Log warnings for number of bins of categorical features (#4448)

* log warnings when number of bins of categorical features exceeds the configured maximum number of bins

* log only one warning information for all categorical features

* Add #include <memory> for unique_ptr

* remove useless param description

Author: shiyu1994

include/LightGBM/dataset_loader.h

@@ -7,6 +7,7 @@
 
 #include <LightGBM/dataset.h>
 
+#include <memory>
 #include <string>
 #include <unordered_set>
 #include <vector>
@@ -63,6 +64,16 @@ class DatasetLoader {
   /*! \brief Check can load from binary file */
   std::string CheckCanLoadFromBin(const char* filename);
 
+  /*! \brief Check the number of bins for categorical features.
+   * The number of bins for categorical features may exceed the configured maximum value.
+   * Log warnings when such cases happen.
+   *
+   * \param bin_mappers the bin_mappers of all features
+   * \param max_bin max_bin from Config
+   * \param max_bin_by_feature max_bin_by_feature from Config
+   */
+  void CheckCategoricalFeatureNumBin(const std::vector<std::unique_ptr<BinMapper>>& bin_mappers, const int max_bin, const std::vector<int>& max_bin_by_feature) const;
+
   const Config& config_;
   /*! \brief Random generator*/
   Random random_;

src/io/dataset_loader.cpp

@@ -805,6 +805,7 @@ Dataset* DatasetLoader::ConstructFromSampleData(double** sample_values,
       cp_ptr += bin_mappers[i]->SizesInByte();
     }
   }
+  CheckCategoricalFeatureNumBin(bin_mappers, config_.max_bin, config_.max_bin_by_feature);
   auto dataset = std::unique_ptr<Dataset>(new Dataset(num_data));
   dataset->Construct(&bin_mappers, num_total_features, forced_bin_bounds, sample_indices, sample_values, num_per_col, num_col, total_sample_size, config_);
   if (dataset->has_raw()) {
@@ -1184,6 +1185,7 @@ void DatasetLoader::ConstructBinMappersFromTextData(int rank, int num_machines,
       cp_ptr += bin_mappers[i]->SizesInByte();
     }
   }
+  CheckCategoricalFeatureNumBin(bin_mappers, config_.max_bin, config_.max_bin_by_feature);
   dataset->Construct(&bin_mappers, dataset->num_total_features_, forced_bin_bounds, Common::Vector2Ptr<int>(&sample_indices).data(),
                      Common::Vector2Ptr<double>(&sample_values).data(),
                      Common::VectorSize<int>(sample_indices).data(), static_cast<int>(sample_indices.size()), sample_data.size(), config_);
@@ -1463,4 +1465,44 @@ std::vector<std::vector<double>> DatasetLoader::GetForcedBins(std::string forced
   return forced_bins;
 }
 
+void DatasetLoader::CheckCategoricalFeatureNumBin(
+  const std::vector<std::unique_ptr<BinMapper>>& bin_mappers,
+  const int max_bin, const std::vector<int>& max_bin_by_feature) const {
+  bool need_warning = false;
+  if (bin_mappers.size() < 1024) {
+    for (size_t i = 0; i < bin_mappers.size(); ++i) {
+      const int max_bin_for_this_feature = max_bin_by_feature.empty() ? max_bin : max_bin_by_feature[i];
+      if (bin_mappers[i]->bin_type() == BinType::CategoricalBin && bin_mappers[i]->num_bin() > max_bin_for_this_feature) {
+        need_warning = true;
+        break;
+      }
+    }
+  } else {
+    const int num_threads = OMP_NUM_THREADS();
+    std::vector<bool> thread_need_warning(num_threads, false);
+    Threading::For<size_t>(0, bin_mappers.size(), 1,
+      [&bin_mappers, &thread_need_warning, &max_bin_by_feature, max_bin] (int thread_index, size_t start, size_t end) {
+        for (size_t i = start; i < end; ++i) {
+          thread_need_warning[thread_index] = false;
+          const int max_bin_for_this_feature = max_bin_by_feature.empty() ? max_bin : max_bin_by_feature[i];
+          if (bin_mappers[i]->bin_type() == BinType::CategoricalBin && bin_mappers[i]->num_bin() > max_bin_for_this_feature) {
+            thread_need_warning[thread_index] = true;
+            break;
+          }
+        }
+      });
+    for (int thread_index = 0; thread_index < num_threads; ++thread_index) {
+      if (thread_need_warning[thread_index]) {
+        need_warning = true;
+        break;
+      }
+    }
+  }
+
+  if (need_warning) {
+    Log::Warning("Categorical features with more bins than the configured maximum bin number found.");
+    Log::Warning("For categorical features, max_bin and max_bin_by_feature may be ignored with a large number of categories.");
+  }
+}
+
 }  // namespace LightGBM

