void Backpropagation::trainOnlineCV(Mlp& network,
MATRIX& trainingInputs,
VECTOR& trainingTargets,
MATRIX& testInputs,
VECTOR& testTargets)
{
VECTOR trainingOutputs(trainingTargets.size(), 0.0);
VECTOR testOutputs(testTargets.size(), 0.0);
while(error > tolerance && testCount < maxTestCount)
{
VECTOR::iterator output = trainingOutputs.begin();
VECTOR::iterator target = trainingTargets.begin();
for(MATRIX::iterator input = trainingInputs.begin();
input != trainingInputs.end();
++input, ++target, ++output)
{
*output = network(*input);
double err = *output - *target;
getWeightUpdates(network, *input, err);
applyWeightUpdates(network);
++iteration;
if(iteration >= maxIterations)
break;
}
++epoch;
error = mse(trainingTargets, trainingOutputs);
// Early-stopping using test (cross-validation) error
testOutputs = network(testInputs);
testError = mse(testTargets, testOutputs);
if(testError < minTestError)
{
// Preserve test error and network weights
minTestError = testError;
W = network.W;
V = network.V;
biasW = network.biasW;
biasV = network.biasV;
testCount = 0;
}
else
{
++testCount;
}
}
network.W = W;
network.V = V;
network.biasW = biasW;
network.biasV = biasV;
testError = minTestError;
}
void AutoSubClassingClassifier::train(
const vector<Mat>& segmentedHands,
const vector<int>& expectedClass) {
assert(segmentedHands.size() == expectedClass.size());
this->subclassesOfTrainingBase.clear();
int numberOfClasses = *max_element(expectedClass.begin(), expectedClass.end()) + 1;
vector<Mat> samplesPerClass(numberOfClasses);
for (int i = 0; i < segmentedHands.size(); i++) {
Mat currentCaracteristicVector = this->caracteristicVector(segmentedHands[i]);
if (samplesPerClass[expectedClass[i]].empty()) {
samplesPerClass[expectedClass[i]] = currentCaracteristicVector;
} else {
vconcat(samplesPerClass[expectedClass[i]], currentCaracteristicVector, samplesPerClass[expectedClass[i]]);
}
}
int accumulatedNumberOfClasses = 0;
for (int i = 0; i < samplesPerClass.size(); i++) {
Mat labels;
int clusterCount = this->numberOfSubclasses[i];
kmeans(samplesPerClass[i], clusterCount, labels, TermCriteria(CV_TERMCRIT_ITER|CV_TERMCRIT_EPS, 0.0001, 10000), 5, KMEANS_RANDOM_CENTERS);
this->subclassesOfTrainingBase.push_back(labels + (Mat::ones(samplesPerClass[i].rows, 1, CV_32S) * accumulatedNumberOfClasses));
accumulatedNumberOfClasses += this->numberOfSubclasses[i];
}
Mat trainingInputs(segmentedHands.size(), this->caracteristicVectorLength(), CV_32F);
Mat trainingOutputs(segmentedHands.size(), 1, CV_32F);
int firstIndex = 0;
for (int i = 0; i < samplesPerClass.size(); i++) {
int lastIndex = firstIndex + samplesPerClass[i].rows;
Mat samples = samplesPerClass[i];
Mat range = trainingInputs.rowRange(firstIndex, lastIndex);
samples.copyTo(range);
Mat outputs = this->subclassesOfTrainingBase[i];
outputs.copyTo(trainingOutputs.rowRange(firstIndex, lastIndex));
firstIndex = lastIndex;
}
this->statisticalModel->train(trainingInputs, trainingOutputs);
}
