Merge pull request #4166 from durovo/feature/platt

Fix platt scaling

examples/undocumented/libshogun/classifier_libsvm_probabilities.cpp

@@ -1,9 +1,10 @@
+#include <iostream>
 #include <shogun/base/init.h>
+#include <shogun/classifier/svm/LibSVM.h>
+#include <shogun/evaluation/SigmoidCalibration.h>
 #include <shogun/features/DenseFeatures.h>
 #include <shogun/kernel/LinearKernel.h>
 #include <shogun/labels/BinaryLabels.h>
-#include <shogun/classifier/svm/LibSVM.h>
-#include <iostream>
 
 using namespace shogun;
 
@@ -50,45 +51,53 @@ int main(int argc, char** argv)
 
     //create train labels
     CLabels* labels=new CBinaryLabels(labelVector);
-
-    //create train features
-    CDenseFeatures<float64_t>* features=new CDenseFeatures<float64_t>();
-    SG_REF(features);
-    features->set_feature_matrix(featureMatrix);
-
-    //create linear kernel
-    CLinearKernel* kernel=new CLinearKernel();
-    SG_REF(kernel);
-    kernel->init(features, features);
-
-    //create svm classifier by LibSVM
-    CLibSVM* svm=new CLibSVM(svm_C,kernel, labels);
-    SG_REF(svm);
-    svm->train();
-
-    //classify data points
-    CBinaryLabels* out_labels=CLabelsFactory::to_binary(svm->apply());
-
-    /*convert scores to calibrated probabilities  by fitting a sigmoid function
-    using the method described in Lin, H., Lin, C., and Weng,  R. (2007). A note
-    on Platt's probabilistic outputs for support vector machines.
-    See BinaryLabels documentation for details*/
-    out_labels->scores_to_probabilities();
-
-    //display output labels and probabilities
-    for (int32_t i=0; i<num_samples; i++)
-    {
-        SG_SPRINT("out[%d]=%f (%f)\n", i, out_labels->get_label(i),
-            out_labels->get_value(i));
-    }
-
-    //clean up
-    SG_UNREF(out_labels);
-    SG_UNREF(kernel);
-    SG_UNREF(features);
-    SG_UNREF(svm);
-
-    exit_shogun();
-
-    return 0;
+	SG_REF(labels);
+
+	// create train features
+	CDenseFeatures<float64_t>* features = new CDenseFeatures<float64_t>();
+	SG_REF(features);
+	features->set_feature_matrix(featureMatrix);
+
+	// create linear kernel
+	CLinearKernel* kernel = new CLinearKernel();
+	SG_REF(kernel);
+	kernel->init(features, features);
+
+	// create svm classifier by LibSVM
+	CLibSVM* svm = new CLibSVM(svm_C, kernel, labels);
+	SG_REF(svm);
+	svm->train();
+
+	// classify data points
+	CBinaryLabels* out_labels = CLabelsFactory::to_binary(svm->apply());
+
+	/*convert scores to calibrated probabilities  by fitting a sigmoid function
+	using the method described in Lin, H., Lin, C., and Weng,  R. (2007). A note
+	on Platt's probabilistic outputs for support vector machines.
+	See BinaryLabels documentation for details*/
+	CSigmoidCalibration* sigmoid_calibration = new CSigmoidCalibration();
+	sigmoid_calibration->fit_binary(
+	    out_labels, CLabelsFactory::to_binary(labels));
+	CBinaryLabels* calibrated_labels =
+	    sigmoid_calibration->calibrate_binary(out_labels);
+
+	// display output labels and probabilities
+	for (int32_t i = 0; i < num_samples; i++)
+	{
+		SG_SPRINT(
+		    "out[%d]=%f (%f)\n", i, out_labels->get_label(i),
+		    calibrated_labels->get_value(i));
+	}
+
+	// clean up
+	SG_UNREF(sigmoid_calibration);
+	SG_UNREF(calibrated_labels);
+	SG_UNREF(out_labels);
+	SG_UNREF(kernel);
+	SG_UNREF(features);
+	SG_UNREF(svm);
+
+	exit_shogun();
+
+	return 0;
 }

examples/undocumented/libshogun/labels_binary_fit_sigmoid.cpp

@@ -5,23 +5,32 @@
  */
 
 #include <shogun/base/init.h>
+#include <shogun/evaluation/SigmoidCalibration.h>
 #include <shogun/labels/BinaryLabels.h>
 
 using namespace shogun;
 
 void test_sigmoid_fitting()
 {
 	CBinaryLabels* labels=new CBinaryLabels(10);
+	CBinaryLabels* predictions = new CBinaryLabels(10);
 	labels->set_values(SGVector<float64_t>(labels->get_num_labels()));
 
 	for (index_t i=0; i<labels->get_num_labels(); ++i)
-		labels->set_value(i%2==0 ? 1 : -1, i);
-
-	labels->get_values().display_vector("scores");
-	labels->scores_to_probabilities();
-	labels->get_values().display_vector("probabilities");
-
-
+	{
+		predictions->set_value(i % 2 == 0 ? 1 : -1, i);
+		labels->set_value(i % 4 == 0 ? 1 : -1, i);
+	}
+	predictions->get_values().display_vector("scores");
+	CSigmoidCalibration* sigmoid_calibration = new CSigmoidCalibration();
+	sigmoid_calibration->fit_binary(predictions, labels);
+	CBinaryLabels* calibrated_labels =
+	    sigmoid_calibration->calibrate_binary(predictions);
+	calibrated_labels->get_values().display_vector("probabilities");
+
+	SG_UNREF(predictions);
+	SG_UNREF(sigmoid_calibration);
+	SG_UNREF(calibrated_labels);
 	SG_UNREF(labels);
 }
 

src/shogun/evaluation/Calibration.h

@@ -0,0 +1,69 @@
+/*
+ * This software is distributed under BSD 3-clause license (see LICENSE file).
+ *
+ * Authors: Heiko Strathmann, Dhruv Arya
+ */
+
+#ifndef _CALIBRATION_H__
+#define _CALIBRATION_H__
+
+#include <shogun/lib/config.h>
+
+#include <shogun/machine/Machine.h>
+
+namespace shogun
+{
+	/** @brief Base class for all calibration methods. Call fit_binary() or
+	fit_multiclass() to
+	fit the parameters on the predictions and the true labels. Call calibrate to
+	calibrate predictions. **/
+	class CCalibration : public CSGObject
+	{
+	public:
+		/** Constructor. */
+		CCalibration()
+		{
+		}
+		/** Destructor. */
+		virtual ~CCalibration()
+		{
+		}
+
+		virtual const char* get_name() const
+		{
+			return "Calibration";
+		}
+
+		/** Fit calibration parameters for binary labels.
+		* @param predictions The predictions outputted by the machine
+		* @param targets The true labels corresponding to the predictions
+		* @return Indicates whether the calibration was succesful
+		**/
+		virtual bool
+		fit_binary(CBinaryLabels* predictions, CBinaryLabels* targets) = 0;
+
+		/** Calibrate binary predictions based on parameters learned by calling
+		*fit.
+		* @param predictions The predictions outputted by the machine
+		* @return Calibrated binary labels
+		**/
+		virtual CBinaryLabels* calibrate_binary(CBinaryLabels* predictions) = 0;
+
+		/** Fit calibration parameters for multiclass labels.
+		* @param predictions The predictions outputted by the machine
+		* @param targets The true labels corresponding to the predictions
+		* @return Indicates whether the calibration was succesful
+		**/
+		virtual bool fit_multiclass(
+		    CMulticlassLabels* predictions, CMulticlassLabels* targets) = 0;
+
+		/** Calibrate multiclass predictions based on parameters learned by
+		*calling fit.
+		* @param predictions The predictions outputted by the machine
+		* @return Calibrated binary labels
+		**/
+		virtual CMulticlassLabels*
+		calibrate_multiclass(CMulticlassLabels* predictions) = 0;
+	};
+}
+#endif