Added an option to solve the L2-loss version of the SVM objective function.

dlib/svm/svm_c_linear_dcd_trainer.h

@@ -47,7 +47,8 @@ namespace dlib
             verbose(false),
             have_bias(true),
             last_weight_1(false),
-            do_shrinking(true)
+            do_shrinking(true),
+            do_svm_l2(false)
         {
         }
 
@@ -61,7 +62,8 @@ namespace dlib
             verbose(false),
             have_bias(true),
             last_weight_1(false),
-            do_shrinking(true)
+            do_shrinking(true),
+            do_svm_l2(false)
         {
             // make sure requires clause is not broken
             DLIB_ASSERT(0 < C_,
@@ -104,6 +106,13 @@ namespace dlib
             bool enabled
         ) { do_shrinking = enabled; }
 
+        bool solving_svm_l2_problem (
+        ) const { return do_svm_l2; }
+
+        void solve_svm_l2_problem (
+            bool enabled
+        ) { do_svm_l2 = enabled; }
+
         void be_verbose (
         )
         {
@@ -497,11 +506,15 @@ namespace dlib
             std::vector<long>& index = state.index;
             const long dims = state.dims;
 
+
             unsigned long active_size = index.size();
 
             scalar_type PG_max_prev = std::numeric_limits<scalar_type>::infinity();
             scalar_type PG_min_prev = -std::numeric_limits<scalar_type>::infinity();
 
+            const scalar_type Dii_pos = 1/(2*Cpos);
+            const scalar_type Dii_neg = 1/(2*Cneg);
+
             // main loop
             for (unsigned long iter = 0; iter < max_iterations; ++iter)
             {
@@ -521,8 +534,16 @@ namespace dlib
                 {
                     const long i = index[ii];
 
-                    const scalar_type G = y(i)*dot(w, x(i)) - 1;
+                    scalar_type G = y(i)*dot(w, x(i)) - 1;
+                    if (do_svm_l2)
+                    {
+                        if (y(i) > 0)
+                            G += Dii_pos*alpha[i];
+                        else
+                            G += Dii_neg*alpha[i];
+                    }
                     const scalar_type C = (y(i) > 0) ? Cpos : Cneg;
+                    const scalar_type U = do_svm_l2 ? std::numeric_limits<scalar_type>::infinity() : C;
 
                     scalar_type PG = 0;
                     if (alpha[i] == 0)
@@ -539,7 +560,7 @@ namespace dlib
                         if (G < 0)
                             PG = G;
                     }
-                    else if (alpha[i] == C)
+                    else if (alpha[i] == U)
                     {
                         if (G < PG_min_prev)
                         {
@@ -567,7 +588,7 @@ namespace dlib
                     if (std::abs(PG) > 1e-12)
                     {
                         const scalar_type alpha_old = alpha[i];
-                        alpha[i] = std::min(std::max(alpha[i] - G/state.Q[i], (scalar_type)0.0), C);
+                        alpha[i] = std::min(std::max(alpha[i] - G/state.Q[i], (scalar_type)0.0), U);
                         const scalar_type delta = (alpha[i]-alpha_old)*y(i);
                         add_to(w, x(i), delta);
                         if (have_bias && !last_weight_1)
@@ -660,6 +681,7 @@ namespace dlib
         bool have_bias; // having a bias means we pretend all x vectors have an extra element which is always -1.
         bool last_weight_1;
         bool do_shrinking;
+        bool do_svm_l2;
 
     }; // end of class svm_c_linear_dcd_trainer
 

dlib/svm/svm_c_linear_dcd_trainer_abstract.h

@@ -67,6 +67,7 @@ namespace dlib
                 - #forces_last_weight_to_1() == false
                 - #includes_bias() == true
                 - #shrinking_enabled() == true
+                - #solving_svm_l2_problem() == false
         !*/
 
         explicit svm_c_linear_dcd_trainer (
@@ -86,6 +87,7 @@ namespace dlib
                 - #forces_last_weight_to_1() == false
                 - #includes_bias() == true
                 - #shrinking_enabled() == true
+                - #solving_svm_l2_problem() == false
         !*/
 
         bool includes_bias (
@@ -140,6 +142,23 @@ namespace dlib
                 - #shrinking_enabled() == enabled
         !*/
 
+        bool solving_svm_l2_problem (
+        ) const; 
+        /*!
+            ensures
+                - returns true if this solver will solve the L2 version of the SVM
+                  objective function.  That is, if solving_svm_l2_problem()==true then this
+                  object, rather than using the hinge loss, uses the squared hinge loss.
+        !*/
+
+        void solve_svm_l2_problem (
+            bool enabled
+        ); 
+        /*!
+            ensures
+                - #solving_svm_l2_problem() == enabled
+        !*/
+
         void be_verbose (
         );
         /*!

dlib/test/svm_c_linear_dcd.cpp

@@ -440,6 +440,65 @@ namespace
 
 // ----------------------------------------------------------------------------------------
 
+    void test_l2_version ()
+    {
+        typedef std::map<unsigned long,double> sample_type;
+        typedef sparse_linear_kernel<sample_type> kernel_type;
+
+        svm_c_linear_dcd_trainer<kernel_type> linear_trainer;
+        linear_trainer.set_c(10);
+        linear_trainer.set_epsilon(1e-5);
+
+        std::vector<sample_type> samples;
+        std::vector<double> labels;
+
+        // make an instance of a sample vector so we can use it below
+        sample_type sample;
+
+
+        // Now let's go into a loop and randomly generate 10000 samples.
+        double label = +1;
+        for (int i = 0; i < 1000; ++i)
+        {
+            // flip this flag
+            label *= -1;
+
+            sample.clear();
+
+            // now make a random sparse sample with at most 10 non-zero elements
+            for (int j = 0; j < 10; ++j)
+            {
+                int idx = std::rand()%100;
+                double value = static_cast<double>(std::rand())/RAND_MAX;
+
+                sample[idx] = label*value;
+            }
+
+            // Also save the samples we are generating so we can let the svm_c_linear_trainer
+            // learn from them below.  
+            samples.push_back(sample);
+            labels.push_back(label);
+        }
+
+        decision_function<kernel_type> df = linear_trainer.train(samples, labels);
+
+        sample.clear();
+        sample[4] = 0.3;
+        sample[10] = 0.9;
+        DLIB_TEST(df(sample) > 0);
+
+        sample.clear();
+        sample[83] = -0.3;
+        sample[26] = -0.9;
+        sample[58] = -0.7;
+        DLIB_TEST(df(sample) < 0);
+
+        sample.clear();
+        sample[0] = -0.2;
+        sample[9] = -0.8;
+        DLIB_TEST(df(sample) < 0);
+    }
+
     class tester_svm_c_linear_dcd : public tester
     {
     public:
@@ -474,6 +533,8 @@ namespace
             print_spinner();
             test_sparse_1_sample(-1);
             print_spinner();
+
+            test_l2_version();
         }
     } a;