diff --git a/doc/design/block.md b/doc/design/block.md
index be8800122035984df281692fc40009c397565046..4d5dd4ba95a686d18b2339c69f0316c340681909 100644
--- a/doc/design/block.md
+++ b/doc/design/block.md
@@ -55,17 +55,23 @@ Let us consolidate the discussion by presenting some examples.
 The following C++ programs shows how blocks are used with the `if-else` structure:
 
 ```c++
+namespace pd = paddle;
+
 int x = 10;
-int y = 20;
-int out;
+int y = 1;
+int z = 10;
 bool cond = false;
+int o1, o2;
 if (cond) {
   int z = x + y;
-  out = softmax(z);
+  o1 = z;
+  o2 = pd::layer::softmax(z);
 } else {
-  int z = fc(x);
-  out = z;
+  int d = pd::layer::fc(z);
+  o1 = d;
+  o2 = d+1;
 }
+
 ```
 
 An equivalent PaddlePaddle program from the design doc of the [IfElseOp operator](./if_else_op.md) is as follows:
@@ -73,57 +79,55 @@ An equivalent PaddlePaddle program from the design doc of the [IfElseOp operator
 ```python
 import paddle as pd
 
-x = var(10)
-y = var(20)
-cond = var(false)
-ie = pd.create_ifelseop(inputs=[x], output_num=1)
+x = minibatch([10, 20, 30]) # shape=[None, 1]
+y = var(1) # shape=[1], value=1
+z = minibatch([10, 20, 30]) # shape=[None, 1]
+cond = larger_than(x, 15) # [false, true, true]
+
+ie = pd.ifelse()
 with ie.true_block():
-    x = ie.inputs(true, 0)
-    z = operator.add(x, y)
-    ie.set_output(true, 0, operator.softmax(z))
+    d = pd.layer.add_scalar(x, y)
+    ie.output(d, pd.layer.softmax(d))
 with ie.false_block():
-    x = ie.inputs(false, 0)
-    z = layer.fc(x)
-    ie.set_output(true, 0, operator.softmax(z))
-out = b(cond)
+    d = pd.layer.fc(z)
+    ie.output(d, d+1)
+o1, o2 = ie(cond)
 ```
 
-In both examples, the left branch computes `softmax(x+y)` and the right branch computes `fc(x)`.
+In both examples, the left branch computes `x+y` and `softmax(x+y)`, the right branch computes `x+1` and `fc(x)`.
 
 A difference is that variables in the C++ program contain scalar values, whereas those in the PaddlePaddle programs are mini-batches of instances.  The `ie.input(true, 0)` invocation returns instances in the 0-th input, `x`, that corresponds to true values in `cond` as the local variable `x`, where `ie.input(false, 0)` returns instances corresponding to false values.
 
+
 ### Blocks with `for` and `RNNOp`
 
 The following RNN model from the [RNN design doc](./rnn.md)
 
 ```python
-x = sequence([10, 20, 30])
-m = var(0)
-W = tensor()
-U = tensor()
-
-rnn = create_rnn(inputs=[input])
-with rnn.stepnet() as net:
-  x = net.set_inputs(0)
-  h = net.add_memory(init=m)
-  fc_out = pd.matmul(W, x)
-  hidden_out = pd.matmul(U, h.pre(n=1))
-  sum = pd.add_two(fc_out, hidden_out)
-  act = pd.sigmoid(sum)
-  h.update(act)                       # update memory with act
-  net.set_outputs(0, act, hidden_out) # two outputs
-
+x = sequence([10, 20, 30]) # shape=[None, 1]
+m = var(0) # shape=[1]
+W = var(0.314, param=true) # shape=[1]
+U = var(0.375, param=true) # shape=[1]
+
+rnn = pd.rnn()
+with rnn.step():
+  h = rnn.memory(init = m)
+  hh = rnn.previous_memory(h)
+  a = layer.fc(W, x)
+  b = layer.fc(U, hh)  
+  s = pd.add(a, b)
+  act = pd.sigmoid(s)
+  rnn.update_memory(h, act)
+  rnn.output(a, b)
 o1, o2 = rnn()
-print o1, o2
 ```
-
 has its equivalent C++ program as follows
 
 ```c++
 int* x = {10, 20, 30};
-int m = 0;
-int W = some_value();
-int U = some_other_value();
+int* m = {0};
+int* W = {0.314};
+int* U = {0.375};
 
 int mem[sizeof(x) / sizeof(x[0]) + 1];
 int o1[sizeof(x) / sizeof(x[0]) + 1];
@@ -131,20 +135,16 @@ int o2[sizeof(x) / sizeof(x[0]) + 1];
 for (int i = 1; i <= sizeof(x)/sizeof(x[0]); ++i) {
   int x = x[i-1];
   if (i == 1) mem[0] = m;
-  int fc_out = W * x;
-  int hidden_out = Y * mem[i-1];
-  int sum = fc_out + hidden_out;
+  int a = W * x;
+  int b = Y * mem[i-1];
+  int s = fc_out + hidden_out;
   int act = sigmoid(sum);
   mem[i] = act;
   o1[i] = act;
   o2[i] = hidden_out;
 }
-
-print_array(o1);
-print_array(o2);
 ```
 
-
 ## Compilation and Execution
 
 Like TensorFlow programs, a PaddlePaddle program is written in Python.  The first part describes a neural network as a protobuf message, and the rest part executes the message for training or inference.
@@ -210,11 +210,11 @@ a = pd.Varaible(shape=[20, 20])
 b = pd.fc(a, params=["fc.w", "fc.b"])
 
 rnn = pd.create_rnn()
-with rnn.stepnet() as net:
-    x = net.set_inputs(a)
+with rnn.stepnet()
+    x = a.as_step_input()
     # reuse fc's parameter
     fc_without_b = pd.get_variable("fc.w")
-    net.set_outputs(fc_without_b)
+    rnn.output(fc_without_b)
 
 out = rnn()
 ```
diff --git a/doc/design/if_else_op.md b/doc/design/if_else_op.md
index 954a19c0733358c235eae3cffe134c23dac94c95..26d140f06db4ecefa86be015eaa731ffddc6910c 100644
--- a/doc/design/if_else_op.md
+++ b/doc/design/if_else_op.md
@@ -1,41 +1,51 @@
-IfOp should have only one branch. An IfOp operator takes a `cond` variable whose value must be a vector of N boolean elements. Its return value has N instances. If cond[i] == True, input instance input[i] will go through true_block() and generate output[i]; otherwise it will produce output from false_bloack().
+# The `IfElse` Operator
 
-```python
-import paddle as pd
+PaddlePaddle's `IfElse` operator differs from TensorFlow's:
 
-x = var()
-y = var()
-cond = var()
-default_value = var()
-b = pd.create_ifelseop(inputs=[x], output_num=1)
-with b.true_block():
-    x = b.inputs(0)
-    z = operator.add(x, y)
-    b.set_output(0, operator.softmax(z))
-
-with b.false_block():
-    x = b.inputs(0)
-    z = layer.fc(x)
-    b.set_output(0, operator.softmax(z))
-
-out = b(cond)
-```
+- the TensorFlow version takes a scalar boolean value as the condition so that the whole mini-batch goes to either the true or the false branch, whereas
+- the PaddlePaddle version takes a vector of boolean value as the condition, and instances corresponding to true values go to the true branch, those corresponding to false values go to the false branch.
+
+## Example
+
+The following PaddlePaddle program shows the usage of the IfElse operator:
 
-If only true_block is set in an IfElseOp, a special case is that we can have a default value for false as:
 ```python
 import paddle as pd
 
-x = var()
-y = var()
-cond = var()
-default_value = var()
-b = pd.create_ifelseop(inputs=[x], output_num=1, default_value)
-
-with b.true_block():
-    x = b.inputs(0)
-    z = operator.add(x, y)
-    b.set_output(0, operator.softmax(z))
+x = minibatch([10, 20, 30]) # shape=[None, 1]
+y = var(1) # shape=[1], value=1
+z = minibatch([10, 20, 30]) # shape=[None, 1]
+cond = larger_than(x, 15) # [false, true, true]
+
+ie = pd.ifelse()
+with ie.true_block():
+    d = pd.layer.add(x, y)
+    ie.output(d, pd.layer.softmax(d))
+with ie.false_block():
+    d = pd.layer.fc(z)
+    ie.output(d, d+1)
+o1, o2 = ie(cond)
+```
 
-out = b(cond)
+A challenge to implement the `IfElse` operator is to infer those variables to be split, or, say, to identify the variable of the mini-batch or those derived from the mini-batch.
+
+An equivalent C++ program is as follows:
+
+```c++
+namespace pd = paddle;
+
+int x = 10;
+int y = 1;
+int z = 10;
+bool cond = false;
+int o1, o2;
+if (cond) {
+  int d = x + y;
+  o1 = z;
+  o2 = pd::layer::softmax(z);
+} else {
+  int d = pd::layer::fc(z);
+  o1 = d;
+  o2 = d+1;
+}
 ```
-where default_value is a list of vars for `cond` == False.
diff --git a/doc/design/program.md b/doc/design/program.md
index fb8f86ac07af403c9fee015f2a3adbfaa3c6d631..bd2456787c4e336d357a65255a8274a7c9e465cc 100644
--- a/doc/design/program.md
+++ b/doc/design/program.md
@@ -1,8 +1,10 @@
-# Design Doc: ProgramDesc
+# Design Doc: PaddlePaddle Programs
 
-The basic structure of a PaddlePaddle program is some nested blocks, as a C++ or Java program.
+## Compile and Execution
+
+A PaddlePaddle program consists of two parts -- the first generates a `ProgramDesc` protobuf message that describes the program, and the second runs this message using a C++ class `Executor`.
 
-As described in [graph.md](./graph.md), the first five lines of the following PaddlePaddle program
+A simple example PaddlePaddle program can be found in [graph.md](./graph.md):
 
 ```python
 x = layer.data("images")
@@ -13,36 +15,112 @@ optimize(cost)
 train(cost, reader=mnist.train())
 ```
 
-generates, or compiles, a PaddelPaddle program, which is represented by the following protobuf message:
+The first five lines of the following PaddlePaddle program generates, or, compiles, the `ProgramDesc` message.  The last line runs it.
 
-```protobuf
-message ProgramDesc {
-  repeated BlockDesc blocks = 1;
+## Programs and Blocks
+
+The basic structure of a PaddlePaddle program is some nested blocks, as a C++ or Java program.
+
+- program: some nested blocks
+- [block](./block.md):
+  - some local variable definitions, and
+  - a sequence of operators
+
+The concept of block comes from usual programs.  For example, the following C++ program has three blocks:
+
+```c++
+int main() { // block 0
+  int i = 0;
+  if (i < 10) { // block 1
+    for (int j = 0; j < 10; j++) { // block 2
+    }
+  }
+  return 0;
 }
+```
+
+The following PaddlePaddle program has three blocks:
+
+```python
+import paddle as pd  // block 0
+
+x = minibatch([10, 20, 30]) # shape=[None, 1]
+y = var(1) # shape=[1], value=1
+z = minibatch([10, 20, 30]) # shape=[None, 1]
+cond = larger_than(x, 15) # [false, true, true]
 
+ie = pd.ifelse()
+with ie.true_block():  // block 1
+    d = pd.layer.add_scalar(x, y)
+    ie.output(d, pd.layer.softmax(d))
+with ie.false_block():  // block 2
+    d = pd.layer.fc(z)
+    ie.output(d, d+1)
+o1, o2 = ie(cond)
+```
+
+## `BlockDesc` and `ProgramDesc`
+
+All protobuf messages are defined in `framework.proto`.
+
+`BlockDesc` is straight-forward -- it includes local variable definitions, `vars`, and a sequence of operators, `ops`.
+
+```protobuf
 message BlockDesc {
   required int32 parent = 1;
   repeated VarDesc vars = 2;
   repeated OpDesc ops = 3;
 }
+```
+
+The parent ID indicates the parent block so that operators in a block can refer to variables defined locally and also those defined in their ancestor blocks.
+
+All hierarchical blocks in a program are flattened and stored in an array. The block ID is the index of the block in this array.
+
+```protobuf
+message ProgramDesc {
+  repeated BlockDesc blocks = 1;
+}
+```
+
+
+### Global Block
 
+The global block is the first one in the above array.
+
+## Operators that Use Blocks
+
+In the above example, the operator `IfElseOp` has two blocks -- the true branch and the false branch.
+
+The definition of `OpDesc` shows that an operator could have some attributes:
+
+```protobuf
 message OpDesc {
   AttrDesc attrs = 1;
   ...
 }
+```
+
+and an attribute could be of type block, which is, in fact, a block ID as described above:
 
+```
 message AttrDesc {
-  required AttrType type = 1;
+  required string name = 1;
 
-  // index into ProgramDesc::blocks when type==BLOCK
-  optional int32 block = 2;
+  enum AttrType {
+    INT = 1,
+    STRING = 2,
+    ...
+    BLOCK = ...
+  }
+  required AttrType type = 2;
+
+  optional int32 block = 10; // when type == BLOCK
   ...
 }
 ```
 
-When each of the first five lines runs, related Python function, e.g., `layer.fc`, calls C++ InferShape functions.  This InferShape function needs to access the properties of VarDesc's accessed by the current OpDesc. These VarDesc's might not be defined in the current block, but in some ancestor blocks.  This requires that we can trace the parent of a block.
-
-A nested block is often an attribute of an operator, most likely, an IfElseOp or a WhileOp.  In above solution, all blocks are in `ProgramDesc::blocks`, this implicitly assigns a zero-based ID to each block -- the index of the block in `ProgramDesc::blocks`.  So that `AttrDesc::block` could be an integer block ID.
+## InferShape
 
