Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into warpctc

45cf2341 · wanghaoshuang · 137f0dfc · df9c13a8 · 45cf2341 · 45cf2341
100 changed file
--- a/adversarial/README.md
+++ b/adversarial/README.md
+# Advbox
+
+Advbox is a Python toolbox to create adversarial examples that fool neural networks. It requires Python and paddle.
+
+## How to use
+
+1. train a model and save it's parameters. (like fluid_mnist.py)
+2. load the parameters which is trained in step1, then reconstruct the model.(like mnist_tutorial_fgsm.py)
+3. use advbox to generate the adversarial sample.
--- a/adversarial/advbox/__init__.py
+++ b/adversarial/advbox/__init__.py
+# Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved
+#
+# 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.
+"""
+   A set of tools for generating adversarial example on paddle platform 
+"""
--- a/adversarial/advbox/attacks/base.py
+++ b/adversarial/advbox/attacks/base.py
+"""
+The base model of the model.
+"""
+from abc import ABCMeta, abstractmethod
+
+
+class Attack(object):
+    """
+    Abstract base class for adversarial attacks. `Attack` represent an adversarial attack
+    which search an adversarial example. subclass should implement the _apply() method.
+
+    Args:
+        model(Model): an instance of the class advbox.base.Model.
+
+    """
+    __metaclass__ = ABCMeta
+
+    def __init__(self, model):
+        self.model = model
+
+    def __call__(self, image_label):
+        """
+        Generate the adversarial sample.
+
+        Args:
+        image_label(list): The image and label tuple list with one element.
+        """
+        adv_img = self._apply(image_label)
+        return adv_img
+
+    @abstractmethod
+    def _apply(self, image_label):
+        """
+        Search an adversarial example.
+
+        Args:
+        image_batch(list): The image and label tuple list with one element.
+        """
+        raise NotImplementedError
--- a/adversarial/advbox/attacks/gradientsign.py
+++ b/adversarial/advbox/attacks/gradientsign.py
+"""
+This module provide the attack method for FGSM's implement.
+"""
+from __future__ import division
+import numpy as np
+from collections import Iterable
+from .base import Attack
+
+
+class GradientSignAttack(Attack):
+    """
+    This attack was originally implemented by Goodfellow et al. (2015) with the
+    infinity norm (and is known as the "Fast Gradient Sign Method"). This is therefore called
+    the Fast Gradient Method.
+    Paper link: https://arxiv.org/abs/1412.6572
+    """
+
+    def _apply(self, image_label, epsilons=1000):
+        assert len(image_label) == 1
+        pre_label = np.argmax(self.model.predict(image_label))
+
+        min_, max_ = self.model.bounds()
+        gradient = self.model.gradient(image_label)
+        gradient_sign = np.sign(gradient) * (max_ - min_)
+
+        if not isinstance(epsilons, Iterable):
+            epsilons = np.linspace(0, 1, num=epsilons + 1)
+
+        for epsilon in epsilons:
+            adv_img = image_label[0][0].reshape(
+                gradient_sign.shape) + epsilon * gradient_sign
+            adv_img = np.clip(adv_img, min_, max_)
+            adv_label = np.argmax(self.model.predict([(adv_img, 0)]))
+            if pre_label != adv_label:
+                return adv_img
+
+
+FGSM = GradientSignAttack
--- a/adversarial/advbox/models/__init__.py
+++ b/adversarial/advbox/models/__init__.py
+# Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved
+#
+# 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.
+"""
+Paddle model for target of attack 
+"""
--- a/adversarial/advbox/models/base.py
+++ b/adversarial/advbox/models/base.py
+"""
+The base model of the model.
+"""
+from abc import ABCMeta
+import abc
+
+abstractmethod = abc.abstractmethod
+
+
+class Model(object):
+    """
+    Base class of model to provide attack.
+
+
+    Args:
+        bounds(tuple): The lower and upper bound for the image pixel.
+        channel_axis(int): The index of the axis that represents the color channel.
+        preprocess(tuple): Two element tuple used to preprocess the input. First
+            substract the first element, then divide the second element.
+    """
+    __metaclass__ = ABCMeta
+
+    def __init__(self, bounds, channel_axis, preprocess=None):
+        assert len(bounds) == 2
+        assert channel_axis in [0, 1, 2, 3]
+
+        if preprocess is None:
+            preprocess = (0, 1)
+        self._bounds = bounds
+        self._channel_axis = channel_axis
+        self._preprocess = preprocess
+
+    def bounds(self):
+        """
+        Return the upper and lower bounds of the model.
+        """
+        return self._bounds
+
+    def channel_axis(self):
+        """
+        Return the channel axis of the model.
+        """
+        return self._channel_axis
+
+    def _process_input(self, input_):
+        res = input_
+        sub, div = self._preprocess
+        if sub != 0:
+            res = input_ - sub
+        assert div != 0
+        if div != 1:
+            res /= div
+        return res
+
+    @abstractmethod
+    def predict(self, image_batch):
+        """
+        Calculate the prediction of the image batch.
+
+        Args:
+            image_batch(numpy.ndarray): image batch of shape (batch_size, height, width, channels).
+
+        Return:
+            numpy.ndarray: predictions of the images with shape (batch_size, num_of_classes).
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def num_classes(self):
+        """
+        Determine the number of the classes
+
+        Return:
+            int: the number of the classes
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def gradient(self, image_batch):
+        """
+        Calculate the gradient of the cross-entropy loss w.r.t the image.
+
+        Args:
+            image_batch(list): The image and label tuple list.
+
+        Return:
+            numpy.ndarray: gradient of the cross-entropy loss w.r.t the image with
+                the shape (height, width, channel).
+        """
+        raise NotImplementedError
--- a/adversarial/advbox/models/paddle.py
+++ b/adversarial/advbox/models/paddle.py
+from __future__ import absolute_import
+
+import numpy as np
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+from paddle.v2.fluid.framework import program_guard
+
+from .base import Model
+
+
+class PaddleModel(Model):
+    """
+    Create a PaddleModel instance.
+    When you need to generate a adversarial sample, you should construct an instance of PaddleModel.
+
+    Args:
+        program(paddle.v2.fluid.framework.Program): The program of the model which generate the adversarial sample.
+        input_name(string): The name of the input.
+        logits_name(string): The name of the logits.
+        predict_name(string): The name of the predict.
+        cost_name(string): The name of the loss in the program.
+    """
+
+    def __init__(self,
+                 program,
+                 input_name,
+                 logits_name,
+                 predict_name,
+                 cost_name,
+                 bounds,
+                 channel_axis=3,
+                 preprocess=None):
+        super(PaddleModel, self).__init__(
+            bounds=bounds, channel_axis=channel_axis, preprocess=preprocess)
+
+        if preprocess is None:
+            preprocess = (0, 1)
+
+        self._program = program
+        self._place = fluid.CPUPlace()
+        self._exe = fluid.Executor(self._place)
+
+        self._input_name = input_name
+        self._logits_name = logits_name
+        self._predict_name = predict_name
+        self._cost_name = cost_name
+
+        # gradient
+        loss = self._program.block(0).var(self._cost_name)
+        param_grads = fluid.backward.append_backward(
+            loss, parameter_list=[self._input_name])
+        self._gradient = dict(param_grads)[self._input_name]
+
+    def predict(self, image_batch):
+        """
+            Predict the label of the image_batch.
+
+            Args:
+                image_batch(list): The image and label tuple list.
+            Return:
+                numpy.ndarray: predictions of the images with shape (batch_size, num_of_classes).
+        """
+        feeder = fluid.DataFeeder(
+            feed_list=[self._input_name, self._logits_name],
+            place=self._place,
+            program=self._program)
+        predict_var = self._program.block(0).var(self._predict_name)
+        predict = self._exe.run(self._program,
+                                feed=feeder.feed(image_batch),
+                                fetch_list=[predict_var])
+        return predict
+
+    def num_classes(self):
+        """
+            Calculate the number of classes of the output label. 
+
+        Return:
+            int: the number of classes
+        """
+        predict_var = self._program.block(0).var(self._predict_name)
+        assert len(predict_var.shape) == 2
+        return predict_var.shape[1]
+
+    def gradient(self, image_batch):
+        """
+        Calculate the gradient of the loss w.r.t the input.
+
+        Args:
+            image_batch(list): The image and label tuple list.
+        Return:
+            list: The list of the gradient of the image.
+        """
+        feeder = fluid.DataFeeder(
+            feed_list=[self._input_name, self._logits_name],
+            place=self._place,
+            program=self._program)
+
+        grad, = self._exe.run(self._program,
+                              feed=feeder.feed(image_batch),
+                              fetch_list=[self._gradient])
+        return grad
--- a/adversarial/fluid_mnist.py
+++ b/adversarial/fluid_mnist.py
+"""
+CNN on mnist data using fluid api of paddlepaddle
+"""
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+
+
+def mnist_cnn_model(img):
+    """
+    Mnist cnn model
+
+    Args:
+        img(Varaible): the input image to be recognized
+
+    Returns:
+        Variable: the label prediction
+    """
+    conv_pool_1 = fluid.nets.simple_img_conv_pool(
+        input=img,
+        num_filters=20,
+        filter_size=5,
+        pool_size=2,
+        pool_stride=2,
+        act='relu')
+
+    conv_pool_2 = fluid.nets.simple_img_conv_pool(
+        input=conv_pool_1,
+        num_filters=50,
+        filter_size=5,
+        pool_size=2,
+        pool_stride=2,
+        act='relu')
+
+    logits = fluid.layers.fc(input=conv_pool_2, size=10, act='softmax')
+    return logits
+
+
+def main():
+    """
+    Train the cnn model on mnist datasets
+    """
+    img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32')
+    label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+    logits = mnist_cnn_model(img)
+    cost = fluid.layers.cross_entropy(input=logits, label=label)
+    avg_cost = fluid.layers.mean(x=cost)
+    optimizer = fluid.optimizer.Adam(learning_rate=0.01)
+    optimizer.minimize(avg_cost)
+
+    accuracy = fluid.evaluator.Accuracy(input=logits, label=label)
+
+    BATCH_SIZE = 50
+    PASS_NUM = 3
+    ACC_THRESHOLD = 0.98
+    LOSS_THRESHOLD = 10.0
+    train_reader = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.mnist.train(), buf_size=500),
+        batch_size=BATCH_SIZE)
+
+    place = fluid.CPUPlace()
+    exe = fluid.Executor(place)
+    feeder = fluid.DataFeeder(feed_list=[img, label], place=place)
+    exe.run(fluid.default_startup_program())
+
+    for pass_id in range(PASS_NUM):
+        accuracy.reset(exe)
+        for data in train_reader():
+            loss, acc = exe.run(fluid.default_main_program(),
+                                feed=feeder.feed(data),
+                                fetch_list=[avg_cost] + accuracy.metrics)
+            pass_acc = accuracy.eval(exe)
+            print("pass_id=" + str(pass_id) + " acc=" + str(acc) + " pass_acc="
+                  + str(pass_acc))
+            if loss < LOSS_THRESHOLD and pass_acc > ACC_THRESHOLD:
+                break
+
+        pass_acc = accuracy.eval(exe)
+        print("pass_id=" + str(pass_id) + " pass_acc=" + str(pass_acc))
+    fluid.io.save_params(
+        exe, dirname='./mnist', main_program=fluid.default_main_program())
+    print('train mnist done')
+
+
+if __name__ == '__main__':
+    main()
--- a/adversarial/mnist_tutorial_fgsm.py
+++ b/adversarial/mnist_tutorial_fgsm.py
+"""
+FGSM demos on mnist using advbox tool.
+"""
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+import matplotlib.pyplot as plt
+import numpy as np
+
+from advbox.models.paddle import PaddleModel
+from advbox.attacks.gradientsign import GradientSignAttack
+
+
+def cnn_model(img):
+    """
+    Mnist cnn model
+    Args:
+        img(Varaible): the input image to be recognized
+    Returns:
+        Variable: the label prediction
+    """
+    #conv1 = fluid.nets.conv2d()
+    conv_pool_1 = fluid.nets.simple_img_conv_pool(
+        input=img,
+        num_filters=20,
+        filter_size=5,
+        pool_size=2,
+        pool_stride=2,
+        act='relu')
+
+    conv_pool_2 = fluid.nets.simple_img_conv_pool(
+        input=conv_pool_1,
+        num_filters=50,
+        filter_size=5,
+        pool_size=2,
+        pool_stride=2,
+        act='relu')
+
+    logits = fluid.layers.fc(input=conv_pool_2, size=10, act='softmax')
+    return logits
+
+
+def main():
+    """
+    Advbox demo which demonstrate how to use advbox.
+    """
+    IMG_NAME = 'img'
+    LABEL_NAME = 'label'
+
+    img = fluid.layers.data(name=IMG_NAME, shape=[1, 28, 28], dtype='float32')
+    # gradient should flow
+    img.stop_gradient = False
+    label = fluid.layers.data(name=LABEL_NAME, shape=[1], dtype='int64')
+    logits = cnn_model(img)
+    cost = fluid.layers.cross_entropy(input=logits, label=label)
+    avg_cost = fluid.layers.mean(x=cost)
+
+    place = fluid.CPUPlace()
+    exe = fluid.Executor(place)
+
+    BATCH_SIZE = 1
+    train_reader = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.mnist.train(), buf_size=500),
+        batch_size=BATCH_SIZE)
+    feeder = fluid.DataFeeder(
+        feed_list=[IMG_NAME, LABEL_NAME],
+        place=place,
+        program=fluid.default_main_program())
+
+    fluid.io.load_params(
+        exe, "./mnist/", main_program=fluid.default_main_program())
+
+    # advbox demo
+    m = PaddleModel(fluid.default_main_program(), IMG_NAME, LABEL_NAME,
+                    logits.name, avg_cost.name, (-1, 1))
+    att = GradientSignAttack(m)
+    for data in train_reader():
+        # fgsm attack
+        adv_img = att(data)
+        plt.imshow(n[0][0], cmap='Greys_r')
+        plt.show()
+        #np.save('adv_img', adv_img)
+        break
+
+
+if __name__ == '__main__':
+    main()
--- a/benchmark/cluster/README.md
+++ b/benchmark/cluster/README.md
+# Cluster Training Benchmark
+
+## Setup
+
+- Platform
+  - Kubernetes: v1.6.2
+  - Linux Kernel: v3.10.0
+
+- Resource
+  - CPU: 10 Cores per Pod
+  - Memory: 5GB per Pod
+
+- Docker Image
+
+  We use different base Docker Image to run the benchmark on Kubernetes:
+  - PaddlePaddle v2: paddlepaddle/paddle:0.11.0
+  - PaddlePaddle Fluid: paddlepaddle/paddle:[commit-id]
+  - TensorFlow: tensorflow/tensorflow:1.5.0-rc0
+
+- Model
+  vgg16 is used in this benchmark.
+
+## Cases
+
+- Variable
+  - Batch Size of training data.
+  - PServer count of the training job.
+  - The number of trainers.
+
+- Invariant
+  - The resource of trainer/pserver Pod.
+
+### Measure the Performance for Different Batch Size
+
+- PServer Count: 40
+- Trainer Count: 100
+- Metrics: mini-batch / sec
+
+| Batch Size | 32 | 64 | 128 | 256 |
+| -- | -- | -- | -- | -- |
+| PaddlePaddle Fluid | - | - | - | - |
+| PaddlePaddle v2 | - | - | - | - |
+| TensorFlow | - | - | - | - |
+
+### Measure the Performance for Different PServer Count
+
+- Trainer Count: 100
+- Batch Size: 64
+- Metrics: mini-batch / sec
+
+| PServer Count | 10 | 20 | 40 | 60 |
+| -- | -- | -- | -- | -- |
+| PaddlePaddle Fluid | - | - | - | - |
+| PaddlePaddle v2 | - | - | - | - |
+| TensorFlow | - | - | - | - |
+
+### Measure Parallel Efficiency By Increasing Trainer Count
+
+- PServer Count: 20
+- Batch Size: 64
+- Metrics:
+
+$S = \div(T1, TN)$
+
+which S is the ratio of T1 over TN, training time of 1 and N trainers.
+The parallel efficiency is:
+
+$E = \div(S, N)$
+
+| Trainer Counter | 1 | 10 | 20 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 |
+| -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- |
+| PaddlePaddle Fluid | - | - | - | - | - | - | - | - | - | - | - |
+| PaddlePaddle v2 | - | - | - | - | - | - | - | - | - | - | - | - |
+| TensorFlow | - | - | - | - | - | - | - | - | - | - | - | - | - |
+
+## Reproduce the benchmark
+
+TODO
--- a/doc/api/v2/fluid.rst
+++ b/doc/api/v2/fluid.rst
@@ -15,4 +15,4 @@ Fluid
    fluid/param_attr.rst
    fluid/profiler.rst
    fluid/regularizer.rst
-
+    fluid/io.rst
--- a/doc/api/v2/fluid/io.rst
+++ b/doc/api/v2/fluid/io.rst
+===========
+IO
+===========
+
+
+
+is_parameter
+-----------
+..  autofunction:: paddle.v2.fluid.io.is_parameter
+    :noindex:
--- a/doc/api/v2/fluid/layers.rst
+++ b/doc/api/v2/fluid/layers.rst
@@ -38,6 +38,16 @@ elementwise_add
 ..  autofunction:: paddle.v2.fluid.layers.elementwise_add
    :noindex:

+elementwise_sub
+---------------
+..  autofunction:: paddle.v2.fluid.layers.elementwise_sub
+    :noindex:
+
+elementwise_mul
+---------------
+..  autofunction:: paddle.v2.fluid.layers.elementwise_mul
+    :noindex:
+
 elementwise_div
 ---------------
 ..  autofunction:: paddle.v2.fluid.layers.elementwise_div
@@ -348,3 +358,132 @@ reduce_min
 ..  autofunction:: paddle.v2.fluid.layers.reduce_min
    :noindex:

+logsigmoid
+----------
+..  autofunction:: paddle.v2.fluid.layers.logsigmoid
+    :noindex:
+
+exp
+---
+..  autofunction:: paddle.v2.fluid.layers.exp
+    :noindex:
+
+relu
+----
+..  autofunction:: paddle.v2.fluid.layers.relu
+    :noindex:
+
+tanh
+----
+..  autofunction:: paddle.v2.fluid.layers.tanh
+    :noindex:
+
+tanh_shrink
+-----------
+..  autofunction:: paddle.v2.fluid.layers.tanh_shrink
+    :noindex:
+
+softshrink
+----------
+..  autofunction:: paddle.v2.fluid.layers.softshrink
+    :noindex:
+
+sqrt
+----
+..  autofunction:: paddle.v2.fluid.layers.sqrt
+    :noindex:
+
+abs
+----
+..  autofunction:: paddle.v2.fluid.layers.abs
+    :noindex:
+
+ceil
+----
+..  autofunction:: paddle.v2.fluid.layers.ceil
+    :noindex:
+
+floor
+-----
+..  autofunction:: paddle.v2.fluid.layers.floor
+    :noindex:
+
+round
+-----
+..  autofunction:: paddle.v2.fluid.layers.round
+    :noindex:
+
+reciprocal
+----------
+..  autofunction:: paddle.v2.fluid.layers.reciprocal
+    :noindex:
+
+log
+---
+..  autofunction:: paddle.v2.fluid.layers.log
+    :noindex:
+
+square
+------
+..  autofunction:: paddle.v2.fluid.layers.square
+    :noindex:
+
+softplus
+--------
+..  autofunction:: paddle.v2.fluid.layers.softplus
+    :noindex:
+
+softsign
+---------
+..  autofunction:: paddle.v2.fluid.layers.softsign
+    :noindex:
+
+brelu
+-----
+..  autofunction:: paddle.v2.fluid.layers.brelu
+    :noindex:
+
+leaky_relu
+----------
+..  autofunction:: paddle.v2.fluid.layers.leaky_relu
+    :noindex:
+
+soft_relu
+---------
+..  autofunction:: paddle.v2.fluid.layers.soft_relu
+    :noindex:
+
+elu
+----
+..  autofunction:: paddle.v2.fluid.layers.elu
+    :noindex:
+
+relu6
+-----
+..  autofunction:: paddle.v2.fluid.layers.relu6
+    :noindex:
+
+pow
+----
+..  autofunction:: paddle.v2.fluid.layers.pow
+    :noindex:
+
+hard_shrink
+-----------
+..  autofunction:: paddle.v2.fluid.layers.hard_shrink
+    :noindex:
+
+thresholded_relu
+----------------
+..  autofunction:: paddle.v2.fluid.layers.thresholded_relu
+    :noindex:
+
+hard_sigmoid
+-------------
+..  autofunction:: paddle.v2.fluid.layers.hard_sigmoid
+    :noindex:
+
+swish
+------
+..  autofunction:: paddle.v2.fluid.layers.swish
+    :noindex:
--- a/doc/design/block.md
+++ b/doc/design/block.md
@@ -202,8 +202,8 @@ This `OpDesc` value is in the `ops` field of the `BlockDesc` value representing

 During the generation of the Protobuf message, the Block should store VarDesc (the Protobuf message which describes Variable) and OpDesc (the Protobuf message which describes Operator).

-VarDesc in a block should have its name scope to avoid local variables affect parent block's name scope.
-Child block's name scopes should inherit the parent's so that OpDesc in child block can reference a VarDesc that stored in parent block. For example:
+VarDesc in a block should have its name scope to avoid local variables affecting parent block's name scope.
+Child block's name scopes should inherit the parent's so that OpDesc in child block can reference a VarDesc that is stored in the parent block. For example:

 ```python
 a = pd.Variable(shape=[20, 20])

--- a/doc/design/memory_optimization.md
+++ b/doc/design/memory_optimization.md
@@ -5,28 +5,28 @@

 In a lecture from Andrew Ng, he attributes the recent sucess of AI due to a combination of these:

- availability of Big Data
- supercomputing power to process this Big Data over very large neural networks
- modern algorithms
+- Availability of Big Data
+- Supercomputing power to process this Big Data over very large neural networks
+- Modern algorithms

 Following graph shows the details:

 ![](images/deep_learning.png)

-Larger model usually brings better performance. However, GPU memory is certain limited. For example, the memory size of a GTX TITAN X is only 12GB. To train complex and large model, we have to take care of memory using. Besides, memory optimization is also necessary in both online/mobile inference. 
+Larger model usually bring better performance. However, GPU memory is limited. For example, the memory size of a GTX TITAN X is only 12GB. To train complex and large models, we have to take care of memory usage. Besides, memory optimization is also necessary in both online/mobile inference. 

 ## Solution

 ### Basic Strategy

-There are some basic strategies to make memory optimization, including in-place operation and memory sharing.
+There are some basic strategies to improve memory usage, including in-place operations and memory sharing.

 #### In-place Operation
 In a relu activation operator： 

 $y = \max(x, 0)$

-If the variable x is not used in any other operator, we can make an in-place operation. In other words, the memory block of variable y and variable x are the same. In-place operation will save 50% memory occupancy immediately.
+If the variable x is not used in any other operator, we can make an in-place operation. In other words, the memory block of variable y and variable x will be the same. In-place operations will save 50% memory occupancy immediately.

 #### Memory Sharing

@@ -40,18 +40,18 @@ d = op2(a)
 e = op3(d, f)
 ```

-In this case, variable a is no longer used, and op2 does not support in-place operation. After op2 finished, we can put the memory of variable a to a memory pool. Then, variable e can share the memory of variable a from the pool.
+In this case, variable a is no longer used, and op2 does not support in-place operation. After op2 finishes, we can put the memory of variable a to a memory pool. Then, variable e can share the memory of variable a from the pool.


 ### Live Variable Analysis

-It's not enough to only have some basic strategies. The prerequisite of memory optimization is to know if a variable is still "live" after an operation.
+It's not enough to only have some basic strategies. The pre-requisite of memory optimization is to know if a variable is still "live" after an operation.

 In our design, the neural network topology is defined as a program. Luckily, [live variable analysis](https://en.wikipedia.org/wiki/Live_variable_analysis) is a classic problem in compilers which can be used in many stages, such as register allocation. 

-In compilers, the front end of the compilers translates programs into an intermediate language with an unbounded number of temporaries. This program must run on a machine with a bounded number of registers. Two temporaries a and b can fit into the same register, if a and b are never "in use" at the same time. Thus, many temporaries can fit in few registers; if they don't all fit, the excess temporaries can be kept in memory.
+In compilers, the front end of the compiler translates programs into an intermediate language with an unbounded number of temporary variables. This program must run on a machine with a bounded number of registers. Two temporary variables a and b can fit into the same register, if a and b are never "in use" at the same time. Thus, many temporary variables can fit in few registers; if they don't all fit, the excess tempory variables can be kept in memory.

