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;
}
void MnistDataSet<M>::read() {
std::string path = "/home/local/datasets/MNIST";
ifstream ftraind((path + "/train-images.idx3-ubyte").c_str());
ifstream ftrainl((path + "/train-labels.idx1-ubyte").c_str());
ifstream ftestd ((path + "/t10k-images.idx3-ubyte").c_str());
ifstream ftestl ((path + "/t10k-labels.idx1-ubyte").c_str());
char buf[16];
ftraind.read(buf,16); ftrainl.read(buf, 8);
ftestd.read(buf,16); ftestl.read(buf, 8);
tensor<unsigned char, host_memory_space> traind(extents[n_trainData][n_dim]);
tensor<unsigned char, host_memory_space> trainl(extents[n_trainData]);
tensor<unsigned char, host_memory_space> testd(extents[n_testData][n_dim]);
tensor<unsigned char, host_memory_space> testl(extents[n_testData]);
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());
tensor<unsigned char, M> train_d(extents[n_trainData][n_dim]);
tensor<unsigned char, M> train_l(extents[n_trainData]);
tensor<unsigned char, M> test_d(extents[n_testData][n_dim]);
tensor<unsigned char, M> test_l(extents[n_testData]);
train_d = traind;
train_l = trainl;
test_d = testd;
test_l = testl;
// conversion to float:
convert(this->X_train, train_d);
convert(this->Y_train, train_l);
convert(this->X_test, test_d);
convert(this->Y_test, test_l);
}