Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fix-7195

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.

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 
+"""

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

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

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 
+"""

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

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

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()

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()

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:
 

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)

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
 

paddle/gserver/layers/MKLDNNLayer.cpp

@@ -132,6 +132,8 @@ void MKLDNNLayer::reshapeInput(int& batchsize,
   if (w != 0) {
     width = w;
   }
+  height = height != 0 ? height : 1;
+  width = width != 0 ? width : 1;
 }
 
 void MKLDNNLayer::reshapeOutput(size_t height, size_t width) {

paddle/gserver/layers/MKLDNNLayer.h

@@ -98,6 +98,8 @@ protected:
 public:
   explicit MKLDNNLayer(const LayerConfig& config)
       : Layer(config),
+        ih_(0),
+        iw_(0),
         condition_(0),
         needResetBwd_(true),
         outputOnlyMKLDNN_(false),

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)

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

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

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

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

paddle/operators/detail/recv_impl.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 "send_recv_impl.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;
-  }
-  condition_.notify_all();
-}
-
-}  // namespace detail
-}  // namespace operators
-}  // namespace paddle

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.

paddle/operators/detail/send_recv_impl.h → paddle/operators/detail/sendrecvop_utils.cc

@@ -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:

paddle/operators/detail/send_impl.cc → paddle/operators/detail/sendrecvop_utils.h

@@ -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

paddle/operators/edit_distance_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/edit_distance_op.h"
+
+namespace paddle {
+namespace operators {
+
+class EditDistanceOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Hyps"), "Input(Hyps) shouldn't be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Refs"), "Input(Refs) shouldn't be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) shouldn't be null.");
+    auto hyp_dims = ctx->GetInputDim("Hyps");
+    auto ref_dims = ctx->GetInputDim("Refs");
+    PADDLE_ENFORCE(hyp_dims.size() == 2 && hyp_dims[1] == 1,
+                   "Input(Hyps) must be a 2-D LoDTensor with the 2nd dimension "
+                   "equal to 1.");
+    PADDLE_ENFORCE(ref_dims.size() == 2 && ref_dims[1] == 1,
+                   "Input(Refs) must be a 2-D LoDTensor with the 2nd dimension "
+                   "equal to 1.");
+    ctx->SetOutputDim("Out", ctx->GetInputDim("Refs"));
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    return framework::OpKernelType(framework::proto::DataType::FP32,
+                                   ctx.device_context());
+  }
+};
+
+class EditDistanceOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  EditDistanceOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Hyps",
+             "(2-D LoDTensor<int>, 2nd dim. equal to 1) "
+             "The indices for hypothesis strings.");
+    AddInput("Refs",
+             "(2-D LoDTensor<int>, 2nd dim. equal to 1) "
+             "The indices for reference strings.");
+    AddAttr<bool>("normalized",
+                  "(bool, default false) Indicated whether to normalize "
+                  "the edit distance by the length of reference string.")
+        .SetDefault(false);
+    AddOutput("Out",
+              "(2-D Tensor with shape [`batch_size` x 1]) "
+              "The output edit distances of EditDistance operator.");
+    AddComment(R"DOC(
+
+EditDistance operator computes the edit distances between a batch of hypothesis
+strings and their references.
+
+Edit distance, also called Levenshtein distance, measures how dissimilar two strings 
+are by counting the minimum number of operations to transform one string into anthor. 
+Here the operations include insertion, deletion, and substitution. For example, 
+given hypothesis string A = "kitten" and reference B = "sitting", the edit distance 
+is 3 for A will be transformed into B at least after two substitutions and one 
+insertion:
+  
+   "kitten" -> "sitten" -> "sittin" -> "sitting"
+
+Input(Hyps) is a LoDTensor consisting of all the hypothesis strings with the total 
+number denoted by `batch_size`, and the separation is specified by the LoD information. 
+And the `batch_size` reference strings are arranged in order in the same way in the 
+LoDTensor Input(Refs).
+
+Output(Out) contains the `batch_size` results and each stands for the edit stance 
+for a pair of strings respectively. If Attr(normalized) is true, the edit distance 
+will be divided by the length of reference string.
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OPERATOR(edit_distance, ops::EditDistanceOp, ops::EditDistanceOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    edit_distance, ops::EditDistanceKernel<paddle::platform::CPUPlace, float>);

