Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Span class. (#3548)

Browse Source 
* Add basic Span class based on ISO++20.

* Use Span<Entry const> instead of Inst in SparsePage.

* Add DeviceSpan in HostDeviceVector, use it in regression obj.
This commit is contained in:
trivialfis
2018-08-14 13:58:11 +08:00
committed by Rory Mitchell
parent 2b7a1c5780
commit 2c502784ff
28 changed files with 1927 additions and 138 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/linear/coordinate_common.h
View File

@@ -69,7 +69,7 @@ inline std::pair<double, double> GetGradient(int group_idx, int num_group, int f
while (iter->Next()) {
auto &batch = iter->Value();
auto col = batch[fidx];
const auto ndata = static_cast<bst_omp_uint>(col.length);
const auto ndata = static_cast<bst_omp_uint>(col.size());
for (bst_omp_uint j = 0; j < ndata; ++j) {
const bst_float v = col[j].fvalue;
auto &p = gpair[col[j].index * num_group + group_idx];
@@ -100,7 +100,7 @@ inline std::pair<double, double> GetGradientParallel(int group_idx, int num_grou
while (iter->Next()) {
auto &batch = iter->Value();
auto col = batch[fidx];
const auto ndata = static_cast<bst_omp_uint>(col.length);
const auto ndata = static_cast<bst_omp_uint>(col.size());
#pragma omp parallel for schedule(static) reduction(+ : sum_grad, sum_hess)
for (bst_omp_uint j = 0; j < ndata; ++j) {
const bst_float v = col[j].fvalue;
@@ -159,7 +159,7 @@ inline void UpdateResidualParallel(int fidx, int group_idx, int num_group,
auto &batch = iter->Value();
auto col = batch[fidx];
// update grad value
const auto num_row = static_cast<bst_omp_uint>(col.length);
const auto num_row = static_cast<bst_omp_uint>(col.size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint j = 0; j < num_row; ++j) {
GradientPair &p = (*in_gpair)[col[j].index * num_group + group_idx];
@@ -331,7 +331,7 @@ class GreedyFeatureSelector : public FeatureSelector {
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nfeat; ++i) {
const auto col = batch[i];
const bst_uint ndata = col.length;
const bst_uint ndata = col.size();
auto &sums = gpair_sums_[group_idx * nfeat + i];
for (bst_uint j = 0u; j < ndata; ++j) {
const bst_float v = col[j].fvalue;
@@ -399,7 +399,7 @@ class ThriftyFeatureSelector : public FeatureSelector {
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nfeat; ++i) {
const auto col = batch[i];
const bst_uint ndata = col.length;
const bst_uint ndata = col.size();
for (bst_uint gid = 0u; gid < ngroup; ++gid) {
auto &sums = gpair_sums_[gid * nfeat + i];
for (bst_uint j = 0u; j < ndata; ++j) {

8
src/linear/updater_gpu_coordinate.cu
View File

@@ -118,13 +118,13 @@ class DeviceShard {
return e1.index < e2.index;
};
auto column_begin =
std::lower_bound(col.data, col.data + col.length,
std::lower_bound(col.data(), col.data() + col.size(),
Entry(row_begin, 0.0f), cmp);
auto column_end =
std::upper_bound(col.data, col.data + col.length,
std::upper_bound(col.data(), col.data() + col.size(),
Entry(row_end, 0.0f), cmp);
column_segments.push_back(
std::make_pair(column_begin - col.data, column_end - col.data));
std::make_pair(column_begin - col.data(), column_end - col.data()));
row_ptr_.push_back(row_ptr_.back() + column_end - column_begin);
}
ba_.Allocate(device_idx, param.silent, &data_, row_ptr_.back(), &gpair_,
@@ -134,7 +134,7 @@ class DeviceShard {
auto col = batch[fidx];
auto seg = column_segments[fidx];
dh::safe_cuda(cudaMemcpy(
data_.Data() + row_ptr_[fidx], col.data + seg.first,
data_.Data() + row_ptr_[fidx], col.data() + seg.first,
sizeof(Entry) * (seg.second - seg.first), cudaMemcpyHostToDevice));
}
// Rescale indices with respect to current shard

12
src/linear/updater_shotgun.cc
View File

@@ -92,10 +92,10 @@ class ShotgunUpdater : public LinearUpdater {
auto col = batch[ii];
for (int gid = 0; gid < ngroup; ++gid) {
double sum_grad = 0.0, sum_hess = 0.0;
for (bst_uint j = 0; j < col.length; ++j) {
GradientPair &p = gpair[col[j].index * ngroup + gid];
for (auto& c : col) {
GradientPair &p = gpair[c.index * ngroup + gid];
if (p.GetHess() < 0.0f) continue;
const bst_float v = col[j].fvalue;
const bst_float v = c.fvalue;
sum_grad += p.GetGrad() * v;
sum_hess += p.GetHess() * v * v;
}
@@ -107,10 +107,10 @@ class ShotgunUpdater : public LinearUpdater {
if (dw == 0.f) continue;
w += dw;
// update grad values
for (bst_uint j = 0; j < col.length; ++j) {
GradientPair &p = gpair[col[j].index * ngroup + gid];
for (auto& c : col) {
GradientPair &p = gpair[c.index * ngroup + gid];
if (p.GetHess() < 0.0f) continue;
p += GradientPair(p.GetHess() * col[j].fvalue * dw, 0);
p += GradientPair(p.GetHess() * c.fvalue * dw, 0);
}
}
}