine` carries class-specific information about
the latent variable :math:`z`.
An instance of :py:class:`bob.learn.em.ISVBase` can be initialized as
follows, given an existing GMM:
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> isv_base = bob.learn.em.ISVBase(gmm,2) # dimension of U is equal to 2
>>> isv_base.u = U
>>> isv_base.d = d
Next, this :py:class:`bob.learn.em.ISVBase` can be shared by several
instances of :py:class:`bob.learn.em.ISVMachine`, the initialization being
as follows:
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> m = bob.learn.em.ISVMachine(isv_base)
>>> m.z = numpy.array([3,4,1,2,0,1], 'float64')
Once the :py:class:`bob.learn.em.ISVMachine` has been configured for a
specific class, the log-likelihood (score) that an input sample belongs to the
enrolled class, can be estimated, by first computing the GMM sufficient
statistics of this input sample, and then calling the
``__call__`` on the sufficient statistics.
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> gs = bob.learn.em.GMMStats(2,3)
>>> gmm.acc_statistics(sample, gs)
>>> score = m(gs)
As with other machines you can save and re-load machines of this type using
:py:meth:`bob.learn.em.ISVMachine.save` and the class constructor
respectively.
Total Variability (i-vectors)
=============================
Total Variability (TV) modelling [4]_ is a front-end initially introduced for
speaker recognition, which aims at describing samples by vectors of low
dimensionality called ``i-vectors``. The model consists of a subspace :math:`T`
and a residual diagonal covariance matrix :math:`\Sigma`, that are then used to
extract i-vectors, and is built upon the GMM approach.
An instance of the class :py:class:`bob.learn.em.IVectorMachine` carries
information about these two matrices. This can be initialized as follows:
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> m = bob.learn.em.IVectorMachine(gmm, 2)
>>> m.t = numpy.array([[1.,2],[4,1],[0,3],[5,8],[7,10],[11,1]])
>>> m.sigma = numpy.array([1.,2.,1.,3.,2.,4.])
Once the :py:class:`bob.learn.em.IVectorMachine` has been set, the
extraction of an i-vector :math:`w_{ij}` can be done in two steps, by first
extracting the GMM sufficient statistics, and then estimating the i-vector:
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> gs = bob.learn.em.GMMStats(2,3)
>>> gmm.acc_statistics(sample, gs)
>>> w_ij = m(gs)
As with other machines you can save and re-load machines of this type using
:py:meth:`bob.learn.em.IVectorMachine.save` and the class constructor
respectively.
Probabilistic Linear Discriminant Analysis (PLDA)
=================================================
Probabilistic Linear Discriminant Analysis [5]_ [6]_ is a probabilistic model
that incorporates components describing both between-class and within-class
variations. Given a mean :math:`\mu`, between-class and within-class subspaces
:math:`F` and :math:`G` and residual noise :math:`\epsilon` with zero mean and
diagonal covariance matrix :math:`\Sigma`, the model assumes that a sample
:math:`x_{i,j}` is generated by the following process:
.. math::
x_{i,j} = \mu + F h_{i} + G w_{i,j} + \epsilon_{i,j}
Information about a PLDA model (:math:`\mu`, :math:`F`, :math:`G` and
:math:`\Sigma`) are carried out by an instance of the class
:py:class:`bob.learn.em.PLDABase`.
.. doctest::
>>> ### This creates a PLDABase container for input feature of dimensionality 3,
>>> ### and with subspaces F and G of rank 1 and 2 respectively.
>>> pldabase = bob.learn.em.PLDABase(3,1,2)
Class-specific information (usually from enrollment samples) are contained in
an instance of :py:class:`bob.learn.em.PLDAMachine`, that must be attached
to a given :py:class:`bob.learn.em.PLDABase`. Once done, log-likelihood
computations can be performed.
.. doctest::
>>> plda = bob.learn.em.PLDAMachine(pldabase)
>>> samples = numpy.array([[3.5,-3.4,102], [4.5,-4.3,56]], dtype=numpy.float64)
>>> loglike = plda.compute_log_likelihood(samples)
Trainers
--------
In the previous section, the concept of a `machine` was introduced. A `machine`
is fed by some input data, processes it and returns an output. Machines can be
learnt using trainers in |project|.
Expectation Maximization
========================
Each one of the following trainers has their own `initialize`, `eStep` and `mStep` methods in order to train the respective machines. A Gaussian mixture model (`GMM <http://en.wikipedia.org/wiki/Mixture_model>`_)
For example, to train a K-Means with 10 iterations you can use the following steps. is a probabilistic model for density estimation. It assumes that all the data
points are generated from a mixture of a finite number of Gaussian
distributions. More formally, a GMM can be defined as:
:math:`P(x|\Theta) = \sum_{c=0}^{C} \omega_c \mathcal{N}(x | \mu_c, \sigma_c)`
, where :math:`\Theta = \{ \omega_c, \mu_c, \sigma_c \}`.
.. doctest:: This statistical model is defined in the class
:options: +NORMALIZE_WHITESPACE :py:class:`bob.learn.em.GMMMachine` as bellow.
>>> data = numpy.array([[3,-3,100], [4,-4,98], [3.5,-3.5,99], [-7,7,-100], [-5,5,-101]], dtype='float64') #Data
>>> kmeans_machine = bob.learn.em.KMeansMachine(2, 3) # Create a machine with k=2 clusters with a dimensionality equal to 3
>>> kmeans_trainer = bob.learn.em.KMeansTrainer() #Creating the k-means machine
>>> max_iterations = 10
>>> kmeans_trainer.initialize(kmeans_machine, data) #Initilizing the means with random values
>>> for i in range(max_iterations):
... kmeans_trainer.e_step(kmeans_machine, data)
... kmeans_trainer.m_step(kmeans_machine, data)
>>> print(kmeans_machine.means)
[[ -6. 6. -100.5]
[ 3.5 -3.5 99. ]]
With that granularity you can train your K-Means (or any trainer procedure) with your own convergence criteria.
Furthermore, to make the things even simpler, it is possible to train the K-Means (and have the same example as above) using the wrapper :py:class:`bob.learn.em.train` as in the example below:
.. doctest:: .. doctest::
:options: +NORMALIZE_WHITESPACE :options: +NORMALIZE_WHITESPACE
>>> data = numpy.array([[3,-3,100], [4,-4,98], [3.5,-3.5,99], [-7,7,-100], [-5,5,-101]], dtype='float64') #Data >>> import bob.learn.em
>>> kmeans_machine = bob.learn.em.KMeansMachine(2, 3) # Create a machine with k=2 clusters with a dimensionality equal to 3 >>> # Create a GMM with k=2 Gaussians with the dimensionality of 3
>>> kmeans_trainer = bob.learn.em.KMeansTrainer() #Creating the k-means machine >>> gmm_machine = bob.learn.em.GMMMachine(2, 3)
>>> max_iterations = 10
>>> bob.learn.em.train(kmeans_trainer, kmeans_machine, data, max_iterations = 10) #wrapper for the em trainer
>>> print(kmeans_machine.means)
[[ -6. 6. -100.5]
[ 3.5 -3.5 99. ]]
There are plenty of ways to estimate :math:`\Theta`; the next subsections
explains some that are implemented in Bob.
K-means
=======
**k-means** [7]_ is a clustering method, which aims to partition a set of Maximum likelihood Estimator (MLE)
observations into :math:`k` clusters. This is an `unsupervised` technique. As ==================================
for **PCA** [1]_, which is implemented in the :py:class:`bob.learn.linear.PCATrainer` .. _mle:
class, the training data is passed in a 2D :py:class:`numpy.ndarray` container.
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> data = numpy.array([[3,-3,100], [4,-4,98], [3.5,-3.5,99], [-7,7,-100], [-5,5,-101]], dtype='float64')
The training procedure will learn the `means` for the In statistics, maximum likelihood estimation (MLE) is a method of estimating
:py:class:`bob.learn.em.KMeansMachine`. The number :math:`k` of `means` is given the parameters of a statistical model given observations by finding the
when creating the `machine`, as well as the dimensionality of the features. :math:`\Theta` that maximizes :math:`P(x|\Theta)` for all :math:`x` in your
dataset [10]_. This optimization is done by the **Expectation-Maximization**
(EM) algorithm [8]_ and it is implemented by
:py:class:`bob.learn.em.ML_GMMTrainer`.
.. doctest:: A very nice explanation of EM algorithm for the maximum likelihood estimation
:options: +NORMALIZE_WHITESPACE can be found in this
`Mathematical Monk <https://www.youtube.com/watch?v=AnbiNaVp3eQ>`_ YouTube
video.
>>> kmeans = bob.learn.em.KMeansMachine(2, 3) # Create a machine with k=2 clusters with a dimensionality equal to 3 Follow bellow an snippet on how to train a GMM using the maximum likelihood
estimator.
Then training procedure for `k-means` is an **Expectation-Maximization**-based
[8]_ algorithm. There are several options that can be set such as the maximum
number of iterations and the criterion used to determine if the convergence has
occurred. After setting all of these options, the training procedure can then
be called.
.. doctest:: .. doctest::
:options: +NORMALIZE_WHITESPACE :options: +NORMALIZE_WHITESPACE
>>> kmeansTrainer = bob.learn.em.KMeansTrainer() >>> import bob.learn.em
>>> import numpy
>>> bob.learn.em.train(kmeansTrainer, kmeans, data, max_iterations = 200, convergence_threshold = 1e-5) # Train the KMeansMachine >>> data = numpy.array(
>>> print(kmeans.means) ... [[3,-3,100],
... [4,-4,98],
... [3.5,-3.5,99],
... [-7,7,-100],
... [-5,5,-101]], dtype='float64')
>>> # Create a kmeans model (machine) m with k=2 clusters
>>> # with a dimensionality equal to 3
>>> gmm_machine = bob.learn.em.GMMMachine(2, 3)
>>> # Using the MLE trainer to train the GMM:
>>> # True, True, True means update means/variances/weights at each
>>> # iteration
>>> gmm_trainer = bob.learn.em.ML_GMMTrainer(True, True, True)
>>> # Setting some means to start the training.
>>> # In practice, the output of kmeans is a good start for the MLE training
>>> gmm_machine.means = numpy.array(
... [[ -4., 2.3, -10.5],
... [ 2.5, -4.5, 59. ]])
>>> max_iterations = 200
>>> convergence_threshold = 1e-5
>>> # Training
>>> bob.learn.em.train(gmm_trainer, gmm_machine, data,
... max_iterations=max_iterations,
... convergence_threshold=convergence_threshold)
>>> print(gmm_machine.means)
[[ -6. 6. -100.5] [[ -6. 6. -100.5]
[ 3.5 -3.5 99. ]] [ 3.5 -3.5 99. ]]
Bellow follow an intuition of the GMM trained the maximum likelihood estimator
using the Iris flower
`dataset <https://en.wikipedia.org/wiki/Iris_flower_data_set>`_.
