void WeakDiscreteTreeLearner::sortIndexesBasedOnDataPositions(indices_t &positions,
const size_t featureIndex) const
{
const FeaturesResponses &featuresResponses = _trainingData->getFeatureResponses();
std::sort(positions.begin(), positions.end(), comparator(featuresResponses, featureIndex));
return;
}
/* compute the information needed to update the current context stats. */
void SkipCTS::getContextInfo(zobhash_t &hash, const indices_t &idxs) const {
// compute the hash
for (int k=0; k < idxs.size(); k++) {
size_t x = idxs[k];
hash ^= s_zobtbl[x][m_context[x]];
}
}
void tuner_t::map(const indices_t& indices, strings_t& values) const
{
assert(indices.size() == m_params.size());
assert(values.size() == m_params.size());
for (size_t i = 0; i < m_params.size(); ++ i)
{
const auto& param = m_params[i];
assert(indices[i] < param.m_values.size());
values[i] = param.m_values[indices[i]];
}
}
bool init_index(const vector<string> &index_dirs, indices_t &indices)
{
Logger logger = Logger::getInstance("main");
for (vector<string>::const_iterator i=index_dirs.begin(); i!=index_dirs.end(); ++i) {
if (!i->empty()) {
LOG4CPLUS_DEBUG(logger, "Opening index at " << *i);
argos::index::index_ptr_t idx=argos::index::index_ptr_t(new Index(i->c_str()));
if(!idx->is_open())
{
LOG4CPLUS_FATAL(logger, "Failed to open index " << *i);
return false;
}
indices.insert(indices_t::value_type(get_index_name(idx), idx));
}
}
return true;
}
int WeakDiscreteTreeLearner::findThreshold(
const weights_t &weights,
const indices_t & sortedIndices,
const size_t featureIndex,
double &errorMin, int &thresholdMin, int &alphaMin, int &splitIndex) const
{
const FeaturesResponses &featuresResponses = _trainingData->getFeatureResponses();
const FeaturesResponses::const_reference featureResponses = featuresResponses[featureIndex];
//getSum of data
double sumPos = 0;
double sumNeg = 0;
errorMin = std::numeric_limits<int>::max();
for (size_t i = 0; i < sortedIndices.size(); ++i)
{
const int index = sortedIndices[i];
if (_classes[index] == _negativeClass)
{
sumNeg += weights[index];
}
else
{
sumPos += weights[index];
}
}
if (_verbose > 3)
{
std::cout << "sumneg: " << sumNeg << " sumpos: " << sumPos << " both: " << sumPos + sumNeg << "\n";
}
if (sumNeg == 0)
{
thresholdMin = featureResponses[sortedIndices[0]] - 1;
alphaMin = -1;
errorMin = 0;
return -1;
}
if (sumPos == 0)
{
thresholdMin = featureResponses[sortedIndices[0]] - 1;
alphaMin = 1;
errorMin = 0;
return -1;
}
//positives left
double positiveLeftError = sumPos;
double negativeLeftError = 0;
//positives right
double negativeRightError = sumNeg;
double positiveRightError = 0;
int minIndex = -1;
// go over all sorted data elements
for (size_t i = 0; i < sortedIndices.size(); ++i)
{
const int index = sortedIndices[i];
const int threshold = featureResponses[index];
if (_classes[index] == _negativeClass)
{
negativeLeftError += weights[index];
negativeRightError -= weights[index];
}
else
{
positiveLeftError -= weights[index];
positiveRightError += weights[index];
}
double error = 0;
int alpha = 0;
if (positiveLeftError + negativeLeftError < positiveRightError + negativeRightError)
{
alpha = 1;
error = positiveLeftError + negativeLeftError;
}
else
{
alpha = -1;
error = positiveRightError + negativeRightError;
}
bool cond = false;
if (i < sortedIndices.size() - 1)
{
cond = (error < errorMin && threshold != featureResponses[sortedIndices[i+1]]);
}
else
{
cond = (error < errorMin) ;
}
if (cond)
{
errorMin = error;
thresholdMin = threshold;
alphaMin = alpha;
minIndex = i;
splitIndex = i + 1;
}
}// end of "for each sorted data element"
