Skip to content

D
dlib
Commit 7fb32196 authored by Davis King's avatar Davis King
Browse files
Options

I just split the program up into a bunch of files. 

--HG--
extra : convert_revision : svn%3Afdd8eb12-d10e-0410-9acb-85c331704f74/trunk%404107
parent f2ec4594
Show whitespace changes
Inline Side-by-side
Showing with 597 additions and 424 deletions
tools/mltool/CMakeLists.txt
View file @ 7fb32196
...@@ -15,6 +15,11 @@ PROJECT(${target_name}) ...@@ -15,6 +15,11 @@ PROJECT(${target_name})
# cmake that they are part of our target (which is the exectuable named mltool) # cmake that they are part of our target (which is the exectuable named mltool)
ADD_EXECUTABLE(${target_name} ADD_EXECUTABLE(${target_name}
src/main.cpp src/main.cpp
src/regression.cpp
src/regression.h
src/common.h
src/option_range.h
src/option_range.cpp
) )
......
tools/mltool/src/common.h 0 → 100644
View file @ 7fb32196
// The contents of this file are in the public domain.
// See LICENSE_FOR_EXAMPLE_PROGRAMS.txt (in trunk/examples)
// Authors:
// Gregory Sharp
// Davis King
#ifndef DLIB_MLTOOL_COMMoN_H__
#include "dlib/cmd_line_parser.h"
#include <map>
#include "dlib/matrix.h"
typedef dlib::cmd_line_parser<char>::check_1a_c clp;
typedef std::map<unsigned long, double> sparse_sample_type;
typedef dlib::matrix<sparse_sample_type::value_type::second_type,0,1> dense_sample_type;
#endif // DLIB_MLTOOL_COMMoN_H__
tools/mltool/src/main.cpp
View file @ 7fb32196
// The contents of this file are in the public domain. // The contents of this file are in the public domain.
// See LICENSE_FOR_EXAMPLE_PROGRAMS.txt (in dlib) // See LICENSE_FOR_EXAMPLE_PROGRAMS.txt (in trunk/examples)
// Authors:
// Gregory Sharp
// Davis King
/* /*
This is a command line program that can try different regression This is a command line program that can try different regression
algorithms on a libsvm-formatted data set. algorithms on a libsvm-formatted data set.
*/ */
#include "regression.h"
#include <iostream> #include <iostream>
#include <map> #include <map>
#include <vector> #include <vector>
#include <math.h>
#include <float.h>
#include "dlib/cmd_line_parser.h"
#include "dlib/data_io.h" #include "dlib/data_io.h"
#include "dlib/mlp.h"
#include "dlib/svm.h" #include "dlib/svm.h"
using namespace dlib; using namespace dlib;
typedef dlib::cmd_line_parser<char>::check_1a_c clp; // ----------------------------------------------------------------------------------------
typedef std::map<unsigned long, double> sparse_sample_type;
typedef dlib::matrix< sparse_sample_type::value_type::second_type,0,1
> dense_sample_type;
/* exp10() is not in C/C++ standard */
double
exp10_ (double m)
{
return exp (2.3025850929940456840179914546844 * m);
}
/* ---------------------------------------------------------------------
option_range class
--------------------------------------------------------------------- */
struct option_range {
public:
bool log_range;
float min_value;
float max_value;
float incr;
public:
option_range () {
log_range = false;
min_value = 0;
max_value = 100;
incr = 10;
}
void set_option (clp& parser, std::string const& option,
float default_val);
float get_min_value ();
float get_max_value ();
float get_next_value (float curr_val);
};
void
option_range::set_option (
clp& parser,
std::string const& option,
float default_val
)
{
int rc;
/* No option specified */
if (!parser.option (option)) {
log_range = 0;
min_value = default_val;
max_value = default_val;
incr = 1;
return;
}
/* Range specified */
rc = sscanf (parser.option(option).argument().c_str(), "%f:%f:%f",
&min_value, &incr, &max_value);
if (rc == 3) {
log_range = 1;
return;
}
/* Single value specified */
if (rc == 1) {
log_range = 0;
max_value = min_value;
incr = 1;
return;
}
else {
std::cerr << "Error parsing option" << option << "\n";
exit (-1);
}
}
float
option_range::get_min_value ()
{
if (log_range) {
return exp10_ (min_value);
} else {
return min_value;
}
}
float
option_range::get_max_value ()
{
if (log_range) {
return exp10_ (max_value);
} else {
return max_value;
}
}
float
option_range::get_next_value (float curr_value)
{
if (log_range) {
curr_value = log10 (curr_value);
curr_value += incr;
curr_value = exp10_ (curr_value);
} else {
curr_value += incr;
}
return curr_value;
}
/* ---------------------------------------------------------------------
