BEAT - tutorial/isv/2

This algorithm is a legacy one. The API has changed since its implementation. New versions and forks will need to be updated.

Endpoint Groups 1

Algorithms have at least one input and one output. All algorithm endpoints are organized in groups. Groups are used by the platform to indicate which inputs and outputs are synchronized together. The first group is automatically synchronized with the channel defined by the block in which the algorithm is deployed.

Unnamed group

Endpoint Name	Data Format	Nature
ubm	tutorial/gmm/1	Input
statistics	tutorial/gmm_statistics/1	Input
client_id	system/uint64/1	Input
isvbase	tutorial/isvbase/1	Output

Parameters 4

Parameters allow users to change the configuration of an algorithm when scheduling an experiment

Name	Type	Default
isv-training-iterations	uint32	10
relevance-factor	float64	4.0
subspace-dimension-of-u	uint32	50
init-seed	uint32	0

import bob
import numpy

def gmm_from_data(data):
    """Unmangles a bob.machine.GMMMachine from a BEAT Data object"""

dim_c, dim_d = data.means.shape
    gmm = bob.machine.GMMMachine(dim_c, dim_d)
    gmm.weights = data.weights
    gmm.means = data.means
    gmm.variances = data.variances
    gmm.variance_thresholds = data.variance_thresholds
    return gmm

def stats_from_data(data):
    """Unmangles a bob.machine.GMMStats from a BEAT Data object"""

dim_c, dim_d = data.sum_px.shape
    stat = bob.machine.GMMStats(dim_c, dim_d)
    stat.t = long(data.t)
    stat.n = data.n
    stat.sum_px = data.sum_px
    stat.sum_pxx = data.sum_pxx
    return stat

class Algorithm:

def __init__(self):
        self.isv_training_iterations = 10
        self.relevance_factor = 4.
        self.subspace_dimension_of_u = 50
        self.init_seed = 0
        self.data = {}
        self.ubm = None

def setup(self, parameters):
        self.isv_training_iterations = parameters.get('isv-training-iterations', self.isv_training_iterations)
        self.relevance_factor = parameters.get('relevance-factor', self.relevance_factor)
        self.subspace_dimension_of_u = parameters.get('subspace-dimension-of-u', self.subspace_dimension_of_u)
        self.init_seed = parameters.get('init-seed', self.init_seed)
        return True

def process(self, inputs, outputs):

# retrieve the UBM once
        if self.ubm is None:
            self.ubm = gmm_from_data(inputs['ubm'].data)

stats = stats_from_data(inputs["statistics"].data)
        c_id = inputs["client_id"].data.value
        if c_id in self.data.keys(): self.data[c_id].append(stats)
        else: self.data[c_id] = [stats]

if not(inputs.hasMoreData()):
            # create array set used for training
            training_set = [v for k,v in self.data.iteritems()]

isvbase = bob.machine.ISVBase(self.ubm, int(self.subspace_dimension_of_u))
            trainer = bob.trainer.ISVTrainer(int(self.isv_training_iterations), self.relevance_factor)
            trainer.rng = bob.core.random.mt19937(int(self.init_seed))
            trainer.train(isvbase, training_set)

# outputs data
            outputs["isvbase"].write({
                'subspace_u':              isvbase.u,
                'subspace_d':              isvbase.d,
            })