Maximum likelihood for Gaussian mixture model .. plot:: plot/plot_ML.py
============================================= :include-source: False
A Gaussian **mixture model** (GMM) [9]_ is a common probabilistic model. In
order to train the parameters of such a model it is common to use a
**maximum-likelihood** (ML) approach [10]_. To do this we use an
**Expectation-Maximization** (EM) algorithm [8]_. Let's first start by creating
a :py:class:`bob.learn.em.GMMMachine`. By default, all of the Gaussian's have
zero-mean and unit variance, and all the weights are equal. As a starting
point, we could set the mean to the one obtained with **k-means** [7]_.
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> gmm = bob.learn.em.GMMMachine(2,3) # Create a machine with 2 Gaussian and feature dimensionality 3
>>> gmm.means = kmeans.means # Set the means to the one obtained with k-means
The |project| class to learn the parameters of a GMM [9]_ using ML [10]_ is
:py:class:`bob.learn.em.ML_GMMTrainer`. It uses an **EM**-based [8]_ algorithm
and requires the user to specify which parameters of the GMM are updated at
each iteration (means, variances and/or weights). In addition, and as for
**k-means** [7]_, it has parameters such as the maximum number of iterations
and the criterion used to determine if the parameters have converged.
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> trainer = bob.learn.em.ML_GMMTrainer(True, True, True) # update means/variances/weights at each iteration
>>> bob.learn.em.train(trainer, gmm, data, max_iterations = 200, convergence_threshold = 1e-5)
>>> print(gmm) # doctest: +SKIP
MAP-adaptation for Gaussian mixture model
=========================================
|project| also supports the training of GMMs [9]_ using a **maximum a
posteriori** (MAP) approach [11]_. MAP is closely related to the ML [10]_
technique but it incorporates a prior on the quantity that we want to estimate.
In our case, this prior is a GMM [9]_. Based on this prior model and some
training data, a new model, the MAP estimate, will be `adapted`.
Let's consider that the previously trained GMM [9]_ is our prior model.
.. doctest:: Maximum a posteriori Estimator (MAP)
:options: +NORMALIZE_WHITESPACE ====================================
.. _map:
>>> print(gmm) # doctest: +SKIP Closely related to the MLE, Maximum a posteriori probability (MAP) is an
estimate that equals the mode of the posterior distribution by incorporating in
its loss function a prior distribution [11]_. Commonly this prior distribution
(the values of :math:`\Theta`) is estimated with MLE. This optimization is done
by the **Expectation-Maximization** (EM) algorithm [8]_ and it is implemented
by :py:class:`bob.learn.em.MAP_GMMTrainer`.
The training data used to compute the MAP estimate [11]_ is again stored in a A compact way to write relevance MAP adaptation is by using GMM supervector
2D :py:class:`numpy.ndarray` container. notation (this will be useful in the next subsections). The GMM supervector
notation consists of taking the parameters of :math:`\Theta` (weights, means
.. doctest:: and covariance matrices) of a GMM and create a single vector or matrix to
:options: +NORMALIZE_WHITESPACE represent each of them. For each Gaussian component :math:`c`, we can
represent the MAP adaptation as the following :math:`\mu_i = m + d_i`, where
:math:`m` is our prior and :math:`d_i` is the class offset.
>>> dataMAP = numpy.array([[7,-7,102], [6,-6,103], [-3.5,3.5,-97]], dtype='float64') Follow bellow an snippet on how to train a GMM using the MAP estimator.
The |project| class used to perform MAP adaptation training [11]_ is
:py:class:`bob.learn.em.MAP_GMMTrainer`. As with the ML estimate [10]_, it uses
an **EM**-based [8]_ algorithm and requires the user to specify which parts of
the GMM are adapted at each iteration (means, variances and/or weights). In
addition, it also has parameters such as the maximum number of iterations and
the criterion used to determine if the parameters have converged, in addition
to this there is also a relevance factor which indicates the importance we give
to the prior. Once the trainer has been created, a prior GMM [9]_ needs to be
set.
.. doctest:: .. doctest::
:options: +NORMALIZE_WHITESPACE :options: +NORMALIZE_WHITESPACE
>>> relevance_factor = 4. >>> import bob.learn.em
>>> trainer = bob.learn.em.MAP_GMMTrainer(gmm, relevance_factor=relevance_factor, update_means=True, update_variances=False, update_weights=False) # mean adaptation only >>> import numpy
>>> gmmAdapted = bob.learn.em.GMMMachine(2,3) # Create a new machine for the MAP estimate >>> data = numpy.array(
>>> bob.learn.em.train(trainer, gmmAdapted, dataMAP, max_iterations = 200, convergence_threshold = 1e-5) ... [[3,-3,100],
>>> print(gmmAdapted) # doctest: +SKIP ... [4,-4,98],
... [3.5,-3.5,99],
... [-7,7,-100],
... [-5,5,-101]], dtype='float64')
>>> # Creating a fake prior
>>> prior_gmm = bob.learn.em.GMMMachine(2, 3)
>>> # Set some random means for the example
>>> prior_gmm.means = numpy.array(
... [[ -4., 2.3, -10.5],
... [ 2.5, -4.5, 59. ]])
>>> # Creating the model for the adapted GMM
>>> adapted_gmm = bob.learn.em.GMMMachine(2, 3)
>>> # Creating the MAP trainer
>>> gmm_trainer = bob.learn.em.MAP_GMMTrainer(prior_gmm, relevance_factor=4)
>>>
>>> max_iterations = 200
>>> convergence_threshold = 1e-5
>>> # Training
>>> bob.learn.em.train(gmm_trainer, adapted_gmm, data,
... max_iterations=max_iterations,
... convergence_threshold=convergence_threshold)
>>> print(adapted_gmm.means)
[[ -4.667 3.533 -40.5 ]
[ 2.929 -4.071 76.143]]
Bellow follow an intuition of the GMM trained with the MAP estimator using the
Iris flower `dataset <https://en.wikipedia.org/wiki/Iris_flower_data_set>`_.
.. plot:: plot/plot_MAP.py
:include-source: False
Session Variability Modeling with Gaussian Mixture Models
---------------------------------------------------------
In the aforementioned GMM based algorithms there is no explicit modeling of
session variability. This section will introduce some session variability
algorithms built on top of GMMs.
GMM statistics
==============
Before introduce session variability for GMM based algorithms, we must
introduce a component called :py:class:`bob.learn.em.GMMStats`. This component
is useful for some computation in the next sections.
:py:class:`bob.learn.em.GMMStats` is a container that solves the Equations 8, 9
and 10 in [Reynolds2000]_ (also called, zeroth, first and second order GMM
statistics).
Given a GMM (:math:`\Theta`) and a set of samples :math:`x_t` this component
accumulates statistics for each Gaussian component :math:`c`.
Follow bellow a 1-1 relationship between statistics in [Reynolds2000]_ and the
properties in :py:class:`bob.learn.em.GMMStats`:
- Eq (8) is :py:class:`bob.learn.em.GMMStats.n`:
:math:`n_c=\sum\limits_{t=1}^T Pr(c | x_t)` (also called responsibilities)
- Eq (9) is :py:class:`bob.learn.em.GMMStats.sum_px`:
:math:`E_c(x)=\frac{1}{n(c)}\sum\limits_{t=1}^T Pr(c | x_t)x_t`
- Eq (10) is :py:class:`bob.learn.em.GMMStats.sum_pxx`:
:math:`E_c(x^2)=\frac{1}{n(c)}\sum\limits_{t=1}^T Pr(c | x_t)x_t^2`
where :math:`T` is the number of samples used to generate the stats.
The snippet bellow shows how to compute accumulated these statistics given a
prior GMM.
Joint Factor Analysis
=====================
The training of the subspace :math:`U`, :math:`V` and :math:`D` of a Joint
Factor Analysis model, is performed in two steps. First, GMM sufficient
statistics of the training samples should be computed against the UBM GMM. Once
done, we get a training set of GMM statistics:
.. doctest:: .. doctest::
:options: +NORMALIZE_WHITESPACE :options: +NORMALIZE_WHITESPACE
>>> F1 = numpy.array( [0.3833, 0.4516, 0.6173, 0.2277, 0.5755, 0.8044, 0.5301, 0.9861, 0.2751, 0.0300, 0.2486, 0.5357]).reshape((6,2)) >>> import bob.learn.em
>>> F2 = numpy.array( [0.0871, 0.6838, 0.8021, 0.7837, 0.9891, 0.5341, 0.0669, 0.8854, 0.9394, 0.8990, 0.0182, 0.6259]).reshape((6,2)) >>> import numpy
>>> F=[F1, F2] >>> numpy.random.seed(10)
>>>
>>> N1 = numpy.array([0.1379, 0.1821, 0.2178, 0.0418]).reshape((2,2)) >>> data = numpy.array(
>>> N2 = numpy.array([0.1069, 0.9397, 0.6164, 0.3545]).reshape((2,2)) ... [[0, 0.3, -0.2],
>>> N=[N1, N2] ... [0.4, 0.1, 0.15],
... [-0.3, -0.1, 0],
>>> gs11 = bob.learn.em.GMMStats(2,3) ... [1.2, 1.4, 1],
>>> gs11.n = N1[:,0] ... [0.8, 1., 1]], dtype='float64')
>>> gs11.sum_px = F1[:,0].reshape(2,3) >>> # Creating a fake prior with 2 Gaussians of dimension 3
>>> gs12 = bob.learn.em.GMMStats(2,3) >>> prior_gmm = bob.learn.em.GMMMachine(2, 3)
>>> gs12.n = N1[:,1] >>> prior_gmm.means = numpy.vstack((numpy.random.normal(0, 0.5, (1, 3)),
>>> gs12.sum_px = F1[:,1].reshape(2,3) ... numpy.random.normal(1, 0.5, (1, 3))))
>>> # All nice and round diagonal covariance
>>> gs21 = bob.learn.em.GMMStats(2,3) >>> prior_gmm.variances = numpy.ones((2, 3)) * 0.5
>>> gs21.n = N2[:,0] >>> prior_gmm.weights = numpy.array([0.3, 0.7])
>>> gs21.sum_px = F2[:,0].reshape(2,3) >>> # Creating the container
>>> gs22 = bob.learn.em.GMMStats(2,3) >>> gmm_stats_container = bob.learn.em.GMMStats(2, 3)
>>> gs22.n = N2[:,1] >>> for d in data:
>>> gs22.sum_px = F2[:,1].reshape(2,3) ... prior_gmm.acc_statistics(d, gmm_stats_container)
>>>
>>> # Printing the responsibilities
>>> print(gmm_stats_container.n/gmm_stats_container.t)
[ 0.429 0.571]
>>> TRAINING_STATS = [[gs11, gs12], [gs21, gs22]]
In the following, we will allocate a :py:class:`bob.learn.em.JFABase` machine, Inter-Session Variability
that will then be trained. =========================
.. _isv:
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> jfa_base = bob.learn.em.JFABase(gmm, 2, 2) # the dimensions of U and V are both equal to 2 Inter-Session Variability (ISV) modeling [3]_ [2]_ is a session variability
modeling technique built on top of the Gaussian mixture modeling approach. It
hypothesizes that within-class variations are embedded in a linear subspace in
the GMM means subspace and these variations can be suppressed by an offset w.r.t
each mean during the MAP adaptation.
