int main (int argc, char **argv)
{
Onyx::LinearLaRank::Classifier *classifier;
Onyx::Example::Dataset datasetTrain;
Onyx::Example::Dataset datasetTest;
std::chrono::time_point<std::chrono::system_clock> start, end; // Timings
std::string filenameTrainingData;
std::string filenameTrainingLabels;
std::string filenameTestData;
std::string filenameTestLabels;
std::string saveClassifier;
std::string loadClassifier;
bool doTraining = true; // Enable all by default
bool doTesting = true; // Enable by default
bool doOnlineTesting = false; // Disable by default
unsigned int numEpochs = 10;
double C = 1.0;
double tau = 0.0001;
// Random number generator (with random seed)
std::mt19937 random_number_generator(std::random_device{}());
// *** Print banner ***
std::cout << "Onyx v.1.0, (C) 2015 Rok Mandeljc <*****@*****.**>" << std::endl;
std::cout << std::endl;
// *** Command-line parser ***
boost::program_options::options_description commandLineArguments("Onyx - Online Classifier Application");
boost::program_options::variables_map optionsMap;
boost::program_options::positional_options_description positionalArguments;
boost::program_options::options_description argArguments("Arguments");
argArguments.add_options()
("help", "produce help message")
("training-data", boost::program_options::value<std::string>(&filenameTrainingData), "name of training .data file")
("training-labels", boost::program_options::value<std::string>(&filenameTrainingLabels), "name of training .labels file")
("test-data", boost::program_options::value<std::string>(&filenameTestData), "name of test .data file")
("test-labels", boost::program_options::value<std::string>(&filenameTestLabels), "name of test .labels file")
("save-classifier", boost::program_options::value<std::string>(&saveClassifier), "optional filename to store classifier")
("load-classifier", boost::program_options::value<std::string>(&loadClassifier), "optional filename to load classifier")
("epochs", boost::program_options::value<unsigned int>(&numEpochs)->default_value(numEpochs), "number of epochs for (re)training")
("training", boost::program_options::value<bool>(&doTraining)->default_value(doTraining), "enable training (if training data is available)")
("test", boost::program_options::value<bool>(&doTesting)->default_value(doTesting), "enable testing (if testing data is available)")
("online-test", boost::program_options::value<bool>(&doOnlineTesting)->default_value(doOnlineTesting), "enable on-line testing (if testing data is available)")
;
commandLineArguments.add(argArguments);
boost::program_options::options_description argParameters("Algorithm parameters");
argParameters.add_options()
("C", boost::program_options::value<double>(&C)->default_value(C), "SVM regularization parameter")
("tau", boost::program_options::value<double>(&tau)->default_value(tau), "tolerance for choosing new support vectors")
;
commandLineArguments.add(argParameters);
positionalArguments.add("training-data", 1);
positionalArguments.add("training-labels", 1);
positionalArguments.add("test-data", 1);
positionalArguments.add("test-labels", 1);
// Parse command-line
try {
boost::program_options::store(boost::program_options::command_line_parser(argc, argv).options(commandLineArguments).positional(positionalArguments).run(), optionsMap);
} catch (std::exception &error) {
std::cout << commandLineArguments << std::endl << std::endl;
std::cout << "Command-line error: " << error.what() << std::endl;
return -1;
}
// Display help?
