float CCMat::getAccuracy(void) const
{
float res = 0;
Mat cMat = getConfusionMatrix();
for (byte s = 0; s < m_pConfusionMatrix->getNumStates(); s++)
res += cMat.at<float>(s, s);
return res;
}
int annTrain::train(std::string imagesDir, int networkInputSize, float testRatio)
{
std::cout << "Reading training set..." << std::endl;
uint64 start = ofGetElapsedTimeMillis();
std::vector<std::string> files = getFilesInDirectory(imagesDir);
std::random_shuffle(files.begin(), files.end());
cv::Mat img;
for (auto it = files.begin(); it != files.end(); ++it)
{
std::string filename = *it;
//std::cout << "Reading image " << filename << "..." << std::endl;
img = cv::imread(filename, 0);
if (img.empty())
{
std::cerr << "WARNING: Could not read image." << std::endl;
continue;
}
std::string classname = getClassName(filename);
cv::Mat descriptors = getDescriptors(img);
processClassAndDesc(classname, descriptors);
}
std::cout << " Seconds : " << (ofGetElapsedTimeMillis() - start) / 1000.0 << std::endl;
std::cout << "Creating vocabulary..." << std::endl;
start = ofGetElapsedTimeMillis();
cv::Mat labels;
cv::Mat vocabulary;
// Use k-means to find k centroids (the words of our vocabulary)
cv::kmeans(descriptorsSet, networkInputSize, labels, cv::TermCriteria(cv::TermCriteria::EPS + cv::TermCriteria::MAX_ITER, 10, 0.01), 1, cv::KMEANS_PP_CENTERS, vocabulary);
// No need to keep it on memory anymore
descriptorsSet.release();
std::cout << " Seconds : " << (ofGetElapsedTimeMillis() - start) / 1000.0 << std::endl;
// Convert a set of local features for each image in a single descriptors
// using the bag of words technique
std::cout << "Getting histograms of visual words..." << std::endl;
int* ptrLabels = (int*)(labels.data);
int size = labels.rows * labels.cols;
for (int i = 0; i < size; i++)
{
int label = *ptrLabels++;
ImageData* data = descriptorsMetadata[i];
data->bowFeatures.at<float>(label)++;
}
// Filling matrixes to be used by the neural network
std::cout << "Preparing neural network..." << std::endl;
std::set<ImageData*> uniqueMetadata(descriptorsMetadata.begin(), descriptorsMetadata.end());
for (auto it = uniqueMetadata.begin(); it != uniqueMetadata.end(); )
{
ImageData* data = *it;
cv::Mat normalizedHist;
cv::normalize(data->bowFeatures, normalizedHist, 0, data->bowFeatures.rows, cv::NORM_MINMAX, -1, cv::Mat());
trainSamples.push_back(normalizedHist);
trainResponses.push_back(getClassCode(classes, data->classname));
delete *it; // clear memory
it++;
}
descriptorsMetadata.clear();
// Training neural network
std::cout << "Training neural network..." << std::endl;
start = ofGetElapsedTimeMillis();
mlp = getTrainedNeuralNetwork(trainSamples, trainResponses);
std::cout << " Seconds : " << (ofGetElapsedTimeMillis() - start) / 1000.0 << std::endl;
// We can clear memory now
trainSamples.release();
trainResponses.release();
// Train FLANN
std::cout << "Training FLANN..." << std::endl;
start = ofGetElapsedTimeMillis();
flann = cv::Ptr<cv::FlannBasedMatcher>(new cv::FlannBasedMatcher());
flann->add(vocabulary);
flann->train();
std::cout << " Seconds : " << (ofGetElapsedTimeMillis() - start) / 1000.0 << std::endl;
// Reading test set
std::cout << "Reading test set..." << std::endl;
start = ofGetElapsedTimeMillis();
readImagesToTest(files.begin() + (size_t)(files.size() * testRatio), files.end());
std::cout << " Seconds : " << (ofGetElapsedTimeMillis() - start) / 1000.0 << std::endl;
// Get confusion matrix of the test set
std::vector<std::vector<int> > confusionMatrix = getConfusionMatrix();
// how accurate is our model
std::cout << "Confusion matrix " << std::endl;
printConfusionMatrix(confusionMatrix, classes);
std::cout << "Accuracy " << getAccuracy(confusionMatrix) << std::endl;
// now save everything
std::cout << "saving models" << std::endl;
saveModels(vocabulary, classes);
return 0;
}