 With this design, the InferShape function should take the following parameters:
 
diff --git a/doc/design/tensor_array.md b/doc/design/tensor_array.md
index a0419ec002159893b035fae1300fce489e68936a..8378e97bf7cfaae54c36b1b92e202b16e4fe1e28 100644
--- a/doc/design/tensor_array.md
+++ b/doc/design/tensor_array.md
@@ -1,39 +1,250 @@
 # Design for TensorArray
+This design doc presents the necessity of a new C++ class `TensorArray`.
+In addition to the very simple C++ implementation
+
+```c++
+class TensorArray {
+ public:
+  explicit TensorArray(const LoDTensor&);
+  explicit TensorArray(size_t size);
+
+ private:
+  vector<LoDTensor> values_;
+};
+```
+
+We also need to expose it to PaddlePaddle's Python API,
+because users would want to use it with our very flexible operators `WhileLoop`.
+An example for a RNN based on dynamic operators is 
+
+```python
+input = pd.data(...)
+num_steps = Var(12)
+
+TensorArray states(size=num_steps)
+TensorArray step_inputs(unstack_from=input)
+TensorArray step_outputs(size=num_steps)
+
+W = Tensor(...)
+U = Tensor(...)
+default_state = some_op()
+
+step = Var(1)
+
+wloop = paddle.create_whileloop(loop_vars=[step])
+with wloop.frame():
+    wloop.break_if(pd.equal(step, num_steps)
+    pre_state = states.read(step-1, default_state)
+    step_input = step_inputs.read(step)
+    state = pd.sigmoid(pd.matmul(U, pre_state) + pd.matmul(W, step_input))
+    states.write(step, state)
+    step_outputs.write(step, state) # output state
+    step.update(state+1)
+
+output = step_outputs.stack()
+```
+
+## Background
+Steps are one of the core concepts of RNN. In each time step of RNN, there should be several input segments, states, and output segments; all these components act like arrays, for example, call `states[step_id]` will get the state in `step_id`th time step.
+
+An RNN can be implemented with the following pseudocode
+
+```c++
+Array states;
+Array input_segments;
+Array output_segments;
+Parameter W, U;
+
+step = 1
+seq_len = 12
+while_loop {
+   if (step == seq_len) break;
+    states[step] = sigmoid(W * states[step-1] + U * input_segments[step]);
+    output_segments[step] = states[step] // take state as output
+   step++;
+}
+```
+According to the [RNN roadmap](https://github.com/PaddlePaddle/Paddle/issues/4561), there are several different RNNs that PaddlePaddle will eventually support.
+
+Currently, the basic RNN implementation supported by PaddlePaddle is the `recurrent_op` which takes tensors as input and splits them into `input_segments`.
+
+
+Since a tensor cannot store variable-length sequences directly, PaddlePaddle implements the tensor with level of details (`LoDTensor` for short).
+Segmenting the `LoDTensor` is much more complicated than splitting a tensor, that makes it necessary to refactor the `recurrent_op` with `LoDTensor` segmenting support.
+
+As the next step in RNN support, `dynamic_recurrent_op` should be introduced to handle inputs with variable-length sequences.
+
+The implementation is similar to `recurrent_op`. 
+The key difference is the way **the original input `LoDTensors` and outupts are split to get the `input_segments` and the `output_segments`.**
+
+
+Though it can't be built over `recurrent_op` or `dynamic_recurrent_op` directly,
+the logic behind splitting a tensor or a LoD tensor into `input_segments` remains the same.
+
+## Why `TensorArray`
+The logic behind splitting the inputs to segments, states and outputs is similar and can be shared in a seperate module.
+
+The array of `states`, `input_segments` and `output_segments` would be exposed to users when writing a dynamic RNN model similar to the above pseudo codes. 
+
+So there should be an array-like container, which can store the segments of a tensor or LoD tensor.
+
+**This container can store an array of tensors and provides several methods to split a tensor or a LoD tensor** .
+This is where the notion of `TensorArray` comes from.
+
+## Introduce TensorArray to uniform all the three RNNs
 TensorArray as a new concept is borrowed from TensorFlow, 
 it is meant to be used with dynamic iteration primitives such as `while_loop` and `map_fn`.
 
 This concept can be used to support our new design of dynamic operations, and help to refactor some existing variant-sentence-related layers, 
-such as `RecurrentGradientMachine`.
+such as `recurrent_op`, `RecurrentGradientMachine`.
 
 In [our design for dynamic RNN](https://github.com/PaddlePaddle/Paddle/pull/4401), 
 `TensorArray` is used to segment inputs and store states in all time steps.
 By providing some methods similar to a C++ array,
-the definition of some state-based dynamic models such as RNN could be more natural and highly flexible.
-
-## Dynamic-Related Methods
-Some basic methods should be proposed as follows:
-
-### stack()
-Pack the values in a `TensorArray` into a tensor with rank one higher than each tensor in `values`.
-### unstack(axis=0)
-Unpacks the given dimension of a rank-`R` tensor into rank-`(R-1)` tensors.
-### concat()
-Return the values in the `TensorArray` as a concatenated Tensor.
-### write(index, value, data_shared=true)
-Write value into index of the TensorArray.
-### read(index)
-Read the value at location `index` in the `TensorArray`.
-### size()
-Return the number of values.
+the definition of some state-based dynamic models such as RNN can be more natural and highly flexible.
+
+## Dynamic-operations on TensorArray
+
+`TensorArray` will be used directly when defining dynamic models, so some operators listed below should be implemented
+
+```python
+# several helper operators for TensorArray
+def tensor_array_stack(ta, tensor):
+    '''
+    get a tensor array `ta`, return a packed `tensor`.
+    '''
+    pass
+
+def tensor_array_unstack(tensor, ta):
+    '''
+    get a `tensor`, unstack it and get a tensor array `ta`.
+    '''
+    pass
+
+def tensor_array_write(ta, index, tensor, data_shared):
+    '''
+    get a `tensor` and a scalar tensor `index`, write `tensor` into index-th
+    value of the tensor array `ta`.
+    `data_shared` is an attribute that specifies whether to copy or reference the tensors.
+    '''
+    pass
+
+def tensor_array_read(ta, index, tensor):
+    '''
+    get a tensor array `ta`, a scalar tensor `index`, read the index-th value of
+    `ta` and return as the `tensor`.
+    '''
+    pass
+
+def tensor_array_size(ta, tensor):
+    '''
+    get a tensor array `ta`, return the size of `ta` and return as the scalar `tensor`.
+    '''
+    pass
+```
+
+It is trivial for users to use so many low-level operators, so some helper methods should be proposed in python wrapper to make `TensorArray` easier to use, 
+for example
+
+```python
+class TensorArray:
+    def __init__(self, name):
+        self.name = name
+        self.desc = TensorArrayDesc()
+
+    def stack(self, name=None):
+        '''
+        Pack the values in a `TensorArray` into a tensor with rank one higher
+        than each tensor in `values`.
+        `stack` can be used to split tensor into time steps for RNN or whileloop.
+
+        @name: str
+            the name of the variable to output.
+        '''
+        tensor = NewVar(name)
+        tensor_array_stack(self.name, tensor)
+        return tensor
+
+    def unstack(self, input):
+        '''
+        Unpacks the given dimension of a rank-`R` tensor into rank-`(R-1)` tensors.
+        `unstack` can be used to concatenate all the time steps for RNN or whileloop.
+
+        @input: str
+            the name of input tensor
+        '''
+        tensor_array_unstack(tensor, self.name)
+
+    def write(self, index, value, data_shared=True):
+        '''
+        Write value into index of the TensorArray.
+        If `data_shared` is set to True, than the index-th value in TensorArray will
+        be shared with the tensor passed in.
+
+        @index: str
+            name of a scalar tensor
+        @value: str
+            name of a tensor
+        @data_shared: bool
+        '''
+        tensor_array_write(self.name, index, value, data_shared)
+
+    def read(self, index, output):
+        '''
+        Read the value at location `index` in the `TensorArray`.
+
+        @index: str
+            name of a scalar tensor
+        @output:
+            name of a output variable
+        '''
+        tensor_array_read(self.name, index, output)
+
+
+    def size(self, output):
+        '''
+        Return the number of values.
+
+        @output: str
+            name of a scalar tensor
+        '''
+        tensor_array_size(self.name, output)
+```
 
 ## LoDTensor-related Supports
-The `RecurrentGradientMachine` in Paddle serves as a flexible RNN layer; it takes variant length sequences as input, 
-because each step of RNN could only take a tensor-represented batch of data as input, 
+The `RecurrentGradientMachine` in Paddle serves as a flexible RNN layer; it takes varience-length sequences as input, and output sequences too.
+
+Since each step of RNN can only take a tensor-represented batch of data as input, 
 some preprocess should be taken on the inputs such as sorting the sentences by their length in descending order and cut each word and pack to new batches.
 
-Such cut-like operations can be embedded into `TensorArray` as general methods called `unpack` and `pack`.
+Such cut-like operations can be embedded into `TensorArray` as general methods called `unpack` and `pack`,
+these two operations are similar to `stack` and `unstack` except that they operate on variable-length sequences formated as a LoD tensor rather than a tensor.
+
+Some definitions are like
+
+```python
+def unpack(level):
+    '''
+    Split LodTensor in some `level` and generate batches, if set `sort_by_length`,
+    will sort by length.
 
-With these two methods, a variant-sentence-RNN can be implemented like
+    Returns:
+        - a new `TensorArray`, whose values are LodTensors and represents batches
+          of data.
+        - an int32 Tensor, which stores the map from the new batch's indices to
+          original LoDTensor
+    '''
+    pass
+
+def pack(level, indices_map):
+    '''
+    Recover the original LoD-arranged LoDTensor with the values in a `TensorArray`
+    and `level` and `indices_map`.
+    '''
+    pass
+```
+
+With these two methods, a varience-length sentence supported RNN can be implemented like
 
 ```c++
 // input is the varient-length data
@@ -58,16 +269,3 @@ LoDTensor rnn_output = ta.pack(ta, indice_map);
 ```
 the code above shows that by embedding the LoDTensor-related preprocess operations into `TensorArray`,
 the implementation of a RNN that supports varient-length sentences is far more concise than `RecurrentGradientMachine` because the latter mixes all the codes together, hard to read and extend.
-
-
-some details are as follows.
-
-### unpack(level, sort_by_length)
-Split LodTensor in some `level` and generate batches, if set `sort_by_length`, will sort by length.
-
-Returns:
-
-- a new `TensorArray`, whose values are LodTensors and represents batches of data.
-- an int32 Tensor, which stores the map from the new batch's indices to original LoDTensor
-### pack(level, indices_map)
-Recover the original LoD-arranged LoDTensor with the values in a `TensorArray` and `level` and `indices_map`.
diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt
index 7dc9d5c804a348e310989b0aca0683ce383c5447..586ecc70e6fc76d48600875ed46363944a219dd8 100644
--- a/paddle/framework/CMakeLists.txt
+++ b/paddle/framework/CMakeLists.txt
@@ -26,10 +26,8 @@ cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto proto_desc)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope proto_desc)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
 
-cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS operator proto_desc)
-cc_library(op_registry SRCS op_registry.cc DEPS grad_op_builder op_proto_maker op_info)
+cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator)
 cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
-cc_test(grad_op_builder_test SRCS grad_op_builder_test.cc DEPS grad_op_builder op_registry sum_op)
 
 py_proto_compile(framework_py_proto SRCS framework.proto)
 # Generate an empty __init__.py to make framework_py_proto as a valid python module.
diff --git a/paddle/framework/backward.cc b/paddle/framework/backward.cc
index c0188c0e55a6db04a86d80c216fd162ad5259593..c970e01dd19d80e9a47f315a05a920ba15585c90 100644
--- a/paddle/framework/backward.cc
+++ b/paddle/framework/backward.cc
@@ -13,10 +13,13 @@
    limitations under the License. */
 
 #include "paddle/framework/backward.h"
+#include "paddle/operators/net_op.h"
 
+#include <deque>
 #include <list>
 #include <memory>
 
+#include "paddle/framework/block_desc.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/net_op.h"
 #include "paddle/operators/recurrent_op.h"
@@ -24,6 +27,35 @@
 namespace paddle {
 namespace framework {
 
+static inline std::unique_ptr<OperatorBase> CreateGradOp(
+    const OperatorBase& op) {
+  OpDescBind op_desc;
+  op_desc.SetInputMap(op.Inputs());
+  op_desc.SetOutputMap(op.Outputs());
+  op_desc.SetType(op.Type());
+  op_desc.SetAttrMap(op.Attrs());
+  auto& info = OpInfoMap::Instance().Get(op.Type());
+  auto grad_descs = info.GradOpMaker()(op_desc);
+  std::vector<std::unique_ptr<OperatorBase>> grad_ops;
+  grad_ops.reserve(grad_descs.size());
+  std::transform(grad_descs.begin(), grad_descs.end(),
+                 std::back_inserter(grad_ops),
+                 [](const std::unique_ptr<OpDescBind>& grad_desc) {
+                   return OpRegistry::CreateOp(*grad_desc);
+                 });
+  PADDLE_ENFORCE(!grad_ops.empty());
+  if (grad_ops.size() == 1) {
+    return std::move(grad_ops[0]);
+  } else {
+    auto net_op = new operators::NetOp();
+    for (auto& grad_op : grad_ops) {
+      net_op->AppendOp(std::move(grad_op));
+    }
+    net_op->CompleteAddOp();
+    return std::unique_ptr<OperatorBase>(net_op);
+  }
+}
+
 template <typename Map, typename T>
 static void ForEachVarName(const Map& names, T callback) {
   for (auto& name : names) {
@@ -171,7 +203,7 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
       net->InsertOp(pos.first + 1, std::move(pos.second));
     }
   } else {
-    std::unique_ptr<OperatorBase> grad_op(OpRegistry::CreateGradOp(forwardOp));
+    std::unique_ptr<OperatorBase> grad_op(CreateGradOp(forwardOp));
 
     ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net, &grad_op](
                                           const std::string& grad_input) {
@@ -240,5 +272,145 @@ std::unique_ptr<OperatorBase> Backward(
   return BackwardRecursive(forwardOp, no_grad_names, uid);
 }
 