global functions
--------------------------------------------------------------------- */
static void static void
parse_args (clp& parser, int argc, char* argv[]) parse_args (clp& parser, int argc, char* argv[])
{ {
...@@ -141,7 +36,7 @@ parse_args (clp& parser, int argc, char* argv[]) ...@@ -141,7 +36,7 @@ parse_args (clp& parser, int argc, char* argv[])
"Learning kernel (for krls,krr,svr methods): {lin,rbk}.",1); "Learning kernel (for krls,krr,svr methods): {lin,rbk}.",1);
parser.add_option ("in","A libsvm-formatted file to test.",1); parser.add_option ("in","A libsvm-formatted file to test.",1);
parser.add_option ("normalize", parser.add_option ("normalize",
"Normalize the sample inputs to zero-mean unit variance?"); "Normalize the sample inputs to zero-mean unit variance.");
parser.add_option ("train-best", parser.add_option ("train-best",
"Train and save a network using best parameters", 1); "Train and save a network using best parameters", 1);
...@@ -169,6 +64,17 @@ parse_args (clp& parser, int argc, char* argv[]) ...@@ -169,6 +64,17 @@ parse_args (clp& parser, int argc, char* argv[])
const char* one_time_opts[] = {"a", "h", "help", "in"}; const char* one_time_opts[] = {"a", "h", "help", "in"};
parser.check_one_time_options(one_time_opts); parser.check_one_time_options(one_time_opts);
const char* valid_kernels[] = {"lin", "rbk"};
const char* valid_algs[] = {"krls", "krr", "mlp", "svr"};
parser.check_option_arg_range("a", valid_algs);
parser.check_option_arg_range("k", valid_kernels);
parser.check_option_arg_range("rbk-gamma", 1e-200, 1e200);
parser.check_option_arg_range("krls-tolerance", 1e-200, 1e200);
parser.check_option_arg_range("mlp-hidden-units", 1, 10000000);
parser.check_option_arg_range("mlp-num-iterations", 1, 10000000);
parser.check_option_arg_range("svr-c", 1e-200, 1e200);
parser.check_option_arg_range("svr-epsilon-insensitivity", 1e-200, 1e200);
// Check if the -h option was given // Check if the -h option was given
if (parser.option("h") || parser.option("help")) { if (parser.option("h") || parser.option("help")) {
std::cout << "Usage: dlib_test [-a algorithm] --in input_file\n"; std::cout << "Usage: dlib_test [-a algorithm] --in input_file\n";
...@@ -196,314 +102,7 @@ parse_args (clp& parser, int argc, char* argv[]) ...@@ -196,314 +102,7 @@ parse_args (clp& parser, int argc, char* argv[])
} }
} }
static const char* // ----------------------------------------------------------------------------------------
get_kernel (
clp& parser
)
{
const char* kernel = "rbk";
if (parser.option ("k")) {
kernel = parser.option("k").argument().c_str();
}
return kernel;
}
static void
get_rbk_gamma (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
option_range& range
) {
float default_gamma = 3.0 / compute_mean_squared_distance (
randomly_subsample (dense_samples, 2000));
range.set_option (parser, "rbk-gamma", default_gamma);
}
static void
get_krls_tolerance (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
option_range& range
)
{
float default_krls_tolerance = 0.001;
range.set_option (parser, "krls-tolerance", default_krls_tolerance);
}
static double
get_mlp_hidden_units (
clp& parser,
std::vector<dense_sample_type>& dense_samples
)
{
int num_hidden = 5;
if (parser.option ("mlp-hidden-units")) {
num_hidden = sa = parser.option("mlp-hidden-units").argument();
}
return num_hidden;
}
static double
get_mlp_num_iterations (
clp& parser,
std::vector<dense_sample_type>& dense_samples
)
{
int num_iterations = 5000;
if (parser.option ("mlp-num-iterations")) {
num_iterations = sa = parser.option("mlp-num-iterations").argument();
}
return num_iterations;
}
static void
get_svr_c (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
option_range& range
)
{
float default_svr_c = 1000.;
range.set_option (parser, "svr-c", default_svr_c);
}
static double
get_svr_epsilon_insensitivity (
clp& parser,
std::vector<dense_sample_type>& dense_samples
)
{
// Epsilon-insensitive regression means we do regression but stop
// trying to fit a data point once it is "close enough" to its
// target value. This parameter is the value that controls what
// we mean by "close enough". In this case, I'm saying I'm happy
// if the resulting regression function gets within 0.001 of the
// target value.
