// Class: ReadMLP
// Automatically generated by MethodBase::MakeClass
//

/* configuration options =====================================================

#GEN -*-*-*-*-*-*-*-*-*-*-*- general info -*-*-*-*-*-*-*-*-*-*-*-

Method         : MLP::MLP
TMVA Release   : 4.2.0         [262656]
ROOT Release   : 5.34/14       [336398]
Creator        : hepg
Date           : Sun Aug 16 07:58:38 2015
Host           : Linux whale1 2.6.32-358.el6.x86_64 #1 SMP Fri Feb 22 03:57:22 CET 2013 x86_64 x86_64 x86_64 GNU/Linux
Dir            : /misc/users/hepg/cowan/tmva/train
Training events: 20000
Analysis type  : [Classification]


#OPT -*-*-*-*-*-*-*-*-*-*-*-*- options -*-*-*-*-*-*-*-*-*-*-*-*-

# Set by User:
HiddenLayers: "3" [Specification of hidden layer architecture]
V: "False" [Verbose output (short form of "VerbosityLevel" below - overrides the latter one)]
H: "True" [Print method-specific help message]
# Default:
NCycles: "500" [Number of training cycles]
NeuronType: "sigmoid" [Neuron activation function type]
RandomSeed: "1" [Random seed for initial synapse weights (0 means unique seed for each run; default value '1')]
EstimatorType: "MSE" [MSE (Mean Square Estimator) for Gaussian Likelihood or CE(Cross-Entropy) for Bernoulli Likelihood]
NeuronInputType: "sum" [Neuron input function type]
VerbosityLevel: "Default" [Verbosity level]
VarTransform: "None" [List of variable transformations performed before training, e.g., "D_Background,P_Signal,G,N_AllClasses" for: "Decorrelation, PCA-transformation, Gaussianisation, Normalisation, each for the given class of events ('AllClasses' denotes all events of all classes, if no class indication is given, 'All' is assumed)"]
CreateMVAPdfs: "False" [Create PDFs for classifier outputs (signal and background)]
IgnoreNegWeightsInTraining: "False" [Events with negative weights are ignored in the training (but are included for testing and performance evaluation)]
TrainingMethod: "BP" [Train with Back-Propagation (BP), BFGS Algorithm (BFGS), or Genetic Algorithm (GA - slower and worse)]
LearningRate: "2.000000e-02" [ANN learning rate parameter]
DecayRate: "1.000000e-02" [Decay rate for learning parameter]
TestRate: "10" [Test for overtraining performed at each #th epochs]
EpochMonitoring: "False" [Provide epoch-wise monitoring plots according to TestRate (caution: causes big ROOT output file!)]
Sampling: "1.000000e+00" [Only 'Sampling' (randomly selected) events are trained each epoch]
SamplingEpoch: "1.000000e+00" [Sampling is used for the first 'SamplingEpoch' epochs, afterwards, all events are taken for training]
SamplingImportance: "1.000000e+00" [ The sampling weights of events in epochs which successful (worse estimator than before) are multiplied with SamplingImportance, else they are divided.]
SamplingTraining: "True" [The training sample is sampled]
SamplingTesting: "False" [The testing sample is sampled]
ResetStep: "50" [How often BFGS should reset history]
Tau: "3.000000e+00" [LineSearch "size step"]
BPMode: "sequential" [Back-propagation learning mode: sequential or batch]
BatchSize: "-1" [Batch size: number of events/batch, only set if in Batch Mode, -1 for BatchSize=number_of_events]
ConvergenceImprove: "1.000000e-30" [Minimum improvement which counts as improvement (<0 means automatic convergence check is turned off)]
ConvergenceTests: "-1" [Number of steps (without improvement) required for convergence (<0 means automatic convergence check is turned off)]
UseRegulator: "False" [Use regulator to avoid over-training]
UpdateLimit: "10000" [Maximum times of regulator update]
CalculateErrors: "False" [Calculates inverse Hessian matrix at the end of the training to be able to calculate the uncertainties of an MVA value]
WeightRange: "1.000000e+00" [Take the events for the estimator calculations from small deviations from the desired value to large deviations only over the weight range]
##