paddle/operators/edit_distance_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 <algorithm>
+#include "paddle/framework/op_registry.h"
+#include "paddle/platform/cuda_helper.h"
+#include "paddle/platform/gpu_info.h"
+
+namespace paddle {
+namespace operators {
+
+using platform::PADDLE_CUDA_NUM_THREADS;
+
+template <typename T>
+__global__ void FillFirstRow(T* dist, const int N) {
+  int idx = blockDim.x * blockIdx.x + threadIdx.x;
+  if (idx < N + 1) {
+    dist[idx] = idx;
+  }
+}
+
+template <typename T>
+__global__ void FillFirstColumn(T* dist, const int M, const int N) {
+  int idx = blockDim.x * blockIdx.x + threadIdx.x;
+  if (idx < M + 1) {
+    dist[idx * (N + 1)] = idx;
+  }
+}
+
+template <typename T>
+__global__ void Levenshtein(T* dist, const int* x1, const int* x2, const int M,
+                            const int N, const int start) {
+  int idx = blockDim.x * blockIdx.x + threadIdx.x;
+  int offset = N;
+  int index = start + idx * offset;
+  int row = index / (N + 1);
+  int col = index % (N + 1);
+  if (row > 0 && col > 0 && row < M + 1 && col < N + 1) {
+    int cost = x1[row - 1] == x2[col - 1] ? 0 : 1;
+    int dels = dist[(row - 1) * (N + 1) + col] + 1;
+    int ins = dist[row * (N + 1) + col - 1] + 1;
+    int subs = dist[(row - 1) * (N + 1) + (col - 1)] + cost;
+    dist[index] = min(dels, min(ins, subs));
+  }
+}
+
+template <typename T>
+__global__ void SetOutput(T* out, const T* dist, const int M, const int N,
+                          bool normalized) {
+  int idx = blockDim.x * blockIdx.x + threadIdx.x;
+  if (idx == 0) {
+    out[0] = normalized ? dist[M * (N + 1) + N] / N : dist[M * (N + 1) + N];
+  }
+}
+
+template <typename Place, typename T>
+class EditDistanceGPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const {
+    auto* out_t = ctx.Output<framework::Tensor>("Out");
+
+    auto* x1_t = ctx.Input<framework::LoDTensor>("Hyps");
+    auto* x2_t = ctx.Input<framework::LoDTensor>("Refs");
+
+    auto normalized = ctx.Attr<bool>("normalized");
+    auto stream = reinterpret_cast<const platform::CUDADeviceContext&>(
+                      ctx.device_context())
+                      .stream();
+
+    auto hyp_lod = x1_t->lod()[0];
+    auto ref_lod = x2_t->lod()[0];
+    PADDLE_ENFORCE(
+        hyp_lod.size() == ref_lod.size(),
+        "Input(Hyps) and Input(Refs) must have the same batch size.");
+    for (size_t i = 1; i < ref_lod.size(); ++i) {
+      PADDLE_ENFORCE(ref_lod[i] > ref_lod[i - 1],
+                     "Reference string %d is empty.", i);
+    }
+
+    auto num_strs = hyp_lod.size() - 1;
+    out_t->Resize({static_cast<int64_t>(num_strs), 1});
+    out_t->mutable_data<T>(ctx.GetPlace());
+    auto out = out_t->data<T>();
+
+    T distance = 0.0;
+    for (size_t num = 0; num < num_strs; num++) {
+      auto m = static_cast<int64_t>(hyp_lod[num + 1] - hyp_lod[num]);
+      auto n = static_cast<int64_t>(ref_lod[num + 1] - ref_lod[num]);
+      if (m == 0 || n == 0) {
+        distance = std::max(m, n);
+        if (normalized) {
+          PADDLE_ENFORCE(n > 0,
+                         "The reference string (#%d) cannot be empty "
+                         "when Attr(normalized) is enabled.",
+                         n);
+          distance = distance / n;
+        }
+        memory::Copy(boost::get<Place>(ctx.GetPlace()), out + num,
+                     platform::CPUPlace(), &distance, sizeof(T), stream);
+      } else {
+        framework::Tensor dist_t;
+        dist_t.Resize({m + 1, n + 1});
+        dist_t.mutable_data<T>(ctx.GetPlace());
+        auto dist = dist_t.data<T>();
+        auto x1 = x1_t->data<int>() + hyp_lod[num];
+        auto x2 = x2_t->data<int>() + ref_lod[num];
+
+        FillFirstColumn<T><<<1 + m / PADDLE_CUDA_NUM_THREADS,
+                             PADDLE_CUDA_NUM_THREADS, 0, stream>>>(dist, m, n);
+
+        FillFirstRow<T><<<1 + n / PADDLE_CUDA_NUM_THREADS,
+                          PADDLE_CUDA_NUM_THREADS, 0, stream>>>(dist, n);
+        // Compute the elements of distance matrix in the anti-diagonal diretion
+        for (int64_t slice = 2; slice < m + n + 1; ++slice) {
+          int z_m = slice < m + 1 ? 0 : slice - m;
+          int z_n = slice < n + 1 ? 0 : slice - n;
+          int size = slice - (z_m + z_n) + 1;  // number of elments in the same
+                                               // anti-diagonal line to update
+          // the start index at which computes from
+          int start = slice < n + 1 ? slice : (z_n + 1) * (n + 1) - 1;
+          Levenshtein<T><<<1 + (size - 1) / PADDLE_CUDA_NUM_THREADS,
+                           PADDLE_CUDA_NUM_THREADS, 0, stream>>>(dist, x1, x2,
+                                                                 m, n, start);
+        }
+        SetOutput<T><<<1, 1, 0, stream>>>(out + num, dist, m, n, normalized);
+      }
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_CUDA_KERNEL(
+    edit_distance,
+    ops::EditDistanceGPUKernel<paddle::platform::CUDAPlace, float>);