Next, we initialize a trainer, which is an instance of In this generative model each sample is assumed to have been generated by a GMM
:py:class:`bob.learn.em.JFATrainer`, as follows: mean supervector with the following shape:
:math:`\mu_{i, j} = m + Ux_{i, j} + D_z{i}`, where :math:`m` is our prior,
:math:`Ux_{i, j}` is the session offset that we want to suppress and
:math:`D_z{i}` is the class offset (with all session effects suppressed).
.. doctest:: All possible sources of session variations is embedded in this matrix
:options: +NORMALIZE_WHITESPACE :math:`U`. Follow bellow an intuition of what is modeled with :math:`U` in the
Iris flower `dataset <https://en.wikipedia.org/wiki/Iris_flower_data_set>`_.
The arrows :math:`U_{1}`, :math:`U_{2}` and :math:`U_{3}` are the directions of
the within class variations, with respect to each Gaussian component, that will
be suppressed a posteriori.
>>> jfa_trainer = bob.learn.em.JFATrainer() .. plot:: plot/plot_ISV.py
:include-source: False
The training process is started by calling the
:py:meth:`bob.learn.em.train`.
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> bob.learn.em.train_jfa(jfa_trainer, jfa_base, TRAINING_STATS, max_iterations=10) The ISV statistical model is stored in this container
:py:class:`bob.learn.em.ISVBase` and the training is performed by
:py:class:`bob.learn.em.ISVTrainer`. The snippet bellow shows how to train a
Intersession variability modeling.
Once the training is finished (i.e. the subspaces :math:`U`, :math:`V` and
:math:`D` are estimated), the JFA model can be shared and used by several
class-specific models. As for the training samples, we first need to extract
GMM statistics from the samples. These GMM statistics are manually defined in
the following.
.. doctest:: .. doctest::
:options: +NORMALIZE_WHITESPACE :options: +NORMALIZE_WHITESPACE
>>> Ne = numpy.array([0.1579, 0.9245, 0.1323, 0.2458]).reshape((2,2)) >>> import bob.learn.em
>>> Fe = numpy.array([0.1579, 0.1925, 0.3242, 0.1234, 0.2354, 0.2734, 0.2514, 0.5874, 0.3345, 0.2463, 0.4789, 0.5236]).reshape((6,2)) >>> import numpy
>>> gse1 = bob.learn.em.GMMStats(2,3) >>> numpy.random.seed(10)
>>> gse1.n = Ne[:,0] >>>
>>> gse1.sum_px = Fe[:,0].reshape(2,3) >>> # Generating some fake data
>>> gse2 = bob.learn.em.GMMStats(2,3) >>> data_class1 = numpy.random.normal(0, 0.5, (10, 3))
>>> gse2.n = Ne[:,1] >>> data_class2 = numpy.random.normal(-0.2, 0.2, (10, 3))
>>> gse2.sum_px = Fe[:,1].reshape(2,3) >>> data = [data_class1, data_class2]
>>> gse = [gse1, gse2]
>>> # Creating a fake prior with 2 gaussians of dimension 3
>>> prior_gmm = bob.learn.em.GMMMachine(2, 3)
>>> prior_gmm.means = numpy.vstack((numpy.random.normal(0, 0.5, (1, 3)),
... numpy.random.normal(1, 0.5, (1, 3))))
>>> # All nice and round diagonal covariance
>>> prior_gmm.variances = numpy.ones((2, 3)) * 0.5
>>> prior_gmm.weights = numpy.array([0.3, 0.7])
>>> # The input the the ISV Training is the statistics of the GMM
>>> gmm_stats_per_class = []
>>> for d in data:
... stats = []
... for i in d:
... gmm_stats_container = bob.learn.em.GMMStats(2, 3)
... prior_gmm.acc_statistics(i, gmm_stats_container)
... stats.append(gmm_stats_container)
... gmm_stats_per_class.append(stats)
>>> # Finally doing the ISV training
>>> subspace_dimension_of_u = 2
>>> relevance_factor = 4
>>> isvbase = bob.learn.em.ISVBase(prior_gmm, subspace_dimension_of_u)
>>> trainer = bob.learn.em.ISVTrainer(relevance_factor)
>>> bob.learn.em.train(trainer, isvbase, gmm_stats_per_class,
... max_iterations=50)
>>> # Printing the session offset w.r.t each Gaussian component
>>> print(isvbase.u)
[[-0.01 -0.027]
[-0.002 -0.004]
[ 0.028 0.074]
[ 0.012 0.03 ]
[ 0.033 0.085]
[ 0.046 0.12 ]]
Class-specific enrollment can then be perfomed as follows. This will estimate
the class-specific latent variables :math:`y` and :math:`z`:
.. doctest:: Joint Factor Analysis
:options: +NORMALIZE_WHITESPACE =====================
.. _jfa:
>>> m = bob.learn.em.JFAMachine(jfa_base) Joint Factor Analysis (JFA) [1]_ [2]_ is an extension of ISV. Besides the
>>> jfa_trainer.enroll(m, gse, 5) # where 5 is the number of enrollment iterations within-class assumption (modeled with :math:`U`), it also hypothesize that
between class variations are embedded in a low rank rectangular matrix
:math:`V`. In the supervector notation, this modeling has the following shape:
:math:`\mu_{i, j} = m + Ux_{i, j} + Vy_{i} + D_z{i}`.
More information about the training process can be found in [12]_ and [13]_. Follow bellow an intuition of what is modeled with :math:`U` and :math:`V` in
the Iris flower
`dataset <https://en.wikipedia.org/wiki/Iris_flower_data_set>`_. The arrows
:math:`V_{1}`, :math:`V_{2}` and :math:`V_{3}` are the directions of the
between class variations with respect to each Gaussian component that will be
added a posteriori.
Inter-Session Variability .. plot:: plot/plot_JFA.py
========================= :include-source: False
The training of the subspace :math:`U` and :math:`D` of an Inter-Session The JFA statistical model is stored in this container
Variability model, is performed in two steps. As for JFA, GMM sufficient :py:class:`bob.learn.em.JFABase` and the training is performed by
statistics of the training samples should be computed against the UBM GMM. Once :py:class:`bob.learn.em.JFATrainer`. The snippet bellow shows how to train a
done, we get a training set of GMM statistics. Next, we will allocate an Intersession variability modeling.
:py:class:`bob.learn.em.ISVBase` machine, that will then be trained.
.. doctest:: .. doctest::
:options: +NORMALIZE_WHITESPACE :options: +NORMALIZE_WHITESPACE
>>> isv_base = bob.learn.em.ISVBase(gmm, 2) # the dimensions of U is equal to 2 >>> import bob.learn.em
>>> import numpy
Next, we initialize a trainer, which is an instance of >>> numpy.random.seed(10)
:py:class:`bob.learn.em.ISVTrainer`, as follows: >>>
>>> # Generating some fake data
.. doctest:: >>> data_class1 = numpy.random.normal(0, 0.5, (10, 3))
:options: +NORMALIZE_WHITESPACE >>> data_class2 = numpy.random.normal(-0.2, 0.2, (10, 3))
>>> data = [data_class1, data_class2]
>>> # Creating a fake prior with 2 Gaussians of dimension 3
>>> prior_gmm = bob.learn.em.GMMMachine(2, 3)
>>> prior_gmm.means = numpy.vstack((numpy.random.normal(0, 0.5, (1, 3)),
... numpy.random.normal(1, 0.5, (1, 3))))
>>> # All nice and round diagonal covariance
>>> prior_gmm.variances = numpy.ones((2, 3)) * 0.5
>>> prior_gmm.weights = numpy.array([0.3, 0.7])
>>>
>>> # The input the the JFA Training is the statistics of the GMM
>>> gmm_stats_per_class = []
>>> for d in data:
... stats = []
... for i in d:
... gmm_stats_container = bob.learn.em.GMMStats(2, 3)
... prior_gmm.acc_statistics(i, gmm_stats_container)
... stats.append(gmm_stats_container)
... gmm_stats_per_class.append(stats)
>>>
>>> # Finally doing the JFA training
>>> subspace_dimension_of_u = 2
>>> subspace_dimension_of_v = 2
>>> relevance_factor = 4
>>> jfabase = bob.learn.em.JFABase(prior_gmm, subspace_dimension_of_u,
... subspace_dimension_of_v)
>>> trainer = bob.learn.em.JFATrainer()
>>> bob.learn.em.train_jfa(trainer, jfabase, gmm_stats_per_class,
... max_iterations=50)
>>> # Printing the session offset w.r.t each Gaussian component
>>> print(jfabase.v)
[[ 0.003 -0.006]
[ 0.041 -0.084]
[-0.261 0.53 ]
[-0.252 0.51 ]
[-0.387 0.785]
[-0.36 0.73 ]]
Total variability Modelling
===========================
.. _ivector:
Total Variability (TV) modeling [4]_ is a front-end initially introduced for
speaker recognition, which aims at describing samples by vectors of low
dimensionality called ``i-vectors``. The model consists of a subspace :math:`T`
and a residual diagonal covariance matrix :math:`\Sigma`, that are then used to
extract i-vectors, and is built upon the GMM approach. In the supervector
notation this modeling has the following shape: :math:`\mu = m + Tv`.
