Skip to content

D
dlib
Commit 04991b7d authored by Davis King's avatar Davis King
Browse files
Options

Made this example program use the new find_max_global() instead of grid search 

and BOBYQA.  This greatly simplifies the example.
parent fc6cce9f
Hide whitespace changes
Inline Side-by-side
Showing with 111 additions and 222 deletions
examples/model_selection_ex.cpp
View file @ 04991b7d
// The contents of this file are in the public domain. See LICENSE_FOR_EXAMPLE_PROGRAMS.txt // The contents of this file are in the public domain. See LICENSE_FOR_EXAMPLE_PROGRAMS.txt
/* /*
This is an example that shows some reasonable ways you can perform This is an example that shows how you can perform model selection with the
model selection with the dlib C++ Library. dlib C++ Library.
It will create a simple set of data and then show you how to use It will create a simple dataset and show you how to use cross validation and
the cross validation and optimization routines to determine good model global optimization to determine good parameters for the purpose of training
parameters for the purpose of training an svm to classify the sample data. an svm to classify the data.
The data used in this example will be 2 dimensional data and will The data used in this example will be 2 dimensional data and will come from a
come from a distribution where points with a distance less than 10 distribution where points with a distance less than 10 from the origin are
from the origin are labeled +1 and all other points are labeled labeled +1 and all other points are labeled as -1.
as -1.
As an side, you should probably read the svm_ex.cpp and matrix_ex.cpp example As an side, you should probably read the svm_ex.cpp and matrix_ex.cpp example
programs before you read this one. programs before you read this one.
*/ */
#include <iostream> #include <iostream>
#include <dlib/svm.h> #include <dlib/svm.h>
#include <dlib/global_optimization.h>
using namespace std; using namespace std;
using namespace dlib; using namespace dlib;
// The svm functions use column vectors to contain a lot of the data on which they
// operate. So the first thing we do here is declare a convenient typedef.
// This typedef declares a matrix with 2 rows and 1 column. It will be the int main() try
// object that contains each of our 2 dimensional samples. {
typedef matrix<double, 2, 1> sample_type; // The svm functions use column vectors to contain a lot of the data on which they
// operate. So the first thing we do here is declare a convenient typedef.
// This is a typedef for the type of kernel we are going to use in this example. // This typedef declares a matrix with 2 rows and 1 column. It will be the
// In this case I have selected the radial basis kernel that can operate on our // object that contains each of our 2 dimensional samples.
// 2D sample_type objects typedef matrix<double, 2, 1> sample_type;
typedef radial_basis_kernel<sample_type> kernel_type;
// Now we make objects to contain our samples and their respective labels.
std::vector<sample_type> samples;
std::vector<double> labels;
class cross_validation_objective // Now let's put some data into our samples and labels objects. We do this
{ // by looping over a bunch of points and labeling them according to their
/*! // distance from the origin.
WHAT THIS OBJECT REPRESENTS for (double r = -20; r <= 20; r += 0.8)
This object is a simple function object that takes a set of model
parameters and returns a number indicating how "good" they are. It
does this by performing 10 fold cross validation on our dataset
and reporting the accuracy.
See below in main() for how this object gets used.
!*/
public:
cross_validation_objective (
const std::vector<sample_type>& samples_,
const std::vector<double>& labels_
) : samples(samples_), labels(labels_) {}
double operator() (
const matrix<double>& params
) const
{ {
// Pull out the two SVM model parameters. Note that, in this case, for (double c = -20; c <= 20; c += 0.8)
// I have setup the parameter search to operate in log scale so we have {
// to remember to call exp() to put the parameters back into a normal scale. sample_type samp;
const double gamma = exp(params(0)); samp(0) = r;
const double nu = exp(params(1)); samp(1) = c;
samples.push_back(samp);
// if this point is less than 10 from the origin
if (sqrt(r*r + c*c) <= 10)
labels.push_back(+1);
else
labels.push_back(-1);
}
}
// Make an SVM trainer and tell it what the parameters are supposed to be. cout << "Generated " << samples.size() << " points" << endl;
// Here we normalize all the samples by subtracting their mean and dividing by their
// standard deviation. This is generally a good idea since it often heads off
// numerical stability problems and also prevents one large feature from smothering
// others. Doing this doesn't matter much in this example so I'm just doing this here
// so you can see an easy way to accomplish this with the library.
vector_normalizer<sample_type> normalizer;
// let the normalizer learn the mean and standard deviation of the samples
normalizer.train(samples);
// now normalize each sample
for (unsigned long i = 0; i < samples.size(); ++i)
samples[i] = normalizer(samples[i]);
// Now that we have some data we want to train on it. We are going to train a
// binary SVM with the RBF kernel to classify the data. However, there are two
// parameters to the training. These are the nu and gamma parameters. Our choice
// for these parameters will influence how good the resulting decision function is.
