Fix model slicing. (#7149)

* Use correct pointer.
* Remove best_iteration/best_score.

python-package/xgboost/training.py

@@ -103,7 +103,9 @@ def _train_internal(params, dtrain,
         # Due to compatibility with version older than 1.4, these attributes are added
         # to Python object even if early stopping is not used.
         bst.best_iteration = bst.num_boosted_rounds() - 1
+        bst.set_attr(best_iteration=str(bst.best_iteration))
         bst.best_ntree_limit = (bst.best_iteration + 1) * num_parallel_tree
+        bst.set_attr(best_ntree_limit=str(bst.best_ntree_limit))
 
     # Copy to serialise and unserialise booster to reset state and free
     # training memory

src/gbm/gbtree.cc

@@ -443,6 +443,8 @@ void GBTree::Slice(int32_t layer_begin, int32_t layer_end, int32_t step,
   CHECK(p_gbtree);
   GBTreeModel &out_model = p_gbtree->model_;
   auto layer_trees = this->LayerTrees();
+  CHECK_NE(this->model_.learner_model_param->num_feature, 0);
+  CHECK_NE(layer_trees, 0);
 
   layer_end = layer_end == 0 ? model_.trees.size() / layer_trees : layer_end;
   CHECK_GT(layer_end, layer_begin);
@@ -453,7 +455,13 @@ void GBTree::Slice(int32_t layer_begin, int32_t layer_end, int32_t step,
   std::vector<int32_t> &out_trees_info = out_model.tree_info;
   out_trees_info.resize(layer_trees * n_layers);
   out_model.param.num_trees = out_model.trees.size();
-  CHECK(this->model_.trees_to_update.empty());
+  if (!this->model_.trees_to_update.empty()) {
+    CHECK_EQ(this->model_.trees_to_update.size(), this->model_.trees.size())
+        << "Not all trees are updated, "
+        << this->model_.trees_to_update.size() - this->model_.trees.size()
+        << " trees remain.  Slice the model before making update if you only "
+           "want to update a portion of trees.";
+  }
 
   *out_of_bound = detail::SliceTrees(
       layer_begin, layer_end, step, this->model_, tparam_, layer_trees,

src/learner.cc

@@ -1024,22 +1024,37 @@ class LearnerImpl : public LearnerIO {
   Learner *Slice(int32_t begin_layer, int32_t end_layer, int32_t step,
                  bool *out_of_bound) override {
     this->Configure();
+    CHECK_NE(this->learner_model_param_.num_feature, 0);
     CHECK_GE(begin_layer, 0);
     auto *out_impl = new LearnerImpl({});
+    out_impl->learner_model_param_ = this->learner_model_param_;
+    out_impl->generic_parameters_ = this->generic_parameters_;
     auto gbm = std::unique_ptr<GradientBooster>(GradientBooster::Create(
-        this->tparam_.booster, &this->generic_parameters_,
+        this->tparam_.booster, &out_impl->generic_parameters_,
-        &this->learner_model_param_));
+        &out_impl->learner_model_param_));
     this->gbm_->Slice(begin_layer, end_layer, step, gbm.get(), out_of_bound);
     out_impl->gbm_ = std::move(gbm);
+
     Json config { Object() };
     this->SaveConfig(&config);
     out_impl->mparam_ = this->mparam_;
     out_impl->attributes_ = this->attributes_;
-    out_impl->learner_model_param_ = this->learner_model_param_;
     out_impl->SetFeatureNames(this->feature_names_);
     out_impl->SetFeatureTypes(this->feature_types_);
     out_impl->LoadConfig(config);
     out_impl->Configure();
+    CHECK_EQ(out_impl->learner_model_param_.num_feature, this->learner_model_param_.num_feature);
+    CHECK_NE(out_impl->learner_model_param_.num_feature, 0);
+
+    auto erase_attr = [&](std::string attr) {
+      // Erase invalid attributes.
+      auto attr_it = out_impl->attributes_.find(attr);
+      if (attr_it != out_impl->attributes_.cend()) {
+        out_impl->attributes_.erase(attr_it);
+      }
+    };
+    erase_attr("best_iteration");
+    erase_attr("best_score");
     return out_impl;
   }
 

tests/cpp/gbm/test_gbtree.cc

@@ -397,6 +397,10 @@ std::pair<Json, Json> TestModelSlice(std::string booster) {
     j++;
   }
 
+  // CHECK sliced model doesn't have dependency on old one
+  learner.reset();
+  CHECK_EQ(sliced->GetNumFeature(), kCols);
+
   return std::make_pair(model, sliced_model);
 }
 

tests/python/test_cli.py

@@ -99,6 +99,8 @@ eval[test] = {data_path}
             # CLI model doesn't contain feature info.
             booster.feature_names = None
             booster.feature_types = None
+            booster.set_attr(best_iteration=None)
+            booster.set_attr(best_ntree_limit=None)
 
             booster.save_model(model_out_py)
             py_predt = booster.predict(data)

tests/python/test_data_iterator.py

@@ -114,7 +114,8 @@ def run_data_iterator(
     if tree_method != "gpu_hist":
         rtol = 1e-1  # flaky
     else:
-        np.testing.assert_allclose(it_predt, arr_predt, rtol=1e-3)
+        # Model can be sensitive to quantiles, use 1e-2 to relax the test.
+        np.testing.assert_allclose(it_predt, arr_predt, rtol=1e-2)
         rtol = 1e-6
 
     np.testing.assert_allclose(