svm_ex.cpp 11.7 KB
Newer Older
1
// The contents of this file are in the public domain. See LICENSE_FOR_EXAMPLE_PROGRAMS.txt
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
/*

    This is an example illustrating the use of the support vector machine
    utilities from the dlib C++ Library.  

    This example creates a simple set of data to train on and then shows
    you how to use the cross validation and svm training functions
    to find a good decision function that can classify examples in our
    data set.


    The data used in this example will be 2 dimensional data and will
    come from a distribution where points with a distance less than 10
    from the origin are labeled +1 and all other points are labeled
    as -1.
        
*/


#include <iostream>
22
#include <dlib/svm.h>
23
24
25
26
27
28
29

using namespace std;
using namespace dlib;


int main()
{
Davis King's avatar
Davis King committed
30
    // The svm functions use column vectors to contain a lot of the data on which they
Davis King's avatar
Davis King committed
31
    // operate. So the first thing we do here is declare a convenient typedef.  
32

Davis King's avatar
Davis King committed
33
34
35
36
37
    // This typedef declares a matrix with 2 rows and 1 column.  It will be the object that
    // contains each of our 2 dimensional samples.   (Note that if you wanted more than 2
    // features in this vector you can simply change the 2 to something else.  Or if you
    // don't know how many features you want until runtime then you can put a 0 here and
    // use the matrix.set_size() member function)
38
39
    typedef matrix<double, 2, 1> sample_type;

Davis King's avatar
Davis King committed
40
41
42
    // This is a typedef for the type of kernel we are going to use in this example.  In
    // this case I have selected the radial basis kernel that can operate on our 2D
    // sample_type objects
43
44
45
    typedef radial_basis_kernel<sample_type> kernel_type;


Davis King's avatar
Davis King committed
46
47
48
    // Now we make objects to contain our samples and their respective labels.
    std::vector<sample_type> samples;
    std::vector<double> labels;
49

Davis King's avatar
Davis King committed
50
    // Now let's put some data into our samples and labels objects.  We do this by looping
Davis King's avatar
Davis King committed
51
52
    // over a bunch of points and labeling them according to their distance from the
    // origin.
53
54
55
56
    for (int r = -20; r <= 20; ++r)
    {
        for (int c = -20; c <= 20; ++c)
        {
Davis King's avatar
Davis King committed
57
58
59
60
            sample_type samp;
            samp(0) = r;
            samp(1) = c;
            samples.push_back(samp);
61
62
63

            // if this point is less than 10 from the origin
            if (sqrt((double)r*r + c*c) <= 10)
Davis King's avatar
Davis King committed
64
                labels.push_back(+1);
65
            else
Davis King's avatar
Davis King committed
66
                labels.push_back(-1);
67
68
69
70
71

        }
    }


Davis King's avatar
Davis King committed
72
73
74
75
76
    // 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.  
77
78
79
    vector_normalizer<sample_type> normalizer;
    // let the normalizer learn the mean and standard deviation of the samples
    normalizer.train(samples);
Davis King's avatar
Davis King committed
80
81
    // now normalize each sample
    for (unsigned long i = 0; i < samples.size(); ++i)
82
        samples[i] = normalizer(samples[i]); 
Davis King's avatar
Davis King committed
83
84


Davis King's avatar
Davis King committed
85
86
87
88
89
90
91
92
93
94
    // 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.
95
96
97
    randomize_samples(samples, labels);


Davis King's avatar
Davis King committed
98
    // The nu parameter has a maximum value that is dependent on the ratio of the +1 to -1
99
100
101
    // labels in the training data.  This function finds that value.
    const double max_nu = maximum_nu(labels);

Davis King's avatar
Davis King committed
102
103
104
    // here we make an instance of the svm_nu_trainer object that uses our kernel type.
    svm_nu_trainer<kernel_type> trainer;

