SVMSGD class added

modules/ml/include/opencv2/ml.hpp

@@ -1513,6 +1513,127 @@ CV_EXPORTS void randMVNormal( InputArray mean, InputArray cov, int nsamples, Out
 CV_EXPORTS void createConcentricSpheresTestSet( int nsamples, int nfeatures, int nclasses,
                                                 OutputArray samples, OutputArray responses);
 
+/****************************************************************************************\
+*                        Stochastic Gradient Descent SVM Classifier                      *
+\****************************************************************************************/
+
+/*!
+@brief Stochastic Gradient Descent SVM classifier
+
+SVMSGD provides a fast and easy-to-use implementation of the SVM classifier using the Stochastic Gradient Descent approach, as presented in @cite bottou2010large.
+The gradient descent show amazing performance for large-scale problems, reducing the computing time. This allows a fast and reliable online update of the classifier for each new feature which
+is fundamental when dealing with variations of data over time (like weather and illumination changes in videosurveillance, for example).
+
+First, create the SVMSGD object. To enable the online update, a value for updateFrequency should be defined.
+
+Then the SVM model can be trained using the train features and the correspondent labels.
+
+After that, the label of a new feature vector can be predicted using the predict function. If the updateFrequency was defined in the constructor, the predict function will update the weights automatically.
+
+@code
+// Initialize object
+SVMSGD SvmSgd;
+
+// Train the Stochastic Gradient Descent SVM
+SvmSgd.train(trainFeatures, labels);
+
+// Predict label for the new feature vector (1xM)
+predictedLabel = SvmSgd.predict(newFeatureVector);
+@endcode
+
+*/
+class CV_EXPORTS_W SVMSGD {
+
+    public:
+        /** @brief SGDSVM constructor.
+
+        @param lambda regularization
+        @param learnRate learning rate
+        @param nIterations number of training iterations
+
+        */
+        SVMSGD(float lambda = 0.000001, float learnRate = 2, uint nIterations = 100000);
+
+        /** @brief SGDSVM constructor.
+
+        @param updateFrequency online update frequency
+        @param learnRateDecay learn rate decay over time: learnRate = learnRate * learnDecay
+        @param lambda regularization
+        @param learnRate learning rate
+        @param nIterations number of training iterations
+
+        */
+        SVMSGD(uint updateFrequency, float learnRateDecay = 1, float lambda = 0.000001, float learnRate = 2, uint nIterations = 100000);
+        virtual ~SVMSGD();
+        virtual SVMSGD* clone() const;
+
+        /** @brief Train the SGDSVM classifier.
+
+        The function trains the SGDSVM classifier using the train features and the correspondent labels (-1 or 1).
+
+        @param trainFeatures features used for training. Each row is a new sample.
+        @param labels mat (size Nx1 with N = number of features) with the label of each training feature.
+
+        */
+        virtual void train(cv::Mat trainFeatures, cv::Mat labels);
+
+        /** @brief Predict the label of a new feature vector.
+
+        The function predicts and returns the label of a new feature vector, using the previously trained SVM model.
+
+        @param newFeature new feature vector used for prediction
+
+        */
+        virtual float predict(cv::Mat newFeature);
+
+        /** @brief Returns the weights of the trained model.
+
+        */
+        virtual std::vector<float> getWeights(){ return _weights; };
+
+        /** @brief Sets the weights of the trained model.
+
+        @param weights weights used to predict the label of a new feature vector.
+
+        */
+        virtual void setWeights(std::vector<float> weights){ _weights = weights; };
+
+    private:
+        void updateWeights();
+        void generateRandomIndex();
+        float calcInnerProduct(float *rowDataPointer);
+        void updateWeights(float innerProduct, float *rowDataPointer, int label);
+
+        // Vector with SVM weights
+        std::vector<float> _weights;
+
+        // Random index generation
+        long long int _randomNumber;
+        unsigned int _randomIndex;
+
+        // Number of features and samples
+        unsigned int _nFeatures;
+        unsigned int _nTrainSamples;
+
+        // Parameters for learning
+        float _lambda;  //regularization
+        float _learnRate;  //learning rate
+        unsigned int _nIterations; //number of training iterations
+
+        // Vars to control the features slider matrix
+        bool _onlineUpdate;
+        bool _initPredict;
+        uint _slidingWindowSize;
+        uint _predictSlidingWindowSize;
+        float* _labelSlider;
+        float _learnRateDecay;
+
+        // Mat with features slider and correspondent counter
+        unsigned int _sliderCounter;
+        cv::Mat _featuresSlider;
+
+};
+
 //! @} ml
 
 }