source: src/Potentials/PotentialTrainer.cpp@ 82e5fb

/*
 * Project: MoleCuilder
 * Description: creates and alters molecular systems
 * Copyright (C)  2014 Frederik Heber. All rights reserved.
 *
 *
 *   This file is part of MoleCuilder.
 *
 *    MoleCuilder is free software: you can redistribute it and/or modify
 *    it under the terms of the GNU General Public License as published by
 *    the Free Software Foundation, either version 2 of the License, or
 *    (at your option) any later version.
 *
 *    MoleCuilder is distributed in the hope that it will be useful,
 *    but WITHOUT ANY WARRANTY; without even the implied warranty of
 *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *    GNU General Public License for more details.
 *
 *    You should have received a copy of the GNU General Public License
 *    along with MoleCuilder.  If not, see <http://www.gnu.org/licenses/>.
 */

/*
 * PotentialTrainer.cpp
 *
 *  Created on: Sep 11, 2014
 *      Author: heber
 */

// include config.h
#ifdef HAVE_CONFIG_H
#include <config.h>
#endif

// needs to come before MemDebug due to placement new
#include <boost/archive/text_iarchive.hpp>

//#include "CodePatterns/MemDebug.hpp"

#include "PotentialTrainer.hpp"

#include <algorithm>
#include <boost/lambda/lambda.hpp>
#include <boost/filesystem.hpp>
#include <fstream>
#include <sstream>

#include "CodePatterns/Assert.hpp"
#include "CodePatterns/Log.hpp"

#include "Element/element.hpp"
#include "Fragmentation/Homology/HomologyContainer.hpp"
#include "Fragmentation/Homology/HomologyGraph.hpp"
#include "FunctionApproximation/Extractors.hpp"
#include "FunctionApproximation/FunctionApproximation.hpp"
#include "FunctionApproximation/FunctionModel.hpp"
#include "FunctionApproximation/TrainingData.hpp"
#include "FunctionApproximation/writeDistanceEnergyTable.hpp"
#include "Potentials/CompoundPotential.hpp"
#include "Potentials/RegistrySerializer.hpp"
#include "Potentials/SerializablePotential.hpp"

PotentialTrainer::PotentialTrainer()
{}

PotentialTrainer::~PotentialTrainer()
{}

bool PotentialTrainer::operator()(
    const HomologyContainer &_homologies,
    const HomologyGraph &_graph,
    const boost::filesystem::path &_trainingfile,
    const boost::filesystem::path &_errorfile,
    const unsigned int _maxiterations,
    const double _threshold,
    const unsigned int _best_of_howmany) const
{
  // fit potential
  CompoundPotential compound(_graph);
  FunctionModel &model = assert_cast<FunctionModel &>(compound);

  if (compound.begin() == compound.end()) {
    ELOG(1, "Could not find any suitable potentials for the compound potential.");
    return false;
  }

  /******************** TRAINING ********************/
  // fit potential
  FunctionModel::parameters_t bestparams(model.getParameterDimension(), 0.);
  {
    // Afterwards we go through all of this type and gather the distance and the energy value
    TrainingData data(model.getSpecificFilter());
    data(_homologies.getHomologousGraphs(_graph));

    // check data
    const TrainingData::FilteredInputVector_t &inputs = data.getTrainingInputs();
    for (TrainingData::FilteredInputVector_t::const_iterator iter = inputs.begin();
        iter != inputs.end(); ++iter)
      if (((*iter).empty()) || ((*iter).front().empty())) {
        ELOG(1, "At least one of the training inputs is empty! Correct fragment and potential charges selected?");
        return false;
      }
    const TrainingData::OutputVector_t &outputs = data.getTrainingOutputs();
    for (TrainingData::OutputVector_t::const_iterator iter = outputs.begin();
        iter != outputs.end(); ++iter)
      if ((*iter).empty()) {
        ELOG(1, "At least one of the training outputs is empty! Correct fragment and potential charges selected?");
        return false;
      }

    // print distances and energies if desired for debugging
    if (!data.getTrainingInputs().empty()) {
      // print which distance is which
      size_t counter=1;
      if (DoLog(3)) {
        const FunctionModel::arguments_t &inputs = data.getAllArguments()[0];
        for (FunctionModel::arguments_t::const_iterator iter = inputs.begin();
            iter != inputs.end(); ++iter) {
          const argument_t &arg = *iter;
          LOG(3, "DEBUG: distance " << counter++ << " is between (#"
              << arg.indices.first << "c" << arg.types.first << ","
              << arg.indices.second << "c" << arg.types.second << ").");
        }
      }