if (optionsMap.count("help")) {
std::cout << commandLineArguments << std::endl;
return 1;
}
// Validate
try {
boost::program_options::notify(optionsMap);
} catch (std::exception &error) {
std::cout << commandLineArguments << std::endl << std::endl;
std::cout << "Argument error: " << error.what() << std::endl;
return -1;
}
bool trainingDataAvailable = !filenameTrainingData.empty() && !filenameTrainingLabels.empty();
bool testingDataAvailable = !filenameTestData.empty() && !filenameTestLabels.empty();
doTraining = doTraining && trainingDataAvailable;
doTesting = doTesting && testingDataAvailable;
doOnlineTesting = doOnlineTesting && testingDataAvailable;
if (!doTraining && !doTesting && !doOnlineTesting) {
std::cout << "Doing neither training nor testing; nothing to do!" << std::endl;
return 1;
}
if (!doTraining && loadClassifier.empty()) {
std::cout << "Neither training dataset nor pre-trained classifier provided!" << std::endl;
return -1;
}
// *** Load datasets ***
if (doTraining) {
std::cout << "Loading training dataset..." << std::endl;
try {
datasetTrain.load(filenameTrainingData, filenameTrainingLabels);
} catch (std::exception &error) {
std::cout << "Failed to load training dataset: " << error.what() << std::endl;
return -2;
}
std::cout << "Loaded training set:" << std::endl;
std::cout << " data file: " << filenameTrainingData << std::endl;
std::cout << " labels file: " << filenameTrainingLabels << std::endl;
std::cout << " samples: " << datasetTrain.numSamples << std::endl;
std::cout << " features: " << datasetTrain.numFeatures << std::endl;
std::cout << " classes: " << datasetTrain.numClasses << std::endl;
std::cout << std::endl;
}
if (doTesting || doOnlineTesting) {
std::cout << "Loading testing dataset..." << std::endl;
try {
datasetTest.load(filenameTestData, filenameTestLabels);
} catch (std::exception &error) {
std::cout << "Failed to load testing dataset: " << error.what() << std::endl;
return -2;
}
std::cout << "Loaded testing set:" << std::endl;
std::cout << " data file: " << filenameTestData << std::endl;
std::cout << " labels file: " << filenameTestLabels << std::endl;
std::cout << " samples: " << datasetTest.numSamples << std::endl;
std::cout << " features: " << datasetTest.numFeatures << std::endl;
std::cout << " classes: " << datasetTest.numClasses << std::endl;
std::cout << std::endl;
}
// *** Classifier ***
classifier = Onyx::LinearLaRank::create_classifier();
if (!loadClassifier.empty()) {
// Load from file
std::cout << "Loading classifier from file: " << loadClassifier << std::endl;
std::cout << std::endl;
std::ifstream stream(loadClassifier, std::ios::binary);
classifier->loadFromStream(stream);
} else {
// Create new classifier
std::cout << "Creating new classifier..." << std::endl;
std::cout << std::endl;
classifier->setC(C);
classifier->setTau(tau);
}
// *** Training (with optional testing) ***
if (doTraining) {
start = std::chrono::system_clock::now();
int sampleRatio = datasetTrain.numSamples / 10;
std::vector<float> trainError(numEpochs, 0.0);
std::vector<float> testError(numEpochs, 0.0);
std::vector<int> sampleIndices(datasetTrain.numSamples);
std::iota(sampleIndices.begin(), sampleIndices.end(), 0);
for (unsigned int epoch = 0; epoch < numEpochs; epoch++) {
std::cout << "Epoch " << epoch << std::endl;
// *** Train ***
// Randomly permute the sample indices
std::vector<std::vector<int>::iterator> shuffledSampleIndices(sampleIndices.size());
std::iota(shuffledSampleIndices.begin(), shuffledSampleIndices.end(), sampleIndices.begin());
std::shuffle(shuffledSampleIndices.begin(), shuffledSampleIndices.end(), random_number_generator);
for (unsigned int s = 0; s < shuffledSampleIndices.size(); s++) {
int idx = *shuffledSampleIndices[s];
const Eigen::VectorXf &sampleFeature = datasetTrain.features[idx];
int sampleLabel = datasetTrain.labels[idx];
// Estimate training error
int label = classifier->predict(sampleFeature);
if (label != sampleLabel) {
trainError[epoch]++;
}
// Update
classifier->update(sampleFeature, sampleLabel, 1.0f);
// Print progress
if (s && sampleRatio && s % sampleRatio == 0) {
