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
7 changed files with 199 additions and 111 deletions
Download Patch File Download Diff File Expand all files Collapse all files

65
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;
}
@@ -99,39 +99,46 @@ 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_);
// loop over each leaf
common::ParallelFor(quantiles.size(), ctx->Threads(), [&](size_t k) {
auto nidx = h_node_idx[k];
CHECK(tree[nidx].IsLeaf());
CHECK_LT(k + 1, h_node_ptr.size());
size_t n = h_node_ptr[k + 1] - h_node_ptr[k];
auto h_row_set = common::Span<size_t const>{ridx}.subspan(h_node_ptr[k], n);
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];
CHECK(tree[nidx].IsLeaf());
CHECK_LT(k + 1, h_node_ptr.size());
size_t n = h_node_ptr[k + 1] - h_node_ptr[k];
auto h_row_set = common::Span<size_t const>{ridx}.subspan(h_node_ptr[k], n);
auto h_labels = info.labels.HostView().Slice(linalg::All(), IdxY(info, group_idx));
auto h_weights = linalg::MakeVec(&info.weights_);
auto h_labels = info.labels.HostView().Slice(linalg::All(), IdxY(info, group_idx));
auto h_weights = linalg::MakeVec(&info.weights_);
auto iter = common::MakeIndexTransformIter([&](size_t i) -> float {
auto row_idx = h_row_set[i];
return h_labels(row_idx) - h_predt(row_idx, group_idx);
});
auto w_it = common::MakeIndexTransformIter([&](size_t i) -> float {
auto row_idx = h_row_set[i];
return h_weights(row_idx);
auto iter = common::MakeIndexTransformIter([&](size_t i) -> float {
auto row_idx = h_row_set[i];
return h_labels(row_idx) - h_predt(row_idx, group_idx);
});
auto w_it = common::MakeIndexTransformIter([&](size_t i) -> float {
auto row_idx = h_row_set[i];
return h_weights(row_idx);
});
float q{0};
if (info.weights_.Empty()) {
q = common::Quantile(ctx, alpha, iter, iter + h_row_set.size());
} else {
q = common::WeightedQuantile(ctx, alpha, iter, iter + h_row_set.size(), w_it);
}
if (std::isnan(q)) {
CHECK(h_row_set.empty());
}
quantiles.at(k) = q;
});
}
float q{0};
if (info.weights_.Empty()) {
q = common::Quantile(ctx, alpha, iter, iter + h_row_set.size());
} else {
q = common::WeightedQuantile(ctx, alpha, iter, iter + h_row_set.size(), w_it);
}
if (std::isnan(q)) {
CHECK(h_row_set.empty());
}
quantiles.at(k) = q;
});
if (info.IsVerticalFederated()) {
collective::Broadcast(static_cast<void*>(quantiles.data()), quantiles.size() * sizeof(float),
0);
}
UpdateLeafValues(&quantiles, nidx, learning_rate, p_tree);
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

14
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()};
collective::Allreduce<collective::Operation::kMax>(&n_leaf, 1);
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)); });
collective::Allreduce<collective::Operation::kSum>(n_valids.data(), n_valids.size());
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
collective::Allreduce<collective::Operation::kSum>(quantiles.data(), quantiles.size());
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]);

11
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;
collective::Allreduce<collective::Operation::kSum>(&meanq, 1);
collective::Allreduce<collective::Operation::kSum>(&sw, 1);
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);

6
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; });
collective::Allreduce<collective::Operation::kSum>(out.Values().data(), out.Values().size());
collective::Allreduce<collective::Operation::kSum>(&w, 1);
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.