      // print table
      if (_trainingfile.string().empty()) {
        LOG(3, "DEBUG: I gathered the following training data:\n" <<
            _detail::writeDistanceEnergyTable(data.getDistanceEnergyTable()));
      } else {
        std::ofstream trainingstream(_trainingfile.string().c_str());
        if (trainingstream.good()) {
          LOG(3, "DEBUG: Writing training data to file " <<
              _trainingfile.string() << ".");
          trainingstream << _detail::writeDistanceEnergyTable(data.getDistanceEnergyTable());
        }
        trainingstream.close();
      }
    }

    if ((_threshold < 1.) && (_best_of_howmany))
      ELOG(2, "threshold parameter always overrules max_runs, both are specified.");
    // now perform the function approximation by optimizing the model function
    FunctionApproximation approximator(data, model, _threshold, _maxiterations);
    if (model.isBoxConstraint() && approximator.checkParameterDerivatives()) {
      double l2error = std::numeric_limits<double>::max();
      // seed with current time
      srand((unsigned)time(0));
      unsigned int runs=0;
      // threshold overrules max_runs
      const double threshold = _threshold;
      const unsigned int max_runs = (threshold >= 1.) ? _best_of_howmany : std::numeric_limits<unsigned int>::max();
      LOG(1, "INFO: Maximum runs is " << max_runs << " and threshold set to " << threshold << ".");
      do {
        // generate new random initial parameter values
        model.setParametersToRandomInitialValues(data);
        LOG(1, "INFO: Initial parameters of run " << runs << " are "
            << model.getParameters() << ".");
        approximator(FunctionApproximation::ParameterDerivative);
        LOG(1, "INFO: Final parameters of run " << runs << " are "
            << model.getParameters() << ".");
        const double new_l2error = data.getL2Error(model);
        if (new_l2error < l2error) {
          // store currently best parameters
          l2error = new_l2error;
          bestparams = model.getParameters();
          LOG(1, "STATUS: New fit from run " << runs
              << " has better error of " << l2error << ".");
        }
      } while (( ++runs < max_runs) && (l2error > threshold));
      // reset parameters from best fit
      model.setParameters(bestparams);
      LOG(1, "INFO: Best parameters with L2 error of "
          << l2error << " are " << model.getParameters() << ".");
    } else {
      return false;
    }

    // create a map of each fragment with error.
    HomologyContainer::range_t fragmentrange = _homologies.getHomologousGraphs(_graph);
    TrainingData::L2ErrorConfigurationIndexMap_t WorseFragmentMap =
        data.getWorstFragmentMap(model, fragmentrange);
    if (_errorfile.string().empty()) {
      LOG(0, "RESULT: WorstFragmentMap " << WorseFragmentMap << ".");
    } else {
      std::ofstream errorstream(_errorfile.string().c_str());
      if (errorstream.good()) {
        LOG(3, "DEBUG: Writing error data to file " <<
            _errorfile.string() << ".");
        errorstream << "step\terror" << std::endl;
        // resort into step as key
        typedef std::map< size_t, double > step_error_t;
        step_error_t step_error;
        for (TrainingData::L2ErrorConfigurationIndexMap_t::const_reverse_iterator iter = WorseFragmentMap.rbegin();
            iter != WorseFragmentMap.rend(); ++iter)
          step_error.insert( std::make_pair(iter->second, iter->first) );
        for (step_error_t::const_iterator iter = step_error.begin();
            iter != step_error.end(); ++iter)
          errorstream << iter->first << "\t" << iter->second << std::endl;
      }
      errorstream.close();
    }
  }

  return true;
}

HomologyGraph PotentialTrainer::getFirstGraphwithSpecifiedElements(
    const HomologyContainer &homologies,
    const SerializablePotential::ParticleTypes_t &types)
{
  ASSERT( !types.empty(),
      "getFirstGraphwithSpecifiedElements() - charges is empty?");

  // convert into count map
  Extractors::elementcounts_t counts_per_element =
      Extractors::_detail::getElementCounts(types);
  ASSERT( !counts_per_element.empty(),
      "getFirstGraphwithSpecifiedElements() - element counts are empty?");
  LOG(1, "DEBUG: counts_per_element is " << counts_per_element << ".");
  // we want to check each (unique) key only once
  HomologyContainer::const_key_iterator olditer = homologies.key_end();
  for (HomologyContainer::const_key_iterator iter =
      homologies.key_begin(); iter != homologies.key_end();
      iter = homologies.getNextKey(iter)) {
    // if it's the same as the old one, skip it
    if (olditer == iter)
      continue;
    else
      olditer = iter;
    // check whether we have the same set of atomic numbers
    const HomologyGraph::nodes_t &nodes = (*iter).getNodes();
    Extractors::elementcounts_t nodes_counts_per_element;
    for (HomologyGraph::nodes_t::const_iterator nodeiter = nodes.begin();
        nodeiter != nodes.end(); ++nodeiter) {
      const Extractors::element_t elem = nodeiter->first.getAtomicNumber();
      const std::pair<Extractors::elementcounts_t::iterator, bool> inserter =
          nodes_counts_per_element.insert( std::make_pair(elem, (Extractors::count_t)nodeiter->second ) );
      if (!inserter.second)
        inserter.first->second += (Extractors::count_t)nodeiter->second;
    }
    LOG(1, "DEBUG: Node (" << *iter << ")'s counts_per_element is " << nodes_counts_per_element << ".");
    if (counts_per_element == nodes_counts_per_element)
      return *iter;
  }
  return HomologyGraph();
}

SerializablePotential::ParticleTypes_t PotentialTrainer::getNumbersFromElements(
    const std::vector<const element *> &fragment)
{
  SerializablePotential::ParticleTypes_t fragmentnumbers;
  std::transform(fragment.begin(), fragment.end(), std::back_inserter(fragmentnumbers),
      boost::bind(&element::getAtomicNumber, _1));
  return fragmentnumbers;
}