ModelCalcerWrapper::CalcFlatTransposed

catboost/libs/model/model.h

@@ -960,6 +960,33 @@ class TFullModel {
         CalcFlat(featureRefs, results, featureInfo);
     }
 
+    /**
+     * Call CalcFlatTransposed on all model trees
+     * @param features
+     * @param results
+     */
+    void CalcFlatTransposed(
+        TConstArrayRef<TConstArrayRef<float>> features,
+        TArrayRef<double> results,
+        const TFeatureLayout* featureInfo = nullptr
+    ) const {
+        CalcFlatTransposed(features, 0, GetTreeCount(), results, featureInfo);
+    }
+
+    /**
+     * Call CalcFlatTransposed on all model trees
+     * @param features
+     * @param results
+     */
+    void CalcFlatTransposed(
+        TConstArrayRef<TVector<float>> features,
+        TArrayRef<double> results,
+        const TFeatureLayout* featureInfo = nullptr
+    ) const {
+        TVector<TConstArrayRef<float>> featureRefs{features.begin(), features.end()};
+        CalcFlatTransposed(featureRefs, results, featureInfo);
+    }
+
     /**
      * Same as CalcFlat method but for one object
      * @param[in] features flat features array reference. First dimension is object index, second dimension is

catboost/libs/model_interface/c_api.cpp

@@ -361,6 +361,20 @@ CATBOOST_API bool CalcModelPredictionFlat(ModelCalcerHandle* modelHandle, size_t
     return true;
 }
 
+CATBOOST_API bool CalcModelPredictionFlatTransposed(ModelCalcerHandle* modelHandle, size_t docCount, const float** floatFeatures, size_t floatFeaturesSize, double* result, size_t resultSize) {
+    try {
+        TVector<TConstArrayRef<float>> featuresVec(floatFeaturesSize);
+        for (size_t i = 0; i < floatFeaturesSize; ++i) {
+            featuresVec[i] = TConstArrayRef<float>(floatFeatures[i], docCount);
+        }
+        FULL_MODEL_PTR(modelHandle)->CalcFlatTransposed(featuresVec, TArrayRef<double>(result, resultSize));
+    } catch (...) {
+        Singleton<TErrorMessageHolder>()->Message = CurrentExceptionMessage();
+        return false;
+    }
+    return true;
+}
+
 CATBOOST_API bool CalcModelPrediction(
         ModelCalcerHandle* modelHandle,
         size_t docCount,

catboost/libs/model_interface/c_api.h

@@ -138,7 +138,7 @@ CATBOOST_API bool SetPredictionTypeString(ModelCalcerHandle* modelHandle, const
  * Flat here means that float features and categorical feature are in the same float array.
  * @param calcer model handle
  * @param docCount number of objects
- * @param floatFeatures array of array of float (first dimension is object index, second if feature index)
+ * @param floatFeatures array of array of float (first dimension is object index, second is feature index)
  * @param floatFeaturesSize float values array size
  * @param result pointer to user allocated results vector
  * @param resultSize Result size should be equal to modelApproxDimension * docCount
@@ -151,6 +151,25 @@ CATBOOST_API bool CalcModelPredictionFlat(
     const float** floatFeatures, size_t floatFeaturesSize,
     double* result, size_t resultSize);
 
+
+/**
+ * **Use this method only if you really understand what you want.**
+ * Calculate raw model predictions on transposed dataset layout
+ * @param calcer model handle
+ * @param docCount number of objects
+ * @param floatFeatures array of array of float (first dimension is feature index, second is object index)
+ * @param floatFeaturesSize float values array size
+ * @param result pointer to user allocated results vector
+ * @param resultSize Result size should be equal to modelApproxDimension * docCount
+ * (e.g. for non multiclass models should be equal to docCount)
+ * @return false if error occured
+ */
+CATBOOST_API bool CalcModelPredictionFlatTransposed(
+    ModelCalcerHandle* modelHandle,
+    size_t docCount,
+    const float** floatFeatures, size_t floatFeaturesSize,
+    double* result, size_t resultSize);
+
 /**
  * Calculate raw model predictions on float features and string categorical feature values
  * @param calcer model handle

catboost/libs/model_interface/calcer.exports

@@ -24,6 +24,7 @@ C CalcModelPredictionText
 C CalcModelPredictionTextAndEmbeddings
 C CalcModelPredictionSingle
 C CalcModelPredictionFlat
+C CalcModelPredictionFlatTransposed
 C CalcModelPredictionWithHashedCatFeatures
 C CalcModelPredictionWithHashedCatFeaturesAndTextFeatures
 C CalcModelPredictionWithHashedCatFeaturesAndTextAndEmbeddingFeatures

catboost/libs/model_interface/wrapped_calcer.h

@@ -213,6 +213,7 @@ class ModelCalcerWrapper {
     std::vector<double> CalcFlat(const std::vector<std::vector<float>>& features) const {
         std::vector<double> result(features.size() * DimensionsCount);
         std::vector<const float*> ptrsVector;
+        ptrsVector.reserve(features.size());
         size_t flatVecSize = 0;
         for (const auto& flatVec : features) {
             flatVecSize = flatVec.size();
@@ -225,6 +226,28 @@ class ModelCalcerWrapper {
         return result;
     }
 
+    /**
+     * Evaluate model on transposed dataset layout.
+     * **WARNING** currently supports only singleclass models.
+     * @param transposedFeatures
+     * @return vector of raw prediction values
+     */
+    std::vector<double> CalcFlatTransposed(const std::vector<std::vector<float>>& transposedFeatures) const {
+        std::vector<const float*> ptrsVector;
+        ptrsVector.reserve(transposedFeatures.size());
+        size_t docCount = 0;
+        for (const auto& feature : transposedFeatures) {
+            docCount = feature.size();
+            // TODO(kirillovs): add check that all docCount are equal
+            ptrsVector.push_back(feature.data());
+        }
+        std::vector<double> result(docCount * DimensionsCount);
+        if (!CalcModelPredictionFlatTransposed(CalcerHolder.get(), docCount, ptrsVector.data(), transposedFeatures.size(), result.data(), result.size())) {
+            throw std::runtime_error(GetErrorString());
+        }
+        return result;
+    }
+
     /**
      * Evaluate model on vectors of float, categorical, text and embedding feature values.
      * **WARNING** categorical and text features string values should not contain zero bytes in the middle of the string (latter this could be changed).