bool NeuralDetector::Train(std::vector<Image_t> images)
{
size_t images_count = images.size();
cv::Mat trainData(images_count, ROWS * COLS, CV_32FC1);
cv::Mat trainOutput(images_count, 1, CV_32FC1);
for (size_t i = 0; i < images_count; ++i) {
(images[i].image.reshape(0, 1)).copyTo(trainData.row(i));
if (images[i].info == SI_UNDEF) {
return false;
}
trainOutput.at<float>(i, 0) = (images[i].info == SI_EXBNZ ? 1 : -1);
}
machineBrain.train(
trainData,
trainOutput,
cv::Mat(),
cv::Mat(),
CvANN_MLP_TrainParams(
cvTermCriteria(
CV_TERMCRIT_ITER | CV_TERMCRIT_EPS,
MAX_ITER,
MLP_EPSILON
),
CvANN_MLP_TrainParams::BACKPROP,
PARAM1,
PARAM2
)
);
machineBrain.save("savednet.txt");
std::cout << "Saved\n";
return true;
}
Impl(const std::string& learnpath, const std::string& allchars)
{
std::set<char> goodChars(allchars.begin(), allchars.end());
std::ifstream train((learnpath + "/train.txt").c_str());
std::vector< std::pair<char, std::string> > samples;
char symbol;
std::string imageFile;
while (train >> symbol >> imageFile) {
if (goodChars.find(symbol) == goodChars.end()) continue;
samples.push_back(std::make_pair(symbol, imageFile));
}
cv::Mat trainData(samples.size(), FEATURE_COUNT, cv::DataType<float>::type);
cv::Mat trainClasses(samples.size(), 1, cv::DataType<float>::type);
for (size_t i = 0; i < samples.size(); ++i) {
std::string path = learnpath + samples[i].second;
cv::Mat inp = cv::imread(path, 0), out, canny;
if (inp.empty()) continue;
cv::Canny(inp, canny, 100, 50, 3);
std::vector< std::vector<cv::Point> > contours;
cv::findContours(canny, contours, CV_RETR_LIST, CV_CHAIN_APPROX_SIMPLE, cv::Point(0, 0));
int maxx = -1, maxy = -1, minx = 1e9, miny = 1e9;
for (size_t j = 0; j < contours.size(); ++j) {
cv::Rect r = cv::boundingRect(contours[j]);
if (r.x + r.width > maxx) maxx = r.x + r.width;
if (r.y + r.height > maxy) maxy = r.y + r.height;
if (r.x < minx) minx = r.x;
if (r.y < miny) miny = r.y;
}
cv::Rect bound(minx, miny, maxx - minx, maxy - miny);
cv::resize(cv::Mat(inp, bound), out, cv::Size(10, 16), 0, 0, cv::INTER_CUBIC);
for (int j = 0; j < FEATURE_COUNT; ++j) trainData.at<float>(i, j) = out.data[j];
trainClasses.at<float>(i, 0) = samples[i].first;
}
oracle.train(trainData, trainClasses);
}
Mat extractFeature(vector<string> imgName)
{
Mat trainData(0,1764, CV_32FC1);
for (int i=0; i != imgName.size(); i++) {
Mat img=imread(imgName[i].c_str(),1);
resize(img, img, Size(64,64));
HOGDescriptor hog=HOGDescriptor(Size(64,64), Size(16,16), Size(8,8), Size(8,8), 9);
vector<float> descriptors;
hog.compute(img, descriptors);
Mat descriptorsMat(1,descriptors.size(),CV_32F);
for (int j=0; j<descriptors.size(); j++) {
descriptorsMat.at<float>(0, j)=descriptors[j];
}
trainData.push_back(descriptorsMat);
}
return trainData;
}
int main()
{
// init variables
double error = 0.;
int truecnt = 0;
int times,timed;
// print useful info for reference
std::cout << "\n" << "hidden neurons: " << "\t \t" << HIDDEN << std::endl;
// init random number generator
srand((int)time(NULL));
// create network
std::cout << "initializing network..." << "\t \t";
NeuralNet DigitNet;
NeuralLayer * pHiddenLayer1 = new NeuralTanhLayer(INPUT,HIDDEN);
DigitNet.addLayer( pHiddenLayer1 );
NeuralLayer * pOutputLayer = new NeuralSoftmaxLayer(HIDDEN,OUTPUT);
DigitNet.addLayer( pOutputLayer );
// set output type:
// SCALAR = tanh or sigmoid output layer (use one output neuron)
// PROB = softmax output layer, 1-of-N output encoding (use two output neurons)
const unsigned int outType = PROB;
// set learning rate, momentum, decay rate
const double learningRate = 0.15;
const double momentum = 0.0;
const double decayRate = 0.0;
