void GenSOM::train(const multi_img &input, ProgressObserver *po)
{
std::cout << "Start feeding" << std::endl;
// matrices hold shuffled sequences of the input for number of iterations
int maxIter = config.maxIter;
cv::Mat_<int> shuffledY(1, maxIter);
cv::Mat_<int> shuffledX(1, maxIter);
// generate random sequence of the input x,y range
cv::RNG rng(config.seed);
rng.fill(shuffledY, cv::RNG::UNIFORM,
cv::Scalar(0), cv::Scalar(input.height));
rng.fill(shuffledX, cv::RNG::UNIFORM,
cv::Scalar(0), cv::Scalar(input.width));
// output percentage
unsigned int hundred = std::max<unsigned int>(maxIter/100, 100);
int percent = 1;
if (config.verbosity > 0) {
std::cout << " 0 %";
std::cout.flush();
}
long sumOfUpdates = 0;
// starting training (notifier needed by OpenCL impl.)
notifyTrainingStart();
cv::MatConstIterator_<int> itY = shuffledY.begin();
cv::MatConstIterator_<int> itX = shuffledX.begin();
for (int curIter = 0; curIter < maxIter; ++curIter, ++itX, ++itY)
{
// feed one sample
const multi_img::Pixel &vec = input(*itY, *itX);
sumOfUpdates += trainSingle(vec, curIter, maxIter);
// print progress (and maybe exit)
if ((config.verbosity > 0 || po) && (config.maxIter > 100)
&& ((curIter % hundred) == 0) && (percent < 100)) {
// send progress updates to observer or stdout
if (po) {
bool cont = po->update(curIter / (float)config.maxIter);
if (!cont) {
std::cerr << "Aborting training" << std::endl;
return;
}
} else {
std::cout << "\r " << (percent < 10 ? " " : "")
<< percent << " %";
std::cout.flush();
}
if (config.verbosity >= 2) {
std::cout << " Feed #" << curIter
<< ", avg. updates per iteration in the last "
<< (maxIter / hundred) << " iterations: "
<< ((double)sumOfUpdates / hundred)
<< std::endl;
sumOfUpdates = 0;
}
// print som each 20%
if (config.verbosity >= 3 && (percent % 20) == 0)
{
// TODO
/*multi_img somimg = som->export_2d();
std::ostringstream filename, filename2;
filename << "debug_som_";
filename << std::setfill('0') << std::setw(2);
filename << (ctr / (config.maxIter / 20));
somimg.write_out(filename.str());
filename2 << filename.str();
filename << "-rgb.png";
cv::imwrite(filename.str(), somimg.bgr() * 255.f);
if (!updateMap.empty()) {
cv::Mat3f upBGR(updateMap.rows, updateMap.cols);
cv::cvtColor(updateMap, upBGR, CV_GRAY2BGR);
for (cv::Mat3f::iterator it = upBGR.begin(); it != upBGR.end();++it)
if (*it == cv::Vec3f(0.f))
*it = cv::Vec3f(0.f, 0.f, 1.f);
filename2 << "-updates.png";
cv::imwrite(filename2.str(), upBGR * 255.f);
}*/
}
percent++;
}
}
if (config.verbosity > 0)
std::cout << "\r100 %" <<std::endl;
// finished training (notifier needed by OpenCL impl.)
notifyTrainingEnd();
std::cout <<"# Feeding done" <<std::endl;
if (!config.somFile.empty()) {
std::cout << "# writing SOM in binary format to \""
<< config.somFile << "\"" << std::endl;
saveFile(config.somFile);
}
if (config.verbosity > 2) {
cv::imwrite("debug_som.png", bgr(input.meta, input.maxval)*255.f);
}
}
void ViTrainer::trainSingle(ViNeuralNetwork *network)
{
setNetwork(network);
trainSingle();
}