-Therefore, the compiler needs to analyze the intermediate-representation program to determine which temporaries are in use at the same time. We say a variable is "live" if it holds a value that may be needed in the future, so this analysis is called liveness analysis. 
+Therefore, the compiler needs to analyze the intermediate-representation program to determine which temporary variables are in use at the same time. We say a variable is "live" if it holds a value that may be needed in the future, so this analysis is called liveness analysis. 

 We can leran these techniques from compilers. There are mainly two stages to make live variable analysis:

@@ -60,7 +60,7 @@ We can leran these techniques from compilers. There are mainly two stages to mak


 #### Control Flow Graph
-To preform analyses on a program, it is often useful to make a control flow graph. A [control flow graph](https://en.wikipedia.org/wiki/Control_flow_graph) (CFG) in computer science is a representation, using graph notation, of all paths that might be traversed through a program during its execution. Each statement in the program is a node in the flow graph; if statemment x can be followed by statement y, there is an egde from x to y.
+To perform analysis on a program, it is often useful to make a control flow graph. A [control flow graph](https://en.wikipedia.org/wiki/Control_flow_graph) (CFG) in computer science is a representation, using graph notation, of all paths that might be traversed through a program during its execution. Each statement in the program is a node in the flow graph; if statemment x can be followed by statement y, there is an egde from x to y.

 Following is the flow graph for a simple loop.

@@ -68,18 +68,18 @@ Following is the flow graph for a simple loop.

 #### Dataflow Analysis

-liveness of variable "flows" around the edges of the control flow graph; determining the live range of each variable is an example of a dataflow problem. [Dataflow analysis](https://en.wikipedia.org/wiki/Data-flow_analysis) is a technique for gathering information about the possible set of values calculated at various points in a computer program.
+Liveness of variable "flows" around the edges of the control flow graph; determining the live range of each variable is an example of a dataflow problem. [Dataflow analysis](https://en.wikipedia.org/wiki/Data-flow_analysis) is a technique for gathering information about the possible set of values calculated at various points in a computer program.

 A simple way to perform data-flow analysis of programs is to set up dataflow equations for each node of the control flow graph and solve them by repeatedly calculating the output from the input locally at each node until the whole system stabilizes.

 - Flow Graph Terminology

-A flow graph node has out-edges that lead to sucessor nodes, and in-edges that come from presucessor nodes. The set *pred[n]* is all the predecessors of node n, and *succ[n]* is the set of sucessors.
+A flow graph node has out-edges that lead to sucessor nodes, and in-edges that come from predecessor nodes. The set *pred[n]* is all the predecessors of node n, and *succ[n]* is the set of sucessors.
 In former control flow graph, the out-edges of node 5 are 5 --> 6 and 5 --> 2, and *succ[5]* = {2, 6}. The in-edges of 2 are 5 --> 2 and 1 --> 2, and *pred[2]* = {1, 5}.

 - Uses and Defs

-An assignmemt to a variable or temporary defines that variable. An occurence of a variable on the right-hand side of an assginment(or in other expressions) uses the variable. We can speak the *def* of a variable as the set of graph nodes that define it; or the *def* of a graph node as the set of variables that it defines; and the similarly for the *use* of a variable or graph node. In former control flow graph, *def(3)* = {c}, *use(3)* = {b, c}.
+An assignmemt to a variable or temporary defines that variable. An occurence of a variable on the right-hand side of an assginment(or in other expressions) uses the variable. We can define the *def* of a variable as the set of graph nodes that define it; or the *def* of a graph node as the set of variables that it defines; and the similarly for the *use* of a variable or graph node. In former control flow graph, *def(3)* = {c}, *use(3)* = {b, c}.

 - Liveness

@@ -168,9 +168,9 @@ class ControlFlowGraph(object):
        return self._program
 ```

-#### make dataflow analysis
+#### Make dataflow analysis

-We follow guide from compilers and try to solve the dataflow equation to get liveness of every variable. If the live-in of an operator node is different from the live-out, then we can make memory sharing. 
+We follow the guide from compilers and try to solve the dataflow equation to get liveness of every variable. If the live-in of an operator node is different from the live-out, then we can make memory sharing. 

 For example:


--- a/doc/design/operator_kernel_type.md
+++ b/doc/design/operator_kernel_type.md
 # Design Doc: The Keys of Operator Kernel Type
 ## Problem
-An operator can have different kernel implementations, and each operator will have a map to store the related kernels. Fluid uses `OpKernelType` as a key to identify a unique Kernel. Before an operator runs, an certain kernel must be chosen by a key of `OpKernelType`. Currently, `OpKernelType` is defined as follows:
+An operator can have different kernel implementations, and each operator will have a map to store the related kernels. Fluid uses `OpKernelType` as a key to identify a unique kernel. Before an operator runs, a certain type of kernel must be chosen via a key of `OpKernelType`. Currently, `OpKernelType` is defined as follows:

 ```cpp
 struct OpKernelType {
@@ -10,13 +10,13 @@ struct OpKernelType {
 ```
 For more details, please refer to [codes](https://github.com/PaddlePaddle/Paddle/blob/2d5ec16bc8a09fb8e0f62c89b116b0cd1d333907/paddle/framework/operator.h#L348-L374) in github.

-It contains two keys, `Place` and `DataType`. And these two keys will be hashed to a unique key to represent a certain type of kernel. However, these two keys are not enough. We need a more complete representation of `OpKernelType`. 
+It contains two keys, `Place` and `DataType`. And these two keys will be hashed to a unique key to represent a certain type of kernel. However, these two keys do not provide enough information. We need a more complete representation of `OpKernelType`.

-We often implement a kernel of an operator with some computing library in certain device(place). Please remind that computing library and device are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices. 
+We often implement a kernel of an operator with some computing library on certain device(place). Please note that computing library and device do not have a one-to-one correspondence. A device can have a lot of computing libraries and a computing library can also support different devices.

-For example, Eigen library can support Nvidia GPU/AMD GPU/CPU. And MKLDNN library can support Intel CPU/Intel FPGA. Both `Place` and `Library` should be a key of `OpKernelType`.
+For example, Eigen library supports Nvidia GPU/AMD GPU/CPU and MKLDNN library supports Intel CPU/Intel FPGA. Both `Place` and `Library` should be a key of `OpKernelType`.

-It's obvious that different DataTypes, like fp64/fp32/int8 will have different kernels. But the data layout of a Tensor will also lead to different implementation. Please refer to the batch norm operator [kernels](https://github.com/PaddlePaddle/Paddle/blob/a948fac4d0ad7e0412d373b8aabeb711c2899563/paddle/operators/batch_norm_op.cc#L180-L209). Data Layout should also be taken into consideration.
+Different DataTypes, such as fp64/fp32/int8, will obviously have different kernels. But different data layout of a Tensor will also lead to different implementations. Please refer to the batch norm operator [kernels](https://github.com/PaddlePaddle/Paddle/blob/a948fac4d0ad7e0412d373b8aabeb711c2899563/paddle/operators/batch_norm_op.cc#L180-L209) as an example. Data layout should also be taken into consideration.

 ## Solution

@@ -31,17 +31,17 @@ struct OpKernelType {
 };
 ```

-Following is the details:
+The details are as follows:

 ### Place

-`Place` is defined as follows:
+`Place` is defined as:

 ```cpp
 typedef boost::variant<CUDAPlace, ROCmPlace, FPGAPlace, CPUPlace> Place;
 ```

-`Place` is to represent the device memory where data is locating.
+`Place` represents the device memory where data is located.


 ### Library
@@ -52,10 +52,10 @@ One operator kernel is usually implemented based on one library. `Library` is de
 enum Library { Plain, MKLDNN, CUDNN };
 ```

-We use `Plain` enumerator to represent default library. Since most operators in Fluid are implemented based on `Eigen` library, we take `Eigen` library as the `Plain` enumerator.
-A library usually has a corresponding `DeviceContext` which contains some handles needed by computation. Fluid now have two default DeviceContexts in CPU and CUDA, `CPUDeviceContext` and `CUDADeviceContext`. `CPUDeviceContext` contains a Eigen library handle and `CDUADeviceContext` contains a Eigen library handle and cuBLAS handle.
+We use `Plain` enumerator to represent default library. Since most operators in Fluid are implemented based on the `Eigen` library, we take `Eigen` library as the `Plain` enumerator.
+A library usually has a corresponding `DeviceContext` which contains some handles needed for computation. Fluid now has two default DeviceContexts for CPU and CUDA, namely, `CPUDeviceContext` and `CUDADeviceContext`. `CPUDeviceContext` contains an Eigen library handle and `CDUADeviceContext` contains an Eigen library handle and a cuBLAS handle.

-If we want to support new Library, a new enumerator need to be added to `Library` and a new corresponding `LibraryDeviceContext` will be created.
+If we want to support new library, a new enumerator need to be added to `Library` and a corresponding new `LibraryDeviceContext` need to be created.


 ### DataType
@@ -67,15 +67,15 @@ If we want to support new Library, a new enumerator need to be added to `Library

 Actually, a Tensor is a view of a block of memory. Besides a pointer to the memory, we also have to get some other descriptions of this block of memory, such as shape(ddim), stride, and layout.

-Different layout leads to different implementation of operator kernel. There are mainly 4 principles we have to follow to support layout in our fluid framework.
+Different layout leads to different implementation of the operator kernel. There are mainly 4 principles we have to follow to support layout in our Fluid framework.

- We take layout as a data member of Tensor. Layout is actually a enum variable. If fluid is built with MKLDNN, then, the memory format in MKLDNN will be added into this enum variable too.
+- We take layout as a data member of Tensor. Layout is actually a enum variable. If Fluid is built with MKLDNN, then the memory format in MKLDNN will also be added into this enum variable.

- Users have to set layout for input data. And some operators like fill_constant/random, also have to set layout of generating data. Of course, we can have some default layout, like NCHW.
+- Users have to set layout for input data. And some operators like fill_constant/random, also have to set layout for generating data. Of course, we can have some default layout, like NCHW.

- The inference of Layout is at run-time, not compile-time.
+- The inference of Layout is at run-time, not at compile-time.

- Every operator have to implement different kernels for different layouts. Let's take MKLDNN as an example, if we want to implement a MKLDNN convolution operator, we have to realize all the kernels for different layout, list at [here](http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html). And we will have a special macro to do registering kernels for MKLDNN operators.
+- Every operator has to implement different kernels for different layouts. Let's take MKLDNN as an example. If we want to implement an MKLDNN convolution operator, we have to implement all the kernels for different layouts, which are listed [here](http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html). And we will have a special macro to  register kernels for MKLDNN operators.

 `Layout` is also defined as a enum variable:


--- a/doc/design/python_api.md
+++ b/doc/design/python_api.md
@@ -279,6 +279,26 @@ class LayerHelper(object):
    return tmp
 ```

+### Return value of layer functions
+
+The layer will return a Variable, which is also the output of an operator.  However, outputs of a layer function have more attributes than an operator. There are parameter variables, and their gradient variables need to return. To return them is useful. For example,
+
+1. Users can debug the network by printing parameter gradients.
+2. Users can append attributes to a parameter, such as, `param.stop_gradient=True` will make a parameter stop generate the gradient. We can fix the parameter value during training by using this attribute.
+
+However, it is good to return a Variable for layers, since all layers and operators use Variables as their parameters. We can just append a `param` field and a `grad` field for layer function since the Python is dynamic typing.
+
+The sample usage is
+
+```python
+data = fluid.layers.data(...)
+hidden = fluid.layers.fc(data, ...)
+...
+
+executor.run(fetch_list=[hidden.param, hidden.param.grad], ...)
+```
+
+
 ## Optimizer

 [Optimizer Design Doc](./optimizer.md)
--- a/doc/design/var_desc.md
+++ b/doc/design/var_desc.md
 ## Background
-PaddlePaddle divides the description of neural network computation graph into two stages: compile time and runtime.
+PaddlePaddle divides the description of neural network computation into two stages: compile time and runtime. At compile time, the neural network computation is described as a `ProgramDesc` whereas at runtime an `Executor` interprets the `ProgramDesc` to compute the operations.

-PaddlePaddle use proto message to describe compile time graph because
+PaddlePaddle use proto message to describe compile time program because

-1. Computation graph should be able to be saved to a file.
-1. In distributed training, the graph will be serialized and send to multiple workers.
+1. The computation program description must be serializable and saved in a file.
+1. During distributed training, the sreialized program will be sent to multiple workers. It should also be possible to break the program into different components, each of which can be executed on different workers.

-The computation graph is constructed by Data Node and Operation Node. The concept to represent them is in the table below.
+The computation `Program` consists of nested `Blocks`. Each `Block` will consist of data(i.e. `Variable`)  and  `Operations`. The concept to represent them is in the table below.

 | |compile time|runtime|
 |---|---|---|

--- a/doc/getstarted/build_and_install/build_from_source_cn.rst
+++ b/doc/getstarted/build_and_install/build_from_source_cn.rst
@@ -32,6 +32,16 @@ PaddlePaddle主要使用 `CMake <https://cmake.org>`_ 以及GCC, G++作为编译

   pip install build/python/dist/*.whl

+如果机器中已经安装过PaddlePaddle，有两种方法：
+
+.. code-block:: bash
+
+   1. 先卸载之前的版本，再重新安装
+   pip uninstall paddlepaddle
+   pip install build/python/dist/*.whl
+
+   2. 直接升级到更新的版本
+   pip install build/python/dist/*.whl -U

 .. _run_test:


--- a/doc/getstarted/build_and_install/build_from_source_en.rst
+++ b/doc/getstarted/build_and_install/build_from_source_en.rst
@@ -36,6 +36,16 @@ machine or copy it to the target machine.

   pip install build/python/dist/*.whl

+If the machine has installed PaddlePaddle before, there are two methods:
+
+.. code-block:: bash
+
+   1. uninstall and reinstall
+   pip uninstall paddlepaddle
+   pip install build/python/dist/*.whl
+
+   2. upgrade directly
+   pip install build/python/dist/*.whl -U

 .. _run_test:


--- a/doc/howto/index_cn.rst
+++ b/doc/howto/index_cn.rst
@@ -9,6 +9,7 @@

  usage/cmd_parameter/index_cn.rst
  usage/cluster/cluster_train_cn.md
+  usage/capi/index_cn.rst