DigitNet.setParams(learningRate,momentum,decayRate,outType);
std::cout << "done" << std::endl;
// load training and test data
std::cout << "loading data..." << "\t \t \t";
std::vector< std::vector<double> > bigData( DATA_SIZE,std::vector<double>(INPUT+1,0.0) );
loadFromFile(bigData,"train.txt");
std::vector< std::vector<double> > trainData( TRAIN_SIZE,std::vector<double>(INPUT+1,0.0) );
std::vector< std::vector<double> > testData( TEST_SIZE,std::vector<double>(INPUT+1,0.0) );
buildData(bigData,trainData,TRAIN_SIZE,testData,TEST_SIZE);
std::cout << "done" << std::endl;
// loop over training data points and train net
// slice off first column of each row (example)
times=(int)time(NULL); // init time counter
std::cout << "\n" << "training examples: " << "\t \t" << TRAIN_SIZE << std::endl;
std::cout << "learning rate: " << "\t \t \t" << learningRate << std::endl;
std::cout << "momentum: " << "\t \t \t" << momentum << std::endl;
std::cout << "weight decay: " << "\t \t \t" << decayRate << std::endl;
std::cout << "training network..." << "\t \t";
for(int i=0;i<TRAIN_SIZE;++i)
{
std::vector<double> data = trainData[i]; // extract data point
double label = data[0]; // extract point label
data.erase(data.begin());
std::vector<double> nLabel = encode((int)label); // encode to 1-of-N
std::vector<double> outputs = DigitNet.runNet(data);
error = DigitNet.trainNet(data,nLabel,outType); // train net, return MSE
// decode output and compare to correct output
if( decode(outputs) == (int)label )
truecnt++;
}
// stop timer and print out useful info
timed=(int)time(NULL);
times=timed-times;
std::cout << "done" << std::endl;
std::cout << "training time: " << "\t \t \t" << times << " seconds " << std::endl;
std::cout << "training accuracy: " << "\t \t" << truecnt*100./TRAIN_SIZE << "%" << std::endl;
// test net on test data
times=(int)time(NULL); // init time counter
std::cout << "\n" << "test points: " << "\t \t \t" << TEST_SIZE << std::endl;
std::cout << "testing network..." << "\t \t";
truecnt = 0;
for(int i=0;i<TEST_SIZE;++i)
{
std::vector<double> data = testData[i]; // extract data point
double label = data[0]; // extract label
data.erase(data.begin());
std::vector<double> outputs = DigitNet.runNet(data); // run net
// decode output and compare to correct output
if( decode(outputs) == (int)label )
truecnt++;
}
// stop timer and print out useful info
timed=(int)time(NULL);
times=timed-times;
std::cout << "done" << std::endl;
std::cout << "testing time: " << "\t \t \t" << times << " seconds " << std::endl;
std::cout << "test accuracy: " << "\t \t \t" << truecnt*100./TEST_SIZE << "% " << std::endl;
// save weights to reuse net in the future
DigitNet.saveNet();
}
void testMakeAppearanceColor()
{
std::cout << "mappearanceTest testMakeAppearanceColor" << std::endl;
aam::Vertices2DList verticesA(TOTAL_POINTS_NUMBER),
verticesB(TOTAL_POINTS_NUMBER);
aam::Vertices2DList meansVertices(TOTAL_POINTS_NUMBER);
aam::Vertices2DList basePoints(TOTAL_POINTS_NUMBER);
for (int i = 0; i < TOTAL_POINTS_NUMBER; i++)
{
verticesA[i].y = verticesAData[i * 2];
verticesA[i].x = verticesAData[i * 2 + 1];
verticesB[i].y = verticesBData[i * 2];
verticesB[i].x = verticesBData[i * 2 + 1];
meansVertices[i].y = meansData[i];
meansVertices[i].x = meansData[i + TOTAL_POINTS_NUMBER];
basePoints[i].y = basePointsData[2 * i];
basePoints[i].x = basePointsData[2 * i + 1];
}
aam::ShapeData shapeData;
shapeData.xMean = meansVertices;
shapeData.lines = boost::assign::list_of<int>
(0) (38) (37) (36) (31) (32) (33) (12) (11) (10) (9)
(8) (7) (6) (5) (4) (3) (2) (1);
shapeData.textureSize = cv::Size(255, 255);
try
{
