////////////////////////////////////////////////////////////////////////////
//! author="Kieron J Messer"
//! lib=RbfUtil
//! date="4/7/99"
//! rcsid="$Id: doRbfSelect.cc,v 1.9 2000/12/12 13:27:00 ees2km Exp $"

#include "amma/Option.hh"
#include "amma/Filename.hh"
#include "amma/EntryPnt.hh"
#include "amma/Num/RbfUtil.hh"
#include "amma/DP/FileFormatIO.hh"

extern void NumRBFIncludeIOFunc();

int doRbfSelect(int argc, char **argv)
{  
  NumRBFIncludeIOFunc();

  OptionC     opt(argc,argv);
  FilenameC   TrainData = opt.String("i",NULL, "input NumVLDataSet data");
  FilenameC   TestData = opt.String("test",NULL, "input NumVLDataSet data");

  FilenameC   OutFile = opt.String("o", NULL, "output file");
  FilenameC   ResFile = opt.String("r", NULL, "results file");
  StringC     ModelSel = opt.String("sel", "", "choose which model selection algorithm to use");
  IntT        StartFrom = opt.Int("start", 1, "start at");
  IntT        StopAt = opt.Int("stop", 10, "stop at");
  IntT        Cycles = opt.Int("cyc", 10, "number of cycles of EM algorithm");
  IntT        Samps = opt.Int("n", 0, "number of samples to use in training step");
  IntT        ValSamps = opt.Int("nvals", 0, "number of samples to use in validation step");
  IntT        LocSamps = opt.Int("nwin", 0, "number of window placements");
  IntT        Seed = opt.Int("seed", 0, "number of window placements");
  RealT       PLo =  opt.Real("plo", 0.1, "starting learning rate for ema lag");
  RealT       PHi = opt.Real("phi", 0.25, "tolerance on predicted performance");
  BooleanT    Full = opt.Boolean("full", FALSE, "use full or diagonal covariance matiex");
  IntT        NCall = opt.Int("ncall", 100, "number of calls in vegas");
  IntT        Itmx = opt.Int("itmx", 10, "number of iterations in vegas");
  StringC     EmAlg = opt.String("emalg", "Normal", "Em algorithm [Normal||Ppca]");
  IntT        Ppca = opt.Int("ppca", 0, "number of dimensions to ue for Ppca");
  StringC     ValAlg = opt.String("valg", "SampleSame", "Validation algorithm [Same||Diff||Joint]");
  StringC     ConfAlg = opt.String("calg", "JointConfidence", "Limits algorithm [LineTest]");
  
  opt.Check();

  

  //: set random seed
  randomIdum=Seed;
  
  //: load in data
  NumVLDataSetC data(TrainData);
  data.Shuffle();
  cout.precision(6);

  // NumVLDataSetC test(TestData);
  
  //: set rbf options
  RbfOptionC rbfOpt;
  rbfOpt.StartFrom(StartFrom);
  rbfOpt.StopAt(StopAt);
  rbfOpt.Cycles(Cycles);
  rbfOpt.TrainSamples(Samps);
  rbfOpt.ValSamples(ValSamps);
  rbfOpt.LocSamples(LocSamps);
  rbfOpt.ProbLo(PLo);
  rbfOpt.ProbHi(PHi);
  rbfOpt.Full(Full);
  rbfOpt.NCall(NCall);
  rbfOpt.Itmx(Itmx);
  rbfOpt.EmAlgorithm(EmAlg);  
  rbfOpt.PpcaDim(Ppca);  
  rbfOpt.ValAlgorithm(ValAlg);  
  rbfOpt.ConfAlgorithm(ConfAlg);
  rbfOpt.ModelSelAlgorithm(ModelSel);
  cout << rbfOpt << endl;

  //: compute the model
  NormalRBFuncC rbf = RbfUtil::ModelSelect(data.InputSet(), rbfOpt);
  //DListC<NormalRBFuncC> rbfs = RbfUtil::GreedyEM(data.InputSet(), 10);
  //NormalRBFuncC rbf = RbfUtil::ModelSelect(rbfs, test.InputSet(), rbfOpt);
  
  //cout << "best rbf: " << rbf.Size() << endl;
  //IntT i=1;
  //cout << endl;
  //for(DLIterC<NormalRBFuncC>It(rbfs);It.IsElm();It.Next()) {
  //  cout << i << ": " << It.Data().NegLogLikelihood(test.InputSet())/(RealT)test.Size() << endl;
  //}

  //ofstream ofs(OutFile, ios::app);
  //ofs << rbf.Size() << " " << rbf.NegLogLikelihood(test.InputSet())/(RealT)test.Size() << endl;
  //ofs.close();
  //: output model
  if(opt.IsOnCommandLine("o")) StdIO::Save(OutFile, rbf);
  else cout << rbf << endl;
 
  return 0;
}

//: This puts a wrapper around the main program that catches
//: exceptions and turns them into readable error messages.

AMMA_ENTRY_POINT(doRbfSelect);