predict.cpp

#include "libgrt_util.h"
#include "cmdline.h"

int main(int argc, char *argv[]) 
{
  static bool is_running = true;
  cmdline::parser c;

  c.add<int>   ("verbose",    'v', "verbosity level: 0-4", false, 0);
  c.add        ("help",       'h', "print this message");
  c.add        ("likelihood", 'l', "print label_prediction likelihood instead of label and prediction");
  c.add        ("null",       'n', "draw labels randomly from the set of labels (for testing the chain)");
  c.footer     ("[classifier-model-file] [filename]...");

  /* parse the classifier-common arguments */
  if (!c.parse(argc,argv,true) || c.exist("help")) {
    cerr << c.usage() << "\n" << (c.exist("help") ? c.error() : "") << "\n" ;
    return -1;
  }

  set_verbosity(c.get<int>("verbose"));

  /* wait until first data has arrived before trying to read the
   * classifier, to catch cases where the training has not yet been
   * completed, and he classifier has not yet been written to disk */
  istream &in = grt_fileinput(c,1);
  in.peek(); // block until data there

  // check if the model input file exists and is size>0
  if (c.rest().size()) {
    ifstream test(c.rest()[0], ifstream::binary | ifstream::ate);
    if (test.tellg() < 1) {
      cerr << "unable to open model: " << c.rest()[0] << endl;
      return -1;
    }
  }

  /* load a classification model */
  ifstream fin; fin.open(c.rest().size() ? c.rest()[0] : "");
  istream &model = c.rest().size() ? fin : cin;

  /* read and predict on input */
  Classifier *classifier = loadClassifierFromFile(model);

  if (classifier == NULL && c.rest().size() > 0)
    // retry 5 times with 1 sec wait in between
    for (int i=0; i<5 && classifier==NULL; i++, usleep(1000*1000)) {
      ifstream fin(c.rest()[0]);
      classifier = loadClassifierFromFile(fin);
    }

  if (classifier == NULL) {
    cerr << "unable to load classification model giving up" << endl;
    return -1;
  }

  /* prepare input */
  string data_type = classifier->getTimeseriesCompatible() ? "timeseries" : "classification";
  CsvIOSample io(data_type);

  while( in >> io && is_running ) {
    UINT prediction = 0, label = 0;
    string s_prediction, s_label;
    bool result = false;

    switch(io.type) {
    case TIMESERIES:
      result = classifier->predict(io.t_data.getData());
      label = io.t_data.getClassLabel();
      s_label = classifier->getClassNameForLabel(label);
      prediction = classifier->getPredictedClassLabel();
      s_prediction = classifier->getClassNameForLabel(prediction);
      break;
    case CLASSIFICATION:
      result = classifier->predict(io.c_data.getSample());
      label = io.c_data.getClassLabel();
      s_label = classifier->getClassNameForLabel(label);
      prediction   = classifier->getPredictedClassLabel();
      s_prediction = classifier->getClassNameForLabel(prediction);
      break;
    default:
      cerr << "unknown input type" << endl;
      return -1;
    }

    if (!result) {
      cerr << "prediction failed (wrong input type?)" << endl;
      return -1;
    }

    if (label == 0) s_label = "NULL";
    if (prediction == 0) s_prediction = "NULL";

    /*
     * replace the prediction with a random choice from the labelset
     */
    if (c.exist("null")) {
      UINT index = (UINT) round(drand48() * (classifier->getNumClasses()-1));
      s_prediction = index == 0 ? "NULL" : classifier->getClassNameForLabel(index);
    }

    if (c.exist("likelihood"))
      cout << s_label << "\t" << s_prediction << "\t" << classifier->getMaximumLikelihood() << endl;
    else
      cout << s_label << "\t" << s_prediction << endl;
  }

  cout << endl;
  return 0;
}