double epsilon_insensitivity = 0.001;
if (parser.option ("svr-epsilon-insensitivity")) {
epsilon_insensitivity
= sa = parser.option("svr-epsilon-insensitivity").argument();
}
return epsilon_insensitivity;
}
static void
krls_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
)
{
typedef radial_basis_kernel<dense_sample_type> kernel_type;
option_range gamma_range, krls_tol_range;
get_rbk_gamma (parser, dense_samples, gamma_range);
get_krls_tolerance (parser, dense_samples, krls_tol_range);
// Split into training set and testing set
float training_pct = 0.8;
unsigned int training_samples = (unsigned int) floor (
training_pct * dense_samples.size());
for (float krls_tol = krls_tol_range.get_min_value();
krls_tol <= krls_tol_range.get_max_value();
krls_tol = krls_tol_range.get_next_value (krls_tol))
{
for (float gamma = gamma_range.get_min_value();
gamma <= gamma_range.get_max_value();
gamma = gamma_range.get_next_value (gamma))
{
krls<kernel_type> net (kernel_type(gamma), krls_tol);
// Krls doesn't seem to come with any batch training function
for (unsigned int j = 0; j < training_samples; j++) {
net.train (dense_samples[j], labels[j]);
}
// Test the performance (sorry, no cross-validation)
double total_err = 0.0;
for (unsigned int j = training_samples + 1;
j < dense_samples.size(); j++)
{
double diff = net(dense_samples[j]) - labels[j];
total_err += diff * diff;
}
double testset_error = total_err
/ (dense_samples.size() - training_samples);
printf ("%3.6f %3.6f %3.9f\n", krls_tol, gamma, testset_error);
}
}
}
static void
krr_rbk_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
)
{
typedef radial_basis_kernel<dense_sample_type> kernel_type;
krr_trainer<kernel_type> trainer;
option_range gamma_range;
double best_gamma = DBL_MAX;
float best_loo = FLT_MAX;
get_rbk_gamma (parser, dense_samples, gamma_range);
for (float gamma = gamma_range.get_min_value();
gamma <= gamma_range.get_max_value();
gamma = gamma_range.get_next_value (gamma))
{
// LOO cross validation
double loo_error;
if (parser.option("verbose")) {
trainer.set_search_lambdas(logspace(-9, 4, 100));
trainer.be_verbose();
}
trainer.set_kernel (kernel_type (gamma));
trainer.train (dense_samples, labels, loo_error);
if (loo_error < best_loo) {
best_loo = loo_error;
best_gamma = gamma;
}
printf ("10^%f %9.6f\n", log10(gamma), loo_error);
}
printf ("Best result: gamma=10^%f (%g), loo_error=%9.6f\n",
log10(best_gamma), best_gamma, best_loo);
if (parser.option("train-best")) {
printf ("Training network with best parameters\n");
trainer.set_kernel (kernel_type (best_gamma));
decision_function<kernel_type> best_network =
trainer.train (dense_samples, labels);
std::ofstream fout (parser.option("train-best").argument().c_str(),
std::ios::binary);
serialize (best_network, fout);
fout.close();
}
}
static void
krr_lin_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
)
{
typedef linear_kernel<dense_sample_type> kernel_type;
krr_trainer<kernel_type> trainer;
// LOO cross validation
double loo_error;
trainer.train(dense_samples, labels, loo_error);
std::cout << "mean squared LOO error: " << loo_error << std::endl;
}
static void
krr_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
)
{
const char* kernel = get_kernel (parser);
if (!strcmp (kernel, "lin")) {
krr_lin_test (parser, dense_samples, labels);
} else if (!strcmp (kernel, "rbk")) {
krr_rbk_test (parser, dense_samples, labels);
} else {
fprintf (stderr, "Unknown kernel type: %s\n", kernel);
exit (-1);
}
}
static void
mlp_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
)
{
// Create a multi-layer perceptron network.
const int num_input = dense_samples[0].size();
int num_hidden = get_mlp_hidden_units (parser, dense_samples);
printf ("Creating ANN with size (%d, %d)\n", num_input, num_hidden);
mlp::kernel_1a_c net (num_input, num_hidden);
// Dlib barfs if output values are not normalized to [0,1]
double label_min = *(std::min_element (labels.begin(), labels.end()));
double label_max = *(std::max_element (labels.begin(), labels.end()));
std::vector<double>::iterator it;
for (it = labels.begin(); it != labels.end(); it++) {
(*it) = ((*it) - label_min) / (label_max - label_min);
}
// Split into training set and testing set
float training_pct = 0.8;
unsigned int training_samples = (unsigned int) floor (
training_pct * dense_samples.size());
// Dlib doesn't seem to come with any batch training functions for mlp.
// Also, note that only backprop is supported.
int num_iterations = get_mlp_num_iterations (parser, dense_samples);
for (int i = 0; i < num_iterations; i++) {
for (unsigned int j = 0; j < training_samples; j++) {
net.train (dense_samples[j], labels[j]);
}
}
// Test the performance (sorry, no cross-validation) */
double total_err = 0.0;
for (unsigned int j = training_samples + 1; j < dense_samples.size(); j++)
{
double diff = net(dense_samples[j]) - labels[j];
diff = diff * (label_max - label_min);
total_err += diff * diff;
}
std::cout
<< "MSE (no cross-validation): "
<< total_err / (dense_samples.size() - training_samples) << std::endl;
}
static void
svr_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
)
{
typedef radial_basis_kernel<dense_sample_type> kernel_type;
svr_trainer<kernel_type> trainer;
option_range gamma_range, svr_c_range;
get_rbk_gamma (parser, dense_samples, gamma_range);
get_svr_c (parser, dense_samples, svr_c_range);
double epsilon_insensitivity = get_svr_epsilon_insensitivity (
parser, dense_samples);
trainer.set_epsilon_insensitivity (epsilon_insensitivity);
for (float svr_c = svr_c_range.get_min_value();
svr_c <= svr_c_range.get_max_value();
svr_c = svr_c_range.get_next_value (svr_c))
{
trainer.set_c (svr_c);
for (float gamma = gamma_range.get_min_value();
gamma <= gamma_range.get_max_value();
gamma = gamma_range.get_next_value (gamma))
{
double cv_error;
trainer.set_kernel (kernel_type (gamma));
cv_error = cross_validate_regression_trainer (trainer,
dense_samples, labels, 10);
printf ("%3.6f %3.6f %3.9f\n", svr_c, gamma, cv_error);
}
}
}
int int
main (int argc, char* argv[]) main (int argc, char* argv[])
...@@ -581,3 +180,6 @@ main (int argc, char* argv[]) ...@@ -581,3 +180,6 @@ main (int argc, char* argv[])
exit (-1); exit (-1);
} }
} }
// ----------------------------------------------------------------------------------------
tools/mltool/src/option_range.cpp 0 → 100644
View file @ 7fb32196
// The contents of this file are in the public domain.