paddle/operators/edit_distance_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 <algorithm>
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class EditDistanceKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const {
+    auto* out_t = ctx.Output<framework::Tensor>("Out");
+
+    auto* x1_t = ctx.Input<framework::LoDTensor>("Hyps");
+    auto* x2_t = ctx.Input<framework::LoDTensor>("Refs");
+
+    auto normalized = ctx.Attr<bool>("normalized");
+
+    auto hyp_lod = x1_t->lod()[0];
+    auto ref_lod = x2_t->lod()[0];
+    PADDLE_ENFORCE(
+        hyp_lod.size() == ref_lod.size(),
+        "Input(Hyps) and Input(Refs) must have the same batch size.");
+    for (size_t i = 1; i < ref_lod.size(); ++i) {
+      PADDLE_ENFORCE(ref_lod[i] > ref_lod[i - 1],
+                     "Reference string %d is empty.", i);
+    }
+    auto num_strs = hyp_lod.size() - 1;
+
+    out_t->Resize({static_cast<int64_t>(num_strs), 1});
+    out_t->mutable_data<float>(ctx.GetPlace());
+    auto out = out_t->data<T>();
+
+    T distance = 0.0;
+    for (size_t num = 0; num < num_strs; ++num) {
+      auto m = static_cast<int64_t>(hyp_lod[num + 1] - hyp_lod[num]);
+      auto n = static_cast<int64_t>(ref_lod[num + 1] - ref_lod[num]);
+
+      if (m == 0) {
+        distance = n;
+      } else if (n == 0) {
+        distance = m;
+      } else {
+        framework::Tensor dist_t;
+        dist_t.Resize({m + 1, n + 1});
+        dist_t.mutable_data<T>(ctx.GetPlace());
+        auto dist = dist_t.data<T>();
+        auto x1 = x1_t->data<int>() + hyp_lod[num];
+        auto x2 = x2_t->data<int>() + ref_lod[num];
+        for (int64_t i = 0; i < m + 1; ++i) {
+          dist[i * (n + 1)] = i;
+        }
+        for (int64_t j = 0; j < n + 1; ++j) {
+          dist[j] = j;
+        }
+        for (int64_t i = 1; i < m + 1; ++i) {
+          for (int64_t j = 1; j < n + 1; ++j) {
+            int cost = x1[i - 1] == x2[j - 1] ? 0 : 1;
+            int dels = dist[(i - 1) * (n + 1) + j] + 1;
+            int ins = dist[i * (n + 1) + (j - 1)] + 1;
+            int subs = dist[(i - 1) * (n + 1) + (j - 1)] + cost;
+            dist[i * (n + 1) + j] = std::min(dels, std::min(ins, subs));
+          }
+        }
+        distance = dist[m * (n + 1) + n];
+      }
+
+      if (normalized) {
+        PADDLE_ENFORCE(n > 0,
+                       "The reference string (#%d) cannot be empty "
+                       "when Attr(normalized) is enabled.",
+                       n);
+        distance = distance / n;
+      }
+      out[num] = distance;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

