Customized average fixes (nullptr and network) (#739)

* Created objectives and metrics xentropy and xentropy1

* Some coment and code cleanup.

* Added Kullback-Leibler version of metric. Changed some warning messages.

* Fixed sign error in KL-divergence calc.

* Removed __PRETTY_FUNCTION__.

* Fixed better name for alternative xentropy parameterization.
Documented details on the objectives / metrics in code comments.

* Common code for label interval checks. Cleanups.

* Use common utility for various weight property checks.

* Added code for customizable initial average to boost from.

* Fixed spelling error in aliases.

* Fixed nullptr check error in custom average.

* Added network code for customized init score.

* Restructured so that "standard" and "custom" init auto scores share aux.
function.

* Updated name for aux. automatic initial score fetcher.

src/boosting/gbdt.cpp

@@ -393,13 +393,30 @@ void GBDT::UpdateScoreOutOfBag(const Tree* tree, const int cur_tree_id) {
   #endif
 }
 
-double LabelAverage(const float* label, data_size_t num_data) {
-  double sum_label = 0.0f;
-  #pragma omp parallel for schedule(static) reduction(+:sum_label)
-  for (data_size_t i = 0; i < num_data; ++i) {
-    sum_label += label[i];
+/* If the custom "average" is implemented it will be used inplace of the label average (if enabled)
+ *
+ * An improvement to this is to have options to explicitly choose
+ * (i) standard average
+ * (ii) custom average if available
+ * (iii) any user defined scalar bias (e.g. using a new option "init_score" that overrides (i) and (ii) )
+ *
+ * (i) and (ii) could be selected as say "auto_init_score" = 0 or 1 etc..
+ *
+ */
+double ObtainAutomaticInitialScore(const ObjectiveFunction* objf, const float* label, data_size_t num_data) {
+  double init_score = 0.0f;
+  bool got_custom = false;
+  if (objf != nullptr) {
+    got_custom = objf->GetCustomAverage(&init_score);
+  }
+  if (!got_custom) {
+    double sum_label = 0.0f;
+    #pragma omp parallel for schedule(static) reduction(+:sum_label)
+    for (data_size_t i = 0; i < num_data; ++i) {
+      sum_label += label[i];
+    }
+    init_score = sum_label / num_data;
   }
-  double init_score = sum_label / num_data;
   if (Network::num_machines() > 1) {
     double global_init_score = 0.0f;
     Network::Allreduce(reinterpret_cast<char*>(&init_score),
@@ -426,20 +443,15 @@ double LabelAverage(const float* label, data_size_t num_data) {
 }
 
 bool GBDT::TrainOneIter(const score_t* gradient, const score_t* hessian, bool is_eval) {
-  // boosting from average prediction. It doesn't work well for classification, remove it for now.
+  // boosting from average label; or customized "average" if implemented for the current objective
   if (models_.empty()
       && gbdt_config_->boost_from_average
       && !train_score_updater_->has_init_score()
       && num_class_ <= 1
       && objective_function_ != nullptr
       && objective_function_->BoostFromAverage()) {
-    double init_score = 0.0f;
-    // First try to poll the optional custom average score calculation for the specific objective
-    if (!objective_function_->GetCustomAverage(&init_score)) {
-      // otherwise compute a standard label average
-      auto label = train_data_->metadata().label();
-      init_score = LabelAverage(label, num_data_);
-    }
+    auto label = train_data_->metadata().label();
+    double init_score = ObtainAutomaticInitialScore(objective_function_, label, num_data_);
     std::unique_ptr<Tree> new_tree(new Tree(2));
     new_tree->Split(0, 0, BinType::NumericalBin, 0, 0, 0, init_score, init_score, 0, 0, -1, 0, 0, 0);
     train_score_updater_->AddScore(init_score, 0);

src/objective/xentropy_objective.hpp

@@ -110,7 +110,7 @@ public:
     if (initscore == nullptr) return false;
     double suml = 0.0f;
     double sumw = 0.0f;
-    if (weights_ == nullptr) {
+    if (weights_ != nullptr) {
       for (data_size_t i = 0; i < num_data_; ++i) {
         suml += label_[i] * weights_[i];
         sumw += weights_[i];