diff --git a/paddle/operators/array_to_lod_tensor_op.cc b/paddle/operators/array_to_lod_tensor_op.cc
index 6cd9c06b8ae3d3b17be83268c2f5d4002705b111..c0903bb4e5ca7f160e19eefab99af7e3e4a8ed76 100644
--- a/paddle/operators/array_to_lod_tensor_op.cc
+++ b/paddle/operators/array_to_lod_tensor_op.cc
@@ -140,6 +140,23 @@ class ArrayToLoDTensorInferShape : public framework::InferShapeBase {
                    "ArrayToLoDTensorOp must has input X.");
     PADDLE_ENFORCE(context->HasInput("RankTable"),
                    "ArrayToLoDTensorOp must has input RankTable.");
+    context->SetOutputDim("Out", context->GetInputDim("X"));
+  }
+};
+
+class ArrayToLoDTensorGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDescBind> Apply() const override {
+    auto *grad_op = new framework::OpDescBind();
+    grad_op->SetType("lod_tensor_to_array");
+    grad_op->SetInput("X", OutputGrad("Out"));
+    grad_op->SetInput("RankTable", Input("RankTable"));
+    grad_op->SetOutput("Out", InputGrad("X"));
+    grad_op->SetAttrMap(Attrs());
+    return std::unique_ptr<framework::OpDescBind>(grad_op);
   }
 };
 
@@ -149,4 +166,5 @@ class ArrayToLoDTensorInferShape : public framework::InferShapeBase {
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(array_to_lod_tensor, ops::ArrayToLoDTensorOp,
                   ops::ArrayToLoDTensorOpProtoMaker,
-                  ops::ArrayToLoDTensorInferShape);
+                  ops::ArrayToLoDTensorInferShape,
+                  ops::ArrayToLoDTensorGradMaker);
diff --git a/paddle/operators/lod_tensor_to_array_op.cc b/paddle/operators/lod_tensor_to_array_op.cc
index 5f02f5e8a12831a33683cdc53cf0feb7cb908da5..58af35564d83b9699af4f7783fb6367ff9590682 100644
--- a/paddle/operators/lod_tensor_to_array_op.cc
+++ b/paddle/operators/lod_tensor_to_array_op.cc
@@ -133,6 +133,22 @@ class LoDTensorToArrayInferVarType : public framework::VarTypeInference {
   }
 };
 
+class LoDTensorToArrayGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDescBind> Apply() const override {
+    auto *grad_op = new framework::OpDescBind();
+    grad_op->SetType("array_to_lod_tensor");
+    grad_op->SetInput("X", OutputGrad("Out"));
+    grad_op->SetInput("RankTable", Input("RankTable"));
+    grad_op->SetOutput("Out", InputGrad("X"));
+    grad_op->SetAttrMap(Attrs());
+    return std::unique_ptr<framework::OpDescBind>(grad_op);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -140,4 +156,5 @@ namespace ops = paddle::operators;
 REGISTER_OPERATOR(lod_tensor_to_array, ops::LoDTensorToArrayOp,
                   ops::LoDTensorToArrayOpProtoMaker,
                   ops::LoDTensorToArrayInferShape,
-                  ops::LoDTensorToArrayInferVarType);
+                  ops::LoDTensorToArrayInferVarType,
+                  ops::LoDTensorToArrayGradMaker);
diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc
index 78b4bbca84d4670aba73222f1d679604d7516b02..dcc5b4286f4ac833268a779a9a7edd2ed119ffff 100644
--- a/paddle/operators/mean_op.cc
+++ b/paddle/operators/mean_op.cc
@@ -51,6 +51,7 @@ class MeanGradOp : public framework::OperatorWithKernel {
 
   void InferShape(framework::InferShapeContext* ctx) const override {
     ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
+    ctx->ShareLoD("X", framework::GradVarName("X"));
   }
 };
 
