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

Move num_parallel_tree to model parameter. (#7751)

Browse Source 
The size of forest should be a property of model itself instead of a training
hyper-parameter.
This commit is contained in:
Jiaming Yuan
2022-03-29 02:32:42 +08:00
committed by GitHub
parent 8b3ecfca25
commit 3c9b04460a
11 changed files with 158 additions and 101 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
tests/cpp/gbm/test_gbtree.cc
View File

@@ -207,7 +207,7 @@ TEST(GBTree, JsonIO) {
ASSERT_EQ(get<Integer>(get<Object>(get<Array>(gbtree_model["trees"]).front()).at("id")), 0);
ASSERT_EQ(get<Array>(gbtree_model["tree_info"]).size(), 1ul);
auto j_train_param = model["config"]["gbtree_train_param"];
auto j_train_param = model["config"]["gbtree_model_param"];
ASSERT_EQ(get<String>(j_train_param["num_parallel_tree"]), "1");
}
@@ -337,6 +337,13 @@ std::pair<Json, Json> TestModelSlice(std::string booster) {
Json sliced_config {Object()};
sliced->SaveConfig(&sliced_config);
// Only num trees is changed
if (booster == "gbtree") {
sliced_config["learner"]["gradient_booster"]["gbtree_model_param"]["num_trees"] = String("60");
} else {
sliced_config["learner"]["gradient_booster"]["gbtree"]["gbtree_model_param"]["num_trees"] =
String("60");
}
CHECK_EQ(sliced_config, config);
auto get_trees = [&](Json const& model) {