npoint_mlpack::ResamplingSplitter::ResamplingSplitter(arma::mat& data,
arma::colvec& weights,
int num_x_regions,
int num_y_regions,
int num_z_regions,
ResamplingHelper& helper)
:
resampling_helper_(helper),
num_resampling_regions_(num_x_regions * num_y_regions * num_z_regions),
data_all_mat_(data.memptr(), data.n_rows, data.n_cols, false),
data_all_weights_(weights),
data_mats_(num_resampling_regions_),
data_weights_(num_resampling_regions_),
num_x_partitions_(num_x_regions),
num_y_partitions_(num_y_regions),
num_z_partitions_(num_z_regions),
x_step_(resampling_helper_.x_size() / (double)num_x_partitions_),
y_step_(resampling_helper_.y_size() / (double)num_y_partitions_),
z_step_(resampling_helper_.z_size() / (double)num_z_partitions_),
num_points_(data.n_cols),
num_points_per_region_(num_resampling_regions_, 0)
{
for (size_t i = 0; i < num_resampling_regions_; i++) {
data_mats_[i] = new arma::mat;
data_weights_[i] = new arma::colvec;
}
SplitData_();
} // constructor
typename std::enable_if<
HasWeightsCheck<T, arma::mat&(T::*)()>::value, size_t>::type
LayerWeights(T& layer,
arma::mat& weights,
size_t offset,
arma::mat& /* unused */)
{
layer.Weights() = arma::mat(weights.memptr() + offset,
layer.Weights().n_rows, layer.Weights().n_cols, false, false);
return layer.Weights().n_elem;
}
typename std::enable_if<
HasGradientCheck<T, arma::cube&(T::*)()>::value, size_t>::type
LayerGradients(T& layer,
arma::mat& gradients,
size_t offset,
arma::cube& /* unused */)
{
layer.Gradient() = arma::cube(gradients.memptr() + offset,
layer.Weights().n_rows, layer.Weights().n_cols,
layer.Weights().n_slices, false, false);
return layer.Weights().n_elem;
}
typename std::enable_if<
HasWeightsCheck<T, arma::mat&(T::*)()>::value, size_t>::type
LayerWeights(InitializationRuleType& initializeRule,
T& layer,
arma::mat& weights,
size_t offset,
arma::mat& /* output */)
{
layer.Weights() = arma::mat(weights.memptr() + offset,
layer.Weights().n_rows, layer.Weights().n_cols, false, false);
initializeRule.Initialize(layer.Weights(), layer.Weights().n_rows,
layer.Weights().n_cols);
return layer.Weights().n_elem;
}
void Convolution2DMethodTest(const arma::mat input,
const arma::mat filter,
const arma::mat output)
{
arma::mat convOutput;
ConvolutionFunction::Convolution(input, filter, convOutput);
// Check the outut dimension.
bool b = (convOutput.n_rows == output.n_rows) &&
(convOutput.n_cols == output.n_cols);
BOOST_REQUIRE_EQUAL(b, 1);
const double* outputPtr = output.memptr();
const double* convOutputPtr = convOutput.memptr();
for (size_t i = 0; i < output.n_elem; i++, outputPtr++, convOutputPtr++)
BOOST_REQUIRE_CLOSE(*outputPtr, *convOutputPtr, 1e-3);
}
void CLgemm::dgemm(
arma::mat &res,
arma::mat const &left,
arma::mat const &right,
double alpha,
double beta,
bool transL,
bool transR)
{
timer.tic();
if (transL == false && transR == false)
{
//Check if res is correctly sized
if (res.n_rows != left.n_rows || res.n_cols != right.n_cols)
throw std::string("Badly shaped \"res\" in CLgemm::dgemm");
int m = left.n_rows;
int n = right.n_cols;
int p = left.n_cols;
int work = m * n * p;
if (work < 35000000) //TODO: A more carefull analysis of this decision should be done.
{
res = alpha * left * right + beta * res;
} else
{
//The memptr can not be const in cl, however it is expected not to change.
cl_double *A_p = const_cast<double *> (left.memptr());
cl_double *B_p = const_cast<double *> (right.memptr());
cl_double *C_p = res.memptr();
//Create CL buffers, pointing to Host memory.