+// ====================================  //
+
+static bool AllGradInSet(const std::vector<std::string>& names,
+                         const std::unordered_set<std::string>& set) {
+  for (const std::string& name : names) {
+    if (!set.count(GradVarName(name))) {
+      return false;
+    }
+  }
+  return true;
+}
+
+std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
+    const std::unique_ptr<OpDescBind>& op_desc,
+    std::unordered_set<std::string>& no_grad_vars) {
+  std::vector<std::unique_ptr<OpDescBind>> grad_op_descs;
+  // All input gradients of forwarding operator do not need to calculat.
+  const std::vector<std::string>& inputs = op_desc->InputArgumentNames();
+  if (AllGradInSet(inputs, no_grad_vars)) {
+    return grad_op_descs;  // empty vector
+  }
+  // All output gradients of forwarding operator do not need to calculate.
+  const std::vector<std::string>& outputs = op_desc->OutputArgumentNames();
+  if (AllGradInSet(outputs, no_grad_vars)) {
+    for (const std::string& name : inputs) {
+      no_grad_vars.insert(GradVarName(name));
+    }
+    return grad_op_descs;  // empty vector
+  }
+
+  grad_op_descs = OpRegistry::CreateGradOpDescs(*op_desc);
+
+  std::list<std::unique_ptr<OpDescBind>> pending_fill_zeros_ops;
+  for (auto& desc : grad_op_descs) {
+    for (const std::string& in_name : desc->InputArgumentNames()) {
+      if (no_grad_vars.count(in_name)) {
+        std::string prefix = in_name.substr(
+            0, in_name.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
+        std::string new_name = prefix + kZeroVarSuffix;
+        desc->Rename(in_name, new_name);
+        std::unique_ptr<OpDescBind> fill_zeros_op(new OpDescBind(
+            "fill_zeros_like", {{"X", {prefix}}}, {{"Y", {new_name}}}, {}));
+        pending_fill_zeros_ops.push_back(std::move(fill_zeros_op));
+      }
+    }
+    for (const std::string& out_name : desc->OutputArgumentNames()) {
+      if (no_grad_vars.count(out_name)) {
+        desc->Rename(out_name, kEmptyVarName);
+      }
+    }
+  }
+
+  for (auto& p : pending_fill_zeros_ops) {
+    grad_op_descs.insert(grad_op_descs.begin(), std::move(p));
+  }
+  return grad_op_descs;
+}
+
+std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
+    ProgramDescBind& program_desc, int block_idx,
+    std::unordered_set<std::string>& no_grad_vars) {
+  BlockDescBind* cur_block = program_desc.Block(block_idx);
+  std::deque<std::unique_ptr<OpDescBind>>& op_descs = cur_block->ops_;
+  std::unordered_map<std::string, std::vector<size_t>> dup_out_ops;
+  size_t grad_desc_idx = 0;
+  std::vector<std::unique_ptr<OpDescBind>> backward_descs;
+  for (auto it = op_descs.rbegin(); it != op_descs.rend(); ++it) {
+    std::vector<std::unique_ptr<OpDescBind>> op_grads =
+        MakeOpGrad(*it, no_grad_vars);
+
+    if ((*it)->Type() == "recurrent") {
+      PADDLE_ENFORCE_EQ(
+          op_grads.size(), size_t(1),
+          "rnn_op's gradient process should contain only one op.");
+      int step_block_idx = (*it)->GetBlockAttr("stop_block");
+      auto backward_block_op_descs =
+          MakeBlockBackward(program_desc, step_block_idx, no_grad_vars);
+      BlockDescBind* backward_block = program_desc.AppendBlock(*cur_block);
+      for (auto& ptr : backward_block_op_descs) {
+        backward_block->ops_.push_back(std::move(ptr));
+      }
+      op_grads[0]->SetBlockAttr("step_block", *backward_block);
+    }
+
+    for (const auto& desc : op_grads) {
+      for (const std::string& out_name : desc->OutputArgumentNames()) {
+        dup_out_ops[out_name].emplace_back(grad_desc_idx);
+      }
+      ++grad_desc_idx;
+    }
+    std::transform(
+        op_grads.begin(), op_grads.end(), std::back_inserter(backward_descs),
+        [](std::unique_ptr<OpDescBind>& ptr) { return std::move(ptr); });
+  }
+  // Check whether some variables are written more than once
+  std::list<std::pair<size_t, std::unique_ptr<OpDescBind>>> pending_sum_ops;
+  for (const auto& dup : dup_out_ops) {
+    const std::string& out_name = dup.first;
+    const std::vector<size_t> dup_op = dup.second;
+    if (out_name != kEmptyVarName && dup_op.size() > 1) {
+      std::vector<std::string> sum_op_inputs;
+      for (size_t i = 0; i < dup_op.size(); ++i) {
+        std::string new_name = out_name + "@RENAME@" + std::to_string(i);
+        backward_descs[dup_op[i]]->Rename(out_name, new_name);
+        sum_op_inputs.emplace_back(new_name);
+      }
+      std::unique_ptr<OpDescBind> sum_op(new OpDescBind(
+          "sum", {{"X", sum_op_inputs}}, {{"Out", {out_name}}}, {}));
+      pending_sum_ops.push_back({dup_op.back(), std::move(sum_op)});
+    }
+  }
+  pending_sum_ops.sort(
+      [](const std::pair<size_t, std::unique_ptr<OpDescBind>>& a,
+         const std::pair<size_t, std::unique_ptr<OpDescBind>>& b) {
+        return a.first > b.first;
+      });
+  for (auto& p : pending_sum_ops) {
+    backward_descs.insert(backward_descs.begin() + p.first + 1,
+                          std::move(p.second));
+  }
+  return backward_descs;
+}
+
+void AppendBackward(ProgramDescBind& program_desc,
+                    const std::unordered_set<std::string>& no_grad_vars) {
+  std::unordered_set<std::string> no_grad_var_names;
+  no_grad_var_names.reserve(no_grad_vars.size() + 1);
+  no_grad_var_names.insert(std::string(kEmptyVarName) + kGradVarSuffix);
+  for (auto& name : no_grad_vars) {
+    no_grad_var_names.insert(GradVarName(name));
+  }
+  const int root_block_idx = 0;
+  auto backward_op_descs =
+      MakeBlockBackward(program_desc, root_block_idx, no_grad_var_names);
+  auto& forw_op_descs = program_desc.Block(root_block_idx)->ops_;
+  for (auto& ptr : backward_op_descs) {
+    forw_op_descs.push_back(std::move(ptr));
+  }
+}
+
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/framework/backward.h b/paddle/framework/backward.h
index 1ecf69881b3126c2904920b9f4b77bfcccc9cf86..7ffe4c28103f9d6a9f179422d1beb86106ef786e 100644
--- a/paddle/framework/backward.h
+++ b/paddle/framework/backward.h
@@ -13,8 +13,11 @@
    limitations under the License. */
 
 #pragma once
+
 #include <unordered_set>
-#include "operator.h"
+#include "paddle/framework/operator.h"
+#include "paddle/framework/program_desc.h"
+
 namespace paddle {
 namespace framework {
 
@@ -23,5 +26,9 @@ namespace framework {
 extern std::unique_ptr<OperatorBase> Backward(
     const OperatorBase& forwardOp,
     const std::unordered_set<std::string>& no_grad_vars);
+
+void AppendBackward(ProgramDescBind& program_desc,
+                    const std::unordered_set<std::string>& no_grad_vars);
+
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/framework/backward_test.cc b/paddle/framework/backward_test.cc
index a36e7bde8c61c6a8bfc6eefd46ce5e09907476ef..30225a4a99d993c4f12a5e0d276bda18acbc360e 100644
--- a/paddle/framework/backward_test.cc
+++ b/paddle/framework/backward_test.cc
@@ -15,30 +15,42 @@
 #include "paddle/framework/backward.h"
 
 #include <gtest/gtest.h>
+#include "paddle/framework/block_desc.h"
+#include "paddle/framework/op_desc.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/net_op.h"
 
 namespace paddle {
 namespace framework {
 
-using OperatorBase = framework::OperatorBase;
-using OpProtoAndCheckerMaker = framework::OpProtoAndCheckerMaker;
-using OpProto = framework::OpProto;
-using OpAttrChecker = framework::OpAttrChecker;
-using Scope = framework::Scope;
 using DeviceContext = platform::DeviceContext;
 
 class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
  public:
   RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "Input X of Add").NotInGradient();
-    AddInput("b", "Bias of Add").NotInGradient();
-    AddOutput("Out", "Out of Add").NotInGradient();
+    AddInput("X", "Input X of Add");
+    AddInput("b", "Bias of Add");
+    AddOutput("Out", "Out of Add");
     AddComment("Add Op");
   }
 };
 
+class RowWiseAddGradMaker : public SingleGradOpDescMaker {
+ public:
+  using SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<OpDescBind> Apply() const override {
+    auto grad_op = new OpDescBind();
+    grad_op->SetInput(GradVarName("Out"), OutputGrad("Out"));
+    grad_op->SetOutput(GradVarName("X"), InputGrad("X"));
+    grad_op->SetOutput(GradVarName("b"), InputGrad("b"));
+    grad_op->SetType("rowwise_add_grad");
+    return std::unique_ptr<OpDescBind>(grad_op);
+  }
+};
+
 class MulOpMaker : public OpProtoAndCheckerMaker {
  public:
   MulOpMaker(OpProto *proto, OpAttrChecker *op_checker)
@@ -137,10 +149,20 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   SumOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "the input tensors of sum operator.")
-        .AsDuplicable()
-        .NotInGradient();
-    AddOutput("Out", "the output tensor of sum operator.").NotInGradient();
+    AddInput("X", "the input tensors of sum operator.").AsDuplicable();
+    AddOutput("Out", "the output tensor of sum operator.");
+    AddComment("");
+  }
+};
+
+class MultInOutOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  MultInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "x");
+    AddInput("H", "h");
+    AddOutput("Y", "y");
+    AddOutput("Z", "z");
     AddComment("");
   }
 };
@@ -151,8 +173,9 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker {
 namespace f = paddle::framework;
 namespace ops = paddle::operators;
 using EnforceNotMet = paddle::platform::EnforceNotMet;
-REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, rowwise_add_grad,
-            f::NOP);
+REGISTER_OPERATOR(rowwise_add, f::NOP, f::RowWiseAddOpMaker,
+                  f::RowWiseAddGradMaker);
+REGISTER_OPERATOR(rowwise_add_grad, f::NOP);
 REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP);
 REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP);
 REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker);
@@ -161,17 +184,7 @@ REGISTER_OP(sum, f::NOP, f::SumOpMaker, sum_grad, f::NOP);
 REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker);
 REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad,
             f::NOP);
-
-TEST(Backward, simple_op_grad) {
-  auto fwd = f::OpRegistry::CreateOp(
-      "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
-  ASSERT_NE(fwd, nullptr);
-  auto gop = f::OpRegistry::CreateGradOp(*fwd);
-  ASSERT_EQ(1UL, gop->Inputs().size());
-  ASSERT_EQ("rowwise_add_grad", gop->Type());
-  ASSERT_EQ(f::GradVarName("x"), gop->Output(f::GradVarName("X")));
-  ASSERT_EQ(f::GradVarName("b"), gop->Output(f::GradVarName("b")));
-}
+REGISTER_OP(mult_in_out, f::NOP, f::MultInOutOpMaker, mult_in_out_grad, f::NOP);
 
 TEST(Backward, simple_op_not_need_grad) {
   auto fwd = f::OpRegistry::CreateOp(
@@ -289,17 +302,6 @@ TEST(Backward, net_shared_weight) {
   ASSERT_EQ("sum", bwd_net->ops_[2]->Type());
 }
 
