Auto set `objective=None` when using fobj (#1231)

* fix the bug

* fix R

R-package/R/lgb.Booster.R

@@ -209,7 +209,9 @@ Booster <- R6Class(
       
       # Check if objective is empty
       if (is.null(fobj)) {
-        
+        if (private$set_objective_to_none) {
+          stop("lgb.Booster.update: cannot update due to null objective function")
+        }
         # Boost iteration from known objective
         ret <- lgb.call("LGBM_BoosterUpdateOneIter_R", ret = NULL, private$handle)
         
@@ -219,7 +221,10 @@ Booster <- R6Class(
         if (!is.function(fobj)) {
           stop("lgb.Booster.update: fobj should be a function")
         }
-        
+        if (!private$set_objective_to_none) {
+          self$reset_parameter(objective="none")
+          private$set_objective_to_none = TRUE
+        }
         # Perform objective calculation
         gpair <- fobj(private$inner_predict(1), private$train_set)
         
@@ -449,7 +454,7 @@ Booster <- R6Class(
     init_predictor = NULL,
     eval_names = NULL,
     higher_better_inner_eval = NULL,
-    
+    set_objective_to_none = FALSE,
     # Predict data
     inner_predict = function(idx) {
       

python-package/lightgbm/basic.py

@@ -1282,6 +1282,7 @@ class Booster(object):
         self.__need_reload_eval_info = True
         self.__train_data_name = "training"
         self.__attr = {}
+        self.__set_objective_to_none = False
         self.best_iteration = -1
         self.best_score = {}
         params = {} if params is None else params
@@ -1516,12 +1517,17 @@ class Booster(object):
             self.__inner_predict_buffer[0] = None
         is_finished = ctypes.c_int(0)
         if fobj is None:
+            if self.__set_objective_to_none:
+                raise ValueError('Cannot update due to null objective function.')
             _safe_call(_LIB.LGBM_BoosterUpdateOneIter(
                 self.handle,
                 ctypes.byref(is_finished)))
             self.__is_predicted_cur_iter = [False for _ in range_(self.__num_dataset)]
             return is_finished.value == 1
         else:
+            if not self.__set_objective_to_none:
+                self.reset_parameter({"objective": "none"})
+                self.__set_objective_to_none = True
             grad, hess = fobj(self.__inner_predict(0), self.train_set)
             return self.__boost(grad, hess)
 

src/boosting/gbdt.cpp

@@ -375,10 +375,10 @@ double GBDT::BoostFromAverage() {
 }
 
 bool GBDT::TrainOneIter(const score_t* gradients, const score_t* hessians) {
-  auto init_score = BoostFromAverage();
+  double init_score = 0.0f;
   // boosting first
   if (gradients == nullptr || hessians == nullptr) {
-
+    init_score = BoostFromAverage();
     #ifdef TIMETAG
     auto start_time = std::chrono::steady_clock::now();
     #endif