void testEigenConstMap()
{
Vector_t x(4);
x << 1, 2, 3, 4;
std::cout << x << std::endl;
const Vector_t y = x;
std::cout << y << std::endl;
Matrix_t X = Eigen::Map<const Matrix_t>(y.data(), 2, 2);
std::cout << X << std::endl;
}
void testEigenMap()
{
Vector_t theta;
theta.setZero(4 * 5);
for (int i = 0; i < theta.size(); ++i)
theta(i) = i;
std::cout << theta << std::endl;
Eigen::Map<Matrix_t> Theta(theta.data(), 4, 5); // reshape
std::cout << Theta << std::endl;
Vector_t theta2(Eigen::Map<Vector_t>(Theta.data(), 5 * 4));
std::cout << theta2 << std::endl;
}
void testStlDriver()
{
const int numPatches = 200000; // 200000
const int patchWidth = 9;
const int numFeatures = 50;
const double lambda = 0.0005f;
const double epsilon = 1e-2;
Config config;
config.setValue("addBiasTerm", false);
config.setValue("meanStddNormalize", false);
config.setValue("configurePolicyTesting", false);
config.setValue("trainingMeanAndStdd", false);
updateMNISTConfig(config);
if (false)
{
MNISTSamplePatchesUnlabeledDataFunction mnistUnlabeled(numPatches, patchWidth);
SoftICACostFunction sfc(numFeatures, lambda, epsilon);
LIBLBFGSOptimizer lbfgs(200); // 1000
Driver drv1(&config, &mnistUnlabeled, &sfc, &lbfgs);
const Vector_t optThetaRica = drv1.drive();
Matrix_t Wrica(
Eigen::Map<const Matrix_t>(optThetaRica.data(), numFeatures, pow(patchWidth, 2)));
std::ofstream ofs_wrica("../W2.txt");
ofs_wrica << Wrica << std::endl;
}
Matrix_t Wrica;
// debug: read off the values
std::ifstream in("/home/sam/School/online/stanford_dl_ex/W2.txt");
if (in.is_open())
{
std::string str;
int nbRows = 0;
while (std::getline(in, str))
{
if (str.size() == 0)
continue;
std::istringstream iss(str);
std::vector<double> tokens //
{ std::istream_iterator<double> { iss }, std::istream_iterator<double> { } };
Wrica.conservativeResize(nbRows + 1, tokens.size());
for (size_t i = 0; i < tokens.size(); ++i)
Wrica(nbRows, i) = tokens[i];
++nbRows;
}
}
else
{
std::cerr << "file W.txt failed" << std::endl;
exit(EXIT_FAILURE);
}
const int imageDim = 28;
Eigen::Vector2i imageConfig;
imageConfig << imageDim, imageDim;
const int numFilters = numFeatures;
const int poolDim = 5;
const int filterDim = patchWidth;
const int convDim = (imageDim - filterDim + 1);
assert(convDim % poolDim == 0);
const int outputDim = (convDim / poolDim);
StlFilterFunction stlFilterFunction(filterDim, Wrica);
SigmoidFunction sigmoidFunction;
ConvolutionFunction convolutionFunction(&stlFilterFunction, &sigmoidFunction);
MeanPoolFunction meanPoolFunction(numFilters, outputDim);
MNISTSamplePatchesLabeledDataFunction mnistLabeled(&convolutionFunction, &meanPoolFunction,
imageConfig, numFilters, poolDim, outputDim);
SoftmaxCostFunction mnistcf(0.01f);
LIBLBFGSOptimizer lbfgs2(300);
config.setValue("configurePolicyTesting", false);
config.setValue("trainingMeanAndStdd", true);
config.setValue("meanStddNormalize", true);
config.setValue("addBiasTerm", true);
config.setValue("numGrd", true);
config.setValue("training_accuracy", true);
config.setValue("testing_accuracy", true);
//config.setValue("addBiasTerm", false);
Driver drv2(&config, &mnistLabeled, &mnistcf, &lbfgs2);
drv2.drive();
}
void testConvolutionAndPool()
{
if (true)
{
const int filterDim = 8;
const int imageDim = 28;
const int poolDim = 3;
const int numFilters = 100;
const int outputDim = (imageDim - filterDim + 1) / poolDim;
RandomFilterFunction rff(filterDim, numFilters);
rff.configure();
SigmoidFunction sf;
ConvolutionFunction cf(&rff, &sf);
MeanPoolFunction pf(numFilters, outputDim);
//MeanPoolFunction mpf;
Eigen::Vector2i configImageDim;
configImageDim << imageDim, imageDim;
MNISTDataFunction mnistdf;
Config config;
updateMNISTConfig(config);