-TEST(Backward, op_register_grad_not_for_network) {
-  auto fwd =
-      f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {"b"}}},
-                              {{"mul_result", {"mul_out"}},
-                               {"add_result", {"add_out"}},
-                               {"Out", {"out1"}}},
-                              {{"temporary_index", std::vector<int>{0, 1}}});
-
-  ASSERT_THROW(f::OpRegistry::CreateGradOp(*fwd), EnforceNotMet);
-}
-
 TEST(Backward, op_all_input_are_not_need) {
   auto fwd = f::OpRegistry::CreateOp(
       "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
@@ -402,3 +404,293 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
   EXPECT_EQ(bwd_net->ops_[2]->Inputs(all).size(), 0UL);
   EXPECT_EQ(bwd_net->ops_[2]->Outputs(all).size(), 0UL);
 }
+
+// =================================== //
+
+f::ProgramDesc *GetNewProgramDesc() {
+  auto *program_desc = new f::ProgramDesc();
+  auto *root_block = program_desc->add_blocks();
+  root_block->set_idx(0);
+  root_block->set_parent_idx(-1);
+  return program_desc;
+}
+
+TEST(Backward, simple_single_op) {
+  f::ProgramDesc *program_desc = GetNewProgramDesc();
+  f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);
+  f::BlockDescBind *block = program.Block(0);
+  f::OpDescBind *op = block->AppendOp();
+  op->SetType("rowwise_add");
+  op->SetInput("X", {"x"});
+  op->SetInput("b", {"b"});
+  op->SetOutput("Out", {"out"});
+
+  AppendBackward(program, {});
+
+  ASSERT_EQ(block->AllOps().size(), 2UL);
+  f::OpDescBind *grad_op = block->AllOps()[1];
+  EXPECT_EQ(grad_op->Type(), "rowwise_add_grad");
+  ASSERT_EQ(grad_op->InputNames().size(), 1UL);
+  ASSERT_EQ(grad_op->OutputNames().size(), 2UL);
+  EXPECT_EQ(grad_op->Input(f::GradVarName("Out")),
+            std::vector<std::string>({f::GradVarName("out")}));
+  EXPECT_EQ(grad_op->Output(f::GradVarName("X")),
+            std::vector<std::string>({f::GradVarName("x")}));
+  EXPECT_EQ(grad_op->Output(f::GradVarName("b")),
+            std::vector<std::string>({f::GradVarName("b")}));
+}
+
+TEST(Backward, simple_mult_op) {
+  f::ProgramDesc *program_desc = GetNewProgramDesc();
+  f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);
+  f::BlockDescBind *block = program.Block(0);
+  f::OpDescBind *op1 = block->AppendOp();
+  op1->SetType("rowwise_add");
+  op1->SetInput("X", {"x1"});
+  op1->SetInput("b", {"b1"});
+  op1->SetOutput("Out", {"out1"});
+
+  f::OpDescBind *op2 = block->AppendOp();
+  op2->SetType("mul");
+  op2->SetInput("X", {"out1"});
+  op2->SetInput("Y", {"y2"});
+  op2->SetOutput("Out", {"out2"});
+
+  f::OpDescBind *op3 = block->AppendOp();
+  op3->SetType("rowwise_add");
+  op3->SetInput("X", {"out2"});
+  op3->SetInput("b", {"b3"});
+  op3->SetOutput("Out", {"out3"});
+
+  AppendBackward(program, {});
+
+  ASSERT_EQ(block->AllOps().size(), 6UL);
+  f::OpDescBind *grad_op1 = block->AllOps()[5];
+  EXPECT_EQ(grad_op1->Type(), "rowwise_add_grad");
+  ASSERT_EQ(grad_op1->InputNames().size(), 1UL);
+  ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
+  EXPECT_EQ(grad_op1->Input(f::GradVarName("Out")),
+            std::vector<std::string>({f::GradVarName("out1")}));
+  EXPECT_EQ(grad_op1->Output(f::GradVarName("X")),
+            std::vector<std::string>({f::GradVarName("x1")}));
+  EXPECT_EQ(grad_op1->Output(f::GradVarName("b")),
+            std::vector<std::string>({f::GradVarName("b1")}));
+
+  f::OpDescBind *grad_op2 = block->AllOps()[4];
+  EXPECT_EQ(grad_op2->Type(), "mul_grad");
+  ASSERT_EQ(grad_op2->InputNames().size(), 4UL);
+  ASSERT_EQ(grad_op2->OutputNames().size(), 2UL);
+  EXPECT_EQ(grad_op2->Input("X"), std::vector<std::string>({"out1"}));
+  EXPECT_EQ(grad_op2->Input("Y"), std::vector<std::string>({"y2"}));
+  EXPECT_EQ(grad_op2->Input("Out"), std::vector<std::string>({"out2"}));
+  EXPECT_EQ(grad_op2->Input(f::GradVarName("Out")),
+            std::vector<std::string>({f::GradVarName("out2")}));
+  EXPECT_EQ(grad_op2->Output(f::GradVarName("X")),
+            std::vector<std::string>({f::GradVarName("out1")}));
+  EXPECT_EQ(grad_op2->Output(f::GradVarName("Y")),
+            std::vector<std::string>({f::GradVarName("y2")}));
+
+  f::OpDescBind *grad_op3 = block->AllOps()[3];
+  EXPECT_EQ(grad_op3->Type(), "rowwise_add_grad");
+  ASSERT_EQ(grad_op3->InputNames().size(), 1UL);
+  ASSERT_EQ(grad_op3->OutputNames().size(), 2UL);
+  EXPECT_EQ(grad_op3->Input(f::GradVarName("Out")),
+            std::vector<std::string>({f::GradVarName("out3")}));
+  EXPECT_EQ(grad_op3->Output(f::GradVarName("X")),
+            std::vector<std::string>({f::GradVarName("out2")}));
+  EXPECT_EQ(grad_op3->Output(f::GradVarName("b")),
+            std::vector<std::string>({f::GradVarName("b3")}));
+}
+
+TEST(Backward, intermedia_var_no_grad) {
+  f::ProgramDesc *program_desc = GetNewProgramDesc();
+  f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);
+  f::BlockDescBind *block = program.Block(0);
+  f::OpDescBind *op1 = block->AppendOp();
+  op1->SetType("rowwise_add");
+  op1->SetInput("X", {"x1"});
+  op1->SetInput("b", {"b1"});
+  op1->SetOutput("Out", {"out1"});
+
+  f::OpDescBind *op2 = block->AppendOp();
+  op2->SetType("mul");
+  op2->SetInput("X", {"x2"});
+  op2->SetInput("Y", {"y2"});
+  op2->SetOutput("Out", {"out2"});
+
+  f::OpDescBind *op3 = block->AppendOp();
+  op3->SetType("rowwise_add");
+  op3->SetInput("X", {"out2"});
+  op3->SetInput("b", {"b3"});
+  op3->SetOutput("Out", {"out3"});
+
+  f::OpDescBind *op4 = block->AppendOp();
+  op4->SetType("mul");
+  op4->SetInput("X", {"out1"});
+  op4->SetInput("Y", {"out3"});
+  op4->SetOutput("Out", {"out4"});
+
+  AppendBackward(program, {"out3"});
+
+  ASSERT_EQ(block->AllOps().size(), 6UL);
+  f::OpDescBind *grad_op1 = block->AllOps()[5];
+  EXPECT_EQ(grad_op1->Type(), "rowwise_add_grad");
+  ASSERT_EQ(grad_op1->InputNames().size(), 1UL);
+  ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
+  EXPECT_EQ(grad_op1->Input(f::GradVarName("Out")),
+            std::vector<std::string>({f::GradVarName("out1")}));
+  EXPECT_EQ(grad_op1->Output(f::GradVarName("X")),
+            std::vector<std::string>({f::GradVarName("x1")}));
+  EXPECT_EQ(grad_op1->Output(f::GradVarName("b")),
+            std::vector<std::string>({f::GradVarName("b1")}));
+
+  f::OpDescBind *grad_op4 = block->AllOps()[4];
+  EXPECT_EQ(grad_op4->Type(), "mul_grad");
+  ASSERT_EQ(grad_op4->InputNames().size(), 4UL);
+  ASSERT_EQ(grad_op4->OutputNames().size(), 2UL);
+  EXPECT_EQ(grad_op4->Input("X"), std::vector<std::string>({"out1"}));
+  EXPECT_EQ(grad_op4->Input("Y"), std::vector<std::string>({"out3"}));
+  EXPECT_EQ(grad_op4->Input("Out"), std::vector<std::string>({"out4"}));
+  EXPECT_EQ(grad_op4->Input(f::GradVarName("Out")),
+            std::vector<std::string>({f::GradVarName("out4")}));
+  EXPECT_EQ(grad_op4->Output(f::GradVarName("X")),
+            std::vector<std::string>({f::GradVarName("out1")}));
+  EXPECT_EQ(grad_op4->Output(f::GradVarName("Y")),
+            std::vector<std::string>({f::kEmptyVarName}));
+}
+
+TEST(Backward, var_no_grad) {
+  f::ProgramDesc *program_desc = GetNewProgramDesc();
+  f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);
+  f::BlockDescBind *block = program.Block(0);
+  f::OpDescBind *op1 = block->AppendOp();
+  op1->SetType("mult_in_out");
+  op1->SetInput("X", {"x1"});
+  op1->SetInput("H", {"h1"});
+  op1->SetOutput("Y", {"y1"});
+  op1->SetOutput("Z", {"z1"});
+
+  f::OpDescBind *op2 = block->AppendOp();
+  op2->SetType("mult_in_out");
+  op2->SetInput("X", {"y1"});
+  op2->SetInput("H", {"z1"});
+  op2->SetOutput("Y", {"y2"});
+  op2->SetOutput("Z", {"z2"});
+
+  AppendBackward(program, {"z1"});
+
+  ASSERT_EQ(block->AllOps().size(), 5UL);
+  f::OpDescBind *grad_op2 = block->AllOps()[2];
+  ASSERT_EQ(grad_op2->Type(), "mult_in_out_grad");
+  ASSERT_EQ(grad_op2->InputNames().size(), 6UL);
+  ASSERT_EQ(grad_op2->OutputNames().size(), 2UL);
+  EXPECT_EQ(grad_op2->Input("X"), std::vector<std::string>({"y1"}));
+  EXPECT_EQ(grad_op2->Input("H"), std::vector<std::string>({"z1"}));
+  EXPECT_EQ(grad_op2->Input("Y"), std::vector<std::string>({"y2"}));
+  EXPECT_EQ(grad_op2->Input("Z"), std::vector<std::string>({"z2"}));
+  EXPECT_EQ(grad_op2->Input(f::GradVarName("Y")),
+            std::vector<std::string>({f::GradVarName("y2")}));
+  EXPECT_EQ(grad_op2->Input(f::GradVarName("Z")),
+            std::vector<std::string>({f::GradVarName("z2")}));
+  EXPECT_EQ(grad_op2->Output(f::GradVarName("X")),
+            std::vector<std::string>({f::GradVarName("y1")}));
+  EXPECT_EQ(grad_op2->Output(f::GradVarName("H")),
+            std::vector<std::string>({f::kEmptyVarName}));
+
+  f::OpDescBind *fill_zero_op = block->AllOps()[3];
+  ASSERT_EQ(fill_zero_op->Type(), "fill_zeros_like");
+  ASSERT_EQ(fill_zero_op->InputNames().size(), 1UL);
+  ASSERT_EQ(fill_zero_op->OutputNames().size(), 1UL);
+  EXPECT_EQ(fill_zero_op->Input("X"), std::vector<std::string>({"z1"}));
+  EXPECT_EQ(fill_zero_op->Output("Y"),
+            std::vector<std::string>({std::string("z1") + f::kZeroVarSuffix}));
+
+  f::OpDescBind *grad_op1 = block->AllOps()[4];
+  ASSERT_EQ(grad_op1->Type(), "mult_in_out_grad");
+  ASSERT_EQ(grad_op1->InputNames().size(), 6UL);
+  ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
+  EXPECT_EQ(grad_op1->Input("X"), std::vector<std::string>({"x1"}));
+  EXPECT_EQ(grad_op1->Input("H"), std::vector<std::string>({"h1"}));
+  EXPECT_EQ(grad_op1->Input("Y"), std::vector<std::string>({"y1"}));
+  EXPECT_EQ(grad_op1->Input("Z"), std::vector<std::string>({"z1"}));
+  EXPECT_EQ(grad_op1->Input(f::GradVarName("Y")),
+            std::vector<std::string>({f::GradVarName("y1")}));
+  EXPECT_EQ(grad_op1->Input(f::GradVarName("Z")),
+            std::vector<std::string>({std::string("z1") + f::kZeroVarSuffix}));
+  EXPECT_EQ(grad_op1->Output(f::GradVarName("X")),
+            std::vector<std::string>({f::GradVarName("x1")}));
+  EXPECT_EQ(grad_op1->Output(f::GradVarName("H")),
+            std::vector<std::string>({f::GradVarName("h1")}));
+}
+
+TEST(Backward, shared_var) {
+  f::ProgramDesc *program_desc = GetNewProgramDesc();
+  f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);
+  f::BlockDescBind *block = program.Block(0);
+  f::OpDescBind *op1 = block->AppendOp();
+  op1->SetType("rowwise_add");
+  op1->SetInput("X", {"x1"});
+  op1->SetInput("b", {"b1"});
+  op1->SetOutput("Out", {"out1"});
+
+  f::OpDescBind *op2 = block->AppendOp();
+  op2->SetType("mul");
+  op2->SetInput("X", {"out1"});
+  op2->SetInput("Y", {"y2"});
+  op2->SetOutput("Out", {"out2"});
+
+  f::OpDescBind *op3 = block->AppendOp();
+  op3->SetType("rowwise_add");
+  op3->SetInput("X", {"out1"});
+  op3->SetInput("b", {"b3"});
+  op3->SetOutput("Out", {"out3"});
+
+  AppendBackward(program, {});
+
+  ASSERT_EQ(block->AllOps().size(), 7UL);
+  f::OpDescBind *grad_op3 = block->AllOps()[3];
+  ASSERT_EQ(grad_op3->Type(), "rowwise_add_grad");
+  ASSERT_EQ(grad_op3->InputNames().size(), 1UL);
+  ASSERT_EQ(grad_op3->OutputNames().size(), 2UL);
+  EXPECT_EQ(grad_op3->Input(f::GradVarName("Out")),
+            std::vector<std::string>({f::GradVarName("out3")}));
+  EXPECT_EQ(grad_op3->Output(f::GradVarName("X")),
+            std::vector<std::string>({f::GradVarName("out1") + "@RENAME@0"}));
+  EXPECT_EQ(grad_op3->Output(f::GradVarName("b")),
+            std::vector<std::string>({f::GradVarName("b3")}));
+
+  f::OpDescBind *grad_op4 = block->AllOps()[4];
+  ASSERT_EQ(grad_op4->Type(), "mul_grad");
+  ASSERT_EQ(grad_op4->InputNames().size(), 4UL);
+  ASSERT_EQ(grad_op4->OutputNames().size(), 2UL);
+  EXPECT_EQ(grad_op4->Input("X"), std::vector<std::string>({"out1"}));
+  EXPECT_EQ(grad_op4->Input("Y"), std::vector<std::string>({"y2"}));
+  EXPECT_EQ(grad_op4->Input("Out"), std::vector<std::string>({"out2"}));
+  EXPECT_EQ(grad_op4->Input(f::GradVarName("Out")),
+            std::vector<std::string>({f::GradVarName("out2")}));
+  EXPECT_EQ(grad_op4->Output(f::GradVarName("X")),
+            std::vector<std::string>({f::GradVarName("out1") + "@RENAME@1"}));
+  EXPECT_EQ(grad_op4->Output(f::GradVarName("Y")),
+            std::vector<std::string>({f::GradVarName("y2")}));
+
+  f::OpDescBind *sum_op = block->AllOps()[5];
+  ASSERT_EQ(sum_op->Type(), "sum");
+  ASSERT_EQ(sum_op->InputNames().size(), 1UL);
+  ASSERT_EQ(sum_op->OutputNames().size(), 1UL);
+  EXPECT_EQ(sum_op->Input("X"),
+            std::vector<std::string>({f::GradVarName("out1") + "@RENAME@0",
+                                      f::GradVarName("out1") + "@RENAME@1"}));
+  EXPECT_EQ(sum_op->Output("Out"),
+            std::vector<std::string>({f::GradVarName("out1")}));
+
+  f::OpDescBind *grad_op1 = block->AllOps()[6];
+  ASSERT_EQ(grad_op1->Type(), "rowwise_add_grad");
+  ASSERT_EQ(grad_op1->InputNames().size(), 1UL);
+  ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
+  EXPECT_EQ(grad_op1->Input(f::GradVarName("Out")),
+            std::vector<std::string>({f::GradVarName("out1")}));
+  EXPECT_EQ(grad_op1->Output(f::GradVarName("X")),
+            std::vector<std::string>({f::GradVarName("x1")}));
+  EXPECT_EQ(grad_op1->Output(f::GradVarName("b")),
+            std::vector<std::string>({f::GradVarName("b1")}));
+}
\ No newline at end of file
diff --git a/paddle/framework/block_desc.h b/paddle/framework/block_desc.h
index b646ad5f3b67e22b933a56d88e4a1bf6e74a124e..2de270f60ec2ae981335f1adb204cfc3bf78c622 100644
--- a/paddle/framework/block_desc.h
+++ b/paddle/framework/block_desc.h
@@ -32,6 +32,14 @@ class ProgramDescBind;
 
 class BlockDescBind {
  public:
+  friend std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
+      ProgramDescBind &program_desc, int block_idx,
+      std::unordered_set<std::string> &no_grad_vars);
+
+  friend void AppendBackward(
+      ProgramDescBind &program_desc,
+      const std::unordered_set<std::string> &no_grad_vars);
+
   BlockDescBind(ProgramDescBind *prog, BlockDesc *desc)
       : prog_(prog), desc_(desc), need_update_(false) {}
 
diff --git a/paddle/framework/framework.proto b/paddle/framework/framework.proto
index d696cdf5ad68118c98577fcaf1aed1cf013df296..ac2827e54773f811eb855c092e3c0ed2fab06dd3 100644
--- a/paddle/framework/framework.proto
+++ b/paddle/framework/framework.proto
@@ -66,7 +66,6 @@ message OpProto {
 
     optional bool duplicable = 3 [ default = false ];
     optional bool intermediate = 4 [ default = false ];
-    optional bool not_in_gradient = 5 [ default = false ];
   }
 
   // AttrProto describes the C++ type Attribute.
@@ -116,4 +115,7 @@ message BlockDesc {
   repeated OpDesc ops = 4;
 }
 