 开发标准
 --------

--- a/doc/howto/read_source.md
+++ b/doc/howto/read_source.md
@@ -26,16 +26,16 @@ sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
 sgd_optimizer.minimize(avg_cost)
 ```

- Variables: `x`,  `y`, `y_predict`, `cost` and `avg_cost`. [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/framework.py#L93)
- Layers: `fluid.layers.data`, `fluid.layers.fc` and `fluid.layers.mean` are layers. [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/layers.py)
+- Variables: `x`,  `y`, `y_predict`, `cost` and `avg_cost`. [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/framework.py#)
+- Layers: `fluid.layers.data`, `fluid.layers.fc` and `fluid.layers.mean` are layers. [Python](https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/v2/fluid/layers)
  - Every Layer has one or more operators and variables/parameters
    - All the operators are defined at [`paddle/operators/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators). Other worth-looking files:
      - Base class: [`paddle/framework/operator.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h)
      - Operator Registration: [`paddle/framework/op_registry.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h) 
      - Operator Lookup: [`paddle/framework/op_info.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_info.h)
 - Optimizer: `fluid.optimizer.SGD`. It does the following
-  - Add backward operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/backward.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/backward.cc)]
-  - Add optimizer operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/optimizer.py), [C++](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/optimizer)]
+  - Add backward operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/backward.py)]
+  - Add optimizer operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/optimizer.py)]

 # Run Time


--- a/doc/howto/usage/capi/compile_paddle_lib_cn.md
+++ b/doc/howto/usage/capi/compile_paddle_lib_cn.md
+## 编译 PaddlePaddle 预测库
+
+### 概述
+
+使用 C-API 进行预测依赖于将 PaddlePaddle 核心代码编译成链接库，只需在编译时需配制下面这些编译选项：
+
+必须配置选项：
+- `WITH_C_API`，必须配置为`ON`。
+
+推荐配置选项：
+- `WITH_PYTHON`，推荐配置为`OFF`
+- `WITH_SWIG_PY`，推荐配置为`OFF`
+- `WITH_GOLANG`，推荐设置为`OFF`
+
+可选配置选项：
+- `WITH_GPU`，可配置为`ON/OFF`
+- `WITH_MKL`，可配置为`ON/OFF`
+
+对推荐配置中的选项建议按照设置，以避免链接不必要的库。其它可选编译选项按需进行设定。
+
+下面的代码片段从github拉取最新代码，配制编译选项（需要将PADDLE_ROOT替换为PaddlePaddle预测库的安装路径）：
+
+```shell
+PADDLE_ROOT=/path/of/capi
+git clone https://github.com/PaddlePaddle/Paddle.git
+cd Paddle
+mkdir build
+cd build
+cmake -DCMAKE_INSTALL_PREFIX=$PADDLE_ROOT \
+      -DCMAKE_BUILD_TYPE=Release \
+      -DWITH_C_API=ON \
+      -DWITH_SWIG_PY=OFF \
+      -DWITH_GOLANG=OFF \
+      -DWITH_PYTHON=OFF \
+      -DWITH_MKL=OFF \
+      -DWITH_GPU=OFF  \
+      ..
+```
+
+执行上述代码生成Makefile文件后，执行：`make && make install`。成功编译后，使用C-API所需的依赖（包括：（1）编译出的PaddlePaddle预测库和头文件；（2）第三方链接库和头文件）均会存放于`PADDLE_ROOT`目录中。
+
+编译成功后在 `PADDLE_ROOT` 下会看到如下目录结构（包括了编译出的PaddlePaddle头文件和链接库，以及第三方依赖链接库和头文件（如果需要，由链接方式决定））：
+
+```text
+├── include
+│   └── paddle
+│       ├── arguments.h
+│       ├── capi.h
+│       ├── capi_private.h
+│       ├── config.h
+│       ├── error.h
+│       ├── gradient_machine.h
+│       ├── main.h
+│       ├── matrix.h
+│       ├── paddle_capi.map
+│       └── vector.h
+├── lib
+│   ├── libpaddle_capi_engine.a
+│   ├── libpaddle_capi_layers.a
+│   ├── libpaddle_capi_shared.so
+│   └── libpaddle_capi_whole.a
+└── third_party
+    ├── gflags
+    │   ├── include
+    │   │   └── gflags
+    │   │       ├── gflags_completions.h
+    │   │       ├── gflags_declare.h
+    │   │       ...
+    │   └── lib
+    │       └── libgflags.a
+    ├── glog
+    │   ├── include
+    │   │   └── glog
+    │   │       ├── config.h
+    │   │       ...
+    │   └── lib
+    │       └── libglog.a
+    ├── openblas
+    │   ├── include
+    │   │   ├── cblas.h
+    │   │   ...
+    │   └── lib
+    │       ...
+    ├── protobuf
+    │   ├── include
+    │   │   └── google
+    │   │       └── protobuf
+    │   │           ...
+    │   └── lib
+    │       └── libprotobuf-lite.a
+    └── zlib
+        ├── include
+        │   ...
+        └── lib
+            ...
+
+```
+
+### 链接说明
+
+目前提供三种链接方式：
+
+1. 链接`libpaddle_capi_shared.so` 动态库
+    - 使用 PaddlePaddle C-API 开发预测程序链接`libpaddle_capi_shared.so`时，需注意：
+        1. 如果编译时指定编译CPU版本，且使用`OpenBLAS`数学库，在使用C-API开发预测程序时，只需要链接`libpaddle_capi_shared.so`这一个库。
+        1. 如果是用编译时指定CPU版本，且使用`MKL`数学库，由于`MKL`库有自己独立的动态库文件，在使用PaddlePaddle C-API开发预测程序时，需要自己链接MKL链接库。
+        1. 如果编译时指定编译GPU版本，CUDA相关库会在预测程序运行时动态装载，需要将CUDA相关的库设置到`LD_LIBRARY_PATH`环境变量中。
+    - 这种方式最为简便，链接相对容易，**在无特殊需求情况下，推荐使用此方式**。
+
+2. 链接静态库 `libpaddle_capi_whole.a`
+    - 使用PaddlePaddle C-API 开发预测程序链接`libpaddle_capi_whole.a`时，需注意：
+        1. 需要指定`-Wl,--whole-archive`链接选项。
+        1. 需要显式地链接 `gflags`、`glog`、`libz`、`protobuf` 等第三方库，可在`PADDLE_ROOT/third_party`下找到。
+        1. 如果在编译 C-API 时使用OpenBLAS数学库，需要显示地链接`libopenblas.a`。
+        1. 如果在编译 C-API 是使用MKL数学库，需要显示地链接MKL的动态库。
+
+3. 链接静态库 `libpaddle_capi_layers.a`和`libpaddle_capi_engine.a`
+    - 使用PaddlePaddle C-API 开发预测程序链接`libpaddle_capi_whole.a`时，需注意：
+        1. 这种链接方式主要用于移动端预测。
+        1. 为了减少生成链接库的大小把`libpaddle_capi_whole.a`拆成以上两个静态链接库。
+        1. 需指定`-Wl,--whole-archive -lpaddle_capi_layers` 和 `-Wl,--no-whole-archive -lpaddle_capi_engine` 进行链接。
+        1. 第三方依赖库需要按照与方式2同样方法显示地进行链接。
--- a/doc/howto/usage/capi/images/csr.png
+++ b/doc/howto/usage/capi/images/csr.png
--- a/doc/howto/usage/capi/images/sequence_data.png
+++ b/doc/howto/usage/capi/images/sequence_data.png
--- a/doc/howto/usage/capi/images/workflow_of_CAPI.png
+++ b/doc/howto/usage/capi/images/workflow_of_CAPI.png
--- a/doc/howto/usage/capi/index_cn.rst
+++ b/doc/howto/usage/capi/index_cn.rst
+PaddlePaddle C-API
+==================
+
+..  toctree::
+  :maxdepth: 1
+
+  compile_paddle_lib_cn.md
+  organization_of_the_inputs_cn.md
+  workflow_of_capi_cn.md
--- a/doc/howto/usage/capi/organization_of_the_inputs_cn.md
+++ b/doc/howto/usage/capi/organization_of_the_inputs_cn.md
+## 输入/输出数据组织
+
+这篇文档介绍在使用 PaddlePaddle C-API 时如何组织输入数据，以及如何解析神经网络前向计算的输出结果。
+
+### 输入/输出数据类型
+在C-API中，按照基本数据类型在PaddlePaddle内部的定义和实现，输入数据可分为：
+1. 一维整型数组
+1. 二维浮点型矩阵
+    - 稠密矩阵
+    - 稀疏矩阵
+
+说明：
+1. 一维数组**仅支持整型值**；
+    - 常用于自然语言处理任务，例如：表示词语在词典中的序号；
+    - 分类任务中类别标签；
+1. 逻辑上高于二维的数据（例如含有多个通道的图片，视频等）在程序实现中都会转化为二维矩阵，转化方法在相应的领域都有通用解决方案，需要使用者自己了解并完成转化；
+1. 二维矩阵可以表示行向量和列向量，任何时候如果需要浮点型数组（向量），都应使用C-API中的矩阵来表示，而不是C-API中的一维数组。
+1. 不论是一维整型数组还是二维浮点数矩阵，**为它们附加上序列信息将变成序列输入。PaddlePaddle 会通过判数据是否附带有序列信息来判断一个向量/矩阵是否是一个序列**。当非序列输入时，无需关心和处理序列信息。关于什么是“序列信息”，下文会详细进行介绍。
+
+### 基本使用概念
+
+- 在PaddlePaddle内部，神经网络中一个计算层的输入/输出被组织为一个 `Argument` 结构体，如果神经网络有多个输入或者多个输入，每一个输入/输入都会对应有自己的`Argument`。
+- `Argument` 并不真正“存储”数据，而是将输入/输出信息有机地组织在一起。
+- 在`Argument`内部由`IVector`（对应着上文提到的一维整型数组）和`Matrix`（对应着上文提到的二维浮点型矩阵）来实际存储数据；由 `Sequence Start Positions` (下文详细解释) 来描述输入/输出的序列信息。
+
+- **注**：
+    1. 这篇文档之后部分将会统一使用`argument`来特指PaddlePaddle中神经网络计算层一个输入/输出数据。
+    1. 使用`paddle_ivector`来特指PaddlePaddle中的一维整型数组。
+    1. 使用`paddle_matrix`来特指PaddlePaddle中的二维浮点型矩阵。
+
+### 组织输入数据
+- 一维整型数组
+
+    概念上可以将`paddle_ivector`理解为一个一维的整型数组，通常用于表示离散的类别标签，或是在自然语言处理任务中表示词语在字典中的序号。下面的代码片段创建了含有三个元素`1`、`2`、`3`的`paddle_ivector`。
+    ```c
+    int ids[] = {1, 2, 3};
+     paddle_ivector ids_array =
+         paddle_ivector_create(ids, sizeof(ids) / sizeof(int), false, false);
+     CHECK(paddle_arguments_set_ids(in_args, 0, ids_array));
+    ```
+
+- **稠密矩阵**
+    - 一个`m×n`的稠密矩阵是一个由`m`行`n`列元素排列成的矩形阵列，矩阵里的元素是浮点数。对神经网络来说，矩阵的高度`m`是一次预测接受的样本数目，宽度$n$是神经网络定义时，`paddle.layer.data`的`size`。
+    - 下面的代码片段创建了一个高度为1，宽度为`layer_size`的稠密矩阵，矩阵中每个元素的值随机生成。
+
+    ```c
+    paddle_matrix mat = paddle_matrix_create(
+                            /* height = batch size */ 1,
+                            /* width = dimensionality of the data layer */ layer_size,
+                            /* whether to use GPU */ false);
+
+    paddle_real* array;
+    // Get the pointer pointing to the start address of the first row of the
+    // created matrix.
+    CHECK(paddle_matrix_get_row(mat, 0, &array));
+
+    // Fill the matrix with a randomly generated test sample.
+    srand(time(0));
+    for (int i = 0; i < layer_size; ++i) {
+      array[i] = rand() / ((float)RAND_MAX);
+    }
+
+    // Assign the matrix to the argument.
+    CHECK(paddle_arguments_set_value(in_args, 0, mat));
+    ```
+
+- **稀疏矩阵**
+
+  PaddlePaddle C-API 中 稀疏矩阵使用[CSR（Compressed Sparse Row Format）](https://en.wikipedia.org/wiki/Sparse_matrix#Compressed_sparse_row_(CSR,_CRS_or_Yale_format))格式存储。下图是CSR存储稀疏矩阵的示意图。
+  <p align="center">
+  <img src="https://user-images.githubusercontent.com/5842774/34159369-009fd328-e504-11e7-9e08-36bc6dc5e505.png" width=700><br> 图1. 稀疏矩阵存储示意图
+  </p>
+
+  CSR存储格式通过：（1）非零元素的值（上图中的`values`）；（2）行偏移(上图中的`row offsets`)：每一行元素在`values`中的起始偏移，`row offsets`中元素个数总是等于行数 + 1；（3）非零元素的列号（上图中的`column indices`）来确定稀疏矩阵的内容。
+
+  在PaddlePaddle C-API中，通过调用以下接口创建稀疏矩阵：
+
+  ```c
+  PD_API paddle_matrix paddle_matrix_create_sparse(
+      uint64_t height, uint64_t width, uint64_t nnz, bool isBinary, bool useGpu);
+  ```
+
+  1. 创建稀疏矩阵时需要显示地指定矩阵的（1）高度（`height`，在神经网络中等于一次预测处理的样本数）（2）宽度（`width`，`paddle.layer.data`的`size`）以及（3）非零元个数（`nnz`）。
+  1. 当上述接口第4个参数`isBinary`指定为`true`时，**只需要设置行偏移（`row_offset`）和列号(`colum indices`)，不需要提供元素值（`values`）**，这时行偏移和列号指定的元素默认其值为1。
+
+  下面的代码片段创建了一个CPU上的二值稀疏矩阵：
+
+  ```c
+  paddle_matrix mat = paddle_matrix_create_sparse(1, layer_size, nnz, true, false);
+  int colIndices[] = {9, 93, 109};  // layer_size here is greater than 109.
+  int rowOffset[] = {0, sizeof(colIndices) / sizeof(int)};
+
+  CHECK(paddle_matrix_sparse_copy_from(mat,
+                                 rowOffset,
+                                 sizeof(rowOffset) / sizeof(int),
+                                 colIndices,
+                                 (colIndices) / sizeof(int),
+                                 NULL /*values array is NULL.*/,
+                                 0 /*size of the value arrary is 0.*/));
+  CHECK(paddle_arguments_set_value(in_args, 0, mat));
+  ```
+  下面的代码片段在创建了一个CPU上的带元素值的稀疏矩阵：
+  ```c
+  paddle_matrix mat = paddle_matrix_create_sparse(1, layer_size, nnz, false, false);
+  int colIndices[] = {9, 93, 109};  // layer_size here is greater than 109.
+  int rowOffset[] = {0, sizeof(colIndices) / sizeof(int)};
+  float values[] = {0.5, 0.5, 0.5};
+
+  CHECK(paddle_matrix_sparse_copy_from(mat,
+                                 rowOffset,
+                                 sizeof(rowOffset) / sizeof(int),
+                                 colIndices,
+                                 sizeof(colIndices) / sizeof(int),
+                                 values,
+                                 sizeof(values) / sizeof(float)));
+  ```
+  注意事项：
+  1. 移动端预测**不支持**稀疏矩阵及相关的接口。
+
+### 组织序列信息
+
+多个排成一列的元素（可以是整型、浮点数、浮点数向量等）构成一个序列，元素之间的顺序是序列所携带的重要信息。不同序列可能会含有不同数目个元素。在 PaddlePaddle 中，序列输入/输出数据是在上文介绍的**数据输入（一维整型数组，二维浮点数矩阵）基础上，附加上序列信息**。下面详细解释什么是“序列信息”。
+
+我们将神经网络一次计算接受的所有输入样本称之为一个`batch`（可以含有一条或多条样本），每一个序列在整个`batch`中的偏移，就是PaddlePaddle中所指的**序列信息**，称之为“sequence start positions”。PaddlePaddle 支持两种序列类型：
+
+1. 单层序列
+    - 序列中的每一个元素是非序列，是进行计算的基本单位，不可再进行拆分。
+    - 例如：自然语言中的句子是一个序列，序列中的元素是词语；
+1. 双层序列
+    - 序列中的每一个元素又是一个序列。
+    - 例如：自然语言中的段落是一个双层序列；段落是由句子构成的序列；句子是由词语构成的序列。
+    - 双层序列在处理长序列的任务或是构建层级模型时会发挥作用。
+
+这篇文档之后部分会统一使用`sequence_start_positions`来特指：PaddlePaddle中神经网络计算层输入/输出所携带的序列信息。
+
+对双层序列来讲，不仅要提供每一个外层序列在整个`batch`中的偏移，每一个外层序列又含有若干个内层序列，需要同时提供每一个内层序列在整个`batch`中的偏移。也就是说：**双层序列需要设置分别为外层序列和内层序列分别设置`sequence_start_positions`信息**。
+
+**注：**
+1. 不论序列中的元素在内存中占用多少实际存储空间，`sequence_start_positions`表示的偏移是以“序列中的一个元素”作为统计的基本单位，而不是相对`batch`起始存储地址以数据的存储大小为单位的偏移。
+1. 非序列输入不携带`sequence_start_positions`，非序列输入无需构造`sequence_start_positions`。
+1. **不论是单层序列还是双层序列的序列信息，都使用`paddle_ivector`（也就是PaddlePaddle中的一维整型数组）来存储。**
+
+图2 是PaddlePaddle中单层序列和双层序列存储示意图。
+<p align="center">
+<img src="https://user-images.githubusercontent.com/5842774/34159714-1f81a9be-e505-11e7-8a8a-4902146ec899.png" width=800><br>图2. 序列输入示意图
+</p>
+
+- 单层序列
+
+    图2 (a) 展示了一个含有4个序列的`batch`输入：
+    1. 4个序列的长度分别为：5、3、2、4；
+    1. 这时的`sequence_start_positions`为：`[0, 5, 8, 10, 14]`；
+    1. 本地训练. 不论数据域是`paddle_ivector`类型还是`paddle_matrix`类型，都可以通过调用下面的接口为原有的数据输入附加上序列信息，使之变为一个单层序列输入，代码片段如下：
+
+    ```c
+    int seq_pos_array[] = {0, 5, 8, 10, 14};
+    paddle_ivector seq_pos = paddle_ivector_create(
+        seq_pos_array, sizeof(seq_pos_array) / sizeof(int), false, false);
+    // Suppose the network only has one input data layer.
+    CHECK(paddle_arguments_set_sequence_start_pos(in_args, 0, 0, seq_pos));
+    ```
+
+- 双层序列
+
+    图2 (b) 展示了一个含有4个序列的`batch`输入；
+    1. 4个序列的长度分别为：5、3、2、4；这四个序列又分别含有3、2、1、2个子序列；
+    1. 这时的需要同时提供：
+        - 外层序列在`batch`中的起始偏移`：[0, 5, 8, 10, 14]`；
+        - 内层序列在`batch`中的起始偏移：`[0, 2, 3, 5, 7， 8， 10， 13， 14]`；
+    1. 不论数据域是`paddle_ivector`类型还是`paddle_matrix`类型，这时需要调用创建序列信息和为`argument`设置序列信息的接口**两次**，分别为数据输入添加外层序列和内层序列的序列信息，使之变为一个双层序列输入，代码片段如下：
+    ```c
+    // set the sequence start positions for the outter sequences.
+    int outter_seq_pos_array[] = {0, 5, 8, 10, 14};
+    paddle_ivector seq_pos =
+        paddle_ivector_create(outter_seq_pos_array,
+                              sizeof(outter_pos_array) / sizeof(int),
+                              false,
+                              false);
+    // The third parameter of this API indicates the sequence level.
+    // 0 for the outter sequence. 1 for the inner sequence.
+    // If the input is a sequence not the nested sequence, the third parameter is
+    // fixed to be 0.
+    CHECK(paddle_arguments_set_sequence_start_pos(in_args, 0, 0, seq_pos));
+
+    // set the sequence start positions for the outter sequences.
+    int inner_seq_pos_array[] = {0, 2, 3, 5, 7， 8， 10， 13， 14};
+    paddle_ivector seq_pos = paddle_ivector_create(
+        inner_pos_array, sizeof(inner_pos_array) / sizeof(int), false, false);
+    // The third parameter of this API indicates the sequence level.
+    // 0 for the outter sequence. 1 for the inner sequence.
+    CHECK(paddle_arguments_set_sequence_start_pos(in_args, 0, 1, seq_pos));
+    ```
+
+注意事项：
+1. 当一个`batch`中含有多个序列，**不支持序列长度为`0`的序列（也就是空输入）** 作为输入。不同计算层对空输入的处理策略有可能不同，潜在会引起未定义行为，或者引起行时错误，请在输入时进行合法性检查。
+
+### Python 端数据类型说明
+
+下表列出了Python端训练接口暴露的数据类型（`paddle.layer.data`函数`type`字段的取值）对应于调用C-API需要创建的数据类型：
+
+<html>
+<table border="2" frame="border">
+<table>
+<thead>
+<tr>
+<th style="text-align:left">Python 端数据类型</th>
+<th style="text-align:left">C-API 输入数据类型</th>
+</tr>
+</thead>
+<tbody>
+<tr>
+<td style="text-align:left">paddle.data_type.integer_value</td>
+<td style="text-align:left">整型数组，无需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.dense_vector</td>
+<td style="text-align:left">浮点型稠密矩阵，无需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.sparse_binary_vector</td>
+<td style="text-align:left">浮点型稀疏矩阵，无需提供非零元的值，默认为1，无需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.sparse_vector</td>
+<td style="text-align:left">浮点型稀疏矩阵，需提供非零元的值，无需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.integer_value_sequence</td>
+<td style="text-align:left">整型数组，需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.dense_vector_sequence</td>
+<td style="text-align:left">浮点型稠密矩阵，需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.sparse_binary_vector_sequence</td>
+<td style="text-align:left">浮点型稀疏矩阵，无需提供非零元的值，默认为1，需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.sparse_vector_sequence</td>
+<td style="text-align:left">浮点型稀疏矩阵，需提供非零元的值，需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.integer_value_sub_sequence</td>
+<td style="text-align:left">整型数组，需附加双层序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.dense_vector_sub_sequence</td>
+<td style="text-align:left">浮点型稠密矩阵，需附加双层序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.sparse_binary_vector_sub_sequence</td>
+<td style="text-align:left">浮点型稀疏矩阵，无需提供非零元的值，默认为1，需附加双层序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.sparse_vector_sub_sequence</td>
+<td style="text-align:left">浮点型稀疏矩阵，需提供非零元的值，需附加双层序列信息</td>
+</tr>
+</tbody>
+</table>
+</html>
+<br>
+
+
+### 输出数据
+
+PaddlePaddle中一个计算层的输出数据组织方式和输入数据组织方式完全相同。一个输出数据同样被组织为一个`argument`，`argument`通过`paddle_matrix`或`paddle_ivector`存数数据，如果输出是一个序列，那么会携带有`sequence_start_positions`信息。调用C-API相关接口，读取需要的结果即可。
+
+### 总结
+
+- 在PaddlePaddle内部，神经网络中一个计算层的输入/输出被组织为`argument`。
+- `argument`并不真正“存储”数据，而是将输入/输出信息有机地组织在一起。
+- 在`argument`内部由`paddle_ivector`（一维整型数组）和`paddle_matrix`（二维浮点型矩阵）来实际存储数据。
+如果是一个序列输入/输出由 `sequence start positions` 来记录输入/输出的序列信息。
+
+于是，在组织神经网络输入时，需要思考完成以下工作：
+1. 为每一个输入/输出创建`argument`。
+    - C-API 中操作`argument`的接口请查看[argument.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/arguments.h)。
+1. 为每一个`argument`创建`paddle_matrix`或者`paddle_ivector`来存储数据。
+    - C-API 中操作`paddle_ivector`的接口请查看 [vector.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/vector.h)。
+    - C-API 中操作`paddle_matrix`的接口请查看[matrix.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/matrix.h)。
+1. 如果输入是序列数据，需要创建并填写`sequence_start_positions`信息。
+    - 通过调用 [`paddle_arguments_set_sequence_start_pos`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/arguments.h#L137) 来为一个`argument`添加序列信息。
+    - 通过调用 [`paddle_arguments_get_sequence_start_pos`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/arguments.h#L150) 来读取一个`argument`添加序列信息。
+    - 接口说明请查看 [argument.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/arguments.h) 文件。
--- a/doc/howto/usage/capi/workflow_of_capi_cn.md
+++ b/doc/howto/usage/capi/workflow_of_capi_cn.md
+## C-API 使用流程
+
+这篇文档介绍 PaddlePaddle C-API 整体使用流程。
+
+### 使用流程
+
+使用 C-API 的工作流程如图1所示，分为（1）准备预测模型和（2）预测程序开发两大部分。
+
+<p align="center">
+<img src="https://user-images.githubusercontent.com/5842774/34658453-365f73ea-f46a-11e7-9b3f-0fd112b27bae.png" width=500><br> 图1. C-API使用流程示意图
+</p>
+
+- 准备预测模型
+    1. 只将神经网络结构进行序列化。
+        - 只对神经网络结构进行序列化，加载模型需同时指定：网络结构的序列化结果和模型参数存储目录。
+    1. 将网络结构定义和训练结束存储下来的模型参数文件（多个）合并入一个文件。
+        - 神经网络模型结构和训练好的模型将被序列化合并入一个文件。
+        - 预测时只需加载一个文件便于发布。
+    - **注意**：以上两种方式只需选择其一即可。
+- 调用 C-API 开发预测序
+    1. 初始化PaddlePaddle运行环境。
+    1. 加载预测模型。
+    1. 创建神经网络输入，组织输入数据。
+    1. 进行前向计算，获得计算结果。
+    1. 清理和结束。
+
+### 准备预测模型
+
+准备预测模型部分，我们以手写数字识别任务为例进行介绍。手写数字识别任务定义了一个含有[两个隐层的简单全连接网络](https://github.com/PaddlePaddle/book/blob/develop/02.recognize_digits/README.cn.md#softmax回归softmax-regression)，网络接受一幅图片作为输入，将图片分类到 0 ~ 9 类别标签之一。完整代码可以查看[此目录](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense) 中的相关脚本。
+
+调用C-API开发预测程序需要一个训练好的模型，运行[MNIST手写数字识别目录](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)下的[mnist_v2.py](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/examples/model_inference/dense/mnist_v2.py)脚本，在终端执行`python mnist_v2.py`，会使用 PaddlePaddle 内置的 [MNIST 数据集](http://yann.lecun.com/exdb/mnist/)进行训练。训练好的模型默认保存在当前运行目录下的`models`目录中。
+
+下面，我们将训练结束后存储下来的模型转换成预测模型。
+
+1. 序列化神经网络模型配置
+
+    PaddlePaddle 使用 protobuf 来传输网络配置文件中定义的网络结构和相关参数，使用 C-API 进行预测时，需要将网络结构使用 protobuf 进行序列化，写入文件中。
+
+    调用[`paddle.utils.dump_v2_config`](https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/utils/dump_v2_config.py)中的`dump_v2_config`函数能够将使用 PaddlePaddle V2 API 定义的神经网络结构 dump 到指定文件中，示例代码如下：
+
+    ```python
+    from paddle.utils.dump_v2_config import dump_v2_config
+    from mnist_v2 import network
+
+    predict = network(is_infer=True)
+    dump_v2_config(predict, "trainer_config.bin", True)
+    ```
+
+    对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)这个示例，[`mnist_v2.py`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense/mnist_v2.py)脚本集成了序列化神经网络结构的过程，可以直接运行 `python mnist_v2.py --task dump_config` 对神经网络结构进行序列化，结果会写入当前运行目录下的`trainer_config.bin`文件中。
+
+    使用这种方式，需要**在运行时将神经网络的多个可学习参数放在同一个目录中**，C-API可以通过分别指定序列化后的网络结构文件和参数目录来加载训练好的模型。
+
+2. 合并模型文件(可选)
+
+    一些情况为了便于发布，希望能够将序列化后的神经网络结构和训练好的模型参数打包进一个文件。对于这样的需求，可以使用`paddle.utils.merge_model`中的`merge_v2_model`接口对神经网络结构和训练好的参数进行序列化，将序列化结果写入一个文件内。
+
+    代码示例如下：
+
+    ```python
+    from paddle.utils.merge_model import merge_v2_modelss
+    from mnist_v2 import network
+
+    net = network(is_infer=True)
+    param_file = "models/params_pass_4.tar"
+    output_file = "output.paddle.model"
+    merge_v2_model(net, param_file, output_file)
+    ```
+    对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)这个示例，可直接运行 `python` [merge_v2_model.py](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense/merge_v2_model.py)。序列化结果会写入当前运行目录下的`output.paddle.model`文件中。使用这种方式，运行时C-API可以通过指定`output.paddle.model`文件的路径来加载预测模型。
+
+#### 注意事项
+1. 为使用C-API，在调用`dump_v2_config`序列化神经网络结构时，参数`binary`必须指定为`True`。
+1. **预测使用的网络结构往往不同于训练**，通常需要去掉网络中的：（1）类别标签层；（2）损失函数层；（3）`evaluator`等，只留下核心计算层，请注意是否需要修改网络结构。
+1. 预测时，可以获取网络中定义的任意多个（大于等于一个）层前向计算的结果，需要哪些层的计算结果作为输出，就将这些层加入一个Python list中，作为调用`dump_v2_config`的第一个参数。
+
+### 编写预测代码
+
+预测代码更多详细示例代码请参考[C-API使用示例](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference) 目录下的代码示例。这一节对图1中预测代码编写的5个步骤进行介绍和说明。
+
+#### step 1. 初始化PaddlePaddle运行环境
+第一步需调用[`paddle_init`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/main.h#L27) 初始化PaddlePaddle运行环境，该接口接受两个参数：参数的个数和参数列表。
+
+#### step2. 加载模型
+
+这里介绍C-API使用中的一个重要概念：Gradient Machine。
+
+概念上，在 PaddlePaddle 内部，一个GradientMachine类的对象管理着一组计算层（PaddlePaddle Layers）来完成前向和反向计算，并处理与之相关的所有细节。在调用C-API预测时，只需进行前向计算而无需调用反向计算。这篇文档之后部分会使用`gradient machine`来特指调用PaddlePaddle C-API创建的GradientMachine类的对象。每一个 `gradient machine` 都会管理维护一份训练好的模型，下面是C-API提供的，两种常用的模型加载方式：
+
+1. 调用[`paddle_gradient_machine_load_parameter_from_disk`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/gradient_machine.h#L61)接口，从磁盘加载预测模型。这时`gradient machine`会独立拥有一份训练好的模型；
+1. 调用[`paddle_gradient_machine_create_shared_param`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/gradient_machine.h#L88)接口，与其它`gradient machine`的共享已经加载的预测模型。这种情况多出现在使用多线程预测时，通过多个线程共享同一个模型来减少内存开销。可参考[此示例](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/examples/model_inference/multi_thread/main.c)。
+
+- 注意事项
+    1. 使用PaddlePaddle V2 API训练，模型中所有可学习参数会被存为一个压缩文件，需要手动进行解压，将它们放在同一目录中，C-API不会直接加载 V2 API 存储的压缩文件。
+    1. 如果使用`merge model`方式将神经网络结构和训练好的参数序列化到一个文件，请参考此[示例](https://github.com/PaddlePaddle/Mobile/blob/develop/Demo/linux/paddle_image_recognizer.cpp#L59)。
+    1. 通过灵活使用以上两个接口，加载模型可其它多种方式，例如也可在程序运行过程中再加载另外一个模型。
+
+#### step 3. 创建神经网络输入，组织输入数据
+
+基本使用概念：
+- 在PaddlePaddle内部，神经网络中一个计算层的输入输出被组织为一个 `Argument` 结构体，如果神经网络有多个输入或者多个输出，每一个输入/输出都会对应有自己的`Argument`。
+- `Argument` 并不真正“存储”数据，而是将输入/输出数据有机地组织在一起。
+- 在`Argument`内部由：1. `Matrix`（二维矩阵，存储浮点类型输入/输出）；2. `IVector`（一维数组，**仅用于存储整型值**，多用于自然语言处理任务）来实际存储数据。
+
+C-API支持的所有输入数据类型和他们的组织方式，请参考“输入/输出数据组织”一节。
+
+这篇文档的之后部分会使用`argument`来特指PaddlePaddle C-API中神经网络的一个输入/输出，使用`paddle_matrix`**特指**`argument`中用于存储数据的`Matrix`类的对象。
+
+在组织神经网络输入，获取输出时，需要思考完成以下工作：
+1. 为每一个输入/输出创建`argument`；
+1. 为每一个`argument`创建`paddle_matrix`来存储数据；
+
+与输入不同的是，不需在使用C-API时为输出`argument`的`paddle_matrix`对象分配空间。前向计算之后PaddlePaddle内部已经分配/管理了每个计算层输出的存储空间。
+
+#### step 4. 前向计算
+
+完成上述准备之后，通过调用 [`paddle_gradient_machine_forward`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/gradient_machine.h#L73) 接口完成神经网络的前向计算。
+
+#### step 5. 清理
+
+结束预测之后，对使用的中间变量和资源进行清理和释放。
--- a/paddle/capi/examples/model_inference/dense/main.c
+++ b/paddle/capi/examples/model_inference/dense/main.c
@@ -3,59 +3,82 @@

 #include "../common/common.h"

+// Modify this path as needed.
 #define CONFIG_BIN "./trainer_config.bin"
+// Modify this path as needed.
+// This demo assumes that merged model is not used, then this path is the
+// directory storing all the trained parameters.
+// If the model is trained by PaddlePaddle V2 API, the model is saved as
+// a compressed file. You need to uncompress the compressed file first.
+#define MODEL_PATH "models/pass_4"

 int main() {
-  // Initalize Paddle
+  // Initalize the PaddlePaddle runtime environment.
  char* argv[] = {"--use_gpu=False"};
  CHECK(paddle_init(1, (char**)argv));

-  // Reading config binary file. It is generated by `convert_protobin.sh`
+  // Read the binary configuration file generated by `convert_protobin.sh`
  long size;
  void* buf = read_config(CONFIG_BIN, &size);

