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

Dmatrix refactor stage 1 (#3301)

Browse Source 
* Use sparse page as singular CSR matrix representation

* Simplify dmatrix methods

* Reduce statefullness of batch iterators

* BREAKING CHANGE: Remove prob_buffer_row parameter. Users are instead recommended to sample their dataset as a preprocessing step before using XGBoost.
This commit is contained in:
Rory Mitchell
2018-06-07 10:25:58 +12:00
committed by GitHub
parent 286dccb8e8
commit a96039141a
47 changed files with 650 additions and 1036 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
src/linear/coordinate_common.h
View File

@@ -65,10 +65,10 @@ inline std::pair<double, double> GetGradient(int group_idx, int num_group, int f
const std::vector<GradientPair> &gpair,
DMatrix *p_fmat) {
double sum_grad = 0.0, sum_hess = 0.0;
dmlc::DataIter<ColBatch> *iter = p_fmat->ColIterator({static_cast<bst_uint>(fidx)});
auto iter = p_fmat->ColIterator();
while (iter->Next()) {
const ColBatch &batch = iter->Value();
ColBatch::Inst col = batch[0];
auto batch = iter->Value();
auto col = batch[fidx];
const auto ndata = static_cast<bst_omp_uint>(col.length);
for (bst_omp_uint j = 0; j < ndata; ++j) {
const bst_float v = col[j].fvalue;
@@ -96,10 +96,10 @@ inline std::pair<double, double> GetGradientParallel(int group_idx, int num_grou
const std::vector<GradientPair> &gpair,
DMatrix *p_fmat) {
double sum_grad = 0.0, sum_hess = 0.0;
dmlc::DataIter<ColBatch> *iter = p_fmat->ColIterator({static_cast<bst_uint>(fidx)});
auto iter = p_fmat->ColIterator();
while (iter->Next()) {
const ColBatch &batch = iter->Value();
ColBatch::Inst col = batch[0];
auto batch = iter->Value();
auto col = batch[fidx];
const auto ndata = static_cast<bst_omp_uint>(col.length);
#pragma omp parallel for schedule(static) reduction(+ : sum_grad, sum_hess)
for (bst_omp_uint j = 0; j < ndata; ++j) {
@@ -154,10 +154,10 @@ inline void UpdateResidualParallel(int fidx, int group_idx, int num_group,
float dw, std::vector<GradientPair> *in_gpair,
DMatrix *p_fmat) {
if (dw == 0.0f) return;
dmlc::DataIter<ColBatch> *iter = p_fmat->ColIterator({static_cast<bst_uint>(fidx)});
auto iter = p_fmat->ColIterator();
while (iter->Next()) {
const ColBatch &batch = iter->Value();
ColBatch::Inst col = batch[0];
auto batch = iter->Value();
auto col = batch[fidx];
// update grad value
const auto num_row = static_cast<bst_omp_uint>(col.length);
#pragma omp parallel for schedule(static)
@@ -325,12 +325,12 @@ class GreedyFeatureSelector : public FeatureSelector {
const bst_omp_uint nfeat = model.param.num_feature;
// Calculate univariate gradient sums
std::fill(gpair_sums_.begin(), gpair_sums_.end(), std::make_pair(0., 0.));
dmlc::DataIter<ColBatch> *iter = p_fmat->ColIterator();
auto iter = p_fmat->ColIterator();
while (iter->Next()) {
const ColBatch &batch = iter->Value();
auto batch = iter->Value();
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nfeat; ++i) {
const ColBatch::Inst col = batch[i];
const auto col = batch[i];
const bst_uint ndata = col.length;
auto &sums = gpair_sums_[group_idx * nfeat + i];
for (bst_uint j = 0u; j < ndata; ++j) {
@@ -392,13 +392,13 @@ class ThriftyFeatureSelector : public FeatureSelector {
}
// Calculate univariate gradient sums
std::fill(gpair_sums_.begin(), gpair_sums_.end(), std::make_pair(0., 0.));
dmlc::DataIter<ColBatch> *iter = p_fmat->ColIterator();
auto iter = p_fmat->ColIterator();
while (iter->Next()) {
const ColBatch &batch = iter->Value();
auto batch = iter->Value();
// column-parallel is usually faster than row-parallel
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nfeat; ++i) {
const ColBatch::Inst col = batch[i];
const auto col = batch[i];
const bst_uint ndata = col.length;
for (bst_uint gid = 0u; gid < ngroup; ++gid) {
auto &sums = gpair_sums_[gid * nfeat + i];