>>> isv_trainer = bob.learn.em.ISVTrainer(relevance_factor=4.) # 4 is the relevance factor Follow bellow an intuition of the data from the Iris flower
`dataset <https://en.wikipedia.org/wiki/Iris_flower_data_set>`_, embedded in
the iVector space.
The training process is started by calling the .. plot:: plot/plot_iVector.py
:py:meth:`bob.learn.em.train`. :include-source: False
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> bob.learn.em.train(isv_trainer, isv_base, TRAINING_STATS, max_iterations=10) The iVector statistical model is stored in this container
:py:class:`bob.learn.em.IVectorMachine` and the training is performed by
Once the training is finished (i.e. the subspaces :math:`V` and :math:`D` are :py:class:`bob.learn.em.IVectorTrainer`. The snippet bellow shows how to train
estimated), the ISV model can be shared and used by several class-specific a Total variability modeling.
models. As for the training samples, we first need to extract GMM statistics
from the samples. Class-specific enrollment can then be perfomed, which will
estimate the class-specific latent variable :math:`z`:
.. doctest:: .. doctest::
:options: +NORMALIZE_WHITESPACE :options: +NORMALIZE_WHITESPACE
>>> m = bob.learn.em.ISVMachine(isv_base) >>> import bob.learn.em
>>> isv_trainer.enroll(m, gse, 5) # where 5 is the number of iterations >>> import numpy
>>> numpy.random.seed(10)
More information about the training process can be found in [14]_ and [13]_. >>>
>>> # Generating some fake data
>>> data_class1 = numpy.random.normal(0, 0.5, (10, 3))
Total Variability (i-vectors) >>> data_class2 = numpy.random.normal(-0.2, 0.2, (10, 3))
============================= >>> data = [data_class1, data_class2]
>>>
>>> # Creating a fake prior with 2 gaussians of dimension 3
>>> prior_gmm = bob.learn.em.GMMMachine(2, 3)
>>> prior_gmm.means = numpy.vstack((numpy.random.normal(0, 0.5, (1, 3)),
... numpy.random.normal(1, 0.5, (1, 3))))
>>> # All nice and round diagonal covariance
>>> prior_gmm.variances = numpy.ones((2, 3)) * 0.5
>>> prior_gmm.weights = numpy.array([0.3, 0.7])
>>>
>>> # The input the the TV Training is the statistics of the GMM
>>> gmm_stats_per_class = []
>>> for d in data:
... for i in d:
... gmm_stats_container = bob.learn.em.GMMStats(2, 3)
... prior_gmm.acc_statistics(i, gmm_stats_container)
... gmm_stats_per_class.append(gmm_stats_container)
>>>
>>> # Finally doing the TV training
>>> subspace_dimension_of_t = 2
>>>
>>> ivector_trainer = bob.learn.em.IVectorTrainer(update_sigma=True)
>>> ivector_machine = bob.learn.em.IVectorMachine(
... prior_gmm, subspace_dimension_of_t, 10e-5)
>>> # train IVector model
>>> bob.learn.em.train(ivector_trainer, ivector_machine,
... gmm_stats_per_class, 500)
>>>
>>> # Printing the session offset w.r.t each Gaussian component
>>> print(ivector_machine.t)
[[ 0.11 -0.203]
[-0.124 0.014]
[ 0.296 0.674]
[ 0.447 0.174]
[ 0.425 0.583]
[ 0.394 0.794]]
Linear Scoring
==============
.. _linearscoring:
In :ref:`MAP <map>` adaptation, :ref:`ISV <isv>` and :ref:`JFA <jfa>` a
traditional way to do scoring is via the log-likelihood ratio between the
adapted model and the prior as the following:
The training of the subspace :math:`T` and :math:`\Sigma` of a Total .. math::
Variability model, is performed in two steps. As for JFA and ISV, GMM score = ln(P(x | \Theta)) - ln(P(x | \Theta_{prior})),
sufficient statistics of the training samples should be computed against the
UBM GMM. Once done, we get a training set of GMM statistics. Next, we will
allocate an instance of :py:class:`bob.learn.em.IVectorMachine`, that will
then be trained.
.. doctest::
:options: +NORMALIZE_WHITESPACE
>>> m = bob.learn.em.IVectorMachine(gmm, 2)
>>> m.variance_threshold = 1e-5
(with :math:`\Theta` varying for each approach).
Next, we initialize a trainer, which is an instance of A simplification proposed by [Glembek2009]_, called linear scoring,
:py:class:`bob.learn.em.IVectorTrainer`, as follows: approximate this ratio using a first order Taylor series as the following:
.. doctest:: .. math::
:options: +NORMALIZE_WHITESPACE score = \frac{\mu - \mu_{prior}}{\sigma_{prior}} f * (\mu_{prior} + U_x),
>>> ivec_trainer = bob.learn.em.IVectorTrainer(update_sigma=True) where :math:`\mu` is the the GMM mean supervector (of the prior and the adapted
>>> TRAINING_STATS_flatten = [gs11, gs12, gs21, gs22] model), :math:`\sigma` is the variance, supervector, :math:`f` is the first
order GMM statistics (:py:class:`bob.learn.em.GMMStats.sum_px`) and
:math:`U_x`, is possible channel offset (:ref:`ISV <isv>`).
The training process is started by calling the This scoring technique is implemented in :py:func:`bob.learn.em.linear_scoring`.
:py:meth:`bob.learn.em.train`. The snippet bellow shows how to compute scores using this approximation.
.. doctest:: .. doctest::
:options: +NORMALIZE_WHITESPACE :options: +NORMALIZE_WHITESPACE
>>> bob.learn.em.train(ivec_trainer, m, TRAINING_STATS_flatten, max_iterations=10) >>> import bob.learn.em
>>> import numpy
>>> # Defining a fake prior
>>> prior_gmm = bob.learn.em.GMMMachine(3, 2)
>>> prior_gmm.means = numpy.array([[1, 1], [2, 2.1], [3, 3]])
>>> # Defining a fake prior
>>> adapted_gmm = bob.learn.em.GMMMachine(3,2)
>>> adapted_gmm.means = numpy.array([[1.5, 1.5], [2.5, 2.5], [2, 2]])
>>> # Defining an input
>>> input = numpy.array([[1.5, 1.5], [1.6, 1.6]])
>>> #Accumulating statistics of the GMM
>>> stats = bob.learn.em.GMMStats(3, 2)
>>> prior_gmm.acc_statistics(input, stats)
>>> score = bob.learn.em.linear_scoring(
... [adapted_gmm], prior_gmm, [stats], [],
... frame_length_normalisation=True)
>>> print(score)
[[ 0.254]]
More information about the training process can be found in [15]_.
Probabilistic Linear Discriminant Analysis (PLDA) Probabilistic Linear Discriminant Analysis (PLDA)
================================================= -------------------------------------------------
Probabilistic Linear Discriminant Analysis [16]_ is a probabilistic model that Probabilistic Linear Discriminant Analysis [16]_ is a probabilistic model that
incorporates components describing both between-class and within-class incorporates components describing both between-class and within-class
...@@ -699,7 +594,7 @@ diagonal covariance matrix :math:`\Sigma`, the model assumes that a sample ...@@ -699,7 +594,7 @@ diagonal covariance matrix :math:`\Sigma`, the model assumes that a sample
x_{i,j} = \mu + F h_{i} + G w_{i,j} + \epsilon_{i,j} x_{i,j} = \mu + F h_{i} + G w_{i,j} + \epsilon_{i,j}
An Expectaction-Maximization algorithm can be used to learn the parameters of An Expectation-Maximization algorithm can be used to learn the parameters of
this model :math:`\mu`, :math:`F` :math:`G` and :math:`\Sigma`. As these this model :math:`\mu`, :math:`F` :math:`G` and :math:`\Sigma`. As these
parameters can be shared between classes, there is a specific container class parameters can be shared between classes, there is a specific container class
for this purpose, which is :py:class:`bob.learn.em.PLDABase`. The process is for this purpose, which is :py:class:`bob.learn.em.PLDABase`. The process is
...@@ -711,8 +606,14 @@ dimensionality 3. ...@@ -711,8 +606,14 @@ dimensionality 3.
.. doctest:: .. doctest::
:options: +NORMALIZE_WHITESPACE :options: +NORMALIZE_WHITESPACE
>>> data1 = numpy.array([[3,-3,100], [4,-4,50], [40,-40,150]], dtype=numpy.float64) >>> data1 = numpy.array(
>>> data2 = numpy.array([[3,6,-50], [4,8,-100], [40,79,-800]], dtype=numpy.float64) ... [[3,-3,100],
... [4,-4,50],
... [40,-40,150]], dtype=numpy.float64)
>>> data2 = numpy.array(
... [[3,6,-50],
... [4,8,-100],
... [40,79,-800]], dtype=numpy.float64)
>>> data = [data1,data2] >>> data = [data1,data2]
Learning a PLDA model can be performed by instantiating the class Learning a PLDA model can be performed by instantiating the class
...@@ -721,8 +622,8 @@ Learning a PLDA model can be performed by instantiating the class ...@@ -721,8 +622,8 @@ Learning a PLDA model can be performed by instantiating the class
.. doctest:: .. doctest::
>>> ### This creates a PLDABase container for input feature of dimensionality 3, >>> # This creates a PLDABase container for input feature of dimensionality
>>> ### and with subspaces F and G of rank 1 and 2 respectively. >>> # 3 and with subspaces F and G of rank 1 and 2, respectively.
>>> pldabase = bob.learn.em.PLDABase(3,1,2) >>> pldabase = bob.learn.em.PLDABase(3,1,2)
>>> trainer = bob.learn.em.PLDATrainer() >>> trainer = bob.learn.em.PLDATrainer()
...@@ -739,21 +640,27 @@ obtained by calling the ...@@ -739,21 +640,27 @@ obtained by calling the
.. doctest:: .. doctest::
>>> plda = bob.learn.em.PLDAMachine(pldabase) >>> plda = bob.learn.em.PLDAMachine(pldabase)
>>> samples = numpy.array([[3.5,-3.4,102], [4.5,-4.3,56]], dtype=numpy.float64) >>> samples = numpy.array(
... [[3.5,-3.4,102],
... [4.5,-4.3,56]], dtype=numpy.float64)
>>> loglike = plda.compute_log_likelihood(samples) >>> loglike = plda.compute_log_likelihood(samples)
If separate models for different classes need to be enrolled, each of them with If separate models for different classes need to be enrolled, each of them with
a set of enrolment samples, then, several instances of a set of enrollment samples, then, several instances of
:py:class:`bob.learn.em.PLDAMachine` need to be created and enrolled using :py:class:`bob.learn.em.PLDAMachine` need to be created and enrolled using
the :py:meth:`bob.learn.em.PLDATrainer.enroll()` method as follows. the :py:meth:`bob.learn.em.PLDATrainer.enroll()` method as follows.