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_;
 };

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];
-      }
-    }
-    // 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];
+
+    // FIXME(gongwb): DeviceContext?
+    auto ctx = platform::CPUDeviceContext();
+    for (size_t i = 0; i < ins.size(); i++) {
+      client_.AsyncSendVariable(epmap[i], ctx, scope, ins[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")

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;

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>);

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>);

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

paddle/scripts/submit_local.sh.in

@@ -92,6 +92,9 @@ function threads_config() {
     if [ -z "$OPENBLAS_NUM_THREADS" ]; then
       export OPENBLAS_NUM_THREADS=$threads
     fi
+    if [ $threads -gt 1 ] && [ -z "$OPENBLAS_MAIN_FREE" ]; then
+      export OPENBLAS_MAIN_FREE=1
+    fi
   fi
   
 }

python/paddle/v2/__init__.py

@@ -62,12 +62,15 @@ __all__ = [
 cp.begin_parse()
 
 
-def set_omp_mkl_env_vars(trainer_count):
+def set_env_vars(trainer_count):
     '''Auto set CPU environment if have not set before.
+       For MKL:
          export KMP_AFFINITY, OMP_DYNAMIC according to the Hyper Threading status.
          export OMP_NUM_THREADS, MKL_NUM_THREADS according to trainer_count.
+       For OpenBLAS:
+         export OPENBLAS_NUM_THREADS, OPENBLAS_MAIN_FREE according to trainer_count. 
     '''
-    import platform
+    import platform, paddle
     if not platform.system() in ['Linux', 'Darwin']:
         return
 
@@ -103,6 +106,7 @@ def set_omp_mkl_env_vars(trainer_count):
 
     num_cores = num_physical_cores()
     num_processors = num_logical_processors()
+    if paddle.version.mkl() == 'ON':
         if num_processors > num_cores:  # Hyper Threading is enabled
             set_env("OMP_DYNAMIC", "true")
             set_env("KMP_AFFINITY", "granularity=fine,compact,1,0")
@@ -111,8 +115,13 @@ def set_omp_mkl_env_vars(trainer_count):
             set_env("KMP_AFFINITY", "granularity=fine,compact,0,0")
     threads = num_processors / trainer_count
     threads = '1' if threads < 1 else str(threads)
+    if paddle.version.mkl() == 'ON':
         set_env("OMP_NUM_THREADS", threads)
         set_env("MKL_NUM_THREADS", threads)
+    else:
+        set_env("OPENBLAS_NUM_THREADS", threads)
+        if threads > 1:
+            set_env("OPENBLAS_MAIN_FREE", '1')
 
 
 def init(**kwargs):
@@ -129,7 +138,7 @@ def init(**kwargs):
     for key in args_dict.keys():
         args.append('--%s=%s' % (key, str(args_dict[key])))
 