cl::Buffer A_cl(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, sizeof (*A_p) * left.n_elem, A_p);
cl::Buffer B_cl(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, sizeof (*B_p) * right.n_elem, B_p);
cl::Buffer C_cl(context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR, sizeof (*C_p) * res.n_elem, C_p);
const size_t amd_M = left.n_rows;
const size_t amd_N = right.n_cols;
const size_t amd_P = right.n_rows;
if (left.n_cols != right.n_rows)
throw std::string("CLgemm: left and right matrix dimensions not compatible.");
const clAmdBlasOrder amd_order = clAmdBlasColumnMajor;
const clAmdBlasTranspose amd_transAll = clAmdBlasNoTrans;
const cl_double amd_alpha = alpha;
const cl_double amd_beta = beta;
const size_t amd_lda = amd_M;
const size_t amd_ldb = amd_P;
const size_t amd_ldc = amd_M;
cl::Event e;
clAmdBlasDgemm(amd_order, amd_transAll, amd_transAll, amd_M, amd_N, amd_P,
amd_alpha, A_cl(), amd_lda, B_cl(), amd_ldb, amd_beta, C_cl(), amd_ldc,
1, &queue(), 0, NULL, &e());
queue.enqueueReadBuffer(C_cl, true, 0, sizeof (*C_p) * res.n_elem, C_p); //BLOCKING
//clReleaseMemObject();
}
} else if (transL == true && transR == false)
res = alpha * trans(left) * right + beta * res;
else if (transL == false && transR == true)
res = alpha * left * trans(right) + beta * res;
else if (transL == true && transR == true)
res = alpha * trans(left) * trans(right) + beta * res;
tot_time += timer.toc();
return;
}
// [[Rcpp::export]]
arma::mat sgd(arma::mat& coords,
arma::ivec& targets_i, // vary randomly
arma::ivec& sources_j, // ordered
arma::ivec& ps, // N+1 length vector of indices to start of each row j in vector is
arma::vec& weights, // w{ij}
const double& gamma,
const double& rho,
const arma::uword& n_samples,
const int& M,
const double& alpha,
const Rcpp::Nullable<Rcpp::NumericVector> momentum,
const bool& useDegree,
const Rcpp::Nullable<Rcpp::NumericVector> seed,
const Rcpp::Nullable<Rcpp::NumericVector> threads,
const bool verbose) {
#ifdef _OPENMP
checkCRAN(threads);
#endif
const dimidxtype D = coords.n_rows;
const vertexidxtype N = coords.n_cols;
const edgeidxtype E = targets_i.n_elem;
Visualizer* v;
if (momentum.isNull()) v = new Visualizer(
sources_j.memptr(), targets_i.memptr(), coords.memptr(),
D, N, E,
rho, n_samples,
M, alpha, gamma);
else {
float moment = NumericVector(momentum)[0];
if (moment < 0) throw Rcpp::exception("Momentum cannot be negative.");
if (moment > 0.95) throw Rcpp::exception("Bad things happen when momentum is > 0.95.");
v = new MomentumVisualizer(
sources_j.memptr(), targets_i.memptr(), coords.memptr(),
D, N, E,
rho, n_samples, moment,
M, alpha, gamma);
}
distancetype* negweights = new distancetype[N];
std::fill(negweights, negweights + N, 0);
if (useDegree) {
std::for_each(targets_i.begin(), targets_i.end(), [&negweights](const sword& e) {negweights[e]++;});
} else {
for (vertexidxtype p = 0; p < N; ++p) {
for (edgeidxtype e = ps[p]; e != ps[p + 1]; ++e) {
negweights[p] += weights[e];
}
}
}
std::for_each(negweights, negweights + N, [](distancetype& weight) {weight = pow(weight, 0.75);});
v -> initAlias(weights.memptr(), negweights, seed);
delete[] negweights;
const uword batchSize = BATCHSIZE;
#ifdef _OPENMP
const unsigned int ts = omp_get_max_threads();
#else
const unsigned int ts = 2;
#endif
Progress progress(max((uword) ts, n_samples / BATCHSIZE), verbose);
#ifdef _OPENMP
#pragma omp parallel for
#endif
for (unsigned int t = 0; t < ts; ++t) {
v->thread(progress, batchSize);
}
delete v;
return coords;
}
void Distribution::generate_distribution3D(arma::mat& dist,
int n_p,
double bin_edge,
int N,
bool rerun) {
int x_i, y_i, z_i, r_i, n, n_tot;
double x, y, z, r, dr, dr_R, stretch, mean_r;
int dim = 3;
using namespace arma;
ucube distribution(N, N, N);
uvec radial_dist = zeros<uvec > (N);
mat tot_dist;
vec R = arma::sqrt(sum(dist % dist, 1));
n = dist.n_rows;
#ifdef MPI_ON
ivec n_list = zeros<ivec > (n_nodes);
MPI_Allgather(&n, 1, MPI_INT, n_list.memptr(), 1, MPI_INT, MPI_COMM_WORLD);
n_tot = accu(n_list);
#else
n_tot = n;
#endif
detect_deadlock(dist, n_p, dim, n);
//On fly calculation during QMC initialized by a binedge 0.