.. doctest:: .. doctest::
>>> plda1 = bob.learn.em.PLDAMachine(pldabase) >>> plda1 = bob.learn.em.PLDAMachine(pldabase)
>>> samples1 = numpy.array([[3.5,-3.4,102], [4.5,-4.3,56]], dtype=numpy.float64) >>> samples1 = numpy.array(
... [[3.5,-3.4,102],
... [4.5,-4.3,56]], dtype=numpy.float64)
>>> trainer.enroll(plda1, samples1) >>> trainer.enroll(plda1, samples1)
>>> plda2 = bob.learn.em.PLDAMachine(pldabase) >>> plda2 = bob.learn.em.PLDAMachine(pldabase)
>>> samples2 = numpy.array([[3.5,7,-49], [4.5,8.9,-99]], dtype=numpy.float64) >>> samples2 = numpy.array(
... [[3.5,7,-49],
... [4.5,8.9,-99]], dtype=numpy.float64)
>>> trainer.enroll(plda2, samples2) >>> trainer.enroll(plda2, samples2)
Afterwards, the joint log-likelihood of the enrollment samples and of one or Afterwards, the joint log-likelihood of the enrollment samples and of one or
...@@ -766,12 +673,14 @@ separately for each model. ...@@ -766,12 +673,14 @@ separately for each model.
>>> l1 = plda1.compute_log_likelihood(sample) >>> l1 = plda1.compute_log_likelihood(sample)
>>> l2 = plda2.compute_log_likelihood(sample) >>> l2 = plda2.compute_log_likelihood(sample)
In a verification scenario, there are two possible hypotheses: 1. In a verification scenario, there are two possible hypotheses:
:math:`x_{test}` and :math:`x_{enroll}` share the same class. 2.
:math:`x_{test}` and :math:`x_{enroll}` are from different classes. Using the #. :math:`x_{test}` and :math:`x_{enroll}` share the same class.
methods :py:meth:`bob.learn.em.PLDAMachine.log_likelihood_ratio` or #. :math:`x_{test}` and :math:`x_{enroll}` are from different classes.
its alias ``__call__`` function, the corresponding
log-likelihood ratio will be computed, which is defined in more formal way by: Using the methods :py:meth:`bob.learn.em.PLDAMachine.log_likelihood_ratio` or
its alias ``__call__`` function, the corresponding log-likelihood ratio will be
computed, which is defined in more formal way by:
:math:`s = \ln(P(x_{test},x_{enroll})) - \ln(P(x_{test})P(x_{enroll}))` :math:`s = \ln(P(x_{test},x_{enroll})) - \ln(P(x_{test})P(x_{enroll}))`
.. doctest:: .. doctest::
...@@ -785,6 +694,99 @@ log-likelihood ratio will be computed, which is defined in more formal way by: ...@@ -785,6 +694,99 @@ log-likelihood ratio will be computed, which is defined in more formal way by:
os.chdir(current_directory) os.chdir(current_directory)
shutil.rmtree(temp_dir) shutil.rmtree(temp_dir)
Score Normalization
-------------------
Score normalization aims to compensate statistical variations in output scores
due to changes in the conditions across different enrollment and probe samples.
This is achieved by scaling distributions of system output scores to better
facilitate the application of a single, global threshold for authentication.
Bob has implemented 3 different strategies to normalize scores and these
strategies are presented in the next subsections.
Z-Norm
======
.. _znorm:
Given a score :math:`s_i`, Z-Norm [Auckenthaler2000]_ and [Mariethoz2005]_
(zero-normalization) scales this value by the mean (:math:`\mu`) and standard
deviation (:math:`\sigma`) of an impostor score distribution. This score
distribution can be computed before hand and it is defined as the following.
.. math::
zs_i = \frac{s_i - \mu}{\sigma}
This scoring technique is implemented in :py:func:`bob.learn.em.znorm`. Follow
bellow an example of score normalization using :py:func:`bob.learn.em.znorm`.
.. plot:: plot/plot_Znorm.py
:include-source: True
.. note::
Observe how the scores were scaled in the plot above.
T-Norm
======
.. _tnorm:
T-norm [Auckenthaler2000]_ and [Mariethoz2005]_ (Test-normalization) operates
in a probe-centric manner. If in the Z-Norm :math:`\mu` and :math:`\sigma` are
estimated using an impostor set of models and its scores, the t-norm computes
these statistics using the current probe sample against at set of models in a
co-hort :math:`\Theta_{c}`. A co-hort can be any semantic organization that is
sensible to your recognition task, such as sex (male and females), ethnicity,
age, etc and is defined as the following.
.. math::
zs_i = \frac{s_i - \mu}{\sigma}
where, :math:`s_i` is :math:`P(x_i | \Theta)` (the score given the claimed
model), :math:`\mu = \frac{ \sum\limits_{i=0}^{N} P(x_i | \Theta_{c}) }{N}`
(:math:`\Theta_{c}` are the models of one co-hort) and :math:`\sigma` is the
standard deviation computed using the same criteria used to compute
:math:`\mu`.
This scoring technique is implemented in :py:func:`bob.learn.em.tnorm`. Follow
bellow an example of score normalization using :py:func:`bob.learn.em.tnorm`.
.. plot:: plot/plot_Tnorm.py
:include-source: True
.. note::
T-norm introduces extra computation during scoring, as the probe samples
need to be compared to each cohort model in order to have :math:`\mu` and
:math:`\sigma`.
ZT-Norm
=======
.. _ztnorm:
ZT-Norm [Auckenthaler2000]_ and [Mariethoz2005]_ consists in the application of
:ref:`Z-Norm <znorm>` followed by a :ref:`T-Norm <tnorm>` and it is implemented
in :py:func:`bob.learn.em.ztnorm`.
Follow bellow an example of score normalization using
:py:func:`bob.learn.em.ztnorm`.
.. plot:: plot/plot_ZTnorm.py
:include-source: True
.. note::
Observe how the scores were scaled in the plot above.
.. Place here your external references .. Place here your external references
.. include:: links.rst .. include:: links.rst
.. [1] http://dx.doi.org/10.1109/TASL.2006.881693 .. [1] http://dx.doi.org/10.1109/TASL.2006.881693
......
doc/index.rst
View file @ 61a31463
.. vim: set fileencoding=utf-8 : .. vim: set fileencoding=utf-8 :
.. Tiago de Freitas Pereira <tiago.pereira@idiap.ch>
.. Tue 17 Feb 2015 13:50:06 CET
..
.. Copyright (C) 2011-2014 Idiap Research Institute, Martigny, Switzerland
.. _bob.learn.em: .. _bob.learn.em:
...@@ -10,12 +6,12 @@ ...@@ -10,12 +6,12 @@
Expectation Maximization Machine Learning Tools Expectation Maximization Machine Learning Tools
================================================ ================================================
The EM algorithm is an iterative method that estimates parameters for statistical models, where the model depends on unobserved latent variables. The EM iteration alternates between performing an expectation (E) step, which creates a function for the expectation of the log-likelihood evaluated using the current estimate for the parameters, and a maximization (M) step, which computes parameters maximizing the expected log-likelihood found on the E step. These parameter-estimates are then used to determine the distribution of the latent variables in the next E step [WikiEM]_. This package is a part of Bob_. It implements a general EM algorithm and
includes implementations of the following algorithms:
The package includes the machine definition per se and a selection of different trainers for specialized purposes: - K-Means
- Maximum Likelihood (ML) - Maximum Likelihood (ML)
- Maximum a Posteriori (MAP) - Maximum a Posteriori (MAP)
- K-Means
- Inter Session Variability Modelling (ISV) - Inter Session Variability Modelling (ISV)
- Joint Factor Analysis (JFA) - Joint Factor Analysis (JFA)
- Total Variability Modeling (iVectors) - Total Variability Modeling (iVectors)
...@@ -31,7 +27,7 @@ Documentation ...@@ -31,7 +27,7 @@ Documentation
guide guide
py_api py_api
References References
----------- -----------
...@@ -47,7 +43,9 @@ References ...@@ -47,7 +43,9 @@ References
.. [Roweis1998] Roweis, Sam. "EM algorithms for PCA and SPCA." Advances in neural information processing systems (1998): 626-632. .. [Roweis1998] Roweis, Sam. "EM algorithms for PCA and SPCA." Advances in neural information processing systems (1998): 626-632.
.. [WikiEM] `Expectation Maximization <http://en.wikipedia.org/wiki/Expectation%E2%80%93maximization_algorithm>`_ .. [WikiEM] `Expectation Maximization <http://en.wikipedia.org/wiki/Expectation%E2%80%93maximization_algorithm>`_
.. [Glembek2009] Glembek, Ondrej, et al. "Comparison of scoring methods used in speaker recognition with joint factor analysis." Acoustics, Speech and Signal Processing, 2009. ICASSP 2009. IEEE International Conference on. IEEE, 2009.
.. [Auckenthaler2000] Auckenthaler, Roland, Michael Carey, and Harvey Lloyd-Thomas. "Score normalization for text-independent speaker verification systems." Digital Signal Processing 10.1 (2000): 42-54.
.. [Mariethoz2005] Mariethoz, Johnny, and Samy Bengio. "A unified framework for score normalization techniques applied to text-independent speaker verification." IEEE signal processing letters 12.7 (2005): 532-535.