-  // Create a gradient machine for inference.
+  // Create the gradient machine for inference.
  paddle_gradient_machine machine;
  CHECK(paddle_gradient_machine_create_for_inference(&machine, buf, (int)size));
-  CHECK(paddle_gradient_machine_randomize_param(machine));

-  // Loading parameter. Uncomment the following line and change the directory.
-  // CHECK(paddle_gradient_machine_load_parameter_from_disk(machine,
-  //                                                "./some_where_to_params"));
+  // Load the trained model. Modify the parameter MODEL_PATH to set the correct
+  // path of the trained model.
+  CHECK(paddle_gradient_machine_load_parameter_from_disk(machine, MODEL_PATH));
+
+  // Inputs and outputs of the network are organized as paddle_arguments object
+  // in C-API. In the comments below, "argument" specifically means one input of
+  // the neural network in PaddlePaddle C-API.
  paddle_arguments in_args = paddle_arguments_create_none();

-  // There is only one input of this network.
+  // There is only one data layer in this demo MNIST network, invoke this
+  // function to create one argument.
  CHECK(paddle_arguments_resize(in_args, 1));

-  // Create input matrix.
-  paddle_matrix mat = paddle_matrix_create(/* sample_num */ 1,
-                                           /* size */ 784,
-                                           /* useGPU */ false);
-  srand(time(0));
+  // Each argument needs one matrix or one ivector (integer vector, for sparse
+  // index input, usually used in NLP task) to holds the real input data.
+  // In the comments below, "matrix" specifically means the object needed by
+  // argument to hold the data. Here we create the matrix for the above created
+  // agument to store the testing samples.
+  paddle_matrix mat =
+      paddle_matrix_create(/* height = batch size */ 1,
+                           /* width = dimensionality of the data layer */ 784,
+                           /* whether to use GPU */ false);

  paddle_real* array;
-
-  // Get First row.
+  // Get the pointer pointing to the start address of the first row of the
+  // created matrix.
  CHECK(paddle_matrix_get_row(mat, 0, &array));

+  // Fill the matrix with a randomly generated test sample.
+  srand(time(0));
  for (int i = 0; i < 784; ++i) {
    array[i] = rand() / ((float)RAND_MAX);
  }

+  // Assign the matrix to the argument.
  CHECK(paddle_arguments_set_value(in_args, 0, mat));

+  // Create the output argument.
  paddle_arguments out_args = paddle_arguments_create_none();
+
+  // Invoke the forward computation.
  CHECK(paddle_gradient_machine_forward(machine,
                                        in_args,
                                        out_args,
-                                        /* isTrain */ false));
-  paddle_matrix prob = paddle_matrix_create_none();
+                                        /* is train taks or not */ false));

+  // Create the matrix to hold the forward result of the neural network.
+  paddle_matrix prob = paddle_matrix_create_none();
+  // Access the matrix of the output argument, the predicted result is stored in
+  // which.
  CHECK(paddle_arguments_get_value(out_args, 0, prob));

  uint64_t height;
  uint64_t width;
-
  CHECK(paddle_matrix_get_shape(prob, &height, &width));
  CHECK(paddle_matrix_get_row(prob, 0, &array));

@@ -68,6 +91,7 @@ int main() {
  }
  printf("\n");

+  // The cleaning up.
  CHECK(paddle_matrix_destroy(prob));
  CHECK(paddle_arguments_destroy(out_args));
  CHECK(paddle_matrix_destroy(mat));

--- a/paddle/capi/examples/model_inference/dense/merge_v2_model.py
+++ b/paddle/capi/examples/model_inference/dense/merge_v2_model.py
+from paddle.utils.merge_model import merge_v2_model
+
+from mnist_v2 import network
+
+net = network(is_infer=True)
+param_file = "models/params_pass_4.tar"
+output_file = "output.paddle.model"
+merge_v2_model(net, param_file, output_file)
--- a/paddle/capi/examples/model_inference/dense/mnist_v2.py
+++ b/paddle/capi/examples/model_inference/dense/mnist_v2.py
+import os
+import sys
+import gzip
+import logging
+import argparse
+from PIL import Image
+import numpy as np
+
+import paddle.v2 as paddle
+from paddle.utils.dump_v2_config import dump_v2_config
+
+logger = logging.getLogger("paddle")
+logger.setLevel(logging.INFO)
+
+
+def multilayer_perceptron(img, layer_size, lbl_dim):
+    for idx, size in enumerate(layer_size):
+        hidden = paddle.layer.fc(input=(img if not idx else hidden),
+                                 size=size,
+                                 act=paddle.activation.Relu())
+    return paddle.layer.fc(input=hidden,
+                           size=lbl_dim,
+                           act=paddle.activation.Softmax())
+
+
+def network(input_dim=784, lbl_dim=10, is_infer=False):
+    images = paddle.layer.data(
+        name='pixel', type=paddle.data_type.dense_vector(input_dim))
+
+    predict = multilayer_perceptron(
+        images, layer_size=[128, 64], lbl_dim=lbl_dim)
+
+    if is_infer:
+        return predict
+    else:
+        label = paddle.layer.data(
+            name='label', type=paddle.data_type.integer_value(lbl_dim))
+        return paddle.layer.classification_cost(input=predict, label=label)
+
+
+def main(task="train", use_gpu=False, trainer_count=1, save_dir="models"):
+    if task == "train":
+        if not os.path.exists(save_dir):
+            os.mkdir(save_dir)
+
+        paddle.init(use_gpu=use_gpu, trainer_count=trainer_count)
+        cost = network()
+        parameters = paddle.parameters.create(cost)
+        optimizer = paddle.optimizer.Momentum(
+            learning_rate=0.1 / 128.0,
+            momentum=0.9,
+            regularization=paddle.optimizer.L2Regularization(rate=0.0005 * 128))
+
+        trainer = paddle.trainer.SGD(cost=cost,
+                                     parameters=parameters,
+                                     update_equation=optimizer)
+
+        def event_handler(event):
+            if isinstance(event, paddle.event.EndIteration):
+                if event.batch_id % 100 == 0:
+                    logger.info("Pass %d, Batch %d, Cost %f, %s" %
+                                (event.pass_id, event.batch_id, event.cost,
+                                 event.metrics))
+            if isinstance(event, paddle.event.EndPass):
+                with gzip.open(
+                        os.path.join(save_dir, "params_pass_%d.tar" %
+                                     event.pass_id), "w") as f:
+                    trainer.save_parameter_to_tar(f)
+
+        trainer.train(
+            reader=paddle.batch(
+                paddle.reader.shuffle(
+                    paddle.dataset.mnist.train(), buf_size=8192),
+                batch_size=128),
+            event_handler=event_handler,
+            num_passes=5)
+    elif task == "dump_config":
+        predict = network(is_infer=True)
+        dump_v2_config(predict, "trainer_config.bin", True)
+    else:
+        raise RuntimeError(("Error value for parameter task. "
+                            "Available options are: train and dump_config."))
+
+
+def parse_cmd():
+    parser = argparse.ArgumentParser(
+        description="PaddlePaddle MNIST demo for CAPI.")
+    parser.add_argument(
+        "--task",
+        type=str,
+        required=False,
+        help=("A string indicating the taks type. "
+              "Available options are: \"train\", \"dump_config\"."),
+        default="train")
+    parser.add_argument(
+        "--use_gpu",
+        type=bool,
+        help=("A bool flag indicating whether to use GPU device or not."),
+        default=False)
+    parser.add_argument(
+        "--trainer_count",
+        type=int,
+        help=("This parameter is only used in training task. It indicates "
+              "how many computing threads are created in training."),
+        default=1)
+    parser.add_argument(
+        "--save_dir",
+        type=str,
+        help=("This parameter is only used in training task. It indicates "
+              "path of the directory to save the trained models."),
+        default="models")
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = parse_cmd()
+    main(args.task, args.use_gpu, args.trainer_count, args.save_dir)
--- a/paddle/capi/examples/model_inference/sparse_binary/main.c
+++ b/paddle/capi/examples/model_inference/sparse_binary/main.c
 #include <paddle/capi.h>
 #include <time.h>
+
 #include "../common/common.h"

 #define CONFIG_BIN "./trainer_config.bin"
@@ -9,16 +10,18 @@ int main() {
  char* argv[] = {"--use_gpu=False"};
  CHECK(paddle_init(1, (char**)argv));

-  // Reading config binary file. It is generated by `convert_protobin.sh`
+  // Read the binary configuration file which is generated by
+  // `convert_protobin.sh`
  long size;
  void* buf = read_config(CONFIG_BIN, &size);

-  // Create a gradient machine for inference.
+  // Create the gradient machine for inference.
  paddle_gradient_machine machine;
  CHECK(paddle_gradient_machine_create_for_inference(&machine, buf, (int)size));
  CHECK(paddle_gradient_machine_randomize_param(machine));

-  // Loading parameter. Uncomment the following line and change the directory.
+  // Load the trained parameters. Uncomment the following line and change the
+  // directory as needed.
  // CHECK(paddle_gradient_machine_load_parameter_from_disk(machine,
  //                                                "./some_where_to_params"));
  paddle_arguments in_args = paddle_arguments_create_none();
@@ -26,7 +29,7 @@ int main() {
  // There is only one input of this network.
  CHECK(paddle_arguments_resize(in_args, 1));

-  // Create input matrix.
+  // Create the input matrix.
  paddle_matrix mat = paddle_matrix_create_sparse(1, 784, 3, true, false);
  srand(time(0));
  paddle_real* array;

--- a/paddle/framework/CMakeLists.txt
+++ b/paddle/framework/CMakeLists.txt
@@ -47,7 +47,7 @@ cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
 cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
 cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
-    shape_inference data_transform)
+    shape_inference data_transform lod_tensor)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry init)
 cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog)


--- a/paddle/framework/init.cc
+++ b/paddle/framework/init.cc
@@ -11,6 +11,7 @@ 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. */
+#include <string.h>  // for strdup
 #include <algorithm>
 #include <string>

@@ -60,7 +61,9 @@ void InitDevices() {
 }

 void InitGLOG(const std::string &prog_name) {
-  google::InitGoogleLogging(prog_name.c_str());
+  // glog will not hold the ARGV[0] inside.
+  // Use strdup to alloc a new string.
+  google::InitGoogleLogging(strdup(prog_name.c_str()));
  google::InstallFailureSignalHandler();
 }


--- a/paddle/framework/lod_tensor.cc
+++ b/paddle/framework/lod_tensor.cc
@@ -69,6 +69,12 @@ std::ostream &operator<<(std::ostream &os, const LoDTensor &t) {
  return os;
 }

+std::string LoDToString(const LoD &lod) {
+  std::ostringstream stream;
+  stream << lod;
+  return stream.str();
+}
+
 LoD SliceInLevel(const LoD &in, size_t level, size_t elem_begin,
                 size_t elem_end) {
  PADDLE_ENFORCE_LT(level, in.size());

--- a/paddle/framework/lod_tensor.h
+++ b/paddle/framework/lod_tensor.h
@@ -60,6 +60,8 @@ using LoD = std::vector<Vector<size_t>>;
 std::ostream& operator<<(std::ostream& os, const LoD& lod);
 std::ostream& operator<<(std::ostream& os, const LoDTensor& t);

+std::string LoDToString(const LoD& lod);
+
 LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin,
                 size_t elem_end);
 /*

--- a/paddle/framework/operator.cc
+++ b/paddle/framework/operator.cc
@@ -11,6 +11,7 @@ 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. */
+#include <gflags/gflags.h>
 #include <glog/logging.h>

 #include <algorithm>
@@ -21,6 +22,10 @@ limitations under the License. */
 #include "paddle/framework/shape_inference.h"
 #include "paddle/framework/var_type.h"

+DEFINE_bool(op_sync, false,
+            "Default cuda is asynchronous device, set to True will"
+            "force op run in synchronous mode.");
+
 namespace paddle {
 namespace framework {

@@ -75,7 +80,9 @@ static DDim GetDims(const Scope& scope, const std::string& name) {
  Variable* var = scope.FindVar(name);
  if (var == nullptr) {
    return DDim({-1});
-  } else if (var->IsType<LoDTensor>()) {
+  }
+
+  if (var->IsType<LoDTensor>()) {
    return var->Get<LoDTensor>().dims();
  } else if (var->IsType<SelectedRows>()) {
    return var->Get<SelectedRows>().GetCompleteDims();
@@ -84,6 +91,21 @@ static DDim GetDims(const Scope& scope, const std::string& name) {
  }
 }

+static LoD GetLoD(const Scope& scope, const std::string& name) {
+  Variable* var = scope.FindVar(name);
+  auto default_lod = LoD({{}});
+
+  if (var == nullptr) {
+    return default_lod;
+  }
+
+  if (var->IsType<LoDTensor>()) {
+    return var->Get<LoDTensor>().lod();
+  } else {
+    return default_lod;
+  }
+}
+
 std::string OperatorBase::Input(const std::string& name) const {
  auto& ins = Inputs(name);
  PADDLE_ENFORCE_LE(ins.size(), 1UL,
@@ -125,7 +147,8 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const {
    for (size_t i = 0; i < input.second.size(); ++i) {
      ss << input.second[i];
      if (scope) {
-        ss << "(" << GetDims(*scope, input.second[i]) << ")";
+        ss << "[" << GetDims(*scope, input.second[i]) << "]";
+        ss << "(" << GetLoD(*scope, input.second[i]) << ")";
      }
      if (i != input.second.size() - 1) {
        ss << ", ";
@@ -144,7 +167,8 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const {
    for (size_t i = 0; i < output.second.size(); ++i) {
      ss << output.second[i];
      if (scope) {
-        ss << "(" << GetDims(*scope, output.second[i]) << ")";
+        ss << "[" << GetDims(*scope, output.second[i]) << "]";
+        ss << "(" << GetLoD(*scope, output.second[i]) << ")";
      }
      if (i != output.second.size() - 1) {
        ss << ", ";
@@ -542,8 +566,14 @@ void OperatorWithKernel::Run(const Scope& scope,

  auto kernel_iter = kernels.find(expected_kernel_key);

-  kernel_iter->second->Compute(ExecutionContext(
-      *this, new_scope, *pool.Get(expected_kernel_key.place_)));
+  auto* new_dev_ctx = pool.Get(expected_kernel_key.place_);
+  kernel_iter->second->Compute(
+      ExecutionContext(*this, new_scope, *new_dev_ctx));
+
+  /*For profiling/benchmark only*/
+  if (FLAGS_op_sync) {
+    new_dev_ctx->Wait();
+  }
 }

 proto::DataType OperatorWithKernel::IndicateDataType(

--- a/paddle/framework/tensor_util.h
+++ b/paddle/framework/tensor_util.h
@@ -116,8 +116,8 @@ inline void Copy(const Tensor& src, const platform::Place& dst_place,
 * @param[in] src        The external tensor.
 * @param[in] ctx        The device context contains device resources.
 *
- * * @note    CopyFromVector assumes that the tensor has been resized
- *            before invoking.
+ * * @note    CopyFromVector will resize dst to an 1D tensor with the same
+ *            size as src.
 */
 template <typename T>
 inline void CopyFromVector(const std::vector<T>& src,

--- a/paddle/operators/CMakeLists.txt
+++ b/paddle/operators/CMakeLists.txt
@@ -135,6 +135,7 @@ op_library(detection_output_op DEPS softmax)
 op_library(sequence_softmax_op DEPS softmax)
 op_library(sum_op DEPS selected_rows_functor)
 op_library(sgd_op DEPS selected_rows_functor)
+op_library(print_op DEPS lod_tensor)
 op_library(adagrad_op DEPS selected_rows_functor)
 op_library(conv_op DEPS vol2col)
 op_library(pool_op DEPS pooling)

--- a/paddle/operators/assign_value_op.cc
+++ b/paddle/operators/assign_value_op.cc
+/* Copyright (c) 2018 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. */
+
+#include "paddle/operators/assign_value_op.h"
+
+namespace paddle {
+namespace operators {
+
+class AssignValueOp : public framework::OperatorWithKernel {
+ public:
+  AssignValueOp(const std::string &type,
+                const framework::VariableNameMap &inputs,
+                const framework::VariableNameMap &outputs,
+                const framework::AttributeMap &attrs)
+      : OperatorWithKernel(type, inputs, outputs, attrs) {}
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of AssignValueOp should not be null.");
+    auto shape = ctx->Attrs().Get<std::vector<int>>("shape");
+    ctx->SetOutputDim("Out", framework::make_ddim(shape));
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    return framework::OpKernelType(
+        framework::proto::DataType(ctx.Attr<int>("dtype")), ctx.GetPlace());
+  }
+};
+
+class AssignValueOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  AssignValueOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddOutput("Out", "(Tensor) Output tensor of assign_value operator.");
+    AddAttr<std::vector<int>>("shape",
+                              "(vector<int>) "
+                              "Shape of values.");
+    AddAttr<int>("dtype", "data type of values")
+        .InEnum({framework::proto::DataType::INT32,
+                 framework::proto::DataType::FP32});
+    AddAttr<std::vector<float>>("fp32_values", "store the float values")
+        .SetDefault({});
+    AddAttr<std::vector<int>>("int32_values", "store the int values")
+        .SetDefault({});
+    AddComment(R"DOC(
+AssignValue operator
+
+$$Out = values$$
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OPERATOR(assign_value, ops::AssignValueOp, ops::AssignValueOpMaker);
+REGISTER_OP_CPU_KERNEL(assign_value, ops::AssignValueKernel<int>,
+                       ops::AssignValueKernel<float>);
--- a/paddle/operators/assign_value_op.cu.cc
+++ b/paddle/operators/assign_value_op.cu.cc
+/* 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.
+Indicesou 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. */
+
+#include "paddle/operators/assign_value_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(assign_value, ops::AssignValueKernel<int>,
+                        ops::AssignValueKernel<float>);
--- a/paddle/operators/detail/recv_impl.cc
+++ b/paddle/operators/detail/recv_impl.cc
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+/* Copyright (c) 2018 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.
@@ -12,54 +12,39 @@ 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. */

-#include "send_recv_impl.h"
+#pragma once
+
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/platform/enforce.h"

 namespace paddle {
 namespace operators {
-namespace detail {
-
-Status SendRecvServerImpl::SendVariable(ServerContext *context,
-                                        const VariableMessage *in_var,
-                                        VoidMessage *out_var) {
-  MessageWithName msg_with_name =
-      std::make_pair(in_var->varname(), std::move(*in_var));
-  var_recv_queue_.Push(std::move(msg_with_name));
-  return Status::OK;
-}
-
-Status SendRecvServerImpl::GetVariable(ServerContext *context,
-                                       const VariableMessage *in_var,
-                                       VariableMessage *out_var) {
-  std::string get_var_name = in_var->varname();
-  auto *var = scope_->FindVar(get_var_name);
-
-  SerializeToMessage(get_var_name, var, platform::CPUDeviceContext(), out_var);
-  return Status::OK;
-}
-
-Status SendRecvServerImpl::Wait(ServerContext *context,
-                                const VoidMessage *in_var,
-                                VoidMessage *out_var) {
-  {
-    std::unique_lock<std::mutex> lock(this->mutex_);
-    condition_.wait(lock, [=] { return this->done_ == true; });
-  }
-  return Status::OK;
-}
-
-void SendRecvServerImpl::Reset() {
-  std::lock_guard<std::mutex> lock(this->mutex_);
-  done_ = false;
-}

-void SendRecvServerImpl::Done() {
-  {
-    std::lock_guard<std::mutex> lock(this->mutex_);
-    done_ = true;
+template <typename T>
+class AssignValueKernel : public framework::OpKernel<T> {
+ public:
+  virtual void Compute(const framework::ExecutionContext& ctx) const {
+    auto shape = ctx.Attr<std::vector<int>>("shape");
+    auto* out = ctx.Output<framework::Tensor>("Out");
+    int dtype = ctx.Attr<int>("dtype");
+    const char* value_name = nullptr;
+    switch (dtype) {
+      case framework::proto::DataType::INT32:
+        value_name = "int32_values";
+        break;
+      case framework::proto::DataType::FP32:
+        value_name = "fp32_values";
+        break;
+      default:
+        PADDLE_THROW("Unsupported dtype for assign_value_op: %d", dtype);
+        break;
+    }
+    auto values = ctx.Attr<std::vector<T>>(value_name);
+    framework::CopyFromVector(values, ctx.device_context(), out);
+    out->Resize(framework::make_ddim(shape));
  }
-  condition_.notify_all();
-}
+};

-}  // namespace detail
 }  // namespace operators
 }  // namespace paddle
--- a/paddle/operators/detail/CMakeLists.txt
+++ b/paddle/operators/detail/CMakeLists.txt
-grpc_library(sendrecvop_grpc SRCS recv_impl.cc send_impl.cc PROTO send_recv.proto DEPS lod_tensor selected_rows)
+grpc_library(sendrecvop_grpc SRCS sendrecvop_utils.cc grpc_client.cc grpc_server.cc PROTO send_recv.proto DEPS lod_tensor selected_rows)
--- a/paddle/operators/detail/grpc_client.cc
+++ b/paddle/operators/detail/grpc_client.cc
+/* 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. */
+
+#include "grpc_client.h"
+namespace paddle {
+namespace operators {
+namespace detail {
+
+bool RPCClient::AsyncSendVariable(const std::string& ep,
+                                  const platform::DeviceContext& ctx,
+                                  const framework::Scope& scope,
+                                  const std::string& var_name,
+                                  int64_t time_out) {
+  sendrecv::VariableMessage req;
+  auto* var = scope.FindVar(var_name);
+  SerializeToMessage(var_name, var, ctx, &req);
+
+  // varhandle
+  VarHandle var_h;
+  var_h.ep = ep;
+  var_h.scope = &scope;
+  var_h.name = var_name;
+  var_h.ctx = &ctx;
+
+  // stub context
+  auto ch = GetChannel(ep);
+  SendProcessor* s = new SendProcessor(ch);
+  s->Prepare(var_h, time_out);
+  s->response_call_back_ = NULL;
+
+  auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_);
+  rpc->Finish(&s->reply_, &s->status_, (void*)s);
+
+  req_count_++;
+
+  return true;
+}
+
+void ProcGetResponse(const VarHandle& var_h,
+                     const sendrecv::VariableMessage& ret_msg) {
+  auto* outvar = var_h.scope->FindVar(var_h.name);
+
+  std::istringstream iss(ret_msg.serialized());
+  DeserializeFromMessage(ret_msg, *var_h.ctx, outvar);
+}
+
+bool RPCClient::AsyncGetVariable(const std::string& ep,
+                                 const platform::DeviceContext& ctx,
+                                 const framework::Scope& scope,
+                                 const std::string& var_name,
+                                 int64_t time_out) {
+  sendrecv::VariableMessage req;
+  req.set_varname(var_name);
+
+  auto* var = scope.FindVar(var_name);
+  SerializeToMessage(var_name, var, ctx, &req);
+
+  // varhandle
+  VarHandle var_h;
+  var_h.ep = ep;
+  var_h.scope = &scope;
+  var_h.name = var_name;
+  var_h.ctx = &ctx;
+
+  // stub context
+  auto ch = GetChannel(ep);
+  GetProcessor* s = new GetProcessor(ch);
+  s->Prepare(var_h, time_out);
+  s->response_call_back_ = ProcGetResponse;
+
+  auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_);
+  rpc->Finish(&s->reply_, &s->status_, (void*)s);
+
+  req_count_++;
+
+  return true;
+}
+
+bool RPCClient::wait() {
+  bool ok = true;
+
+  while (true) {
+    if (req_count_ <= 0) {
+      break;
+    }
+
+    if (!Proceed()) {
+      LOG(ERROR) << "Get meets CompletionQueue error";
+      return false;
+    }
+  }
+
+  return ok;
+}
+
+bool RPCClient::Proceed() {
+  void* tag = NULL;
+  bool ok = false;
+
+  // request counts.
+  if (!cq_.Next(&tag, &ok)) {
+    return false;
+  }
+  req_count_--;
+
+  GPR_ASSERT(ok);
+  PADDLE_ENFORCE(tag);
+
+  // TODO(gongwb): add more retries.
+  ClientBase* c = static_cast<ClientBase*>(tag);
