Merge pull request #752 from iglesias/master

Loss included in multiclass argmax
shogun-toolbox · Aug 22, 2012 · 8cec5d0 · 8cec5d0
2 parents 50ba31c + bbb3457
commit 8cec5d0
Show file tree

Hide file tree

Showing 6 changed files with 68 additions and 9 deletions.
diff --git a/examples/undocumented/libshogun/Makefile b/examples/undocumented/libshogun/Makefile
@@ -96,6 +96,7 @@ TARGETS = basic_minimal \
 		  statistics_linear_time_mmd_kernel_choice \
 		  statistics_hsic \
 		  classifier_featureblocklogisticregression \
+		  so_multiclass_BMRM \
 
 all: $(TARGETS)
 

diff --git a/examples/undocumented/libshogun/so_multiclass.cpp b/examples/undocumented/libshogun/so_multiclass.cpp
@@ -13,6 +13,8 @@
 #include <shogun/structure/MulticlassSOLabels.h>
 #include <shogun/structure/MulticlassModel.h>
 #include <shogun/structure/PrimalMosekSOSVM.h>
+#include <shogun/structure/DualLibQPBMSOSVM.h>
+#include <shogun/lib/Time.h>
 
 using namespace shogun;
 
@@ -59,14 +61,37 @@ void gen_rand_data(SGVector< float64_t > labs, SGMatrix< float64_t > feats)
 	fclose(pfile);
 }
 
+void read_data(SGVector< float64_t > labs, SGMatrix< float64_t > feats)
+{
+	FILE* pfile = fopen(FNAME, "r");
+	int32_t label, idx;
+	float32_t value;
+
+	for ( int32_t i = 0 ; i < NUM_SAMPLES*NUM_CLASSES ; ++i )
+	{
+		fscanf(pfile, "%d", &label);
+
+		labs[i] = label;
+
+		for ( int32_t j = 0 ; j < DIMS ; ++j )
+		{
+			fscanf(pfile, "%d:%f", &idx, &value);
+			feats[i*DIMS + j] = value;
+		}
+	}
+
+	fclose(pfile);
+}
+
 int main(int argc, char ** argv)
 {
 	init_shogun_with_defaults();
 
 	SGVector< float64_t > labs(NUM_CLASSES*NUM_SAMPLES);
 	SGMatrix< float64_t > feats(DIMS, NUM_CLASSES*NUM_SAMPLES);
 
-	gen_rand_data(labs, feats);
+	//gen_rand_data(labs, feats);
+	read_data(labs, feats);
 
 	// Create train labels
 	CMulticlassSOLabels* labels = new CMulticlassSOLabels(labs);
@@ -82,11 +107,20 @@ int main(int argc, char ** argv)
 	CHingeLoss* loss = new CHingeLoss();
 
 	// Create SO-SVM
-	CPrimalMosekSOSVM* sosvm = new CPrimalMosekSOSVM(model, loss, labels, features);
+	CPrimalMosekSOSVM* sosvm = new CPrimalMosekSOSVM(model, loss, labels);
+	CDualLibQPBMSOSVM* bundle = new CDualLibQPBMSOSVM(model, loss, labels, 1000);
+	bundle->set_verbose(false);
 	SG_REF(sosvm);
+	SG_REF(bundle);
 
+	CTime start;
+	float64_t t1;
 	sosvm->train();
+	SG_SPRINT(">>>> PrimalMosekSOSVM trained in %9.4f\n", (t1 = start.cur_time_diff(false)));
+	bundle->train();
+	SG_SPRINT(">>>> BMRM trained in %9.4f\n", start.cur_time_diff(false)-t1);
 	CStructuredLabels* out = CStructuredLabels::obtain_from_generic(sosvm->apply());
+	CStructuredLabels* bout = CStructuredLabels::obtain_from_generic(bundle->apply());
 
 	// Create liblinear svm classifier with L2-regularized L2-loss
 	CLibLinear* svm = new CLibLinear(L2R_L2LOSS_SVC);
@@ -127,16 +161,18 @@ int main(int argc, char ** argv)
 	SG_REF(multiclass_evaluator);
 
