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

Make objectives work with vertical distributed and federated learning (#9002)

Browse Source
This commit is contained in:
Rong Ou 2023-04-03 02:07:42 -07:00 committed by GitHub
parent 720a8c3273
commit 15e073ca9d
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
7 changed files with 199 additions and 111 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
src/objective/adaptive.cc
View File

@ -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, learning_rate, p_tree);
UpdateLeafValues(&quantiles, nidx, info, learning_rate, p_tree);
return;
}
@ -99,6 +99,7 @@ void UpdateTreeLeafHost(Context const* ctx, std::vector<bst_node_t> const& posit
auto h_predt = linalg::MakeTensorView(ctx, predt.ConstHostSpan(), info.num_row_,
predt.Size() / info.num_row_);
if (!info.IsVerticalFederated() || collective::GetRank() == 0) {
// loop over each leaf
common::ParallelFor(quantiles.size(), ctx->Threads(), [&](size_t k) {
auto nidx = h_node_idx[k];
@ -130,8 +131,14 @@ void UpdateTreeLeafHost(Context const* ctx, std::vector<bst_node_t> const& posit
}
quantiles.at(k) = q;
});
}
UpdateLeafValues(&quantiles, nidx, learning_rate, p_tree);
if (info.IsVerticalFederated()) {
collective::Broadcast(static_cast<void*>(quantiles.data()), quantiles.size() * sizeof(float),
0);
}
UpdateLeafValues(&quantiles, nidx, info, learning_rate, p_tree);
}
#if !defined(XGBOOST_USE_CUDA)

4
src/objective/adaptive.cu
View File

@ -151,7 +151,7 @@ void UpdateTreeLeafDevice(Context const* ctx, common::Span<bst_node_t const> pos
if (nptr.Empty()) {
std::vector<float> quantiles;
UpdateLeafValues(&quantiles, nidx.ConstHostVector(), learning_rate, p_tree);
UpdateLeafValues(&quantiles, nidx.ConstHostVector(), info, learning_rate, p_tree);
}
HostDeviceVector<float> quantiles;
@ -186,7 +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(), learning_rate, p_tree);
UpdateLeafValues(&quantiles.HostVector(), nidx.ConstHostVector(), info, learning_rate, p_tree);
}
} // namespace detail
} // namespace obj

8
src/objective/adaptive.h
View File

@ -36,13 +36,15 @@ 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,
float learning_rate, RegTree* p_tree) {
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{h_node_idx.size()};
if (info.IsRowSplit()) {
collective::Allreduce<collective::Operation::kMax>(&n_leaf, 1);
}
CHECK(quantiles.empty() || quantiles.size() == n_leaf);
if (quantiles.empty()) {
quantiles.resize(n_leaf, std::numeric_limits<float>::quiet_NaN());
@ -52,12 +54,16 @@ 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)); });
if (info.IsRowSplit()) {
collective::Allreduce<collective::Operation::kSum>(n_valids.data(), n_valids.size());
}
// 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
if (info.IsRowSplit()) {
collective::Allreduce<collective::Operation::kSum>(quantiles.data(), quantiles.size());
}
for (size_t i = 0; i < n_leaf; ++i) {
if (n_valids[i] > 0) {
quantiles[i] /= static_cast<float>(n_valids[i]);

7
src/objective/quantile_obj.cu
View File

@ -35,7 +35,10 @@ class QuantileRegression : public ObjFunction {
bst_target_t Targets(MetaInfo const& info) const override {
auto const& alpha = param_.quantile_alpha.Get();
CHECK_EQ(alpha.size(), alpha_.Size()) << "The objective is not yet configured.";
CHECK_EQ(info.labels.Shape(1), 1) << "Multi-target is not yet supported by the quantile loss.";
if (!info.IsVerticalFederated() || collective::GetRank() == 0) {
CHECK_EQ(info.labels.Shape(1), 1)
<< "Multi-target is not yet supported by the quantile loss.";
}
CHECK(!alpha.empty());
// We have some placeholders for multi-target in the quantile loss. But it's not
// supported as the gbtree doesn't know how to slice the gradient and there's no 3-dim
@ -167,8 +170,10 @@ class QuantileRegression : public ObjFunction {
common::Mean(ctx_, *base_score, &temp);
double meanq = temp(0) * sw;
if (info.IsRowSplit()) {
collective::Allreduce<collective::Operation::kSum>(&meanq, 1);
collective::Allreduce<collective::Operation::kSum>(&sw, 1);
}
meanq /= (sw + kRtEps);
base_score->Reshape(1);
base_score->Data()->Fill(meanq);

2
src/objective/regression_obj.cu
View File

@ -728,8 +728,10 @@ class MeanAbsoluteError : public ObjFunction {
std::transform(linalg::cbegin(out), linalg::cend(out), linalg::begin(out),
[w](float v) { return v * w; });
if (info.IsRowSplit()) {
collective::Allreduce<collective::Operation::kSum>(out.Values().data(), out.Values().size());
collective::Allreduce<collective::Operation::kSum>(&w, 1);
}
if (common::CloseTo(w, 0.0)) {
// Mostly for handling empty dataset test.

