[YDF] Fix training logs crash

yggdrasil_decision_forests/model/gradient_boosted_trees/gradient_boosted_trees.cc

@@ -986,9 +986,14 @@ metric::proto::EvaluationResults TrainingLogToEvaluationResults(
   evaluation.set_loss_value(eval_set == TrainingLogEvaluationSet::kValidation
                                 ? log_entry.validation_loss()
                                 : log_entry.training_loss());
-
-  for (int metrix_idx = 0;
-       metrix_idx < training_logs.secondary_metric_names_size(); metrix_idx++) {
+  int secondary_metric_size =
+      eval_set == TrainingLogEvaluationSet::kValidation
+          ? log_entry.validation_secondary_metrics_size()
+          : log_entry.training_secondary_metrics_size();
+  secondary_metric_size = std::min(secondary_metric_size,
+                                   training_logs.secondary_metric_names_size());
+
+  for (int metrix_idx = 0; metrix_idx < secondary_metric_size; metrix_idx++) {
     const auto& metric_name = training_logs.secondary_metric_names(metrix_idx);
     const auto metric_value =
         eval_set == TrainingLogEvaluationSet::kValidation

yggdrasil_decision_forests/port/python/ydf/model/gradient_boosted_trees_model/gradient_boosted_trees_model_test.py

@@ -92,6 +92,30 @@ def test_training_logs(self):
         self.assertAlmostEqual(training_evaluation.loss, 0.5057407)
         self.assertAlmostEqual(training_evaluation.accuracy, 0.89436144)
 
+  def test_training_logs_with_newly_trained_model(self):
+    dataset = {
+        "x": np.array([0, 0, 1, 1] * 20),
+        "y": np.array([0, 0, 1, 1] * 20),
+    }
+    model = specialized_learners.GradientBoostedTreesLearner(
+        label="y",
+        num_trees=5,
+        validation_ratio=0.5,
+    ).train(dataset)
+
+    training_logs = model.training_logs()
+    self.assertLen(training_logs, 5)
+
+    for log in training_logs:
+      # Check validation evaluation
+      self.assertIsNotNone(log.evaluation)
+      self.assertTrue(hasattr(log.evaluation, "loss"))
+      self.assertIsInstance(log.evaluation.loss, float)
+
+      # Check training evaluation
+      self.assertIsNotNone(log.training_evaluation)
+      self.assertTrue(hasattr(log.training_evaluation, "loss"))
+
   def test_empty_training_logs(self):
     # This model has no training logs.
     training_logs = self.adult_binary_class_gbdt.training_logs()

yggdrasil_decision_forests/port/python/ydf/model/gradient_boosted_trees_model/gradient_boosted_trees_wrapper.cc

@@ -17,6 +17,7 @@
 
 #include <pybind11/numpy.h>
 
+#include <cmath>
 #include <cstring>
 #include <memory>
 #include <utility>
@@ -94,18 +95,21 @@ std::vector<GBTCCTrainingLogEntry> GradientBoostedTreesCCModel::training_logs()
   const auto& label_col_spec = gbt_model_->label_col_spec();
   logs.reserve(training_logs.entries_size());
   for (const auto& entry : training_logs.entries()) {
-    const auto& validation_evaluation =
-        model::gradient_boosted_trees::internal::TrainingLogToEvaluationResults(
-            entry, training_logs, gbt_model_->task(), label_col_spec,
-            gbt_model_->loss_config(), gbt_model_->GetLossName(),
-            model::gradient_boosted_trees::internal::TrainingLogEvaluationSet::
-                kValidation);
-    const auto& training_evaluation =
+    const auto training_evaluation =
         model::gradient_boosted_trees::internal::TrainingLogToEvaluationResults(
             entry, training_logs, gbt_model_->task(), label_col_spec,
             gbt_model_->loss_config(), gbt_model_->GetLossName(),
             model::gradient_boosted_trees::internal::TrainingLogEvaluationSet::
                 kTraining);
+    metric::proto::EvaluationResults validation_evaluation;
+    if (!std::isnan(gbt_model_->validation_loss())) {
+      validation_evaluation = model::gradient_boosted_trees::internal::
+          TrainingLogToEvaluationResults(
+              entry, training_logs, gbt_model_->task(), label_col_spec,
+              gbt_model_->loss_config(), gbt_model_->GetLossName(),
+              model::gradient_boosted_trees::internal::
+                  TrainingLogEvaluationSet::kValidation);
+    }
     logs.push_back({.iteration = entry.number_of_trees(),
                     .validation_evaluation = validation_evaluation,
                     .training_evaluation = training_evaluation});