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

merge changes Jun092023

Browse Source
This commit is contained in:
amdsc21
2023-06-09 22:41:33 +02:00
parent 35cde3b1b2 152e2fb072
commit 5f78360949
63 changed files with 1299 additions and 243 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/metric/elementwise_metric.cu
View File

@@ -499,9 +499,13 @@ class QuantileError : public MetricNoCache {
const char* Name() const override { return "quantile"; }
void LoadConfig(Json const& in) override {
auto const& name = get<String const>(in["name"]);
CHECK_EQ(name, "quantile");
FromJson(in["quantile_loss_param"], &param_);
auto const& obj = get<Object const>(in);
auto it = obj.find("quantile_loss_param");
if (it != obj.cend()) {
FromJson(it->second, &param_);
auto const& name = get<String const>(in["name"]);
CHECK_EQ(name, "quantile");
}
}
void SaveConfig(Json* p_out) const override {
auto& out = *p_out;

24
src/predictor/cpu_predictor.cc
View File

@@ -152,7 +152,7 @@ void PredictByAllTrees(gbm::GBTreeModel const &model, std::uint32_t const tree_b
} else {
for (std::size_t i = 0; i < block_size; ++i) {
out_predt(predict_offset + i, gid) +=
scalar::PredValueByOneTree<true>(thread_temp[offset + i], tree, cats);
scalar::PredValueByOneTree<false>(thread_temp[offset + i], tree, cats);
}
}
}
@@ -430,8 +430,7 @@ class ColumnSplitHelper {
<< "column-split prediction is only supported for distributed training";
for (auto const &batch : p_fmat->GetBatches<SparsePage>()) {
CHECK_EQ(out_preds->size(),
p_fmat->Info().num_row_ * model_.learner_model_param->num_output_group);
CHECK_EQ(out_preds->size(), p_fmat->Info().num_row_ * (tree_end_ - tree_begin_));
PredictBatchKernel<SparsePageView, kBlockOfRowsSize, true>(SparsePageView{&batch}, out_preds);
}
}
@@ -543,8 +542,12 @@ class ColumnSplitHelper {
for (size_t tree_id = tree_begin_; tree_id < tree_end_; ++tree_id) {
auto const gid = model_.tree_info[tree_id];
for (size_t i = 0; i < block_size; ++i) {
preds[(predict_offset + i) * num_group + gid] +=
PredictOneTree<predict_leaf>(tree_id, batch_offset + i);
auto const result = PredictOneTree<predict_leaf>(tree_id, batch_offset + i);
if constexpr (predict_leaf) {
preds[(predict_offset + i) * (tree_end_ - tree_begin_) + tree_id] = result;
} else {
preds[(predict_offset + i) * num_group + gid] += result;
}
}
}
}
@@ -645,6 +648,9 @@ class CPUPredictor : public Predictor {
void PredictDMatrix(DMatrix *p_fmat, std::vector<bst_float> *out_preds,
gbm::GBTreeModel const &model, int32_t tree_begin, int32_t tree_end) const {
if (p_fmat->Info().IsColumnSplit()) {
CHECK(!model.learner_model_param->IsVectorLeaf())
<< "Predict DMatrix with column split" << MTNotImplemented();
ColumnSplitHelper helper(this->ctx_->Threads(), model, tree_begin, tree_end);
helper.PredictDMatrix(p_fmat, out_preds);
return;
@@ -743,6 +749,8 @@ class CPUPredictor : public Predictor {
unsigned tree_end) const override {
auto proxy = dynamic_cast<data::DMatrixProxy *>(p_m.get());
CHECK(proxy)<< "Inplace predict accepts only DMatrixProxy as input.";
CHECK(!p_m->Info().IsColumnSplit())
<< "Inplace predict support for column-wise data split is not yet implemented.";
auto x = proxy->Adapter();
if (x.type() == typeid(std::shared_ptr<data::DenseAdapter>)) {
this->DispatchedInplacePredict<data::DenseAdapter, kBlockOfRowsSize>(
@@ -773,6 +781,9 @@ class CPUPredictor : public Predictor {
out_preds->resize(model.learner_model_param->num_output_group);
if (is_column_split) {
CHECK(!model.learner_model_param->IsVectorLeaf())
<< "Predict instance with column split" << MTNotImplemented();
ColumnSplitHelper helper(this->ctx_->Threads(), model, 0, ntree_limit);
helper.PredictInstance(inst, out_preds);
return;
@@ -802,6 +813,9 @@ class CPUPredictor : public Predictor {
preds.resize(info.num_row_ * ntree_limit);
if (p_fmat->Info().IsColumnSplit()) {
CHECK(!model.learner_model_param->IsVectorLeaf())
<< "Predict leaf with column split" << MTNotImplemented();
ColumnSplitHelper helper(n_threads, model, 0, ntree_limit);
helper.PredictLeaf(p_fmat, &preds);
return;

