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

Use Predictor for dart. (#6693)

Browse Source 
* Use normal predictor for dart booster.
* Implement `inplace_predict` for dart.
* Enable `dart` for dask interface now that it's thread-safe.
* categorical data should be working out of box for dart now.

The implementation is not very efficient as it has to pull back the data and
apply weight for each tree, but still a significant improvement over previous
implementation as now we no longer binary search for each sample.

* Fix output prediction shape on dataframe.
This commit is contained in:
Jiaming Yuan
2021-02-09 23:30:19 +08:00
committed by GitHub
parent dbf7e9d3cb
commit e8c5c53e2f
13 changed files with 246 additions and 180 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
src/predictor/gpu_predictor.cu
View File

@@ -599,21 +599,9 @@ class GPUPredictor : public xgboost::Predictor {
int device = generic_param_->gpu_id;
CHECK_GE(device, 0) << "Set `gpu_id' to positive value for processing GPU data.";
auto* out_preds = &predts->predictions;
if (out_preds->Size() == 0 && dmat->Info().num_row_ != 0) {
CHECK_EQ(predts->version, 0);
}
if (tree_end == 0) {
tree_end = model.trees.size();
}
if (predts->version == 0) {
// out_preds->Size() can be non-zero as it's initialized here before any tree is
// built at the 0^th iterator.
this->InitOutPredictions(dmat->Info(), out_preds, model);
}
if (tree_end - tree_begin == 0) {
return;
}
this->DevicePredictInternal(dmat, out_preds, model, tree_begin, tree_end);
}
@@ -788,7 +776,7 @@ class GPUPredictor : public xgboost::Predictor {
protected:
void InitOutPredictions(const MetaInfo& info,
HostDeviceVector<bst_float>* out_preds,
const gbm::GBTreeModel& model) const {
const gbm::GBTreeModel& model) const override {
size_t n_classes = model.learner_model_param->num_output_group;
size_t n = n_classes * info.num_row_;
const HostDeviceVector<bst_float>& base_margin = info.base_margin_;