Learner SGDE OnOff

This example shows how to perform offline/online-classification using sparse grid density estimation and matrix decomposition methods.

It creates an instance of LearnerSGDEOnOff and runs the function train() where the main functionality is implemented.

Currently, only binary classification with class labels -1 and 1 is possible.

The example provides the option to execute several runs over differently ordered data and perform a 5-fold cross-validation within each run. Therefore, already randomly ordered and partitioned data is required. Average results from several runs might be more reliable in an online-learning scenario, because the ordering of the data points seen by the learner can affect the result.

int main(int argc, char *argv[]) {
#ifdef USE_MPI
#ifdef USE_GSL
 
  omp_set_num_threads(1);
 
  std::cout << "LearnerSGDEOnOffParallelTest" << std::endl;
 
  if (argc != 5) {
    std::cout << "Usage:" << std::endl
              << "learnerSGDEOnOffParallelTest <trainDataFile> "
              << "<testDataFile> <batchSize> <refPeriod>" << std::endl;
    return -1;
  }

Specify the number of runs to perform. If only one specific example should be executed, set totalSets=1.

  //  size_t totalSets = 1;
  //  size_t totalFolds = 1;  // set to 5 to perform 5-fold cv
  //  double avgError = 0.0;
  //  double avgErrorFolds = 0.0;
  //  for (size_t numSets = 0; numSets < totalSets; numSets++) {
  //    /**
  //     * A vector to compute average classification error throughout
  //     * the learning process. The length of the vector determines
  //     * the total number of error observations.
  //     */
  //    sgpp::base::DataVector avgErrorsFolds(51, 0.0);
  //
  //    for (size_t numFolds = 0; numFolds < totalFolds; numFolds++) {

Get the training, test and validation data

  sgpp::datadriven::Dataset trainDataset = loadDataset(argv[1]);
  sgpp::datadriven::Dataset testDataset = loadDataset(argv[2]);
 
  // if fixed validation data should be used (required for convergence
  // monitor):
  /*filename = "";  // specify file containing validation data here
  // load validation samples
  std::cout << "# loading file: " << filename << std::endl;
  sgpp::datadriven::Dataset validationDataset =
      sgpp::datadriven::ARFFTools::readARFF(filename); */

Specify the number of classes and the corresponding class labels.

  size_t classNum = 2;
  sgpp::base::DataVector classLabels(classNum);
  classLabels[0] = -1;
  classLabels[1] = 1;

The grid configuration.

  std::cout << "# create grid config" << std::endl;
  sgpp::base::RegularGridConfiguration gridConfig;
  gridConfig.dim_ = trainDataset.getDimension();
  gridConfig.level_ = 3;
  gridConfig.type_ = sgpp::base::GridType::Linear;
  // gridConfig.type_ = sgpp::base::GridType::ModLinear;

Configure regularization.

  std::cout << "# create regularization config" << std::endl;
  sgpp::datadriven::RegularizationConfiguration regularizationConfig{};
  regularizationConfig.type_ = sgpp::datadriven::RegularizationType::Identity;
  // initial regularization parameter lambda
  regularizationConfig.lambda_ = 0.01;

Select the desired decomposition type for the offline step. Note: Refinement/Coarsening only possible for Cholesky decomposition.

  sgpp::datadriven::MatrixDecompositionType dt;
  std::string decompType;
  // choose "LU decomposition"
  // dt = MatrixDecompositionType::DBMatDecompLU;
  // decompType = "LU decomposition";
  // choose"Eigen decomposition"
  // dt = MatrixDecompositionType::DBMatDecompEigen;
  // decompType = "Eigen decomposition";
  // choose "Cholesky decomposition"
  //      dt = sgpp::datadriven::MatrixDecompositionType::Chol;
  //      decompType = "Cholesky decomposition";
  //      dt = sgpp::datadriven::MatrixDecompositionType::IChol;
  //      decompType = "Incomplete Cholesky decomposition";
  dt = sgpp::datadriven::MatrixDecompositionType::DenseIchol;
  decompType = "Incomplete Cholesky decomposition on Dense Matrix";
  std::cout << "Decomposition type: " << decompType << std::endl;
  sgpp::datadriven::DensityEstimationConfiguration densityEstimationConfig;
  densityEstimationConfig.decomposition_ = dt;
  densityEstimationConfig.iCholSweepsDecompose_ = 2;
  densityEstimationConfig.iCholSweepsRefine_ = 2;

Configure adaptive refinement (if Cholesky is chosen). As refinement monitor the periodic monitor or the convergence monitor can be chosen. Possible refinement indicators are surplus refinement, data-based refinement, zero-crossings-based refinement.

