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

Use DART tree weights when computing SHAPs (#5050)

Browse Source 
This PR fixes tree weights in dart being ignored when computing contributions.

* Fix ellpack page source link.
* Add tree weights to compute contribution.
This commit is contained in:
Kodi Arfer
2019-12-03 06:55:53 -05:00
committed by Jiaming Yuan
parent 64f4361b47
commit f2277e7106
12 changed files with 88 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/data/ellpack_page.cc
View File

@@ -10,6 +10,8 @@ namespace xgboost {
class EllpackPageImpl {};
EllpackPage::EllpackPage() = default;
EllpackPage::EllpackPage(DMatrix* dmat, const BatchParam& param) {
LOG(FATAL) << "Internal Error: XGBoost is not compiled with CUDA but EllpackPage is required";
}

2
src/data/ellpack_page_source.cc
View File

@@ -3,8 +3,8 @@
*/
#ifndef XGBOOST_USE_CUDA
#include <xgboost/data.h>
#include "ellpack_page_source.h"
namespace xgboost {
namespace data {

6
src/data/ellpack_page_source.cu
View File

@@ -1,14 +1,14 @@
/*!
* Copyright 2019 XGBoost contributors
*/
#include "ellpack_page_source.h"
#include <memory>
#include <utility>
#include <vector>
#include "../common/hist_util.h"
#include "ellpack_page_source.h"
#include "sparse_page_source.h"
#include "ellpack_page.cuh"
namespace xgboost {

1
src/data/ellpack_page_source.h
View File

@@ -9,7 +9,6 @@
#include <memory>
#include <string>
#include "sparse_page_source.h"
#include "../common/timer.h"
namespace xgboost {

17
src/gbm/gbtree.cc
View File

@@ -348,6 +348,23 @@ class Dart : public GBTree {
return GBTree::UseGPU();
}
void PredictContribution(DMatrix* p_fmat,
std::vector<bst_float>* out_contribs,
unsigned ntree_limit, bool approximate, int condition,
unsigned condition_feature) override {
CHECK(configured_);
cpu_predictor_->PredictContribution(p_fmat, out_contribs, model_,
ntree_limit, &weight_drop_, approximate);
}
void PredictInteractionContributions(DMatrix* p_fmat,
std::vector<bst_float>* out_contribs,
unsigned ntree_limit, bool approximate) override {
CHECK(configured_);
cpu_predictor_->PredictInteractionContributions(p_fmat, out_contribs, model_,
ntree_limit, &weight_drop_, approximate);
}
protected:
friend class GBTree;
// internal prediction loop

5
src/gbm/gbtree.h
View File

@@ -216,7 +216,8 @@ class GBTree : public GradientBooster {
unsigned ntree_limit, bool approximate, int condition,
unsigned condition_feature) override {
CHECK(configured_);
cpu_predictor_->PredictContribution(p_fmat, out_contribs, model_, ntree_limit, approximate);
cpu_predictor_->PredictContribution(p_fmat, out_contribs, model_,
ntree_limit, nullptr, approximate);
}
void PredictInteractionContributions(DMatrix* p_fmat,
@@ -224,7 +225,7 @@ class GBTree : public GradientBooster {
unsigned ntree_limit, bool approximate) override {
CHECK(configured_);
cpu_predictor_->PredictInteractionContributions(p_fmat, out_contribs, model_,
ntree_limit, approximate);
ntree_limit, nullptr, approximate);
}
std::vector<std::string> DumpModel(const FeatureMap& fmap,

22
src/predictor/cpu_predictor.cc
View File

@@ -257,6 +257,7 @@ class CPUPredictor : public Predictor {
void PredictContribution(DMatrix* p_fmat, std::vector<bst_float>* out_contribs,
const gbm::GBTreeModel& model, unsigned ntree_limit,
std::vector<bst_float>* tree_weights,
bool approximate,
int condition,
unsigned condition_feature) override {
@@ -296,16 +297,23 @@ class CPUPredictor : public Predictor {
bst_float* p_contribs =
&contribs[(row_idx * ngroup + gid) * ncolumns];
feats.Fill(batch[i]);
std::vector<bst_float> this_tree_contribs;
this_tree_contribs.resize(ncolumns);
// calculate contributions
for (unsigned j = 0; j < ntree_limit; ++j) {
std::fill(this_tree_contribs.begin(), this_tree_contribs.end(), 0);
if (model.tree_info[j] != gid) {
continue;
}
if (!approximate) {
model.trees[j]->CalculateContributions(feats, root_id, p_contribs,
model.trees[j]->CalculateContributions(feats, root_id, &this_tree_contribs[0],
condition, condition_feature);
} else {
model.trees[j]->CalculateContributionsApprox(feats, root_id, p_contribs);
model.trees[j]->CalculateContributionsApprox(feats, root_id, &this_tree_contribs[0]);
}
for (int ci = 0 ; ci < ncolumns ; ++ci) {
p_contribs[ci] += this_tree_contribs[ci] *
(tree_weights == nullptr ? 1 : (*tree_weights)[j]);
}
}
feats.Drop(batch[i]);
@@ -322,6 +330,7 @@ class CPUPredictor : public Predictor {
void PredictInteractionContributions(DMatrix* p_fmat, std::vector<bst_float>* out_contribs,
const gbm::GBTreeModel& model, unsigned ntree_limit,
std::vector<bst_float>* tree_weights,
bool approximate) override {
const MetaInfo& info = p_fmat->Info();
const int ngroup = model.param.num_output_group;
@@ -340,10 +349,13 @@ class CPUPredictor : public Predictor {
// Compute the difference in effects when conditioning on each of the features on and off
// see: Axiomatic characterizations of probabilistic and
// cardinal-probabilistic interaction indices
PredictContribution(p_fmat, &contribs_diag, model, ntree_limit, approximate, 0, 0);
PredictContribution(p_fmat, &contribs_diag, model, ntree_limit,
tree_weights, approximate, 0, 0);
for (size_t i = 0; i < ncolumns + 1; ++i) {
PredictContribution(p_fmat, &contribs_off, model, ntree_limit, approximate, -1, i);
PredictContribution(p_fmat, &contribs_on, model, ntree_limit, approximate, 1, i);
PredictContribution(p_fmat, &contribs_off, model, ntree_limit,
tree_weights, approximate, -1, i);
PredictContribution(p_fmat, &contribs_on, model, ntree_limit,
tree_weights, approximate, 1, i);
for (size_t j = 0; j < info.num_row_; ++j) {
for (int l = 0; l < ngroup; ++l) {

2
src/predictor/gpu_predictor.cu
View File

@@ -397,6 +397,7 @@ class GPUPredictor : public xgboost::Predictor {
void PredictContribution(DMatrix* p_fmat,
std::vector<bst_float>* out_contribs,
const gbm::GBTreeModel& model, unsigned ntree_limit,
std::vector<bst_float>* tree_weights,
bool approximate, int condition,
unsigned condition_feature) override {
LOG(FATAL) << "Internal error: " << __func__
@@ -407,6 +408,7 @@ class GPUPredictor : public xgboost::Predictor {
std::vector<bst_float>* out_contribs,
const gbm::GBTreeModel& model,
unsigned ntree_limit,
std::vector<bst_float>* tree_weights,
bool approximate) override {
LOG(FATAL) << "Internal error: " << __func__
<< " is not implemented in GPU Predictor.";