105
    // Now we loop over some different nu and gamma values to see how good they are.  Note
Davis King's avatar
Davis King committed
106
107
    // that this is a very simple way to try out a few possible parameter choices.  You
    // should look at the model_selection_ex.cpp program for examples of more sophisticated
108
    // strategies for determining good parameter choices.
109
    cout << "doing cross validation" << endl;
110
    for (double gamma = 0.00001; gamma <= 1; gamma *= 5)
111
    {
112
        for (double nu = 0.00001; nu < max_nu; nu *= 5)
113
        {
Davis King's avatar
Davis King committed
114
115
116
117
            // tell the trainer the parameters we want to use
            trainer.set_kernel(kernel_type(gamma));
            trainer.set_nu(nu);

118
            cout << "gamma: " << gamma << "    nu: " << nu;
Davis King's avatar
Davis King committed
119
120
121
122
            // Print out the cross validation accuracy for 3-fold cross validation using
            // the current gamma and nu.  cross_validate_trainer() returns a row vector.
            // The first element of the vector is the fraction of +1 training examples
            // correctly classified and the second number is the fraction of -1 training
123
            // examples correctly classified.
Davis King's avatar
Davis King committed
124
            cout << "     cross validation accuracy: " << cross_validate_trainer(trainer, samples, labels, 3);
125
126
127
128
        }
    }


Davis King's avatar
Davis King committed
129
130
    // From looking at the output of the above loop it turns out that a good value for nu
    // and gamma for this problem is 0.15625 for both.  So that is what we will use.
131

Davis King's avatar
Davis King committed
132
133
134
135
    // Now we train on the full set of data and obtain the resulting decision function.  We
    // use the value of 0.15625 for nu and gamma.  The decision function will return values
    // >= 0 for samples it predicts are in the +1 class and numbers < 0 for samples it
    // predicts to be in the -1 class.
136
137
    trainer.set_kernel(kernel_type(0.15625));
    trainer.set_nu(0.15625);
138
139
140
    typedef decision_function<kernel_type> dec_funct_type;
    typedef normalized_function<dec_funct_type> funct_type;

Davis King's avatar
Davis King committed
141
142
143
    // Here we are making an instance of the normalized_function object.  This object
    // provides a convenient way to store the vector normalization information along with
    // the decision function we are going to learn.  
144
145
146
    funct_type learned_function;
    learned_function.normalizer = normalizer;  // save normalization information
    learned_function.function = trainer.train(samples, labels); // perform the actual SVM training and save the results
147
148

    // print out the number of support vectors in the resulting decision function
149
    cout << "\nnumber of support vectors in our learned_function is " 
Davis King's avatar
Davis King committed
150
         << learned_function.function.basis_vectors.size() << endl;
151

Davis King's avatar
Davis King committed
152
    // Now let's try this decision_function on some samples we haven't seen before.
153
154
155
156
    sample_type sample;

    sample(0) = 3.123;
    sample(1) = 2;
Davis King's avatar
Davis King committed
157
    cout << "This is a +1 class example, the classifier output is " << learned_function(sample) << endl;
158
159
160

    sample(0) = 3.123;
    sample(1) = 9.3545;
Davis King's avatar
Davis King committed
161
    cout << "This is a +1 class example, the classifier output is " << learned_function(sample) << endl;
162
163
164

    sample(0) = 13.123;
    sample(1) = 9.3545;
Davis King's avatar
Davis King committed
165
    cout << "This is a -1 class example, the classifier output is " << learned_function(sample) << endl;
166
167
168

    sample(0) = 13.123;
    sample(1) = 0;
Davis King's avatar
Davis King committed
169
    cout << "This is a -1 class example, the classifier output is " << learned_function(sample) << endl;
170
171


Davis King's avatar
Davis King committed
172
173
    // We can also train a decision function that reports a well conditioned probability
    // instead of just a number > 0 for the +1 class and < 0 for the -1 class.  An example
Davis King's avatar
Davis King committed
174
    // of doing that follows:
175
176
177
178
179
180
    typedef probabilistic_decision_function<kernel_type> probabilistic_funct_type;  
    typedef normalized_function<probabilistic_funct_type> pfunct_type;

    pfunct_type learned_pfunct; 
    learned_pfunct.normalizer = normalizer;
    learned_pfunct.function = train_probabilistic_decision_function(trainer, samples, labels, 3);
181
    // Now we have a function that returns the probability that a given sample is of the +1 class.  
182