Indices and tables Indices and tables
......
doc/links.rst
View file @ 61a31463
...@@ -12,6 +12,7 @@ ...@@ -12,6 +12,7 @@
.. _blitz++: http://www.oonumerics.org/blitz .. _blitz++: http://www.oonumerics.org/blitz
.. _bob's idiap guide: https://gitlab.idiap.ch/bob/bob/wikis/Using-Bob-at-Idiap .. _bob's idiap guide: https://gitlab.idiap.ch/bob/bob/wikis/Using-Bob-at-Idiap
.. _bob's website: https://www.idiap.ch/software/bob .. _bob's website: https://www.idiap.ch/software/bob
.. _bob: https://www.idiap.ch/software/bob
.. _boost: http://www.boost.org .. _boost: http://www.boost.org
.. _buildbot: http://trac.buildbot.net .. _buildbot: http://trac.buildbot.net
.. _buildout: http://pypi.python.org/pypi/zc.buildout/ .. _buildout: http://pypi.python.org/pypi/zc.buildout/
......
doc/plot/plot_ISV.py 0 → 100755
View file @ 61a31463
import bob.db.iris
import bob.learn.em
import bob.learn.linear
import matplotlib.pyplot as plt
import numpy
numpy.random.seed(2) # FIXING A SEED
def train_ubm(features, n_gaussians):
"""
Train UBM
**Parameters**
features: 2D numpy array with the features
n_gaussians: Number of Gaussians
"""
input_size = features.shape[1]
kmeans_machine = bob.learn.em.KMeansMachine(int(n_gaussians), input_size)
ubm = bob.learn.em.GMMMachine(int(n_gaussians), input_size)
# The K-means clustering is firstly used to used to estimate the initial
# means, the final variances and the final weights for each gaussian
# component
kmeans_trainer = bob.learn.em.KMeansTrainer('RANDOM_NO_DUPLICATE')
bob.learn.em.train(kmeans_trainer, kmeans_machine, features)
# Getting the means, weights and the variances for each cluster. This is a
# very good estimator for the ML
(variances, weights) = kmeans_machine.get_variances_and_weights_for_each_cluster(features)
means = kmeans_machine.means
# initialize the UBM with the output of kmeans
ubm.means = means
ubm.variances = variances
ubm.weights = weights
# Creating the ML Trainer. We will adapt only the means
trainer = bob.learn.em.ML_GMMTrainer(
update_means=True, update_variances=False, update_weights=False)
bob.learn.em.train(trainer, ubm, features)
return ubm
def isv_train(features, ubm):
"""
Train U matrix
**Parameters**
features: List of :py:class:`bob.learn.em.GMMStats` organized by class
n_gaussians: UBM (:py:class:`bob.learn.em.GMMMachine`)
"""
stats = []
for user in features:
user_stats = []
for f in user:
s = bob.learn.em.GMMStats(ubm.shape[0], ubm.shape[1])
ubm.acc_statistics(f, s)
user_stats.append(s)
stats.append(user_stats)
relevance_factor = 4
subspace_dimension_of_u = 1
isvbase = bob.learn.em.ISVBase(ubm, subspace_dimension_of_u)
trainer = bob.learn.em.ISVTrainer(relevance_factor)
# trainer.rng = bob.core.random.mt19937(int(self.init_seed))
bob.learn.em.train(trainer, isvbase, stats, max_iterations=50)
return isvbase
# GENERATING DATA
data_per_class = bob.db.iris.data()
setosa = numpy.column_stack(
(data_per_class['setosa'][:, 0], data_per_class['setosa'][:, 3]))
versicolor = numpy.column_stack(
(data_per_class['versicolor'][:, 0], data_per_class['versicolor'][:, 3]))
virginica = numpy.column_stack(
(data_per_class['virginica'][:, 0], data_per_class['virginica'][:, 3]))
data = numpy.vstack((setosa, versicolor, virginica))
# TRAINING THE PRIOR
ubm = train_ubm(data, 3)
isvbase = isv_train([setosa, versicolor, virginica], ubm)
# Variability direction
u0 = isvbase.u[0:2, 0] / numpy.linalg.norm(isvbase.u[0:2, 0])
u1 = isvbase.u[2:4, 0] / numpy.linalg.norm(isvbase.u[2:4, 0])
u2 = isvbase.u[4:6, 0] / numpy.linalg.norm(isvbase.u[4:6, 0])
figure, ax = plt.subplots()
plt.scatter(setosa[:, 0], setosa[:, 1], c="darkcyan", label="setosa")
plt.scatter(versicolor[:, 0], versicolor[:, 1],
c="goldenrod", label="versicolor")
plt.scatter(virginica[:, 0], virginica[:, 1], c="dimgrey", label="virginica")
plt.scatter(ubm.means[:, 0], ubm.means[:, 1], c="blue",
marker="x", label="centroids - mle")
# plt.scatter(ubm.means[:, 0], ubm.means[:, 1], c="blue",
# marker=".", label="within class varibility", s=0.01)
ax.arrow(ubm.means[0, 0], ubm.means[0, 1], u0[0], u0[1],
fc="k", ec="k", head_width=0.05, head_length=0.1)
ax.arrow(ubm.means[1, 0], ubm.means[1, 1], u1[0], u1[1],
fc="k", ec="k", head_width=0.05, head_length=0.1)
ax.arrow(ubm.means[2, 0], ubm.means[2, 1], u2[0], u2[1],
fc="k", ec="k", head_width=0.05, head_length=0.1)
plt.text(ubm.means[0, 0] + u0[0], ubm.means[0, 1] +
u0[1] - 0.1, r'$\mathbf{U}_1$', fontsize=15)
plt.text(ubm.means[1, 0] + u1[0], ubm.means[1, 1] +
u1[1] - 0.1, r'$\mathbf{U}_2$', fontsize=15)
plt.text(ubm.means[2, 0] + u2[0], ubm.means[2, 1] +
u2[1] - 0.1, r'$\mathbf{U}_3$', fontsize=15)
plt.xticks([], [])
plt.yticks([], [])
# plt.grid(True)
plt.xlabel('Sepal length')
plt.ylabel('Petal width')
plt.legend()
plt.tight_layout()
plt.show()
doc/plot/plot_JFA.py 0 → 100755
View file @ 61a31463
import bob.db.iris
import bob.learn.em
import bob.learn.linear
import matplotlib.pyplot as plt
import numpy
numpy.random.seed(2) # FIXING A SEED
def train_ubm(features, n_gaussians):
"""
Train UBM
**Parameters**
features: 2D numpy array with the features
n_gaussians: Number of Gaussians
"""
input_size = features.shape[1]
kmeans_machine = bob.learn.em.KMeansMachine(int(n_gaussians), input_size)
ubm = bob.learn.em.GMMMachine(int(n_gaussians), input_size)
# The K-means clustering is firstly used to used to estimate the initial
# means, the final variances and the final weights for each gaussian
# component
kmeans_trainer = bob.learn.em.KMeansTrainer('RANDOM_NO_DUPLICATE')
bob.learn.em.train(kmeans_trainer, kmeans_machine, features)
# Getting the means, weights and the variances for each cluster. This is a
# very good estimator for the ML
(variances, weights) = kmeans_machine.get_variances_and_weights_for_each_cluster(features)
means = kmeans_machine.means
# initialize the UBM with the output of kmeans
ubm.means = means
ubm.variances = variances
ubm.weights = weights
# Creating the ML Trainer. We will adapt only the means
trainer = bob.learn.em.ML_GMMTrainer(
update_means=True, update_variances=False, update_weights=False)
bob.learn.em.train(trainer, ubm, features)
return ubm
def jfa_train(features, ubm):
"""
Trains U and V matrix
**Parameters**
features: List of :py:class:`bob.learn.em.GMMStats` organized by class
n_gaussians: UBM (:py:class:`bob.learn.em.GMMMachine`)
"""
stats = []
for user in features:
user_stats = []
for f in user:
s = bob.learn.em.GMMStats(ubm.shape[0], ubm.shape[1])
ubm.acc_statistics(f, s)
user_stats.append(s)
stats.append(user_stats)
subspace_dimension_of_u = 1
subspace_dimension_of_v = 1
jfa_base = bob.learn.em.JFABase(
ubm, subspace_dimension_of_u, subspace_dimension_of_v)
trainer = bob.learn.em.JFATrainer()
# trainer.rng = bob.core.random.mt19937(int(self.init_seed))
bob.learn.em.train_jfa(trainer, jfa_base, stats, max_iterations=50)
return jfa_base
# GENERATING DATA
data_per_class = bob.db.iris.data()
setosa = numpy.column_stack(
(data_per_class['setosa'][:, 0], data_per_class['setosa'][:, 3]))
versicolor = numpy.column_stack(
(data_per_class['versicolor'][:, 0], data_per_class['versicolor'][:, 3]))
virginica = numpy.column_stack(
(data_per_class['virginica'][:, 0], data_per_class['virginica'][:, 3]))
data = numpy.vstack((setosa, versicolor, virginica))
# TRAINING THE PRIOR
ubm = train_ubm(data, 3)
jfa_base = jfa_train([setosa, versicolor, virginica], ubm)
# Variability direction U
u0 = jfa_base.u[0:2, 0] / numpy.linalg.norm(jfa_base.u[0:2, 0])
u1 = jfa_base.u[2:4, 0] / numpy.linalg.norm(jfa_base.u[2:4, 0])
u2 = jfa_base.u[4:6, 0] / numpy.linalg.norm(jfa_base.u[4:6, 0])
# Variability direction V
v0 = jfa_base.v[0:2, 0] / numpy.linalg.norm(jfa_base.v[0:2, 0])
v1 = jfa_base.v[2:4, 0] / numpy.linalg.norm(jfa_base.v[2:4, 0])
v2 = jfa_base.v[4:6, 0] / numpy.linalg.norm(jfa_base.v[4:6, 0])
figure, ax = plt.subplots()
plt.scatter(setosa[:, 0], setosa[:, 1], c="darkcyan", label="setosa")
plt.scatter(versicolor[:, 0], versicolor[:, 1],
c="goldenrod", label="versicolor")
plt.scatter(virginica[:, 0], virginica[:, 1], c="dimgrey", label="virginica")
plt.scatter(ubm.means[:, 0], ubm.means[:, 1], c="blue",
marker="x", label="centroids - mle")
# plt.scatter(ubm.means[:, 0], ubm.means[:, 1], c="blue",
# marker=".", label="within class varibility", s=0.01)
# U
ax.arrow(ubm.means[0, 0], ubm.means[0, 1], u0[0], u0[1],
fc="k", ec="k", head_width=0.05, head_length=0.1)
ax.arrow(ubm.means[1, 0], ubm.means[1, 1], u1[0], u1[1],
fc="k", ec="k", head_width=0.05, head_length=0.1)
ax.arrow(ubm.means[2, 0], ubm.means[2, 1], u2[0], u2[1],
fc="k", ec="k", head_width=0.05, head_length=0.1)
plt.text(ubm.means[0, 0] + u0[0], ubm.means[0, 1] +
u0[1] - 0.1, r'$\mathbf{U}_1$', fontsize=15)
plt.text(ubm.means[1, 0] + u1[0], ubm.means[1, 1] +
u1[1] - 0.1, r'$\mathbf{U}_2$', fontsize=15)
plt.text(ubm.means[2, 0] + u2[0], ubm.means[2, 1] +
u2[1] - 0.1, r'$\mathbf{U}_3$', fontsize=15)
# V
ax.arrow(ubm.means[0, 0], ubm.means[0, 1], v0[0], v0[1],
fc="k", ec="k", head_width=0.05, head_length=0.1)
ax.arrow(ubm.means[1, 0], ubm.means[1, 1], v1[0], v1[1],
fc="k", ec="k", head_width=0.05, head_length=0.1)
ax.arrow(ubm.means[2, 0], ubm.means[2, 1], v2[0], v2[1],
fc="k", ec="k", head_width=0.05, head_length=0.1)
plt.text(ubm.means[0, 0] + v0[0], ubm.means[0, 1] +
v0[1] - 0.1, r'$\mathbf{V}_1$', fontsize=15)
plt.text(ubm.means[1, 0] + v1[0], ubm.means[1, 1] +
v1[1] - 0.1, r'$\mathbf{V}_2$', fontsize=15)
plt.text(ubm.means[2, 0] + v2[0], ubm.means[2, 1] +
v2[1] - 0.1, r'$\mathbf{V}_3$', fontsize=15)
plt.xticks([], [])
plt.yticks([], [])
# plt.grid(True)
plt.xlabel('Sepal length')
plt.ylabel('Petal width')
plt.legend(loc=2)
plt.ylim([-1, 3.5])
plt.tight_layout()
# plt.show()
doc/plot/plot_MAP.py 0 → 100644
View file @ 61a31463
import matplotlib.pyplot as plt
import bob.db.iris
import bob.learn.em
import numpy
numpy.random.seed(10)
data_per_class = bob.db.iris.data()
setosa = numpy.column_stack(
(data_per_class['setosa'][:, 0], data_per_class['setosa'][:, 3]))
versicolor = numpy.column_stack(
(data_per_class['versicolor'][:, 0], data_per_class['versicolor'][:, 3]))
virginica = numpy.column_stack(
(data_per_class['virginica'][:, 0], data_per_class['virginica'][:, 3]))
data = numpy.vstack((setosa, versicolor, virginica))
# Two clusters with a feature dimensionality of 3
mle_machine = bob.learn.em.GMMMachine(3, 2)
mle_machine.means = numpy.array([[5, 3], [4, 2], [7, 3.]])
