add options to turn off categorical sorting and sorting by gradients …

…(not updates)
interpretml · Dec 31, 2024 · 8bb0f56 · 8bb0f56
1 parent 15c55ac
commit 8bb0f56
Show file tree

Hide file tree

Showing 3 changed files with 39 additions and 23 deletions.
diff --git a/docs/benchmarks/ebm-benchmark.ipynb b/docs/benchmarks/ebm-benchmark.ipynb
@@ -206,8 +206,9 @@
     "    max_samples = None\n",
     "    n_calibration_folds = 4  # 4 uses all cores on the containers\n",
     "\n",
-    "    from interpret.develop import set_option\n",
     "    from interpret.glassbox import ExplainableBoostingClassifier, ExplainableBoostingRegressor\n",
+    "    from interpret.develop import set_option\n",
+    "    from interpret.utils._native import Native\n",
     "    from xgboost import XGBClassifier, XGBRegressor, XGBRFClassifier, XGBRFRegressor\n",
     "    from lightgbm import LGBMClassifier, LGBMRegressor\n",
     "    from catboost import CatBoostClassifier, CatBoostRegressor\n",
@@ -235,6 +236,9 @@
     "    import re\n",
     "    import random\n",
     "\n",
+    "    # turn off AVX512F support since it adds variability in results given spotty architecture support\n",
+    "    set_option(\"acceleration\", ~Native.AccelerationFlags_AVX512F)\n",
+    "    \n",
     "    random.seed(seed)\n",
     "    np.random.seed(seed)\n",
     "\n",
@@ -996,7 +1000,7 @@
     "#results_df = results_df[results_df['type'] != 'multiclass']\n",
     "#results_df = results_df[results_df['type'] != 'regression']\n",
     "#\n",
-    "#results_df = results_df[(results_df['method'] != 'ebm') | (results_df['meta'] == '{}')]\n",
+    "#results_df = results_df[(results_df['method'] != 'ebm') | (results_df['meta'] == '{\"interactions\": 0}') | (results_df['meta'] == '{}')]\n",
     "#results_df = results_df[((results_df['method'] == 'ebm') & (results_df['meta'] == '{}')) | (results_df['method'] == 'xgb')]\n",
     "#\n",
     "#results_df = results_df[\n",

diff --git a/shared/libebm/PartitionOneDimensionalBoosting.cpp b/shared/libebm/PartitionOneDimensionalBoosting.cpp
@@ -919,17 +919,24 @@ template<bool bHessian, size_t cCompilerScores> class CompareBin final {
       // NEVER check for exact equality (as a precondition is ok), since then we'd violate the weak ordering rule
       // https://medium.com/@shiansu/strict-weak-ordering-and-the-c-stl-f7dcfa4d4e07
 
-      const bool bUpdateWithHessian = bHessian && m_bHessianRuntime;
-
-      const FloatCalc hess1 =
-            static_cast<FloatCalc>(bUpdateWithHessian ? lhs->GetGradientPairs()[0].GetHess() : lhs->GetWeight());
-      const FloatCalc val1 =
-            static_cast<FloatCalc>(lhs->GetGradientPairs()[0].m_sumGradients) / (hess1 + m_categoricalSmoothing);
-
-      const FloatCalc hess2 =
-            static_cast<FloatCalc>(bUpdateWithHessian ? rhs->GetGradientPairs()[0].GetHess() : rhs->GetWeight());
-      const FloatCalc val2 =
-            static_cast<FloatCalc>(rhs->GetGradientPairs()[0].m_sumGradients) / (hess2 + m_categoricalSmoothing);
+      EBM_ASSERT(!std::isnan(m_categoricalSmoothing));
+
+      FloatCalc val1 = static_cast<FloatCalc>(lhs->GetGradientPairs()[0].m_sumGradients);
+      FloatCalc val2 = static_cast<FloatCalc>(rhs->GetGradientPairs()[0].m_sumGradients);
+      if(!std::isinf(m_categoricalSmoothing)) {
+         FloatCalc hess1;
+         FloatCalc hess2;
+         const bool bUpdateWithHessian = bHessian && m_bHessianRuntime;
+         if(!bUpdateWithHessian) {
+            hess1 = static_cast<FloatCalc>(lhs->GetWeight());
+            hess2 = static_cast<FloatCalc>(rhs->GetWeight());
+         } else {
+            hess1 = static_cast<FloatCalc>(lhs->GetGradientPairs()[0].GetHess());
+            hess2 = static_cast<FloatCalc>(rhs->GetGradientPairs()[0].GetHess());
+         }
+         val1 /= (hess1 + m_categoricalSmoothing);
+         val2 /= (hess2 + m_categoricalSmoothing);
+      }
 
       if(val1 == val2) {
          return lhs < rhs;
@@ -1130,17 +1137,22 @@ template<bool bHessian, size_t cCompilerScores> class PartitionOneDimensionalBoo
             return error;
          }
 
-         // shuffle
-         while(size_t{1} != cRemaining) {
-            const size_t iSwap = pRng->NextFast(cRemaining);
-            auto* const pTemp = apBins[iSwap];
-            --cRemaining;
-            apBins[iSwap] = apBins[cRemaining];
-            apBins[cRemaining] = pTemp;
+         const bool bShuffle = 1 != cCompilerScores || std::isnan(categoricalSmoothing);
+         const bool bSort = 1 == cCompilerScores && !std::isnan(categoricalSmoothing);
+
+         EBM_ASSERT(bShuffle || bSort);
+
+         if(bShuffle) {
+            while(size_t{1} != cRemaining) {
+               const size_t iSwap = pRng->NextFast(cRemaining);
+               auto* const pTemp = apBins[iSwap];
+               --cRemaining;
+               apBins[iSwap] = apBins[cRemaining];
+               apBins[cRemaining] = pTemp;
+            }
          }
 
-         static constexpr bool bSingleScore = 1 == cCompilerScores;
-         if(bSingleScore) {
+         if(bSort) {
             // there isn't a single key to sort on with multiple grad/hess pairs, so use random ordering otherwise.
             std::sort(apBins,
                   ppBin,

diff --git a/shared/libebm/tests/boosting_unusual_inputs.cpp b/shared/libebm/tests/boosting_unusual_inputs.cpp
@@ -2380,7 +2380,7 @@ static double RandomizedTesting(const AccelerationFlags acceleration) {
 }
 
 TEST_CASE("stress test, boosting") {
-   const double expected = 15044531333054.148;
+   const double expected = 15111161995602.100;
 
    double validationMetricExact = RandomizedTesting(AccelerationFlags_NONE);
    CHECK(validationMetricExact == expected);