////////////////////////////////////////////////////////////////////////////////
// 'processFeatureVector' reads a feature vector and modifies the classifier's
// data members accordingly
////////////////////////////////////////////////////////////////////////////////
void
MaxEnt::processFeatureVector(const FeatureVector& features){
// Class=2; // what does this do? Need to be there?
en.f.resize(C);
en.fs.resize(C);
en.Ny.resize(C);
size_t c=0; // class
en.Ny[c]=0.0; // strength of class (first num after @)
// for each feature
for (FeatureVector::const_iterator feat = features.begin();
feat!=features.end();feat++)
{
// these loops may have to be reversed !!
// the classes may need to be looped over first, and features within
// instead of classes within feature loop
StringXML feature;
double value = feat->getValue();
// for each class
for (unsigned int i=0;i<C;i++){
stringstream ss;
ss << "cat" << i << "_" << feat->getFeature();
feature = ss.str();
if (value!=0) {
// set values
if (f2s.count(feature)!=0) { // check whether feature exists
(s2f)[f2s[feature]].second+=(en.Ny[i]!=0);
en.f[i].push_back(make_pair(f2s[feature], value));
en.fs[i]+=value;
F=max(F, en.fs[i]);
}
}
}
}
}
ErrorStruct StrongClassifier::errorForFeatures(const TrainingData &features, bool printStats) const {
ErrorStruct e;
for (int i=0; i<features.size(); i++) {
FeatureVector feature = *(features.feature(i));
if (decide(feature)) {
feature.val() == POS ? e.true_pos++ : e.false_pos++;//it is really positive
} else {
feature.val() == NEG ? e.true_neg++ : e.false_neg++;//it is really negative
}
}
// if all 10 samples, 3 is misclassified, error is 3/10
e.error = (e.false_pos + e.false_neg) / ((float)features.size());
if (printStats) {
std::cout << e.true_pos << " true positives" << std::endl;
std::cout << e.false_pos << " false positives" << std::endl;
std::cout << e.true_neg << " true negatives" << std::endl;
std::cout << e.false_neg << " false negatives" << std::endl;
std::cout << e.error * 100 << "% error" << std::endl;
std::cout << std::endl;
}
return e;
}
bool StrongClassifier::decide(const FeatureVector &featureVector) const {
std::vector<float> features(featureVector.size());
for (int i=0; i<featureVector.size(); i++) {
features[i] = featureVector.at(i);
}
return decide(features);
}
realv Layer::signalWeighting(FeatureVector _signal, Mat _weights) {
if (_signal.getLength() != ((uint) _weights.cols)) {
throw length_error("Layer : Uncorrect length between signal and weights");
}
realv sum = 0;
for (uint i = 0; i < _signal.getLength(); i++) {
sum += _signal[i] * _weights.at<realv>(0, i);
}
return sum;
}
double Slic::_ComputeDistance(const FeatureVector &vec1, const FeatureVector &vec2)
{
assert(vec1.size() == vec2.size());
double dis = 0;
for (unsigned int i=0;i<vec1.size();++i)
{
dis += (vec1[i]-vec2[i])*(vec1[i]-vec2[i]);
}
return dis;
}
realv SEMeasurer::totalError(FeatureVector _output, FeatureVector _target) {
if (_output.getLength() != _target.getLength()) {
throw length_error("SE : Output and target do not have the same size");
}
realv result = 0;
for (uint i = 0; i < _output.getLength(); i++) {
result += (_output[i] - _target[i]) * (_output[i] - _target[i]);
}
err = result;
return err;
}
ErrorVector SEMeasurer::errorPerUnit(FeatureVector _output, FeatureVector _target) {
if (_output.getLength() != _target.getLength()) {
throw length_error("SEMeasurer : Output and target do not have the same size");
}
ErrorVector result(_output.getLength());
for (uint i = 0; i < _output.getLength(); i++) {
result[i] = (_output[i] - _target[i]) * (_output[i] - _target[i]);
}
errPerUnit = result;
return result;
}
/**
* Determines the Random Approximation of GVC
* (Determines the optimal Regularization Factor)
*/
double KerDenSOM::randApproxGVC(const TS* _examples, const FuzzyMap* _som, double _dataSD, double _reg)
{
unsigned j, vv, cc;
double num, den, r;
FeatureVector VV;
VV.resize(dim, 0.0);
FuzzyMap tmpSOM(*_som);
TS tmpTS(*_examples);
num = 0;
for (vv = 0; vv < numVectors; vv++)
{
for (j = 0; j < dim; j++)
VV[j] = 0.0;
for (cc = 0; cc < numNeurons; cc++)
{
for (j = 0; j < dim; j++)
VV[j] += (_examples->theItems[vv][j] - _som->theItems[cc][j]) * _som->memb[vv][cc];
}
for (j = 0; j < dim; j++)
num += VV[j] * VV[j];
}
init_random_generator();
for (vv = 0; vv < numVectors; vv++)
{
for (j = 0; j < dim; j++)
tmpTS.theItems[vv][j] += rnd_gaus() * _dataSD;
}
updateV(&tmpSOM, &tmpTS, _reg);
den = 0.0;
init_random_generator();
for (vv = 0; vv < numVectors; vv++)
{
for (j = 0; j < dim; j++)
VV[j] = 0.0;
for (cc = 0; cc < numNeurons; cc++)
{
for (j = 0; j < dim; j++)
VV[j] += (tmpTS.theItems[vv][j] - tmpSOM.theItems[cc][j] - _examples->theItems[vv][j] + _som->theItems[cc][j]) * _som->memb[vv][cc];
}
r = 0.;
for (j = 0; j < dim; j++)
r += VV[j] * rnd_gaus() * _dataSD;
den += r * r;
}
if (den != 0)
return (double) num / den;
else
return 0.;
}
inline
Tp dot_product(const FeatureVector<Tp1, Alloc1>& x, Iterator first, Iterator last, Tp __dot)
{
typedef FeatureVector<Tp1, Alloc1> feature_vector1_type;
for (/**/; first != last; ++ first) {
typename feature_vector1_type::const_iterator iter = x.find(first->first);
if (iter != x.end())
__dot += iter->second * first->second;
}
return __dot;
}
inline
Tp dot_product(Iterator first, Iterator last, const FeatureVector<Tp2, Alloc2>& y, Tp __dot)
{
typedef FeatureVector<Tp2, Alloc2> feature_vector2_type;
for (/**/; first != last; ++ first) {
typename feature_vector2_type::const_iterator iter = y.find(first->first);
if (iter != y.end())
__dot += first->second * iter->second;
}
return __dot;
}
void SegmentList::normalizeFeatures() {
if (isEmpty()) return;
FeatureVector min = at(0)->features();
FeatureVector max = at(0)->features();
foreach (Segment const * const segment, *this) {
min.setCompwiseMin(segment->features());
max.setCompwiseMax(segment->features());
}
max = max - min;
foreach (Segment * const segment, *this) {
segment->features() = (segment->features() - min) / max;
}
}
/**************************************
* Fucntion: is_classifier_correct
* -------------------------------
* returns true if weak classifier (wc) correctly identified the
* feature vector (fv), false otherwise.