+// Please refer to
+// https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/program.md
+// for more details.
 message ProgramDesc { repeated BlockDesc blocks = 1; }
diff --git a/paddle/framework/grad_op_builder.cc b/paddle/framework/grad_op_builder.cc
deleted file mode 100644
index 3661ce41beba1328d1b1cdd9f0f913e693af9cff..0000000000000000000000000000000000000000
--- a/paddle/framework/grad_op_builder.cc
+++ /dev/null
@@ -1,97 +0,0 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOpArgType::OUT WARRANTIES OR CONDITIONS OF ANY KOpArgType::IND, either
-express or implied. See the License for the specific language governing
-permissions and limitations under the License. */
-
-#include "paddle/framework/grad_op_builder.h"
-#include "paddle/framework/op_registry.h"
-
-namespace paddle {
-namespace framework {
-enum class OpArgType { IN, OUT };
-
-static void TransOpArg(const OperatorBase* src_op, const OpArgType& src_type,
-                       bool is_grad, VariableNameMap* vars) {
-  const auto& src_inout =
-      src_type == OpArgType::IN ? src_op->Inputs() : src_op->Outputs();
-  auto& dst_inout = *vars;
-  auto& proto = OpInfoMap::Instance().Get(src_op->Type()).Proto();
-  const auto& src_arg_list =
-      src_type == OpArgType::IN ? proto.inputs() : proto.outputs();
-  for (const auto& arg : src_arg_list) {
-    if (arg.not_in_gradient() && !is_grad) continue;
-    const std::string src_name = arg.name();
-    std::string dst_name = is_grad ? GradVarName(src_name) : src_name;
-    dst_inout[dst_name].reserve(src_inout.at(src_name).size());
-    for (auto& var_name : src_inout.at(src_name)) {
-      std::string s = is_grad ? GradVarName(var_name) : var_name;
-      dst_inout[dst_name].emplace_back(s);
-    }
-  }
-}
-
-OperatorBase* BuildGradOp(const OperatorBase* op) {
-  auto& info = OpInfoMap::Instance().Get(op->Type());
-  PADDLE_ENFORCE(info.HasGradientOp());
-
-  VariableNameMap inputs;
-  VariableNameMap outputs;
-  TransOpArg(op, OpArgType::IN, false, &inputs);   // I
-  TransOpArg(op, OpArgType::OUT, false, &inputs);  // O
-  TransOpArg(op, OpArgType::OUT, true, &inputs);   // OG
-  TransOpArg(op, OpArgType::IN, true, &outputs);   // IG
-
-  auto& grad_info = OpInfoMap::Instance().Get(info.grad_op_type_);
-  return grad_info.Creator()(info.grad_op_type_, inputs, outputs, op->Attrs());
-}
-
-static void TransOpDescArg(const OpDescBind* src_op, const OpArgType& src_type,
-                           bool is_grad, OpDescBind* dst_op,
-                           const OpArgType& dst_type) {
-  PADDLE_ENFORCE(dst_op != nullptr,
-                 "Protobuf desc of gradient op must be initialized first.");
-  const auto& proto = OpInfoMap::Instance().Get(src_op->Type()).Proto();
-  const auto& src_arg_list =
-      src_type == OpArgType::IN ? proto.inputs() : proto.outputs();
-  for (const auto& arg : src_arg_list) {
-    if (arg.not_in_gradient() && !is_grad) continue;
-    const std::string src_name = arg.name();
-    std::vector<std::string> vars = src_type == OpArgType::IN
-                                        ? src_op->Input(src_name)
-                                        : src_op->Output(src_name);
-    if (is_grad) {
-      for (std::string& var : vars) {
-        var = GradVarName(var);
-      }
-    }
-    std::string dst_name = is_grad ? GradVarName(src_name) : src_name;
-    dst_type == OpArgType::IN ? dst_op->SetInput(dst_name, vars)
-                              : dst_op->SetOutput(dst_name, vars);
-  }
-}
-
-void CompleteGradOpDesc(const OpDescBind* forw_op, OpDescBind* grad_op) {
-  auto& info = OpInfoMap::Instance().Get(forw_op->Type());
-  PADDLE_ENFORCE(info.HasGradientOp());
-
-  grad_op->SetType(info.grad_op_type_);
-
-  TransOpDescArg(forw_op, OpArgType::IN, false, grad_op, OpArgType::IN);
-  TransOpDescArg(forw_op, OpArgType::OUT, false, grad_op, OpArgType::IN);
-  TransOpDescArg(forw_op, OpArgType::OUT, true, grad_op, OpArgType::IN);
-  TransOpDescArg(forw_op, OpArgType::IN, true, grad_op, OpArgType::OUT);
-
-  grad_op->SetAttrMap(forw_op->GetAttrMap());
-}
-
-}  // namespace framework
-}  // namespace paddle
diff --git a/paddle/framework/grad_op_builder.h b/paddle/framework/grad_op_builder.h
deleted file mode 100644
index b601406061f9f8f24302251c2144b07b6e65717f..0000000000000000000000000000000000000000
--- a/paddle/framework/grad_op_builder.h
+++ /dev/null
@@ -1,28 +0,0 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#pragma once
-
-#include "paddle/framework/op_desc.h"
-#include "paddle/framework/operator.h"
-
-namespace paddle {
-namespace framework {
-
-OperatorBase* BuildGradOp(const OperatorBase* op);
-
-void CompleteGradOpDesc(const OpDescBind* forw_op, OpDescBind* grad_op);
-
-}  // namespace framework
-}  // namespace paddle
diff --git a/paddle/framework/grad_op_builder_test.cc b/paddle/framework/grad_op_builder_test.cc
deleted file mode 100644
index 793780ea44b80a0d750d35cbea601bfbfd5ccda4..0000000000000000000000000000000000000000
--- a/paddle/framework/grad_op_builder_test.cc
+++ /dev/null
@@ -1,186 +0,0 @@
-#include "paddle/framework/grad_op_builder.h"
-#include <gtest/gtest.h>
-#include "paddle/framework/op_registry.h"
-#include "paddle/framework/operator.h"
-
-USE_OP(sum);
-
-namespace paddle {
-namespace framework {
-
-class MutiInOutOpMaker : public OpProtoAndCheckerMaker {
- public:
-  MutiInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("In1", "a single input");
-    AddInput("In2_mult", "a multiple input").AsDuplicable();
-    AddInput("In3", "another single input");
-    AddOutput("Out1", "a single output");
-    AddOutput("Out2_mult", "a multiple output").AsDuplicable();
-    AddComment("test op with multiple inputs and outputs");
-  }
-};
-
-class IOIgnoredOpMaker : public OpProtoAndCheckerMaker {
- public:
-  IOIgnoredOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("In1", "a single input");
-    AddInput("In2_mult", "a multiple input").AsDuplicable().NotInGradient();
-    AddInput("In3_mult", "another multiple input").AsDuplicable();
-    AddOutput("Out1_mult", "a multiple output").AsDuplicable();
-    AddOutput("Out2", "a single output").NotInGradient();
-    AddComment("op with inputs and outputs ignored in gradient calculating");
-  }
-};
-
-}  // namespace framework
-}  // namespace paddle
-
-namespace f = paddle::framework;
-
-REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker, mult_io_grad, f::NOP);
-REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker, io_ignored_grad, f::NOP);
-
-TEST(GradOpBuilder, MutiInOut) {
-  std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(
-      "mult_io", {{"In1", {"in1"}},
-                  {"In2_mult", {"in2_1", "in2_2", "in2_3"}},
-                  {"In3", {"in3"}}},
-      {{"Out1", {"out1"}}, {"Out2_mult", {"out2_1", "out2_2"}}}, {}));
-  std::shared_ptr<f::OperatorBase> grad_test_op =
-      f::OpRegistry::CreateGradOp(*test_op);
-
-  ASSERT_EQ(grad_test_op->Inputs().size(), 3UL + 2UL + 2UL);
-  EXPECT_EQ(grad_test_op->Input("In1"), "in1");
-  EXPECT_EQ(grad_test_op->Inputs("In2_mult"),
-            std::vector<std::string>({"in2_1", "in2_2", "in2_3"}));
-  EXPECT_EQ(grad_test_op->Input("In3"), "in3");
-  EXPECT_EQ(grad_test_op->Input("Out1"), "out1");
-  EXPECT_EQ(grad_test_op->Inputs("Out2_mult"),
-            std::vector<std::string>({"out2_1", "out2_2"}));
-  EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out1")),
-            f::GradVarName("out1"));
-  EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out2_mult")),
-            std::vector<std::string>(
-                {f::GradVarName("out2_1"), f::GradVarName("out2_2")}));
-
-  ASSERT_EQ(grad_test_op->Outputs().size(), 3UL);
-  EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
-  EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
-            std::vector<std::string>({f::GradVarName("in2_1"),
-                                      f::GradVarName("in2_2"),
-                                      f::GradVarName("in2_3")}));
-  EXPECT_EQ(grad_test_op->Output(f::GradVarName("In3")), f::GradVarName("in3"));
-}
-
-TEST(GradOpBuilder, IOIgnoredInGradient) {
-  std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(
-      "io_ignored", {{"In1", {"in1"}},
-                     {"In2_mult", {"in2_1", "in2_2"}},
-                     {"In3_mult", {"in3_1", "in3_2"}}},
-      {{"Out1_mult", {"out1_1", "out1_2"}}, {"Out2", {"out2"}}}, {}));
-  std::shared_ptr<f::OperatorBase> grad_test_op =
-      f::OpRegistry::CreateGradOp(*test_op);
-
-  // 'In2' and 'Out2' are ignored in gradient calculating
-  ASSERT_EQ(grad_test_op->Inputs().size(), 2UL + 1UL + 2UL);
-  EXPECT_EQ(grad_test_op->Input("In1"), "in1");
-  EXPECT_EQ(grad_test_op->Inputs("In3_mult"),
-            std::vector<std::string>({"in3_1", "in3_2"}));
-  EXPECT_EQ(grad_test_op->Inputs("Out1_mult"),
-            std::vector<std::string>({"out1_1", "out1_2"}));
-  EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out1_mult")),
-            std::vector<std::string>(
-                {f::GradVarName("out1_1"), f::GradVarName("out1_2")}));
-  EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out2")),
-            f::GradVarName("out2"));
-
-  ASSERT_EQ(grad_test_op->Outputs().size(), 3UL);
-  EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
-  EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
-            std::vector<std::string>(
-                {f::GradVarName("in2_1"), f::GradVarName("in2_2")}));
-  EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In3_mult")),
-            std::vector<std::string>(
-                {f::GradVarName("in3_1"), f::GradVarName("in3_2")}));
-}
-
-TEST(GradOpDescBuilder, MutiInOut) {
-  f::OpDescBind *forw_op = new f::OpDescBind();
-  forw_op->SetType("mult_io");
-  forw_op->SetInput("In1", {"in1"});
-  forw_op->SetInput("In2_mult", {"in2_1", "in2_2", "in2_3"});
-  forw_op->SetInput("In3", {"in3"});
-  forw_op->SetOutput("Out1", {"out1"});
-  forw_op->SetOutput("Out2_mult", {"out2_1", "out2_2"});
-
-  f::OpDescBind *grad_op = new f::OpDescBind();
-  f::CompleteGradOpDesc(forw_op, grad_op);
-
-  EXPECT_EQ(grad_op->Type(), "mult_io_grad");
-  ASSERT_EQ(grad_op->InputNames().size(), 3UL + 2UL + 2UL);
-  EXPECT_EQ(grad_op->Input("In1"), std::vector<std::string>({"in1"}));
-  EXPECT_EQ(grad_op->Input("In2_mult"),
-            std::vector<std::string>({"in2_1", "in2_2", "in2_3"}));
-  EXPECT_EQ(grad_op->Input("In3"), std::vector<std::string>({"in3"}));
-  EXPECT_EQ(grad_op->Input("Out1"), std::vector<std::string>({"out1"}));
-  EXPECT_EQ(grad_op->Input("Out2_mult"),
-            std::vector<std::string>({"out2_1", "out2_2"}));
-  EXPECT_EQ(grad_op->Input(f::GradVarName("Out1")),
-            std::vector<std::string>({f::GradVarName("out1")}));
-  EXPECT_EQ(grad_op->Input(f::GradVarName("Out2_mult")),
-            std::vector<std::string>(
-                {f::GradVarName("out2_1"), f::GradVarName("out2_2")}));
-
-  ASSERT_EQ(grad_op->OutputNames().size(), 3UL);
-  EXPECT_EQ(grad_op->Output(f::GradVarName("In1")),
-            std::vector<std::string>({f::GradVarName("in1")}));
-  EXPECT_EQ(grad_op->Output(f::GradVarName("In2_mult")),
-            std::vector<std::string>({f::GradVarName("in2_1"),
-                                      f::GradVarName("in2_2"),
-                                      f::GradVarName("in2_3")}));
-  EXPECT_EQ(grad_op->Output(f::GradVarName("In3")),
-            std::vector<std::string>({f::GradVarName("in3")}));
-  delete forw_op;
-  delete grad_op;
-}
-
-TEST(GradOpDescBuilder, IOIgnoredInGradient) {
-  f::OpDescBind *forw_op = new f::OpDescBind();
-  forw_op->SetType("io_ignored");
-  forw_op->SetInput("In1", {"in1"});
-  forw_op->SetInput("In2_mult", {"in2_1", "in2_2"});
-  forw_op->SetInput("In3_mult", {"in3_1", "in3_2"});
-  forw_op->SetOutput("Out1_mult", {"out1_1", "out1_2"});
-  forw_op->SetOutput("Out2", {"out2"});
-
-  f::OpDescBind *grad_op = new f::OpDescBind();
-  f::CompleteGradOpDesc(forw_op, grad_op);
-
-  EXPECT_EQ(grad_op->Type(), "io_ignored_grad");
-  // 'In2' and 'Out2' are ignored in gradient calculating
-  ASSERT_EQ(grad_op->InputNames().size(), 2UL + 1UL + 2UL);
-  EXPECT_EQ(grad_op->Input("In1"), std::vector<std::string>({"in1"}));
-  EXPECT_EQ(grad_op->Input("In3_mult"),
-            std::vector<std::string>({"in3_1", "in3_2"}));
-  EXPECT_EQ(grad_op->Input("Out1_mult"),
-            std::vector<std::string>({"out1_1", "out1_2"}));
-  EXPECT_EQ(grad_op->Input(f::GradVarName("Out1_mult")),
-            std::vector<std::string>(
-                {f::GradVarName("out1_1"), f::GradVarName("out1_2")}));
-  EXPECT_EQ(grad_op->Input(f::GradVarName("Out2")),
-            std::vector<std::string>({f::GradVarName("out2")}));
-
-  ASSERT_EQ(grad_op->OutputNames().size(), 3UL);
-  EXPECT_EQ(grad_op->Output(f::GradVarName("In1")),
-            std::vector<std::string>({f::GradVarName("in1")}));
-  EXPECT_EQ(grad_op->Output(f::GradVarName("In2_mult")),
-            std::vector<std::string>(
-                {f::GradVarName("in2_1"), f::GradVarName("in2_2")}));
-  EXPECT_EQ(grad_op->Output(f::GradVarName("In3_mult")),
-            std::vector<std::string>(
-                {f::GradVarName("in3_1"), f::GradVarName("in3_2")}));
-  delete forw_op;
-  delete grad_op;
-}
diff --git a/paddle/framework/op_desc.cc b/paddle/framework/op_desc.cc
index 852f0f1eb8ea09b336d656afea4f5b146c8d45de..02aa74a8420a5c685c88d7cb0b487284814b3690 100644
--- a/paddle/framework/op_desc.cc
+++ b/paddle/framework/op_desc.cc
@@ -18,6 +18,15 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+OpDescBind::OpDescBind(const std::string &type, const VariableNameMap &inputs,
+                       const VariableNameMap &outputs,
+                       const AttributeMap &attrs) {
+  op_desc_.set_type(type);
+  inputs_ = inputs;
+  outputs_ = outputs;
+  attrs_ = attrs;
+}
+
 OpDesc *OpDescBind::Proto() {
   Sync();
   return &op_desc_;
@@ -31,6 +40,14 @@ const std::vector<std::string> &OpDescBind::Input(
   return it->second;
 }
 