// See LICENSE_FOR_EXAMPLE_PROGRAMS.txt (in trunk/examples)
// Authors:
// Gregory Sharp
// Davis King
#include "option_range.h"
#include <iostream>
// ----------------------------------------------------------------------------------------
/* exp10() is not in C/C++ standard */
double
exp10_ (double m)
{
return exp (2.3025850929940456840179914546844 * m);
}
// ----------------------------------------------------------------------------------------
void
option_range::set_option (
clp& parser,
std::string const& option,
float default_val
)
{
int rc;
/* No option specified */
if (!parser.option (option)) {
log_range = 0;
min_value = default_val;
max_value = default_val;
incr = 1;
return;
}
/* Range specified */
rc = sscanf (parser.option(option).argument().c_str(), "%f:%f:%f",
&min_value, &incr, &max_value);
if (rc == 3) {
log_range = 1;
return;
}
/* Single value specified */
if (rc == 1) {
log_range = 0;
max_value = min_value;
incr = 1;
return;
}
else {
std::cerr << "Error parsing option" << option << "\n";
exit (-1);
}
}
// ----------------------------------------------------------------------------------------
float
option_range::get_min_value ()
{
if (log_range) {
return exp10_ (min_value);
} else {
return min_value;
}
}
// ----------------------------------------------------------------------------------------
float
option_range::get_max_value ()
{
if (log_range) {
return exp10_ (max_value);
} else {
return max_value;
}
}
// ----------------------------------------------------------------------------------------
float
option_range::get_next_value (float curr_value)
{
if (log_range) {
curr_value = log10 (curr_value);
curr_value += incr;
curr_value = exp10_ (curr_value);
} else {
curr_value += incr;
}
return curr_value;
}
// ----------------------------------------------------------------------------------------
tools/mltool/src/option_range.h 0 → 100644
View file @ 7fb32196
// The contents of this file are in the public domain.
// See LICENSE_FOR_EXAMPLE_PROGRAMS.txt (in trunk/examples)
// Authors:
// Gregory Sharp
// Davis King
#ifndef DLIB_MLTOOL_OPTION_RaNGE_H__
#define DLIB_MLTOOL_OPTION_RaNGE_H__
#include "common.h"
#include <string>
// ----------------------------------------------------------------------------------------
struct option_range
{
/*!
WHAT THIS OBJECT REPRESENTS
This object is a tool for representing a range of possible parameter values.
The range is determined based on the contents of the command line.
!*/
public:
bool log_range;
float min_value;
float max_value;
float incr;
public:
option_range () {
log_range = false;
min_value = 0;
max_value = 100;
incr = 10;
}
void set_option (clp& parser, std::string const& option,
float default_val);
float get_min_value ();
float get_max_value ();
float get_next_value (float curr_val);
};
// ----------------------------------------------------------------------------------------
#endif // DLIB_MLTOOL_OPTION_RaNGE_H__
tools/mltool/src/regression.cpp 0 → 100644
View file @ 7fb32196
// The contents of this file are in the public domain.
// See LICENSE_FOR_EXAMPLE_PROGRAMS.txt (in trunk/examples)
// Authors:
// Gregory Sharp
// Davis King
#include "regression.h"
#include "dlib/mlp.h"
#include "dlib/svm.h"
#include "option_range.h"
#include "dlib/string.h"
#include <cmath>
#include <cfloat>
using namespace dlib;
// ----------------------------------------------------------------------------------------
static const char*
get_kernel (
clp& parser
)
{
const char* kernel = "rbk";
if (parser.option ("k")) {
kernel = parser.option("k").argument().c_str();
}
return kernel;
}
// ----------------------------------------------------------------------------------------
static void
get_rbk_gamma (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
option_range& range
) {
float default_gamma = 3.0 / compute_mean_squared_distance (
randomly_subsample (dense_samples, 2000));
range.set_option (parser, "rbk-gamma", default_gamma);
}
// ----------------------------------------------------------------------------------------
static void
get_krls_tolerance (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
option_range& range
)
{
float default_krls_tolerance = 0.001;
range.set_option (parser, "krls-tolerance", default_krls_tolerance);
}
// ----------------------------------------------------------------------------------------
static double
get_mlp_hidden_units (
clp& parser,
std::vector<dense_sample_type>& dense_samples
)
{
int num_hidden = 5;
if (parser.option ("mlp-hidden-units")) {
num_hidden = sa = parser.option("mlp-hidden-units").argument();
}
return num_hidden;
}
// ----------------------------------------------------------------------------------------
static double
get_mlp_num_iterations (
clp& parser,
std::vector<dense_sample_type>& dense_samples
)
{
int num_iterations = 5000;
if (parser.option ("mlp-num-iterations")) {
num_iterations = sa = parser.option("mlp-num-iterations").argument();
}
return num_iterations;
}
// ----------------------------------------------------------------------------------------
static void
get_svr_c (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
option_range& range
)
{
float default_svr_c = 1000.;
range.set_option (parser, "svr-c", default_svr_c);
}
// ----------------------------------------------------------------------------------------
static double
get_svr_epsilon_insensitivity (
clp& parser,
std::vector<dense_sample_type>& dense_samples
)
{
// Epsilon-insensitive regression means we do regression but stop
// trying to fit a data point once it is "close enough" to its
// target value. This parameter is the value that controls what
// we mean by "close enough". In this case, I'm saying I'm happy
// if the resulting regression function gets within 0.001 of the
// target value.
double epsilon_insensitivity = 0.001;
if (parser.option ("svr-epsilon-insensitivity")) {
epsilon_insensitivity
= sa = parser.option("svr-epsilon-insensitivity").argument();
}
return epsilon_insensitivity;
}
// ----------------------------------------------------------------------------------------
void
krls_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
)
{
typedef radial_basis_kernel<dense_sample_type> kernel_type;
option_range gamma_range, krls_tol_range;
get_rbk_gamma (parser, dense_samples, gamma_range);
get_krls_tolerance (parser, dense_samples, krls_tol_range);
// Split into training set and testing set
float training_pct = 0.8;
unsigned int training_samples = (unsigned int) floor (
training_pct * dense_samples.size());
for (float krls_tol = krls_tol_range.get_min_value();
krls_tol <= krls_tol_range.get_max_value();
krls_tol = krls_tol_range.get_next_value (krls_tol))
{
for (float gamma = gamma_range.get_min_value();
gamma <= gamma_range.get_max_value();
gamma = gamma_range.get_next_value (gamma))
{
krls<kernel_type> net (kernel_type(gamma), krls_tol);
// Krls doesn't seem to come with any batch training function
for (unsigned int j = 0; j < training_samples; j++) {
net.train (dense_samples[j], labels[j]);
}
// Test the performance (sorry, no cross-validation)
double total_err = 0.0;
for (unsigned int j = training_samples + 1;
j < dense_samples.size(); j++)
{
double diff = net(dense_samples[j]) - labels[j];
total_err += diff * diff;
}
double testset_error = total_err
/ (dense_samples.size() - training_samples);
printf ("%3.6f %3.6f %3.9f\n", krls_tol, gamma, testset_error);
}
}
}
// ----------------------------------------------------------------------------------------
static void
krr_rbk_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
)
{
typedef radial_basis_kernel<dense_sample_type> kernel_type;
krr_trainer<kernel_type> trainer;
option_range gamma_range;
double best_gamma = DBL_MAX;
float best_loo = FLT_MAX;
get_rbk_gamma (parser, dense_samples, gamma_range);
for (float gamma = gamma_range.get_min_value();
gamma <= gamma_range.get_max_value();
gamma = gamma_range.get_next_value (gamma))
{
// LOO cross validation
double loo_error;
if (parser.option("verbose")) {
trainer.set_search_lambdas(logspace(-9, 4, 100));
trainer.be_verbose();
}
trainer.set_kernel (kernel_type (gamma));
trainer.train (dense_samples, labels, loo_error);
if (loo_error < best_loo) {
best_loo = loo_error;
best_gamma = gamma;
}
printf ("10^%f %9.6f\n", log10(gamma), loo_error);
}
printf ("Best result: gamma=10^%f (%g), loo_error=%9.6f\n",
log10(best_gamma), best_gamma, best_loo);
if (parser.option("train-best")) {
printf ("Training network with best parameters\n");
trainer.set_kernel (kernel_type (best_gamma));
decision_function<kernel_type> best_network =
trainer.train (dense_samples, labels);
std::ofstream fout (parser.option("train-best").argument().c_str(),
std::ios::binary);
serialize (best_network, fout);
fout.close();
}
}
// ----------------------------------------------------------------------------------------
static void
krr_lin_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
)
{
typedef linear_kernel<dense_sample_type> kernel_type;
krr_trainer<kernel_type> trainer;
// LOO cross validation
double loo_error;
trainer.train(dense_samples, labels, loo_error);
std::cout << "mean squared LOO error: " << loo_error << std::endl;
}
// ----------------------------------------------------------------------------------------
void
krr_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
)
{
const char* kernel = get_kernel (parser);
if (!strcmp (kernel, "lin")) {
krr_lin_test (parser, dense_samples, labels);
} else if (!strcmp (kernel, "rbk")) {
krr_rbk_test (parser, dense_samples, labels);
} else {
fprintf (stderr, "Unknown kernel type: %s\n", kernel);
exit (-1);
}
}
// ----------------------------------------------------------------------------------------
void
mlp_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
)
{
// Create a multi-layer perceptron network.