diff --git a/python/paddle/v2/framework/layers.py b/python/paddle/v2/framework/layers.py
index 22540b2b9731e95cbc5ca903c9b2fa4de1a59502..4c6703cd8b1e7a959ae70ec76d7578eda3a81f25 100644
--- a/python/paddle/v2/framework/layers.py
+++ b/python/paddle/v2/framework/layers.py
@@ -87,7 +87,8 @@ def data(name,
          type=core.VarDesc.VarType.LOD_TENSOR,
          append_batch_size=True,
          main_program=None,
-         startup_program=None):
+         startup_program=None,
+         stop_gradient=True):
     helper = LayerHelper('data', **locals())
     shape = list(shape)
     for i in xrange(len(shape)):
@@ -101,7 +102,11 @@ def data(name,
         shape = [-1] + shape  # append batch size as -1
 
     return helper.create_global_variable(
-        name=name, shape=shape, dtype=data_type, type=type, stop_gradient=True)
+        name=name,
+        shape=shape,
+        dtype=data_type,
+        type=type,
+        stop_gradient=stop_gradient)
 
 
 def _convert_(name):
@@ -845,7 +850,8 @@ def lod_tensor_to_array(x, table, main_program=None):
     helper = LayerHelper("lod_tensor_to_array", **locals())
     array = helper.create_variable(
         name=unique_name("lod_tensor_to_array"),
-        type=core.VarDesc.VarType.LOD_TENSOR_ARRAY)
+        type=core.VarDesc.VarType.LOD_TENSOR_ARRAY,
+        dtype=x.data_type)
     helper.append_op(
         type='lod_tensor_to_array',
         inputs={'X': x,
diff --git a/python/paddle/v2/framework/tests/test_lod_tensor_array_ops.py b/python/paddle/v2/framework/tests/test_lod_tensor_array_ops.py
index 61a5fcf07d2af18bf904c74d5c0d4c4eb462154c..e9713666b3f64d7a39afadab7da6b22f149b8cf8 100644
--- a/python/paddle/v2/framework/tests/test_lod_tensor_array_ops.py
+++ b/python/paddle/v2/framework/tests/test_lod_tensor_array_ops.py
@@ -4,6 +4,7 @@ import numpy
 import paddle.v2.framework.layers as layers
 from paddle.v2.framework.framework import Program
 from paddle.v2.framework.executor import Executor
+from paddle.v2.framework.backward import append_backward_ops
 
 
 class TestCPULoDTensorArrayOps(unittest.TestCase):
@@ -123,5 +124,42 @@ class TestCPULoDTensorArrayOps(unittest.TestCase):
         self.assertEqual(actual.lod(), expect.lod())
 
 
+class TestCPULoDTensorArrayOpGrad(unittest.TestCase):
+    def test_grad(self):
+        place = core.CPUPlace()
+        program = Program()
+
+        x = layers.data(
+            name='x',
+            shape=[1],
+            data_type='float32',
+            main_program=program,
+            stop_gradient=False)
+        table = layers.lod_rank_table(x, level=0, main_program=program)
+        array = layers.lod_tensor_to_array(x, table, main_program=program)
+        result = layers.array_to_lod_tensor(array, table, main_program=program)
+
+        mean = layers.mean(x=result, main_program=program)
+
+        append_backward_ops(mean)
+
+        tensor = core.LoDTensor()
+        tensor.set(numpy.arange(10).reshape(10, 1).astype('float32'), place)
+        tensor.set_lod([[0, 3, 9, 10]])
+
+        g_vars = program.global_block().var(x.name + "@GRAD")
+
+        exe = Executor(place)
+        g_out = [
+            item.sum()
+            for item in map(
+                numpy.array,
+                exe.run(program, feed={'x': tensor}, fetch_list=[g_vars]))
+        ]
+        g_out_sum = numpy.array(g_out).sum()
+
+        self.assertAlmostEqual(1.0, g_out_sum, delta=0.1)
+
+
 if __name__ == '__main__':
     unittest.main()