config.setValue("addBiasTerm", false);
config.setValue("meanStddNormalize", false);
mnistdf.configure(&config);
Convolutions* convolvedFeatures = nullptr;
for (int i = 0; i < 13; ++i)
convolvedFeatures = cf.conv(mnistdf.getTrainingX().topRows<199>(), configImageDim);
assert(convolvedFeatures->unordered_map.size() == 199);
for (auto i = convolvedFeatures->unordered_map.begin();
i != convolvedFeatures->unordered_map.end(); ++i)
assert((int )i->second.size() == rff.getWeights().cols());
// Validate convoluations
Matrix_t ConvImages = mnistdf.getTrainingX().topRows<8>();
for (int i = 0; i < 1000; ++i)
{
const int filterNum = rand() % rff.getWeights().cols();
const int imageNum = rand() % 8;
const int imageRow = rand() % (configImageDim(0) - rff.getConfig()(0) + 1);
const int imageCol = rand() % (configImageDim(1) - rff.getConfig()(1) + 1);
Vector_t im = ConvImages.row(imageNum);
Eigen::Map<Matrix_t> Image(im.data(), configImageDim(0), configImageDim(1));
Matrix_t Patch = Image.block(imageRow, imageCol, rff.getConfig()(0), rff.getConfig()(1));
// Filter
Eigen::Map<Matrix_t> W(rff.getWeights().col(filterNum).data(), rff.getConfig()(0),
rff.getConfig()(1));
const double b = rff.getBiases()(filterNum);
double feature = Patch.cwiseProduct(W).sum() + b;
feature = 1.0f / (1.0f + exp(-feature));
if (fabs(
feature - convolvedFeatures->unordered_map[imageNum][filterNum]->X(imageRow, imageCol))
> 1e-9)
{
std::cout << "Convolved feature does not match test feature: " << i << std::endl;
std::cout << "Filter Number: " << filterNum << std::endl;
std::cout << "Image Number: " << imageNum << std::endl;
std::cout << "Image Row: " << imageRow << std::endl;
std::cout << "Image Col: " << imageCol << std::endl;
std::cout << "Convolved feature: "
<< convolvedFeatures->unordered_map[imageNum][filterNum]->X(imageRow, imageCol)
<< std::endl;
std::cout << "Test feature: " << feature << std::endl;
std::cout << "Convolved feature does not match test feature" << std::endl;
exit(EXIT_FAILURE);
}
}
// Pool
Poolings* pooling = nullptr;
for (int i = 0; i < 13; ++i)
pooling = pf.pool(convolvedFeatures, poolDim);
assert((int )pooling->unordered_map.size() == 199);
for (auto iter = pooling->unordered_map.begin(); iter != pooling->unordered_map.end(); ++iter)
{
assert(iter->second.size() == (size_t )rff.getWeights().cols());
for (auto iter2 = iter->second.begin(); iter2 != iter->second.end(); ++iter2)
{
assert(iter2->second->X.rows() == (configImageDim(0) - rff.getConfig()(0) + 1) / 3);
assert(iter2->second->X.rows() == 7);
assert(iter2->second->X.cols() == (configImageDim(0) - rff.getConfig()(0) + 1) / 3);
assert(iter2->second->X.cols() == 7);
}
}
}
if (true)
{
// test pool function
Vector_t testVec(64);
for (int i = 0; i < testVec.size(); ++i)
testVec(i) = i + 1;
Eigen::Map<Matrix_t> TestMatrix(testVec.data(), 8, 8);
std::cout << "TestMatrix: " << std::endl;
std::cout << TestMatrix << std::endl;
Matrix_t ExpectedMatrix(2, 2);
ExpectedMatrix(0, 0) = TestMatrix.block(0, 0, 4, 4).array().mean();
ExpectedMatrix(0, 1) = TestMatrix.block(0, 4, 4, 4).array().mean();
ExpectedMatrix(1, 0) = TestMatrix.block(4, 0, 4, 4).array().mean();
ExpectedMatrix(1, 1) = TestMatrix.block(4, 4, 4, 4).array().mean();
std::cout << "Expected: " << std::endl;
std::cout << ExpectedMatrix << std::endl;
Convolutions cfs;
Convolution xcf;
xcf.X = TestMatrix;
cfs.unordered_map[0].insert(std::make_pair(0, (&xcf)));
MeanPoolFunction testMpf(1, 2);
Poolings* pfs = testMpf.pool(&cfs, 4);
assert(pfs->unordered_map.size() == 1);
assert(pfs->unordered_map[0].size() == 1);
Matrix_t PX = pfs->unordered_map[0][0]->X;
std::cout << "Obtain: " << std::endl;
std::cout << PX << std::endl;
}
}