*/
bool AdaBooster::is_classifier_correct(WeakClassifier &wc, FeatureVector &fv){
// check if threshold is greater than (or equal to) feature
bool guess = ( wc.threshold() >= fv.at(wc.dimension()) );
// if classifier is flipped, negate guess
guess = wc.isFlipped() ? !guess : guess;
// find actual value of point
bool real = ( fv.val() == POS );
// return if guess and real agree
return ( real == guess );
}
PERIPHERAL_ERROR GetFeatures(const JOYSTICK_INFO* joystick, const char* controller_id,
unsigned int* feature_count, JOYSTICK_FEATURE** features)
{
if (!joystick || !controller_id || !feature_count || !features)
return PERIPHERAL_ERROR_INVALID_PARAMETERS;
FeatureVector featureVector;
if (CStorageManager::Get().GetFeatures(ADDON::Joystick(*joystick), controller_id, featureVector))
{
*feature_count = featureVector.size();
ADDON::JoystickFeatures::ToStructs(featureVector, features);
return PERIPHERAL_NO_ERROR;
}
return PERIPHERAL_ERROR_FAILED;
}
int RecBase :: recognizeEuclidean(FeatureVector& charvec, int gHint){
int cat,rank,i;
double d,d0;
initResultTable();
d0 = DBL_MAX;
for ( cat=0 ; cat<n_cat ; cat++ ){
// d = charvec.distEuclidean2(dic[cat]);
d = charvec.distEuclidean2(dic[cat],d0);
if ( d >= d0 ) continue;
for ( rank=0 ; rank<n_top ; rank++ ){
if ( d >= resultTable[rank].dist ) continue;
for ( i=n_top-2 ; i>=rank ; i-- ){
resultTable[i+1] = resultTable[i];
}
resultTable[rank].dist = d;
resultTable[rank].id = cat;
d0 = resultTable[n_top-1].dist;
break;
}
}
for ( i=0 ; i<n_top ; i++ ){
resultTable[i].dist = sqrt(resultTable[i].dist);
}
return(0);
}
inline
Tp1 dot_product(const FeatureVector<Tp1, Alloc1>& x, const WeightVector<Tp, Alloc>& w, const FeatureVector<Tp2, Alloc2>& y)
{
typedef FeatureVector<Tp1, Alloc1> feature_vector1_type;
typedef WeightVector<Tp, Alloc> weight_vector_type;
typedef FeatureVector<Tp2, Alloc2> feature_vector2_type;
if (x.empty() || y.empty()) return Tp1();
if (static_cast<const void*>(&x) == static_cast<const void*>(&y))
return details::__inner_product(x.begin(), x.end(), w, Tp1());
if (x.size() < y.size())
return dot_product(x.begin(), x.end(), w, y, Tp1());
else
return dot_product(x, w, y.begin(), y.end(), Tp1());
}
void VectorDataSet::weightedSum(FeatureVector& result, const std::vector<int> &patterns,
const std::vector<double> &alpha)
{
int pIndex;
// ASA FeatureVector is an alias for vector<double>
vector<double> weight_temp (numFeatures,0);
for (unsigned int i = 0; i < patterns.size(); i++){
pIndex = patterns[i];
for (int j = 0; j != X[pIndex].size(); ++j){
weight_temp[j] += X[pIndex][j] * alpha[i];
}
}
result.clear();
result.initialize(weight_temp);
}
inline
Tp1 dot_product(const FeatureVector<Tp1, Alloc1>& x, const FeatureVector<Tp2, Alloc2>& y)
{
typedef FeatureVector<Tp1, Alloc1> feature_vector1_type;
typedef FeatureVector<Tp2, Alloc2> feature_vector2_type;
if (x.empty() || y.empty()) return Tp1();
// if the same address, we are identical!
