[BP] Fix num_boosted_rounds for linear model. (#7538) (#7559)

* Add note.

* Fix n boosted rounds.

src/gbm/gblinear_model.cc

@@ -18,6 +18,7 @@ void GBLinearModel::SaveModel(Json* p_out) const {
     j_weights[i] = weight[i];
   }
   out["weights"] = std::move(j_weights);
+  out["boosted_rounds"] = Json{this->num_boosted_rounds};
 }
 
 void GBLinearModel::LoadModel(Json const& in) {
@@ -27,6 +28,13 @@ void GBLinearModel::LoadModel(Json const& in) {
   for (size_t i = 0; i < n_weights; ++i) {
     weight[i] = get<Number const>(j_weights[i]);
   }
+  auto const& obj = get<Object const>(in);
+  auto boosted_rounds = obj.find("boosted_rounds");
+  if (boosted_rounds != obj.cend()) {
+    this->num_boosted_rounds = get<Integer const>(boosted_rounds->second);
+  } else {
+    this->num_boosted_rounds = 0;
+  }
 }
 
 DMLC_REGISTER_PARAMETER(DeprecatedGBLinearModelParam);

tests/python-gpu/test_gpu_linear.py

@@ -1,7 +1,6 @@
 import sys
-from hypothesis import strategies, given, settings, assume
+from hypothesis import strategies, given, settings, assume, note
 import pytest
-import numpy
 import xgboost as xgb
 sys.path.append("tests/python")
 import testing as tm
@@ -17,10 +16,14 @@
     'top_k': strategies.integers(1, 10),
 })
 
+
 def train_result(param, dmat, num_rounds):
     result = {}
-    xgb.train(param, dmat, num_rounds, [(dmat, 'train')], verbose_eval=False,
-              evals_result=result)
+    booster = xgb.train(
+        param, dmat, num_rounds, [(dmat, 'train')], verbose_eval=False,
+        evals_result=result
+    )
+    assert booster.num_boosted_rounds() == num_rounds
     return result
 
 
@@ -33,6 +36,7 @@ def test_gpu_coordinate(self, param, num_rounds, dataset):
         param['updater'] = 'gpu_coord_descent'
         param = dataset.set_params(param)
         result = train_result(param, dataset.get_dmat(), num_rounds)['train'][dataset.metric]
+        note(result)
         assert tm.non_increasing(result)
 
     # Loss is not guaranteed to always decrease because of regularisation parameters
@@ -49,6 +53,7 @@ def test_gpu_coordinate_regularised(self, param, num_rounds, dataset, alpha, lam
         param['lambda'] = lambd
         param = dataset.set_params(param)
         result = train_result(param, dataset.get_dmat(), num_rounds)['train'][dataset.metric]
+        note(result)
         assert tm.non_increasing([result[0], result[-1]])
 
     @pytest.mark.skipif(**tm.no_cupy())

tests/python/test_linear.py

@@ -32,6 +32,7 @@ def test_coordinate(self, param, num_rounds, dataset, coord_param):
         param.update(coord_param)
         param = dataset.set_params(param)
         result = train_result(param, dataset.get_dmat(), num_rounds)['train'][dataset.metric]
+        note(result)
         assert tm.non_increasing(result, 5e-4)
 
     # Loss is not guaranteed to always decrease because of regularisation parameters
@@ -48,6 +49,7 @@ def test_coordinate_regularised(self, param, num_rounds, dataset, coord_param, a
         param.update(coord_param)
         param = dataset.set_params(param)
         result = train_result(param, dataset.get_dmat(), num_rounds)['train'][dataset.metric]
+        note(result)
         assert tm.non_increasing([result[0], result[-1]])
 
     @given(parameter_strategy, strategies.integers(10, 50),
@@ -57,6 +59,7 @@ def test_shotgun(self, param, num_rounds, dataset):
         param['updater'] = 'shotgun'
         param = dataset.set_params(param)
         result = train_result(param, dataset.get_dmat(), num_rounds)['train'][dataset.metric]
+        note(result)
         # shotgun is non-deterministic, so we relax the test by only using first and last
         # iteration.
         if len(result) > 2:
@@ -75,4 +78,5 @@ def test_shotgun_regularised(self, param, num_rounds, dataset, alpha, lambd):
         param['lambda'] = lambd
         param = dataset.set_params(param)
         result = train_result(param, dataset.get_dmat(), num_rounds)['train'][dataset.metric]
+        note(result)
         assert tm.non_increasing([result[0], result[-1]])