Skip to content

GitLab

Commit c4c83bc7 authored by Guolin Ke's avatar Guolin Ke
Browse files

fix bug in finding best split for categorical feature.

parent ef778069
Show whitespace changes
Inline Side-by-side
Showing with 22 additions and 15 deletions
src/treelearner/feature_histogram.hpp
View file @ c4c83bc7
...@@ -131,12 +131,12 @@ public: ...@@ -131,12 +131,12 @@ public:
output->left_output = CalculateSplittedLeafOutput(best_sum_left_gradient, best_sum_left_hessian); output->left_output = CalculateSplittedLeafOutput(best_sum_left_gradient, best_sum_left_hessian);
output->left_count = best_left_count; output->left_count = best_left_count;
output->left_sum_gradient = best_sum_left_gradient; output->left_sum_gradient = best_sum_left_gradient;
output->left_sum_hessian = best_sum_left_hessian; output->left_sum_hessian = best_sum_left_hessian - kEpsilon;
output->right_output = CalculateSplittedLeafOutput(sum_gradient - best_sum_left_gradient, output->right_output = CalculateSplittedLeafOutput(sum_gradient - best_sum_left_gradient,
sum_hessian - best_sum_left_hessian); sum_hessian - best_sum_left_hessian);
output->right_count = num_data - best_left_count; output->right_count = num_data - best_left_count;
output->right_sum_gradient = sum_gradient - best_sum_left_gradient; output->right_sum_gradient = sum_gradient - best_sum_left_gradient;
output->right_sum_hessian = sum_hessian - best_sum_left_hessian; output->right_sum_hessian = sum_hessian - best_sum_left_hessian - kEpsilon;
output->gain = best_gain - gain_shift; output->gain = best_gain - gain_shift;
} else { } else {
output->feature = meta_->feature_idx; output->feature = meta_->feature_idx;
...@@ -148,6 +148,9 @@ public: ...@@ -148,6 +148,9 @@ public:
SplitInfo* output) { SplitInfo* output) {
double best_gain = kMinScore; double best_gain = kMinScore;
uint32_t best_threshold = static_cast<uint32_t>(meta_->num_bin); uint32_t best_threshold = static_cast<uint32_t>(meta_->num_bin);
data_size_t best_left_count = 0;
double best_sum_left_gradient = 0.0f;
double best_sum_left_hessian = 0.0f;
double gain_shift = GetLeafSplitGain(sum_gradient, sum_hessian); double gain_shift = GetLeafSplitGain(sum_gradient, sum_hessian);
double min_gain_shift = gain_shift + meta_->tree_config->min_gain_to_split; double min_gain_shift = gain_shift + meta_->tree_config->min_gain_to_split;
is_splittable_ = false; is_splittable_ = false;
...@@ -179,6 +182,9 @@ public: ...@@ -179,6 +182,9 @@ public:
// better split point // better split point
if (current_gain > best_gain) { if (current_gain > best_gain) {
best_threshold = static_cast<uint32_t>(t + bias); best_threshold = static_cast<uint32_t>(t + bias);
best_sum_left_gradient = data_[t].sum_gradients;
best_sum_left_hessian = data_[t].sum_hessians + kEpsilon;
best_left_count = data_[t].cnt;
best_gain = current_gain; best_gain = current_gain;
} }
} }
...@@ -186,7 +192,7 @@ public: ...@@ -186,7 +192,7 @@ public:
if (bias == 1) { if (bias == 1) {
t = meta_->num_bin - 1 - bias; t = meta_->num_bin - 1 - bias;
double sum_bin0_gradient = sum_gradient; double sum_bin0_gradient = sum_gradient;
double sum_bin0_hessian = sum_hessian; double sum_bin0_hessian = sum_hessian - 2 * kEpsilon;
data_size_t cnt_bin0 = num_data; data_size_t cnt_bin0 = num_data;
for (; t >= 0; --t) { for (; t >= 0; --t) {
sum_bin0_gradient -= data_[t].sum_gradients; sum_bin0_gradient -= data_[t].sum_gradients;
...@@ -207,6 +213,9 @@ public: ...@@ -207,6 +213,9 @@ public:
// better split point // better split point
if (current_gain > best_gain) { if (current_gain > best_gain) {
best_threshold = static_cast<uint32_t>(0); best_threshold = static_cast<uint32_t>(0);
best_sum_left_gradient = sum_bin0_gradient;
best_sum_left_hessian = sum_bin0_hessian + kEpsilon;
best_left_count = cnt_bin0;
best_gain = current_gain; best_gain = current_gain;
} }
} }
...@@ -216,17 +225,15 @@ public: ...@@ -216,17 +225,15 @@ public:
// update split information // update split information
output->feature = meta_->feature_idx; output->feature = meta_->feature_idx;
output->threshold = best_threshold; output->threshold = best_threshold;
output->left_output = CalculateSplittedLeafOutput(data_[best_threshold].sum_gradients, output->left_output = CalculateSplittedLeafOutput(best_sum_left_gradient, best_sum_left_hessian);
data_[best_threshold].sum_hessians + kEpsilon); output->left_count = best_left_count;
output->left_count = data_[best_threshold].cnt; output->left_sum_gradient = best_sum_left_gradient;
output->left_sum_gradient = data_[best_threshold].sum_gradients; output->left_sum_hessian = best_sum_left_hessian - kEpsilon;
output->left_sum_hessian = data_[best_threshold].sum_hessians + kEpsilon; output->right_output = CalculateSplittedLeafOutput(sum_gradient - best_sum_left_gradient,
sum_hessian - best_sum_left_hessian);
output->right_output = CalculateSplittedLeafOutput(sum_gradient - data_[best_threshold].sum_gradients, output->right_count = num_data - best_left_count;
sum_hessian - data_[best_threshold].sum_hessians - kEpsilon); output->right_sum_gradient = sum_gradient - best_sum_left_gradient;
output->right_count = num_data - data_[best_threshold].cnt; output->right_sum_hessian = sum_hessian - best_sum_left_hessian - kEpsilon;
output->right_sum_gradient = sum_gradient - data_[best_threshold].sum_gradients;
output->right_sum_hessian = sum_hessian - data_[best_threshold].sum_hessians - kEpsilon;
output->gain = best_gain - gain_shift; output->gain = best_gain - gain_shift;
} else { } else {
output->feature = meta_->feature_idx; output->feature = meta_->feature_idx;
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment