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

adding feature contributions to R and gblinear (#2295)

Browse Source 
* [gblinear] add features contribution prediction; fix DumpModel bug

* [gbtree] minor changes to PredContrib

* [R] add feature contribution prediction to R

* [R] bump up version; update NEWS

* [gblinear] fix the base_margin issue; fixes #1969

* [R] list of matrices as output of multiclass feature contributions

* [gblinear] make order of DumpModel coefficients consistent: group index changes the fastest
This commit is contained in:
Vadim Khotilovich
2017-05-21 06:41:51 -05:00
committed by Yuan (Terry) Tang
parent e5e721722e
commit b52db87d5c
10 changed files with 255 additions and 60 deletions
Download Patch File Download Diff File Expand all files Collapse all files

74
src/gbm/gblinear.cc
View File

@@ -180,10 +180,6 @@ class GBLinear : public GradientBooster {
<< "GBLinear::Predict ntrees is only valid for gbtree predictor";
std::vector<bst_float> &preds = *out_preds;
const std::vector<bst_float>& base_margin = p_fmat->info().base_margin;
if (base_margin.size() != 0) {
CHECK_EQ(preds.size(), base_margin.size())
<< "base_margin.size does not match with prediction size";
}
preds.resize(0);
// start collecting the prediction
dmlc::DataIter<RowBatch> *iter = p_fmat->RowIterator();
@@ -218,45 +214,87 @@ class GBLinear : public GradientBooster {
this->Pred(inst, dmlc::BeginPtr(*out_preds), gid, base_margin_);
}
}
void PredictLeaf(DMatrix *p_fmat,
std::vector<bst_float> *out_preds,
unsigned ntree_limit) override {
LOG(FATAL) << "gblinear does not support predict leaf index";
LOG(FATAL) << "gblinear does not support prediction of leaf index";
}
void PredictContribution(DMatrix* p_fmat,
std::vector<bst_float>* out_contribs,
unsigned ntree_limit) override {
LOG(FATAL) << "gblinear does not support predict contributions";
if (model.weight.size() == 0) {
model.InitModel();
}
CHECK_EQ(ntree_limit, 0U)
<< "GBLinear::PredictContribution: ntrees is only valid for gbtree predictor";
const std::vector<bst_float>& base_margin = p_fmat->info().base_margin;
const int ngroup = model.param.num_output_group;
const size_t ncolumns = model.param.num_feature + 1;
// allocate space for (#features + bias) times #groups times #rows
std::vector<bst_float>& contribs = *out_contribs;
contribs.resize(p_fmat->info().num_row * ncolumns * ngroup);
// make sure contributions is zeroed, we could be reusing a previously allocated one
std::fill(contribs.begin(), contribs.end(), 0);
// start collecting the contributions
dmlc::DataIter<RowBatch>* iter = p_fmat->RowIterator();
iter->BeforeFirst();
while (iter->Next()) {
const RowBatch& batch = iter->Value();
// parallel over local batch
const bst_omp_uint nsize = static_cast<bst_omp_uint>(batch.size);
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nsize; ++i) {
const RowBatch::Inst &inst = batch[i];
size_t row_idx = static_cast<size_t>(batch.base_rowid + i);
// loop over output groups
for (int gid = 0; gid < ngroup; ++gid) {
bst_float *p_contribs = &contribs[(row_idx * ngroup + gid) * ncolumns];
// calculate linear terms' contributions
for (bst_uint c = 0; c < inst.length; ++c) {
if (inst[c].index >= model.param.num_feature) continue;
p_contribs[inst[c].index] = inst[c].fvalue * model[inst[c].index][gid];
}
// add base margin to BIAS
p_contribs[ncolumns - 1] = model.bias()[gid] +
((base_margin.size() != 0) ? base_margin[row_idx * ngroup + gid] : base_margin_);
}
}
}
}
std::vector<std::string> DumpModel(const FeatureMap& fmap,
bool with_stats,
std::string format) const override {
const int ngroup = model.param.num_output_group;
const unsigned nfeature = model.param.num_feature;
std::stringstream fo("");
if (format == "json") {
fo << " { \"bias\": [" << std::endl;
for (int i = 0; i < model.param.num_output_group; ++i) {
if (i != 0) fo << "," << std::endl;
fo << " " << model.bias()[i];
for (int gid = 0; gid < ngroup; ++gid) {
if (gid != 0) fo << "," << std::endl;
fo << " " << model.bias()[gid];
}
fo << std::endl << " ]," << std::endl
<< " \"weight\": [" << std::endl;
for (int i = 0; i < model.param.num_output_group; ++i) {
for (unsigned j = 0; j < model.param.num_feature; ++j) {
if (i != 0 || j != 0) fo << "," << std::endl;
fo << " " << model[i][j];
for (unsigned i = 0; i < nfeature; ++i) {
for (int gid = 0; gid < ngroup; ++gid) {
if (i != 0 || gid != 0) fo << "," << std::endl;
fo << " " << model[i][gid];
}
}
fo << std::endl << " ]" << std::endl << " }";
} else {
fo << "bias:\n";
for (int i = 0; i < model.param.num_output_group; ++i) {
fo << model.bias()[i] << std::endl;
for (int gid = 0; gid < ngroup; ++gid) {
fo << model.bias()[gid] << std::endl;
}
fo << "weight:\n";
for (int i = 0; i < model.param.num_output_group; ++i) {
for (unsigned j = 0; j <model.param.num_feature; ++j) {
fo << model[i][j] << std::endl;
for (unsigned i = 0; i < nfeature; ++i) {
for (int gid = 0; gid < ngroup; ++gid) {
fo << model[i][gid] << std::endl;
}
}
}

