Initial support for column-wise data split (#8468)

include/xgboost/data.h

@@ -112,6 +112,9 @@ class MetaInfo {
   void Validate(int32_t device) const;
 
   MetaInfo Slice(common::Span<int32_t const> ridxs) const;
+
+  MetaInfo Copy() const;
+
   /*!
    * \brief Get weight of each instances.
    * \param i Instance index.
@@ -620,6 +623,15 @@ class DMatrix {
 
   virtual DMatrix *Slice(common::Span<int32_t const> ridxs) = 0;
 
+  /**
+   * \brief Slice a DMatrix by columns.
+   *
+   * @param start The position of the first column
+   * @param size The number of columns in the slice
+   * @return DMatrix containing the slice of columns
+   */
+  virtual DMatrix *SliceCol(std::size_t start, std::size_t size) = 0;
+
  protected:
   virtual BatchSet<SparsePage> GetRowBatches() = 0;
   virtual BatchSet<CSCPage> GetColumnBatches() = 0;

src/data/data.cc

@@ -381,6 +381,12 @@ MetaInfo MetaInfo::Slice(common::Span<int32_t const> ridxs) const {
   return out;
 }
 
+MetaInfo MetaInfo::Copy() const {
+  MetaInfo out;
+  out.Extend(*this, /*accumulate_rows=*/true, /*check_column=*/false);
+  return out;
+}
+
 namespace {
 template <int32_t D, typename T>
 void CopyTensorInfoImpl(Context const& ctx, Json arr_interface, linalg::Tensor<T, D>* p_out) {
@@ -777,8 +783,10 @@ DMatrix *TryLoadBinary(std::string fname, bool silent) {
 
 DMatrix* DMatrix::Load(const std::string& uri, bool silent, DataSplitMode data_split_mode,
                        const std::string& file_format) {
-  CHECK(data_split_mode == DataSplitMode::kRow || data_split_mode == DataSplitMode::kNone)
+  CHECK(data_split_mode == DataSplitMode::kRow ||
-      << "Precondition violated; data split mode can only be 'row' or 'none'";
+        data_split_mode == DataSplitMode::kCol ||
+        data_split_mode == DataSplitMode::kNone)
+      << "Precondition violated; data split mode can only be 'row', 'col', or 'none'";
   std::string fname, cache_file;
   size_t dlm_pos = uri.find('#');
   if (dlm_pos != std::string::npos) {
@@ -878,7 +886,20 @@ DMatrix* DMatrix::Load(const std::string& uri, bool silent, DataSplitMode data_s
    * partitioned data will fail the train/val validation check
    * since partitioned data not knowing the real number of features. */
   collective::Allreduce<collective::Operation::kMax>(&dmat->Info().num_col_, 1);
-  return dmat;
+
+  if (data_split_mode == DataSplitMode::kCol) {
+    if (!cache_file.empty()) {
+      LOG(FATAL) << "Column-wise data split is not support for external memory.";
+    }
+    auto slice_cols = (dmat->Info().num_col_ + 1) / npart;
+    auto slice_start = slice_cols * partid;
+    auto size = std::min(slice_cols, dmat->Info().num_col_ - slice_start);
+    auto* sliced = dmat->SliceCol(slice_start, size);
+    delete dmat;
+    return sliced;
+  } else {
+    return dmat;
+  }
 }
 
 template <typename DataIterHandle, typename DMatrixHandle, typename DataIterResetCallback,

src/data/iterative_dmatrix.h

@@ -86,6 +86,10 @@ class IterativeDMatrix : public DMatrix {
     LOG(FATAL) << "Slicing DMatrix is not supported for Quantile DMatrix.";
     return nullptr;
   }
+  DMatrix *SliceCol(std::size_t start, std::size_t size) override {
+    LOG(FATAL) << "Slicing DMatrix columns is not supported for Quantile DMatrix.";
+    return nullptr;
+  }
   BatchSet<SparsePage> GetRowBatches() override {
     LOG(FATAL) << "Not implemented.";
     return BatchSet<SparsePage>(BatchIterator<SparsePage>(nullptr));

src/data/proxy_dmatrix.h

@@ -87,6 +87,10 @@ class DMatrixProxy : public DMatrix {
     LOG(FATAL) << "Slicing DMatrix is not supported for Proxy DMatrix.";
     return nullptr;
   }
+  DMatrix* SliceCol(std::size_t start, std::size_t size) override {
+    LOG(FATAL) << "Slicing DMatrix columns is not supported for Proxy DMatrix.";
+    return nullptr;
+  }
   BatchSet<SparsePage> GetRowBatches() override {
     LOG(FATAL) << "Not implemented.";
     return BatchSet<SparsePage>(BatchIterator<SparsePage>(nullptr));

src/data/simple_dmatrix.cc

@@ -45,6 +45,29 @@ DMatrix* SimpleDMatrix::Slice(common::Span<int32_t const> ridxs) {
   return out;
 }
 