if (static_cast<const void*>(&x) == static_cast<const void*>(&y))
return details::__inner_product(x.begin(), x.end(), Tp1());
if (x.size() < y.size())
return dot_product(x.begin(), x.end(), y, Tp1());
else
return dot_product(x, y.begin(), y.end(), Tp1());
}
FeatureSet FeatureSet::clone() const
{
FeatureSet other;
for (FeatureSet::const_iterator i = begin(); i != end(); ++i) {
FeatureVector cv = i->clone();
// check that all data is the same:
if (cv.size() != i->size())
throw std::runtime_error("ssssssssss");
for (unsigned int idx=0; idx<cv.size(); ++idx) {
if (cv[idx] != (*i)[idx]) {
cout << "(" << cv[idx] << "!=" << (*i)[idx] << ")" << flush;
throw std::runtime_error("??????????");
}
}
if (cv.getTag<TagTrueClassLabel>().label != i->getTag<TagTrueClassLabel>().label) {
cout << "(" << cv.getTag<TagTrueClassLabel>().label << "!=" << i->getTag<TagTrueClassLabel>().label << ")" << flush;
throw std::runtime_error("!!!!!!!!!!!");
}
other.push_back(cv);
}
return other;
}
int main(int argc, char** argv)
{
static const size_t K = 3;
typedef double Scalar;
typedef Eigen::Matrix<Scalar, 1, K> Feature;
typedef std::vector<Feature, Eigen::aligned_allocator<Feature> > FeatureVector;
FeatureVector features(1000*K);
FeatureVector centers;
std::vector<unsigned int> membership;
FILE* data = fopen(argv[1], "r");
for (size_t i = 0; i < features.size(); ++i) {
Feature& f = features[i];
fscanf(data, "%lf %lf %lf", &f[0], &f[1], &f[2]);
}
vt::SimpleKmeans<Feature> kmeans(Feature::Zero());
double sse = kmeans.cluster(features, K, centers, membership);
for (size_t i = 0; i < centers.size(); ++i) {
printf("%f %f %f\n", centers[i][0], centers[i][1], centers[i][2]);
}
printf("sse = %f\n", sse);
unsigned int pattern[K];
pattern[0] = membership[0];
pattern[1] = membership[1];
pattern[2] = membership[2];
for (size_t i = 0; i < membership.size(); i+=K) {
if (membership[i] != pattern[0] ||
membership[i+1] != pattern[1] ||
membership[i+2] != pattern[2]) {
printf("Misassignment!\n");
return -1;
}
}
}
void Slic::_InitCenterFeature(int i, int j, FeatureVector &featureVec)
{
featureVec.clear();
uchar *buffer = new uchar[5*5*_dataSize*_bandCount];
_poSrcDS->RasterIO(GF_Read,j-2,i-2,5,5,buffer,5,5,_dataType,_bandCount,0,0,0,0);
double minEdge = std::numeric_limits<double>::max();
int minN;
int minM;
for (int n=1;n<4;++n)
{
for(int m=1;m<4;++m)
{
double edge = 0;
edge += (buffer[(n-1)*5+m]-buffer[(n+1)*5+m])*(buffer[(n-1)*5+m]-buffer[(n+1)*5+m]);
edge += (buffer[n*5+m-1]-buffer[n*5+m+1])*(buffer[n*5+m-1]-buffer[n*5+m+1]);
if (edge < minEdge)
{
minEdge = edge;
minN = n;
minM = m;
}
}
}
int centerIndex = minN*5+minM;
uchar* p = buffer;
for(int k=0;k<_bandCount;++k,p += (5*5*_dataSize))
{
featureVec.push_back( SRCVAL(p,_dataType,centerIndex)/_regularizer);
}
featureVec.push_back(static_cast<double>(j+minM-2)/_regionSize); //x
featureVec.push_back(static_cast<double>(i+minN-2)/_regionSize); //y
delete []buffer;
}
void ClassifySvmSharedCommand::readSharedRAbundVectors(vector<SharedRAbundVector*>& lookup, GroupMap& designMap, LabeledObservationVector& labeledObservationVector, FeatureVector& featureVector) {
for ( int j = 0; j < lookup.size(); j++ ) {
//i++;
vector<individual> data = lookup[j]->getData();
Observation* observation = new Observation(data.size(), 0.0);
string sharedGroupName = lookup[j]->getGroup();
string treatmentName = designMap.getGroup(sharedGroupName);
//std::cout << "shared group name: " << sharedGroupName << " treatment name: " << treatmentName << std::endl;
//labeledObservationVector.push_back(std::make_pair(treatmentName, observation));
labeledObservationVector.push_back(LabeledObservation(j, treatmentName, observation));
//std::cout << " j=" << j << " label : " << lookup[j]->getLabel() << " group: " << lookup[j]->getGroup();
for (int k = 0; k < data.size(); k++) {
//std::cout << " abundance " << data[k].abundance;
observation->at(k) = double(data[k].abundance);
if ( j == 0) {
featureVector.push_back(Feature(k, m->currentSharedBinLabels[k]));
}
}
//std::cout << std::endl;
// let this happen later?