11
src/gbm/gbtree.cc
View File

@@ -571,6 +571,7 @@ class GBTree : public GradientBooster {
if (ntree_limit == 0 || ntree_limit > trees.size()) {
ntree_limit = static_cast<unsigned>(trees.size());
}
const int ngroup = mparam.num_output_group;
size_t ncolumns = mparam.num_feature + 1;
// allocate space for (number of features + bias) times the number of rows
std::vector<bst_float>& contribs = *out_contribs;
@@ -584,7 +585,7 @@ class GBTree : public GradientBooster {
}
// start collecting the contributions
dmlc::DataIter<RowBatch>* iter = p_fmat->RowIterator();
const std::vector<bst_float>& base_margin = p_fmat->info().base_margin;
const std::vector<bst_float>& base_margin = info.base_margin;
iter->BeforeFirst();
while (iter->Next()) {
const RowBatch& batch = iter->Value();
@@ -596,8 +597,8 @@ class GBTree : public GradientBooster {
unsigned root_id = info.GetRoot(row_idx);
RegTree::FVec &feats = thread_temp[omp_get_thread_num()];
// loop over all classes
for (int gid = 0; gid < mparam.num_output_group; ++gid) {
bst_float *p_contribs = &contribs[(row_idx * mparam.num_output_group + gid) * ncolumns];
for (int gid = 0; gid < ngroup; ++gid) {
bst_float *p_contribs = &contribs[(row_idx * ngroup + gid) * ncolumns];
feats.Fill(batch[i]);
// calculate contributions
for (unsigned j = 0; j < ntree_limit; ++j) {
@@ -607,9 +608,9 @@ class GBTree : public GradientBooster {
trees[j]->CalculateContributions(feats, root_id, p_contribs);
}
feats.Drop(batch[i]);
// add base margin to BIAS feature
// add base margin to BIAS
if (base_margin.size() != 0) {
p_contribs[ncolumns - 1] += base_margin[row_idx * mparam.num_output_group + gid];
p_contribs[ncolumns - 1] += base_margin[row_idx * ngroup + gid];
} else {
p_contribs[ncolumns - 1] += base_margin_;
}