/** 
* @file compressObjectBankFeatures.cpp
* @brief convert ObjectBank features to a sparse histogram representation
* @author Erik Rodner
* @date 01/23/2012

*/
#include <algorithm>
#include "core/basics/Config.h"
#include "vislearning/cbaselib/MultiDataset.h"
#include "vislearning/cbaselib/MutualInformation.h"
#include "vislearning/baselib/Globals.h"

using namespace std;
using namespace NICE;
using namespace OBJREC;

const bool use_standard = false; // experimental setting

Vector transformFeature ( const vector<double> & src )
{
  Vector dst;

  if ( use_standard ) {
    dst = Vector(src);
  } else {
    if ( src.size() != 44604 )
      fthrow(Exception, "This is not a ObjectBank feature! The size is: " << src.size());

    dst.resize ( 177 );
    dst.set(0.0);
    
    // This was a bad idea: taking the maximum
    /*
    for ( uint i = 0 ; i < 177; i++ )
      dst[i] = *max_element( src.begin() + i*252, src.begin() + (i+1)*252 );
    // even a worse idea: summation
    for ( uint i = 0 ; i < 177; i++ )
      for ( uint j = 0 ; j < 252 ; j++ )
        dst[i] += src[j + i*252];
    */
  }

  return dst;
}

void readPlainData ( const Config & conf, const LabeledSet & ls, LabeledSetVector & X, string extension = ".txt" )
{
  string cacheroot = conf.gS("cache", "root");

  X.clear();

  LOOP_ALL_S ( ls ) 
  {
    EACH_S(classno, imgfn);
    Globals::setCurrentImgFN ( imgfn ); 
    string cachefn = Globals::getCacheFilename ( cacheroot, Globals::SORT_CATEGORIES ) + extension;
    
    cerr << "fn: " << imgfn << " cachefn: " << cachefn << endl;

    vector<double> x;
    ifstream ifs ( cachefn.c_str(), ios::in );
    if ( ! ifs.good() )
      fthrow(Exception, "File not found: " << cachefn );
    while ( !ifs.eof() ) 
    {
      double val = 0.0;
      if ( ifs >> val )
        x.push_back(val);
    }
    ifs.close();

    X.add ( classno, transformFeature( x ) );
  }
}

void saveFeatures ( const Config & conf, const map<double, int> & features, const Vector & thresholds, 
                    const LabeledSet & ls, const string & srcExtension, const string & dstExtension )
{
  string cacheroot = conf.gS("cache", "root");

  LOOP_ALL_S ( ls ) 
  {
    EACH_S(classno, imgfn);
    Globals::setCurrentImgFN ( imgfn ); 
    string cachefn = Globals::getCacheFilename ( cacheroot, Globals::SORT_CATEGORIES ) + srcExtension;
    cerr << "processing " << cachefn << endl;

    vector<double> x;
    ifstream ifs ( cachefn.c_str(), ios::in );
    if ( ! ifs.good() )
      fthrow(Exception, "File not found: " << cachefn );
    while ( !ifs.eof() ) {
      double val = 0.0;
      if ( ifs >> val )
        x.push_back(val);
    }
    ifs.close();
   
    Vector xt = transformFeature(x);

    Vector xnew ( features.size() );
    int index = 0;
    for ( map<double, int>::const_iterator j = features.begin(); j != features.end(); j++, index++ )
    {
      int srcIndex = j->second;
      if ( srcIndex >= xt.size() )
        fthrow(Exception, "Bad bug in saveFeatures(...)" );
      xnew[index] = (xt[srcIndex] > thresholds[srcIndex]) ? 1.0 : 0.0;
    }

    // If we do not normalize our features, we pretty much get into 
    // trouble with the minimum kernel...because the vector with only values of "1" is very
    // much similar to every vector
    xnew.normalizeL1();

    string dst_cachefn = Globals::getCacheFilename ( cacheroot, Globals::SORT_CATEGORIES ) + dstExtension;
    ofstream ofs ( dst_cachefn.c_str(), ios::out );
    if ( ! ofs.good() )
      fthrow(Exception, "Unable to write to " << dst_cachefn );
    ofs << xnew << endl;
    ofs.close ();
  }
    
}

/** 
    
    convert ObjectBank features to a sparse histogram representation 
    
*/
int main (int argc, char **argv)
{   
  std::set_terminate(__gnu_cxx::__verbose_terminate_handler);

  Config conf ( argc, argv );
    
  MultiDataset md ( &conf );
  Vector y;
  const LabeledSet *train = md["train"];
  LabeledSetVector trainData;

  readPlainData ( conf, *train, trainData, ".jpg.feat" );
  
  // compute optimal thresholds for thresholding
  MutualInformation mi ( true /*verbose*/ );
  Vector thresholds;
  Vector mis;
  mi.computeThresholdsOverall ( trainData, thresholds, mis );

  cerr << mis << endl;
    
  int numFeatures = conf.gI("main", "d", mis.size() );
  cerr << "Retaining " << numFeatures << " features ..." << endl;
   
  map<double, int> features;
  for ( uint i = 0 ; i < mis.size(); i++ )
    features.insert ( pair<double, int> ( - mis[i], i ) );
  // all features should be now sorted with features[0] being the most informative one

  // remove boring features
  map<double, int>::iterator j = features.begin();
  advance( j, numFeatures );
  features.erase ( j, features.end() );
   
  const LabeledSet *test = md["test"];
  string dstExtention = conf.gS("main", "dstext", ".txt");
  saveFeatures ( conf, features, thresholds, *train, ".jpg.feat", dstExtention );
  saveFeatures ( conf, features, thresholds, *test, ".jpg.feat", dstExtention );
    
  return 0;
}