 	SG_SPRINT("SO-SVM: %5.2f%\n", 100.0*structured_evaluator->evaluate(out, labels));
+	SG_SPRINT("BMRM:   %5.2f%\n", 100.0*structured_evaluator->evaluate(bout, labels));
 	SG_SPRINT("MC:     %5.2f%\n", 100.0*multiclass_evaluator->evaluate(mout, mlabels));
 
 	// Free memory
 	SG_UNREF(multiclass_evaluator);
 	SG_UNREF(structured_evaluator);
 	SG_UNREF(mout);
 	SG_UNREF(mc_svm);
+	SG_UNREF(bundle);
 	SG_UNREF(sosvm);
+	SG_UNREF(bout);
 	SG_UNREF(out);
-
 	exit_shogun();
 
 	return 0;

diff --git a/examples/undocumented/libshogun/so_multiclass_BMRM.cpp b/examples/undocumented/libshogun/so_multiclass_BMRM.cpp
@@ -139,7 +139,6 @@ int main(int argc, char * argv[])
 				model,
 				loss,
 				labels,
-				features,
 				lambda);
 	SG_REF(sosvm);
 

diff --git a/src/shogun/structure/MulticlassModel.cpp b/src/shogun/structure/MulticlassModel.cpp
@@ -81,11 +81,13 @@ CResultSet* CMulticlassModel::argmax(
 	// Find the class that gives the maximum score
 
 	float64_t score = 0, ypred = 0;
-	float64_t max_score = df->dense_dot(feat_idx, w.vector, feats_dim);
+	float64_t max_score = -CMath::INFTY;
 
-	for ( int32_t c = 1 ; c < m_num_classes ; ++c )
+	for ( int32_t c = 0 ; c < m_num_classes ; ++c )
 	{
 		score = df->dense_dot(feat_idx, w.vector+c*feats_dim, feats_dim);
+		if ( training )
+			score += delta_loss(feat_idx, c);
 
 		if ( score > max_score )
 		{
@@ -119,7 +121,24 @@ float64_t CMulticlassModel::delta_loss(CStructuredData* y1, CStructuredData* y2)
 	ASSERT(rn1 != NULL);
 	ASSERT(rn2 != NULL);
 
-	return ( rn1->value == rn2->value ) ? 0 : 1;
+	return delta_loss(rn1->value, rn2->value);
+}
+
+float64_t CMulticlassModel::delta_loss(int32_t y1_idx, float64_t y2)
+{
+	REQUIRE(y1_idx >= 0 || y1_idx < m_labels->get_num_labels(),
+			"The label index must be inside [0, num_labels-1]\n");
+
+	CRealNumber* rn1 = CRealNumber::obtain_from_generic(m_labels->get_label(y1_idx));
+	float64_t ret = delta_loss(rn1->value, y2);
+	SG_UNREF(rn1);
+
+	return ret;
+}
+
+float64_t CMulticlassModel::delta_loss(float64_t y1, float64_t y2)
+{
+	return (y1 == y2) ? 0 : 1;
 }
 
 void CMulticlassModel::init_opt(

diff --git a/src/shogun/structure/MulticlassModel.h b/src/shogun/structure/MulticlassModel.h
@@ -133,6 +133,10 @@ class CMulticlassModel : public CStructuredModel
 	private:
 		void init();
 
+		/** Different flavours of the delta_loss that become handy */
+		float64_t delta_loss(float64_t y1, float64_t y2);
+		float64_t delta_loss(int32_t y1_idx, float64_t y2);
+
 	private:
 		/** number of classes */
 		int32_t m_num_classes;

diff --git a/src/shogun/structure/StructuredModel.cpp b/src/shogun/structure/StructuredModel.cpp
@@ -98,8 +98,8 @@ SGVector< float64_t > CStructuredModel::get_joint_feature_vector(
 
 float64_t CStructuredModel::delta_loss(int32_t ytrue_idx, CStructuredData* ypred)
 {
-	if ( ytrue_idx < 0 || ytrue_idx >= m_labels->get_num_labels() )
-		SG_ERROR("The label index must be inside [0, num_labels-1]\n");
+	REQUIRE(ytrue_idx >= 0 || ytrue_idx < m_labels->get_num_labels(),
+			"The label index must be inside [0, num_labels-1]\n");
 
 	CStructuredData* ytrue = m_labels->get_label(ytrue_idx);
 	float64_t ret = delta_loss(ytrue, ypred);