Davis King's avatar
Davis King committed
183
184
    // print out the number of support vectors in the resulting decision function.  
    // (it should be the same as in the one above)
185
    cout << "\nnumber of support vectors in our learned_pfunct is " 
Davis King's avatar
Davis King committed
186
         << learned_pfunct.function.decision_funct.basis_vectors.size() << endl;
187
188
189

    sample(0) = 3.123;
    sample(1) = 2;
Davis King's avatar
Davis King committed
190
191
    cout << "This +1 class example should have high probability.  Its probability is: " 
         << learned_pfunct(sample) << endl;
192
193
194

    sample(0) = 3.123;
    sample(1) = 9.3545;
Davis King's avatar
Davis King committed
195
196
    cout << "This +1 class example should have high probability.  Its probability is: " 
         << learned_pfunct(sample) << endl;
197
198
199

    sample(0) = 13.123;
    sample(1) = 9.3545;
Davis King's avatar
Davis King committed
200
201
    cout << "This -1 class example should have low probability.  Its probability is: " 
         << learned_pfunct(sample) << endl;
202
203
204

    sample(0) = 13.123;
    sample(1) = 0;
Davis King's avatar
Davis King committed
205
206
    cout << "This -1 class example should have low probability.  Its probability is: " 
         << learned_pfunct(sample) << endl;
207
208


209

Davis King's avatar
Davis King committed
210
211
212
    // Another thing that is worth knowing is that just about everything in dlib is
    // serializable.  So for example, you can save the learned_pfunct object to disk and
    // recall it later like so:
213
    serialize("saved_function.dat") << learned_pfunct;
214

Davis King's avatar
Davis King committed
215
    // Now let's open that file back up and load the function object it contains.
216
    deserialize("saved_function.dat") >> learned_pfunct;
217

Davis King's avatar
Davis King committed
218
219
220
221
222
    // Note that there is also an example program that comes with dlib called the
    // file_to_code_ex.cpp example.  It is a simple program that takes a file and outputs a
    // piece of C++ code that is able to fully reproduce the file's contents in the form of
    // a std::string object.  So you can use that along with the std::istringstream to save
    // learned decision functions inside your actual C++ code files if you want.  
223
224
225



226

Davis King's avatar
Davis King committed
227
    // Lastly, note that the decision functions we trained above involved well over 200
Davis King's avatar
Davis King committed
228
    // basis vectors.  Support vector machines in general tend to find decision functions
Davis King's avatar
Davis King committed
229
230
231
232
233
234
235
236
237
238
    // that involve a lot of basis vectors.  This is significant because the more basis
    // vectors in a decision function, the longer it takes to classify new examples.  So
    // dlib provides the ability to find an approximation to the normal output of a trainer
    // using fewer basis vectors.  

    // Here we determine the cross validation accuracy when we approximate the output using
    // only 10 basis vectors.  To do this we use the reduced2() function.  It takes a
    // trainer object and the number of basis vectors to use and returns a new trainer
    // object that applies the necessary post processing during the creation of decision
    // function objects.
239
240
241
    cout << "\ncross validation accuracy with only 10 support vectors: " 
         << cross_validate_trainer(reduced2(trainer,10), samples, labels, 3);

Davis King's avatar
Davis King committed
242
    // Let's print out the original cross validation score too for comparison.
243
244
245
    cout << "cross validation accuracy with all the original support vectors: " 
         << cross_validate_trainer(trainer, samples, labels, 3);

Davis King's avatar
Davis King committed
246
247
    // When you run this program you should see that, for this problem, you can reduce the
    // number of basis vectors down to 10 without hurting the cross validation accuracy. 
248
249
250


    // To get the reduced decision function out we would just do this:
251
    learned_function.function = reduced2(trainer,10).train(samples, labels);
252
    // And similarly for the probabilistic_decision_function: 
253
    learned_pfunct.function = train_probabilistic_decision_function(reduced2(trainer,10), samples, labels, 3);
254
255
}