return True

xxxxxxxxxx
 
import bob
import numpy
​
​
def gmm_from_data(data):
    """Unmangles a bob.machine.GMMMachine from a BEAT Data object"""
​
    dim_c, dim_d = data.means.shape
    gmm = bob.machine.GMMMachine(dim_c, dim_d)
    gmm.weights = data.weights
    gmm.means = data.means
    gmm.variances = data.variances
    gmm.variance_thresholds = data.variance_thresholds
    return gmm
​
​
​
def stats_from_data(data):
    """Unmangles a bob.machine.GMMStats from a BEAT Data object"""
​
    dim_c, dim_d = data.sum_px.shape
    stat = bob.machine.GMMStats(dim_c, dim_d)
    stat.t = long(data.t)
    stat.n = data.n
    stat.sum_px = data.sum_px
    stat.sum_pxx = data.sum_pxx
    return stat
​
​
class Algorithm:
​
    def __init__(self):
        self.isv_training_iterations = 10
        self.relevance_factor = 4.
        self.subspace_dimension_of_u = 50
        self.init_seed = 0
        self.data = {}
        self.ubm = None
​
    def setup(self, parameters):
        self.isv_training_iterations = parameters.get('isv-training-iterations', self.isv_training_iterations)
        self.relevance_factor = parameters.get('relevance-factor', self.relevance_factor)
        self.subspace_dimension_of_u = parameters.get('subspace-dimension-of-u', self.subspace_dimension_of_u)
        self.init_seed = parameters.get('init-seed', self.init_seed)
        return True
​
​
    def process(self, inputs, outputs):
​
        # retrieve the UBM once
        if self.ubm is None:
            self.ubm = gmm_from_data(inputs['ubm'].data)
​
        stats = stats_from_data(inputs["statistics"].data)
        c_id = inputs["client_id"].data.value
        if c_id in self.data.keys(): self.data[c_id].append(stats)
        else: self.data[c_id] = [stats]
​
        if not(inputs.hasMoreData()):
            # create array set used for training
            training_set = [v for k,v in self.data.iteritems()]
​
            isvbase = bob.machine.ISVBase(self.ubm, int(self.subspace_dimension_of_u))
            trainer = bob.trainer.ISVTrainer(int(self.isv_training_iterations), self.relevance_factor)
            trainer.rng = bob.core.random.mt19937(int(self.init_seed))
            trainer.train(isvbase, training_set)
​
            # outputs data
            outputs["isvbase"].write({
                'subspace_u':              isvbase.u,
                'subspace_d':              isvbase.d,
            })
​
        return True
​

The code for this algorithm in Python
The ruler at 80 columns indicate suggested POSIX line breaks (for readability).
The editor will automatically enlarge to accomodate the entirety of your input
Use keyboard shortcuts for search/replace and faster editing. For example, use Ctrl-F (PC) or Cmd-F (Mac) to search through this box

For a Gaussian Mixture Models (GMM) mean supervector space, computes the within-class variability subspace (U subspace) described in [McCool2013]:

This algorithm relies on the Bob library.

The inputs are:

statistics: A training set of GMM Statistics.
ubm: A GMM corresponding to the Universal Background Model.
client_id: Client (class/subject) identifier as an unsigned 64 bits integer.

The outputs are subspace_u and subspace_d for the session and the client offset respectivelly.

[McCool2013]

McCool, Christopher, et al. "Session variability modelling for face authentication." IET biometrics 2.3 (2013): 117-129.

Experiments

Attestation:

Privacy:

Status:

Name	Databases/Protocols	Analyzers
smarcel/tutorial/full_isv/2/mobio_male-gmm_100Gx10I-isv_50Ux10Ix4R-dct_12Bx8Ox45C-seed101	mobio/1@male	tutorial/eerhter_postperf_iso/1
tutorial/tutorial/full_isv/2/bancaMc_isv_DCT12x8_100G_U50	banca/1@Mc	tutorial/eerhter_postperf_iso/1
tutorial/tutorial/full_isv/2/xm2vtsLp1_isv_DCT12x8_100G_U50	xm2vts/1@lp1	tutorial/eerhter_postperf_iso/1
tutorial/tutorial/full_isv/2/mobioMale_isv_DCT12x8_100G_U50	mobio/1@male	tutorial/eerhter_postperf_iso/1
tutorial/tutorial/full_isv/2/bancaP_isv_DCT12x8_100G_U50	banca/1@P	tutorial/eerhter_postperf_iso/1
tutorial/tutorial/full_isv/2/atnt_isv_DCT12x8_100G_U50	atnt/1@idiap_test_eyepos	tutorial/eerhter_postperf_iso/1

This table shows the number of times this algorithm has been successfully run using the given environment. Note this does not provide sufficient information to evaluate if the algorithm will run when submitted to different conditions.

algorithms tutorial isv 2

Endpoint Groups 1

Unnamed group

Parameters 4

Experiments

algorithms

tutorial

isv

2