if (bin_edge == 0) {
stretch = 3;
//calculate the bin edge and size based on the mean radius
if (!locked) {
mean_r = ErrorEstimator::combine_mean(mean(R), n, n_tot);
} else {
mean_r = 0;
int k = 0;
for (int i = 0; i < n; i++) {
if (is_deadlocked(dist, dim, i)) continue;
mean_r += R(i);
k++;
}
mean_r = ErrorEstimator::combine_mean(mean_r / k, n, n_tot);
}
bin_edge = stretch*mean_r;
// if (node==0) cout << "pre " << mean_r << endl;
}
dr = 2 * bin_edge / (N - 1);
for (int ni = 0; ni < n; ni++) {
if (is_deadlocked(dist, dim, ni)) continue;
x = dist(ni, 0);
y = dist(ni, 1);
z = dist(ni, 2);
x_i = (N / 2 + int(x / dr))*(x > 0) + (N / 2 + int(x / dr) - 1)*(x <= 0);
y_i = (N / 2 + int(y / dr))*(y > 0) + (N / 2 + int(y / dr) - 1)*(y <= 0);
z_i = (N / 2 + int(z / dr))*(z > 0) + (N / 2 + int(z / dr) - 1)*(z <= 0);
if (x_i < 0 || x_i >= N) {
continue;
} else if (y_i < 0 || y_i >= N) {
continue;
} else if (z_i < 0 || z_i >= N) {
continue;
}
distribution(x_i, y_i, z_i)++;
}
dr_R = dr / 2;
for (int ni = 0; ni < n; ni++) {
if (is_deadlocked(dist, dim, ni)) continue;
r = R(ni);
r_i = r / dr_R;
if (r_i >= N) {
continue;
}
radial_dist(r_i)++;
}
#ifdef MPI_ON
if (node == 0) {
MPI_Reduce(MPI_IN_PLACE, distribution.memptr(), N * N*N, MPI_UNSIGNED, MPI_SUM, 0, MPI_COMM_WORLD);
MPI_Reduce(MPI_IN_PLACE, radial_dist.memptr(), N, MPI_UNSIGNED, MPI_SUM, 0, MPI_COMM_WORLD);
if (!rerun) {
ivec displs = zeros<ivec > (n_nodes);
double displ_sum = 0;
for (int j = 0; j < n_nodes; j++) {
displs(j) = displ_sum;
displ_sum += n_list(j);
}
tot_dist = zeros<mat > (n_tot, dim);
for (int i = 0; i < dim; i++) {
MPI_Gatherv(dist.colptr(i), n, MPI_DOUBLE,
tot_dist.colptr(i), n_list.memptr(), displs.memptr(),
MPI_DOUBLE, 0, MPI_COMM_WORLD);
}
// cout << as_scalar(mean(arma::sqrt(sum(tot_dist % tot_dist, 1)))) << endl;
displs.reset();
}
} else {
MPI_Reduce(distribution.memptr(), NULL, N * N*N, MPI_UNSIGNED, MPI_SUM, 0, MPI_COMM_WORLD);
MPI_Reduce(radial_dist.memptr(), NULL, N, MPI_UNSIGNED, MPI_SUM, 0, MPI_COMM_WORLD);
if (!rerun) {
for (int i = 0; i < dim; i++) {
MPI_Gatherv(dist.colptr(i), n, MPI_DOUBLE,
NULL, NULL, NULL,
NULL, 0, MPI_COMM_WORLD);
}
}
radial_dist.reset();
distribution.reset();
}
#else
tot_dist(dist.memptr(), n, dim, false, true);
#endif
n_list.reset();
R.reset();
if (node == 0) {
if (!silent) cout << "3D Distribution calculated using " << n_tot << " samples." << endl;
cube normalized_dist = conv_to< cube>::from(distribution);
vec normalized_radd = conv_to< vec>::from(radial_dist);
vec radial_axis = linspace(0, bin_edge, N);
normalized_dist *= n_p / (accu(normalized_dist) * dr * dr * dr);
//project out a symmetric axis and normalize (skip singularity)
// normalized_radd(span(1, N - 1)) /= radial_axis(span(1, N - 1));
//
// normalized_radd(span(1, N - 1)) /= radial_axis(span(1, N - 1));
//
// normalized_radd(0) = normalized_radd(1);
// cout << normalized_radd.max()/(2*datum::pi*accu(normalized_radd)*dr_R) << endl;
normalized_radd *= n_p / (accu(normalized_radd) * dr_R);
s << path << "walker_positions/dist_out_" << name + suffix << "_edge" << bin_edge << ".arma3D";
normalized_dist.save(s.str());
normalized_dist.reset();
s.str(std::string());
if (!rerun) {
s << path << "walker_positions/dist_rawdata_" << name + suffix << ".arma";
tot_dist.save(s.str());
tot_dist.reset();
s.str(std::string());
}
s << path << "walker_positions/radial_out_" << name + suffix << "_edge" << bin_edge << ".arma";
normalized_radd.save(s.str());
normalized_radd.reset();
s.str(std::string());
radial_axis.reset();
radial_dist.reset();
distribution.reset();
}
}