diff --git a/paddle/framework/lod_tensor.cc b/paddle/framework/lod_tensor.cc
index beb2edee23d5586ef19155ac59a172f23a9794d9..dcde3705d6a6d2d7aa69d43dcd7cfc8bd1c1a49b 100644
--- a/paddle/framework/lod_tensor.cc
+++ b/paddle/framework/lod_tensor.cc
@@ -43,6 +43,22 @@ std::ostream &operator<<(std::ostream &os, const LoD &lod) {
   return os;
 }
 
+std::ostream &operator<<(std::ostream &os, const LoDTensor &t) {
+  PADDLE_ENFORCE(platform::is_cpu_place(t.place()));
+  PADDLE_ENFORCE(t.type().hash_code() == typeid(float).hash_code());
+
+  os << "dim: " << t.dims() << "\n";
+  os << "lod: " << t.lod() << "\n";
+
+  // only print first ten elements
+  int64_t size = t.numel() < 10 ? t.numel() : 10;
+  for (int64_t i = 0; i < size; ++i) {
+    os << t.data<float>()[i] << " ";
+  }
+
+  return os;
+}
+
 LoD SliceLevels(const LoD &in, size_t level_begin, size_t level_end) {
   LoD new_lod;
   new_lod.reserve(level_end - level_begin);
diff --git a/paddle/framework/lod_tensor.h b/paddle/framework/lod_tensor.h
index fae36892f00132c5fb2d9bc017d9b4f24e78b6ea..7e44ffca3b81da990cdef6871ca4d1ac85ce109b 100644
--- a/paddle/framework/lod_tensor.h
+++ b/paddle/framework/lod_tensor.h
@@ -58,6 +58,7 @@ using Vector = thrust::host_vector<
 using LoD = std::vector<Vector<size_t>>;
 
 std::ostream& operator<<(std::ostream& os, const LoD& lod);
+std::ostream& operator<<(std::ostream& os, const LoDTensor& t);
 
 /*
  * Slice levels from a LoD.
diff --git a/paddle/operators/parallel_do_op.cc b/paddle/operators/parallel_do_op.cc
index b15d171b9b6e6debf53c34050b718bf3e3fa851c..fd48c1b54cb9cc45416d2fd8e9b1426d488e5d2d 100644
--- a/paddle/operators/parallel_do_op.cc
+++ b/paddle/operators/parallel_do_op.cc
@@ -156,11 +156,12 @@ class ParallelDoGradOp : public OperatorBase {
 
     for (auto &s : Inputs(framework::GradVarName(kOutputs))) {
       LOG(INFO) << s;
-      LOG(INFO) << scope.FindVar(s)->Get<LoDTensor>().dims();
+      LOG(INFO) << scope.FindVar(s)->Get<LoDTensor>();
       for (auto *sub_scope : sub_scopes) {
-        LOG(INFO) << sub_scope->FindVar(s)->Get<LoDTensor>().dims();
+        LOG(INFO) << sub_scope->FindVar(s)->Get<LoDTensor>();
       }
     }
+
     // exe run
     for (int place_idx = 0; place_idx < places.size(); ++place_idx) {
       VLOG(3) << "Run " << place_idx;