+std::vector<std::string> OpDescBind::InputArgumentNames() const {
+  std::vector<std::string> retv;
+  for (auto &ipt : this->inputs_) {
+    retv.insert(retv.end(), ipt.second.begin(), ipt.second.end());
+  }
+  return retv;
+}
+
 void OpDescBind::SetInput(const std::string &param_name,
                           const std::vector<std::string> &args) {
   need_update_ = true;
@@ -45,6 +62,14 @@ const std::vector<std::string> &OpDescBind::Output(
   return it->second;
 }
 
+std::vector<std::string> OpDescBind::OutputArgumentNames() const {
+  std::vector<std::string> retv;
+  for (auto &ipt : this->outputs_) {
+    retv.insert(retv.end(), ipt.second.begin(), ipt.second.end());
+  }
+  return retv;
+}
+
 void OpDescBind::SetOutput(const std::string &param_name,
                            const std::vector<std::string> &args) {
   need_update_ = true;
@@ -94,6 +119,18 @@ const std::unordered_map<std::string, Attribute> &OpDescBind::GetAttrMap()
   return attrs_;
 }
 
+void OpDescBind::Rename(const std::string &old_name,
+                        const std::string &new_name) {
+  for (auto &input : inputs_) {
+    std::replace(input.second.begin(), input.second.end(), old_name, new_name);
+  }
+  for (auto &output : outputs_) {
+    std::replace(output.second.begin(), output.second.end(), old_name,
+                 new_name);
+  }
+  need_update_ = true;
+}
+
 struct SetAttrDescVisitor : public boost::static_visitor<void> {
   explicit SetAttrDescVisitor(OpDesc::Attr *attr) : attr_(attr) {}
   mutable OpDesc::Attr *attr_;
diff --git a/paddle/framework/op_desc.h b/paddle/framework/op_desc.h
index 397393f796ed11312578859495529f4aedd17e78..b39808dad1de061e896936ec84169cd62e29856d 100644
--- a/paddle/framework/op_desc.h
+++ b/paddle/framework/op_desc.h
@@ -27,6 +27,11 @@ class BlockDescBind;
 
 class OpDescBind {
  public:
+  OpDescBind() {}
+
+  OpDescBind(const std::string &type, const VariableNameMap &inputs,
+             const VariableNameMap &outputs, const AttributeMap &attrs);
+
   OpDesc *Proto();
 
   std::string Type() const { return op_desc_.type(); }
@@ -35,11 +40,15 @@ class OpDescBind {
 
   const std::vector<std::string> &Input(const std::string &name) const;
 
+  std::vector<std::string> InputArgumentNames() const;
+
   void SetInput(const std::string &param_name,
                 const std::vector<std::string> &args);
 
   const std::vector<std::string> &Output(const std::string &name) const;
 
+  std::vector<std::string> OutputArgumentNames() const;
+
   void SetOutput(const std::string &param_name,
                  const std::vector<std::string> &args);
 
@@ -61,6 +70,8 @@ class OpDescBind {
 
   int GetBlockAttr(const std::string &name) const;
 
+  void Rename(const std::string &old_name, const std::string &new_name);
+
   // Only be used in C++
   const AttributeMap &GetAttrMap() const;
 
@@ -70,6 +81,22 @@ class OpDescBind {
   std::vector<std::string> InputNames() const { return MapKeys(inputs_); }
   std::vector<std::string> OutputNames() const { return MapKeys(outputs_); }
 
+  void SetInputMap(const VariableNameMap &input) {
+    this->inputs_ = input;
+    this->need_update_ = true;
+  }
+
+  void SetOutputMap(const VariableNameMap &output) {
+    this->outputs_ = output;
+    this->need_update_ = true;
+  }
+
+  void Sync();
+
+  const VariableNameMap &Inputs() const { return inputs_; }
+
+  const VariableNameMap &Outputs() const { return outputs_; }
+
  private:
   template <typename MapType>
   static std::vector<typename MapType::key_type> MapKeys(const MapType &map) {
@@ -81,8 +108,6 @@ class OpDescBind {
     return ret_val;
   }
 
-  void Sync();
-
   OpDesc op_desc_;
   VariableNameMap inputs_;
   VariableNameMap outputs_;
diff --git a/paddle/framework/op_info.h b/paddle/framework/op_info.h
index 8b7882485ff82860b884d3a8bc3d5543e2426898..c504f69e30bb899c183bd4281d2eadb50fd3b376 100644
--- a/paddle/framework/op_info.h
+++ b/paddle/framework/op_info.h
@@ -17,6 +17,7 @@
 #include <map>
 #include <string>
 #include <unordered_map>
+
 #include "paddle/framework/attribute.h"
 #include "paddle/framework/op_desc.h"
 #include "paddle/framework/type_defs.h"
@@ -27,7 +28,6 @@ namespace framework {
 
 struct OpInfo {
   OpCreator creator_;
-  std::string grad_op_type_;
   GradOpMakerFN grad_op_maker_;
   OpProto* proto_{nullptr};
   OpAttrChecker* checker_{nullptr};
@@ -43,19 +43,19 @@ struct OpInfo {
     return *proto_;
   }
 
-  const OpAttrChecker& Checker() const {
-    PADDLE_ENFORCE_NOT_NULL(checker_,
-                            "Operator Checker has not been registered");
-    return *checker_;
-  }
-
   const OpCreator& Creator() const {
     PADDLE_ENFORCE_NOT_NULL(creator_,
                             "Operator Creator has not been registered");
     return creator_;
   }
 
-  bool HasGradientOp() const { return !grad_op_type_.empty(); }
+  const GradOpMakerFN& GradOpMaker() const {
+    PADDLE_ENFORCE_NOT_NULL(grad_op_maker_,
+                            "Operator GradOpMaker has not been registered.");
+    return grad_op_maker_;
+  }
+
+  const OpAttrChecker* Checker() const { return checker_; }
 };
 
 class OpInfoMap {
diff --git a/paddle/framework/op_proto_maker.h b/paddle/framework/op_proto_maker.h
index 4d55a37db9f0a3deac7b3489c8bc288ea41f4799..a134befd90a1eaeff6f6ea62f11412df63cdc394 100644
--- a/paddle/framework/op_proto_maker.h
+++ b/paddle/framework/op_proto_maker.h
@@ -44,11 +44,6 @@ class OpProtoAndCheckerMaker {
       var_->set_intermediate(true);
       return *this;
     }
-
-    VariableBuilder& NotInGradient() {
-      var_->set_not_in_gradient(true);
-      return *this;
-    }
   };
 
   VariableBuilder AddInput(const std::string& name, const std::string& comment);
diff --git a/paddle/framework/op_registry.cc b/paddle/framework/op_registry.cc
index b0e85dd49f97da4a7f889fde0b5f060954947be8..66043f6e04fdb63b5d11a15c66abc84339e13c9a 100644
--- a/paddle/framework/op_registry.cc
+++ b/paddle/framework/op_registry.cc
@@ -23,7 +23,9 @@ std::unique_ptr<OperatorBase> OpRegistry::CreateOp(
     const std::string& type, const VariableNameMap& inputs,
     const VariableNameMap& outputs, AttributeMap attrs) {
   auto& info = OpInfoMap::Instance().Get(type);
-  info.Checker().Check(attrs);
+  if (info.Checker() != nullptr) {
+    info.Checker()->Check(attrs);
+  }
   auto op = info.Creator()(type, inputs, outputs, attrs);
   return std::unique_ptr<OperatorBase>(op);
 }
@@ -52,9 +54,15 @@ std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
   return CreateOp(op_desc.type(), inputs, outputs, attrs);
 }
 
-std::unique_ptr<OperatorBase> OpRegistry::CreateGradOp(const OperatorBase& op) {
-  PADDLE_ENFORCE(!op.IsNetOp(), "Use framework::Backward to get backward ops");
-  return std::unique_ptr<OperatorBase>(BuildGradOp(&op));
+std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDescBind& op_desc) {
+  return CreateOp(op_desc.Type(), op_desc.Inputs(), op_desc.Outputs(),
+                  op_desc.GetAttrMap());
+}
+
+std::vector<std::unique_ptr<OpDescBind>> OpRegistry::CreateGradOpDescs(
+    const OpDescBind& op_desc) {
+  auto& info = OpInfoMap::Instance().Get(op_desc.Type());
+  return info.grad_op_maker_(op_desc);
 }
 
 }  // namespace framework
diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h
index 958cf581f53b2bb0252b655267f657c8d9ab371c..cce3605fd480c1d79a5969f6a4cb170ea4d879f2 100644
--- a/paddle/framework/op_registry.h
+++ b/paddle/framework/op_registry.h
@@ -23,25 +23,37 @@ limitations under the License. */
 #include "paddle/framework/attribute.h"
 #include "paddle/framework/details/op_registry.h"
 #include "paddle/framework/framework.pb.h"
-#include "paddle/framework/grad_op_builder.h"
+#include "paddle/framework/grad_op_desc_maker.h"
+#include "paddle/framework/op_desc.h"
 #include "paddle/framework/operator.h"
 #include "paddle/framework/scope.h"
 
 namespace paddle {
 namespace framework {
+class Registrar {
+ public:
+  // In our design, various kinds of classes, e.g., operators and kernels,
+  // have their corresponding registry and registrar. The action of
+  // registration is in the constructor of a global registrar variable, which,
+  // however, are not used in the code that calls package framework, and would
+  // be removed from the generated binary file by the linker. To avoid such
+  // removal, we add Touch to all registrar classes and make USE_OP macros to
+  // call this method. So, as long as the callee code calls USE_OP, the global
+  // registrar variable won't be removed by the linker.
+  void Touch() {}
+};
 
 template <typename... ARGS>
-struct OperatorRegistrar {
+struct OperatorRegistrar : public Registrar {
   explicit OperatorRegistrar(const char* op_type) : op_type(op_type) {
     PADDLE_ENFORCE(!OpInfoMap::Instance().Has(op_type),
                    "'%s' is registered more than once.", op_type);
     static_assert(sizeof...(ARGS) != 0,
                   "OperatorRegistrar should be invoked at least by OpClass");
     details::OperatorRegistrarRecursive<0, false, ARGS...>(op_type, &info);
+    OpInfoMap::Instance().Insert(op_type, info);
   }
 
-  ~OperatorRegistrar() { OpInfoMap::Instance().Insert(op_type, info); }
-
   const char* op_type;
 
   OpInfo info;
@@ -67,20 +79,10 @@ class OpRegistry {
 
   static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
 
-  static std::unique_ptr<OperatorBase> CreateGradOp(const OperatorBase& op);
-};
+  static std::vector<std::unique_ptr<OpDescBind>> CreateGradOpDescs(
+      const OpDescBind& op_desc);
 
-class Registrar {
- public:
-  // In our design, various kinds of classes, e.g., operators and kernels,
-  // have their corresponding registry and registrar. The action of
-  // registration is in the constructor of a global registrar variable, which,
-  // however, are not used in the code that calls package framework, and would
-  // be removed from the generated binary file by the linker. To avoid such
-  // removal, we add Touch to all registrar classes and make USE_OP macros to
-  // call this method. So, as long as the callee code calls USE_OP, the global
-  // registrar variable won't be removed by the linker.
-  void Touch() {}
+  static std::unique_ptr<OperatorBase> CreateOp(const OpDescBind& op_desc);
 };
 
 template <typename OpType, typename ProtoMakerType, typename GradOpType>
@@ -138,33 +140,41 @@ class OpKernelRegistrar : public Registrar {
                              __test_global_namespace_##uniq_name##__>::value, \
                 msg)
 
+#define REGISTER_OPERATOR(op_type, op_class, ...)                      \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                      \
+      __reg_op__##op_type,                                             \
+      "REGISTER_OPERATOR must be called in global namespace");         \
+  class _OpClass_##op_type##_ : public op_class {                      \
+   public:                                                             \
+    DEFINE_OP_CLONE_METHOD(_OpClass_##op_type##_);                     \
+    DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_, op_class);            \
+  };                                                                   \
+  static ::paddle::framework::OperatorRegistrar<_OpClass_##op_type##_, \
+                                                ##__VA_ARGS__>         \
+      __op_registrar_##op_type##__(#op_type);                          \
+  int TouchOpRegistrar_##op_type() {                                   \
+    __op_registrar_##op_type##__.Touch();                              \
+    return 0;                                                          \
+  }
+
 /**
  * Macro to register Operator.
  */
-#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type,          \
-                    grad_op_class)                                            \
-  STATIC_ASSERT_GLOBAL_NAMESPACE(                                             \
-      __reg_op__##op_type, "REGISTER_OP must be called in global namespace"); \
-  class _OpClass_##op_type##_ : public op_class {                             \
-   public:                                                                    \
-    DEFINE_OP_CLONE_METHOD(_OpClass_##op_type##_);                            \
-    DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_, op_class);                   \
-  };                                                                          \
-  class _OpGradClass_##op_type##_ : public grad_op_class {                    \
-   public:                                                                    \
-    DEFINE_OP_CLONE_METHOD(_OpGradClass_##op_type##_);                        \
-    DEFINE_OP_CONSTRUCTOR(_OpGradClass_##op_type##_, grad_op_class);          \
-  };                                                                          \
-  static ::paddle::framework::OpRegistrar<                                    \
-      _OpClass_##op_type##_, op_maker_class, _OpGradClass_##op_type##_>       \
-      __op_registrar_##op_type##__(#op_type, #grad_op_type);                  \
-  int TouchOpRegistrar_##op_type() {                                          \
-    __op_registrar_##op_type##__.Touch();                                     \
-    return 0;                                                                 \
-  }
+#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type,           \
+                    grad_op_class)                                             \
+  REGISTER_OPERATOR(grad_op_type, grad_op_class);                              \
+  class _GradOpDescMaker_##grad_op_type##_                                     \
+      : public ::paddle::framework::DefaultGradOpDescMaker {                   \
+    using ::paddle::framework::DefaultGradOpDescMaker::DefaultGradOpDescMaker; \
+                                                                               \
+   protected:                                                                  \
+    virtual std::string GradOpType() const { return #grad_op_type; }           \
+  };                                                                           \
+  REGISTER_OPERATOR(op_type, op_class, _GradOpDescMaker_##grad_op_type##_,     \
+                    op_maker_class);
 
