NextCloud/dlib
1
0
Fork 0
mirror of https://github.com/davisking/dlib.git synced 2024-11-01 10:14:53 +08:00
Code Issues Packages Projects Releases Wiki Activity

Added the option to use an identity matrix prior to the vector_normalizer_frobmetric

Browse Source 
object.
This commit is contained in:
Davis King 2014-09-14 11:27:36 -04:00
parent f4e50eaaec
commit a0a692f91b
3 changed files with 125 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

71
dlib/statistics/vector_normalizer_frobmetric.h
View File

@ -81,8 +81,9 @@ namespace dlib
{
objective (
const std::vector<compact_frobmetric_training_sample>& samples_,
matrix<double,0,0,mem_manager_type>& Aminus_
) : samples(samples_), Aminus(Aminus_) {}
matrix<double,0,0,mem_manager_type>& Aminus_,
const matrix<double,0,1,mem_manager_type>& bias_
) : samples(samples_), Aminus(Aminus_), bias(bias_) {}
double operator()(const matrix<double,0,1,mem_manager_type>& u) const
{
@ -118,12 +119,13 @@ namespace dlib
// computation can make Aminus slightly non-symmetric.
Aminus = make_symmetric(Aminus);
return sum(u) - 0.5*sum(squared(Aminus));
return dot(u,bias) - 0.5*sum(squared(Aminus));
}
private:
const std::vector<compact_frobmetric_training_sample>& samples;
matrix<double,0,0,mem_manager_type>& Aminus;
const matrix<double,0,1,mem_manager_type>& bias;
};
struct derivative
@ -131,8 +133,9 @@ namespace dlib
derivative (
unsigned long num_triples_,
const std::vector<compact_frobmetric_training_sample>& samples_,
matrix<double,0,0,mem_manager_type>& Aminus_
) : num_triples(num_triples_), samples(samples_), Aminus(Aminus_) {}
matrix<double,0,0,mem_manager_type>& Aminus_,
const matrix<double,0,1,mem_manager_type>& bias_
) : num_triples(num_triples_), samples(samples_), Aminus(Aminus_), bias(bias_) {}
matrix<double,0,1,mem_manager_type> operator()(const matrix<double,0,1,mem_manager_type>& ) const
{
@ -158,7 +161,8 @@ namespace dlib
{
for (unsigned long k = 0; k < samples[i].far_vects.size(); ++k)
{
grad(idx++) = 1 + ufar[k]-unear[j];
grad(idx) = bias(idx) + ufar[k]-unear[j];
idx++;
}
}
}
@ -170,6 +174,7 @@ namespace dlib
const unsigned long num_triples;
const std::vector<compact_frobmetric_training_sample>& samples;
matrix<double,0,0,mem_manager_type>& Aminus;
const matrix<double,0,1,mem_manager_type>& bias;
};
@ -245,6 +250,20 @@ namespace dlib
eps = 0.1;
C = 1;
max_iter = 5000;
_use_identity_matrix_prior = false;
}
bool uses_identity_matrix_prior (
) const
{
return _use_identity_matrix_prior;
}
void set_uses_identity_matrix_prior (
bool use_prior
)
{
_use_identity_matrix_prior = use_prior;
}
void be_verbose(
@ -402,27 +421,50 @@ namespace dlib
num_triples += samples[i].near_vects.size()*samples[i].far_vects.size();
matrix<double,0,1,mem_manager_type> u(num_triples);
matrix<double,0,1,mem_manager_type> bias(num_triples);
u = 0;
bias = 1;
// precompute all the anchor_vect to far_vects/near_vects pairs
std::vector<compact_frobmetric_training_sample> data(samples.size());
unsigned long cnt = 0;
std::vector<double> far_norm, near_norm;
for (unsigned long i = 0; i < data.size(); ++i)
{
far_norm.clear();
near_norm.clear();
data[i].far_vects.reserve(samples[i].far_vects.size());
data[i].near_vects.reserve(samples[i].near_vects.size());
for (unsigned long j = 0; j < samples[i].far_vects.size(); ++j)
{
data[i].far_vects.push_back(samples[i].anchor_vect - samples[i].far_vects[j]);
if (_use_identity_matrix_prior)
far_norm.push_back(length_squared(data[i].far_vects.back()));
}
for (unsigned long j = 0; j < samples[i].near_vects.size(); ++j)
{
data[i].near_vects.push_back(samples[i].anchor_vect - samples[i].near_vects[j]);
if (_use_identity_matrix_prior)
near_norm.push_back(length_squared(data[i].near_vects.back()));
}
// Note that this loop only executes if _use_identity_matrix_prior == true.
for (unsigned long j = 0; j < near_norm.size(); ++j)
{
for (unsigned long k = 0; k < far_norm.size(); ++k)
{
bias(cnt++) = 1 - (far_norm[k] - near_norm[j]);
}
}
}
// Now run the main part of the algorithm
matrix<double,0,0,mem_manager_type> Aminus;
find_max_box_constrained(lbfgs_search_strategy(10),
custom_stop_strategy(C, eps, verbose, max_iter),
objective(data, Aminus),
derivative(num_triples, data, Aminus),
objective(data, Aminus, bias),
derivative(num_triples, data, Aminus, bias),
u, 0, C/num_triples);
@ -437,6 +479,9 @@ namespace dlib
if (eigs(i) < tol)
eigs(i) = 0;
}
if (_use_identity_matrix_prior)
tform = matrix_cast<scalar_type>(identity_matrix(Aminus) + diagm(sqrt(eigs))*trans(ed.get_pseudo_v()));
else
tform = matrix_cast<scalar_type>(diagm(sqrt(eigs))*trans(ed.get_pseudo_v()));
// Pre-apply the transform to m so we don't have to do it inside operator()
@ -509,6 +554,7 @@ namespace dlib
double eps;
double C;
unsigned long max_iter;
bool _use_identity_matrix_prior;
// This is just a temporary variable that doesn't contribute to the
// state of this object.
@ -525,7 +571,7 @@ namespace dlib
std::ostream& out
)
{
const int version = 1;
const int version = 2;
serialize(version, out);
serialize(item.m, out);
@ -534,6 +580,7 @@ namespace dlib
serialize(item.eps, out);
serialize(item.C, out);
serialize(item.max_iter, out);
serialize(item._use_identity_matrix_prior, out);
}
// ----------------------------------------------------------------------------------------
@ -548,7 +595,7 @@ namespace dlib
{
int version = 0;
deserialize(version, in);
if (version != 1)
if (version != 1 && version != 2)
throw serialization_error("Unsupported version found while deserializing dlib::vector_normalizer_frobmetric.");
deserialize(item.m, in);
@ -557,6 +604,10 @@ namespace dlib
deserialize(item.eps, in);
deserialize(item.C, in);
deserialize(item.max_iter, in);
if (version == 2)
deserialize(item._use_identity_matrix_prior, in);
else
item._use_identity_matrix_prior = false;
}
// ----------------------------------------------------------------------------------------