  std::cout << "# create adaptive refinement configuration" << std::endl;
  std::string refMonitor;
  // select periodic monitor - perform refinements in fixed intervals
  refMonitor = "periodic";
 
  size_t refPeriod = 0;  // the refinement interval
  parseInputValue(argv[4], refPeriod);
  // select convergence monitor - perform refinements if algorithm has
  // converged
  // (convergence measured with respect to changes of the classification
  // accuracy)
  // refMonitor = "convergence";
  // the convergence threshold
  double accDeclineThreshold = 0.001;
  // number of accuracy measurements which
  // are considered for convergence check
  size_t accDeclineBufferSize = 140;
  // minimum number of iterations before next refinement
  // is allowed to be performed
  size_t minRefInterval = 10;
  std::cout << "Refinement monitor: " << refMonitor << std::endl;
  std::string refType;
  // select surplus refinement
  // refType = "surplus";
  // select data-based refinement
  // refType = "data";
  // select zero-crossings-based refinement
  refType = "zero";
  std::cout << "Refinement type: " << refType << std::endl;
  sgpp::base::AdaptivityConfiguration adaptivityConfig;

Specify number of refinement steps and the max number of grid points to refine each step.

  adaptivityConfig.numRefinements_ = 2;
  adaptivityConfig.numRefinementPoints_ = 7;
  adaptivityConfig.refinementThreshold_ = 0.0;  // only required for surplus refinement
 
  // initial weighting factor
  double beta = 0.0;
  // specify if prior should be used to predict class labels
  bool usePrior = false;
 
  // specify batch size
  // (set to 1 for processing only a single data point each iteration)
  size_t batchSize = 0;
  parseInputValue(argv[3], batchSize);
 
  // Create the MPI Task Scheduling using the round robin algorithm
  sgpp::datadriven::RoundRobinScheduler scheduler(batchSize);

Create the learner.

  std::cout << "# create learner" << std::endl;
  sgpp::datadriven::LearnerSGDEOnOffParallel learner(
      gridConfig, adaptivityConfig, regularizationConfig, densityEstimationConfig, trainDataset,
      testDataset, nullptr, classLabels, classNum, usePrior, beta, scheduler);
 
  // specify max number of passes over traininig data set
  size_t maxDataPasses = 1;

Learn the data.

  MPI_Barrier(MPI_COMM_WORLD);
 
  std::cout << "# start to train the learner" << std::endl;
  sgpp::base::SGppStopwatch stopwatch;
 
  //    CALLGRIND_START_INSTRUMENTATION;
 
  stopwatch.start();
  learner.trainParallel(batchSize, maxDataPasses, refType, refMonitor, refPeriod,
                        accDeclineThreshold, accDeclineBufferSize, minRefInterval);
  double deltaTime = stopwatch.stop();
 
  //    CALLGRIND_STOP_INSTRUMENTATION;
 
  MPI_Barrier(MPI_COMM_WORLD);

Accuracy on test data.

  double acc = learner.getAccuracy();
  if (sgpp::datadriven::MPIMethods::isMaster()) {
    std::cout << "# accuracy (test data): " << acc << std::endl;
    std::cout << "# delta time training: " << deltaTime << std::endl;
 
  } else {
    std::cout << "# accuracy (client, test data): " << acc << std::endl;
    std::cout << "# delta time training (client): " << deltaTime << std::endl;
  }
// store results (classified data, grids, density functions)
// learner.storeResults();
 
//        sgpp::base::DataVector tmp;
//        avgErrorFolds += 1.0 - learner.getAccuracy();
//        learner.getAvgErrors(tmp);
//        avgErrorsFolds.add(tmp);
//    }
//    avgErrorFolds = avgErrorFolds / static_cast<double>(totalFolds);
//    if ((totalSets > 1) && (totalFolds > 1)) {
//      /**
//       * Average accuracy on test data reagarding 5-fold cv.
//       */
//      std::cout << "Average accuracy on test data (set " + std::to_string(numSets + 1) + "): "
//                << (1.0 - avgErrorFolds) << std::endl;
//    }
//    avgError += avgErrorFolds;
//    avgErrorFolds = 0.0;
//    avgErrorsFolds.mult(1.0 / static_cast<double>(totalFolds));
 
// write error evaluation to csv-file
/*std::ofstream output;
output.open("SGDEOnOff_avg_classification_error_"+std::to_string(numSets+1)+".csv");
if (output.fail()) {
  std::cout << "failed to create csv file!" << std::endl;
}
else {
  for (size_t i = 0; i < avgErrorsFolds.getSize(); i++) {
    output << avgErrorsFolds.get(i) << ";" << std::endl;
  }
  output.close();
}*/
//  }
#else
  std::cout << "GSL not enabled at compile time" << std::endl;
#endif  // USE_GSL
#endif  /* USE_MPI */

}
 
#ifdef USE_MPI
sgpp::datadriven::Dataset loadDataset(const std::string &filename) {
  // load test samples
  std::cout << "# loading file: " << filename << std::endl;
  sgpp::datadriven::Dataset dataset = sgpp::datadriven::ARFFTools::readARFFFromFile(filename);
 
  if (dataset.getDimension() <= 0) {
    std::cout << "# Failed to read dataset! " << filename << std::endl;
    exit(-1);
  } else {
    std::cout << "# dataset dimensionality: " << dataset.getDimension() << std::endl;
  }
  return dataset;
}
 
void parseInputValue(char *inputString, size_t &outputValue) {
  std::basic_stringstream<char> argumentParser = std::stringstream(inputString);
  argumentParser >> outputValue;
  if (argumentParser.fail()) {
    throw sgpp::base::application_exception("Failed to parse parameter");
  }
}
#endif /* USE_MPI */