fix a bug

paddle/fluid/operators/reader/create_custom_reader_op.cc

@@ -159,23 +159,24 @@ void CustomReader::ReadNext(std::vector<framework::LoDTensor>* out) {
   // The scope for CustomReader's sub-block should be independent and shouldn't
   // be any other computation scope's child. Otherwise, data preprocessing and
   // compution cannot be concurrent.
-  framework::Scope& exe_scope = scope_.NewScope();
+  framework::Scope* exe_scope = &scope_.NewScope();
   // 1. Copy LoDTensors from underlying reader's output to source variables.
   for (size_t i = 0; i < source_var_names_.size(); ++i) {
-    framework::Variable* var = exe_scope.Var(source_var_names_[i]);
+    framework::Variable* var = exe_scope->Var(source_var_names_[i]);
     framework::LoDTensor* tensor = var->GetMutable<framework::LoDTensor>();
     tensor->ShareDataWith(underlying_outs[i]);
     tensor->set_lod(underlying_outs[i].lod());
   }
   // 2. Run the sub-block.
-  exe_.Run(program_, &exe_scope, sub_block_id_, false, true);
+  exe_.Run(program_, exe_scope, sub_block_id_, false, true);
   // 3. Copy LoDTensors from sink variables to out.
   out->resize(sink_var_names_.size());
   for (size_t i = 0; i < sink_var_names_.size(); ++i) {
-    const auto& tensor = detail::Ref(exe_scope.FindVar(sink_var_names_[i]))
+    const auto& tensor = detail::Ref(exe_scope->FindVar(sink_var_names_[i]))
                              .Get<framework::LoDTensor>();
     framework::TensorCopySync(tensor, platform::CPUPlace(), &(*out)[i]);
   }
+  scope_.DeleteScope(exe_scope);
 }
 
 }  // namespace reader