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

Extract interaction constraint from split evaluator. (#5034)

Browse Source 
*  Extract interaction constraints from split evaluator.

The reason for doing so is mostly for model IO, where num_feature and interaction_constraints are copied in split evaluator. Also interaction constraint by itself is a feature selector, acting like column sampler and it's inefficient to bury it deep in the evaluator chain. Lastly removing one another copied parameter is a win.

*  Enable inc for approx tree method.

As now the implementation is spited up from evaluator class, it's also enabled for approx method.

*  Removing obsoleted code in colmaker.

They are never documented nor actually used in real world. Also there isn't a single test for those code blocks.

*  Unifying the types used for row and column.

As the size of input dataset is marching to billion, incorrect use of int is subject to overflow, also singed integer overflow is undefined behaviour. This PR starts the procedure for unifying used index type to unsigned integers. There's optimization that can utilize this undefined behaviour, but after some testings I don't see the optimization is beneficial to XGBoost.
This commit is contained in:
Jiaming Yuan
2019-11-14 20:11:41 +08:00
committed by GitHub
parent 886bf93ba4
commit 97abcc7ee2
45 changed files with 688 additions and 652 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
src/tree/constraints.cu
View File

@@ -173,7 +173,7 @@ void FeatureInteractionConstraint::ClearBuffers() {
output_buffer_bits_, input_buffer_bits_);
}
common::Span<int32_t> FeatureInteractionConstraint::QueryNode(int32_t node_id) {
common::Span<bst_feature_t> FeatureInteractionConstraint::QueryNode(int32_t node_id) {
if (!has_constraint_) { return {}; }
CHECK_LT(node_id, s_node_constraints_.size());
@@ -184,7 +184,7 @@ common::Span<int32_t> FeatureInteractionConstraint::QueryNode(int32_t node_id) {
auto p_result_buffer = result_buffer_.data();
LBitField64 node_constraints = s_node_constraints_[node_id];
thrust::device_ptr<int32_t> const out_end = thrust::copy_if(
thrust::device_ptr<bst_feature_t> const out_end = thrust::copy_if(
thrust::device,
begin, end,
p_result_buffer,
@@ -197,7 +197,7 @@ common::Span<int32_t> FeatureInteractionConstraint::QueryNode(int32_t node_id) {
return {s_result_buffer_.data(), s_result_buffer_.data() + n_available};
}
__global__ void SetInputBufferKernel(common::Span<int32_t> feature_list_input,
__global__ void SetInputBufferKernel(common::Span<bst_feature_t> feature_list_input,
LBitField64 result_buffer_input) {
uint32_t tid = threadIdx.x + blockIdx.x * blockDim.x;
if (tid < feature_list_input.size()) {
@@ -212,8 +212,8 @@ __global__ void QueryFeatureListKernel(LBitField64 node_constraints,
result_buffer_output &= result_buffer_input;
}
common::Span<int32_t> FeatureInteractionConstraint::Query(
common::Span<int32_t> feature_list, int32_t nid) {
common::Span<bst_feature_t> FeatureInteractionConstraint::Query(
common::Span<bst_feature_t> feature_list, int32_t nid) {
if (!has_constraint_ || nid == 0) {
return feature_list;
}
@@ -238,7 +238,7 @@ common::Span<int32_t> FeatureInteractionConstraint::Query(
LBitField64 local_result_buffer = output_buffer_bits_;
thrust::device_ptr<int32_t> const out_end = thrust::copy_if(
thrust::device_ptr<bst_feature_t> const out_end = thrust::copy_if(
thrust::device,
begin, end,
result_buffer_.data(),
@@ -248,7 +248,7 @@ common::Span<int32_t> FeatureInteractionConstraint::Query(
});
size_t const n_available = std::distance(result_buffer_.data(), out_end);
common::Span<int32_t> result =
common::Span<bst_feature_t> result =
{s_result_buffer_.data(), s_result_buffer_.data() + n_available};
return result;
}
@@ -258,12 +258,12 @@ common::Span<int32_t> FeatureInteractionConstraint::Query(
__global__ void RestoreFeatureListFromSetsKernel(
LBitField64 feature_buffer,
int32_t fid,
bst_feature_t fid,
common::Span<int32_t> feature_interactions,
common::Span<int32_t> feature_interactions_ptr, // of size n interaction set + 1
common::Span<int32_t> interactions_list,
common::Span<int32_t> interactions_list_ptr) {
common::Span<bst_feature_t> interactions_list,
common::Span<size_t> interactions_list_ptr) {
auto const tid_x = threadIdx.x + blockIdx.x * blockDim.x;
auto const tid_y = threadIdx.y + blockIdx.y * blockDim.y;
// painful mapping: fid -> sets related to it -> features related to sets.
@@ -312,7 +312,7 @@ __global__ void InteractionConstraintSplitKernel(LBitField64 feature,
}
void FeatureInteractionConstraint::Split(
int32_t node_id, int32_t feature_id, int32_t left_id, int32_t right_id) {
bst_node_t node_id, bst_feature_t feature_id, bst_node_t left_id, bst_node_t right_id) {
if (!has_constraint_) { return; }
CHECK_NE(node_id, left_id)
<< " Split node: " << node_id << " and its left child: "