+  if (!c->status_.ok()) {
+    delete c;
+    return true;
+  }
+
+  c->Process();
+  delete c;
+  return true;
+}
+
+std::shared_ptr<grpc::Channel> RPCClient::GetChannel(const std::string& ep) {
+  auto it = channels_.find(ep);
+  if (it != channels_.end()) {
+    return it->second;
+  }
+
+  auto ch = std::shared_ptr<grpc::Channel>(
+      grpc::CreateChannel(ep, grpc::InsecureChannelCredentials()));
+
+  channels_[ep] = ch;
+  return ch;
+}
+
+}  // namespace detail
+}  // namespace operators
+}  // namespace paddle
--- a/paddle/operators/detail/grpc_client.h
+++ b/paddle/operators/detail/grpc_client.h
+/* 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 <grpc++/grpc++.h>
+#include <grpc/support/log.h>
+#include <time.h>
+#include <chrono>
+#include <ctime>
+#include <functional>
+#include <iostream>
+#include <map>
+#include <string>
+#include <vector>
+
+#include "paddle/framework/data_type.h"
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/framework/scope.h"
+#include "paddle/framework/selected_rows.h"
+#include "paddle/operators/detail/sendrecvop_utils.h"
+#include "paddle/operators/detail/simple_block_queue.h"
+
+namespace paddle {
+namespace operators {
+namespace detail {
+
+struct VarHandle {
+  std::string ep;
+  const platform::DeviceContext* ctx;
+  const framework::Scope* scope;
+  std::string name;
+
+  std::string String() const {
+    std::ostringstream s;
+    s << "name:[" << name << "] ep:[" << ep << "]";
+    return s.str();
+  }
+};
+
+void ProcGetResponse(const VarHandle& var_h,
+                     const sendrecv::VariableMessage& msg);
+
+class ClientBase {
+ public:
+  explicit ClientBase(std::shared_ptr<grpc::Channel> ch) {
+    stub_ = sendrecv::SendRecvService::NewStub(ch);
+    context_ = NULL;
+  }
+
+  virtual ~ClientBase() {}
+
+  virtual void Prepare(const VarHandle& var_info, int64_t time_out) {
+    context_.reset(new grpc::ClientContext());
+    var_h_ = var_info;
+
+    std::chrono::system_clock::time_point deadline =
+        std::chrono::system_clock::now() + std::chrono::milliseconds(time_out);
+
+    context_->set_deadline(deadline);
+  }
+
+  virtual void Process() = 0;
+
+  std::unique_ptr<sendrecv::SendRecvService::Stub> stub_;
+  std::unique_ptr<grpc::ClientContext> context_;
+  grpc::Status status_;
+  VarHandle var_h_;
+};
+
+typedef std::function<void(const VarHandle&, const sendrecv::VoidMessage&)>
+    RequestSendCallBack;
+
+class SendProcessor : public ClientBase {
+ public:
+  explicit SendProcessor(std::shared_ptr<grpc::Channel> ch) : ClientBase(ch) {}
+
+  virtual ~SendProcessor() {}
+
+  virtual void Process() {
+    if (response_call_back_) {
+      response_call_back_(var_h_, reply_);
+    }
+  }
+
+  sendrecv::VoidMessage reply_;
+  RequestSendCallBack response_call_back_ = NULL;
+};
+
+typedef std::function<void(const VarHandle&, const sendrecv::VariableMessage&)>
+    RequestGetCallBack;
+
+class GetProcessor : public ClientBase {
+ public:
+  explicit GetProcessor(std::shared_ptr<grpc::Channel> ch) : ClientBase(ch) {}
+
+  virtual ~GetProcessor() {}
+
+  virtual void Process() {
+    if (response_call_back_) {
+      response_call_back_(var_h_, reply_);
+    }
+  }
+
+  sendrecv::VariableMessage reply_;
+  RequestGetCallBack response_call_back_ = ProcGetResponse;
+};
+
+class RPCClient {
+ public:
+  bool AsyncSendVariable(const std::string& ep,
+                         const platform::DeviceContext& ctx,
+                         const framework::Scope& scope,
+                         const std::string& var_name,
+                         int64_t time_out = 600 * 1000);
+
+  bool AsyncGetVariable(const std::string& ep,
+                        const platform::DeviceContext& ctx,
+                        const framework::Scope& scope,
+                        const std::string& var_name,
+                        int64_t time_out = 600 * 1000);
+  bool wait();
+
+ private:
+  bool Proceed();
+  std::shared_ptr<grpc::Channel> GetChannel(const std::string& ep);
+
+ private:
+  grpc::CompletionQueue cq_;
+  std::map<std::string, std::shared_ptr<grpc::Channel>> channels_;
+  int64_t req_count_ = 0;
+};
+
+}  // namespace detail
+}  // namespace operators
+}  // namespace paddle
--- a/paddle/operators/detail/grpc_server.cc
+++ b/paddle/operators/detail/grpc_server.cc
+/* 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. */
+
+#include "paddle/operators/detail/grpc_server.h"
+
+using grpc::ServerAsyncResponseWriter;
+
+namespace paddle {
+namespace operators {
+namespace detail {
+
+enum CallStatus { PROCESS = 0, FINISH };
+
+// reference:
+// https://stackoverflow.com/questions/41732884/grpc-multiple-services-in-cpp-async-server
+class RequestBase {
+ public:
+  explicit RequestBase(sendrecv::SendRecvService::AsyncService* service,
+                       grpc::ServerCompletionQueue* cq)
+      : service_(service), cq_(cq), status_(PROCESS) {}
+  virtual ~RequestBase() {}
+  virtual void Process() { assert(false); }
+
+  CallStatus Status() { return status_; }
+  void SetStatus(CallStatus status) { status_ = status; }
+
+ protected:
+  grpc::ServerContext ctx_;
+  sendrecv::SendRecvService::AsyncService* service_;
+  grpc::ServerCompletionQueue* cq_;
+  CallStatus status_;
+};
+
+typedef std::pair<std::string, sendrecv::VariableMessage> MessageWithName;
+
+class RequestSend final : public RequestBase {
+ public:
+  explicit RequestSend(sendrecv::SendRecvService::AsyncService* service,
+                       grpc::ServerCompletionQueue* cq,
+                       SimpleBlockQueue<MessageWithName>* queue)
+      : RequestBase(service, cq), queue_(queue), responder_(&ctx_) {
+    service_->RequestSendVariable(&ctx_, &request_, &responder_, cq_, cq_,
+                                  this);
+  }
+
+  virtual ~RequestSend() {}
+
+  virtual void Process() {
+    MessageWithName msg_with_name =
+        std::make_pair(request_.varname(), std::move(request_));
+    queue_->Push(std::move(msg_with_name));
+    // TODO(gongwb): check var's info.
+    responder_.Finish(reply_, grpc::Status::OK, this);
+  }
+
+ protected:
+  sendrecv::VariableMessage request_;
+  sendrecv::VoidMessage reply_;
+  SimpleBlockQueue<MessageWithName>* queue_;
+  ServerAsyncResponseWriter<sendrecv::VoidMessage> responder_;
+};
+
+class RequestGet final : public RequestBase {
+ public:
+  explicit RequestGet(sendrecv::SendRecvService::AsyncService* service,
+                      grpc::ServerCompletionQueue* cq, framework::Scope* scope)
+      : RequestBase(service, cq), responder_(&ctx_), scope_(scope) {
+    service_->RequestGetVariable(&ctx_, &request_, &responder_, cq_, cq_, this);
+  }
+
+  virtual ~RequestGet() {}
+
+  virtual void Process() {
+    // proc request.
+    std::string var_name = request_.varname();
+    auto* var = scope_->FindVar(var_name);
+    SerializeToMessage(var_name, var, platform::CPUDeviceContext(), &reply_);
+    // TODO(gongwb): check var's info.
+    responder_.Finish(reply_, grpc::Status::OK, this);
+  }
+
+ protected:
+  sendrecv::VariableMessage request_;
+  sendrecv::VariableMessage reply_;
+  ServerAsyncResponseWriter<sendrecv::VariableMessage> responder_;
+  framework::Scope* scope_;
+};
+
+void AsyncGRPCServer::RunSyncUpdate() {
+  grpc::ServerBuilder builder;
+  builder.AddListeningPort(address_, grpc::InsecureServerCredentials());
+  builder.RegisterService(&service_);
+
+  cq_send_ = builder.AddCompletionQueue();
+  cq_get_ = builder.AddCompletionQueue();
+  server_ = builder.BuildAndStart();
+  LOG(INFO) << "Server listening on " << address_ << std::endl;
+
+  std::function<void()> send_register =
+      std::bind(&AsyncGRPCServer::TryToRegisterNewSendOne, this);
+  std::function<void()> get_register =
+      std::bind(&AsyncGRPCServer::TryToRegisterNewGetOne, this);
+
+  t_send_.reset(
+      new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this, false,
+                                cq_send_.get(), "cq_send", send_register)));
+
+  t_get_.reset(
+      new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this, true,
+                                cq_get_.get(), "cq_get", get_register)));
+
+  // wait server
+  server_->Wait();
+  t_send_->join();
+  t_get_->join();
+}
+
+void AsyncGRPCServer::ShutdownQueue() {
+  std::unique_lock<std::mutex> lock(cq_mutex_);
+  cq_send_->Shutdown();
+  cq_get_->Shutdown();
+  is_shut_down_ = true;
+}
+
+// This URL explains why shutdown is complicate:
+// https://stackoverflow.com/questions/35708348/grpc-what-is-the-recommended-way-to-shut-down-an-asynchronous-server-in-c
+void AsyncGRPCServer::ShutDown() {
+  server_->Shutdown();
+  ShutdownQueue();
+}
+
+void AsyncGRPCServer::TryToRegisterNewSendOne() {
+  std::unique_lock<std::mutex> lock(cq_mutex_);
+  if (is_shut_down_) {
+    return;
+  }
+  RequestSend* send =
+      new RequestSend(&service_, cq_send_.get(), &var_recv_queue_);
+  VLOG(4) << "create RequestSend status:" << send->Status();
+}
+
+void AsyncGRPCServer::TryToRegisterNewGetOne() {
+  std::unique_lock<std::mutex> lock(cq_mutex_);
+  if (is_shut_down_) {
+    return;
+  }
+  RequestGet* get = new RequestGet(&service_, cq_get_.get(), scope_);
+  VLOG(4) << "create Requestget status:" << get->Status();
+}
+
+void AsyncGRPCServer::SetFinishOrDelete(RequestBase*& last) {
+  std::unique_lock<std::mutex> lock(cq_mutex_);
+  if (is_shut_down_) {
+    delete last;
+    last = NULL;
+    return;
+  }
+
+  last->SetStatus(FINISH);
+  return;
+}
+
+void AsyncGRPCServer::HandleRequest(bool wait, grpc::ServerCompletionQueue* cq,
+                                    std::string cq_name,
+                                    std::function<void()> TryToRegisterNewOne) {
+  TryToRegisterNewOne();
+
+  void* tag = NULL;
+  bool ok = false;
+  while (true) {
+    if (!cq->Next(&tag, &ok)) {
+      LOG(INFO) << cq_name << " get CompletionQueue shutdown!";
+      break;
+    }
+
+    if (wait && !done_) {
+      Wait();
+    }
+
+    RequestBase* base = (RequestBase*)tag;
+    if (!ok) {
+      VLOG(4) << cq_name << " recv no regular event";
+      TryToRegisterNewOne();
+      delete base;
+      continue;
+    }
+
+    switch (base->Status()) {
+      case PROCESS: {
+        VLOG(4) << cq_name << " status:" << base->Status();
+        TryToRegisterNewOne();
+        base->Process();
+        SetFinishOrDelete(base);
+        break;
+      }
+      case FINISH: {
+        VLOG(4) << cq_name << " status:" << base->Status();
+        delete base;
+        break;
+      }
+      default: { assert(false); }
+    }
+  }
+}
+
+void AsyncGRPCServer::Wait() {
+  std::unique_lock<std::mutex> lock(this->mutex_);
+  condition_.wait(lock, [=] { return this->done_ == true; });
+}
+
+void AsyncGRPCServer::Reset() {
+  std::lock_guard<std::mutex> lock(this->mutex_);
+  done_ = false;
+}
+
+void AsyncGRPCServer::Done() {
+  {
+    std::lock_guard<std::mutex> lock(this->mutex_);
+    done_ = true;
+  }
+  condition_.notify_all();
+}
+
+}  // namespace detail
+}  // namespace operators
+}  // namespace paddle
--- a/paddle/operators/detail/grpc_server.h
+++ b/paddle/operators/detail/grpc_server.h
+/* 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/lod_tensor.h"
+#include "paddle/framework/scope.h"
+#include "paddle/framework/selected_rows.h"
+#include "paddle/framework/var_type.h"
+#include "paddle/operators/detail/simple_block_queue.h"
+
+#include "paddle/operators/detail/send_recv.grpc.pb.h"
+#include "paddle/operators/detail/send_recv.pb.h"
+
+#include <grpc++/grpc++.h>
+#include <grpc/support/log.h>
+#include <thread>
+#include "paddle/operators/detail/sendrecvop_utils.h"
+
+namespace paddle {
+namespace operators {
+namespace detail {
+
+typedef std::pair<std::string, sendrecv::VariableMessage> MessageWithName;
+class RequestBase;
+
+class AsyncGRPCServer final : public sendrecv::SendRecvService::Service {
+ public:
+  explicit AsyncGRPCServer(std::string address) { address_ = address; }
+
+  void RunSyncUpdate();
+
+  void Reset();
+
+  void Done();
+
+  void SetScope(framework::Scope *scope) { scope_ = scope; }
+
+  const MessageWithName Get() { return this->var_recv_queue_.Pop(); }
+
+  void Push(const MessageWithName &msg) { this->var_recv_queue_.Push(msg); }
+
+  void ShutDown();
+
+ protected:
+  void Wait();
+  void HandleRequest(bool wait, grpc::ServerCompletionQueue *cq,
+                     std::string cq_name,
+                     std::function<void()> TryToRegisterNewOne);
+  void TryToRegisterNewSendOne();
+  void TryToRegisterNewGetOne();
+  void SetFinishOrDelete(RequestBase *&last);
+  void ShutdownQueue();
+
+ private:
+  std::mutex cq_mutex_;
+  volatile bool is_shut_down_ = false;
+  std::unique_ptr<grpc::ServerCompletionQueue> cq_send_;
+  std::unique_ptr<grpc::ServerCompletionQueue> cq_get_;
+
+  sendrecv::SendRecvService::AsyncService service_;
+  std::unique_ptr<grpc::Server> server_;
+
+  std::string address_;
+  framework::Scope *scope_;
+  // received variable from RPC, operators fetch variable from this queue.
+  SimpleBlockQueue<MessageWithName> var_recv_queue_;
+
+  // condition of the sub program
+  std::mutex mutex_;
+  volatile mutable bool done_;
+  std::condition_variable condition_;
+
+  std::unique_ptr<std::thread> t_send_;
+  std::unique_ptr<std::thread> t_get_;
+};
+
+};  // namespace detail
+};  // namespace operators
+};  // namespace paddle
--- a/paddle/operators/detail/send_recv.proto
+++ b/paddle/operators/detail/send_recv.proto
@@ -21,8 +21,6 @@ service SendRecvService {
  rpc SendVariable(VariableMessage) returns (VoidMessage) {}
  // Argument VariableMessage for GetVariable should only contain varname.
  rpc GetVariable(VariableMessage) returns (VariableMessage) {}
-  // wait for one execution of the program
-  rpc Wait(VoidMessage) returns (VoidMessage) {}
 }

 // VariableMessage is serialized paddle variable message.

--- a/paddle/operators/detail/send_recv_impl.h
+++ b/paddle/operators/detail/send_recv_impl.h
@@ -12,87 +12,15 @@ 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/lod_tensor.h"
-#include "paddle/framework/scope.h"
-#include "paddle/framework/selected_rows.h"
-#include "paddle/framework/var_type.h"
-#include "paddle/operators/detail/simple_block_queue.h"
-
-#include "paddle/operators/detail/send_recv.grpc.pb.h"
-#include "paddle/operators/detail/send_recv.pb.h"
-
-#include <grpc++/grpc++.h>
-
-using grpc::Channel;
-using grpc::Server;
-using grpc::ServerContext;
-using grpc::ServerReader;
-using grpc::ServerBuilder;
-
-using grpc::ClientContext;
-using grpc::ClientReader;
-using grpc::ClientReaderWriter;
-using grpc::ClientWriter;
-using grpc::Status;
-using sendrecv::SendRecvService;
-using sendrecv::VariableMessage;
-using sendrecv::VoidMessage;
+#include "paddle/operators/detail/sendrecvop_utils.h"

 namespace paddle {
 namespace operators {
 namespace detail {

-typedef std::pair<std::string, sendrecv::VariableMessage> MessageWithName;
-
-class SendRecvServerImpl final : public SendRecvService::Service {
- public:
-  explicit SendRecvServerImpl() {}
-
-  Status SendVariable(ServerContext *context, const VariableMessage *in_var,
-                      VoidMessage *out_var) override;
-  Status GetVariable(ServerContext *context, const VariableMessage *in_var,
-                     VariableMessage *out_var) override;
-  Status Wait(ServerContext *context, const VoidMessage *in_var,
-              VoidMessage *out_var) override;
-  void Reset();
-  void Done();
-  void SetScope(framework::Scope *scope) { scope_ = scope; };
-
-  const MessageWithName Get() { return this->var_recv_queue_.Pop(); }
-
-  void Push(const MessageWithName &msg) { this->var_recv_queue_.Push(msg); }
-
- private:
-  // received variable from RPC, operators fetch variable from this queue.
-  SimpleBlockQueue<MessageWithName> var_recv_queue_;
-  framework::Scope *scope_;
-  // condition of the sub program
-  std::mutex mutex_;
-  bool done_;
-  std::condition_variable condition_;
-};
-
-// RPCClient is a class to send tensors to pserver sub-network
-// using different hashing methods.
-class RPCClient {
- public:
-  RPCClient(std::shared_ptr<Channel> channel)
-      : stub_(SendRecvService::NewStub(channel)) {}
-
-  bool SendVariable(const framework::Scope &scope, const std::string &inname);
-  bool GetVariable(const framework::Scope &scope, const std::string &outname);
-  void Wait();
-
- private:
-  std::unique_ptr<SendRecvService::Stub> stub_;
-};
-
-inline void SerializeToMessage(const std::string &name,
-                               const framework::Variable *var,
-                               const platform::DeviceContext &ctx,
-                               VariableMessage *msg) {
+void SerializeToMessage(const std::string& name, const framework::Variable* var,
+                        const platform::DeviceContext& ctx,
+                        sendrecv::VariableMessage* msg) {
  msg->set_varname(name);
  std::ostringstream oss;
  switch (framework::ToVarType(var->Type())) {
@@ -114,10 +42,9 @@ inline void SerializeToMessage(const std::string &name,
  msg->set_serialized(oss.str());
 }

-inline void DeserializeFromMessage(const VariableMessage &msg,
-                                   const platform::DeviceContext &ctx,
-                                   framework::Variable *var) {
-  using namespace paddle::framework::proto;
+void DeserializeFromMessage(const sendrecv::VariableMessage& msg,
+                            const platform::DeviceContext& ctx,
+                            framework::Variable* var) {
  std::istringstream iss(msg.serialized());
  switch (msg.type()) {
    case sendrecv::VarType::LOD_TENSOR:

--- a/paddle/operators/detail/send_impl.cc
+++ b/paddle/operators/detail/send_impl.cc
@@ -12,56 +12,31 @@ 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. */

-#include "send_recv_impl.h"
+#pragma once
+#include <iostream>
+#include <string>
+#include <vector>
+
+#include "paddle/framework/data_type.h"
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/framework/scope.h"
+#include "paddle/framework/selected_rows.h"
+#include "paddle/framework/var_type.h"
+
+#include "paddle/operators/detail/send_recv.grpc.pb.h"
+#include "paddle/operators/detail/send_recv.pb.h"

 namespace paddle {
 namespace operators {
 namespace detail {

-bool RPCClient::SendVariable(const framework::Scope& scope,
-                             const std::string& inname) {
-  ClientContext context;
-  VariableMessage msg;
-  VoidMessage out_msg;
-  // FIXME(typhoonzero): pass device context to here.
-  auto ctx = platform::CPUDeviceContext();
-  auto* var = scope.FindVar(inname);
-  PADDLE_ENFORCE(var);
-  SerializeToMessage(inname, var, ctx, &msg);
-
-  Status status = stub_->SendVariable(&context, msg, &out_msg);
-  if (!status.ok()) {
-    LOG(ERROR) << "gRPC error: " << status.error_message();
-    return false;
-  }
-  return true;
-}
-
-bool RPCClient::GetVariable(const framework::Scope& scope,
-                            const std::string& outname) {
-  ClientContext context;
-  VariableMessage call_msg, ret_msg;
-  call_msg.set_varname(outname);
-  auto ctx = platform::CPUDeviceContext();
-  Status status = stub_->GetVariable(&context, call_msg, &ret_msg);
-  auto* outvar = scope.FindVar(outname);
-  if (!status.ok()) {
-    LOG(ERROR) << "gRPC error: " << status.error_message();
-    return false;
-  }
-
-  std::istringstream iss(ret_msg.serialized());
-  DeserializeFromMessage(ret_msg, ctx, outvar);
-
-  return true;
-}
-
-void RPCClient::Wait() {
-  ClientContext context;
-  VoidMessage call_msg, ret_msg;
-  stub_->Wait(&context, call_msg, &ret_msg);
-}
+void SerializeToMessage(const std::string& name, const framework::Variable* var,
+                        const platform::DeviceContext& ctx,
+                        sendrecv::VariableMessage* msg);

+void DeserializeFromMessage(const sendrecv::VariableMessage& msg,
+                            const platform::DeviceContext& ctx,
+                            framework::Variable* var);
 }  // namespace detail
 }  // namespace operators
 }  // namespace paddle
--- a/paddle/operators/elementwise_add_op.cc
+++ b/paddle/operators/elementwise_add_op.cc
@@ -21,7 +21,7 @@ class ElementwiseAddOpMaker : public ElementwiseOpMaker {
 public:
  ElementwiseAddOpMaker(OpProto* proto, OpAttrChecker* op_checker)
      : ElementwiseOpMaker(proto, op_checker) {
-    SetComment("Add", "$Out = X + Y$");
+    SetComment("Add", "Out = X + Y");
    AddComment(comment_);
  }
 };

--- a/paddle/operators/elementwise_div_op.cc
+++ b/paddle/operators/elementwise_div_op.cc
@@ -21,7 +21,7 @@ class ElementwiseDivOpMaker : public ElementwiseOpMaker {
 public:
  ElementwiseDivOpMaker(OpProto* proto, OpAttrChecker* op_checker)
      : ElementwiseOpMaker(proto, op_checker) {
-    SetComment("Div", "$Out = X / Y$");
+    SetComment("Div", "Out = X / Y");
    AddComment(comment_);
  }
 };

--- a/paddle/operators/elementwise_mul_op.cc
+++ b/paddle/operators/elementwise_mul_op.cc
@@ -22,7 +22,7 @@ class ElementwiseMulOpMaker : public ElementwiseOpMaker {
 public:
  ElementwiseMulOpMaker(OpProto* proto, OpAttrChecker* op_checker)
      : ElementwiseOpMaker(proto, op_checker) {
-    SetComment("Mul", "$Out = X \\odot\\ Y$");
+    SetComment("Mul", "Out = X \\odot\\ Y");
    AddComment(comment_);
  }
 };

--- a/paddle/operators/elementwise_op.h
+++ b/paddle/operators/elementwise_op.h
@@ -58,7 +58,8 @@ Limited Elementwise {name} Operator.

 The equation is:

-{equation}
+.. math::
+  {equation}

 X is a tensor of any dimension and the dimensions of tensor Y must be smaller than
 or equal to the dimensions of X. 
@@ -71,15 +72,16 @@ For case 2:
 Y will be broadcasted to match the shape of X and axis should be 
 the starting dimension index for broadcasting Y onto X.

-example:
-  shape(X) = (2, 3, 4, 5), shape(Y) = (,)
-  shape(X) = (2, 3, 4, 5), shape(Y) = (5,)
-  shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)
-  shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
-  shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0
+For example
+  .. code-block:: python

-Both the input X and Y can carry the LoD (Level of Details) information,
-or not. But the output only shares the LoD information with input X.
+    shape(X) = (2, 3, 4, 5), shape(Y) = (,)
+    shape(X) = (2, 3, 4, 5), shape(Y) = (5,)
+    shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)
+    shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
+    shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0
+
+Either of the inputs X and Y or none can carry the LoD (Level of Details) information. However, the output only shares the LoD information with input X.

 )DOC";
    AddComment(comment_);

--- a/paddle/operators/elementwise_sub_op.cc
+++ b/paddle/operators/elementwise_sub_op.cc
@@ -21,7 +21,7 @@ class ElementwiseSubOpMaker : public ElementwiseOpMaker {
 public:
  ElementwiseSubOpMaker(OpProto* proto, OpAttrChecker* op_checker)
      : ElementwiseOpMaker(proto, op_checker) {
-    SetComment("Sub", "$Out = X - Y$");
+    SetComment("Sub", "Out = X - Y");
    AddComment(comment_);
  }
 };

--- a/paddle/operators/print_op.cc
+++ b/paddle/operators/print_op.cc
+/* 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. */
+
+#include <algorithm>
+#include <ctime>
+
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+#define CLOG std::cout
+
+struct Formater {
+  std::string message;
+  std::string name;
+  std::vector<int> dims;
+  std::type_index dtype{typeid(char)};
+  framework::LoD lod;
+  int summarize;
+  void* data{nullptr};
+
+  void operator()(size_t size) {
+    PrintMessage();
+    PrintName();
+    PrintDims();
+    PrintDtype();
+    PrintLod();
+    PrintData(size);
+  }
+
+ private:
+  void PrintMessage() { CLOG << std::time(nullptr) << "\t" << message; }
+  void PrintName() {
+    if (!name.empty()) {
+      CLOG << "Tensor[" << name << "]" << std::endl;
+    }
+  }
+  void PrintDims() {
+    if (!dims.empty()) {
+      CLOG << "\tshape: [";
+      for (auto i : dims) {
+        CLOG << i << ",";
+      }