+DMatrix* SimpleDMatrix::SliceCol(std::size_t start, std::size_t size) {
+  auto out = new SimpleDMatrix;
+  SparsePage& out_page = *out->sparse_page_;
+  for (auto const &page : this->GetBatches<SparsePage>()) {
+    auto batch = page.GetView();
+    auto& h_data = out_page.data.HostVector();
+    auto& h_offset = out_page.offset.HostVector();
+    size_t rptr{0};
+    for (auto i = 0; i < this->Info().num_row_; i++) {
+      auto inst = batch[i];
+      auto prev_size = h_data.size();
+      std::copy_if(inst.begin(), inst.end(), std::back_inserter(h_data), [&](Entry e) {
+        return e.index >= start && e.index < start + size;
+      });
+      rptr += h_data.size() - prev_size;
+      h_offset.emplace_back(rptr);
+    }
+    out->Info() = this->Info().Copy();
+    out->Info().num_nonzero_ = h_offset.back();
+  }
+  return out;
+}
+
 BatchSet<SparsePage> SimpleDMatrix::GetRowBatches() {
   // since csr is the default data structure so `source_` is always available.
   auto begin_iter = BatchIterator<SparsePage>(

src/data/simple_dmatrix.h

@@ -35,6 +35,7 @@ class SimpleDMatrix : public DMatrix {
 
   bool SingleColBlock() const override { return true; }
   DMatrix* Slice(common::Span<int32_t const> ridxs) override;
+  DMatrix* SliceCol(std::size_t start, std::size_t size) override;
 
   /*! \brief magic number used to identify SimpleDMatrix binary files */
   static const int kMagic = 0xffffab01;

src/data/sparse_page_dmatrix.h

@@ -107,6 +107,10 @@ class SparsePageDMatrix : public DMatrix {
     LOG(FATAL) << "Slicing DMatrix is not supported for external memory.";
     return nullptr;
   }
+  DMatrix *SliceCol(std::size_t start, std::size_t size) override {
+    LOG(FATAL) << "Slicing DMatrix columns is not supported for external memory.";
+    return nullptr;
+  }
 
  private:
   BatchSet<SparsePage> GetRowBatches() override;

tests/cpp/data/test_simple_dmatrix.cc

@@ -300,6 +300,69 @@ TEST(SimpleDMatrix, Slice) {
   ASSERT_EQ(out->Info().num_nonzero_, ridxs.size() * kCols);  // dense
 }
 
+TEST(SimpleDMatrix, SliceCol) {
+  size_t constexpr kRows {16};
+  size_t constexpr kCols {8};
+  size_t constexpr kClasses {3};
+  auto p_m = RandomDataGenerator{kRows, kCols, 0}.GenerateDMatrix(true);
+  auto& weights = p_m->Info().weights_.HostVector();
+  weights.resize(kRows);
+  std::iota(weights.begin(), weights.end(), 0.0f);
+
+  auto& lower = p_m->Info().labels_lower_bound_.HostVector();
+  auto& upper = p_m->Info().labels_upper_bound_.HostVector();
+  lower.resize(kRows);
+  upper.resize(kRows);
+
+  std::iota(lower.begin(), lower.end(), 0.0f);
+  std::iota(upper.begin(), upper.end(), 1.0f);
+
+  auto& margin = p_m->Info().base_margin_;
+  margin = decltype(p_m->Info().base_margin_){{kRows, kClasses}, GenericParameter::kCpuId};
+
+  size_t constexpr kSlicCols {4};
+  for (auto slice = 0; slice < 2; slice++) {
+    auto const slice_start = slice * kSlicCols;
+    std::unique_ptr<DMatrix> out { p_m->SliceCol(slice_start, kSlicCols) };
+    ASSERT_EQ(out->Info().labels.Size(), kRows);
+    ASSERT_EQ(out->Info().labels_lower_bound_.Size(), kRows);
+    ASSERT_EQ(out->Info().labels_upper_bound_.Size(), kRows);
+    ASSERT_EQ(out->Info().base_margin_.Size(), kRows * kClasses);
+
+    for (auto const &in_batch : p_m->GetBatches<SparsePage>()) {
+      auto in_page = in_batch.GetView();
+      for (auto const &out_batch : out->GetBatches<SparsePage>()) {
+        auto out_page = out_batch.GetView();
+        for (size_t i = 0; i < kRows; ++i) {
+          auto out_inst = out_page[i];
+          auto in_inst = in_page[i];
+          ASSERT_EQ(out_inst.size() * 2, in_inst.size()) << i;
+          for (size_t j = 0; j < kSlicCols; ++j) {
+            ASSERT_EQ(in_inst[slice_start + j].fvalue, out_inst[j].fvalue);
+            ASSERT_EQ(in_inst[slice_start + j].index, out_inst[j].index);
+          }
+
+          ASSERT_EQ(p_m->Info().labels_lower_bound_.HostVector().at(i),
+                    out->Info().labels_lower_bound_.HostVector().at(i));
+          ASSERT_EQ(p_m->Info().labels_upper_bound_.HostVector().at(i),
+                    out->Info().labels_upper_bound_.HostVector().at(i));
+          ASSERT_EQ(p_m->Info().weights_.HostVector().at(i), out->Info().weights_.HostVector().at(i));
+
+          auto out_margin = out->Info().base_margin_.View(GenericParameter::kCpuId);
+          auto in_margin = margin.View(GenericParameter::kCpuId);
+          for (size_t j = 0; j < kClasses; ++j) {
+            ASSERT_EQ(out_margin(i, j), in_margin(i, j));
+          }
+        }
+      }
+    }
+
+    ASSERT_EQ(out->Info().num_col_, out->Info().num_col_);
+    ASSERT_EQ(out->Info().num_row_, kRows);
+    ASSERT_EQ(out->Info().num_nonzero_, kRows * kSlicCols);  // dense
+  }
+}
+
 TEST(SimpleDMatrix, SaveLoadBinary) {
   dmlc::TemporaryDirectory tempdir;
   const std::string tmp_file = tempdir.path + "/simple.libsvm";