 #define REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) \
-  REGISTER_OP(op_type, op_class, op_maker_class, , ::paddle::framework::NOP)
+  REGISTER_OPERATOR(op_type, op_class, op_maker_class)
 
 /**
  * Macro to register OperatorKernel.
diff --git a/paddle/framework/tensor_array.h b/paddle/framework/tensor_array.h
index 22ae6a966f90c47fe8b4bbaaf5eb227c39d84173..94a14c2df492b175cf6a643800937878e95c5f37 100644
--- a/paddle/framework/tensor_array.h
+++ b/paddle/framework/tensor_array.h
@@ -26,6 +26,9 @@ namespace framework {
  * in original lod-tensor.
  */
 struct DySeqMeta {
+  DySeqMeta(size_t begin, size_t end, size_t ori_idx)
+      : begin(begin), end(end), ori_idx(ori_idx) {}
+
   size_t begin;
   size_t end;  // not included
   size_t ori_idx;
diff --git a/paddle/operators/activation_op.cc b/paddle/operators/activation_op.cc
index 5f2ecc2673a5bf87d07e99faf60d585fb02621d0..66e9d2c40138c26975f07cb544e54de6f00d6b09 100644
--- a/paddle/operators/activation_op.cc
+++ b/paddle/operators/activation_op.cc
@@ -97,6 +97,17 @@ class TanhOpMaker : public framework::OpProtoAndCheckerMaker {
   }
 };
 
+class TanhShrinkOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  TanhShrinkOpMaker(framework::OpProto *proto,
+                    framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "Input of TanhShrink operator");
+    AddOutput("Y", "Output of TanhShrink operator");
+    AddComment("TanhShrink activation operator, tanhshrink(x) = x - tanh(x)");
+  }
+};
+
 class SqrtOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   SqrtOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
@@ -235,6 +246,9 @@ REGISTER_OP(relu, ops::ActivationOp, ops::ReluOpMaker, relu_grad,
 REGISTER_OP(tanh, ops::ActivationOp, ops::TanhOpMaker, tanh_grad,
             ops::ActivationOpGrad);
 
+REGISTER_OP(tanh_shrink, ops::ActivationOp, ops::TanhShrinkOpMaker,
+            tanh_shrink_grad, ops::ActivationOpGrad);
+
 REGISTER_OP(sqrt, ops::ActivationOp, ops::SqrtOpMaker, sqrt_grad,
             ops::ActivationOpGrad);
 
diff --git a/paddle/operators/activation_op.h b/paddle/operators/activation_op.h
index dae66cc77d9103a6a2b13e69b3f014f8de313209..245060174224c5e24f75adf4ddc9a6db29101d74 100644
--- a/paddle/operators/activation_op.h
+++ b/paddle/operators/activation_op.h
@@ -146,6 +146,24 @@ struct TanhGradFunctor : public BaseActivationFunctor<T> {
   }
 };
 
+// tanhshrink(x) = x - tanh(x)
+// where tanh(x) = (exp(x) - exp(-x)) / (exp(x) + exp(-x))
+template <typename T>
+struct TanhShrinkFunctor : public BaseActivationFunctor<T> {
+  template <typename Device, typename X, typename Y>
+  void operator()(Device d, X x, Y y) const {
+    y.device(d) = x - x.tanh();
+  }
+};
+
+template <typename T>
+struct TanhShrinkGradFunctor : public BaseActivationFunctor<T> {
+  template <typename Device, typename X, typename Y, typename dY, typename dX>
+  void operator()(Device d, X x, Y y, dY dy, dX dx) const {
+    dx.device(d) = dy * (x.tanh() * x.tanh());
+  }
+};
+
 // sqrt(x) = x^(1/2)
 template <typename T>
 struct SqrtFunctor : public BaseActivationFunctor<T> {
@@ -407,4 +425,5 @@ struct STanhGradFunctor : public BaseActivationFunctor<T> {
   __macro(pow, PowFunctor, PowGradFunctor);                      \
   __macro(stanh, STanhFunctor, STanhGradFunctor);                \
   __macro(softsign, SoftsignFunctor, SoftsignGradFunctor);       \
-  __macro(leaky_relu, LeakyReluFunctor, LeakyReluGradFunctor)
+  __macro(leaky_relu, LeakyReluFunctor, LeakyReluGradFunctor);   \
+  __macro(tanh_shrink, TanhShrinkFunctor, TanhShrinkGradFunctor)
diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc
index d799239d4ed6d230578c77921a1a454b476b63fa..2332c9546b037c94a5a6d30319abda8e23c2b3bb 100644
--- a/paddle/operators/mean_op.cc
+++ b/paddle/operators/mean_op.cc
@@ -36,7 +36,7 @@ class MeanOpMaker : public framework::OpProtoAndCheckerMaker {
   MeanOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input of mean op");
-    AddOutput("Out", "The output of mean op").NotInGradient();
+    AddOutput("Out", "The output of mean op");
     AddComment(R"DOC( Mean Operator
 )DOC");
   }
@@ -52,11 +52,27 @@ class MeanGradOp : public framework::OperatorWithKernel {
   }
 };
 
+class MeanGradMaker : 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("mean_grad");
+    grad_op->SetInput("X", Input("X"));
+    grad_op->SetInput(framework::GradVarName("Out"), OutputGrad("Out"));
+    grad_op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
+    return std::unique_ptr<framework::OpDescBind>(grad_op);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(mean, ops::MeanOp, ops::MeanOpMaker, mean_grad, ops::MeanGradOp);
+REGISTER_OPERATOR(mean, ops::MeanOp, ops::MeanOpMaker, ops::MeanGradMaker);
+REGISTER_OPERATOR(mean_grad, ops::MeanGradOp);
 REGISTER_OP_CPU_KERNEL(mean,
                        ops::MeanKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(mean_grad,
diff --git a/paddle/operators/minus_op.cc b/paddle/operators/minus_op.cc
index ce049d4d7bd96a6758d71b381e6e6b4edbcc8b5c..7057dcbd6e375adef57d17a13afdfade67e938b6 100644
--- a/paddle/operators/minus_op.cc
+++ b/paddle/operators/minus_op.cc
@@ -49,9 +49,9 @@ class MinusOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   MinusOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "The left tensor of minus operator.").NotInGradient();
-    AddInput("Y", "The right tensor of minus operator.").NotInGradient();
-    AddOutput("Out", "The output tensor of minus operator.").NotInGradient();
+    AddInput("X", "The left tensor of minus operator.");
+    AddInput("Y", "The right tensor of minus operator.");
+    AddOutput("Out", "The output tensor of minus operator.");
 
     AddComment(R"DOC(Minus Operator
 
@@ -64,26 +64,35 @@ or not. But the output only shares the LoD with input `X`.
 )DOC");
   }
 };
-template <typename AttrType>
-class MinusGradOp : public NetOp {
+
+class MinusGradMaker : public framework::GradOpDescMakerBase {
  public:
-  MinusGradOp(const std::string &type, const framework::VariableNameMap &inputs,
-              const framework::VariableNameMap &outputs,
-              const framework::AttributeMap &attrs)
-      : NetOp(type, inputs, outputs, attrs) {
-    auto out_grad = Input(framework::GradVarName("Out"));
-    auto x_grad = Output(framework::GradVarName("X"));
-    auto y_grad = Output(framework::GradVarName("Y"));
-
-    // x_grad = out_grad
-    AppendOp(framework::OpRegistry::CreateOp("identity", {{"X", {out_grad}}},
-                                             {{"Y", {x_grad}}}, {}));
-
-    framework::AttributeMap scale_attr;
-    scale_attr["scale"] = static_cast<AttrType>(-1);
-    AppendOp(framework::OpRegistry::CreateOp("scale", {{"X", {out_grad}}},
-                                             {{"Out", {y_grad}}}, scale_attr));
-    CompleteAddOp(false);
+  using framework::GradOpDescMakerBase::GradOpDescMakerBase;
+
+  std::vector<std::unique_ptr<framework::OpDescBind>> operator()()
+      const override {
+    std::vector<std::unique_ptr<framework::OpDescBind>> ops;
+    auto x_g = InputGrad("X");
+    if (!x_g.empty()) {
+      auto *x_g_op = new framework::OpDescBind();
+      x_g_op->SetType("scale");
+      x_g_op->SetInput("X", OutputGrad("Out"));
+      x_g_op->SetOutput("Out", x_g);
+      x_g_op->SetAttr("scale", 1.0f);
+      ops.emplace_back(x_g_op);
+    }
+
+    auto y_g = InputGrad("Y");
+    if (!y_g.empty()) {
+      auto *y_g_op = new framework::OpDescBind();
+      y_g_op->SetType("scale");
+      y_g_op->SetInput("X", OutputGrad("Out"));
+      y_g_op->SetOutput("Out", y_g);
+      y_g_op->SetAttr("scale", -1.0f);
+      ops.emplace_back(y_g_op);
+    }
+
+    return ops;
   }
 };
 
@@ -91,7 +100,6 @@ class MinusGradOp : public NetOp {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(minus, ops::MinusOp, ops::MinusOpMaker, minus_grad,
-            ops::MinusGradOp<float>);
+REGISTER_OPERATOR(minus, ops::MinusOp, ops::MinusOpMaker, ops::MinusGradMaker);
 REGISTER_OP_CPU_KERNEL(minus,
                        ops::MinusKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/pad_op.cc b/paddle/operators/pad_op.cc
index 04ebb14f6ee6c73f48aa2f75811a22f9b8a25006..15aa05f26610be14e4c37be35137a259e00eb947 100644
--- a/paddle/operators/pad_op.cc
+++ b/paddle/operators/pad_op.cc
@@ -56,8 +56,7 @@ class PadOpMaker : public framework::OpProtoAndCheckerMaker {
              "The input should be a k-D tensor(k > 0 and k < 7)");
     AddOutput("Out",
               "The output of pad op."
-              "A tensor with the same shape as X.")
-        .NotInGradient();
+              "A tensor with the same shape as X.");
     AddComment(R"DOC(
 Pad input into output, as specified by paddings and pad_value. The input should be a k-D tensor(k > 0 and k < 7). As an example:
 
@@ -111,11 +110,29 @@ class PadOpGrad : public framework::OperatorWithKernel {
   }
 };
 
+class PadOpGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDescBind> Apply() const override {
+    auto* bind = new framework::OpDescBind();
+    bind->SetInput("X", Input("X"));
+    bind->SetInput(framework::GradVarName("Out"), OutputGrad("Out"));
+    bind->SetOutput(framework::GradVarName("X"), InputGrad("X"));
+    bind->SetAttrMap(Attrs());
+    bind->SetType("pad_grad");
+    return std::unique_ptr<framework::OpDescBind>(bind);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(pad, ops::PadOp, ops::PadOpMaker, pad_grad, ops::PadOpGrad);
+
+REGISTER_OPERATOR(pad, ops::PadOp, ops::PadOpMaker, ops::PadOpGradMaker);
+REGISTER_OPERATOR(pad_grad, ops::PadOpGrad);
 REGISTER_OP_CPU_KERNEL(pad, ops::PadKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(pad_grad,
                        ops::PadGradKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/scale_op.cc b/paddle/operators/scale_op.cc
index e92501e12834b92875f494de401672344f50e3b5..e225aecc270bc17c535c10253c970b888c42e5d3 100644
--- a/paddle/operators/scale_op.cc
+++ b/paddle/operators/scale_op.cc
@@ -41,8 +41,8 @@ class ScaleOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   ScaleOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "The input tensor of scale operator.").NotInGradient();
-    AddOutput("Out", "The output tensor of scale operator.").NotInGradient();
+    AddInput("X", "The input tensor of scale operator.");
+    AddOutput("Out", "The output tensor of scale operator.");
     AddComment(R"DOC(Scale operator
 
 The equation is: Out = scale*X
@@ -52,21 +52,18 @@ The equation is: Out = scale*X
   }
 };
 
-// The operator to calculate gradients of a scale operator is just the scale
-// operator itself.
-// Grad(Out=scale(X)) => Grad(X) = scale(Grad(Out))
-template <typename AttrType>
-class ScaleGradOp : public NetOp {
+class ScaleGradMaker : public framework::SingleGradOpDescMaker {
  public:
-  ScaleGradOp(const std::string &type, const framework::VariableNameMap &inputs,
-              const framework::VariableNameMap &outputs,
-              const framework::AttributeMap &attrs)
-      : NetOp(type, inputs, outputs, attrs) {
-    AppendOp(framework::OpRegistry::CreateOp(
-        "scale", {{"X", {Input(framework::GradVarName("Out"))}}},
-        {{"Out", {Output(framework::GradVarName("X"))}}},
-        {{"scale", Attr<AttrType>("scale")}}));
-    CompleteAddOp(false);
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDescBind> Apply() const override {
+    auto *grad_op = new framework::OpDescBind();
+    grad_op->SetType("scale");
+    grad_op->SetInput("X", OutputGrad("Out"));
+    grad_op->SetOutput("Out", InputGrad("X"));
+    grad_op->SetAttr("scale", GetAttr("scale"));
+    return std::unique_ptr<framework::OpDescBind>(grad_op);
   }
 };
 
@@ -75,7 +72,7 @@ class ScaleGradOp : public NetOp {
 
 namespace ops = paddle::operators;
 
-REGISTER_OP(scale, ops::ScaleOp, ops::ScaleOpMaker<float>, scale_grad,
-            ops::ScaleGradOp<float>);
+REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker<float>,
+                  ops::ScaleGradMaker);
 REGISTER_OP_CPU_KERNEL(scale,
                        ops::ScaleKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/softmax_with_cross_entropy_op.cc b/paddle/operators/softmax_with_cross_entropy_op.cc
index a76489871f30dc8d852b6a783efeff41704fd4a4..42c1ba6fdf1351c43ef78efaaf05c54acb54ce94 100644
--- a/paddle/operators/softmax_with_cross_entropy_op.cc
+++ b/paddle/operators/softmax_with_cross_entropy_op.cc
@@ -14,6 +14,7 @@
 