//delete lookup[j];
}
}
bool CButtonMapXml::Load(void)
{
TiXmlDocument xmlFile;
if (!xmlFile.LoadFile(m_strResourcePath))
{
esyslog("Error opening %s: %s", m_strResourcePath.c_str(), xmlFile.ErrorDesc());
return false;
}
TiXmlElement* pRootElement = xmlFile.RootElement();
if (!pRootElement || pRootElement->NoChildren() || pRootElement->ValueStr() != BUTTONMAP_XML_ROOT)
{
esyslog("Can't find root <%s> tag", BUTTONMAP_XML_ROOT);
return false;
}
const TiXmlElement* pDevice = pRootElement->FirstChildElement(DEVICES_XML_ELEM_DEVICE);
if (!pDevice)
{
esyslog("Can't find <%s> tag", DEVICES_XML_ELEM_DEVICE);
return false;
}
if (!CDeviceXml::Deserialize(pDevice, m_device))
return false;
const TiXmlElement* pController = pDevice->FirstChildElement(BUTTONMAP_XML_ELEM_CONTROLLER);
if (!pController)
{
esyslog("Device \"%s\": can't find <%s> tag", m_device.Name().c_str(), BUTTONMAP_XML_ELEM_CONTROLLER);
return false;
}
// For logging purposes
unsigned int totalFeatureCount = 0;
while (pController)
{
const char* id = pController->Attribute(BUTTONMAP_XML_ATTR_CONTROLLER_ID);
if (!id)
{
esyslog("Device \"%s\": <%s> tag has no attribute \"%s\"", m_device.Name().c_str(),
BUTTONMAP_XML_ELEM_CONTROLLER, BUTTONMAP_XML_ATTR_CONTROLLER_ID);
return false;
}
FeatureVector features;
if (!Deserialize(pController, features))
return false;
if (features.empty())
{
esyslog("Device \"%s\" has no features for controller %s", m_device.Name().c_str(), id);
}
else
{
totalFeatureCount += features.size();
m_buttonMap[id] = std::move(features);
}
pController = pController->NextSiblingElement(BUTTONMAP_XML_ELEM_CONTROLLER);
}
dsyslog("Loaded device \"%s\" with %u controller profiles and %u total features", m_device.Name().c_str(), m_buttonMap.size(), totalFeatureCount);
return true;
}
MLClassPtr Classifier2::ClassifyAImageOneLevel (FeatureVector& example,
double& probability,
kkint32& numOfWinners,
bool& knownClassOneOfTheWinners,
double& breakTie
)
{
probability = 0.0;
double probOfKnownClass = 0.0f;
probability = 0.0;
MLClassPtr origClass = example.MLClass ();
MLClassPtr predictedClass = NULL;
MLClassPtr predictedClass2 = NULL;
kkint32 class1Votes = -1;
kkint32 class2Votes = -1;
double predictedClass2Prob = 0.0f;
trainedModel->Predict (&example,
origClass,
predictedClass,
predictedClass2,
class1Votes,
class2Votes,
probOfKnownClass,
probability,
predictedClass2Prob,
numOfWinners,
knownClassOneOfTheWinners,
breakTie,
log
);
if (predictedClass == NULL)
{
log.Level (-1) << endl << endl
<< "Classifier2::ClassifyAImageOneLevel The trainedModel returned back a NULL pointer for predicted class" << endl
<< endl;
predictedClass = unKnownMLClass;
}
if (subClassifiers)
{
Classifier2Ptr subClassifer = LookUpSubClassifietByClass (predictedClass);
if (subClassifer)
{
double subProbability = 0.0;
kkint32 subNumOfWinners = 0;
double subBreakTie = 0.0;
/**@todo make sure that the following call does not normalize the features. */
MLClassPtr subPrediction
= subClassifer->ClassifyAImageOneLevel (example, subProbability, subNumOfWinners, knownClassOneOfTheWinners, subBreakTie);
if (subPrediction)
{
probability = probability * subProbability;
numOfWinners = numOfWinners + subNumOfWinners;
breakTie += subBreakTie * (1.0 - breakTie);
}
}
}
if (predictedClass->UnDefined ())
predictedClass = noiseMLClass;
example.MLClass (predictedClass);
return predictedClass;
} /* ClassifyAImageOneLevel */
inline
Tp1 dot_product(const WeightVector<Tp1, Alloc1>& x, const FeatureVector<Tp2, Alloc2>& y)
{
return dot_product(x, y.begin(), y.end(), Tp1());
}
void Classifier2::ClassifyAExample (FeatureVector& example,
MLClassPtr& predClass1,
MLClassPtr& predClass2,
kkint32& predClass1Votes,
kkint32& predClass2Votes,
double& knownClassProb,
double& predClass1Prob,
double& predClass2Prob,
kkint32& numOfWinners,
double& breakTie
)
{
bool knownClassOneOfTheWiners = false;
predClass1 = NULL;
predClass2 = NULL;
knownClassProb = -1.0f;
predClass1Prob = -1.0f;
predClass2Prob = -1.0f;
MLClassPtr origClass = example.MLClass ();
trainedModel->Predict (&example,
origClass,
predClass1,
predClass2,
predClass1Votes,
predClass2Votes,
knownClassProb,
predClass1Prob,
predClass2Prob,
numOfWinners,
knownClassOneOfTheWiners,
breakTie,
log
);
if (!predClass1)
predClass1 = noiseMLClass;
if (subClassifiers)
{
Classifier2Ptr subClassifer = LookUpSubClassifietByClass (predClass1);
if (subClassifer)
{
MLClassPtr subPredClass1 = NULL;
MLClassPtr subPredClass2 = NULL;
kkint32 subPredClass1Votes = 0;
kkint32 subPredClass2Votes = 0;
double subKnownClassProb = 0.0;
double subPredClass1Prob = 0.0;
double subPredClass2Prob = 0.0;
kkint32 subNumOfWinners = 0;
double subBreakTie = 0.0;
subClassifer->ClassifyAExample (example, subPredClass1, subPredClass2,
subPredClass1Votes, subPredClass2Votes, subKnownClassProb,
subPredClass1Prob, subPredClass2Prob, subNumOfWinners,
subBreakTie
);
predClass1 = subPredClass1;
predClass1Votes += subPredClass1Votes;
predClass1Prob *= subPredClass1Prob;
knownClassProb *= subKnownClassProb;
numOfWinners += subNumOfWinners;
breakTie += subBreakTie * (1.0 - breakTie);
}
subClassifer = LookUpSubClassifietByClass (predClass2);
if (subClassifer)
{
MLClassPtr subPredClass1 = NULL;
MLClassPtr subPredClass2 = NULL;
kkint32 subPredClass1Votes = 0;
kkint32 subPredClass2Votes = 0;
double subKnownClassProb = 0.0;
double subPredClass1Prob = 0.0;
double subPredClass2Prob = 0.0;
kkint32 subNumOfWinners = 0;
double subBreakTie = 0.0;
subClassifer->ClassifyAExample (example, subPredClass1, subPredClass2,
subPredClass1Votes, subPredClass2Votes, subKnownClassProb,
subPredClass1Prob, subPredClass2Prob, subNumOfWinners,
subBreakTie
);
predClass2 = subPredClass1;
predClass2Votes += subPredClass1Votes;
predClass2Prob *= subPredClass1Prob;
}
}
example.MLClass (predClass1);
return;
} /* ClassifyAExample */
void Pipe::TrainEpoch(int epoch) {
Instance *instance;
Parts *parts = CreateParts();
Features *features = CreateFeatures();
vector<double> scores;
vector<double> gold_outputs;
vector<double> predicted_outputs;
double total_cost = 0.0;
double total_loss = 0.0;
double eta;
int num_instances = instances_.size();
double lambda = 1.0/(options_->GetRegularizationConstant() *
(static_cast<double>(num_instances)));
timeval start, end;
gettimeofday(&start, NULL);
int time_decoding = 0;
int time_scores = 0;
int num_mistakes = 0;
LOG(INFO) << " Iteration #" << epoch + 1;
dictionary_->StopGrowth();
for (int i = 0; i < instances_.size(); i++) {
int t = num_instances * epoch + i;
instance = instances_[i];
MakeParts(instance, parts, &gold_outputs);
MakeFeatures(instance, parts, features);
// If using only supported features, must remove the unsupported ones.