// To test how good a particular choice of these parameters is we can use the
// cross_validate_trainer() function to perform n-fold cross validation on our
// training data. However, there is a problem with the way we have sampled our
// distribution above. The problem is that there is a definite ordering to the
// samples. That is, the first half of the samples look like they are from a
// different distribution than the second half. This would screw up the cross
// validation process, but we can fix it by randomizing the order of the samples
// with the following function call.
randomize_samples(samples, labels);
// And now we get to the important bit. Here we define a function,
// cross_validation_score(), that will do the cross-validation we
// mentioned and return a number indicating how good a particular setting
// of gamma and nu is.
auto cross_validation_score = [&](const double gamma, const double nu)
{
// Make a RBF SVM trainer and tell it what the parameters are supposed to be.
typedef radial_basis_kernel<sample_type> kernel_type;
svm_nu_trainer<kernel_type> trainer; svm_nu_trainer<kernel_type> trainer;
trainer.set_kernel(kernel_type(gamma)); trainer.set_kernel(kernel_type(gamma));
trainer.set_nu(nu); trainer.set_nu(nu);
...@@ -77,189 +109,46 @@ public: ...@@ -77,189 +109,46 @@ public:
matrix<double> result = cross_validate_trainer(trainer, samples, labels, 10); matrix<double> result = cross_validate_trainer(trainer, samples, labels, 10);
cout << "gamma: " << setw(11) << gamma << " nu: " << setw(11) << nu << " cross validation accuracy: " << result; cout << "gamma: " << setw(11) << gamma << " nu: " << setw(11) << nu << " cross validation accuracy: " << result;
// Here I'm returning the harmonic mean between the accuracies of each class. // Now return a number indicating how good the parameters are. Bigger is
// However, you could do something else. For example, you might care a lot more // better in this example. Here I'm returning the harmonic mean between the
// about correctly predicting the +1 class, so you could penalize results that // accuracies of each class. However, you could do something else. For
// didn't obtain a high accuracy on that class. You might do this by using // example, you might care a lot more about correctly predicting the +1 class,
// something like a weighted version of the F1-score (see http://en.wikipedia.org/wiki/F1_score). // so you could penalize results that didn't obtain a high accuracy on that
// class. You might do this by using something like a weighted version of the
// F1-score (see http://en.wikipedia.org/wiki/F1_score).
return 2*prod(result)/sum(result); return 2*prod(result)/sum(result);
} };
const std::vector<sample_type>& samples;
const std::vector<double>& labels;
};
int main()
{
try
{
// Now we make objects to contain our samples and their respective labels.
std::vector<sample_type> samples;
std::vector<double> labels;
// Now let's put some data into our samples and labels objects. We do this
// by looping over a bunch of points and labeling them according to their
// distance from the origin.
for (double r = -20; r <= 20; r += 0.8)
{
for (double c = -20; c <= 20; c += 0.8)
{
sample_type samp;
samp(0) = r;
samp(1) = c;
samples.push_back(samp);
// if this point is less than 10 from the origin
if (sqrt(r*r + c*c) <= 10)
labels.push_back(+1);
else
labels.push_back(-1);
}
}
cout << "Generated " << samples.size() << " points" << endl;
// Here we normalize all the samples by subtracting their mean and dividing by their standard deviation.
// This is generally a good idea since it often heads off numerical stability problems and also
// prevents one large feature from smothering others. Doing this doesn't matter much in this example
// so I'm just doing this here so you can see an easy way to accomplish this with
// the library.
vector_normalizer<sample_type> normalizer;
// let the normalizer learn the mean and standard deviation of the samples
normalizer.train(samples);
// now normalize each sample
for (unsigned long i = 0; i < samples.size(); ++i)
samples[i] = normalizer(samples[i]);
// Now that we have some data we want to train on it. However, there are two parameters to the
// training. These are the nu and gamma parameters. Our choice for these parameters will
// influence how good the resulting decision function is. To test how good a particular choice
// of these parameters is we can use the cross_validate_trainer() function to perform n-fold cross
// validation on our training data. However, there is a problem with the way we have sampled
// our distribution above. The problem is that there is a definite ordering to the samples.
// That is, the first half of the samples look like they are from a different distribution
// than the second half. This would screw up the cross validation process but we can
// fix it by randomizing the order of the samples with the following function call.
randomize_samples(samples, labels);
// The nu parameter has a maximum value that is dependent on the ratio of the +1 to -1
// labels in the training data. This function finds that value. The 0.999 is here because
// the maximum allowable nu is strictly less than the value returned by maximum_nu(). So
// rather than dealing with that below we can just back away from it a little bit here and then
// not worry about it.
const double max_nu = 0.999*maximum_nu(labels);
// The first kind of model selection we will do is a simple grid search. That is, below we just
// generate a fixed grid of points (each point represents one possible setting of the model parameters)
// and test each via cross validation.