119
tests/cpp/plugin/test_federated_learner.cc
View File

@ -13,66 +13,91 @@
namespace xgboost {
void VerifyObjectives(size_t rows, size_t cols, std::vector<float> const &expected_base_scores,
std::vector<Json> const &expected_models) {
auto const world_size = collective::GetWorldSize();
auto const rank = collective::GetRank();
std::shared_ptr<DMatrix> dmat{RandomDataGenerator{rows, cols, 0}.GenerateDMatrix(rank == 0)};
if (rank == 0) {
auto &h_upper = dmat->Info().labels_upper_bound_.HostVector();
auto &h_lower = dmat->Info().labels_lower_bound_.HostVector();
h_lower.resize(rows);
h_upper.resize(rows);
for (size_t i = 0; i < rows; ++i) {
h_lower[i] = 1;
h_upper[i] = 10;
}
}
std::shared_ptr<DMatrix> sliced{dmat->SliceCol(world_size, rank)};
auto i = 0;
for (auto const *entry : ::dmlc::Registry<::xgboost::ObjFunctionReg>::List()) {
std::unique_ptr<Learner> learner{Learner::Create({sliced})};
learner->SetParam("tree_method", "approx");
learner->SetParam("objective", entry->name);
if (entry->name.find("quantile") != std::string::npos) {
learner->SetParam("quantile_alpha", "0.5");
}
if (entry->name.find("multi") != std::string::npos) {
learner->SetParam("num_class", "3");
}
learner->UpdateOneIter(0, sliced);
Json config{Object{}};
learner->SaveConfig(&config);
auto base_score = GetBaseScore(config);
ASSERT_EQ(base_score, expected_base_scores[i]);
Json model{Object{}};
learner->SaveModel(&model);
ASSERT_EQ(model, expected_models[i]);
i++;
}
}
class FederatedLearnerTest : public BaseFederatedTest {
protected:
static auto constexpr kRows{16};
static auto constexpr kCols{16};
};
void VerifyBaseScore(size_t rows, size_t cols, float expected_base_score) {
auto const world_size = collective::GetWorldSize();
auto const rank = collective::GetRank();
std::shared_ptr<DMatrix> Xy_{RandomDataGenerator{rows, cols, 0}.GenerateDMatrix(rank == 0)};
std::shared_ptr<DMatrix> sliced{Xy_->SliceCol(world_size, rank)};
std::unique_ptr<Learner> learner{Learner::Create({sliced})};
TEST_F(FederatedLearnerTest, Objectives) {
std::shared_ptr<DMatrix> dmat{RandomDataGenerator{kRows, kCols, 0}.GenerateDMatrix(true)};
auto &h_upper = dmat->Info().labels_upper_bound_.HostVector();
auto &h_lower = dmat->Info().labels_lower_bound_.HostVector();
h_lower.resize(kRows);
h_upper.resize(kRows);
for (size_t i = 0; i < kRows; ++i) {
h_lower[i] = 1;
h_upper[i] = 10;
}
std::vector<float> base_scores;
std::vector<Json> models;
for (auto const *entry : ::dmlc::Registry<::xgboost::ObjFunctionReg>::List()) {
std::unique_ptr<Learner> learner{Learner::Create({dmat})};
learner->SetParam("tree_method", "approx");
learner->SetParam("objective", "binary:logistic");
learner->UpdateOneIter(0, sliced);
learner->SetParam("objective", entry->name);
if (entry->name.find("quantile") != std::string::npos) {
learner->SetParam("quantile_alpha", "0.5");
}
if (entry->name.find("multi") != std::string::npos) {
learner->SetParam("num_class", "3");
}
learner->UpdateOneIter(0, dmat);
Json config{Object{}};
learner->SaveConfig(&config);
auto base_score = GetBaseScore(config);
ASSERT_EQ(base_score, expected_base_score);
}
base_scores.emplace_back(GetBaseScore(config));
void VerifyModel(size_t rows, size_t cols, Json const& expected_model) {
auto const world_size = collective::GetWorldSize();
auto const rank = collective::GetRank();
std::shared_ptr<DMatrix> Xy_{RandomDataGenerator{rows, cols, 0}.GenerateDMatrix(rank == 0)};
std::shared_ptr<DMatrix> sliced{Xy_->SliceCol(world_size, rank)};
std::unique_ptr<Learner> learner{Learner::Create({sliced})};
learner->SetParam("tree_method", "approx");
learner->SetParam("objective", "binary:logistic");
learner->UpdateOneIter(0, sliced);
Json model{Object{}};
learner->SaveModel(&model);
ASSERT_EQ(model, expected_model);
models.emplace_back(model);
}
TEST_F(FederatedLearnerTest, BaseScore) {
std::shared_ptr<DMatrix> Xy_{RandomDataGenerator{kRows, kCols, 0}.GenerateDMatrix(true)};
std::unique_ptr<Learner> learner{Learner::Create({Xy_})};
learner->SetParam("tree_method", "approx");
learner->SetParam("objective", "binary:logistic");
learner->UpdateOneIter(0, Xy_);
Json config{Object{}};
learner->SaveConfig(&config);
auto base_score = GetBaseScore(config);
ASSERT_NE(base_score, ObjFunction::DefaultBaseScore());
RunWithFederatedCommunicator(kWorldSize, server_address_, &VerifyBaseScore, kRows, kCols,
base_score);
}
TEST_F(FederatedLearnerTest, Model) {
std::shared_ptr<DMatrix> Xy_{RandomDataGenerator{kRows, kCols, 0}.GenerateDMatrix(true)};
std::unique_ptr<Learner> learner{Learner::Create({Xy_})};
learner->SetParam("tree_method", "approx");
learner->SetParam("objective", "binary:logistic");
learner->UpdateOneIter(0, Xy_);
Json model{Object{}};
learner->SaveModel(&model);
RunWithFederatedCommunicator(kWorldSize, server_address_, &VerifyModel, kRows, kCols,
std::cref(model));
RunWithFederatedCommunicator(kWorldSize, server_address_, &VerifyObjectives, kRows, kCols,
base_scores, models);
}
} // namespace xgboost