std::cout << "Epoch: " << epoch << ": ";
std::cout << (10 * s) / sampleRatio << "%";
std::cout << " -> training error: " << trainError[epoch];
std::cout << "/" << s;
std::cout << " = " << trainError[epoch]/s*100 << "%";
std::cout << std::endl;
}
}
if (doTesting) {
// *** Test ***
for (unsigned int s = 0; s < datasetTest.numSamples; s++) {
const Eigen::VectorXf &sampleFeature = datasetTest.features[s];
int sampleLabel = datasetTest.labels[s];
if (classifier->predict(sampleFeature) != sampleLabel) {
testError[epoch]++;
}
}
std::cout << "Test error: " << testError[epoch] << "/" << datasetTest.numSamples << " = " << testError[epoch]/datasetTest.numSamples*100 << "%" << std::endl;
std::cout << std::endl;
}
}
end = std::chrono::system_clock::now();
std::cout << "Elapsed time: " << std::chrono::duration<float>(end - start).count() << " seconds." << std::endl;
}
// *** Save classifier ***
if (!saveClassifier.empty()) {
std::cout << "Saving classifier to file: " << saveClassifier << std::endl;
std::ofstream stream(saveClassifier, std::ios::binary);
classifier->saveToStream(stream);
std::cout << "Done!" << std::endl;
}
// *** Test - only if we didn't do the training ***
if (!doTraining && doTesting) {
float testError = 0.0f;
std::cout << "Performing off-line test..." << std::endl;
// Experiment
start = std::chrono::system_clock::now();
for (unsigned int s = 0; s < datasetTest.numSamples; s++) {
const Eigen::VectorXf &sampleFeature = datasetTest.features[s];
int sampleLabel = datasetTest.labels[s];
if (classifier->predict(sampleFeature) != sampleLabel) {
testError++;
}
}
end = std::chrono::system_clock::now();
std::cout << "Test error: " << testError << "/" << datasetTest.numSamples << " = " << testError/datasetTest.numSamples*100 << "%" << std::endl;
std::cout << "Elapsed time: " << std::chrono::duration<float>(end - start).count() << " seconds." << std::endl;
std::cout << std::endl;
}
// *** Online test ***
if (doOnlineTesting) {
float testError = 0.0f;
std::cout << "Performing on-line test..." << std::endl;
// Randomly permute the sample indices
std::vector<int> sampleIndices(datasetTest.numSamples);
std::iota(sampleIndices.begin(), sampleIndices.end(), 0);
std::vector<std::vector<int>::iterator> shuffledSampleIndices(sampleIndices.size());
std::iota(shuffledSampleIndices.begin(), shuffledSampleIndices.end(), sampleIndices.begin());
std::shuffle(shuffledSampleIndices.begin(), shuffledSampleIndices.end(), random_number_generator);
// Experiment
start = std::chrono::system_clock::now();
for (unsigned int s = 0; s < shuffledSampleIndices.size(); s++) {
int idx = *shuffledSampleIndices[s];
const Eigen::VectorXf &sampleFeature = datasetTest.features[idx];
int sampleLabel = datasetTest.labels[idx];
// Predict
if (classifier->predict(sampleFeature) != sampleLabel) {
testError++;
}
// Update
classifier->update(sampleFeature, sampleLabel);
}
end = std::chrono::system_clock::now();
std::cout << "Online test error: " << testError << "/" << datasetTest.numSamples << " = " << testError/datasetTest.numSamples*100 << "%" << std::endl;
std::cout << "Elapsed time: " << std::chrono::duration<float>(end - start).count() << " seconds." << std::endl;
std::cout << std::endl;
}
// Cleanup
delete classifier;
return 0;
}
///////////////////////////////////////////////////////////
// uses particles, variance, ind
// uses newPoints, newIndices, trees, Ndens
void gibbs2(unsigned int _Ndens, const BallTreeDensity* _trees,
unsigned long Np, unsigned int Niter,
double *_pts, BallTree::index *_ind,
double *_randU, double* _randN)
{
unsigned int i,j,l;
unsigned long s, maxNp;
Ndens = _Ndens; // SET UP GLOBALS
trees = _trees;
newPoints = _pts; newIndices = _ind;
randU = _randU; randN = _randN;
Ndim = trees[0].Ndim(); // dimension of densities
maxNp = 0; // largest # of particles we deal with
for (unsigned int j=0; j<Ndens; j++) // compute Max Np over all densities
if (maxNp < trees[j].Npts()) maxNp = trees[j].Npts();
ind = new BallTree::index[Ndens]; // ALLOCATE GLOBALS
p = new double[maxNp];
Nlevels = (unsigned int) (log((double)maxNp)/log((double)2))+1; // how many levels to a balanced binary tree?