//set the threshold between suceeding values, except the last one(+1)
if (minIndex == int(sortedIndices.size() - 1))
{
thresholdMin = thresholdMin + 1;
}
else if (thresholdMin + 1 == featureResponses[sortedIndices[minIndex+1]])
{
thresholdMin = thresholdMin + 1;
}
else
{
thresholdMin = int((featureResponses[sortedIndices[minIndex]] + featureResponses[sortedIndices[minIndex+1]]) / 2.0);
}
if (_verbose > 3)
{
std::cout << "sorted data:\n";
for (size_t i = 0; i < sortedIndices.size(); ++i)
{
const int index = sortedIndices[i];
const int threshold = featureResponses[index];
std::cout << std::setw(6) << threshold << " ";
}
std::cout << "sorted classes:\n";
for (size_t i = 0; i < sortedIndices.size(); ++i)
{
int ind = sortedIndices[i];
std::cout << std::setprecision(2) << std::setw(6) << _classes[ind]*weights[ind] << " ";
}
std::cout << "threshold: " << thresholdMin << " alpha: " << alphaMin << " error: " << errorMin;
}
return 0;
}
int WeakDiscreteTreeLearner::createNode(
const weights_t &weights,
const indices_t &indices, const size_t start, const size_t end,
TreeNode::shared_ptr &node, double &minError,
const bool isLeft, const int root_node_bottom_height, const int root_node_left_width) const
{
TrainingData::point_t modelWindow = _trainingData->getModelWindow();
const int
shrinking_factor = bootstrapping::integral_channels_computer_t::get_shrinking_factor(),
modelWidth = modelWindow.x() / shrinking_factor ,
modelHeight = modelWindow.y() / shrinking_factor;
//for all features get responses on every image
//find feature with lowest error
const size_t numFeatures = _trainingData->getFeaturesPoolSize();
std::vector<std::pair<double, size_t> >
minErrorsForSearch(numFeatures, std::make_pair(std::numeric_limits<double>::max(), 0));
indices_t indicesCrop;
const int binsize = 1000;
if(start > end)
{
throw std::invalid_argument("WeakDiscreteTreeLearner::createNode received start > end but expected start <= end");
}
indicesCrop.resize(end - start);
std::copy(indices.begin() + start, indices.begin() + end, indicesCrop.begin());
if (indicesCrop.size() == 0)
{
return -1;
}
int return_value = 0;
//find max valid feature index
size_t max_valid_feature_index= 0;
for (size_t featureIndex = numFeatures-1; featureIndex >=0; --featureIndex)
{
if (_trainingData->getFeatureValidity(featureIndex))
{
max_valid_feature_index = featureIndex+1;
break;
}
}
//biasing features not yet supported
double _pushBias = 0;
#pragma omp parallel for reduction(+:return_value) schedule(guided)
for (size_t featureIndex = 0; featureIndex < max_valid_feature_index; ++featureIndex)
{
if (_trainingData->getFeatureValidity(featureIndex)){
const int minv = (*_mins)[featureIndex], maxv = (*_maxs)[featureIndex];
double error = std::numeric_limits<double>::max();
return_value += getErrorEstimate(weights, indicesCrop, featureIndex, binsize, minv, maxv, error);
minErrorsForSearch[featureIndex] = std::make_pair(error, featureIndex);
}
} // end of "for each feature"
//if ( *(std::min_element(minErrorsForSearch.begin(),minErrorsForSearch.end(), sort_pair())) >2)//== mm)
if (return_value < 0)
{
//create dummy node and return;
//node = TreeNode::shared_ptr(new TreeNode(minthr, alphamin, _features[minFeat], minFeat, indicesCrop, splitIndexMin ));
return -1;
}
const int deltaH=0;
//get kth minimal elements
//std::sort(minErrorsForSearch.begin(), minErrorsForSearch.end(), sort_pair());
//thrust::sort(thrust::reinterpret_tag<thrust::omp::tag>(minErrorsForSearch.begin()),
// thrust::reinterpret_tag<thrust::omp::tag>(minErrorsForSearch.end()),
// sort_pair());
std::sort(minErrorsForSearch.begin(),
minErrorsForSearch.end(),
sort_pair());
