mnist_dataset(const std::string& path, const bool out_of_n_coding = true){
std::cout << "Reading MNIST dataset..."<<std::flush;
std::ifstream ftraind((path + "/train-images.idx3-ubyte").c_str(),std::ios::in | std::ios::binary); // image data
std::ifstream ftrainl((path + "/train-labels.idx1-ubyte").c_str(),std::ios::in | std::ios::binary); // label data
std::ifstream ftestd ((path + "/t10k-images.idx3-ubyte").c_str(),std::ios::in | std::ios::binary); // image data
std::ifstream ftestl ((path + "/t10k-labels.idx1-ubyte").c_str(),std::ios::in | std::ios::binary); // label data
assert(ftraind.is_open());
assert(ftrainl.is_open());
assert(ftestd.is_open());
assert(ftestl.is_open());
char buf[16];
ftraind.read(buf,16);
ftrainl.read(buf, 8);
ftestd.read(buf,16);
ftestl.read(buf, 8);
cuv::tensor<unsigned char,cuv::host_memory_space> traind(cuv::extents[60000][784]);
cuv::tensor<unsigned char,cuv::host_memory_space> trainl(cuv::extents[60000]);
cuv::tensor<unsigned char,cuv::host_memory_space> testd(cuv::extents[10000][784]);
cuv::tensor<unsigned char,cuv::host_memory_space> testl(cuv::extents[10000]);
ftraind.read((char*)traind.ptr(), traind.size()); assert(ftraind.good());
ftrainl.read((char*)trainl.ptr(), trainl.size()); assert(ftrainl.good());
ftestd.read((char*)testd.ptr(), testd.size()); assert(ftestd.good());
ftestl.read((char*)testl.ptr(), testl.size()); assert(ftestl.good());
train_data.resize(traind.shape());
test_data.resize(testd.shape());
convert(train_data , traind); // convert data type
convert(test_data , testd); // convert data type
if (out_of_n_coding){
train_labels.resize(cuv::extents[60000][10]);
test_labels.resize(cuv::extents[10000][10]);
train_labels = 0.f;
test_labels = 0.f;
for (unsigned int i = 0; i < trainl.size(); ++i){
train_labels(i, trainl[i]) = 1.f;
}
for (unsigned int i = 0; i < testl.size(); ++i){
test_labels(i, testl[i]) = 1.f;
}
} else {
train_labels.resize(cuv::extents[60000]);
test_labels.resize(cuv::extents[10000]);
for (unsigned int i = 0; i < trainl.size(); ++i){
train_labels(i) = trainl[i];
}
for (unsigned int i = 0; i < testl.size(); ++i){
test_labels(i) = testl[i];
}
}
//train_data = train_data[cuv::indices[cuv::index_range(0,5000)][cuv::index_range()]];
//train_labels = train_labels[cuv::indices[cuv::index_range(0,5000)][cuv::index_range()]];
binary = true;
channels = 1;
image_size = 28;
std::cout << "done."<<std::endl;
}
int main(int argc, char **argv[])
{
string name;
vector<Mat>Images(100), TestImages(50);
vector<Mat> Descriptor(100), TestDescriptor(50), TestPcafeature(50);
vector<vector<KeyPoint>>Keypoints(100), TestKeypoint(50);
Mat histogram = Mat::zeros(100, Cluster, CV_32F);
Mat Testhistogram = Mat::zeros(50, Cluster, CV_32F);
Mat Keyword = Mat::zeros(Cluster, 20, CV_32F);
Mat full_Descriptor, Pcafeature, Pcaduplicate, clusteridx, trainlabels(100, 1, CV_32F);
vector<vector<DMatch>> matches(50);
Mat predicted(Testhistogram.rows, 1, CV_32F);
// Read Training Images.
read_train(Images, name);
//Calculate SIFT features for the Training Images.
calculate_SIFT(Images,Keypoints,Descriptor);
merge_descriptor(full_Descriptor,Descriptor);
//Compute PCA for all the features across all Images.
PCA pca;
perform_PCA(full_Descriptor, Pcafeature, pca);
//Perform K-Means on all the PCA reduced features.
Pcafeature.convertTo(Pcaduplicate, CV_32F);
calculate_Kmeans(Pcaduplicate, clusteridx);
//Calculate the Keywords in the Feature Space.
make_dictionary(clusteridx, Pcaduplicate, Keyword);
//Get the Histogram for each Training Image.
hist(Descriptor, clusteridx, histogram);
//Read Test Image
read_test(TestImages, name);
//Calculate the SIFT feature for all the test Images.
calculate_SIFT(TestImages, TestKeypoint, TestDescriptor);
//Project the SIFT feature of each feature on the lower dimensional PCA plane calculated above.
pca_testProject(TestDescriptor, TestPcafeature, pca);
//Find the Label by searching for keywords closest to current feature.
get_matches(TestPcafeature,Keyword,matches);
//Calculate Histogram for each test Image.
hist_test(TestDescriptor, matches, Testhistogram);
//Perform classification through Knn Classifier.
train_labels(trainlabels);
KNearest knn;
train_classifier(histogram, trainlabels, knn);
test_classify(Testhistogram,predicted,knn);
//Calculate Accuracy for each class.
calculate_accuracy(predicted);
getchar();
return 0;
}