 #include "paddle/operators/softmax_with_cross_entropy_op.h"
 #include <paddle/function/TensorType.h>
+#include <iostream>
 
 namespace paddle {
 namespace operators {
@@ -27,15 +28,14 @@ class SoftmaxWithCrossEntropyOpMaker
     AddInput("Logits",
              "(Tensor, default: Tensor<float>), The unscaled log probabilities "
              "which is a 2-D tensor with shape [N x K]. N is the batch_size, "
-             "and K is the class number.")
-        .NotInGradient();
-    AddInput(
-        "Label",
-        "(Tensor, default: Tensor<int>), The ground truth which is a 2-D "
-        "tensor. "
-        "If softLable is set to 0, Label is a Tensor<int> with shape [N x 1]. "
-        "If softLable is set to 1, Label is a Tensor<float/double> "
-        "with shape [N x K].");
+             "and K is the class number.");
+    AddInput("Label",
+             "(Tensor, default: Tensor<int>), The ground truth which is a 2-D "
+             "tensor. "
+             "If softLable is set to 0, Label is a Tensor<int> with shape [N x "
+             "1]. "
+             "If softLable is set to 1, Label is a Tensor<float/double> "
+             "with shape [N x K].");
     AddOutput(
         "Softmax",
         "(Tensor, default: Tensor<float>), A 2-D tensor with shape [N x K]. "
@@ -163,15 +163,34 @@ class SoftmaxWithCrossEntropyOpGrad : public framework::OperatorWithKernel {
   }
 };
 
+class SoftmaxGradMaker : 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("softmax_with_cross_entropy_grad");
+    grad_op->SetInput("Label", Input("Label"));
+    grad_op->SetInput("Softmax", Output("Softmax"));
+    grad_op->SetInput("Loss", Output("Loss"));
+    grad_op->SetInput(framework::GradVarName("Softmax"), OutputGrad("Softmax"));
+    grad_op->SetInput(framework::GradVarName("Loss"), OutputGrad("Loss"));
+    grad_op->SetOutput(framework::GradVarName("Logits"), InputGrad("Logits"));
+    grad_op->SetAttrMap(Attrs());
+    return std::unique_ptr<framework::OpDescBind>(grad_op);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 
-REGISTER_OP(softmax_with_cross_entropy, ops::SoftmaxWithCrossEntropyOp,
-            ops::SoftmaxWithCrossEntropyOpMaker,
-            softmax_with_cross_entropy_grad,
-            ops::SoftmaxWithCrossEntropyOpGrad);
+REGISTER_OPERATOR(softmax_with_cross_entropy, ops::SoftmaxWithCrossEntropyOp,
+                  ops::SoftmaxWithCrossEntropyOpMaker, ops::SoftmaxGradMaker);
+REGISTER_OPERATOR(softmax_with_cross_entropy_grad,
+                  ops::SoftmaxWithCrossEntropyOpGrad);
 REGISTER_OP_CPU_KERNEL(softmax_with_cross_entropy,
                        ops::SoftmaxWithCrossEntropyKernel<float>);
 REGISTER_OP_CPU_KERNEL(softmax_with_cross_entropy_grad,
diff --git a/paddle/operators/sum_op.cc b/paddle/operators/sum_op.cc
index 7c422b477083fba4661aeb427422abb623b172bb..c701ee8dde26f0fd50fae227d3f345df2d7a119d 100644
--- a/paddle/operators/sum_op.cc
+++ b/paddle/operators/sum_op.cc
@@ -45,10 +45,8 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   SumOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "the input tensors of sum operator.")
-        .AsDuplicable()
-        .NotInGradient();
-    AddOutput("Out", "the output tensor of sum operator.").NotInGradient();
+    AddInput("X", "the input tensors of sum operator.").AsDuplicable();
+    AddOutput("Out", "the output tensor of sum operator.");
     AddComment(R"DOC(
 Sum the input tensors.
 
@@ -58,23 +56,26 @@ or not. But the output only shares the LoD with the first input.
   }
 };
 
-class SumGradOp : public NetOp {
+class SumGradMaker : public framework::GradOpDescMakerBase {
  public:
-  SumGradOp(const std::string& type, const framework::VariableNameMap& inputs,
-            const framework::VariableNameMap& outputs,
-            const framework::AttributeMap& attrs)
-      : NetOp(type, inputs, outputs, attrs) {
-    auto& x_grad_names = Outputs(framework::GradVarName("X"));
-    auto out_grad_name = this->Input(framework::GradVarName("Out"));
+  using framework::GradOpDescMakerBase::GradOpDescMakerBase;
 
-    framework::AttributeMap grad_attrs;
-    grad_attrs["scale"] = 1.0f;
-    for (auto& x_grad_name : x_grad_names) {
-      AppendOp(framework::OpRegistry::CreateOp(
-          "scale", {{"X", {out_grad_name}}}, {{"Out", {x_grad_name}}},
-          grad_attrs));
-    }
-    CompleteAddOp(false);
+  std::vector<std::unique_ptr<framework::OpDescBind>> operator()()
+      const override {
+    auto x_grads = InputGrad("X");
+    std::vector<std::unique_ptr<framework::OpDescBind>> grad_ops;
+    grad_ops.reserve(x_grads.size());
+    auto og = OutputGrad("Out");
+    std::transform(x_grads.begin(), x_grads.end(), std::back_inserter(grad_ops),
+                   [&og](const std::string& x_grad) {
+                     auto* grad_op = new framework::OpDescBind();
+                     grad_op->SetType("scale");
+                     grad_op->SetInput("X", og);
+                     grad_op->SetOutput("Out", {x_grad});
+                     grad_op->SetAttr("scale", 1.0f);
+                     return std::unique_ptr<framework::OpDescBind>(grad_op);
+                   });
+    return grad_ops;
   }
 };
 
@@ -82,5 +83,6 @@ class SumGradOp : public NetOp {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(sum, ops::SumOp, ops::SumOpMaker, sum_grad, ops::SumGradOp);
+
+REGISTER_OPERATOR(sum, ops::SumOp, ops::SumOpMaker, ops::SumGradMaker);
 REGISTER_OP_CPU_KERNEL(sum, ops::SumKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/pybind/CMakeLists.txt b/paddle/pybind/CMakeLists.txt
index 18ecbd1aa34c82d63ae7f8ec1bd8f81b35eee30b..97364f2db9523c0629616692631d8372657a2128 100644
--- a/paddle/pybind/CMakeLists.txt
+++ b/paddle/pybind/CMakeLists.txt
@@ -1,6 +1,6 @@
 if(WITH_PYTHON)
   cc_library(paddle_pybind SHARED
     SRCS pybind.cc exception.cc protobuf.cc
-    DEPS pybind python backward proto_desc
+    DEPS pybind python backward proto_desc tensor_array
     ${GLOB_OP_LIB})
 endif(WITH_PYTHON)
diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc
index 38ba450447386b44ee8abe71c3c8b6427bbc398c..356c4986e2e182e904215f7ebb8cac5146364f8b 100644
--- a/paddle/pybind/pybind.cc
+++ b/paddle/pybind/pybind.cc
@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include "paddle/framework/backward.h"
 #include "paddle/framework/lod_tensor.h"
+#include "paddle/framework/tensor_array.h"
 #include "paddle/operators/cond_op.h"
 #include "paddle/operators/net_op.h"
 #include "paddle/operators/recurrent_op.h"
@@ -286,6 +287,56 @@ All parameter, weight, gradient are variables in Paddle.
         self->CompleteAddOp();
       });
 
+  py::class_<framework::TensorArray>(m, "TensorArray")
+      .def("__init__",
+           [](TensorArray &instance) { new (&instance) TensorArray(); })
+      .def("read",
+           [](TensorArray &self, size_t index) { return self.Read(index); })
+      .def("write", [](TensorArray &self, size_t index,
+                       LoDTensor &value) { self.Write(index, value); })
+      .def("write_shared",
+           [](TensorArray &self, size_t index, const LoDTensor &value) {
+             self.WriteShared(index, value);
+           })
+      .def("size", [](TensorArray &self) { return self.size(); })
+      .def("pack",
+           [](TensorArray &self, size_t level,
+              const std::vector<std::vector<size_t>> &meta_info,
+              const std::vector<std::vector<size_t>> &lod) {
+             std::vector<DySeqMeta> meta;
+             for (auto &info : meta_info) {
+               PADDLE_ENFORCE_EQ(info.size(), 3UL);
+               meta.emplace_back(info[0], info[1], info[2]);
+             }
+#ifndef PADDLE_WITH_CUDA
+             return self.Pack(level, meta, lod);
+#else
+             LoD new_lod;
+             new_lod.reserve(lod.size());
+             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
+             return self.Pack(level, meta, new_lod);
+#endif
+           })
+      .def("unpack",
+           [](TensorArray &self, const LoDTensor &source, int level,
+              bool length_descend) {
+             auto metas = self.Unpack(source, level, length_descend);
+             std::vector<std::vector<size_t>> meta_info;
+             for (auto meta : metas) {
+               meta_info.emplace_back(
+                   std::vector<size_t>({meta.begin, meta.end, meta.ori_idx}));
+             }
+             return meta_info;
+           })
+      .def("stack", [](TensorArray &self) { return self.Stack(); })
+      .def("unstack",
+           [](TensorArray &self, const LoDTensor &source) {
+             return self.Unstack(source);
+           })
+      .def("unstack_shared", [](TensorArray &self, const LoDTensor &source) {
+        return self.UnstackShared(source);
+      });
+
   // recurrent_op
   py::class_<operators::RecurrentOp, OperatorBase>(m, "RecurrentOp")
       .def_static(
diff --git a/python/paddle/v2/framework/tests/test_activation_op.py b/python/paddle/v2/framework/tests/test_activation_op.py
index f232996a55da86399140ea2aad27428926c1a055..701e1a1aeec2746643fbd5432dadfd6bc46f358f 100644
--- a/python/paddle/v2/framework/tests/test_activation_op.py
+++ b/python/paddle/v2/framework/tests/test_activation_op.py
@@ -48,6 +48,21 @@ class TestTanh(OpTest):
         self.check_grad(['X'], 'Y', max_relative_error=0.007)
 
 
+class TestTanhShrink(OpTest):
+    def setUp(self):
+        self.op_type = "tanh_shrink"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [10, 17]).astype("float32")
+        }
+        self.outputs = {'Y': self.inputs['X'] - np.tanh(self.inputs['X'])}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Y', max_relative_error=0.008)
+
+
 class TestSqrt(OpTest):
     def setUp(self):
         self.op_type = "sqrt"
diff --git a/python/paddle/v2/framework/tests/test_tensor_array.py b/python/paddle/v2/framework/tests/test_tensor_array.py
new file mode 100644
index 0000000000000000000000000000000000000000..11f8a01f9224fcbd6dd6cbc8c37cc81036ad3e07
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_tensor_array.py
@@ -0,0 +1,106 @@
+import logging
+import paddle.v2.framework.core as core
+import unittest
+import numpy as np
+
+
+class TestTensorArray(unittest.TestCase):
+    def setUp(self):
+        self.ta = core.TensorArray()
+
+        self.batch_size = 10
+        self.dim = 2
+
+        # create a LoDTensor
+        self.scope = core.Scope()
+        var = self.scope.new_var("test_tensor")
+        self.place = core.CPUPlace()
+        tensor = var.get_tensor()
+        tensor.set_dims([self.batch_size, self.dim])
+        tensor.alloc_float(self.place)
+        tensor_array = np.array(tensor)
+        tensor_array[0, 0] = 0
+        tensor_array[1, 0] = 1
+        tensor_array[2, 0] = 2
+        tensor_array[3, 0] = 3
+        tensor_array[4, 0] = 4
+        tensor_array[5, 0] = 5
+        tensor_array[6, 0] = 6
+        tensor_array[7, 0] = 7
+        tensor_array[8, 0] = 8
+        tensor_array[9, 0] = 9
+
+        lod_py = [[0, 2, 5, 10]]
+        lod_tensor = core.LoDTensor(lod_py)
+        lod_tensor.set(tensor_array, self.place)
+
+        self.py_seq_meta = [[5, 10, 2], [2, 5, 1], [0, 2, 0]]
+
+        self.tensor = lod_tensor
+
+    def test_unstack(self):
+        self.ta.unstack(self.tensor)
+        self.assertEqual(self.tensor.get_dims()[0], self.ta.size())
+
+    def test_read(self):
+        self.ta.unstack(self.tensor)
+        for i in range(self.batch_size):
+            tensor = self.ta.read(i)
+
+    def test_write(self):
+        self.ta.unstack(self.tensor)
+
+        # create a tensor with shape of [1, self.dim]
+        var = self.scope.new_var("hell")
+        tensor = var.get_tensor()
+        tensor.set_dims([1, self.dim])
+        tensor.alloc_float(self.place)
+        tensor_array = np.array(tensor)
+        for i in range(self.dim):
+            tensor_array[0, i] = i
+        tensor.set(tensor_array, self.place)
+
+        self.ta.write(2, tensor)
+
+        ta_tensor = self.ta.read(2)
+        ta_tensor_array = np.array(ta_tensor)
+        self.assertEqual(ta_tensor.get_dims(), [1, self.dim])
+        self.assertTrue((tensor_array == ta_tensor_array).all())
+
+    def test_write_shared(self):
+        self.ta.unstack(self.tensor)
+
+        # create a tensor with shape of [1, self.dim]
+        var = self.scope.new_var("hell")
+        tensor = var.get_tensor()
+        tensor.set_dims([1, self.dim])
+        tensor.alloc_float(self.place)
+        tensor_array = np.array(tensor)
+        for i in range(self.dim):
+            tensor_array[0, i] = i
+        tensor.set(tensor_array, self.place)
+
+        self.ta.write_shared(2, tensor)
+
+        ta_tensor = self.ta.read(2)
+        ta_tensor_array = np.array(ta_tensor)
+        self.assertEqual(ta_tensor.get_dims(), [1, self.dim])
+        self.assertTrue((tensor_array == ta_tensor_array).all())
+
+    def test_unpack(self):
+        meta = self.ta.unpack(self.tensor, 0, True)
+        self.assertEqual(self.ta.size(), 5)
+        self.assertEqual(meta, self.py_seq_meta)
+
+    def test_pack(self):
+        meta = self.ta.unpack(self.tensor, 0, True)
+        print "meta", meta
+        tensor = self.ta.pack(0, meta, self.tensor.lod())
+        print np.array(self.tensor)
+        print np.array(tensor)
+        self.assertTrue((np.array(self.tensor) == np.array(tensor)).all())
+        self.assertTrue(tensor.lod(), self.tensor.lod())
+
+
+if __name__ == '__main__':
+    unittest.main()