// This code generates a 4x4 grid of logarithmically spaced points. The result is a matrix
// with 2 rows and 16 columns where each column represents one of our points.
matrix<double> params = cartesian_product(logspace(log10(5.0), log10(1e-5), 4), // gamma parameter
logspace(log10(max_nu), log10(1e-5), 4) // nu parameter
);
// As an aside, if you wanted to do a grid search over points of dimensionality more than two
// you would just nest calls to cartesian_product(). You can also use linspace() to generate
// linearly spaced points if that is more appropriate for the parameters you are working with.
// Next we loop over all the points we generated and check how good each is.
cout << "Doing a grid search" << endl;
matrix<double> best_result(2,1);
best_result = 0;
double best_gamma = 0.1, best_nu;
for (long col = 0; col < params.nc(); ++col)
{
// pull out the current set of model parameters
const double gamma = params(0, col);
const double nu = params(1, col);
// setup a training object using our current parameters
svm_nu_trainer<kernel_type> trainer;
trainer.set_kernel(kernel_type(gamma));
trainer.set_nu(nu);
// Finally, do 10 fold cross validation and then check if the results are the best we have seen so far.
matrix<double> result = cross_validate_trainer(trainer, samples, labels, 10);
cout << "gamma: " << setw(11) << gamma << " nu: " << setw(11) << nu << " cross validation accuracy: " << result;
// save the best results
if (sum(result) > sum(best_result))
{
best_result = result;
best_gamma = gamma;
best_nu = nu;
}
}
cout << "\n best result of grid search: " << sum(best_result) << endl;
cout << " best gamma: " << best_gamma << " best nu: " << best_nu << endl;
// Grid search is a very simple brute force method. Below we try out the BOBYQA algorithm.
// It is a routine that performs optimization of a function in the absence of derivatives.
cout << "\n\n Try the BOBYQA algorithm" << endl; // The nu parameter has a maximum value that is dependent on the ratio of the +1 to -1
// labels in the training data. This function finds that value. The 0.999 is here
// because the maximum allowable nu is strictly less than the value returned by
// maximum_nu(). So shrinking the limit a little will prevent us from hitting it.
const double max_nu = 0.999*maximum_nu(labels);
// We need to supply a starting point for the optimization. Here we are using the best
// result of the grid search. Generally, you want to try and give a reasonable starting
// point due to the possibility of the optimization getting stuck in a local maxima.
params.set_size(2,1);
params = best_gamma, // initial gamma
best_nu; // initial nu
// We also need to supply lower and upper bounds for the search. // And finally, we call this global optimizer that will search for the best parameters.
matrix<double> lower_bound(2,1), upper_bound(2,1); // It will call cross_validation_score() 50 times with different settings and return
lower_bound = 1e-7, // smallest allowed gamma // the best parameter setting it finds. find_max_global() uses a global optimization
1e-7; // smallest allowed nu // method based on a combination of non-parametric global function modeling and
upper_bound = 100, // largest allowed gamma // quadratic trust region modeling to efficiently find a global maximizer. It usually
max_nu; // largest allowed nu // does a good job with a relatively small number of calls to cross_validation_score().
// In this example, you should observe that it finds settings that give perfect binary
// classification on the data.
auto result = find_max_global(cross_validation_score,
{1e-5, 1e-5}, // lower bound constraints on gamma and nu, respectively
{100, max_nu}, // upper bound constraints on gamma and nu, respectively
max_function_calls(50));
double best_gamma = result.x(0);
double best_nu = result.x(1);
// For the gamma and nu SVM parameters it is generally a good idea to search cout << " best cross-validation score: " << result.y << endl;
// in log space. So I'm just converting them into log space here before cout << " best gamma: " << best_gamma << " best nu: " << best_nu << endl;
// we start the optimization.
params = log(params);
lower_bound = log(lower_bound);
upper_bound = log(upper_bound);
// Finally, ask BOBYQA to look for the best set of parameters. Note that we are using the
// cross validation function object defined at the top of the file.
double best_score = find_max_bobyqa(
cross_validation_objective(samples, labels), // Function to maximize
params, // starting point
params.size()*2 + 1, // See BOBYQA docs, generally size*2+1 is a good setting for this
lower_bound, // lower bound
upper_bound, // upper bound
min(upper_bound-lower_bound)/10, // search radius
0.01, // desired accuracy
100 // max number of allowable calls to cross_validation_objective()
);
// Don't forget to convert back from log scale to normal scale }
params = exp(params); catch (exception& e)
{
cout << " best result of BOBYQA: " << best_score << endl; cout << e.what() << endl;
cout << " best gamma: " << params(0) << " best nu: " << params(1) << endl;
// Also note that the find_max_bobyqa() function only works for optimization problems
// with 2 variables or more. If you only have a single variable then you should use
// the find_max_single_variable() function.
}
catch (exception& e)
{
cout << e.what() << endl;
}
} }
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