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

Require context in aggregators. (#10075)

Browse Source
This commit is contained in:
Jiaming Yuan
2024-02-28 03:12:42 +08:00
committed by GitHub
parent 761845f594
commit 5ac233280e
23 changed files with 190 additions and 144 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/objective/adaptive.cc
View File

@@ -1,5 +1,5 @@
/**
* Copyright 2022-2023 by XGBoost Contributors
* Copyright 2022-2024, XGBoost Contributors
*/
#include "adaptive.h"
@@ -85,7 +85,7 @@ void UpdateTreeLeafHost(Context const* ctx, std::vector<bst_node_t> const& posit
size_t n_leaf = nidx.size();
if (nptr.empty()) {
std::vector<float> quantiles;
UpdateLeafValues(&quantiles, nidx, info, learning_rate, p_tree);
UpdateLeafValues(ctx, &quantiles, nidx, info, learning_rate, p_tree);
return;
}
@@ -100,7 +100,7 @@ void UpdateTreeLeafHost(Context const* ctx, std::vector<bst_node_t> const& posit
predt.Size() / info.num_row_);
collective::ApplyWithLabels(
info, static_cast<void*>(quantiles.data()), quantiles.size() * sizeof(float), [&] {
ctx, info, static_cast<void*>(quantiles.data()), quantiles.size() * sizeof(float), [&] {
// loop over each leaf
common::ParallelFor(quantiles.size(), ctx->Threads(), [&](size_t k) {
auto nidx = h_node_idx[k];
@@ -134,7 +134,7 @@ void UpdateTreeLeafHost(Context const* ctx, std::vector<bst_node_t> const& posit
});
});
UpdateLeafValues(&quantiles, nidx, info, learning_rate, p_tree);
UpdateLeafValues(ctx, &quantiles, nidx, info, learning_rate, p_tree);
}
#if !defined(XGBOOST_USE_CUDA)

9
src/objective/adaptive.cu
View File

@@ -1,5 +1,5 @@
/**
* Copyright 2022-2023 by XGBoost Contributors
* Copyright 2022-2024, XGBoost Contributors
*/
#include <thrust/sort.h>
@@ -150,7 +150,7 @@ void UpdateTreeLeafDevice(Context const* ctx, common::Span<bst_node_t const> pos
if (nptr.Empty()) {
std::vector<float> quantiles;
UpdateLeafValues(&quantiles, nidx.ConstHostVector(), info, learning_rate, p_tree);
UpdateLeafValues(ctx, &quantiles, nidx.ConstHostVector(), info, learning_rate, p_tree);
}
predt.SetDevice(ctx->Device());
@@ -160,7 +160,7 @@ void UpdateTreeLeafDevice(Context const* ctx, common::Span<bst_node_t const> pos
auto t_predt = d_predt.Slice(linalg::All(), group_idx);
HostDeviceVector<float> quantiles;
collective::ApplyWithLabels(info, &quantiles, [&] {
collective::ApplyWithLabels(ctx, info, &quantiles, [&] {
auto d_labels = info.labels.View(ctx->Device()).Slice(linalg::All(), IdxY(info, group_idx));
auto d_row_index = dh::ToSpan(ridx);
auto seg_beg = nptr.DevicePointer();
@@ -186,6 +186,7 @@ void UpdateTreeLeafDevice(Context const* ctx, common::Span<bst_node_t const> pos
w_it + d_weights.size(), &quantiles);
}
});
UpdateLeafValues(&quantiles.HostVector(), nidx.ConstHostVector(), info, learning_rate, p_tree);
UpdateLeafValues(ctx, &quantiles.HostVector(), nidx.ConstHostVector(), info, learning_rate,
p_tree);
}
} // namespace xgboost::obj::detail