const int num_input = dense_samples[0].size();
int num_hidden = get_mlp_hidden_units (parser, dense_samples);
printf ("Creating ANN with size (%d, %d)\n", num_input, num_hidden);
mlp::kernel_1a_c net (num_input, num_hidden);
// Dlib barfs if output values are not normalized to [0,1]
double label_min = *(std::min_element (labels.begin(), labels.end()));
double label_max = *(std::max_element (labels.begin(), labels.end()));
std::vector<double>::iterator it;
for (it = labels.begin(); it != labels.end(); it++) {
(*it) = ((*it) - label_min) / (label_max - label_min);
}
// Split into training set and testing set
float training_pct = 0.8;
unsigned int training_samples = (unsigned int) floor (
training_pct * dense_samples.size());
// Dlib doesn't seem to come with any batch training functions for mlp.
// Also, note that only backprop is supported.
int num_iterations = get_mlp_num_iterations (parser, dense_samples);
for (int i = 0; i < num_iterations; i++) {
for (unsigned int j = 0; j < training_samples; j++) {
net.train (dense_samples[j], labels[j]);
}
}
// Test the performance (sorry, no cross-validation) */
double total_err = 0.0;
for (unsigned int j = training_samples + 1; j < dense_samples.size(); j++)
{
double diff = net(dense_samples[j]) - labels[j];
diff = diff * (label_max - label_min);
total_err += diff * diff;
}
std::cout
<< "MSE (no cross-validation): "
<< total_err / (dense_samples.size() - training_samples) << std::endl;
}
// ----------------------------------------------------------------------------------------
void
svr_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
)
{
typedef radial_basis_kernel<dense_sample_type> kernel_type;
svr_trainer<kernel_type> trainer;
option_range gamma_range, svr_c_range;
get_rbk_gamma (parser, dense_samples, gamma_range);
get_svr_c (parser, dense_samples, svr_c_range);
double epsilon_insensitivity = get_svr_epsilon_insensitivity (
parser, dense_samples);
trainer.set_epsilon_insensitivity (epsilon_insensitivity);
for (float svr_c = svr_c_range.get_min_value();
svr_c <= svr_c_range.get_max_value();
svr_c = svr_c_range.get_next_value (svr_c))
{
trainer.set_c (svr_c);
for (float gamma = gamma_range.get_min_value();
gamma <= gamma_range.get_max_value();
gamma = gamma_range.get_next_value (gamma))
{
double cv_error;
trainer.set_kernel (kernel_type (gamma));
cv_error = cross_validate_regression_trainer (trainer,
dense_samples, labels, 10);
printf ("%3.6f %3.6f %3.9f\n", svr_c, gamma, cv_error);
}
}
}
// ----------------------------------------------------------------------------------------
tools/mltool/src/regression.h 0 → 100644
View file @ 7fb32196
// The contents of this file are in the public domain.
// See LICENSE_FOR_EXAMPLE_PROGRAMS.txt (in trunk/examples)
// Authors:
// Gregory Sharp
// Davis King
#ifndef DLIB_MLTOOL_REGREsSION_H__
#define DLIB_MLTOOL_REGREsSION_H__
#include "common.h"
#include <vector>
void
krr_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
);
void
krls_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
);
void
mlp_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
);
void
svr_test (
clp& parser,
std::vector<dense_sample_type>& dense_samples,
std::vector<double>& labels
);
#endif // DLIB_MLTOOL_REGREsSION_H__
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment