[python-package] Expose ObjectiveFunction class

include/LightGBM/c_api.h

@@ -31,6 +31,7 @@ typedef void* DatasetHandle;  /*!< \brief Handle of dataset. */
 typedef void* BoosterHandle;  /*!< \brief Handle of booster. */
 typedef void* FastConfigHandle; /*!< \brief Handle of FastConfig. */
 typedef void* ByteBufferHandle; /*!< \brief Handle of ByteBuffer. */
+typedef void* ObjectiveFunctionHandle; /*!< \brief Handle of ObjectiveFunction. */
 
 #define C_API_DTYPE_FLOAT32 (0)  /*!< \brief float32 (single precision float). */
 #define C_API_DTYPE_FLOAT64 (1)  /*!< \brief float64 (double precision float). */
@@ -1563,6 +1564,60 @@ LIGHTGBM_C_EXPORT int LGBM_BoosterGetUpperBoundValue(BoosterHandle handle,
 LIGHTGBM_C_EXPORT int LGBM_BoosterGetLowerBoundValue(BoosterHandle handle,
                                                      double* out_results);
 
+/*!
+ * \brief Create an objective function.
+ * \param typ Type of the objective function
+ * \param parameter Parameters for the objective function
+ * \param[out] out Handle pointing to the created objective function
+ * \return 0 when succeed, -1 when failure happens
+ */
+LIGHTGBM_C_EXPORT int LGBM_ObjectiveFunctionCreate(const char *typ,
+                                                   const char *parameter,
+                                                   ObjectiveFunctionHandle *out);
+
+/*!
+ * \brief Initialize an objective function with the dataset.
+ * \param handle Handle of the objective function
+ * \param dataset Handle of the dataset used for initialization
+ * \param[out] num_data Number of data points; this may be modified within the function
+ * \return 0 when succeed, -1 when failure happens
+ */
+LIGHTGBM_C_EXPORT int LGBM_ObjectiveFunctionInit(ObjectiveFunctionHandle handle,
+                                                 DatasetHandle dataset,
+                                                 int *num_data);
+
+/*!
+ * \brief Evaluate the objective function given model scores.
+ * \param handle Handle of the objective function
+ * \param score Array of scores predicted by the model
+ * \param[out] grad Gradient result array
+ * \param[out] hess Hessian result array
+ * \return 0 when succeed, -1 when failure happens
+ */
+LIGHTGBM_C_EXPORT int LGBM_ObjectiveFunctionGetGradients(ObjectiveFunctionHandle handle,
+                                                         const double* score,
+                                                         float* grad,
+                                                         float* hess);
+
+/*!
+ * \brief Free the memory allocated for an objective function.
+ * \param handle Handle of the objective function
+ * \return 0 when succeed, -1 when failure happens
+ */
+LIGHTGBM_C_EXPORT int LGBM_ObjectiveFunctionFree(ObjectiveFunctionHandle handle);
+
+/*!
+ * \brief Convert raw scores to outputs.
+ * \param handle Handle of the objective function
+ * \param num_data Number of data points
+ * \param inputs Array of raw scores
+ * \param[out] outputs Array of outputs
+ */
+LIGHTGBM_C_EXPORT int LGBM_ObjectiveFunctionConvertOutputs(ObjectiveFunctionHandle handle,
+                                                           const int num_data,
+                                                           const double* inputs,
+                                                           double* outputs);
+
 /*!
  * \brief Initialize the network.
  * \param machines List of machines in format 'ip1:port1,ip2:port2'

include/LightGBM/objective_function.h

@@ -67,6 +67,12 @@ class ObjectiveFunction {
   /*! \brief Return the number of positive samples. Return 0 if no binary classification tasks.*/
   virtual data_size_t NumPositiveData() const { return 0; }
 
+  virtual void ConvertOutputs(const int num_data, const double* inputs, double* outputs) const {
+    for (int i = 0; i < num_data; i ++) {
+      ConvertOutput(inputs + i, outputs + i);
+    }
+  }
+
   virtual void ConvertOutput(const double* input, double* output) const {
     output[0] = input[0];
   }

python-package/lightgbm/__init__.py

@@ -6,7 +6,7 @@
 
 from pathlib import Path
 
-from .basic import Booster, Dataset, Sequence, register_logger
+from .basic import Booster, Dataset, ObjectiveFunction, Sequence, register_logger
 from .callback import EarlyStopException, early_stopping, log_evaluation, record_evaluation, reset_parameter
 from .engine import CVBooster, cv, train
 
@@ -31,6 +31,7 @@
 __all__ = [
     "Dataset",
     "Booster",
+    "ObjectiveFunction",
     "CVBooster",
     "Sequence",
     "register_logger",

python-package/lightgbm/basic.py

@@ -5281,3 +5281,162 @@ def __get_eval_info(self) -> None:
                 self.__higher_better_inner_eval = [
                     name.startswith(("auc", "ndcg@", "map@", "average_precision")) for name in self.__name_inner_eval
                 ]
+
+
+class ObjectiveFunction:
+    """
+    ObjectiveFunction in LightGBM.
+
+    This class exposes the builtin objective functions for evaluating gradients and hessians
+    on external datasets. LightGBM does not use this wrapper during its training as it is
+    using the underlying C++ class.
+    """
+
+    def __init__(self, name: str, params: Dict[str, Any]):
+        """
+        Initialize the ObjectiveFunction.
+
+        Parameters
+        ----------
+        name : str
+            The name of the objective function.
+        params : dict
+            Dictionary of parameters for the objective function.
+            These are the parameters that would have been passed to ``booster.train``.
+            The ``name`` should be consistent with the ``params["objective"]`` field.
+        """
+        self.name = name
+        self.params = params
+        self.num_data = None
+        self.num_class = params.get("num_class", 1)
+
+        if "objective" in params and params["objective"] != self.name:
+            raise ValueError('The name should be consistent with the params["objective"] field.')
+
+        self.__create()
+
+    def init(self, dataset: Dataset) -> "ObjectiveFunction":
+        """
+        Initialize the objective function using the provided dataset.
+
+        Parameters
+        ----------
+        dataset : Dataset
+            The dataset object used for initialization.
+
+        Returns
+        -------
+        self : ObjectiveFunction
+            Initialized objective function object.
+        """
+        return self.__init_from_dataset(dataset)
+
+    def convert_outputs(self, scores: np.ndarray) -> np.ndarray:
+        """
+        Convert the raw scores to the final predictions.
+
+        Parameters
+        ----------
+        scores : numpy.ndarray
+            Raw scores from the model.
+
+        Returns
+        -------
+        result : numpy.ndarray
+        """
+        if self._handle is None:
+            raise ValueError("Objective function seems uninitialized")
+
+        if self.num_class == 1:
+            scores = _list_to_1d_numpy(scores, dtype=np.float64, name="scores")
+        else:
+            scores = _data_to_2d_numpy(scores, dtype=np.float64, name="scores")
+
+        num_data = scores.size
+        out_preds = np.zeros_like(scores, dtype=np.float64)
+
+        _safe_call(
+            _LIB.LGBM_ObjectiveFunctionConvertOutputs(
+                self._handle,
+                ctypes.c_int(num_data),
+                scores.ctypes.data_as(ctypes.POINTER(ctypes.c_double)),
+                out_preds.ctypes.data_as(ctypes.POINTER(ctypes.c_double)),
+            )
+        )
+
+        return out_preds
+
+    def get_gradients(self, y_pred: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        Evaluate the objective function given model predictions.
+
+        Parameters
+        ----------
+        y_pred : numpy.ndarray
+            Predicted scores from the model.
+
+        Returns
+        -------
+        (grad, hess) : Tuple[np.ndarray, np.ndarray]
+            A tuple containing gradients and Hessians.
+        """
+        if self._handle is None:
+            raise ValueError("Objective function seems uninitialized")
+
+        if self.num_data is None or self.num_class is None:
+            raise ValueError("ObjectiveFunction was not created properly")
+
+        data_shape = self.num_data * self.num_class
+        grad = np.zeros(dtype=np.float32, shape=data_shape)
+        hess = np.zeros(dtype=np.float32, shape=data_shape)
+
+        _safe_call(
+            _LIB.LGBM_ObjectiveFunctionGetGradients(
+                self._handle,
+                y_pred.ctypes.data_as(ctypes.POINTER(ctypes.c_double)),
+                grad.ctypes.data_as(ctypes.POINTER(ctypes.c_float)),
+                hess.ctypes.data_as(ctypes.POINTER(ctypes.c_float)),
+            )
+        )
+
+        return (grad, hess)
+
+    def __create(self):
+        self._handle = ctypes.c_void_p()
+        _safe_call(
+            _LIB.LGBM_ObjectiveFunctionCreate(
+                _c_str(self.name),
+                _c_str(_param_dict_to_str(self.params)),
+                ctypes.byref(self._handle),
+            )
+        )
+
+    def __init_from_dataset(self, dataset: Dataset) -> "ObjectiveFunction":
+        if dataset._handle is None:
+            raise ValueError("Cannot create ObjectiveFunction from uninitialised Dataset")
+
+        if self._handle is None:
+            raise ValueError("Dealocated ObjectiveFunction cannot be initialized")
+
+        tmp_num_data = ctypes.c_int(0)
+        _safe_call(
+            _LIB.LGBM_ObjectiveFunctionInit(
+                self._handle,
+                dataset._handle,
+                ctypes.byref(tmp_num_data),
+            )
+        )
+        self.num_data = tmp_num_data.value
+        return self
+
+    def __del__(self) -> None:
+        try:
+            self._free_handle()
+        except AttributeError:
+            pass
+
+    def _free_handle(self) -> "ObjectiveFunction":
+        if self._handle is not None:
+            _safe_call(_LIB.LGBM_ObjectiveFunctionFree(self._handle))
+            self._handle = None
+        return self

src/c_api.cpp

@@ -907,6 +907,7 @@ using LightGBM::kZeroThreshold;
 using LightGBM::LGBM_APIHandleException;
 using LightGBM::Log;
 using LightGBM::Network;
+using LightGBM::ObjectiveFunction;
 using LightGBM::Random;
 using LightGBM::ReduceScatterFunction;
 using LightGBM::SingleRowPredictor;
@@ -2747,6 +2748,74 @@ int LGBM_BoosterGetLowerBoundValue(BoosterHandle handle,
   API_END();
 }
 
+LIGHTGBM_C_EXPORT int LGBM_ObjectiveFunctionCreate(const char *typ,
+                                                   const char *parameter,
+                                                   ObjectiveFunctionHandle *out) {
+  API_BEGIN();
+  auto param = Config::Str2Map(parameter);
+  Config config(param);
+  if (config.device_type != std::string("cpu")) {
+    Log::Fatal("Currently the ObjectiveFunction class is only exposed for CPU devices.");
+  } else {
+    *out = ObjectiveFunction::CreateObjectiveFunction(std::string(typ), config);
+  }
+  API_END();
+}
+
+LIGHTGBM_C_EXPORT int LGBM_ObjectiveFunctionInit(ObjectiveFunctionHandle handle,
+                                                 DatasetHandle dataset,
+                                                 int *num_data) {
+  API_BEGIN();
+  ObjectiveFunction* ref_fobj = reinterpret_cast<ObjectiveFunction*>(handle);
+  Dataset* ref_dataset = reinterpret_cast<Dataset*>(dataset);
+  ref_fobj->Init(ref_dataset->metadata(), ref_dataset->num_data());
+  *num_data = ref_dataset->num_data();
+  API_END();
+}
+
+LIGHTGBM_C_EXPORT int LGBM_ObjectiveFunctionGetGradients(ObjectiveFunctionHandle handle,
+                                                         const double* score,
+                                                         float* grad,
+                                                         float* hess) {
+  API_BEGIN();
+  #ifdef SCORE_T_USE_DOUBLE
+  (void) handle;       // UNUSED VARIABLE
+  (void) grad;         // UNUSED VARIABLE
+  (void) hess;         // UNUSED VARIABLE
+  Log::Fatal("Don't support evaluating objective function when SCORE_T_USE_DOUBLE is enabled");
+  #else
+  ObjectiveFunction* ref_fobj = reinterpret_cast<ObjectiveFunction*>(handle);
+  ref_fobj->GetGradients(score, grad, hess);
+  #endif
+  API_END();
+}
+
+LIGHTGBM_C_EXPORT int LGBM_ObjectiveFunctionConvertOutputs(ObjectiveFunctionHandle handle,
+                                                           const int num_data,
+                                                           const double* inputs,
+                                                           double* outputs) {
+  API_BEGIN();
+  #ifdef SCORE_T_USE_DOUBLE
+  (void) handle;       // UNUSED VARIABLE
+  (void) num_data;     // UNUSED VARIABLE
+  (void) inputs;       // UNUSED VARIABLE
+  (void) outputs;      // UNUSED VARIABLE
+  Log::Fatal("Don't support evaluating objective function when SCORE_T_USE_DOUBLE is enabled");
+  #else
+  ObjectiveFunction* ref_fobj = reinterpret_cast<ObjectiveFunction*>(handle);
+  ref_fobj->ConvertOutputs(num_data, inputs, outputs);
+  #endif
+  API_END();
+}
+
+/*!
+ */
+LIGHTGBM_C_EXPORT int LGBM_ObjectiveFunctionFree(ObjectiveFunctionHandle handle) {
+  API_BEGIN();
+  delete reinterpret_cast<ObjectiveFunction*>(handle);
+  API_END();
+}
+
 int LGBM_NetworkInit(const char* machines,
                      int local_listen_port,
                      int listen_time_out,

src/objective/multiclass_objective.hpp

@@ -129,6 +129,12 @@ class MulticlassSoftmax: public ObjectiveFunction {
     }
   }
 
+  void ConvertOutputs(const int num_data, const double* inputs, double* outputs) const override {
+    for (int i = 0; i < num_data; i += num_class_) {
+      ConvertOutput(inputs + i, outputs + i);
+    }
+  }
+
   void ConvertOutput(const double* input, double* output) const override {
     Common::Softmax(input, output, num_class_);
   }
@@ -236,6 +242,12 @@ class MulticlassOVA: public ObjectiveFunction {
     return "multiclassova";
   }
 
+  void ConvertOutputs(const int num_data, const double* inputs, double* outputs) const override {
+    for (int i = 0; i < num_data; i += num_class_) {
+      ConvertOutput(inputs + i, outputs + i);
+    }
+  }
+
   void ConvertOutput(const double* input, double* output) const override {
     for (int i = 0; i < num_class_; ++i) {
       output[i] = 1.0f / (1.0f + std::exp(-sigmoid_ * input[i]));