SGD.h

#ifndef __SGD_H__
#define __SGD_H__
 
#include <MedAlgo/MedAlgo/PredictiveModel.h>
#include <random>
#include <map>
 
using namespace std;
 
class SGD
{
public:
    SGD(PredictiveModel *mdl, double(*loss_funct)(const vector<double> &got, const vector<float> &y, const vector<float> *weights));
    void Learn(const vector<vector<float>> &xData, const vector<float> &yData, int T_Steps, const vector<float> *weights = NULL, bool print_auc = false);
    //use one of these techniques to set params for subgradients
    double(*subGradientI)(int param_number, const vector<double> &param_values, const vector<vector<float>> &x, const vector<float> &y, const vector<float> *weights);
    void set_gradient_params(int samplePointCnt, float h, int minSampForCat = 0);
 
    //use one of them to set learning_rate
    void set_learing_rate(float val);
    void set_learing(float blockVals, float blockDerivate, int T_steps);
    void set_special_step_func(double(*function)(const vector<double> &, const vector<float> &, const vector<double> &, const vector<float> *));
 
    //use to block parameter search, not mandatory: pass val < 0 to force each step projection to abs(val)
    void set_blocking(float val);
 
    //use to round model params - act as some regularization 
    void set_model_precision(double val);
    double get_model_precision();
 
    float get_learing_rate();
    float get_learing_eppsilon(float blockVals, float blockDerivate, int T_steps);
    PredictiveModel *get_model();
    float get_blocking();
 
    size_t output_num; //The number of lines before printing output, 0= no printing
    bool norm_l1;
private:
    PredictiveModel * _model;
    vector<PredictiveModel *> _models_par;
    double(*loss_function)(const vector<double> &got, const vector<float> &y, const vector<float> *weights);
    double(*step_loss_function)(const vector<double> &got, const vector<float> &y, const vector<double> &model_params, const vector<float> *weights);
    float _learning_rate;
    bool has_learing_rate;
    int _minSampForCat;
    double _min_precision;
    map<float, vector<int>> _categoryIndex;
 
    float _h;
    int _sampleSize;
    mt19937 _gen;
 
    float _blocking_val;
 
    //allocate memory for step function once
    vector<float> _sampleY;
    vector<float> _sampleW;
    vector<vector<float>> _sampleX;
    vector<double> _preds_plus;
    vector<double> _preds_minus;
    vector<double> _preds_base;
 
    vector<double> _step(const vector<vector<float>> &xData, const vector<float> &yData, const vector<float> *weights);
    void _projection_step(vector<double> &params);
    void _round_step(vector<double> &params);
};
 
void factor(vector<double> &a, double fact);
void add(vector<double> &a, const vector<double> &b);
double product(const vector<double> &a, const vector<double> &b);
 
#endif