// This is necessary not to mess up the computation of the squared norm
// of the feature difference vector in MIRA.
if (options_->only_supported_features()) {
RemoveUnsupportedFeatures(instance, parts, features);
}
timeval start_scores, end_scores;
gettimeofday(&start_scores, NULL);
ComputeScores(instance, parts, features, &scores);
gettimeofday(&end_scores, NULL);
time_scores += diff_ms(end_scores, start_scores);
if (options_->GetTrainingAlgorithm() == "perceptron" ||
options_->GetTrainingAlgorithm() == "mira" ) {
timeval start_decoding, end_decoding;
gettimeofday(&start_decoding, NULL);
decoder_->Decode(instance, parts, scores, &predicted_outputs);
gettimeofday(&end_decoding, NULL);
time_decoding += diff_ms(end_decoding, start_decoding);
if (options_->GetTrainingAlgorithm() == "perceptron") {
for (int r = 0; r < parts->size(); ++r) {
if (!NEARLY_EQ_TOL(gold_outputs[r], predicted_outputs[r], 1e-6)) {
++num_mistakes;
}
}
eta = 1.0;
} else {
CHECK(false) << "Plain mira is not implemented yet.";
}
MakeGradientStep(parts, features, eta, t, gold_outputs,
predicted_outputs);
} else if (options_->GetTrainingAlgorithm() == "svm_mira" ||
options_->GetTrainingAlgorithm() == "crf_mira" ||
options_->GetTrainingAlgorithm() == "svm_sgd" ||
options_->GetTrainingAlgorithm() == "crf_sgd") {
double loss;
timeval start_decoding, end_decoding;
gettimeofday(&start_decoding, NULL);
if (options_->GetTrainingAlgorithm() == "svm_mira" ||
options_->GetTrainingAlgorithm() == "svm_sgd") {
// Do cost-augmented inference.
double cost;
decoder_->DecodeCostAugmented(instance, parts, scores, gold_outputs,
&predicted_outputs, &cost, &loss);
total_cost += cost;
} else {
// Do marginal inference.
double entropy;
decoder_->DecodeMarginals(instance, parts, scores, gold_outputs,
&predicted_outputs, &entropy, &loss);
CHECK_GE(entropy, 0.0);
}
gettimeofday(&end_decoding, NULL);
time_decoding += diff_ms(end_decoding, start_decoding);
if (loss < 0.0) {
if (!NEARLY_EQ_TOL(loss, 0.0, 1e-9)) {
LOG(INFO) << "Warning: negative loss set to zero: " << loss;
}
loss = 0.0;
}
total_loss += loss;
// Compute difference between predicted and gold feature vectors.
FeatureVector difference;
MakeFeatureDifference(parts, features, gold_outputs, predicted_outputs,
&difference);
// Get the stepsize.
if (options_->GetTrainingAlgorithm() == "svm_mira" ||
options_->GetTrainingAlgorithm() == "crf_mira") {
double squared_norm = difference.GetSquaredNorm();
double threshold = 1e-9;
if (loss < threshold || squared_norm < threshold) {
eta = 0.0;
} else {
eta = loss / squared_norm;
if (eta > options_->GetRegularizationConstant()) {
eta = options_->GetRegularizationConstant();
}
}
} else {
if (options_->GetLearningRateSchedule() == "fixed") {
eta = options_->GetInitialLearningRate();
} else if (options_->GetLearningRateSchedule() == "invsqrt") {
eta = options_->GetInitialLearningRate() /
sqrt(static_cast<double>(t+1));
} else if (options_->GetLearningRateSchedule() == "inv") {
eta = options_->GetInitialLearningRate() /
static_cast<double>(t+1);
} else if (options_->GetLearningRateSchedule() == "lecun") {
eta = options_->GetInitialLearningRate() /
(1.0 + (static_cast<double>(t) / static_cast<double>(num_instances)));
} else {
CHECK(false) << "Unknown learning rate schedule: "
<< options_->GetLearningRateSchedule();
}
// Scale the parameter vector (only for SGD).
double decay = 1 - eta * lambda;
CHECK_GT(decay, 0.0);
parameters_->Scale(decay);
}
MakeGradientStep(parts, features, eta, t, gold_outputs,
predicted_outputs);
} else {
CHECK(false) << "Unknown algorithm: " << options_->GetTrainingAlgorithm();
}
}
// Compute the regularization value (halved squared L2 norm of the weights).