+      CLOG << "]" << std::endl;
+    }
+  }
+  void PrintDtype() {
+    if (dtype.hash_code() != typeid(char).hash_code()) {
+      CLOG << "\tdtype: " << dtype.name() << std::endl;
+    }
+  }
+  void PrintLod() {
+    if (!lod.empty()) {
+      CLOG << "\tLoD: [";
+      for (auto level : lod) {
+        CLOG << "[ ";
+        for (auto i : level) {
+          CLOG << i << ",";
+        }
+        CLOG << " ]";
+      }
+      CLOG << "]" << std::endl;
+    }
+  }
+
+  void PrintData(size_t size) {
+    PADDLE_ENFORCE_NOT_NULL(data);
+    // print float
+    if (dtype.hash_code() == typeid(float).hash_code()) {
+      Display<float>(size);
+    }
+    if (dtype.hash_code() == typeid(double).hash_code()) {
+      Display<double>(size);
+    }
+    if (dtype.hash_code() == typeid(int).hash_code()) {
+      Display<int>(size);
+    }
+    if (dtype.hash_code() == typeid(int64_t).hash_code()) {
+      Display<int64_t>(size);
+    }
+  }
+
+  template <typename T>
+  void Display(size_t size) {
+    auto* d = (T*)data;
+    CLOG << "\tdata: ";
+    if (summarize != -1) {
+      summarize = std::min(size, (size_t)summarize);
+      for (int i = 0; i < summarize; i++) {
+        CLOG << d[i] << ",";
+      }
+    } else {
+      for (size_t i = 0; i < size; i++) {
+        CLOG << d[i] << ",";
+      }
+    }
+    CLOG << std::endl;
+  }
+};
+
+// TODO(ChunweiYan) there should be some other printers for TensorArray
+class TensorPrintOp : public framework::OperatorBase {
+ public:
+  TensorPrintOp(const std::string& type,
+                const framework::VariableNameMap& inputs,
+                const framework::VariableNameMap& outputs,
+                const framework::AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+
+  TensorPrintOp(const TensorPrintOp& o)
+      : framework::OperatorBase(
+            static_cast<const framework::OperatorBase&>(o)) {
+    PADDLE_THROW("Not implemented");
+  }
+
+  void Run(const framework::Scope& scope,
+           const platform::Place& place) const override {
+    // Only run the `first_n` times.
+    int first_n = Attr<int>("first_n");
+    if (first_n > 0 && ++times_ > first_n) return;
+
+    PADDLE_ENFORCE(!Inputs("input").empty(), "input should be set");
+    auto* input_var = scope.FindVar(Input("input"));
+    PADDLE_ENFORCE_NOT_NULL(input_var);
+    auto& tensor = input_var->Get<framework::LoDTensor>();
+
+    // TODO(ChunweiYan) support GPU
+    PADDLE_ENFORCE(platform::is_cpu_place(tensor.place()));
+
+    Formater formater;
+    if (Attr<bool>("print_tensor_name")) {
+      formater.name = Inputs("input").front();
+    }
+    if (Attr<bool>("print_tensor_type")) {
+      formater.dtype = tensor.type();
+    }
+    if (Attr<bool>("print_tensor_shape")) {
+      formater.dims.assign(tensor.dims()[0],
+                           tensor.dims()[tensor.dims().size() - 1]);
+    }
+    if (Attr<bool>("print_tensor_lod")) {
+      formater.lod = tensor.lod();
+    }
+    formater.summarize = Attr<int>("summarize");
+    formater.data = (void*)tensor.data<void>();
+    formater(tensor.numel());
+  }
+
+ private:
+  mutable int times_{0};
+};
+
+class PrintOpProtoAndCheckMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  PrintOpProtoAndCheckMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("input", "the tensor that will be displayed.");
+    AddAttr<int>("first_n", "Only log `first_n` number of times.");
+    AddAttr<std::string>("message", "A string message to print as a prefix.");
+    AddAttr<int>("summarize", "Print this number of elements in the tensor.");
+    AddAttr<bool>("print_tensor_name", "Whether to print the tensor name.");
+    AddAttr<bool>("print_tensor_type", "Whether to print the tensor's dtype.");
+    AddAttr<bool>("print_tensor_shape", "Whether to print the tensor's shape.");
+    AddAttr<bool>("print_tensor_lod", "Whether to print the tensor's lod.");
+    AddComment(R"DOC(
+    Creates a print op that will print when a tensor is accessed.
+
+    Wraps the tensor passed in so that whenever that a tensor is accessed,
+    the message `message` is printed, along with the current value of the
+    tensor `t`.)DOC");
+  }
+};
+
+class InferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext* context) const override {
+    PADDLE_ENFORCE(context->HasInput("input"), "input should be set");
+  }
+};
+
+class InferVarType : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc& op_desc,
+                  framework::BlockDesc* block) const override {}
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OPERATOR(print, paddle::operators::TensorPrintOp,
+                  paddle::operators::PrintOpProtoAndCheckMaker,
+                  paddle::operators::InferShape,
+                  paddle::operators::InferVarType,
+                  paddle::framework::EmptyGradOpMaker);
--- a/paddle/operators/recv_op.cc
+++ b/paddle/operators/recv_op.cc
@@ -24,7 +24,8 @@ limitations under the License. */
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/proto_desc.h"
-#include "paddle/operators/detail/send_recv_impl.h"
+#include "paddle/operators/detail/grpc_server.h"
+#include "paddle/operators/detail/sendrecvop_utils.h"
 #include "paddle/operators/detail/simple_block_queue.h"

 #define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV"
@@ -32,6 +33,11 @@ limitations under the License. */
 namespace paddle {
 namespace operators {

+void RunServer(std::shared_ptr<detail::AsyncGRPCServer> service) {
+  service->RunSyncUpdate();
+  VLOG(4) << "RunServer thread end";
+}
+
 static void CreateTensorFromMessageType(framework::Variable *var,
                                        sendrecv::VarType var_type) {
  if (var_type == sendrecv::VarType::LOD_TENSOR) {
@@ -46,18 +52,6 @@ static void CreateTensorFromMessageType(framework::Variable *var,
  }
 }

-void RunServer(Server **rpc_server,
-               std::shared_ptr<detail::SendRecvServerImpl> service,
-               const std::string &server_address) {
-  ServerBuilder builder;
-  builder.AddListeningPort(server_address, grpc::InsecureServerCredentials());
-  builder.RegisterService(service.get());
-  std::unique_ptr<Server> server(builder.BuildAndStart());
-  *rpc_server = server.get();
-  LOG(INFO) << "Server listening on " << server_address;
-  server->Wait();
-}
-
 class RecvOp : public framework::OperatorBase {
 public:
  RecvOp(const std::string &type, const framework::VariableNameMap &inputs,
@@ -65,10 +59,9 @@ class RecvOp : public framework::OperatorBase {
         const framework::AttributeMap &attrs)
      : OperatorBase(type, inputs, outputs, attrs) {
    if (!rpc_service_) {
-      rpc_service_.reset(new detail::SendRecvServerImpl());
      std::string endpoint = Attr<std::string>("endpoint");
-      server_thread_.reset(
-          new std::thread(RunServer, &rpc_server_, rpc_service_, endpoint));
+      rpc_service_.reset(new detail::AsyncGRPCServer(endpoint));
+      server_thread_.reset(new std::thread(RunServer, rpc_service_));
    }
  }

@@ -76,7 +69,7 @@ class RecvOp : public framework::OperatorBase {
    detail::MessageWithName term_msg;
    term_msg.first = LISTEN_TERMINATE_MESSAGE;
    rpc_service_->Push(term_msg);
-    rpc_server_->Shutdown();
+    rpc_service_->ShutDown();
    server_thread_->join();
  }

@@ -99,10 +92,12 @@ class RecvOp : public framework::OperatorBase {
    auto grad_list = Attr<std::vector<std::string>>("GradList");
    auto trainer_count = Attr<int>("Trainers");
    size_t param_count = param_list.size();
+
    rpc_service_->Reset();
    // TODO(typhoonzero): change this to a while_op for every cluster-batch.
    bool exit_flag = false;
    while (!exit_flag) {
+      // TODO(gognwb): simply this loop.
      // Get from multiple trainers, we don't care about order in which
      // the gradient arrives, just add suffix 0~n then average the gradient.
      for (size_t i = 0; i < param_count * trainer_count; ++i) {
@@ -110,6 +105,7 @@ class RecvOp : public framework::OperatorBase {
        const detail::MessageWithName &v = rpc_service_->Get();
        auto grad_var_name = v.first;
        if (grad_var_name == LISTEN_TERMINATE_MESSAGE) {
+          VLOG(4) << "received LISTEN_TERMINATE_MESSAGE and RunOp.Run() exit";
          exit_flag = true;
          break;
        }
@@ -118,10 +114,12 @@ class RecvOp : public framework::OperatorBase {
        if (it != grad_list.end()) {
          param_var_name = param_list[it - grad_list.begin()];
        } else {
-          LOG(ERROR) << "grad have no paired param found!";
+          LOG(ERROR) << "grad have no paired param found!\"" << grad_var_name
+                     << "\"";
        }
        VLOG(3) << "recved grad: " << grad_var_name
                << " updating param: " << param_var_name;
+
        auto *merged_grad = recv_scope.FindVar(grad_var_name);
        if (merged_grad == nullptr) {
          auto *ptr = recv_scope.Var(grad_var_name);
@@ -141,9 +139,11 @@ class RecvOp : public framework::OperatorBase {
        auto &dev_ctx = *pool.Get(dev_place);
        detail::DeserializeFromMessage(v.second, dev_ctx, var);
      }
+
      if (exit_flag) {
        break;
      }
+
      rpc_service_->Reset();

      std::string program_str = Attr<std::string>("OptimizeProgram");
@@ -158,17 +158,14 @@ class RecvOp : public framework::OperatorBase {
      } catch (std::exception &e) {
        LOG(ERROR) << "run sub program error " << e.what();
      }
+
      rpc_service_->Done();
      grads_counter_.clear();
    }  // while(true)
  }

 protected:
-  // grpc server instance to track status and gracefully shutdown.
-  // borrow an pointer from server thread.
-  Server *rpc_server_{nullptr};
-  // grpc send/recv service implement to register.
-  std::shared_ptr<detail::SendRecvServerImpl> rpc_service_;
+  std::shared_ptr<detail::AsyncGRPCServer> rpc_service_;
  std::shared_ptr<std::thread> server_thread_;
  mutable std::unordered_map<std::string, int> grads_counter_;
 };

--- a/paddle/operators/reorder_lod_tensor_by_rank_op.cc
+++ b/paddle/operators/reorder_lod_tensor_by_rank_op.cc
@@ -26,22 +26,44 @@ class ReorderLoDTensorByRankTableOpProtoMaker
  ReorderLoDTensorByRankTableOpProtoMaker(OpProto *proto,
                                          OpAttrChecker *op_checker)
      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "(LoDTensor) the input lod tensor need to be reordered.");
+    AddInput("X",
+             "(LoDTensor), the input lod tensor to be reordered according to "
+             "Input(RankTable).");
    AddInput("RankTable",
-             "(LoDRankTable) the rank table that input need follow");
-    AddOutput("Out", "(LoDTensor) reordered lod tensor");
-    AddComment(R"DOC(ReorderLoDTensorByRankTable
+             "(LoDRankTable), the rank table according to which Input(X) is "
+             "reordered.");
+    AddOutput("Out", "(LoDTensor), the reordered lod tensor.");
+    AddComment(R"DOC(ReorderLoDTensorByRankTable operator.

-Reorder the input X by the rank of `RankTable`. If `RankTable` is ordered by
-index [3, 0, 2, 1]. Input X will reorder its sequence, the third sequence of
-X will be the first sequence of Output.
-
-NOTE: The RankTable does not need to be calculated by X.
+Input(X) is a batch of sequences. Input(RankTable) stores new orders of the
+input sequence batch. The reorder_lod_tensor_by_rank operator reorders the
+Input(X) according to the information provided by Input(RankTable).

 For example:
-The X = [Seq0, Seq1, Seq2, Seq3]. The indices of RankTable are [3, 0, 2, 1].

-The Out =  [Seq3, Seq0, Seq2, Seq1] with correct LoD information.
+If the indices stored in the Input(RankTable) are [3, 0, 2, 1], the
+Input(X) will be reordered that the fourth sequence in Input(X) will become the
+first one, and then followed by the original first, third, and the second one.
+
+This is:
+X = [Seq0, Seq1, Seq2, Seq3]. The indices in RankTable are [3, 0, 2, 1].
+Out =  [Seq3, Seq0, Seq2, Seq1] with a new LoD information.
+
+If the LoD information of Input(X) is empty, this means Input(X) is not sequence
+data. This is also identical to a batch of sequences where each sequence has a
+fixed length 1. In this case, the reorder_lod_tensor_by_rank operator reorders
+each slice of Input(X) along the first axis according to Input(RankTable).
+
+This is:
+X = [Slice0, Slice1, Slice2, Slice3] and its LoD information is empty. The
+indices in RankTable are [3, 0, 2, 1].
+Out = [Slice3, Slice0, Slice2, Slice1] with no LoD information is appended.
+
+NOTE: This operator sorts Input(X) according to a given LoDRankTable which does
+not need to be calculated according to Input(X). It can be calculated according
+to another different sequence, and then this operator sorts Input(X) according
+to the given LoDRankTable.
+
 )DOC");
  }
 };

--- a/paddle/operators/send_op.cc
+++ b/paddle/operators/send_op.cc
@@ -19,59 +19,45 @@ limitations under the License. */
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/op_registry.h"

-#include "paddle/operators/detail/send_recv_impl.h"
-#include "paddle/operators/detail/simple_block_queue.h"
+#include <future>
+#include "paddle/operators/detail/grpc_client.h"

 namespace paddle {
 namespace operators {

-// TODO(typhoonzero): this is a simple implementation which only send
-// one tensor
 class SendOp : public framework::OperatorBase {
 public:
-  SendOp(const std::string &type, const framework::VariableNameMap &inputs,
-         const framework::VariableNameMap &outputs,
-         const framework::AttributeMap &attrs)
-      : OperatorBase(type, inputs, outputs, attrs) {
-    // init client when the operator is created at runtime.
-    std::vector<std::string> endpoints =
-        Attr<std::vector<std::string>>("endpoints");
-    for (auto ep : endpoints) {
-      client_map_[ep].reset(new detail::RPCClient(
-          grpc::CreateChannel(ep, grpc::InsecureChannelCredentials())));
-    }
-  }
+  SendOp(const std::string& type, const framework::VariableNameMap& inputs,
+         const framework::VariableNameMap& outputs,
+         const framework::AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}

-  void Run(const framework::Scope &scope,
-           const platform::Place &dev_place) const override {
+  void Run(const framework::Scope& scope,
+           const platform::Place& dev_place) const override {
    auto ins = Inputs("X");
    auto outs = Outputs("Out");
    std::vector<std::string> epmap = Attr<std::vector<std::string>>("epmap");
-    // TODO(typhoonzero): use async calls to send multiple variable asyncly.
-    for (size_t i = 0; i < ins.size(); ++i) {
-      bool ret = client_map_[epmap[i]]->SendVariable(scope, ins[i]);
-      if (!ret) {
-        LOG(ERROR) << "send variable error: " << ins[i];
-      }
+
+    // FIXME(gongwb): DeviceContext?
+    auto ctx = platform::CPUDeviceContext();
+    for (size_t i = 0; i < ins.size(); i++) {
+      client_.AsyncSendVariable(epmap[i], ctx, scope, ins[i]);
    }
-    // TODO(typhoonzero): support async optimization
-    client_map_[epmap[0]]->Wait();
-    for (size_t i = 0; i < outs.size(); ++i) {
-      bool ret = client_map_[epmap[i]]->GetVariable(scope, outs[i]);
-      if (!ret) {
-        LOG(ERROR) << "GetVariable error: " << outs[i];
-      }
+
+    for (size_t i = 0; i < outs.size(); i++) {
+      client_.AsyncGetVariable(epmap[i], ctx, scope, outs[i]);
    }
+
+    client_.wait();
  }

- protected:
-  mutable std::unordered_map<std::string, std::shared_ptr<detail::RPCClient>>
-      client_map_;
+ private:
+  mutable detail::RPCClient client_;
 };

 class SendOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
-  SendOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+  SendOpMaker(OpProto* proto, OpAttrChecker* op_checker)
      : OpProtoAndCheckerMaker(proto, op_checker) {
    AddInput("X", "(Tensor) Input tensor to be send").AsDuplicable();
    AddOutput("Out", "(Tensor) Output tensor to get from server")

--- a/paddle/operators/send_recv_op_test.cc
+++ b/paddle/operators/send_recv_op_test.cc
@@ -140,7 +140,7 @@ void StartServerNet(bool is_sparse) {

 TEST(SendRecvOp, CPUDense) {
  std::thread server_thread(StartServerNet, false);
-  sleep(3);  // wait server to start
+  sleep(10);  // wait server to start
  // local net
  f::Scope scope;
  p::CPUPlace place;

--- a/paddle/operators/sequence_erase_op.cc
+++ b/paddle/operators/sequence_erase_op.cc
+/* 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. */
+
+#include "paddle/operators/sequence_erase_op.h"
+
+namespace paddle {
+namespace operators {
+
+class SequenceEraseOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of SequenceEraseOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of SequenceEraseOp should not be null.");
+    auto x_dims = ctx->GetInputDim("X");
+    PADDLE_ENFORCE(x_dims.size() == 2 && x_dims[1] == 1,
+                   "Input(X) of SequenceEraseOp should be a 2-D LoDTensor "
+                   "with the 2nd dimension equal to 1.");
+    ctx->SetOutputDim("Out", x_dims);
+  }
+};
+
+class SequenceEraseOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  SequenceEraseOpMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "(2-D LoDTensor with the 2nd dim. equal to 1) "
+             "Input LoDTensor of SequenceEraseOp.");
+    AddOutput("Out",
+              "(2-D LoDTensor with the 2nd dim. equal to 1) "
+              "Output LoDTensor of SequenceEraseOp.");
+    AddAttr<std::vector<int>>("tokens",
+                              "(vector<int>) Tokens need to be erased from "
+                              "input sequences.");
+    AddComment(R"DOC(
+Sequence Erase Operator.
+
+Sequence erase operator erases tokens specified by Attr(tokens) from the input 
+sequences Input(X), and outputs the remaining data and modifies the LoD 
+information at the same time. For example, given a 2-D LoDTensor
+
+    X = [[2, 2, 6, 1, 3, 9, 6, 1, 0, 1]]^T
+
+with lod = [[0, 3, 6, 10]], there are three sequences in the input:
+   
+     X1 = [[2, 2, 6]]^T, X2 = [[1, 3, 9]]^T and X3 = [[6, 1, 0, 1]]^T.
+
+If the tokens to be erased are Attr(tokens) = [2, 3, 5], after the erasing 
+operation, the three sequences become
+
+    X1' = [[6]]^T, X2' = [[1, 9]]^T and X3' = [[6, 1, 0, 1]]^T.
+
+Hence the LoDTensor Output(Out) should be
+
+    Out = [[6, 1, 9, 6, 1, 0, 1]]^T,
+
+with lod = [[0, 1, 3, 7]].
+
+An example usage for this operator is to remove the special tokens when 
+computing the edit distance between two strings, such as blank, start token, 
+and end token.
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(sequence_erase, ops::SequenceEraseOp,
+                             ops::SequenceEraseOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    sequence_erase,
+    ops::SequenceEraseKernel<paddle::platform::CPUDeviceContext, int32_t>);
--- a/paddle/operators/sequence_erase_op.cu
+++ b/paddle/operators/sequence_erase_op.cu
+/* 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. */
+
+#include <thrust/device_vector.h>
+#include <thrust/host_vector.h>
+#include "paddle/operators/sequence_erase_op.h"
+#include "paddle/platform/cuda_helper.h"
+
+namespace paddle {
+namespace operators {
+using platform::PADDLE_CUDA_NUM_THREADS;
+using LoDTensor = framework::LoDTensor;
+
+template <typename T>
+__global__ void LabelErasedIdx(const T* in_dat, const int in_len,
+                               const T* tokens, const int tokens_len,
+                               int* num_erased) {
+  int index = blockIdx.x * blockDim.x + threadIdx.x;
+  if (index < in_len) {
+    int erased = 0;
+    for (int i = 0; i < tokens_len; ++i) {
+      if (in_dat[index] == tokens[i]) {
+        erased = 1;
+      }
+    }
+    num_erased[index + 1] = erased;
+    if (index == 0) {
+      num_erased[0] = 0;
+    }
+  }
+}
+
+template <typename T>
+__global__ void GetOutLod(const T* num_erased, const int* in_lod,
+                          const int lod_len, int* out_lod0) {
+  int index = blockIdx.x * blockDim.x + threadIdx.x;
+  if (index < lod_len) {
+    out_lod0[index] = in_lod[index] - num_erased[in_lod[index]];
+  }
+}
+
+template <typename T>
+__global__ void SetOutput(const T* in_dat, const int in_len,
+                          const int* num_erased, T* out_dat) {
+  int index = blockIdx.x * blockDim.x + threadIdx.x;
+  if (index < in_len) {
+    if (in_dat[index] != in_dat[index + 1]) {
+      out_dat[index - num_erased[index]] = in_dat[index];
+    }
+  }
+}
+
+template <typename T>
+class SequenceEraseOpCUDAKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* in = ctx.Input<LoDTensor>("X");
+    auto* out = ctx.Output<LoDTensor>("Out");
+
+    auto lod = in->lod();
+    PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now.");
+    PADDLE_ENFORCE_EQ(lod[0].back(), (size_t)in->numel(),
+                      "The actual size mismatches with the LoD information.");
+    auto tokens = ctx.Attr<std::vector<T>>("tokens");
+    auto tokens_len = tokens.size();
+    auto in_len = in->numel();
+    auto in_dat = in->data<T>();
+    auto lod0 = lod[0];
+
+    thrust::host_vector<T> host_tokens(tokens_len);
+    for (size_t i = 0; i < tokens.size(); ++i) {
+      host_tokens[i] = tokens[i];
+    }
+    thrust::device_vector<T> dev_tokens = host_tokens;
+    thrust::device_vector<int> num_erased(in_len + 1);
+
+    T* dev_tokens_ptr = thrust::raw_pointer_cast(dev_tokens.data());
+    int* num_erased_ptr = thrust::raw_pointer_cast(num_erased.data());
+
+    auto stream = ctx.cuda_device_context().stream();
+    LabelErasedIdx<<<(in_len - 1) / PADDLE_CUDA_NUM_THREADS + 1,
+                     PADDLE_CUDA_NUM_THREADS, 0, stream>>>(
+        in_dat, in_len, dev_tokens_ptr, tokens_len, num_erased_ptr);
+    thrust::inclusive_scan(num_erased.begin() + 1, num_erased.end(),
+                           num_erased.begin() + 1);
+
+    // Calc LoD
+    auto lod_len = lod0.size();
+    thrust::host_vector<int> host_lod(lod_len);
+    for (size_t i = 0; i < lod_len; ++i) {
+      host_lod[i] = lod0[i];
+    }
+    thrust::device_vector<int> dev_in_lod = host_lod;
+    thrust::device_vector<int> dev_out_lod(lod_len);
+    int* dev_in_lod_ptr = thrust::raw_pointer_cast(dev_in_lod.data());
+    int* dev_out_lod_ptr = thrust::raw_pointer_cast(dev_out_lod.data());
+    GetOutLod<<<(lod_len - 1) / PADDLE_CUDA_NUM_THREADS + 1,
+                PADDLE_CUDA_NUM_THREADS, 0, stream>>>(
+        num_erased_ptr, dev_in_lod_ptr, lod_len, dev_out_lod_ptr);
+    thrust::host_vector<int> host_out_lod = dev_out_lod;
+    std::vector<int> out_lod0(lod_len, 0);
+    for (size_t i = 0; i < lod_len; i++) {
+      out_lod0[i] = host_out_lod[i];
+    }
+    framework::LoD out_lod;
+    out_lod.push_back(out_lod0);
+    out->set_lod(out_lod);
+
+    // Set output
+    out->Resize({out_lod0.back(), 1});
+    auto out_dat = out->mutable_data<T>(ctx.GetPlace());
+    SetOutput<<<(in_len - 1) / PADDLE_CUDA_NUM_THREADS + 1,
+                PADDLE_CUDA_NUM_THREADS, 0, stream>>>(in_dat, in_len,
+                                                      num_erased_ptr, out_dat);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP_CUDA_KERNEL(sequence_erase,
+                        paddle::operators::SequenceEraseOpCUDAKernel<int32_t>);
--- a/paddle/operators/sequence_erase_op.h
+++ b/paddle/operators/sequence_erase_op.h
+/* 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_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename T>
+class SequenceEraseKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* in = ctx.Input<framework::LoDTensor>("X");
+    auto* out = ctx.Output<framework::LoDTensor>("Out");
+
+    auto lod = in->lod();
+    PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now.");
+    PADDLE_ENFORCE_EQ(lod[0].back(), (size_t)in->numel(),
+                      "The actual size mismatches with the LoD information.");
+    auto tokens = ctx.Attr<std::vector<int>>("tokens");
+    auto in_len = in->numel();
+    auto in_dat = in->data<T>();
+    auto lod0 = lod[0];
+
+    std::vector<size_t> num_erased(in_len + 1, 0);
+    std::vector<size_t> out_lod0(1, 0);
