void ClassifierMLP::Train(std::vector< fvec > samples, ivec labels)
{
u32 sampleCnt = samples.size();
if(!sampleCnt) return;
DEL(mlp);
dim = samples[0].size();
CvMat *layers;
// if(neuronCount == 3) neuronCount = 2; // don't ask me why but 3 neurons mess up everything...
if(!layerCount || neuronCount < 2)
{
layers = cvCreateMat(2,1,CV_32SC1);
cvSet1D(layers, 0, cvScalar(dim));
cvSet1D(layers, 1, cvScalar(1));
}
else
{
layers = cvCreateMat(2+layerCount,1,CV_32SC1);
cvSet1D(layers, 0, cvScalar(dim));
cvSet1D(layers, layerCount+1, cvScalar(1));
FOR(i, layerCount) cvSet1D(layers, i+1, cvScalar(neuronCount));
}
u32 *perm = randPerm(sampleCnt);
CvMat *trainSamples = cvCreateMat(sampleCnt, dim, CV_32FC1);
CvMat *trainLabels = cvCreateMat(labels.size(), 1, CV_32FC1);
CvMat *sampleWeights = cvCreateMat(samples.size(), 1, CV_32FC1);
FOR(i, sampleCnt)
{
FOR(d, dim) cvSetReal2D(trainSamples, i, d, samples[perm[i]][d]);
cvSet1D(trainLabels, i, cvScalar(labels[perm[i]]));
cvSet1D(sampleWeights, i, cvScalar(1));
}
void
Cas1DVanishingPoint::findInteriorVanishingPt(const std::vector<cv::Vec4i> & lines, float & intVPTheta, float & intVPRho) const
{
float p = 0.995f;
float r = 2.0f / lines.size();
float k = log(1.0f - p) / log(1.0f - r * r);
size_t max_iter = 1000;
size_t it = 0;
size_t max_inliers = 2;
float best_theta, best_rho;
while (it < k && it < max_iter)
{
std::vector<size_t> perm = randPerm(lines.size());
std::vector<cv::Vec4i> samples;
samples.push_back(lines[perm[0]]);
samples.push_back(lines[perm[1]]);
cv::Point2f guess = convergeLines(samples);
float th = atan2(guess.y, guess.x);
float rh = hypot(guess.x, guess.y);
if (rh > mInteriorRadius) continue;
size_t inliers = linesSupport(th, rh, lines).size();
// std::cout << cv::Mat(guess) << inliers << std::endl;
// showLines(linesSupport(guess));
if (inliers > max_inliers)
{
max_inliers = inliers;
best_theta = th;
best_rho = rh;
r = (float)max_inliers / lines.size();
k = log(1.0f - p) / log(1.0f - r * r);
it = 0;
}
// std::cout << k << " " << it << std::endl;
it++;
}
intVPTheta = best_theta;
intVPRho = best_rho;
}
void FilterBoostLearner::filter( InputData* pData, int size, bool rejection )
{
pData->clearIndexSet();
const int numExamples = pData->getNumExamples();
const int numClasses = pData->getNumClasses();
set<int> indexSet;
if (size<numExamples) // not whole dataset will be used
{
//random permutation
vector< pair<int,int> > tmpRandomArr( numExamples );
for( int i=0; i < numExamples; i++ )
{
tmpRandomArr[i].first = rand();
tmpRandomArr[i].second = i;
}
sort( tmpRandomArr.begin(), tmpRandomArr.end(), nor_utils::comparePair<1, int, int, less<int> >() );
vector< int > randPerm( numExamples );
for( int i=0; i<numExamples; i++ )
{
randPerm[i] = tmpRandomArr[i].second;
}
//end: random permutation
int iter = 0;
int maxIter = 5 * size;
int wholeIter = 0;
indexSet.clear();
while (1)
{
if ( size<=indexSet.size() ) break;
if ( wholeIter > 5 ) rejection = false;
if ( numExamples <= iter ) {
iter = 0;
wholeIter++;
}
if ( rejection )
{
const vector<Label>& labels = pData->getLabels( randPerm[iter] );
vector<Label>::const_iterator lIt;
AlphaReal scalar = 0.0;
//float scalar = numeric_limits<float>::max();
for ( lIt = labels.begin(); lIt != labels.end(); ++lIt )
{
//if ( scalar > _margins[ randPerm[iter] ][lIt->idx] ) scalar = _margins[ randPerm[iter] ][lIt->idx];
//if ( _margins[ randPerm[iter] ][lIt->idx] < 0.0 ) scalar += _margins[ randPerm[iter] ][lIt->idx];
scalar += (1 / ( 1 + exp(_margins[ randPerm[iter] ][lIt->idx])));
}
AlphaReal qValue = scalar / (AlphaReal) numClasses;
//AlphaReal qValue = 1 / ( 1 + exp( scalar ) );
AlphaReal randNum = (AlphaReal)rand() / RAND_MAX;
if ( randNum < qValue ) indexSet.insert( randPerm[iter] );
}
else
{
indexSet.insert( randPerm[iter] );
}
iter++;
}
// normalize the weights of the labels
set<int>::iterator sIt;
AlphaReal sum = 0.0;
// for each example are in use
for ( sIt = indexSet.begin(); sIt != indexSet.end(); sIt++ )
{
vector<Label>& labels = pData->getLabels(*sIt);
vector<Label>::iterator lIt;
for (lIt = labels.begin(); lIt != labels.end(); ++lIt )
{
lIt->weight = 1 /( 1+exp( _margins[ *sIt ][lIt->idx] ) );
sum += lIt->weight;
}
}
for ( sIt = indexSet.begin(); sIt != indexSet.end(); sIt++ )
{
vector<Label>& labels = pData->getLabels(*sIt);
vector<Label>::iterator lIt;
for (lIt = labels.begin(); lIt != labels.end(); ++lIt )
{
lIt->weight /= sum;
}
}
pData->loadIndexSet( indexSet );
} else { // the whole dataset is used
AlphaReal sum = 0.0;
// for each example are in use
for ( int i = 0; i < pData->getNumExamples(); ++i )
{
vector<Label>& labels = pData->getLabels(i);
vector<Label>::iterator lIt;
for (lIt = labels.begin(); lIt != labels.end(); ++lIt )
{
lIt->weight = 1 /( 1+exp( _margins[ i ][lIt->idx] ) );
sum += lIt->weight;
}
}
for ( int i = 0; i < pData->getNumExamples(); ++i )
{
vector<Label>& labels = pData->getLabels( i );
vector<Label>::iterator lIt;
for (lIt = labels.begin(); lIt != labels.end(); ++lIt )
{
lIt->weight /= sum;
}
}
}
/*
sum = 0.0;
for ( int i=0; i < pData->getNumExamples(); i++ )
{
vector<Label>& labels = pData->getLabels(i);
vector<Label>::iterator lIt;
for (lIt = labels.begin(); lIt != labels.end(); ++lIt )
{
sum += lIt->weight;
}
}
cout << "The size of the dataset: " << pData->getNumExamples() << endl;
cout << "Sum: " << sum << endl;
*/
}