double regularization_value =
lambda * static_cast<double>(num_instances) *
parameters_->GetSquaredNorm() / 2.0;
delete parts;
delete features;
gettimeofday(&end, NULL);
LOG(INFO) << "Time: " << diff_ms(end,start);
LOG(INFO) << "Time to score: " << time_scores;
LOG(INFO) << "Time to decode: " << time_decoding;
LOG(INFO) << "Number of Features: " << parameters_->Size();
if (options_->GetTrainingAlgorithm() == "perceptron" ||
options_->GetTrainingAlgorithm() == "mira") {
LOG(INFO) << "Number of mistakes: " << num_mistakes;
}
LOG(INFO) << "Total Cost: " << total_cost << "\t"
<< "Total Loss: " << total_loss << "\t"
<< "Total Reg: " << regularization_value << "\t"
<< "Total Loss+Reg: " << total_loss + regularization_value << endl;
}
void
vtree_user::compute_features( const std::string& cloud_filename,
FeatureVector &feature_vector )
{
typedef pcl::PointXYZ nx_PointT;
typedef pcl::Normal nx_Normal;
typedef pcl::PointCloud<nx_PointT> nx_PointCloud;
typedef pcl::PointCloud<nx_Normal> nx_PointCloud_normal;
typedef pcl::PointCloud<int> nx_PointCloud_int;
typedef pcl::UniformSampling<nx_PointT> nx_Sampler;
typedef pcl::search::KdTree<nx_PointT> nx_SearchMethod;
typedef pcl::NormalEstimation<nx_PointT, nx_Normal> nx_NormalEst;
#if FEATURE == 1
typedef pcl::FPFHSignature33 nx_FeatureType;
typedef pcl::FPFHEstimation<nx_PointT, nx_Normal, nx_FeatureType> nx_FeatureEst;
#elif FEATURE == 2
typedef pcl::PFHSignature125 nx_FeatureType;
typedef pcl::PFHEstimation<nx_PointT, nx_Normal, nx_FeatureType> nx_FeatureEst;
#elif FEATURE == 3
typedef pcl::VFHSignature308 nx_FeatureType;
typedef pcl::VFHEstimation<nx_PointT, nx_Normal, nx_FeatureType> nx_FeatureEst;
#else
#error A valid feature definition is required!
#endif
typedef pcl::PointCloud<nx_FeatureType> nx_PointCloud_feature;
// load the file
nx_PointCloud::Ptr cld_ptr(new nx_PointCloud);
pcl::io::loadPCDFile<nx_PointT> ( cloud_filename, *cld_ptr);
ROS_INFO("[vtree_user] Starting keypoint extraction...");
clock_t tic = clock();
nx_PointCloud::Ptr keypoints( new nx_PointCloud);
nx_PointCloud_int::Ptr keypoint_idx(new nx_PointCloud_int);
nx_Sampler uniform_sampling;
uniform_sampling.setInputCloud ( cld_ptr );
uniform_sampling.setRadiusSearch ( keypoint_radius_ );
uniform_sampling.compute( *keypoint_idx );
pcl::copyPointCloud ( *cld_ptr, keypoint_idx->points, *keypoints);
ROS_INFO("[vtree_user] No of Keypoints found %d", static_cast<int>(keypoint_idx->size()) );
ROS_INFO("[vtree_user] Keypoint extraction took %f msec.", static_cast<double>((clock()-tic)*1000)/CLOCKS_PER_SEC );
if( keypoints->empty() )
{
ROS_WARN("[vtree_user] No keypoints were found...");
return;
}
// Compute normals for the input cloud
ROS_INFO("[vtree_user] Starting normal extraction...");
tic = clock();
nx_PointCloud_normal::Ptr normals (new nx_PointCloud_normal);
nx_SearchMethod::Ptr search_method_xyz (new nx_SearchMethod);
nx_NormalEst norm_est;
norm_est.setInputCloud ( cld_ptr );
norm_est.setSearchMethod ( search_method_xyz );
norm_est.setRadiusSearch ( normal_radius_ );
norm_est.compute ( *normals );
ROS_INFO("[vtree_user] Normal extraction took %f msec.", static_cast<double>((clock()-tic)*1000)/CLOCKS_PER_SEC );
// Get features at the computed keypoints
ROS_INFO("[vtree_user] Starting feature computation...");
tic = clock();
nx_PointCloud_feature::Ptr features(new nx_PointCloud_feature);
nx_FeatureEst feat_est;
feat_est.setInputCloud ( keypoints );
feat_est.setSearchSurface ( cld_ptr );
feat_est.setInputNormals ( normals );
search_method_xyz.reset(new nx_SearchMethod);
feat_est.setSearchMethod ( search_method_xyz );
feat_est.setRadiusSearch ( feature_radius_ );
feat_est.compute ( *features );
ROS_INFO("[vtree_user] No of Features found %d", static_cast<int>(features->size()) );
ROS_INFO("[vtree_user] Feature computation took %f msec.", static_cast<double>((clock()-tic)*1000)/CLOCKS_PER_SEC );
// Rectify the historgram values to ensure they are in [0,100] and create a document
for( nx_PointCloud_feature::iterator iter = features->begin();
iter != features->end(); ++iter)
{
rectify_histogram( iter->histogram );
feature_vector.push_back( FeatureHist( iter->histogram ) );
}
}
bool CButtonMapXml::Serialize(const FeatureVector& features, TiXmlElement* pElement)
{
if (pElement == NULL)
return false;
for (FeatureVector::const_iterator it = features.begin(); it != features.end(); ++it)
{
const ADDON::JoystickFeature& feature = *it;
if (!IsValid(feature))
continue;
TiXmlElement featureElement(BUTTONMAP_XML_ELEM_FEATURE);
TiXmlNode* featureNode = pElement->InsertEndChild(featureElement);
if (featureNode == NULL)
return false;
TiXmlElement* featureElem = featureNode->ToElement();