cv::Mat im = cv::imread("data/IMM/01-1m.jpg");
std::vector<aam::TrainModelInfo> trainData(2);
trainData[0].setVertices(verticesA);
trainData[0].setImage(im, false);
im = cv::imread("data/IMM/01-2m.jpg");
trainData[1].setVertices(verticesB);
trainData[1].setImage(im, false);
aam::CommonFunctions::delaunay(verticesA, shapeData.triangles);
aam::AppearanceData appearanceData;
aam::AAMFunctions2D::makeAppearanceModel(
trainData, shapeData,
aam::AAMFunctions2D::DEFAULT_PCA_CUT_THRESOLD,
appearanceData);
cv::Mat objectPixels = cv::imread("data/object_pixels_synth.bmp",
cv::IMREAD_GRAYSCALE);
for (int i = 0; i < std::min(3, appearanceData.eigenVectors.cols);
i++)
{
aam::RealMatrix eigv =
appearanceData.eigenVectors.col(i);
aam::RealType minElemR = *std::min_element(eigv.begin(),
eigv.begin() + appearanceData.nPixels);
aam::RealType maxElemR = *std::max_element(eigv.begin(),
eigv.begin() + appearanceData.nPixels);
aam::RealType minElemG = *std::min_element(
eigv.begin() + appearanceData.nPixels,
eigv.begin() + 2 * appearanceData.nPixels);
aam::RealType maxElemG = *std::max_element(
eigv.begin() + appearanceData.nPixels,
eigv.begin() + 2 * appearanceData.nPixels);
aam::RealType minElemB = *std::min_element(
eigv.begin() + 2 * appearanceData.nPixels,
eigv.begin() + 3 * appearanceData.nPixels);
aam::RealType maxElemB = *std::max_element(
eigv.begin() + 2 * appearanceData.nPixels,
eigv.begin() + 3 * appearanceData.nPixels);
cv::Mat result;
aam::AAMFunctions2D::vector2Appearance(
appearanceData.eigenVectors.col(i), objectPixels,
cv::Size(255, 255), result);
std::vector<cv::Mat> channels;
cv::split(result, channels);
channels[0] = (channels[0] - minElemR) / (maxElemR - minElemR);
channels[1] = (channels[1] - minElemG) / (maxElemG - minElemG);
channels[2] = (channels[2] - minElemB) / (maxElemB - minElemB);
std::ostringstream winName;
cv::Mat img;
cv::merge(channels, img);
winName << "eigen appearance " << i + 1;
cv::imshow(winName.str(), img);
}
cv::waitKey(0);
if (appearanceData.basePoints.size() != TOTAL_POINTS_NUMBER)
{
std::cout << "%TEST_FAILED% time=0 testname=testMakeAppearanceColor (mappearanceTest) message=Invalid base points array size" << std::endl;
}
else if (cv::countNonZero(cv::abs(
aam::pointVector2Mat(aam::operator -(appearanceData.basePoints,
basePoints))) > 1e-6))
{
std::cout << "%TEST_FAILED% time=0 testname=testMakeAppearanceColor (mappearanceTest) message=Invalid base points array data" << std::endl;
}
}
catch (std::exception& e)
{
std::cout << "%TEST_FAILED% time=0 testname=testMakeAppearance (mappearanceTest) message=Exception occured: " <<
e.what() << std::endl;
}
}
int main()
{
//learning rate
double alpha = 0.03;
int maxIter = 15000;
int miniBatchSize = 1000;
double noiseRatio = 0.3;
int inputSize = 28*28;
int hiddenSize = 28;
int imgWidth = 8;
char *fileBuf = new char[4096];
bool ret = loadFileToBuf("ParamConfig.ini",fileBuf,4096);
if(ret)
{
getConfigDoubleValue(fileBuf,"noiseRatio:",noiseRatio);
getConfigDoubleValue(fileBuf,"alpha:",alpha);
getConfigIntValue(fileBuf,"maxIter:",maxIter);
getConfigIntValue(fileBuf,"miniBatchSize:",miniBatchSize);
getConfigIntValue(fileBuf,"hiddenSize:",hiddenSize);
getConfigIntValue(fileBuf,"inputSize:",inputSize);
getConfigIntValue(fileBuf,"imgWidth:",imgWidth);
cout << "noiseRatio:" << noiseRatio << endl;
cout << "alpha:" << alpha << endl;
cout << "maxIter:" << maxIter << endl;
cout << "miniBatchSize:" << miniBatchSize << endl;
cout << "hiddenSize:" << hiddenSize << endl;
cout << "inputSize:" << inputSize << endl;
cout << "imgWidth:" << imgWidth << endl;
}
delete []fileBuf;
//set eigen threads
SYSTEM_INFO info;
GetSystemInfo(&info);
