source: src/Actions/PotentialAction/FitPotentialAction.cpp@ 019b96

/*
 * Project: MoleCuilder
 * Description: creates and alters molecular systems
 * Copyright (C)  2013 University of Bonn. All rights reserved.
 * Copyright (C)  2013 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/>.
 */

/*
 * FitPotentialAction.cpp
 *
 *  Created on: Apr 09, 2013
 *      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 <algorithm>
#include <boost/bind.hpp>
#include <boost/filesystem.hpp>
#include <boost/foreach.hpp>
#include <map>
#include <string>
#include <sstream>

#include "Actions/PotentialAction/FitPotentialAction.hpp"

#include "CodePatterns/Log.hpp"

#include "Element/element.hpp"
#include "Fragmentation/Homology/HomologyContainer.hpp"
#include "Fragmentation/Homology/HomologyGraph.hpp"
#include "Fragmentation/Summation/SetValues/Fragment.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/Exceptions.hpp"
#include "Potentials/PotentialDeserializer.hpp"
#include "Potentials/PotentialFactory.hpp"
#include "Potentials/PotentialRegistry.hpp"
#include "Potentials/PotentialSerializer.hpp"
#include "Potentials/SerializablePotential.hpp"

using namespace MoleCuilder;

// and construct the stuff
#include "FitPotentialAction.def"
#include "Action_impl_pre.hpp"
/** =========== define the function ====================== */

HomologyGraph getFirstGraphwithSpecifiedElements(
    const HomologyContainer &homologies,
    const SerializablePotential::ParticleTypes_t &types)
{
  ASSERT( !types.empty(),
      "getFirstGraphwithSpecifiedElements() - charges is empty?");
  // create charges
  Fragment::charges_t charges;
  charges.resize(types.size());
  std::transform(types.begin(), types.end(),
      charges.begin(), boost::lambda::_1);
  // convert into count map
  Extractors::elementcounts_t counts_per_charge =
      Extractors::_detail::getElementCounts(charges);
  ASSERT( !counts_per_charge.empty(),
      "getFirstGraphwithSpecifiedElements() - charge counts are empty?");
  LOG(2, "DEBUG: counts_per_charge is " << counts_per_charge << ".");
  // we want to check each (unique) key only once
  for (HomologyContainer::const_key_iterator iter = homologies.key_begin();
      iter != homologies.key_end(); iter = homologies.getNextKey(iter)) {
    // check if every element has the right number of counts
    Extractors::elementcounts_t::const_iterator countiter = counts_per_charge.begin();
    for (; countiter != counts_per_charge.end(); ++countiter)
      if (!(*iter).hasTimesAtomicNumber(
          static_cast<size_t>(countiter->first),
          static_cast<size_t>(countiter->second))
          )
        break;
    if( countiter == counts_per_charge.end())
      return *iter;
  }
  return HomologyGraph();
}

SerializablePotential::ParticleTypes_t 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;
}


ActionState::ptr PotentialFitPotentialAction::performCall() {
  // fragment specifies the homology fragment to use
  SerializablePotential::ParticleTypes_t fragmentnumbers =
      getNumbersFromElements(params.fragment.get());

  // either charges and a potential is specified or a file
  if (boost::filesystem::exists(params.potential_file.get())) {
    std::ifstream returnstream(params.potential_file.get().string().c_str());
    if (returnstream.good()) {
      try {
        PotentialDeserializer deserialize(returnstream);
        deserialize();
      } catch (SerializablePotentialMissingValueException &e) {
        if (const std::string *key = boost::get_error_info<SerializablePotentialKey>(e))
          STATUS("Missing value when parsing information for potential "+*key+".");
        else
          STATUS("Missing value parsing information for potential with unknown key.");
        return Action::failure;
      } catch (SerializablePotentialIllegalKeyException &e) {
        if (const std::string *key = boost::get_error_info<SerializablePotentialKey>(e))
          STATUS("Illegal key parsing information for potential "+*key+".");
        else
          STATUS("Illegal key parsing information for potential with unknown key.");
        return Action::failure;
      }
    } else {
      STATUS("Failed to parse from "+params.potential_file.get().string()+".");
      return Action::failure;
    }
    returnstream.close();

    LOG(0, "STATUS: I'm training now a set of potentials parsed from "
        << params.potential_file.get().string() << " on a fragment "
        << fragmentnumbers << " on data from World's homologies.");