65
tests/cpp/test_learner.cc
View File

@ -608,31 +608,74 @@ TEST_F(InitBaseScore, InitWithPredict) { this->TestInitWithPredt(); }
TEST_F(InitBaseScore, UpdateProcess) { this->TestUpdateProcess(); }
void TestColumnSplitBaseScore(std::shared_ptr<DMatrix> Xy_, float expected_base_score) {
void TestColumnSplit(std::shared_ptr<DMatrix> dmat, std::vector<float> const& expected_base_scores,
std::vector<Json> const& expected_models) {
auto const world_size = collective::GetWorldSize();
auto const rank = collective::GetRank();
std::shared_ptr<DMatrix> sliced{Xy_->SliceCol(world_size, rank)};
std::shared_ptr<DMatrix> sliced{dmat->SliceCol(world_size, rank)};
auto i = 0;
for (auto const* entry : ::dmlc::Registry<::xgboost::ObjFunctionReg>::List()) {
std::unique_ptr<Learner> learner{Learner::Create({sliced})};
learner->SetParam("tree_method", "approx");
learner->SetParam("objective", "binary:logistic");
learner->SetParam("objective", entry->name);
if (entry->name.find("quantile") != std::string::npos) {
learner->SetParam("quantile_alpha", "0.5");
}
if (entry->name.find("multi") != std::string::npos) {
learner->SetParam("num_class", "3");
}
learner->UpdateOneIter(0, sliced);
Json config{Object{}};
learner->SaveConfig(&config);
auto base_score = GetBaseScore(config);
ASSERT_EQ(base_score, expected_base_score);
ASSERT_EQ(base_score, expected_base_scores[i]);
Json model{Object{}};
learner->SaveModel(&model);
ASSERT_EQ(model, expected_models[i]);
i++;
}
}
TEST_F(InitBaseScore, ColumnSplit) {
std::unique_ptr<Learner> learner{Learner::Create({Xy_})};
TEST(ColumnSplit, Objectives) {
auto constexpr kRows = 10, kCols = 10;
std::shared_ptr<DMatrix> dmat{RandomDataGenerator{kRows, kCols, 0}.GenerateDMatrix(true)};
auto& h_upper = dmat->Info().labels_upper_bound_.HostVector();
auto& h_lower = dmat->Info().labels_lower_bound_.HostVector();
h_lower.resize(kRows);
h_upper.resize(kRows);
for (size_t i = 0; i < kRows; ++i) {
h_lower[i] = 1;
h_upper[i] = 10;
}
std::vector<float> base_scores;
std::vector<Json> models;
for (auto const* entry : ::dmlc::Registry<::xgboost::ObjFunctionReg>::List()) {
std::unique_ptr<Learner> learner{Learner::Create({dmat})};
learner->SetParam("tree_method", "approx");
learner->SetParam("objective", "binary:logistic");
learner->UpdateOneIter(0, Xy_);
learner->SetParam("objective", entry->name);
if (entry->name.find("quantile") != std::string::npos) {
learner->SetParam("quantile_alpha", "0.5");
}
if (entry->name.find("multi") != std::string::npos) {
learner->SetParam("num_class", "3");
}
learner->UpdateOneIter(0, dmat);
Json config{Object{}};
learner->SaveConfig(&config);
auto base_score = GetBaseScore(config);
ASSERT_NE(base_score, ObjFunction::DefaultBaseScore());
base_scores.emplace_back(GetBaseScore(config));
Json model{Object{}};
learner->SaveModel(&model);
models.emplace_back(model);
}
auto constexpr kWorldSize{3};
RunWithInMemoryCommunicator(kWorldSize, &TestColumnSplitBaseScore, Xy_, base_score);
RunWithInMemoryCommunicator(kWorldSize, &TestColumnSplit, dmat, base_scores, models);
}
} // namespace xgboost