if (featureElem == NULL)
return false;
featureElem->SetAttribute(BUTTONMAP_XML_ATTR_FEATURE_NAME, feature.Name());
switch (feature.Type())
{
case JOYSTICK_FEATURE_TYPE_SCALAR:
{
SerializePrimitive(featureElem, feature.Primitive());
break;
}
case JOYSTICK_FEATURE_TYPE_ANALOG_STICK:
{
if (!SerializePrimitiveTag(featureElem, feature.Up(), BUTTONMAP_XML_ELEM_UP))
return false;
if (!SerializePrimitiveTag(featureElem, feature.Down(), BUTTONMAP_XML_ELEM_DOWN))
return false;
if (!SerializePrimitiveTag(featureElem, feature.Right(), BUTTONMAP_XML_ELEM_RIGHT))
return false;
if (!SerializePrimitiveTag(featureElem, feature.Left(), BUTTONMAP_XML_ELEM_LEFT))
return false;
break;
}
case JOYSTICK_FEATURE_TYPE_ACCELEROMETER:
{
if (!SerializePrimitiveTag(featureElem, feature.PositiveX(), BUTTONMAP_XML_ELEM_POSITIVE_X))
return false;
if (!SerializePrimitiveTag(featureElem, feature.PositiveY(), BUTTONMAP_XML_ELEM_POSITIVE_Y))
return false;
if (!SerializePrimitiveTag(featureElem, feature.PositiveZ(), BUTTONMAP_XML_ELEM_POSITIVE_Z))
return false;
break;
}
default:
break;
}
}
return true;
}
bool CButtonMapXml::Deserialize(const TiXmlElement* pElement, FeatureVector& features)
{
const TiXmlElement* pFeature = pElement->FirstChildElement(BUTTONMAP_XML_ELEM_FEATURE);
if (!pFeature)
{
esyslog("Can't find <%s> tag", BUTTONMAP_XML_ELEM_FEATURE);
return false;
}
while (pFeature)
{
const char* name = pFeature->Attribute(BUTTONMAP_XML_ATTR_FEATURE_NAME);
if (!name)
{
esyslog("<%s> tag has no \"%s\" attribute", BUTTONMAP_XML_ELEM_FEATURE, BUTTONMAP_XML_ATTR_FEATURE_NAME);
return false;
}
std::string strName(name);
const TiXmlElement* pUp = nullptr;
const TiXmlElement* pDown = nullptr;
const TiXmlElement* pRight = nullptr;
const TiXmlElement* pLeft = nullptr;
const TiXmlElement* pPositiveX = nullptr;
const TiXmlElement* pPositiveY = nullptr;
const TiXmlElement* pPositiveZ = nullptr;
// Determine the feature type
JOYSTICK_FEATURE_TYPE type;
ADDON::DriverPrimitive primitive;
if (DeserializePrimitive(pFeature, primitive, strName))
{
type = JOYSTICK_FEATURE_TYPE_SCALAR;
}
else
{
pUp = pFeature->FirstChildElement(BUTTONMAP_XML_ELEM_UP);
pDown = pFeature->FirstChildElement(BUTTONMAP_XML_ELEM_DOWN);
pRight = pFeature->FirstChildElement(BUTTONMAP_XML_ELEM_RIGHT);
pLeft = pFeature->FirstChildElement(BUTTONMAP_XML_ELEM_LEFT);
if (pUp || pDown || pRight || pLeft)
{
type = JOYSTICK_FEATURE_TYPE_ANALOG_STICK;
}
else
{
pPositiveX = pFeature->FirstChildElement(BUTTONMAP_XML_ELEM_POSITIVE_X);
pPositiveY = pFeature->FirstChildElement(BUTTONMAP_XML_ELEM_POSITIVE_Y);
pPositiveZ = pFeature->FirstChildElement(BUTTONMAP_XML_ELEM_POSITIVE_Z);
if (pPositiveX || pPositiveY || pPositiveZ)
{
type = JOYSTICK_FEATURE_TYPE_ACCELEROMETER;
}
else
{
esyslog("Feature \"%s\": <%s> tag is not a valid primitive", strName.c_str(), BUTTONMAP_XML_ELEM_FEATURE);
return false;
}
}
}
ADDON::JoystickFeature feature(strName, type);
// Deserialize according to type
switch (type)
{
case JOYSTICK_FEATURE_TYPE_SCALAR:
{
feature.SetPrimitive(primitive);
break;
}
case JOYSTICK_FEATURE_TYPE_ANALOG_STICK:
{
ADDON::DriverPrimitive up;
ADDON::DriverPrimitive down;
ADDON::DriverPrimitive right;
ADDON::DriverPrimitive left;
bool bSuccess = true;
if (pUp && !DeserializePrimitive(pUp, up, strName))
{
esyslog("Feature \"%s\": <%s> tag is not a valid primitive", strName.c_str(), BUTTONMAP_XML_ELEM_UP);
bSuccess = false;
}
if (pDown && !DeserializePrimitive(pDown, down, strName))
{
esyslog("Feature \"%s\": <%s> tag is not a valid primitive", strName.c_str(), BUTTONMAP_XML_ELEM_DOWN);
bSuccess = false;
}
if (pRight && !DeserializePrimitive(pRight, right, strName))
{
esyslog("Feature \"%s\": <%s> tag is not a valid primitive", strName.c_str(), BUTTONMAP_XML_ELEM_RIGHT);
bSuccess = false;
}
if (pLeft && !DeserializePrimitive(pLeft, left, strName))
{
esyslog("Feature \"%s\": <%s> tag is not a valid primitive", strName.c_str(), BUTTONMAP_XML_ELEM_LEFT);
bSuccess = false;
}
if (!bSuccess)
return false;
feature.SetUp(up);
feature.SetDown(down);
feature.SetRight(right);
feature.SetLeft(left);
break;
}
case JOYSTICK_FEATURE_TYPE_ACCELEROMETER:
{
ADDON::DriverPrimitive positiveX;
ADDON::DriverPrimitive positiveY;
ADDON::DriverPrimitive positiveZ;
bool bSuccess = true;
if (pPositiveX && !DeserializePrimitive(pPositiveX, positiveY, strName))
{
esyslog("Feature \"%s\": <%s> tag is not a valid primitive", strName.c_str(), BUTTONMAP_XML_ELEM_POSITIVE_X);
bSuccess = false;
}
if (pPositiveY && !DeserializePrimitive(pPositiveY, positiveY, strName))
{
esyslog("Feature \"%s\": <%s> tag is not a valid primitive", strName.c_str(), BUTTONMAP_XML_ELEM_POSITIVE_Y);
bSuccess = false;