20
src/linear/updater_gpu_coordinate.cu
View File

@@ -81,7 +81,7 @@ struct GPUCoordinateTrainParam
float reg_alpha_denorm;
};
void RescaleIndices(size_t ridx_begin, dh::DVec<SparseBatch::Entry> *data) {
void RescaleIndices(size_t ridx_begin, dh::DVec<Entry> *data) {
auto d_data = data->Data();
dh::LaunchN(data->DeviceIdx(), data->Size(),
[=] __device__(size_t idx) { d_data[idx].index -= ridx_begin; });
@@ -92,14 +92,14 @@ class DeviceShard {
int normalised_device_idx_; // Device index counting from param.gpu_id
dh::BulkAllocator<dh::MemoryType::kDevice> ba_;
std::vector<size_t> row_ptr_;
dh::DVec<SparseBatch::Entry> data_;
dh::DVec<Entry> data_;
dh::DVec<GradientPair> gpair_;
dh::CubMemory temp_;
size_t ridx_begin_;
size_t ridx_end_;
public:
DeviceShard(int device_idx, int normalised_device_idx, const ColBatch &batch,
DeviceShard(int device_idx, int normalised_device_idx, const SparsePage &batch,
bst_uint row_begin, bst_uint row_end,
const GPUCoordinateTrainParam &param,
const gbm::GBLinearModelParam &model_param)
@@ -112,17 +112,17 @@ class DeviceShard {
// this shard
std::vector<std::pair<bst_uint, bst_uint>> column_segments;
row_ptr_ = {0};
for (auto fidx = 0; fidx < batch.size; fidx++) {
for (auto fidx = 0; fidx < batch.Size(); fidx++) {
auto col = batch[fidx];
auto cmp = [](SparseBatch::Entry e1, SparseBatch::Entry e2) {
auto cmp = [](Entry e1, Entry e2) {
return e1.index < e2.index;
};
auto column_begin =
std::lower_bound(col.data, col.data + col.length,
SparseBatch::Entry(row_begin, 0.0f), cmp);
Entry(row_begin, 0.0f), cmp);
auto column_end =
std::upper_bound(col.data, col.data + col.length,
SparseBatch::Entry(row_end, 0.0f), cmp);
Entry(row_end, 0.0f), cmp);
column_segments.push_back(
std::make_pair(column_begin - col.data, column_end - col.data));
row_ptr_.push_back(row_ptr_.back() + column_end - column_begin);
@@ -130,8 +130,8 @@ class DeviceShard {
ba_.Allocate(device_idx, param.silent, &data_, row_ptr_.back(), &gpair_,
(row_end - row_begin) * model_param.num_output_group);
for (int fidx = 0; fidx < batch.size; fidx++) {
ColBatch::Inst col = batch[fidx];
for (int fidx = 0; fidx < batch.Size(); fidx++) {
auto col = batch[fidx];
thrust::copy(col.data + column_segments[fidx].first,
col.data + column_segments[fidx].second,
data_.tbegin() + row_ptr_[fidx]);
@@ -233,7 +233,7 @@ class GPUCoordinateUpdater : public LinearUpdater {
row_begin = row_end;
}
dmlc::DataIter<ColBatch> *iter = p_fmat->ColIterator();
auto iter = p_fmat->ColIterator();
CHECK(p_fmat->SingleColBlock());
iter->Next();
auto batch = iter->Value();

10
src/linear/updater_shotgun.cc
View File

@@ -79,17 +79,17 @@ class ShotgunUpdater : public LinearUpdater {
// lock-free parallel updates of weights
selector_->Setup(*model, in_gpair->HostVector(), p_fmat,
param_.reg_alpha_denorm, param_.reg_lambda_denorm, 0);
dmlc::DataIter<ColBatch> *iter = p_fmat->ColIterator();
auto iter = p_fmat->ColIterator();
while (iter->Next()) {
const ColBatch &batch = iter->Value();
const auto nfeat = static_cast<bst_omp_uint>(batch.size);
auto batch = iter->Value();
const auto nfeat = static_cast<bst_omp_uint>(batch.Size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nfeat; ++i) {
int ii = selector_->NextFeature(i, *model, 0, in_gpair->HostVector(), p_fmat,
param_.reg_alpha_denorm, param_.reg_lambda_denorm);
if (ii < 0) continue;
const bst_uint fid = batch.col_index[ii];
ColBatch::Inst col = batch[ii];
const bst_uint fid = ii;
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) {