src/treelearner/cuda_tree_learner.cpp

@@ -506,10 +506,30 @@ void CUDATreeLearner::InitGPU(int num_gpu) {
   } else {
     Log::Fatal("bin size %d cannot run on GPU", max_num_bin_);
   }
-  if (max_num_bin_ == 65) {
+
+  // ignore the feature groups that contain categorical features when producing warnings about max_bin.
+  // these groups may contain larger number of bins due to categorical features, but not due to the setting of max_bin.
+  int max_num_bin_no_categorical = 0;
+  int cur_feature_group = 0;
+  bool categorical_feature_found = false;
+  for (int inner_feature_index = 0; inner_feature_index < num_features_; ++inner_feature_index) {
+    const int feature_group = train_data_->Feature2Group(inner_feature_index);
+    const BinMapper* feature_bin_mapper = train_data_->FeatureBinMapper(inner_feature_index);
+    if (feature_bin_mapper->bin_type() == BinType::CategoricalBin) {
+      categorical_feature_found = true;
+    }
+    if (feature_group != cur_feature_group || inner_feature_index == num_features_ - 1) {
+      if (!categorical_feature_found) {
+        max_num_bin_no_categorical = std::max(max_num_bin_no_categorical, train_data_->FeatureGroupNumBin(cur_feature_group));
+      }
+      categorical_feature_found = false;
+      cur_feature_group = feature_group;
+    }
+  }
+  if (max_num_bin_no_categorical == 65) {
     Log::Warning("Setting max_bin to 63 is suggested for best performance");
   }
-  if (max_num_bin_ == 17) {
+  if (max_num_bin_no_categorical == 17) {
     Log::Warning("Setting max_bin to 15 is suggested for best performance");
   }
 

src/treelearner/gpu_tree_learner.cpp

@@ -719,10 +719,30 @@ void GPUTreeLearner::InitGPU(int platform_id, int device_id) {
   } else {
     Log::Fatal("bin size %d cannot run on GPU", max_num_bin_);
   }
-  if (max_num_bin_ == 65) {
+
+  // ignore the feature groups that contain categorical features when producing warnings about max_bin.
+  // these groups may contain larger number of bins due to categorical features, but not due to the setting of max_bin.
+  int max_num_bin_no_categorical = 0;
+  int cur_feature_group = 0;
+  bool categorical_feature_found = false;
+  for (int inner_feature_index = 0; inner_feature_index < num_features_; ++inner_feature_index) {
+    const int feature_group = train_data_->Feature2Group(inner_feature_index);
+    const BinMapper* feature_bin_mapper = train_data_->FeatureBinMapper(inner_feature_index);
+    if (feature_bin_mapper->bin_type() == BinType::CategoricalBin) {
+      categorical_feature_found = true;
+    }
+    if (feature_group != cur_feature_group || inner_feature_index == num_features_ - 1) {
+      if (!categorical_feature_found) {
+        max_num_bin_no_categorical = std::max(max_num_bin_no_categorical, train_data_->FeatureGroupNumBin(cur_feature_group));
+      }
+      categorical_feature_found = false;
+      cur_feature_group = feature_group;
+    }
+  }
+  if (max_num_bin_no_categorical == 65) {
     Log::Warning("Setting max_bin to 63 is suggested for best performance");
   }
-  if (max_num_bin_ == 17) {
+  if (max_num_bin_no_categorical == 17) {
     Log::Warning("Setting max_bin to 15 is suggested for best performance");
   }
   ctx_ = boost::compute::context(dev_);