TBNN.cpp

/****************************************************************************
* Copyright (c) 2023, CEA
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
* 1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
* 3. Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
* IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
* OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
*****************************************************************************/
#include <PrePostNN.h>
#include <TBNN.h>
#include <cmath>
#include <cstddef>
#include <iostream>
#include <keras_model.h>
#include <string>
#include <vector>
 
int sgn(double v)
{
  return (v > 0) - (v < 0);
}
 
TBNN::TBNN(string keras_model_file,string preproc_file)
{
  _ppNN = new PrePostNN(preproc_file);
  _model.LoadModel(keras_model_file);
}
 
TBNN::~TBNN()
{
  delete(_ppNN);
}
 
std::vector<double> TBNN::predict(std::vector<double> lambda, std::vector<std::vector<double>> T)
{
  process_lambda(lambda);
  process_T(T);
  applyNN();
  process_b();
  return(_b);
}
 
void TBNN::process_lambda(std::vector<double> lambda)
{
  std::vector<double> lc;                 // lambda centre
  std::vector<double> lcr;                // lambda centre reduit
  size_t nbl = lambda.size();  // nombre d'invariants lambda
 
  lc.resize(nbl);
  lcr.resize(nbl);
  _plambda.resize(nbl);
 
  switch(_ppNN->get_ppl())
    {
    case LNORM:
      // centrage des lambda
      if( _ppNN->get_lmean().size() == nbl )
        for(unsigned int i=0; i<nbl; i++)
          lc[i] = lambda[i] - _ppNN->get_lmean()[i];
      else
        for(unsigned int i=0; i<nbl; i++)
          lc[i] = lambda[i];
      // reduction des lambda
      for(unsigned int i=0; i<nbl; i++)
        _plambda[i] = lc[i] / _ppNN->get_lmax()[i];
      break;
    case LU:
      // puissance alpha
      if( _ppNN->get_alpha().size() == nbl )
        for(unsigned int i=0; i<nbl; i++)
          lc[i] = sgn(lambda[i]) * pow(std::fabs(lambda[i]),_ppNN->get_alpha()[i]);
      else
        for(unsigned int i=0; i<nbl; i++)
          lc[i] = lambda[i];
      // centrage des lambda
      if( _ppNN->get_lmean().size() == nbl )
        for(unsigned int i=0; i<nbl; i++)
          lc[i] = lc[i] - _ppNN->get_lmean()[i];
      // reduction des lambda
      for(unsigned int i=0; i<nbl; i++)
        lcr[i] = lc[i] / _ppNN->get_lmax()[i];
      // multiplication par transpose(au)
      for(unsigned int i=0; i<nbl; i++)
        {
          _plambda[i] = 0.;
          for(unsigned int j=0; j<nbl; j++)
            _plambda[i] += _ppNN->get_lambda_au()[j][i] * lcr[j];
        }
      break;
    case LUS:
      // puissance alpha
      if( _ppNN->get_alpha().size() == nbl )
        for(unsigned int i=0; i<nbl; i++)
          lc[i] = sgn(lambda[i]) * pow(std::fabs(lambda[i]),_ppNN->get_alpha()[i]);
      else
        for(unsigned int i=0; i<nbl; i++)
          lc[i] = lambda[i];
      // centrage des lambda
      if( _ppNN->get_lmean().size() == nbl )
        for(unsigned int i=0; i<nbl; i++)
          lc[i] = lc[i] - _ppNN->get_lmean()[i];
      // reduction des lambda
      for(unsigned int i=0; i<nbl; i++)
        lcr[i] = lc[i] / _ppNN->get_lmax()[i];
      // multiplication par transpose(as)
      for(unsigned int i=0; i<nbl; i++)
        {
          _plambda[i] = 0.;
          for(unsigned int j=0; j<nbl; j++)
            _plambda[i] += _ppNN->get_lambda_as()[j][i] * lcr[j];
        }
      break;
    default:
      // on lance une exception
      cerr << "Mauvaise methode de pre traitement des lambda" << endl;
      break;
    }
}
 
void TBNN::process_T(std::vector<std::vector<double>> T)
{
  size_t nbt = T.size();    // nombre de tenseurs T
  size_t nbb = T[0].size(); // taille de chacun des tenseurs T
 
  _pT.resize(nbt);
  for(unsigned i=0; i<nbt; i++)
    _pT[i].resize(nbb);
 
  // le premier tenseur T0 reste inchange
  for(unsigned int j=0; j<nbb; j++)
    _pT[0][j] = T[0][j];
 
  // pre process de T
  switch(_ppNN->get_ppt())
    {
    case TF:
      // on calcule la norme de Frobenius de chaque tenseur
      for(unsigned int i=1; i<nbt; i++)
        {
          double normf = 0.;
          for(unsigned int j=0; j<nbb; j++)
            normf += T[i][j] * T[i][j];
          normf = sqrt(normf);
          for(unsigned int j=0; j<nbb; j++)
            _pT[i][j] = T[i][j] / (normf + _ppNN->get_t_thresh());
        }
      break;
    case TR:
      // on divise le tenseur Ti par la norme globale
      for(unsigned int i=1; i<nbt; i++)
        for(unsigned int j=0; j<nbb; j++)
          _pT[i][j] = T[i][j] / _ppNN->get_tfn()[i-1];
      break;
    default:
      // on lance une exception
      cerr << "Mauvaise methode de pre traitement des tenseurs T" << endl;
      break;
    }
}
 
void TBNN::process_b()
{
  std::vector<int> iT = _ppNN->get_iT();
  size_t nbt = iT.size();
  size_t nbb = _pT[0].size();
 
  // calcul de _pb a partir de _g et de _pT
  _pb.resize(nbb);
  for(unsigned int i=0; i<nbb; i++)
    {
      _pb[i] = 0;
      for(unsigned int j=0; j<nbt; j++)
        _pb[i] += _g[j] * _pT[iT[j]][i];
    }
 
  // post process de b
  _b.resize(nbb);
  for(unsigned int i=0; i<nbb; i++)
    _b[i] = _ppNN->get_bsigma() * _pb[i];
}
 
void TBNN::applyNN()
{
  std::vector<int> il = _ppNN->get_ilambda();
  std::vector<int> iT = _ppNN->get_iT();
  Tensor in((int)il.size());
  Tensor out;
  size_t nbt = iT.size();
 
  _g.resize(nbt);
 
  // l'entree du reseau est determinee par les indices des lambdas definis dans le vecteur il
  for(unsigned int i=0; i<il.size(); i++)
    in.data_[i] = (float)_plambda[il[i]];
 
  // on fait la prediction a l'aide du reseau de neurones
  _model.Apply(&in,&out);
 
  // on stocke les sorties dans _g
  for(unsigned int i=0; i<nbt; i++) _g[i] = out(i);
 
}