12
src/tree/updater_gpu_hist.cu
View File

@@ -302,7 +302,7 @@ struct GPUHistMakerDevice {
matrix.feature_segments,
matrix.gidx_fvalue_map,
matrix.min_fvalue,
matrix.is_dense
matrix.is_dense && !collective::IsDistributed()
};
auto split = this->evaluator_.EvaluateSingleSplit(inputs, shared_inputs);
return split;
@@ -316,11 +316,11 @@ struct GPUHistMakerDevice {
std::vector<bst_node_t> nidx(2 * candidates.size());
auto h_node_inputs = pinned2.GetSpan<EvaluateSplitInputs>(2 * candidates.size());
auto matrix = page->GetDeviceAccessor(ctx_->gpu_id);
EvaluateSplitSharedInputs shared_inputs{
GPUTrainingParam{param}, *quantiser, feature_types, matrix.feature_segments,
matrix.gidx_fvalue_map, matrix.min_fvalue,
matrix.is_dense
};
EvaluateSplitSharedInputs shared_inputs{GPUTrainingParam{param}, *quantiser, feature_types,
matrix.feature_segments, matrix.gidx_fvalue_map,
matrix.min_fvalue,
// is_dense represents the local data
matrix.is_dense && !collective::IsDistributed()};
dh::TemporaryArray<GPUExpandEntry> entries(2 * candidates.size());
// Store the feature set ptrs so they dont go out of scope before the kernel is called
std::vector<std::shared_ptr<HostDeviceVector<bst_feature_t>>> feature_sets;

7
src/tree/updater_quantile_hist.cc
View File

@@ -419,6 +419,7 @@ class HistBuilder {
CPUExpandEntry InitRoot(DMatrix *p_fmat, linalg::MatrixView<GradientPair const> gpair,
RegTree *p_tree) {
monitor_->Start(__func__);
CPUExpandEntry node(RegTree::kRoot, p_tree->GetDepth(0));
std::size_t page_id = 0;
@@ -434,7 +435,7 @@ class HistBuilder {
{
GradientPairPrecise grad_stat;
if (p_fmat->IsDense()) {
if (p_fmat->IsDense() && !collective::IsDistributed()) {
/**
* Specialized code for dense data: For dense data (with no missing value), the sum
* of gradient histogram is equal to snode[nid]
@@ -475,12 +476,14 @@ class HistBuilder {
node = entries.front();
}
monitor_->Stop(__func__);
return node;
}
void BuildHistogram(DMatrix *p_fmat, RegTree *p_tree,
std::vector<CPUExpandEntry> const &valid_candidates,
linalg::MatrixView<GradientPair const> gpair) {
monitor_->Start(__func__);
std::vector<CPUExpandEntry> nodes_to_build(valid_candidates.size());
std::vector<CPUExpandEntry> nodes_to_sub(valid_candidates.size());
@@ -508,6 +511,7 @@ class HistBuilder {
nodes_to_sub, gpair.Values());
++page_id;
}
monitor_->Stop(__func__);
}
void UpdatePosition(DMatrix *p_fmat, RegTree const *p_tree,
@@ -525,6 +529,7 @@ class HistBuilder {
std::vector<bst_node_t> *p_out_position) {
monitor_->Start(__func__);
if (!task_->UpdateTreeLeaf()) {
monitor_->Stop(__func__);
return;
}
for (auto const &part : partitioner_) {