-    set_omp_mkl_env_vars(kwargs.get('trainer_count', 1))
+    set_env_vars(kwargs.get('trainer_count', 1))
 
     if 'use_gpu' in kwargs:
         cp.g_command_config_args['use_gpu'] = kwargs['use_gpu']

python/paddle/v2/fluid/framework.py

@@ -236,6 +236,9 @@ class Variable(object):
 
     __repr__ = __str__
 
+    def set_desc(self, input):
+        self.desc = input
+
     @property
     def persistable(self):
         return self.desc.persistable()

python/paddle/v2/fluid/layers/nn.py

@@ -14,7 +14,7 @@ __all__ = [
     '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'
+    'sequence_first_step', 'sequence_last_step', 'dropout'
 ]
 
 
@@ -386,6 +386,21 @@ 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**

python/paddle/v2/fluid/layers/ops.py

 from ..registry import register_layer
 
 __activations__ = [
-    'abs',
-    'ceil',
-    'exp',
-    'floor',
-    'log',
-    'relu',
-    'round',
-    'sigmoid',
-    'sqrt',
-    'square',
-    'tanh',
+    'abs', 'tanh', 'sigmoid', 'relu', 'sqrt', 'ceil', 'floor', 'log', 'round'
 ]
 
 __all__ = [
     'mean',
     'mul',
-    'dropout',
     'reshape',
     'scale',
     'transpose',

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 +105,45 @@ 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:
+                    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)))
+                                        %
+                                        (index, x, cache_var, str(cache_shape)))
                                     self.pool.pop(index)
-                            _rename_arg_(self.ops, x, cache_var, begin_idx=i)
-                            self._dataflow_analyze()
+                                    _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