# Creating some random data centered in
map_machine = bob.learn.em.GMMMachine(3, 2)
map_trainer = bob.learn.em.MAP_GMMTrainer(mle_machine, relevance_factor=4)
bob.learn.em.train(map_trainer, map_machine, data, max_iterations=200,
convergence_threshold=1e-5) # Train the KMeansMachine
figure, ax = plt.subplots()
# plt.scatter(data[:, 0], data[:, 1], c="olivedrab", label="new data")
plt.scatter(setosa[:, 0], setosa[:, 1], c="darkcyan", label="setosa")
plt.scatter(versicolor[:, 0], versicolor[:, 1],
c="goldenrod", label="versicolor")
plt.scatter(virginica[:, 0], virginica[:, 1],
c="dimgrey", label="virginica")
plt.scatter(mle_machine.means[:, 0],
mle_machine.means[:, 1], c="blue", marker="x",
label="prior centroids - mle", s=60)
plt.scatter(map_machine.means[:, 0], map_machine.means[:, 1], c="red",
marker="^", label="adapted centroids - map", s=60)
plt.legend()
plt.xticks([], [])
plt.yticks([], [])
ax.set_xlabel("Sepal length")
ax.set_ylabel("Petal width")
plt.tight_layout()
plt.show()
doc/plot/plot_ML.py 0 → 100644
View file @ 61a31463
import bob.learn.em
import bob.db.iris
import numpy
import matplotlib.pyplot as plt
data_per_class = bob.db.iris.data()
setosa = numpy.column_stack(
(data_per_class['setosa'][:, 0], data_per_class['setosa'][:, 3]))
versicolor = numpy.column_stack(
(data_per_class['versicolor'][:, 0], data_per_class['versicolor'][:, 3]))
virginica = numpy.column_stack(
(data_per_class['virginica'][:, 0], data_per_class['virginica'][:, 3]))
data = numpy.vstack((setosa, versicolor, virginica))
# Two clusters with a feature dimensionality of 3
machine = bob.learn.em.GMMMachine(3, 2)
trainer = bob.learn.em.ML_GMMTrainer(True, True, True)
machine.means = numpy.array([[5, 3], [4, 2], [7, 3.]])
bob.learn.em.train(trainer, machine, data, max_iterations=200,
convergence_threshold=1e-5) # Train the KMeansMachine
figure, ax = plt.subplots()
plt.scatter(setosa[:, 0], setosa[:, 1], c="darkcyan", label="setosa")
plt.scatter(versicolor[:, 0], versicolor[:, 1],
c="goldenrod", label="versicolor")
plt.scatter(virginica[:, 0], virginica[:, 1],
c="dimgrey", label="virginica")
plt.scatter(machine.means[:, 0],
machine.means[:, 1], c="blue", marker="x", label="centroids", s=60)
plt.legend()
plt.xticks([], [])
plt.yticks([], [])
ax.set_xlabel("Sepal length")
ax.set_ylabel("Petal width")
plt.tight_layout()
plt.show()
doc/plot/plot_Tnorm.py 0 → 100644
View file @ 61a31463
import matplotlib.pyplot as plt
import bob.learn.em
import numpy
numpy.random.seed(10)
n_clients = 10
n_scores_per_client = 200
# Defining some fake scores for genuines and impostors
impostor_scores = numpy.random.normal(-15.5,
5, (n_scores_per_client, n_clients))
genuine_scores = numpy.random.normal(0.5, 5, (n_scores_per_client, n_clients))
# Defining the scores for the statistics computation
t_scores = numpy.random.normal(-5., 5, (n_scores_per_client, n_clients))
# T - Normalizing
t_norm_impostors = bob.learn.em.tnorm(impostor_scores, t_scores)
t_norm_genuine = bob.learn.em.tnorm(genuine_scores, t_scores)
# PLOTTING
figure = plt.subplot(2, 1, 1)
ax = figure.axes
plt.title("Raw scores", fontsize=8)
plt.hist(impostor_scores.reshape(n_scores_per_client * n_clients),
label='Impostors', normed=True,
color='C1', alpha=0.5, bins=50)
plt.hist(genuine_scores.reshape(n_scores_per_client * n_clients),
label='Genuine', normed=True,
color='C0', alpha=0.5, bins=50)
plt.legend(fontsize=8)
plt.yticks([], [])
figure = plt.subplot(2, 1, 2)
ax = figure.axes
plt.title("T-norm scores", fontsize=8)
plt.hist(t_norm_impostors.reshape(n_scores_per_client * n_clients),
label='T-Norm Impostors', normed=True,
color='C1', alpha=0.5, bins=50)
plt.hist(t_norm_genuine.reshape(n_scores_per_client * n_clients),
label='T-Norm Genuine', normed=True,
color='C0', alpha=0.5, bins=50)
plt.legend(fontsize=8)
plt.yticks([], [])
plt.tight_layout()
plt.show()
doc/plot/plot_ZTnorm.py 0 → 100644
View file @ 61a31463
import matplotlib.pyplot as plt
import bob.learn.em
import numpy
numpy.random.seed(10)
n_clients = 10
n_scores_per_client = 200
# Defining some fake scores for genuines and impostors
impostor_scores = numpy.random.normal(-15.5,
5, (n_scores_per_client, n_clients))
genuine_scores = numpy.random.normal(0.5, 5, (n_scores_per_client, n_clients))
# Defining the scores for the statistics computation
z_scores = numpy.random.normal(-5., 5, (n_scores_per_client, n_clients))
t_scores = numpy.random.normal(-6., 5, (n_scores_per_client, n_clients))
# T-normalizing the Z-scores
zt_scores = bob.learn.em.tnorm(z_scores, t_scores)
# ZT - Normalizing
zt_norm_impostors = bob.learn.em.ztnorm(
impostor_scores, z_scores, t_scores, zt_scores)
zt_norm_genuine = bob.learn.em.ztnorm(
genuine_scores, z_scores, t_scores, zt_scores)
# PLOTTING
figure = plt.subplot(2, 1, 1)
ax = figure.axes
plt.title("Raw scores", fontsize=8)
plt.hist(impostor_scores.reshape(n_scores_per_client * n_clients),
label='Impostors', normed=True,
color='C1', alpha=0.5, bins=50)
plt.hist(genuine_scores.reshape(n_scores_per_client * n_clients),
label='Genuine', normed=True,
color='C0', alpha=0.5, bins=50)
plt.legend(fontsize=8)
plt.yticks([], [])
figure = plt.subplot(2, 1, 2)
ax = figure.axes
plt.title("T-norm scores", fontsize=8)
plt.hist(zt_norm_impostors.reshape(n_scores_per_client * n_clients),
label='T-Norm Impostors', normed=True,
color='C1', alpha=0.5, bins=50)
plt.hist(zt_norm_genuine.reshape(n_scores_per_client * n_clients),
label='T-Norm Genuine', normed=True,
color='C0', alpha=0.5, bins=50)
plt.legend(fontsize=8)
plt.yticks([], [])
plt.tight_layout()
plt.show()
doc/plot/plot_Znorm.py 0 → 100644
View file @ 61a31463
import matplotlib.pyplot as plt
import bob.learn.em
import numpy
numpy.random.seed(10)
n_clients = 10
n_scores_per_client = 200
# Defining some fake scores for genuines and impostors
impostor_scores = numpy.random.normal(-15.5,
5, (n_scores_per_client, n_clients))
genuine_scores = numpy.random.normal(0.5, 5, (n_scores_per_client, n_clients))
# Defining the scores for the statistics computation
z_scores = numpy.random.normal(-5., 5, (n_scores_per_client, n_clients))
# Z - Normalizing
z_norm_impostors = bob.learn.em.znorm(impostor_scores, z_scores)
z_norm_genuine = bob.learn.em.znorm(genuine_scores, z_scores)
# PLOTTING
figure = plt.subplot(2, 1, 1)
ax = figure.axes
plt.title("Raw scores", fontsize=8)
plt.hist(impostor_scores.reshape(n_scores_per_client * n_clients),
label='Impostors', normed=True,
color='C1', alpha=0.5, bins=50)
plt.hist(genuine_scores.reshape(n_scores_per_client * n_clients),
label='Genuine', normed=True,
color='C0', alpha=0.5, bins=50)
plt.legend(fontsize=8)
plt.yticks([], [])
figure = plt.subplot(2, 1, 2)
ax = figure.axes
plt.title("Z-norm scores", fontsize=8)
plt.hist(z_norm_impostors.reshape(n_scores_per_client * n_clients),
label='Z-Norm Impostors', normed=True,
color='C1', alpha=0.5, bins=50)
plt.hist(z_norm_genuine.reshape(n_scores_per_client * n_clients),
label='Z-Norm Genuine', normed=True,
color='C0', alpha=0.5, bins=50)
plt.yticks([], [])
plt.legend(fontsize=8)
plt.tight_layout()
plt.show()
doc/plot/plot_iVector.py 0 → 100755
View file @ 61a31463
import bob.db.iris
import bob.learn.em
import bob.learn.linear
import matplotlib.pyplot as plt
import numpy
numpy.random.seed(2) # FIXING A SEED
def train_ubm(features, n_gaussians):
"""
Train UBM
**Parameters**
features: 2D numpy array with the features
n_gaussians: Number of Gaussians
"""
input_size = features.shape[1]
kmeans_machine = bob.learn.em.KMeansMachine(int(n_gaussians), input_size)
ubm = bob.learn.em.GMMMachine(int(n_gaussians), input_size)
# The K-means clustering is firstly used to used to estimate the initial
# means, the final variances and the final weights for each gaussian
# component
kmeans_trainer = bob.learn.em.KMeansTrainer('RANDOM_NO_DUPLICATE')
bob.learn.em.train(kmeans_trainer, kmeans_machine, features)
# Getting the means, weights and the variances for each cluster. This is a
# very good estimator for the ML
(variances, weights) = kmeans_machine.get_variances_and_weights_for_each_cluster(features)
means = kmeans_machine.means
# initialize the UBM with the output of kmeans
ubm.means = means
ubm.variances = variances
ubm.weights = weights
# Creating the ML Trainer. We will adapt only the means
trainer = bob.learn.em.ML_GMMTrainer(
update_means=True, update_variances=False, update_weights=False)
bob.learn.em.train(trainer, ubm, features)
return ubm
def ivector_train(features, ubm):
"""
Trains T matrix
**Parameters**
features: List of :py:class:`bob.learn.em.GMMStats`
n_gaussians: UBM (:py:class:`bob.learn.em.GMMMachine`)
"""
stats = []
for user in features:
s = bob.learn.em.GMMStats(ubm.shape[0], ubm.shape[1])
for f in user:
ubm.acc_statistics(f, s)
stats.append(s)
subspace_dimension_of_t = 2
ivector_trainer = bob.learn.em.IVectorTrainer(update_sigma=True)
ivector_machine = bob.learn.em.IVectorMachine(
ubm, subspace_dimension_of_t, 10e-5)
# train IVector model
bob.learn.em.train(ivector_trainer, ivector_machine, stats, 500)
return ivector_machine
def acc_stats(data, gmm):
gmm_stats = []
for d in data:
s = bob.learn.em.GMMStats(gmm.shape[0], gmm.shape[1])
gmm.acc_statistics(d, s)
gmm_stats.append(s)
return gmm_stats
def compute_ivectors(gmm_stats, ivector_machine):
"""
Given :py:class:`bob.learn.em.GMMStats` and an T matrix, get the iVectors.