22
dlib/statistics/vector_normalizer_frobmetric_abstract.h
View File

@ -73,6 +73,7 @@ namespace dlib
- get_c() == 1
- get_max_iterations() == 5000
- This object is not verbose
- uses_identity_matrix_prior() == false
WHAT THIS OBJECT REPRESENTS
This object is a tool for performing the FrobMetric distance metric
@ -110,6 +111,27 @@ namespace dlib
- this object is properly initialized
!*/
bool uses_identity_matrix_prior (
) const;
/*!
ensures
- Normally this object will try and find a matrix transform() that
minimizes sum(squared(transform())) but also fits the training data.
However, if #uses_identity_matrix_prior() == true then it will instead
try to find the transformation matrix that minimizes
sum(squared(identity_matrix()-transform())). That is, it will try to
find the matrix most similar to the identity matrix that best fits the
training data.
!*/
void set_uses_identity_matrix_prior (
bool use_prior
);
/*!
ensures
- #uses_identity_matrix_prior() == use_prior
!*/
void be_verbose(
);
/*!

41
dlib/test/statistics.cpp
View File

@ -692,10 +692,51 @@ namespace
}
void prior_frobnorm_test()
{
frobmetric_training_sample<matrix<double,0,1> > sample;
std::vector<frobmetric_training_sample<matrix<double,0,1> > > samples;
matrix<double,3,1> x, near, far;
x = 0,0,0;
near = 1,0,0;
far = 0,1,0;
sample.anchor_vect = x;
sample.near_vects.push_back(near);
sample.far_vects.push_back(far);
samples.push_back(sample);
vector_normalizer_frobmetric<matrix<double,0,1> > trainer;
trainer.set_c(100);
print_spinner();
trainer.train(samples);
matrix<double,3,3> correct;
correct = 0, 0, 0,
0, 1, 0,
0, 0, 0;
dlog << LDEBUG << trainer.transform();
DLIB_TEST(max(abs(trainer.transform()-correct)) < 1e-8);
trainer.set_uses_identity_matrix_prior(true);
print_spinner();
trainer.train(samples);
correct = 1, 0, 0,
0, 2, 0,
0, 0, 1;
dlog << LDEBUG << trainer.transform();
DLIB_TEST(max(abs(trainer.transform()-correct)) < 1e-8);
}
void perform_test (
)
{
prior_frobnorm_test();
dlib::rand rnd;
for (int i = 0; i < 5; ++i)
test_vector_normalizer_frobmetric(rnd);