particles = new double[Ndim*Ndens];
variance = new double[Ndim*Ndens];
dNpts = new unsigned long[Ndens];
levelList = new BallTree::index*[Ndens];
levelListNew = new BallTree::index*[Ndens];
for (j=0;j<Ndens;j++) {
levelList[j] = new BallTree::index[maxNp];
levelListNew[j] = new BallTree::index[maxNp];
}
for (s=0; s<Np; s++) {
levelInit();
initIndices();
calcIndices();
///////////////////////////////////////////////////////////////
// Perform Gibbs sampling only if multiple densities in product
///////////////////////////////////////////////////////////////
samplePoint(newPoints);
for (l=0;l<Nlevels;l++) {
levelDown();
for (i=0;i<Niter;i++) {
sampleIndices(newPoints);
samplePoint(newPoints);
}}
for (unsigned int j=0; j<Ndens; j++) // save and
newIndices[j] = trees[j].getIndexOf(ind[j])+1; // return particle label
newIndices += Ndens; // move pointers to next sample
newPoints += Ndim;
}
for (j=0;j<Ndens;j++) { delete[] levelList[j]; delete[] levelListNew[j]; }
delete[] levelList; delete[] levelListNew;
delete[] dNpts;
delete[] ind; delete[] p; delete[] particles; delete[] variance;
};
Result solve(const Problem& iProblem) const {
// set up initial (empty) result
Result result;
result.mSuccess = false;
result.mNumIterations = 0;
// ensure that there are enough data points to proceed
const int sampleSize = iProblem.getSampleSize();
const int n = iProblem.getNumDataPoints();
if (n < sampleSize) {
return result;
}
const double epsilon = 1e-10;
// best results are currently invalid
double bestScore = -1;
bool success = false;
// start number of iterations as infinite, then reduce as we go
double numIterationsNeeded = 1e10;
int iterationCount = 0;
int skippedSampleCount = 0;
// for random sample index generation
std::vector<int> allIndices(n);
// iterate until adaptive number of iterations are exceeded
while (iterationCount < numIterationsNeeded) {
// determine random sample indices
for (int i = 0; i < n; ++i) {
allIndices[i] = i;
}
for (int i = 0; i < sampleSize; ++i) {
int randIndex = std::rand() % n;
std::swap(allIndices[i], allIndices[randIndex]);
}
std::vector<int> sampleIndices(allIndices.begin(),
allIndices.begin() + sampleSize);
// compute solution on minimal set
typename Problem::Solution solution = iProblem.estimate(sampleIndices);
// compute errors over all data points
std::vector<double> errors2 = iProblem.computeSquaredErrors(solution);
// check whether this is a valid sample
// TODO: this should be done via a method in Problem class, but would
// require changing all existing usages to include that method
if (errors2.size() == 0) {
++skippedSampleCount;
if (skippedSampleCount >= mMaximumIterations) break;
continue;
}
skippedSampleCount = 0;
// compute error threshold to be applied to each term
double thresh = mMaximumError;
if (thresh < 0) {
std::sort(errors2.begin(), errors2.end());
double median = (n % 2 == 0) ?
(0.5*(errors2[n/2]+errors2[n/2+1])) : errors2[n/2];
thresh = 1.4826*std::sqrt(median)*4.6851;
}
thresh *= thresh;
// determine inliers
std::vector<int> inliers;
inliers.reserve(n);
for (int i = 0; i < n; ++i) {
if (errors2[i] <= thresh) {
inliers.push_back(i);
}
}
// if this is the best score, update solution and convergence criteria
double score = inliers.size();
if (score > bestScore) {
bestScore = score;
result.mInliers = inliers;
result.mSolution = solution;
success = true;
double inlierProbability = double(inliers.size()) / n;
double anyOutlierProbability = 1 - pow(inlierProbability,sampleSize);
anyOutlierProbability = std::min(anyOutlierProbability, 1-epsilon);
anyOutlierProbability = std::max(anyOutlierProbability, epsilon);
numIterationsNeeded =
log(1-mGoodSolutionProbability) / log(anyOutlierProbability);
}
// bump up iteration count and terminate if it exceeds hard max
++iterationCount;
if (iterationCount > mMaximumIterations) {
break;
}
}
// finish off result params
result.mSuccess = success;
result.mNumIterations = iterationCount;
// refine result using all inliers if specified
if (result.mSuccess && mRefineUsingInliers) {
result.mSolution = iProblem.estimate(result.mInliers);
}
// done
return result;
}