"""
ivectors = []
for g in gmm_stats:
ivectors.append(ivector_machine(g))
return numpy.array(ivectors)
# GENERATING DATA
data_per_class = bob.db.iris.data()
setosa = numpy.column_stack(
(data_per_class['setosa'][:, 0], data_per_class['setosa'][:, 3]))
versicolor = numpy.column_stack(
(data_per_class['versicolor'][:, 0], data_per_class['versicolor'][:, 3]))
virginica = numpy.column_stack(
(data_per_class['virginica'][:, 0], data_per_class['virginica'][:, 3]))
data = numpy.vstack((setosa, versicolor, virginica))
# TRAINING THE PRIOR
ubm = train_ubm(data, 3)
ivector_machine = ivector_train([setosa, versicolor, virginica], ubm)
# Variability direction U
# t0 = T[0:2, 0] / numpy.linalg.norm(T[0:2, 0])
# t1 = T[2:4, 0] / numpy.linalg.norm(T[2:4, 0])
# t2 = T[4:6, 0] / numpy.linalg.norm(T[4:6, 0])
# figure, ax = plt.subplots()
figure = plt.subplot(2, 1, 1)
ax = figure.axes
plt.title("Raw fetures")
plt.scatter(setosa[:, 0], setosa[:, 1], c="darkcyan", label="setosa")
plt.scatter(versicolor[:, 0], versicolor[:, 1],
c="goldenrod", label="versicolor")
plt.scatter(virginica[:, 0], virginica[:, 1], c="dimgrey", label="virginica")
# plt.grid(True)
# plt.xlabel('Sepal length')
plt.ylabel('Petal width')
plt.legend(loc=2)
plt.ylim([-1, 3.5])
plt.xticks([], [])
plt.yticks([], [])
figure = plt.subplot(2, 1, 2)
ax = figure.axes
ivector_setosa = compute_ivectors(acc_stats(setosa, ubm), ivector_machine)
ivector_versicolor = compute_ivectors(
acc_stats(versicolor, ubm), ivector_machine)
ivector_virginica = compute_ivectors(
acc_stats(virginica, ubm), ivector_machine)
# Whitening iVectors
whitening_trainer = bob.learn.linear.WhiteningTrainer()
whitener_machine = bob.learn.linear.Machine(
ivector_setosa.shape[1], ivector_setosa.shape[1])
whitening_trainer.train(numpy.vstack(
(ivector_setosa, ivector_versicolor, ivector_virginica)), whitener_machine)
ivector_setosa = whitener_machine(ivector_setosa)
ivector_versicolor = whitener_machine(ivector_versicolor)
ivector_virginica = whitener_machine(ivector_virginica)
# LDA ivectors
lda_trainer = bob.learn.linear.FisherLDATrainer()
lda_machine = bob.learn.linear.Machine(
ivector_setosa.shape[1], ivector_setosa.shape[1])
lda_trainer.train([ivector_setosa, ivector_versicolor,
ivector_virginica], lda_machine)
ivector_setosa = lda_machine(ivector_setosa)
ivector_versicolor = lda_machine(ivector_versicolor)
ivector_virginica = lda_machine(ivector_virginica)
# WCCN ivectors
# wccn_trainer = bob.learn.linear.WCCNTrainer()
# wccn_machine = bob.learn.linear.Machine(
# ivector_setosa.shape[1], ivector_setosa.shape[1])
# wccn_trainer.train([ivector_setosa, ivector_versicolor,
# ivector_virginica], wccn_machine)
# ivector_setosa = wccn_machine(ivector_setosa)
# ivector_versicolor = wccn_machine(ivector_versicolor)
# ivector_virginica = wccn_machine(ivector_virginica)
plt.title("First two ivectors")
plt.scatter(ivector_setosa[:, 0],
ivector_setosa[:, 1], c="darkcyan", label="setosa",
marker="x")
plt.scatter(ivector_versicolor[:, 0],
ivector_versicolor[:, 1], c="goldenrod", label="versicolor",
marker="x")
plt.scatter(ivector_virginica[:, 0],
ivector_virginica[:, 1], c="dimgrey", label="virginica",
marker="x")
plt.xticks([], [])
plt.yticks([], [])
# plt.grid(True)
# plt.xlabel('Sepal length')
# plt.ylabel('Petal width')
plt.legend(loc=2)
plt.ylim([-1, 3.5])
plt.tight_layout()
plt.show()
doc/plot/plot_kmeans.py 0 → 100644
View file @ 61a31463
import bob.learn.em
import bob.db.iris
import numpy
import matplotlib.pyplot as plt
data_per_class = bob.db.iris.data()
setosa = numpy.column_stack(
(data_per_class['setosa'][:, 0], data_per_class['setosa'][:, 3]))
versicolor = numpy.column_stack(
(data_per_class['versicolor'][:, 0], data_per_class['versicolor'][:, 3]))
virginica = numpy.column_stack(
(data_per_class['virginica'][:, 0], data_per_class['virginica'][:, 3]))
data = numpy.vstack((setosa, versicolor, virginica))
# Training KMeans
# Two clusters with a feature dimensionality of 3
machine = bob.learn.em.KMeansMachine(3, 2)
trainer = bob.learn.em.KMeansTrainer()
bob.learn.em.train(trainer, machine, data, max_iterations=200,
convergence_threshold=1e-5) # Train the KMeansMachine
# Plotting
figure, ax = plt.subplots()
plt.scatter(setosa[:, 0],
setosa[:, 1], c="darkcyan", label="setosa")
plt.scatter(versicolor[:, 0],
versicolor[:, 1], c="goldenrod", label="versicolor")
plt.scatter(virginica[:, 0],
virginica[:, 1], c="dimgrey", label="virginica")
plt.scatter(machine.means[:, 0],
machine.means[:, 1], c="blue", marker="x", label="centroids",
s=60)
plt.legend()
plt.xticks([], [])
plt.yticks([], [])
ax.set_xlabel("Sepal length")
ax.set_ylabel("Petal width")
plt.tight_layout()
doc/py_api.rst
View file @ 61a31463
...@@ -17,21 +17,21 @@ Trainers ...@@ -17,21 +17,21 @@ Trainers
........ ........
.. autosummary:: .. autosummary::
bob.learn.em.KMeansTrainer bob.learn.em.KMeansTrainer
bob.learn.em.ML_GMMTrainer bob.learn.em.ML_GMMTrainer
bob.learn.em.MAP_GMMTrainer bob.learn.em.MAP_GMMTrainer
bob.learn.em.ISVTrainer bob.learn.em.ISVTrainer
bob.learn.em.JFATrainer bob.learn.em.JFATrainer
bob.learn.em.IVectorTrainer bob.learn.em.IVectorTrainer
bob.learn.em.PLDATrainer bob.learn.em.PLDATrainer
bob.learn.em.EMPCATrainer bob.learn.em.EMPCATrainer
Machines Machines
........ ........
.. autosummary:: .. autosummary::
bob.learn.em.KMeansMachine bob.learn.em.KMeansMachine
bob.learn.em.Gaussian bob.learn.em.Gaussian
bob.learn.em.GMMStats bob.learn.em.GMMStats
...@@ -43,7 +43,7 @@ Machines ...@@ -43,7 +43,7 @@ Machines
bob.learn.em.IVectorMachine bob.learn.em.IVectorMachine
bob.learn.em.PLDABase bob.learn.em.PLDABase
bob.learn.em.PLDAMachine bob.learn.em.PLDAMachine
Functions Functions
--------- ---------
.. autosummary:: .. autosummary::
......