23
src/objective/adaptive.h
View File

@@ -1,5 +1,5 @@
/**
* Copyright 2022-2023 by XGBoost Contributors
* Copyright 2022-2024, XGBoost Contributors
*/
#pragma once
@@ -17,8 +17,7 @@
#include "xgboost/host_device_vector.h" // HostDeviceVector
#include "xgboost/tree_model.h" // RegTree
namespace xgboost {
namespace obj {
namespace xgboost::obj {
namespace detail {
inline void FillMissingLeaf(std::vector<bst_node_t> const& maybe_missing,
std::vector<bst_node_t>* p_nidx, std::vector<size_t>* p_nptr) {
@@ -36,13 +35,14 @@ inline void FillMissingLeaf(std::vector<bst_node_t> const& maybe_missing,
}
}
inline void UpdateLeafValues(std::vector<float>* p_quantiles, std::vector<bst_node_t> const& nidx,
MetaInfo const& info, float learning_rate, RegTree* p_tree) {
inline void UpdateLeafValues(Context const* ctx, std::vector<float>* p_quantiles,
std::vector<bst_node_t> const& nidx, MetaInfo const& info,
float learning_rate, RegTree* p_tree) {
auto& tree = *p_tree;
auto& quantiles = *p_quantiles;
auto const& h_node_idx = nidx;
size_t n_leaf = collective::GlobalMax(info, h_node_idx.size());
size_t n_leaf = collective::GlobalMax(ctx, info, h_node_idx.size());
CHECK(quantiles.empty() || quantiles.size() == n_leaf);
if (quantiles.empty()) {
quantiles.resize(n_leaf, std::numeric_limits<float>::quiet_NaN());
@@ -52,12 +52,16 @@ inline void UpdateLeafValues(std::vector<float>* p_quantiles, std::vector<bst_no
std::vector<int32_t> n_valids(quantiles.size());
std::transform(quantiles.cbegin(), quantiles.cend(), n_valids.begin(),
[](float q) { return static_cast<int32_t>(!std::isnan(q)); });
collective::GlobalSum(info, &n_valids);
auto rc = collective::GlobalSum(ctx, info, linalg::MakeVec(n_valids.data(), n_valids.size()));
collective::SafeColl(rc);
// convert to 0 for all reduce
std::replace_if(
quantiles.begin(), quantiles.end(), [](float q) { return std::isnan(q); }, 0.f);
// use the mean value
collective::GlobalSum(info, &quantiles);
rc = collective::GlobalSum(ctx, info, linalg::MakeVec(quantiles.data(), quantiles.size()));
collective::SafeColl(rc);
for (size_t i = 0; i < n_leaf; ++i) {
if (n_valids[i] > 0) {
quantiles[i] /= static_cast<float>(n_valids[i]);
@@ -105,5 +109,4 @@ inline void UpdateTreeLeaf(Context const* ctx, HostDeviceVector<bst_node_t> cons
predt, alpha, p_tree);
}
}
} // namespace obj
} // namespace xgboost
} // namespace xgboost::obj

6
src/objective/quantile_obj.cu
View File

@@ -1,5 +1,5 @@
/**
* Copyright 2023 by XGBoost contributors
* Copyright 2023-2024, XGBoost contributors
*/
#include <array> // std::array
#include <cstddef> // std::size_t
@@ -170,7 +170,9 @@ class QuantileRegression : public ObjFunction {
double meanq = temp(0) * sw;
std::array<double, 2> dat{meanq, sw};
collective::GlobalSum(info, &dat);
auto rc = collective::GlobalSum(ctx_, info, linalg::MakeVec(dat.data(), dat.size()));
collective::SafeColl(rc);
std::tie(meanq, sw) = std::tuple_cat(dat);
meanq /= (sw + kRtEps);
base_score->Reshape(1);

10
src/objective/regression_obj.cu
View File

@@ -1,5 +1,5 @@
/**
* Copyright 2015-2023 by XGBoost Contributors
* Copyright 2015-2024, XGBoost Contributors
* \file regression_obj.cu
* \brief Definition of single-value regression and classification objectives.
* \author Tianqi Chen, Kailong Chen
@@ -672,8 +672,12 @@ class MeanAbsoluteError : public ObjFunction {
std::transform(linalg::cbegin(out), linalg::cend(out), linalg::begin(out),
[w](float v) { return v * w; });
collective::GlobalSum(info, &out.Values());
collective::GlobalSum(info, &w, 1);
auto rc = collective::Success() << [&] {
return collective::GlobalSum(ctx_, info, out);
} << [&] {
return collective::GlobalSum(ctx_, info, linalg::MakeVec(&w, 1));
};
collective::SafeColl(rc);
if (common::CloseTo(w, 0.0)) {
// Mostly for handling empty dataset test.