//search the feature that is highest up in the image
int miny = std::numeric_limits<int>::max();
int minx = std::numeric_limits<int>::max();
int minFeatureIndex = -1;
double errorth;
if (minErrorsForSearch.size() > 0){
errorth= std::min(0.49999999999, minErrorsForSearch[0].first* (1.0+_pushBias));
}
else{
throw std::runtime_error("minErrorsForSearch has size 0: no features in pool?");
}
for (size_t i = 0; i< minErrorsForSearch.size(); ++i){
const int this_feat_idx = minErrorsForSearch[i].second;
double error= minErrorsForSearch[i].first;
if (_pushBias ==0){
minFeatureIndex = this_feat_idx;
break;
}
if (error > errorth)
break;
const Feature this_feat = _trainingData->getFeature(this_feat_idx);
int y = this_feat.y + this_feat.height;
int x = this_feat.x + this_feat.width;
}
if (minFeatureIndex == -1)
{
//create dummy node and return;
//node = TreeNode::shared_ptr(new TreeNode(minthr, alphamin, _features[minFeat], minFeat, indicesCrop, splitIndexMin ));
return -1;
}
sortIndexesBasedOnDataPositions(indicesCrop, minFeatureIndex);
int splitIndexMin = -1;
int minThreshold = -1;
int alphaMin = -1;
if (findThreshold(weights, indicesCrop, minFeatureIndex, minError, minThreshold,
alphaMin, splitIndexMin) == -1)
{
return -1;
}
if(splitIndexMin < 0)
{
throw std::runtime_error("WeakDiscreteTreeLearner::createNode, findThreshold failed to find an adequate split index");
}
// set the shared_ptr to point towards a new node instance
node.reset(new TreeNode(minThreshold, alphaMin,
_trainingData->getFeature(minFeatureIndex),
minFeatureIndex, indicesCrop, splitIndexMin, isLeft));
return 0;
}
inline
int WeakDiscreteTreeLearner::getErrorEstimate(
const weights_t &weights,
const indices_t& indices, const size_t featureIndex,
int num_bins , const int minv, const int maxv, double &error) const
{
if ((maxv - minv) < num_bins)
{
num_bins = maxv - minv;
}
//std::cout << "binsize: " << binsize << std::endl;
std::vector<double> bin_pos(num_bins + 1, 0), bin_neg(num_bins + 1, 0);
double cumNeg = 0;
double cumPos = 0;
error = std::numeric_limits<double>::max();
const FeaturesResponses &featuresResponses = _trainingData->getFeatureResponses();
for (size_t i = 0; i < indices.size(); ++i)
{
const size_t trainingSampleIndex = indices[i];
//FIXME CHECK this
//if (weights[trainingSampleIndex]< 1E-12)
// continue;
const int featureResponse = featuresResponses[featureIndex][trainingSampleIndex];
const int bin = int(num_bins / double(maxv - minv) * (featureResponse - minv));
//int splitpos = minv + (bin/(double)binsize) *(maxv-minv);
//int bin2= binsize/maxv*((*_featureResp)[pos+ind[i]] - minv);
//assert(bin==bin2);
const int the_class = _classes[indices[i]];
if (the_class == _negativeClass)
{
bin_neg[bin] += weights[indices[i]];
cumNeg += weights[indices[i]];
}
else
{
bin_pos[bin] += weights[indices[i]];
cumPos += weights[indices[i]];
}
}
//run test by setting this to return 0 with error 0
if (cumPos == 0 || cumNeg == 0)
{
return -1;
}
double
//positives left
positivesLeftError = cumPos,
negativesLeftError = 0,
//positives right
negativesRightError = cumNeg,
positivesRightError = 0;
//int minbin = -1;
for (int i = 0; i < num_bins; ++i)
{
positivesLeftError -= bin_pos[i];
negativesLeftError += bin_neg[i];
positivesRightError += bin_pos[i];
negativesRightError -= bin_neg[i];
double binError = 0;
if (positivesLeftError + negativesLeftError < positivesRightError + negativesRightError)
{
binError = positivesLeftError + negativesLeftError;
}
else
{
binError = positivesRightError + negativesRightError;
}
// we keep the min error
if (binError < error)
{
//minbin = i;
error = binError;
}
}
return 0;
}