}
if (pPositiveZ && !DeserializePrimitive(pPositiveZ, positiveZ, strName))
{
esyslog("Feature \"%s\": <%s> tag is not a valid primitive", strName.c_str(), BUTTONMAP_XML_ELEM_POSITIVE_Z);
bSuccess = false;
}
if (!bSuccess)
return false;
feature.SetPositiveX(positiveX);
feature.SetPositiveY(positiveY);
feature.SetPositiveZ(positiveZ);
break;
}
default:
break;
}
features.push_back(feature);
pFeature = pFeature->NextSiblingElement(BUTTONMAP_XML_ELEM_FEATURE);
}
return true;
}
void ClassifySvmSharedCommand::processSharedAndDesignData(vector<SharedRAbundVector*> lookup) {
try {
OutputFilter outputFilter(verbosity);
LabeledObservationVector labeledObservationVector;
FeatureVector featureVector;
readSharedRAbundVectors(lookup, designMap, labeledObservationVector, featureVector);
// optionally remove features with low standard deviation
if ( stdthreshold > 0.0 ) {
FeatureVector removedFeatureVector = applyStdThreshold(stdthreshold, labeledObservationVector, featureVector);
if (removedFeatureVector.size() > 0) {
std::cout << removedFeatureVector.size() << " OTUs were below the stdthreshold of " << stdthreshold << " and were removed" << std::endl;
if ( outputFilter.debug() ) {
std::cout << "the following OTUs were below the standard deviation threshold of " << stdthreshold << std::endl;
for (FeatureVector::iterator i = removedFeatureVector.begin(); i != removedFeatureVector.end(); i++) {
std::cout << " " << i->getFeatureLabel() << std::endl;
}
}
}
}
// apply [0,1] standardization
if ( transformName == "zeroone") {
std::cout << "transforming data to lie within range [0,1]" << std::endl;
transformZeroOne(labeledObservationVector);
}
else {
std::cout << "transforming data to have zero mean and unit variance" << std::endl;
transformZeroMeanUnitVariance(labeledObservationVector);
}
SvmDataset svmDataset(labeledObservationVector, featureVector);
OneVsOneMultiClassSvmTrainer trainer(svmDataset, evaluationFoldCount, trainingFoldCount, *this, outputFilter);
if ( mode == "rfe" ) {
SvmRfe svmRfe;
ParameterRange& linearKernelConstantRange = kernelParameterRangeMap["linear"]["constant"];
ParameterRange& linearKernelSmoCRange = kernelParameterRangeMap["linear"]["smoc"];
RankedFeatureList rankedFeatureList = svmRfe.getOrderedFeatureList(svmDataset, trainer, linearKernelConstantRange, linearKernelSmoCRange);
map<string, string> variables;
variables["[filename]"] = outputDir + m->getRootName(m->getSimpleName(sharedfile));
variables["[distance]"] = lookup[0]->getLabel();
string filename = getOutputFileName("summary", variables);
outputNames.push_back(filename);
outputTypes["summary"].push_back(filename);
m->mothurOutEndLine();
std::ofstream outputFile(filename.c_str());
int n = 0;
int rfeRoundCount = rankedFeatureList.front().getRank();
std::cout << "ordered features:" << std::endl;
std::cout << setw(5) << "index"
<< setw(12) << "OTU"
<< setw(5) << "rank"
<< std::endl;
outputFile << setw(5) << "index"
<< setw(12) << "OTU"
<< setw(5) << "rank"
<< std::endl;
for (RankedFeatureList::iterator i = rankedFeatureList.begin(); i != rankedFeatureList.end(); i++) {
n++;
int rank = rfeRoundCount - i->getRank() + 1;
outputFile << setw(5) << n
<< setw(12) << i->getFeature().getFeatureLabel()
<< setw(5) << rank
<< std::endl;
if ( n <= 20 ) {
std::cout << setw(5) << n
<< setw(12) << i->getFeature().getFeatureLabel()
<< setw(5) << rank
<< std::endl;
}
}
outputFile.close();
}
else {
MultiClassSVM* mcsvm = trainer.train(kernelParameterRangeMap);
map<string, string> variables;
variables["[filename]"] = outputDir + m->getRootName(m->getSimpleName(sharedfile));
variables["[distance]"] = lookup[0]->getLabel();
string filename = getOutputFileName("summary", variables);
outputNames.push_back(filename);
outputTypes["summary"].push_back(filename);
m->mothurOutEndLine();
std::ofstream outputFile(filename.c_str());
printPerformanceSummary(mcsvm, std::cout);
printPerformanceSummary(mcsvm, outputFile);
outputFile << "actual predicted" << std::endl;
for ( LabeledObservationVector::const_iterator i = labeledObservationVector.begin(); i != labeledObservationVector.end(); i++ ) {
Label actualLabel = i->getLabel();
outputFile << i->getDatasetIndex() << " " << actualLabel << " ";
try {
Label predictedLabel = mcsvm->classify(*(i->getObservation()));
outputFile << predictedLabel << std::endl;
}
catch ( MultiClassSvmClassificationTie& m ) {
outputFile << "tie" << std::endl;
std::cout << "classification tie for observation " << i->datasetIndex << " with label " << i->first << std::endl;
}
}
outputFile.close();
delete mcsvm;
}
}
catch (exception& e) {
m->errorOut(e, "ClassifySvmSharedCommand", "processSharedAndDesignData");
exit(1);
}
}