TrainNodeMsRF.h

// Random Forest (based on Microsof Sherwood library) training class interface
// Written by Sergey G. Kosov in 2013 for Project X
#pragma once

#include "TrainNode.h"
#include "SamplesAccumulator.h"

//#ifdef USE_SHERWOOD

namespace MicrosoftResearch { namespace Cambridge { namespace Sherwood {
    class LinearFeatureResponse;
    class HistogramAggregator;
    class DataPointCollection;
    template<class F, class S> class Forest;
    struct TrainingParameters;
}}}

namespace sw = MicrosoftResearch::Cambridge::Sherwood;

namespace DirectGraphicalModels
{
    typedef struct TrainNodeMsRFParams {
        int             max_decision_levels;                        
        int             num_of_candidate_features;                  
        unsigned int    num_of_candidate_thresholds_per_feature;    
        int             num_ot_trees;                               
        bool            verbose;                                    
        size_t          maxSamples;                                 

        TrainNodeMsRFParams() {}
        TrainNodeMsRFParams(int _max_decision_levels, int _num_of_candidate_features, unsigned int _num_of_candidate_thresholds_per_feature, int _num_ot_trees, bool _verbose, int _maxSamples) : max_decision_levels(_max_decision_levels), num_of_candidate_features(_num_of_candidate_features), num_of_candidate_thresholds_per_feature(_num_of_candidate_thresholds_per_feature), num_ot_trees(_num_ot_trees), verbose(_verbose), maxSamples(_maxSamples) {}
    } TrainNodeMsRFParams;

    const TrainNodeMsRFParams TRAIN_NODE_MS_RF_PARAMS_DEFAULT = TrainNodeMsRFParams(
                                                                10,     // Maximum number of the decision levels
                                                                10,     // Number of candidate features
                                                                10,     // Number of candidate thresholds (per feature)
                                                                10,     // Number of trees in the forest (time / accuracy)
                                                                false,  // Verbose mode
                                                                0       // Maximum number of samples to be used in training. 0 means using all the samples
                                                                );
    
    // =========================== Microsoft RF Train Class ===========================
    class CTrainNodeMsRF : public CTrainNode
    {
    public:
        DllExport CTrainNodeMsRF(byte nStates, word nFeatures, TrainNodeMsRFParams params = TRAIN_NODE_MS_RF_PARAMS_DEFAULT);
        DllExport CTrainNodeMsRF(byte nStates, word nFeatures, size_t maxSamples);
        DllExport virtual ~CTrainNodeMsRF(void);

        DllExport void  reset(void);
        DllExport void  save(const std::string &path, const std::string &name = std::string(), short idx = -1) const; 
        DllExport void  load(const std::string &path, const std::string &name = std::string(), short idx = -1); 

        DllExport void  addFeatureVec(const Mat &featureVector, byte gt);
        DllExport void  train(bool doClean = false);


    protected:
        DllExport void saveFile(FILE *pFile) const { }
        DllExport void loadFile(FILE *pFile) { }
        DllExport void calculateNodePotentials(const Mat &featureVector, Mat &potential, Mat &mask) const;


    private:
        void          init(TrainNodeMsRFParams params);                                                 // This function is called by both constructors


    private:
        std::unique_ptr<sw::Forest<sw::LinearFeatureResponse, sw::HistogramAggregator>>     m_pRF;            
        std::unique_ptr<CSamplesAccumulator>                                                m_pSamplesAcc;    
        std::unique_ptr<sw::TrainingParameters>                                             m_pParams;
    };
}
//#endif