Eigen::setNbThreads(info.dwNumberOfProcessors);
MatrixXd trainData(1,inputSize);
DAE dae(inputSize,hiddenSize);
ret = loadMnistData(trainData,"mnist\\train-images-idx3-ubyte");
cout << "Loading training data..." << endl;
if(ret == false)
{
return -1;
}
clock_t start = clock();
//cout << trainData.rows() << " " << trainData.cols() << endl;
MatrixXd showImage = trainData.leftCols(100).transpose();
buildImage(showImage,imgWidth,"data.jpg");
dae.train(trainData,noiseRatio,alpha,maxIter,miniBatchSize);
cout << "End Train" << endl;
MatrixXd hiddenTheta = dae.getTheta();
buildImage(hiddenTheta,imgWidth,"weights.jpg",true);
cout << "Saving hidden neurons" << endl;
dae.saveModel("DAE_Model.txt");
clock_t end = clock();
cout << "The code ran for " << (end - start)/(double)(CLOCKS_PER_SEC*60)
<< " minutes on " << Eigen::nbThreads() << " thread(s)." << endl;
cout << "noiseRatio: " << noiseRatio << endl;
cout << "alpha: " << alpha << endl;
cout << "miniBatchSize: " << miniBatchSize << endl;
system("pause");
return 0;
}
void Learner::learn(const Mat &img, const Mat &imgB, const Mat &img32F, const TYPE_BBOX &ret)
{
TYPE_TRAIN_DATA_SET &nnTrainDataset = detector->trainDataSetNN;
nnTrainDataset.clear();
TYPE_TRAIN_DATA_SET &rfTrainDataset = detector->trainDataSetRF;
rfTrainDataset.clear();
auto &scanBBs = detector->scanBBs;
detector->sortByOverlap(ret, true);
int count = 0;
// P-expert - NN
nnTrainDataset.push_back(make_pair(img32F(scanBBs[0]), CLASS_POS));
// P-expert - RF
int tlx = img.cols, tly = img.rows, brx = 0, bry = 0;
for(int i = 0; i < LEARNER_N_GOOD_BB && scanBBs[i].overlap >= GOODBB_OL; i++)
{
tlx = min(tlx, scanBBs[i].tl().x);
tly = min(tly, scanBBs[i].tl().y);
brx = max(brx, scanBBs[i].br().x);
bry = max(bry, scanBBs[i].br().y);
}
Point tl(tlx, tly), br(brx, bry);
Rect bbHull(tl, br);
int cx, cy;
cx = round((double)(tlx + brx) / 2);
cy = round((double)(tly + bry) / 2);
for(int j = 0; j < LEARNER_N_WARPED; j++)
{
Mat warped;
if(j != 0)
{
patchGenerator(imgB, Point(cx, cy), warped, bbHull.size(), theRNG());
// for optimum in RF::getcode()
Mat tmp(imgB.size() ,CV_8U, Scalar::all(0));
warped.copyTo(tmp(Rect(0, 0, bbHull.size().width, bbHull.size().height)));
warped = tmp;
}
else
{
warped = imgB(bbHull);
}
for(int i = 0; i < LEARNER_N_GOOD_BB && scanBBs[i].overlap >= GOODBB_OL; i++)
{
Rect rect(scanBBs[i].tl() - tl, scanBBs[i].br() - tl);
TYPE_TRAIN_DATA trainData(make_pair(warped(rect), CLASS_POS));
rfTrainDataset.push_back(trainData);
count++;
}
}
// N-expert - NN
int nCountNN = 0;
for(int i = 0; i < detector->scanBBs.size(); i++)
{
TYPE_DETECTOR_SCANBB &sbb = detector->scanBBs[i];
if(sbb.status != DETECTOR_REJECT_VAR)
{
if(sbb.overlap < BADBB_OL)
{
nCountNN++;
TYPE_TRAIN_DATA trainData(make_pair(img32F(sbb), CLASS_NEG));
nnTrainDataset.push_back(trainData);
}
}
if(nCountNN == LEARNER_N_NN_NEG) break;
}
// N-expert - RF
int nCountRF = 0;
for(int i = 0; i < detector->scanBBs.size(); i++)
{
TYPE_DETECTOR_SCANBB &sbb = detector->scanBBs[i];
if(sbb.status != DETECTOR_REJECT_VAR)
{
if(sbb.overlap < BADBB_OL && sbb.posterior >= 0.1)
{
nCountRF++;
TYPE_TRAIN_DATA trainData(make_pair(imgB(sbb), CLASS_NEG));
rfTrainDataset.push_back(trainData);
}
}
}
stringstream info;
info << "Generated 1 positive example for NN, and " << count << " positive example for RF.";
outputInfo("Learner", info.str());
stringstream info2;
info2 << "Generated " << nCountNN << " NN negative sample(s), and " << nCountRF << " RF negative example(s).";
outputInfo("Learner", info2.str());
detector->update();
stringstream info3;
info3 << "Updated detector.";
outputInfo("Learner", info3.str());
detector->nNClassifier.showModel();
}