python/paddle/v2/fluid/tests/test_edit_distance_op.py

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+def Levenshtein(hyp, ref):
+    """ Compute the Levenshtein distance between two strings.
+
+    :param hyp: hypothesis string in index
+    :type hyp: list
+    :param ref: reference string in index
+    :type ref: list
+    """
+    m = len(hyp)
+    n = len(ref)
+    if m == 0:
+        return n
+    if n == 0:
+        return m
+
+    dist = np.zeros((m + 1, n + 1)).astype("float32")
+    for i in range(0, m + 1):
+        dist[i][0] = i
+    for j in range(0, n + 1):
+        dist[0][j] = j
+
+    for i in range(1, m + 1):
+        for j in range(1, n + 1):
+            cost = 0 if hyp[i - 1] == ref[j - 1] else 1
+            deletion = dist[i - 1][j] + 1
+            insertion = dist[i][j - 1] + 1
+            substitution = dist[i - 1][j - 1] + cost
+            dist[i][j] = min(deletion, insertion, substitution)
+    return dist[m][n]
+
+
+class TestEditDistanceOp(OpTest):
+    def setUp(self):
+        self.op_type = "edit_distance"
+        normalized = False
+        x1 = np.array([[0, 12, 3, 5, 8, 2]]).astype("int32")
+        x2 = np.array([[0, 12, 4, 7, 8]]).astype("int32")
+        x1 = np.transpose(x1)
+        x2 = np.transpose(x2)
+        x1_lod = [0, 1, 5]
+        x2_lod = [0, 3, 4]
+
+        num_strs = len(x1_lod) - 1
+        distance = np.zeros((num_strs, 1)).astype("float32")
+        for i in range(0, num_strs):
+            distance[i] = Levenshtein(
+                hyp=x1[x1_lod[i]:x1_lod[i + 1]],
+                ref=x2[x2_lod[i]:x2_lod[i + 1]])
+            if normalized is True:
+                len_ref = x2_lod[i + 1] - x2_lod[i]
+                distance[i] = distance[i] / len_ref
+        self.attrs = {'normalized': normalized}
+        self.inputs = {'Hyps': (x1, [x1_lod]), 'Refs': (x2, [x2_lod])}
+        self.outputs = {'Out': distance}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestEditDistanceOpNormalized(OpTest):
+    def setUp(self):
+        self.op_type = "edit_distance"
+        normalized = True
+        x1 = np.array([[0, 10, 3, 6, 5, 8, 2]]).astype("int32")
+        x2 = np.array([[0, 10, 4, 6, 7, 8]]).astype("int32")
+        x1 = np.transpose(x1)
+        x2 = np.transpose(x2)
+        x1_lod = [0, 1, 3, 6]
+        x2_lod = [0, 2, 3, 5]
+
+        num_strs = len(x1_lod) - 1
+        distance = np.zeros((num_strs, 1)).astype("float32")
+        for i in range(0, num_strs):
+            distance[i] = Levenshtein(
+                hyp=x1[x1_lod[i]:x1_lod[i + 1]],
+                ref=x2[x2_lod[i]:x2_lod[i + 1]])
+            if normalized is True:
+                len_ref = x2_lod[i + 1] - x2_lod[i]
+                distance[i] = distance[i] / len_ref
+        self.attrs = {'normalized': normalized}
+        self.inputs = {'Hyps': (x1, [x1_lod]), 'Refs': (x2, [x2_lod])}
+        self.outputs = {'Out': distance}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/fluid/tests/test_sequence_erase_op.py

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+def sequence_erase(in_seq, lod0, tokens):
+    new_lod0 = [0]
+    out_seq = []
+    for i in range(0, len(lod0) - 1):
+        num_out = 0
+        for dat in in_seq[lod0[i]:lod0[i + 1]]:
+            if dat not in tokens:
+                out_seq.append(dat)
+                num_out += 1
+        new_lod0.append(new_lod0[-1] + num_out)
+    return np.array(out_seq).astype("int32"), new_lod0
+
+
+class TestSequenceEraseOp(OpTest):
+    def setUp(self):
+        self.op_type = "sequence_erase"
+        in_seq = np.random.randint(0, 10, (30, 1)).astype("int32")
+        lod = [[0, 9, 13, 24, 30]]
+        tokens = [2, 3, 5]
+        out_seq, new_lod0 = sequence_erase(in_seq, lod[0], tokens)
+        self.attrs = {'tokens': tokens}
+        self.inputs = {'X': (in_seq, lod)}
+        self.outputs = {'Out': (out_seq, [new_lod0])}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == '__main__':
+    unittest.main()

python/setup.py.in

@@ -31,6 +31,7 @@ patch           = '%(patch)d'
 rc              = '%(rc)d'
 istaged         = %(istaged)s
 commit          = '%(commit)s'
+with_mkl        = '%(with_mkl)s'
 
 def show():
     if istaged:
@@ -41,6 +42,9 @@ def show():
         print 'rc:', rc
     else:
         print 'commit:', commit
+
+def mkl():
+    return with_mkl
 '''
     commit = git_commit()
     with open(filename, 'w') as f:
@@ -51,7 +55,8 @@ def show():
             'rc': RC,
             'version': '${PADDLE_VERSION}',
             'commit': commit,
-            'istaged': ISTAGED})
+            'istaged': ISTAGED,
+            'with_mkl': '@WITH_MKL@'})
 
 write_version_py(filename='@PADDLE_SOURCE_DIR@/python/paddle/version.py')