#VAR -*-*-*-*-*-*-*-*-*-*-*-* variables *-*-*-*-*-*-*-*-*-*-*-*-

NVar 3
x                             x                             x                             x                                                               'F'    [-2.71237683296,3.86315917969]
y                             y                             y                             y                                                               'F'    [-1.74197530746,3.65737724304]
z                             z                             z                             z                                                               'F'    [0.000103908125311,0.999955058098]
NSpec 0


============================================================================ */

#include <vector>
#include <cmath>
#include <string>
#include <iostream>

#ifndef IClassifierReader__def
#define IClassifierReader__def

class IClassifierReader {

 public:

   // constructor
   IClassifierReader() : fStatusIsClean( true ) {}
   virtual ~IClassifierReader() {}

   // return classifier response
   virtual double GetMvaValue( const std::vector<double>& inputValues ) const = 0;

   // returns classifier status
   bool IsStatusClean() const { return fStatusIsClean; }

 protected:

   bool fStatusIsClean;
};

#endif

class ReadMLP : public IClassifierReader {

 public:

   // constructor
   ReadMLP( std::vector<std::string>& theInputVars ) 
      : IClassifierReader(),
        fClassName( "ReadMLP" ),
        fNvars( 3 ),
        fIsNormalised( false )
   {      
      // the training input variables
      const char* inputVars[] = { "x", "y", "z" };

      // sanity checks
      if (theInputVars.size() <= 0) {
         std::cout << "Problem in class \"" << fClassName << "\": empty input vector" << std::endl;
         fStatusIsClean = false;
      }

      if (theInputVars.size() != fNvars) {
         std::cout << "Problem in class \"" << fClassName << "\": mismatch in number of input values: "
                   << theInputVars.size() << " != " << fNvars << std::endl;
         fStatusIsClean = false;
      }

      // validate input variables
      for (size_t ivar = 0; ivar < theInputVars.size(); ivar++) {
         if (theInputVars[ivar] != inputVars[ivar]) {
            std::cout << "Problem in class \"" << fClassName << "\": mismatch in input variable names" << std::endl
                      << " for variable [" << ivar << "]: " << theInputVars[ivar].c_str() << " != " << inputVars[ivar] << std::endl;
            fStatusIsClean = false;
         }
      }

      // initialize min and max vectors (for normalisation)
      fVmin[0] = 0;
      fVmax[0] = 0;
      fVmin[1] = 0;
      fVmax[1] = 0;
      fVmin[2] = 0;
      fVmax[2] = 0;

      // initialize input variable types
      fType[0] = 'F';
      fType[1] = 'F';
      fType[2] = 'F';

      // initialize constants
      Initialize();

   }

   // destructor
   virtual ~ReadMLP() {
      Clear(); // method-specific
   }

   // the classifier response
   // "inputValues" is a vector of input values in the same order as the 
   // variables given to the constructor
   double GetMvaValue( const std::vector<double>& inputValues ) const;

 private:

   // method-specific destructor
   void Clear();

   // common member variables
   const char* fClassName;

   const size_t fNvars;
   size_t GetNvar()           const { return fNvars; }
   char   GetType( int ivar ) const { return fType[ivar]; }

   // normalisation of input variables
   const bool fIsNormalised;
   bool IsNormalised() const { return fIsNormalised; }
   double fVmin[3];
   double fVmax[3];
   double NormVariable( double x, double xmin, double xmax ) const {
      // normalise to output range: [-1, 1]
      return 2*(x - xmin)/(xmax - xmin) - 1.0;
   }

   // type of input variable: 'F' or 'I'
   char   fType[3];

   // initialize internal variables
   void Initialize();
   double GetMvaValue__( const std::vector<double>& inputValues ) const;