+    for (size_t i = 0; i < lod0.size() - 1; ++i) {
+      size_t num_out = 0;
+      for (auto j = lod0[i] + 1; j <= lod0[i + 1]; ++j) {
+        num_erased[j] = num_erased[j - 1];
+        if (std::find(tokens.begin(), tokens.end(), in_dat[j - 1]) !=
+            tokens.end()) {
+          num_erased[j] += 1;
+        } else {
+          num_out += 1;
+        }
+      }
+      out_lod0.push_back(out_lod0.back() + num_out);
+    }
+
+    auto out_len = in_len - num_erased[in_len];
+    out->Resize({static_cast<int64_t>(out_len), 1});
+    auto out_dat = out->mutable_data<T>(ctx.GetPlace());
+
+    for (int64_t i = 0; i < in_len; ++i) {
+      if (num_erased[i] == num_erased[i + 1]) {
+        out_dat[i - num_erased[i]] = in_dat[i];
+      }
+    }
+    framework::LoD out_lod;
+    out_lod.push_back(out_lod0);
+    out->set_lod(out_lod);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
--- a/paddle/operators/shrink_rnn_memory_op.cc
+++ b/paddle/operators/shrink_rnn_memory_op.cc
@@ -45,7 +45,7 @@ class ShrinkRNNMemoryOp : public ArrayOp {
        rank_items.begin();

    auto *out_var = scope.FindVar(Output("Out"));
-    PADDLE_ENFORCE(out_var != nullptr, "Output Out must be set");
+    PADDLE_ENFORCE(out_var != nullptr, "Output(Out) must be set.");
    auto &out_tensor = *out_var->GetMutable<framework::LoDTensor>();

    size_t height = dst_num_rows;
@@ -76,15 +76,17 @@ class ShrinkRNNMemoryOpProtoMaker : public framework::OpProtoAndCheckerMaker {
             "(LoDTensor) The step index. The RNN step memory 'X' will be "
             "shrinked to match the size of the input of the index'th step.");
    AddOutput("Out", "(LoDTensor) The shrinked RNN step memory.");
-    AddComment(
-        R"DOC(
-        In dynamic RNN, we are able to handle sequences of different lengths.
-        Because of the multiple lengths, the size of each step input can be
-        different, which may lead to a mismatching between the input of
-        the current step and the memory generated by the previous one. This
-        operator shrinks memory according to the size of the next step input,
-        to make sure that they can match each other.
-        )DOC");
+    AddComment(R"DOC(
+This operator is used to shrink output batch of memory defined in dynamic RNN.
+
+Dynamic RNN is able to handle variable-length sequences, in which, sequences in
+a mini-batch are sorted by their lengths first. After that, the longest sequence
+becomes the first one in the sorted batch, followed by the second longest, the
+third longest, and so on. Dynamic RNN then slices a batch input timestep by
+timestep from the sorted input. Once any sequence in the input batch reaches its
+end, memory defined in dynamicRNN has to shrink its outputs to adapt to the input
+batch size for the next time step.
+)DOC");
  }
 };


--- a/paddle/scripts/submit_local.sh.in
+++ b/paddle/scripts/submit_local.sh.in
@@ -49,7 +49,18 @@ function cpu_config() {
  if [ "@WITH_MKL@" == "OFF" ]; then
    return 0
  fi
-  ht=`lscpu |grep "per core"|awk -F':' '{print $2}'|xargs`
+  platform="`uname -s`"
+  ht=0
+  if [ $platform == "Linux" ]; then
+    ht=`lscpu |grep "per core"|awk -F':' '{print $2}'|xargs`
+  elif [ $platform == "Darwin" ]; then
+    if [`sysctl -n hw.physicalcpu` -eq `sysctl -n hw.logicalcpu`]; then
+      # HT is OFF
+      ht=1
+    fi
+  else
+    return 0
+  fi
  if [ $ht -eq 1 ]; then # HT is OFF
    if [ -z "$KMP_AFFINITY" ]; then
      export KMP_AFFINITY="granularity=fine,compact,0,0"
@@ -72,7 +83,15 @@ function threads_config() {
  # according to trainer_count and total processors
  # only when MKL enabled
  # auto set OPENBLAS_NUM_THREADS when do not use MKL
-  processors=`grep "processor" /proc/cpuinfo|sort -u|wc -l`
+  platform="`uname -s`"
+  processors=0
+  if [ $platform == "Linux" ]; then
+    processors=`grep "processor" /proc/cpuinfo|sort -u|wc -l`
+  elif [ $platform == "Darwin" ]; then
+    processors=`sysctl -n hw.logicalcpu`
+  else
+    return 0
+  fi
  trainers=`grep -Eo 'trainer_count.[0-9]+' <<< "$@" |grep -Eo '[0-9]+'|xargs`
  if [ -z $trainers ]; then
    trainers=1
@@ -148,11 +167,7 @@ else:
  sys.exit(0)
 EOF

-if [ "`uname -s`" == "Linux" ]; then
-  # only support on linux yet, with mac can use v2
-  cpu_config
-fi
-
+cpu_config
 # echo $KMP_AFFINITY $OMP_DYNAMIC

 case "$1" in

--- a/python/paddle/utils/dump_v2_config.py
+++ b/python/paddle/utils/dump_v2_config.py
+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# 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.
+import collections
+
+from paddle.trainer_config_helpers.layers import LayerOutput
+from paddle.v2.layer import parse_network
+from paddle.proto import TrainerConfig_pb2
+
+__all__ = ["dump_v2_config"]
+
+
+def dump_v2_config(topology, save_path, binary=False):
+    """ Dump the network topology to a specified file.
+
+    This function is only used to dump network defined by using PaddlePaddle V2
+    APIs. This function will NOT dump configurations related to PaddlePaddle
+    optimizer.
+
+    :param topology: The output layers (can be more than one layers given in a
+                     Python List or Tuple) of the entire network. Using the
+                     specified layers (if more than one layer is given) as root,
+                     traversing back to the data layer(s), all the layers
+                     connected to the specified output layers will be dumped.
+                     Layers not connceted to the specified will not be dumped.
+    :type topology: LayerOutput|List|Tuple
+    :param save_path: The path to save the dumped network topology.
+    :type save_path: str
+    :param binary: Whether to dump the serialized network topology or not.
+                   The default value is false. NOTE that, if you call this
+                   function to generate network topology for PaddlePaddle C-API,
+                   a serialized version of network topology is required. When
+                   using PaddlePaddle C-API, this flag MUST be set to True.
+    :type binary: bool
+    """
+
+    if isinstance(topology, LayerOutput):
+        topology = [topology]
+    elif isinstance(topology, collections.Sequence):
+        for out_layer in topology:
+            assert isinstance(out_layer, LayerOutput), (
+                "The type of each element in the parameter topology "
+                "should be LayerOutput.")
+    else:
+        raise RuntimeError("Error input type for parameter topology.")
+
+    model_str = parse_network(topology)
+    with open(save_path, "w") as fout:
+        if binary:
+            fout.write(model_str.SerializeToString())
+        else:
+            fout.write(str(model_str))
--- a/python/paddle/utils/merge_model.py
+++ b/python/paddle/utils/merge_model.py
@@ -30,7 +30,8 @@ def merge_v2_model(net, param_file, output_file):
    which ends with .tar.gz.

    @param  net            The output layer of the network for inference.
-    @param  param_file     Path of the parameters (.tar.gz) which is stored by v2 api.
+    @param  param_file     Path of the parameters (.tar.gz) which is stored by
+                           v2 api.
    @param  output_file    Path of the merged file which will be generated.

    Usage:

--- a/python/paddle/v2/fluid/__init__.py
+++ b/python/paddle/v2/fluid/__init__.py
@@ -23,9 +23,22 @@ from memory_optimization_transpiler import memory_optimize

 Tensor = LoDTensor
 __all__ = framework.__all__ + executor.__all__ + [
-    'io', 'initializer', 'layers', 'nets', 'optimizer', 'backward',
-    'regularizer', 'LoDTensor', 'CPUPlace', 'CUDAPlace', 'Tensor', 'ParamAttr'
-    'DataFeeder', 'clip', 'DistributeTranspiler', 'memory_optimize'
+    'io',
+    'initializer',
+    'layers',
+    'nets',
+    'optimizer',
+    'backward',
+    'regularizer',
+    'LoDTensor',
+    'CPUPlace',
+    'CUDAPlace',
+    'Tensor',
+    'ParamAttr'
+    'DataFeeder',
+    'clip',
+    'DistributeTranspiler',
+    'memory_optimize',
 ]


@@ -58,7 +71,7 @@ def __bootstrap__():

    read_env_flags = ['use_pinned_memory', 'check_nan_inf']
    if core.is_compile_gpu():
-        read_env_flags.append('fraction_of_gpu_memory_to_use')
+        read_env_flags += ['fraction_of_gpu_memory_to_use', 'op_sync']
    core.init_gflags([sys.argv[0]] +
                     ["--tryfromenv=" + ",".join(read_env_flags)])
    core.init_glog(sys.argv[0])

--- a/python/paddle/v2/fluid/backward.py
+++ b/python/paddle/v2/fluid/backward.py
@@ -3,7 +3,10 @@ from . import core
 import collections
 import copy

-__all__ = ['append_backward', 'calc_gradient']
+__all__ = [
+    'append_backward',
+    'calc_gradient',
+]


 def _rename_arg_(op_descs, old_name, new_name, begin_idx=None, end_idx=None):

--- a/python/paddle/v2/fluid/clip.py
+++ b/python/paddle/v2/fluid/clip.py
@@ -3,7 +3,9 @@ import layers
 from . import core

 __all__ = [
-    'GradientClipByValue', 'append_gradient_clip_ops', 'error_clip_callback'
+    'GradientClipByValue',
+    'append_gradient_clip_ops',
+    'error_clip_callback',
 ]



--- a/python/paddle/v2/fluid/default_scope_funcs.py
+++ b/python/paddle/v2/fluid/default_scope_funcs.py
 """
 Default scope function.

-`Paddle` manages Scope as programming language's scope.  It just a 
-thread-local stack of Scope. Top of that stack is current scope, the bottom 
-of that stack is all scopes' parent. 
+`Paddle` manages Scope as programming language's scope.  It just a
+thread-local stack of Scope. Top of that stack is current scope, the bottom
+of that stack is all scopes' parent.

-Invoking `var/find_var`  can `new/find` variable in current scope. 
-Invoking `enter_local_scope/leave_local_scope` can create or destroy local 
-scope. 
+Invoking `var/find_var`  can `new/find` variable in current scope.
+Invoking `enter_local_scope/leave_local_scope` can create or destroy local
+scope.

-A `scoped_function` will take a `function` as input. That function will be 
-invoked in a new local scope. 
+A `scoped_function` will take a `function` as input. That function will be
+invoked in a new local scope.
 """

 import paddle.v2.fluid.core
@@ -19,8 +19,12 @@ import threading
 __tl_scope__ = threading.local()

 __all__ = [
-    'get_cur_scope', 'enter_local_scope', 'leave_local_scope', 'var',
-    'find_var', 'scoped_function'
+    'get_cur_scope',
+    'enter_local_scope',
+    'leave_local_scope',
+    'var',
+    'find_var',
+    'scoped_function',
 ]


@@ -71,7 +75,7 @@ def find_var(name):
 def scoped_function(func):
    """
    invoke `func` in new scope.
-    
+
    :param func: a callable function that will be run in new scope.
    :type func: callable
    """

--- a/python/paddle/v2/fluid/evaluator.py
+++ b/python/paddle/v2/fluid/evaluator.py
@@ -4,7 +4,10 @@ import layers
 from framework import Program, unique_name, Variable, program_guard
 from layer_helper import LayerHelper

-__all__ = ['Accuracy', 'ChunkEvaluator']
+__all__ = [
+    'Accuracy',
+    'ChunkEvaluator',
+]


 def _clone_var_(block, var):
@@ -21,19 +24,19 @@ def _clone_var_(block, var):
 class Evaluator(object):
    """
    Base Class for all evaluators
-    
+
    Args:
-        name(str): The name of evaluator. such as, "accuracy". Used for generate 
+        name(str): The name of evaluator. such as, "accuracy". Used for generate
            temporary variable name.
-        main_program(Program, optional): The evaluator should be added to this 
+        main_program(Program, optional): The evaluator should be added to this
            main_program. Default default_main_program()
-        startup_program(Program, optional):The parameter should be added to this 
+        startup_program(Program, optional):The parameter should be added to this
            startup_program. Default default_startup_program()
-            
+
    Attributes:
-        states(list): The list of state variables. states will be reset to zero 
+        states(list): The list of state variables. states will be reset to zero
            when `reset` is invoked.
-        metrics(list): The list of metrics variables. They will be calculate 
+        metrics(list): The list of metrics variables. They will be calculate
            every mini-batch
    """

@@ -66,14 +69,14 @@ class Evaluator(object):

    def create_state(self, suffix, dtype, shape):
        """
-        Create state variable. 
-        
+        Create state variable.
+
        NOTE: It is not a public API.
-        
+
        Args:
-            suffix(str): the state suffix. 
-            dtype(str|core.DataType): the state data type 
-            shape(tuple|list): the shape of state 
+            suffix(str): the state suffix.
+            dtype(str|core.DataType): the state data type
+            shape(tuple|list): the shape of state

        Returns: State variable

@@ -127,8 +130,8 @@ class Accuracy(Evaluator):

 class ChunkEvaluator(Evaluator):
    """
-    Accumulate counter numbers output by chunk_eval from mini-batches and 
-    compute the precision recall and F1-score using the accumulated counter 
+    Accumulate counter numbers output by chunk_eval from mini-batches and
+    compute the precision recall and F1-score using the accumulated counter
    numbers.
    """


--- a/python/paddle/v2/fluid/framework.py
+++ b/python/paddle/v2/fluid/framework.py
@@ -7,9 +7,15 @@ import proto.framework_pb2 as framework_pb2
 from . import core

 __all__ = [
-    'Block', 'Variable', 'Program', 'Operator', 'default_startup_program',
-    'default_main_program', 'program_guard', 'switch_startup_program',
-    'switch_main_program'
+    'Block',
+    'Variable',
+    'Program',
+    'Operator',
+    'default_startup_program',
+    'default_main_program',
+    'program_guard',
+    'switch_startup_program',
+    'switch_main_program',
 ]

 EMPTY_VAR_NAME = core.kEmptyVarName()
@@ -236,6 +242,9 @@ class Variable(object):

    __repr__ = __str__

+    def set_desc(self, input):
+        self.desc = input
+
    @property
    def persistable(self):
        return self.desc.persistable()

--- a/python/paddle/v2/fluid/initializer.py
+++ b/python/paddle/v2/fluid/initializer.py
 import framework
 import numpy as np

-__all__ = ['Constant', 'Uniform', 'Normal', 'Xavier']
+__all__ = [
+    'Constant',
+    'Uniform',
+    'Normal',
+    'Xavier',
+]


 class Initializer(object):

--- a/python/paddle/v2/fluid/io.py
+++ b/python/paddle/v2/fluid/io.py
@@ -4,13 +4,29 @@ import cPickle as pickle
 from paddle.v2.fluid.framework import Program, Parameter, default_main_program, Variable

 __all__ = [
-    'save_vars', 'save_params', 'save_persistables', 'load_vars', 'load_params',
-    'load_persistables', "save_inference_model", "load_inference_model",
-    "get_inference_program"
+    'save_vars',
+    'save_params',
+    'save_persistables',
+    'load_vars',
+    'load_params',
+    'load_persistables',
+    'save_inference_model',
+    'load_inference_model',
+    'get_inference_program',
 ]


 def is_parameter(var):
+    """Check whether the variable is a Parameter.
+
+    This function checks whether the input variable is a Parameter.
+
+    Args:
+        var : The input variable.
+
+    Returns:
+        boolean result whether the variable is a Parameter.
+    """
    return isinstance(var, Parameter)



--- a/python/paddle/v2/fluid/layers/control_flow.py
+++ b/python/paddle/v2/fluid/layers/control_flow.py
@@ -12,7 +12,7 @@ __all__ = [
    'array_to_lod_tensor', 'increment', 'array_write', 'create_array',
    'less_than', 'array_read', 'shrink_memory', 'array_length', 'IfElse',
    'DynamicRNN', 'ConditionalBlock', 'StaticRNN', 'reorder_lod_tensor_by_rank',
-    'ParallelDo'
+    'ParallelDo', 'Print'
 ]


@@ -110,6 +110,61 @@ def merge_lod_tensor(in_true, in_false, x, mask, level=0):
    return out


+def Print(input,
+          first_n=-1,
+          message=None,
+          summarize=-1,
+          print_tensor_name=True,
+          print_tensor_type=True,
+          print_tensor_shape=True,
+          print_tensor_lod=True):
+    '''
+    **Print operator**
+
+    This creates a print op that will print when a tensor is accessed.
+
+    Wraps the tensor passed in so that whenever that a tensor is accessed,
+    the message `message` is printed, along with the current value of the
+    tensor `t`.
+
+    Args:
+      input(Variable): A Tensor to print.
+      summarize(int): Print this number of elements in the tensor, will print all
+                 if left negative.
+      message(str): A string message to print as a prefix.
+      first_n(int): Only log `first_n` number of times.
+      print_tensor_name(bool): Print the tensor name.
+      print_tensor_type(bool): Print the tensor type.
+      print_tensor_shape(bool): Print the tensor shape.
+      print_tensor_lod(bool): Print the tensor lod.
+
+    Returns:
+      None
+
+    Examples:
+        .. code-block:: python
+
+        value = some_layer(...)
+        Print(value, summarize=10,
+              message="The content of some_layer: ")
+    '''
+    helper = LayerHelper('print', **locals())
+    out = helper.create_tmp_variable(dtype='int32')
+    helper.append_op(
+        type='print',
+        inputs={'input': input},
+        attrs={
+            'first_n': first_n,
+            'summarize': summarize,
+            'message': message or "",
+            'print_tensor_name': print_tensor_name,
+            'print_tensor_type': print_tensor_type,
+            'print_tensor_shape': print_tensor_shape,
+            'print_tensor_lod': print_tensor_lod,
+        })
+    return out
+
+
 class BlockGuard(object):
    """
    BlockGuard class.
@@ -687,11 +742,10 @@ def topk(input, k):


 def lod_tensor_to_array(x, table):
-    """This function performs the operation that converts an LOD_Tensor to
-       an array.
+    """ Convert a LOD_TENSOR to an LOD_TENSOR_ARRAY.

    Args:
-        x (Variable|list): The tensor that needs to be converted to an array.
+        x (Variable|list): The LOD tensor to be converted to a LOD tensor array.
        table (ParamAttr|list): The variable that stores the level of lod
                                which is ordered by sequence length in
                                descending order.
@@ -721,11 +775,10 @@ def lod_tensor_to_array(x, table):


 def array_to_lod_tensor(x, table):
-    """This function performs the operations that converts an array to
-       an LOD_Tensor.
+    """Convert a LoD_Tensor_Aarry to an LoDTensor.

    Args:
-        x (Variable|list): The array that needs to be converted to a tensor.
+        x (Variable|list): The lod tensor array to be converted to a tensor.
        table (ParamAttr|list): The variable that stores the level of lod
                                which is ordered by sequence length in
                                descending order.
@@ -753,7 +806,8 @@ def array_to_lod_tensor(x, table):


 def increment(x, value=1.0, in_place=True):
-    """This function performs an operation that increments each value in the
+    """
+    This function performs an operation that increments each value in the
    input :math:`x` by an amount: :math:`value` as mentioned in the input
    parameter. This operation is performed in-place by default.

@@ -786,17 +840,24 @@ def increment(x, value=1.0, in_place=True):


 def array_write(x, i, array=None):
-    """This function performs the operation to write the data out as an
-    LOD_TENSOR_ARRAY.
+    """
+    This function writes the given input variable to the specified position
+    indicating by the arrary index to an output LOD_TENSOR_ARRAY. If the
+    output LOD_TENSOR_ARRAY is not given(None), a new one will be created and
+    returned.

    Args:
        x (Variable|list): The input tensor from which the data will be read.
-        i (Variable|list): The subscript index in tensor array, that points the
-                           place from which data will be read.
-        array (Variable|list): The data can be read into this variable if
-                               this is assigned.
+        i (Variable|list): The index of the output LOD_TENSOR_ARRAY, pointing to
+                           the position to which the input tensor will be
+                           written.
+        array (Variable|list): The output LOD_TENSOR_ARRAY to which the input
+                               tensor will be written. If this parameter is
+                               NONE, a new LOD_TENSOR_ARRAY will be created and
+                               returned.
+
    Returns:
-        Variable: The tensor type variable that has the data written to it.
+        Variable: The output LOD_TENSOR_ARRAY where the input tensor is written.

    Examples:
        .. code-block::python
@@ -1173,7 +1234,7 @@ class DynamicRNN(object):
        self._assert_in_rnn_block_("step_input")
        if not isinstance(x, Variable):
            raise TypeError(
-                "step_input() can only take a Variable as its input")
+                "step_input() can only take a Variable as its input.")
        parent_block = self._parent_block_()
        if self.lod_rank_table is None:
            self.lod_rank_table = parent_block.create_var(
@@ -1234,8 +1295,8 @@ class DynamicRNN(object):

    def __call__(self, *args, **kwargs):
        if self.status != DynamicRNN.AFTER_RNN:
-            raise ValueError(
-                "Dynamic RNN outputs can only be retrieved after rnn block")
+            raise ValueError(("Output of the dynamic RNN can only be visited "
+                              "outside the rnn block."))
        if len(self.outputs) == 1:
            return self.outputs[0]
        else:

--- a/python/paddle/v2/fluid/layers/nn.py
+++ b/python/paddle/v2/fluid/layers/nn.py
@@ -9,12 +9,33 @@ from ..param_attr import ParamAttr
 from tensor import concat

 __all__ = [
-    'fc', 'embedding', 'dynamic_lstm', 'gru_unit', 'linear_chain_crf',
-    'crf_decoding', 'cos_sim', 'cross_entropy', 'square_error_cost', 'accuracy',
-    'chunk_eval', 'sequence_conv', 'conv2d', 'sequence_pool', 'pool2d',
-    'batch_norm', 'beam_search_decode', 'conv2d_transpose', 'sequence_expand',
-    'lstm_unit', 'reduce_sum', 'reduce_mean', 'reduce_max', 'reduce_min',
-    'sequence_first_step', 'sequence_last_step'
+    'fc',
+    'embedding',
+    'dynamic_lstm',
+    'gru_unit',
+    'linear_chain_crf',
+    'crf_decoding',
+    'cos_sim',
+    'cross_entropy',
+    'square_error_cost',
+    'accuracy',
+    'chunk_eval',
+    'sequence_conv',
+    'conv2d',
+    'sequence_pool',
+    'pool2d',
+    'batch_norm',
+    'beam_search_decode',
+    'conv2d_transpose',
+    'sequence_expand',
+    'lstm_unit',
+    'reduce_sum',
+    'reduce_mean',
+    'reduce_max',
+    'reduce_min',
+    'sequence_first_step',
+    'sequence_last_step',
+    'dropout',
 ]


@@ -248,13 +269,13 @@ def gru_unit(input,
            h_t & = dot((1-u_t), m_t) + dot(u_t, h_{t-1})

    The inputs of gru unit includes :math:`z_t`, :math:`h_{t-1}`. In terms
-    of the equation above, the :math:`z_t` is split into 3 parts - 
-    :math:`xu_t`, :math:`xr_t` and :math:`xm_t`. This means that in order to 
-    implement a full GRU unit operator for an input, a fully 
+    of the equation above, the :math:`z_t` is split into 3 parts -
+    :math:`xu_t`, :math:`xr_t` and :math:`xm_t`. This means that in order to
+    implement a full GRU unit operator for an input, a fully
    connected layer has to be applied, such that :math:`z_t = W_{fc}x_t`.

-    The terms :math:`u_t` and :math:`r_t` represent the update and reset gates 
-    of the GRU cell. Unlike LSTM, GRU has one lesser gate. However, there is 
+    The terms :math:`u_t` and :math:`r_t` represent the update and reset gates
+    of the GRU cell. Unlike LSTM, GRU has one lesser gate. However, there is
    an intermediate candidate hidden output, which is denoted by :math:`m_t`.
    This layer has three outputs :math:`h_t`, :math:`dot(r_t, h_{t-1})`
    and concatenation of :math:`u_t`, :math:`r_t` and :math:`m_t`.
@@ -276,7 +297,7 @@ def gru_unit(input,
        .. code-block:: python

             # assuming we have x_t_data and prev_hidden of size=10
-             x_t = fluid.layers.fc(input=x_t_data, size=30) 
+             x_t = fluid.layers.fc(input=x_t_data, size=30)
             hidden_val, r_h_val, gate_val = fluid.layers.gru_unit(input=x_t,
                                                    hidden = prev_hidden)

@@ -386,6 +407,21 @@ def cos_sim(X, Y, **kwargs):
    return out


+def dropout(x, dropout_prob, is_test=False, seed=0, **kwargs):
+    helper = LayerHelper('dropout', **kwargs)
+    out = helper.create_tmp_variable(dtype=x.dtype)
+    mask = helper.create_tmp_variable(dtype=x.dtype, stop_gradient=True)
+    helper.append_op(
+        type='dropout',
+        inputs={'X': [x]},
+        outputs={'Out': [out],
+                 'Mask': [mask]},
+        attrs={'dropout_prob': dropout_prob,
+               'is_test': is_test,
+               'seed': seed})
+    return out
+
+
 def cross_entropy(input, label, **kwargs):
    """
    **Cross Entropy Layer**
@@ -968,7 +1004,7 @@ def batch_norm(input,
        default_initializer=Constant(1.0))

    bias = helper.create_parameter(
-        attr=helper.param_attr, shape=param_shape, dtype=dtype, is_bias=True)
+        attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)

    mean = helper.create_global_variable(
        dtype=input.dtype,

--- a/python/paddle/v2/fluid/layers/ops.py
+++ b/python/paddle/v2/fluid/layers/ops.py
 from ..registry import register_layer

 __activations__ = [
+    'sigmoid',
+    'logsigmoid',
+    'exp',
+    'relu',
+    'tanh',
+    'tanh_shrink',
+    'softshrink',
+    'sqrt',
    'abs',
    'ceil',
-    'exp',
    'floor',
-    'log',
-    'relu',
    'round',
-    'sigmoid',
-    'sqrt',
+    'reciprocal',
+    'log',
    'square',
-    'tanh',
+    'softplus',
+    'softsign',
+    'brelu',
+    'leaky_relu',
+    'soft_relu',
+    'elu',
+    'relu6',
+    'pow',
+    'stanh',
+    'hard_shrink',
+    'thresholded_relu',
+    'hard_sigmoid',
+    'swish',
 ]

 __all__ = [
    'mean',
    'mul',
-    'dropout',
    'reshape',
    'scale',
    'transpose',

--- a/python/paddle/v2/fluid/layers/tensor.py
+++ b/python/paddle/v2/fluid/layers/tensor.py
 from ..layer_helper import LayerHelper
 from ..param_attr import ParamAttr
+from ..framework import convert_np_dtype_to_dtype_
+from ..framework import Variable
+from ..core import DataType
+import numpy

 __all__ = [
-    'create_tensor', 'create_parameter', 'cast', 'concat', 'sums', 'assign',
-    'fill_constant_batch_size_like', 'fill_constant', 'ones', 'zeros'
+    'create_tensor',
+    'create_parameter',
+    'cast',
+    'concat',
+    'sums',
+    'assign',
+    'fill_constant_batch_size_like',
+    'fill_constant',
+    'ones',
+    'zeros',
 ]


@@ -121,7 +133,7 @@ def assign(input, output):
    This function copies the *input* Variable to the *output* Variable.

    Args:
-        input(Variable): The source variable
+        input(Variable|numpy.ndarray): The source variable
        output(Variable): The destination variable

    Returns:
@@ -134,11 +146,37 @@ def assign(input, output):
          fluid.layers.assign(hidden, out)
    """
    helper = LayerHelper('assign', **locals())
-    helper.append_op(
-        type='scale',
-        inputs={'X': [input]},
-        outputs={'Out': [output]},
-        attrs={'scale': 1.0})
+    if isinstance(input, Variable):
+        helper.append_op(
+            type='scale',
+            inputs={'X': [input]},
+            outputs={'Out': [output]},
+            attrs={'scale': 1.0})
+    elif isinstance(input, numpy.ndarray):
+        dtype = convert_np_dtype_to_dtype_(input.dtype)
+        if dtype == DataType.FP32:
+            value_name = "fp32_values"
+            values = [float(v) for v in input.flat]
+        elif dtype == DataType.INT32:
+            value_name = "int32_values"
+            values = [int(v) for v in input.flat]
+        else:
+            raise ValueError("Unsupported dtype %s", input.dtype)
+        if input.size > 1024 * 1024:
+            raise ValueError("The size of input is too big. Please consider "
+                             "saving it to file and 'load_op' to load it")
+
+        helper.append_op(
+            type='assign_value',
+            outputs={'Out': [output]},
+            attrs={
+                'dtype': dtype,
+                'shape': list(input.shape),
+                value_name: values
+            })
+    else:
+        raise ValueError("Wrong type for assign input: %s" % type(input))
+
    return output


@@ -146,25 +184,26 @@ def fill_constant(shape, dtype, value, out=None):
    """
    **fill_constant**

-    This function creates a tensor of specified *shape* and
-    *dtype*, and initializes this with a constant supplied in *value*.
+    This function creates a tensor with specified `shape` and `dtype`, and
+    initializes it with a constant specifed by `value`.

-    It also sets *stop_gradient* to True.
+    The attribute `stop_gradient` of the created tensor is set to True.

    Args:
-        shape(tuple|list|None): Shape of output tensor
-        dtype(np.dtype|core.DataType|str): Data type of output tensor
-        value(float): Constant value to initialize the output tensor
-        out(Variable): Output Variable to initialize
+        shape(tuple|list|None): Shape of the output tensor.
+        dtype(np.dtype|core.DataType|str): Data type of the output tensor.
+        value(float): The constant value used to initialize the output tensor.
+        out(Variable): The output tensor.

    Returns:
-        Variable: The tensor variable storing the output
+        Variable: The tensor variable storing the output.

    Examples:
        .. code-block:: python

          data = fluid.layers.fill_constant(shape=[1], value=0, dtype='int64')
    """
+
    helper = LayerHelper("fill_constant", **locals())
    if out is None:
        out = helper.create_tmp_variable(dtype=dtype)

--- a/python/paddle/v2/fluid/memory_optimization_transpiler.py
+++ b/python/paddle/v2/fluid/memory_optimization_transpiler.py
@@ -3,6 +3,17 @@ import framework
 from framework import Program, default_main_program, Parameter, Variable
 import backward
 from backward import _rename_arg_
+from . import core
+
+dtype_to_size = {
+    core.DataType.FP16: 2,
+    core.DataType.FP32: 4,
+    core.DataType.FP64: 8,
+    core.DataType.INT16: 2,
+    core.DataType.INT32: 4,
+    core.DataType.INT64: 8,
+    core.DataType.BOOL: 1
+}


 class ControlFlowGraph(object):
@@ -28,18 +39,33 @@ class ControlFlowGraph(object):
        block_size = program_desc.num_blocks()

        # TODO(qijun) handle Program with if/while operators
-        self.global_block = program_desc.block(0)
-        self.op_size = self.global_block.op_size()
+        self.global_block_desc = program_desc.block(0)
+        self.op_size = self.global_block_desc.op_size()

        op_node_connections = [(i, i + 1) for i in range(self.op_size - 1)]
        self._add_connections(op_node_connections)

-        self.ops = [self.global_block.op(i) for i in range(self.op_size)]
+        self.ops = [self.global_block_desc.op(i) for i in range(self.op_size)]

        for i in range(self.op_size):
            self._uses[i].update(self.ops[i].input_arg_names())
            self._defs[i].update(self.ops[i].output_arg_names())

+    def _update_graph(self, old_name, new_name, begin_idx=0):
+        for i in range(begin_idx, self.op_size):
+            if old_name in self._uses[i]:
+                self._uses[i].remove(old_name)
+                self._uses[i].add(new_name)
+            if old_name in self._defs[i]:
+                self._defs[i].remove(old_name)
+                self._defs[i].add(new_name)
+            if old_name in self._live_in[i]:
+                self._live_in[i].remove(old_name)
+                self._live_out[i].add(new_name)
+            if old_name in self._live_out[i]:
+                self._live_out[i].remove(old_name)
+                self._live_out[i].add(new_name)
+
    def _reach_fixed_point(self, live_in, live_out):
        if len(live_in) != len(self._live_in):
            return False
@@ -79,30 +105,47 @@ class ControlFlowGraph(object):
        self.pool = []
        for i in range(self.op_size):
            if self.pool:
-                out_pair = [(x, self.global_block.var(str(x)).shape())
+                out_pair = [(x, self.global_block_desc.var(str(x)).shape())
                            for x in self._defs[i]]
                for x, x_shape in out_pair:
-                    for index, cache_pair in enumerate(self.pool):
-                        cache_var = cache_pair[0]
-                        cache_shape = cache_pair[1]
-                        if x_shape == cache_shape:
-                            print(
-                                "Hit Cache !!!! cache pool index is %d, var name is %s, cached var name is %s, var shape is %s "
-                                % (index, x, cache_var, str(cache_shape)))
-                            self.pool.pop(index)
-                            _rename_arg_(self.ops, x, cache_var, begin_idx=i)
-                            self._dataflow_analyze()
-                            break
+                    if not self.global_block_desc.var(str(x)).persistable():
+                        for index, cache_pair in enumerate(self.pool):
+                            cache_var = cache_pair[0]
+                            cache_shape = cache_pair[1]
+                            if x_shape == cache_shape:
+                                x_dtype = self.global_block_desc.var(str(
+                                    x)).dtype()
+                                cache_dtype = self.global_block_desc.var(
+                                    str(cache_var)).dtype()
+                                # TODO(qijun): actually, we should compare dtype_to_size[x_dtype]
+                                # and dtype_to_size[cache_dtype]
+                                if x_dtype == cache_dtype:
+                                    print(
+                                        ("Hit Cache !!!! cache pool index "
+                                         "is %d, var name is %s, "
+                                         "cached var name is %s, "
+                                         "var shape is %s ") %
+                                        (index, x, cache_var, str(cache_shape)))
+                                    self.pool.pop(index)
+                                    _rename_arg_(
+                                        self.ops, x, cache_var, begin_idx=i)
+                                    self._program.current_block().var(str(
+                                        x)).desc = self.global_block_desc.var(
+                                            str(cache_var))
+                                    self._update_graph(
+                                        x, cache_var, begin_idx=i)
+                                    break

            in_diff, out_diff = self._get_diff(self._live_in[i],
                                               self._live_out[i])
            can_optimize = filter(
-                lambda x: not self.global_block.var(str(x)).persistable(),
+                lambda x: not self.global_block_desc.var(str(x)).persistable(),
                in_diff)
            if can_optimize:
                for var_name in can_optimize:
-                    self.pool.append((
-                        var_name, self.global_block.var(str(var_name)).shape()))
+                    self.pool.append(
+                        (var_name,
+                         self.global_block_desc.var(str(var_name)).shape()))

    def get_program(self):
        return self._program

--- a/python/paddle/v2/fluid/nets.py
+++ b/python/paddle/v2/fluid/nets.py
 import layers

-__all__ = ["simple_img_conv_pool", "sequence_conv_pool"]
+__all__ = [
+    "simple_img_conv_pool",
+    "sequence_conv_pool",
+]


 def simple_img_conv_pool(input,

--- a/python/paddle/v2/fluid/registry.py
+++ b/python/paddle/v2/fluid/registry.py
@@ -8,7 +8,11 @@ import proto.framework_pb2 as framework_pb2
 from framework import OpProtoHolder, Variable, Program, Operator
 from paddle.v2.fluid.layer_helper import LayerHelper, unique_name

-__all__ = ['deprecated', 'register_layer', 'autodoc']
+__all__ = [
+    'deprecated',
+    'register_layer',
+    'autodoc',
+]


 def _convert_(name):
@@ -80,11 +84,10 @@ def _generate_doc_string_(op_proto):


 def register_layer(op_type):
-    """
-    Register an Python layer for an Operator
+    """Register the Python layer for an Operator.

    Args:
-       op_type: The name of the operator to be created
+       op_type: The name of the operator to be created.

    This function takes in the operator type (sigmoid, mean , average etc) and
    creates the operator functionality.
@@ -98,16 +101,16 @@ def register_layer(op_type):

    if len(not_intermediate_outputs) != 1:
        raise ValueError("Only one non intermediate output operator can be",
-                         "automatically generated")
+                         "automatically generated.")

    if not_intermediate_outputs[0].duplicable:
        raise ValueError(
-            "Only non duplicable op can be automatically generated")
+            "Only non duplicable op can be automatically generated.")

    for output in intermediate_outputs:
        if output.duplicable:
            raise ValueError("The op can be automatically generated only when ",
-                             "all intermediate ops are not duplicable")
+                             "all intermediate ops are not duplicable.")

    o_name = not_intermediate_outputs[0].name
    intermediate_output_names = [output.name for output in intermediate_outputs]

--- a/python/paddle/v2/fluid/regularizer.py
+++ b/python/paddle/v2/fluid/regularizer.py
 import framework

-__all__ = ['append_regularization_ops', 'L1Decay', 'L2Decay']
+__all__ = [
+    'append_regularization_ops',
+    'L1Decay',
+    'L2Decay',
+]


 def append_regularization_ops(parameters_and_grads, regularization=None):

--- a/python/paddle/v2/fluid/tests/book_distribute/test_dist_fit_a_line.py
+++ b/python/paddle/v2/fluid/tests/book_distribute/test_dist_fit_a_line.py
+import numpy as np
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+import os
+
+x = fluid.layers.data(name='x', shape=[13], dtype='float32')
+
+y_predict = fluid.layers.fc(input=x, size=1, act=None)
+
+y = fluid.layers.data(name='y', shape=[1], dtype='float32')
+
+cost = fluid.layers.square_error_cost(input=y_predict, label=y)
+avg_cost = fluid.layers.mean(x=cost)
+
+sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
+optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
+
+BATCH_SIZE = 20
+
+train_reader = paddle.batch(
+    paddle.reader.shuffle(
+        paddle.dataset.uci_housing.train(), buf_size=500),
+    batch_size=BATCH_SIZE)
+
+place = fluid.CPUPlace()
+feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
+exe = fluid.Executor(place)
+
+t = fluid.DistributeTranspiler()
+# all parameter server endpoints list for spliting parameters
+pserver_endpoints = os.getenv("PSERVERS")
+# server endpoint for current node
+current_endpoint = os.getenv("SERVER_ENDPOINT")
+# run as trainer or parameter server
+training_role = os.getenv("TRAINING_ROLE",
+                          "TRAINER")  # get the training role: trainer/pserver
+t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2)
+
+if training_role == "PSERVER":
+    if not current_endpoint:
+        print("need env SERVER_ENDPOINT")
+        exit(1)
+    pserver_prog = t.get_pserver_program(current_endpoint, optimize_ops)
+    exe.run(fluid.default_startup_program())
+    exe.run(pserver_prog)
+else:
+    trainer_prog = t.get_trainer_program()
+
+    exe.run(fluid.default_startup_program())
+
+    PASS_NUM = 100
+    for pass_id in range(PASS_NUM):
+        fluid.io.save_persistables(exe, "./fit_a_line.model/")
+        fluid.io.load_persistables(exe, "./fit_a_line.model/")
+        for data in train_reader():
+            avg_loss_value, = exe.run(trainer_prog,
+                                      feed=feeder.feed(data),
+                                      fetch_list=[avg_cost])
+
+            if avg_loss_value[0] < 10.0:
+                exit(0)
+exit(1)
--- a/python/paddle/v2/fluid/tests/book_distribute/test_dist_label_semantic_roles.py
+++ b/python/paddle/v2/fluid/tests/book_distribute/test_dist_label_semantic_roles.py
--- a/python/paddle/v2/fluid/tests/book_distribute/test_understand_sentiment_conv_dist.py
+++ b/python/paddle/v2/fluid/tests/book_distribute/test_understand_sentiment_conv_dist.py
+from __future__ import print_function
+import os
+import numpy as np
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+
+
+def convolution_net(data, label, input_dim, class_dim=2, emb_dim=32,
+                    hid_dim=32):
+    emb = fluid.layers.embedding(input=data, size=[input_dim, emb_dim])
+    conv_3 = fluid.nets.sequence_conv_pool(
+        input=emb,
+        num_filters=hid_dim,
+        filter_size=3,
+        act="tanh",
+        pool_type="sqrt")
+    conv_4 = fluid.nets.sequence_conv_pool(
+        input=emb,
+        num_filters=hid_dim,
+        filter_size=4,
+        act="tanh",
+        pool_type="sqrt")
+    prediction = fluid.layers.fc(input=[conv_3, conv_4],
+                                 size=class_dim,
+                                 act="softmax")
+    cost = fluid.layers.cross_entropy(input=prediction, label=label)
+    avg_cost = fluid.layers.mean(x=cost)
+    adam_optimizer = fluid.optimizer.Adam(learning_rate=0.002)
+    optimize_ops, params_grads = adam_optimizer.minimize(avg_cost)
+    accuracy = fluid.evaluator.Accuracy(input=prediction, label=label)
+    return avg_cost, accuracy, accuracy.metrics[0], optimize_ops, params_grads
+
+
+def to_lodtensor(data, place):
+    seq_lens = [len(seq) for seq in data]
+    cur_len = 0
+    lod = [cur_len]
+    for l in seq_lens:
+        cur_len += l
+        lod.append(cur_len)
+    flattened_data = np.concatenate(data, axis=0).astype("int64")
+    flattened_data = flattened_data.reshape([len(flattened_data), 1])
+    res = fluid.LoDTensor()
+    res.set(flattened_data, place)
+    res.set_lod([lod])
+    return res
+
+
+def main():
+    BATCH_SIZE = 100
+    PASS_NUM = 5
+
+    word_dict = paddle.dataset.imdb.word_dict()
+    dict_dim = len(word_dict)
+    class_dim = 2
+
+    data = fluid.layers.data(
+        name="words", shape=[1], dtype="int64", lod_level=1)
+    label = fluid.layers.data(name="label", shape=[1], dtype="int64")
+    cost, accuracy, acc_out, optimize_ops, params_grads = convolution_net(
+        data, label, input_dim=dict_dim, class_dim=class_dim)
+
+    train_data = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.imdb.train(word_dict), buf_size=1000),
+        batch_size=BATCH_SIZE)
+    place = fluid.CPUPlace()
+    exe = fluid.Executor(place)
+
+    t = fluid.DistributeTranspiler()
+
+    # all parameter server endpoints list for spliting parameters
+    pserver_endpoints = os.getenv("PSERVERS")
+    # server endpoint for current node
+    current_endpoint = os.getenv("SERVER_ENDPOINT")
+    # run as trainer or parameter server
+    training_role = os.getenv(
+        "TRAINING_ROLE", "TRAINER")  # get the training role: trainer/pserver
+    t.transpile(
+        optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2)
+
+    exe.run(fluid.default_startup_program())
+
+    if training_role == "PSERVER":
+        if not current_endpoint:
+            print("need env SERVER_ENDPOINT")
+            exit(1)
+        pserver_prog = t.get_pserver_program(current_endpoint, optimize_ops)
+        exe.run(pserver_prog)
+    elif training_role == "TRAINER":
+        trainer_prog = t.get_trainer_program()
+        feeder = fluid.DataFeeder(feed_list=[data, label], place=place)
+
+        for pass_id in xrange(PASS_NUM):
+            accuracy.reset(exe)
+            for data in train_data():
+                cost_val, acc_val = exe.run(trainer_prog,
+                                            feed=feeder.feed(data),
+                                            fetch_list=[cost, acc_out])
+                pass_acc = accuracy.eval(exe)
+                print("cost=" + str(cost_val) + " acc=" + str(acc_val) +
+                      " pass_acc=" + str(pass_acc))
+                if cost_val < 1.0 and pass_acc > 0.8:
+                    exit(0)
+    else:
+        print("environment var TRAINER_ROLE should be TRAINER os PSERVER")
+
+
+if __name__ == '__main__':
+    main()
--- a/python/paddle/v2/fluid/tests/test_assign_value_op.py
+++ b/python/paddle/v2/fluid/tests/test_assign_value_op.py
+import paddle.v2.fluid as fluid
+import paddle.v2.fluid.layers as layers
+import op_test
+import numpy
+import unittest
+import paddle.v2.fluid.framework as framework
+
+
+class TestAssignValueOp(op_test.OpTest):
+    def setUp(self):
+        self.op_type = "assign_value"
+        x = numpy.random.random(size=(2, 5)).astype(numpy.float32)
+        self.inputs = {}
+        self.outputs = {'Out': x}
+        self.attrs = {
+            'shape': x.shape,
+            'dtype': framework.convert_np_dtype_to_dtype_(x.dtype),
+            'fp32_values': [float(v) for v in x.flat]
+        }
+
+    def test_forward(self):
+        self.check_output()
+
+    def test_assign(self):
+        val = (
+            -100 + 200 * numpy.random.random(size=(2, 5))).astype(numpy.int32)
+        x = layers.create_tensor(dtype="float32")
+        layers.assign(input=val, output=x)
+        exe = fluid.Executor(fluid.CPUPlace())
+        fetched_x = exe.run(fluid.default_main_program(),
+                            feed={},
+                            fetch_list=[x])[0]
+        self.assertTrue(
+            numpy.array_equal(fetched_x, val),
+            "fetch_x=%s val=%s" % (fetched_x, val))
+        self.assertEqual(fetched_x.dtype, val.dtype)
+
+
+if __name__ == '__main__':
+    unittest.main()
--- a/python/paddle/v2/fluid/tests/test_parallel_op.py
+++ b/python/paddle/v2/fluid/tests/test_parallel_op.py
--- a/python/paddle/v2/fluid/tests/test_print_op.py
+++ b/python/paddle/v2/fluid/tests/test_print_op.py
+import unittest
+import numpy as np
+from paddle.v2.fluid.executor import Executor
+import paddle.v2.fluid.core as core
+import paddle.v2.fluid.layers as pd
+
+
+class TestSumOp(unittest.TestCase):
+    def test_tensor(self):
+        i = pd.zeros(shape=[2, 10], dtype='float32')
+
+        pd.Print(i, message="I am a message", summarize=10)
+
+        cpu = core.CPUPlace()
+        exe = Executor(cpu)
+
+        exe.run()
+
+
+if __name__ == '__main__':
+    unittest.main()
--- a/python/paddle/v2/fluid/tests/test_sequence_erase_op.py
+++ b/python/paddle/v2/fluid/tests/test_sequence_erase_op.py
--- a/tools/manylinux1/Dockerfile.android
+++ b/tools/manylinux1/Dockerfile.android
--- a/tools/manylinux1/Dockerfile.x64
+++ b/tools/manylinux1/Dockerfile.x64
--- a/tools/manylinux1/README.md
+++ b/tools/manylinux1/README.md
--- a/tools/manylinux1/build_all.sh
+++ b/tools/manylinux1/build_all.sh
--- a/tools/manylinux1/build_scripts/build.sh
+++ b/tools/manylinux1/build_scripts/build.sh
--- a/tools/manylinux1/build_scripts/build_utils.sh
+++ b/tools/manylinux1/build_scripts/build_utils.sh
--- a/tools/manylinux1/build_scripts/manylinux1-check.py
+++ b/tools/manylinux1/build_scripts/manylinux1-check.py
--- a/tools/manylinux1/build_scripts/python-tag-abi-tag.py
+++ b/tools/manylinux1/build_scripts/python-tag-abi-tag.py
+# Utility script to print the python tag + the abi tag for a Python
+# See PEP 425 for exactly what these are, but an example would be:
+#   cp27-cp27mu
+
+from wheel.pep425tags import get_abbr_impl, get_impl_ver, get_abi_tag
+
+print("{0}{1}-{2}".format(get_abbr_impl(), get_impl_ver(), get_abi_tag()))
--- a/tools/manylinux1/build_scripts/ssl-check.py
+++ b/tools/manylinux1/build_scripts/ssl-check.py