  } else {
    if (params.charges.get().empty()) {
      STATUS("Neither charges nor potential file given!");
      return Action::failure;
    } else {
      // charges specify the potential type
      SerializablePotential::ParticleTypes_t chargenumbers =
          getNumbersFromElements(params.charges.get());

      LOG(0, "STATUS: I'm training now a " << params.potentialtype.get()
          << " potential on charges " << chargenumbers << " on data from World's homologies.");

      // register desired potential and an additional constant one
      {
        EmpiricalPotential *potential =
            PotentialFactory::getInstance().createInstance(
                params.potentialtype.get(),
                chargenumbers);
        // check whether such a potential already exists
        const std::string potential_name = potential->getName();
        if (PotentialRegistry::getInstance().isPresentByName(potential_name)) {
          delete potential;
          potential = PotentialRegistry::getInstance().getByName(potential_name);
        } else
          PotentialRegistry::getInstance().registerInstance(potential);
      }
      {
        EmpiricalPotential *constant =
            PotentialFactory::getInstance().createInstance(
                std::string("constant"),
                SerializablePotential::ParticleTypes_t());
        // check whether such a potential already exists
        const std::string constant_name = constant->getName();
        if (PotentialRegistry::getInstance().isPresentByName(constant_name)) {
          delete constant;
          constant = PotentialRegistry::getInstance().getByName(constant_name);
        } else
          PotentialRegistry::getInstance().registerInstance(constant);
      }
    }
  }

  // parse homologies into container
  HomologyContainer &homologies = World::getInstance().getHomologies();

  // first we try to look into the HomologyContainer
  LOG(1, "INFO: Listing all present homologies ...");
  for (HomologyContainer::container_t::const_iterator iter =
      homologies.begin(); iter != homologies.end(); ++iter) {
    LOG(1, "INFO: graph " << iter->first
        << " has Fragment " << iter->second.fragment
        << ", associated energy " << iter->second.energy
        << ", and sampled grid integral " << iter->second.charge_distribution.integral()
        << ".");
  }

  // then we ought to pick the right HomologyGraph ...
  const HomologyGraph graph = getFirstGraphwithSpecifiedElements(homologies,fragmentnumbers);
  if (graph != HomologyGraph()) {
    LOG(1, "First representative graph containing fragment "
        << fragmentnumbers << " is " << graph << ".");
  } else {
    STATUS("Specific fragment "+toString(fragmentnumbers)+" not found in homologies!");
    return Action::failure;
  }

  // fit potential
  FunctionModel *model = new CompoundPotential(graph);
  ASSERT( model != NULL,
      "PotentialFitPotentialAction::performCall() - model is NULL.");

  /******************** 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));

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

    if ((params.threshold.get() < 1) && (params.best_of_howmany.isSet()))
      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);
    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 = params.threshold.get();
      const unsigned int max_runs = (threshold >= 1.) ?
          (params.best_of_howmany.isSet() ? params.best_of_howmany.get() : 1) : 0;
      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 {
      STATUS("No required parameter derivatives for a box constraint minimization known.");
      return Action::failure;
    }

    // create a map of each fragment with error.
    HomologyContainer::range_t fragmentrange = homologies.getHomologousGraphs(graph);
    TrainingData::L2ErrorConfigurationIndexMap_t WorseFragmentMap =
        data.getWorstFragmentMap(*model, fragmentrange);
    LOG(0, "RESULT: WorstFragmentMap " << WorseFragmentMap << ".");

    // print fitted potentials
    std::stringstream potentials;
    PotentialSerializer serialize(potentials);
    serialize();
    LOG(1, "STATUS: Resulting parameters are " << std::endl << potentials.str());
    std::ofstream returnstream(params.potential_file.get().string().c_str());
    if (returnstream.good()) {
      returnstream << potentials.str();
    }
  }
  delete model;

  return Action::success;
}

ActionState::ptr PotentialFitPotentialAction::performUndo(ActionState::ptr _state) {
  return Action::success;
}

ActionState::ptr PotentialFitPotentialAction::performRedo(ActionState::ptr _state){
  return Action::success;
}

bool PotentialFitPotentialAction::canUndo() {
  return false;
}

bool PotentialFitPotentialAction::shouldUndo() {
  return false;
}
/** =========== end of function ====================== */