   // private members (method specific)

   double ActivationFnc(double x) const;
   double OutputActivationFnc(double x) const;

   int fLayers;
   int fLayerSize[3];
   double fWeightMatrix0to1[4][4];   // weight matrix from layer 0 to 1
   double fWeightMatrix1to2[1][4];   // weight matrix from layer 1 to 2

   double * fWeights[3];
};

inline void ReadMLP::Initialize()
{
   // build network structure
   fLayers = 3;
   fLayerSize[0] = 4; fWeights[0] = new double[4]; 
   fLayerSize[1] = 4; fWeights[1] = new double[4]; 
   fLayerSize[2] = 1; fWeights[2] = new double[1]; 
   // weight matrix from layer 0 to 1
   fWeightMatrix0to1[0][0] = 1.06392518265394;
   fWeightMatrix0to1[1][0] = 1.37643504656176;
   fWeightMatrix0to1[2][0] = 2.34030360839801;
   fWeightMatrix0to1[0][1] = 3.87729972339743;
   fWeightMatrix0to1[1][1] = -1.87080656595418;
   fWeightMatrix0to1[2][1] = -2.19317519450234;
   fWeightMatrix0to1[0][2] = 2.7098697806131;
   fWeightMatrix0to1[1][2] = 0.944867197233981;
   fWeightMatrix0to1[2][2] = -1.42651195526084;
   fWeightMatrix0to1[0][3] = -5.181580900468;
   fWeightMatrix0to1[1][3] = -3.50697717673637;
   fWeightMatrix0to1[2][3] = -3.39062345122844;
   // weight matrix from layer 1 to 2
   fWeightMatrix1to2[0][0] = -1.05073267797386;
   fWeightMatrix1to2[0][1] = 0.091066020714422;
   fWeightMatrix1to2[0][2] = -1.00538254117898;
   fWeightMatrix1to2[0][3] = 1.0517186446555;
}

inline double ReadMLP::GetMvaValue__( const std::vector<double>& inputValues ) const
{
   if (inputValues.size() != (unsigned int)fLayerSize[0]-1) {
      std::cout << "Input vector needs to be of size " << fLayerSize[0]-1 << std::endl;
      return 0;
   }

   for (int l=0; l<fLayers; l++)
      for (int i=0; i<fLayerSize[l]; i++) fWeights[l][i]=0;

   for (int l=0; l<fLayers-1; l++)
      fWeights[l][fLayerSize[l]-1]=1;

   for (int i=0; i<fLayerSize[0]-1; i++)
      fWeights[0][i]=inputValues[i];

   // layer 0 to 1
   for (int o=0; o<fLayerSize[1]-1; o++) {
      for (int i=0; i<fLayerSize[0]; i++) {
         double inputVal = fWeightMatrix0to1[o][i] * fWeights[0][i];
         fWeights[1][o] += inputVal;
      }
      fWeights[1][o] = ActivationFnc(fWeights[1][o]);
   }
   // layer 1 to 2
   for (int o=0; o<fLayerSize[2]; o++) {
      for (int i=0; i<fLayerSize[1]; i++) {
         double inputVal = fWeightMatrix1to2[o][i] * fWeights[1][i];
         fWeights[2][o] += inputVal;
      }
      fWeights[2][o] = OutputActivationFnc(fWeights[2][o]);
   }

   return fWeights[2][0];
}

double ReadMLP::ActivationFnc(double x) const {
   // sigmoid
   return 1.0/(1.0+exp(-x));
}
double ReadMLP::OutputActivationFnc(double x) const {
   // identity
   return x;
}
   
// Clean up
inline void ReadMLP::Clear() 
{
   // clean up the arrays
   for (int lIdx = 0; lIdx < 3; lIdx++) {
      delete[] fWeights[lIdx];
   }
}
   inline double ReadMLP::GetMvaValue( const std::vector<double>& inputValues ) const
   {
      // classifier response value
      double retval = 0;

      // classifier response, sanity check first
      if (!IsStatusClean()) {
         std::cout << "Problem in class \"" << fClassName << "\": cannot return classifier response"
                   << " because status is dirty" << std::endl;
         retval = 0;
      }
      else {
         if (IsNormalised()) {
            // normalise variables
            std::vector<double> iV;
            iV.reserve(inputValues.size());
            int ivar = 0;
            for (std::vector<double>::const_iterator varIt = inputValues.begin();
                 varIt != inputValues.end(); varIt++, ivar++) {
               iV.push_back(NormVariable( *varIt, fVmin[ivar], fVmax[ivar] ));
            }
            retval = GetMvaValue__( iV );
         }
         else {
            retval = GetMvaValue__( inputValues );
         }
      }

      return retval;
   }