fix bug in feature fraction (#1099)

* fix feature fraction * fix bugs.
microsoft · Dec 4, 2017 · 699d438 · 699d438
1 parent a957bd6
commit 699d438
Show file tree

Hide file tree

Showing 3 changed files with 19 additions and 6 deletions.
diff --git a/include/LightGBM/dataset.h b/include/LightGBM/dataset.h
@@ -361,7 +361,15 @@ class Dataset {
   inline uint64_t NumTotalBin() const {
     return group_bin_boundaries_.back();
   }
-
+  inline std::vector<int> ValidFeatureIndices() const {
+    std::vector<int> ret;
+    for (int i = 0; i < num_total_features_; ++i) {
+      if (used_feature_map_[i] >= 0) {
+        ret.push_back(i);
+      }
+    }
+    return ret;
+  }
   void ReSize(data_size_t num_data);
 
   void CopySubset(const Dataset* fullset, const data_size_t* used_indices, data_size_t num_used_indices, bool need_meta_data);

diff --git a/src/treelearner/serial_tree_learner.cpp b/src/treelearner/serial_tree_learner.cpp
@@ -78,6 +78,7 @@ void SerialTreeLearner::Init(const Dataset* train_data, bool is_constant_hessian
   // initialize data partition
   data_partition_.reset(new DataPartition(num_data_, tree_config_->num_leaves));
   is_feature_used_.resize(num_features_);
+  valid_feature_indices_ = train_data_->ValidFeatureIndices();
   // initialize ordered gradients and hessians
   ordered_gradients_.resize(num_data_);
   ordered_hessians_.resize(num_data_);
@@ -237,16 +238,19 @@ void SerialTreeLearner::BeforeTrain() {
   histogram_pool_.ResetMap();
 
   if (tree_config_->feature_fraction < 1) {
-    int used_feature_cnt = static_cast<int>(train_data_->num_total_features()*tree_config_->feature_fraction);
+    int used_feature_cnt = static_cast<int>(valid_feature_indices_.size()*tree_config_->feature_fraction);
+    // at least use one feature
+    used_feature_cnt = std::max(used_feature_cnt, 1);
     // initialize used features
     std::memset(is_feature_used_.data(), 0, sizeof(int8_t) * num_features_);
     // Get used feature at current tree
-    auto used_feature_indices = random_.Sample(train_data_->num_total_features(), used_feature_cnt);
-    int omp_loop_size = static_cast<int>(used_feature_indices.size());
+    auto sampled_indices = random_.Sample(valid_feature_indices_.size(), used_feature_cnt);
+    int omp_loop_size = static_cast<int>(sampled_indices.size());
     #pragma omp parallel for schedule(static, 512) if (omp_loop_size >= 1024)
     for (int i = 0; i < omp_loop_size; ++i) {
-      int inner_feature_index = train_data_->InnerFeatureIndex(used_feature_indices[i]);
-      if (inner_feature_index < 0) { continue; }
+      int used_feature = valid_feature_indices_[sampled_indices[i]];
+      int inner_feature_index = train_data_->InnerFeatureIndex(used_feature);
+      CHECK(inner_feature_index >= 0);
       is_feature_used_[inner_feature_index] = 1;
     }
   } else {

diff --git a/src/treelearner/serial_tree_learner.h b/src/treelearner/serial_tree_learner.h
@@ -125,6 +125,7 @@ class SerialTreeLearner: public TreeLearner {
   std::unique_ptr<LeafSplits> smaller_leaf_splits_;
   /*! \brief stores best thresholds for all feature for larger leaf */
   std::unique_ptr<LeafSplits> larger_leaf_splits_;
+  std::vector<int> valid_feature_indices_;
 
 #ifdef USE_GPU
   /*! \brief gradients of current iteration, ordered for cache optimized, aligned to 4K page */