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

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()

benchmark/IntelOptimizedPaddle.md

@@ -7,11 +7,11 @@ Machine:
 
 System: CentOS release 6.3 (Final), Docker 1.12.1.
 
-PaddlePaddle: (TODO: will rerun after 0.11.0)
-- paddlepaddle/paddle:latest (for MKLML and MKL-DNN)
+PaddlePaddle:
+- paddlepaddle/paddle:0.11.0 (for MKLML and MKL-DNN)
   - MKL-DNN tag v0.11
   - MKLML 2018.0.1.20171007
-- paddlepaddle/paddle:latest-openblas (for OpenBLAS)
+- paddlepaddle/paddle:0.11.0-openblas (for OpenBLAS)
   - OpenBLAS v0.2.20
 	 
 On each machine, we will test and compare the performance of training on single node using MKL-DNN / MKLML / OpenBLAS respectively.
@@ -56,15 +56,15 @@ Input image size - 3 * 224 * 224, Time: images/second
 
 <img src="figs/googlenet-cpu-train.png" width="500">
 
-- Alexnet
+- AlexNet
 
 | BatchSize    | 64     | 128    | 256    |
 |--------------|--------| ------ | -------|
-| OpenBLAS     | 2.13   | 2.45   | 2.68   | 
+| OpenBLAS     | 45.62  | 72.79  | 107.22 | 
 | MKLML        | 66.37  | 105.60 | 144.04 |
 | MKL-DNN      | 399.00 | 498.94 | 626.53 | 
 
-chart TBD
+<img src="figs/alexnet-cpu-train.png" width="500">
 
 #### Inference
 Test on batch size 1, 2, 4, 8, 16 on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
@@ -72,36 +72,41 @@ Test on batch size 1, 2, 4, 8, 16 on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
 
 | BatchSize | 1     | 2     | 4     | 8     | 16    |
 |-----------|-------|-------|-------|-------|-------|
-| OpenBLAS  | 1.07  | 1.08  | 1.06  | 0.88  | 0.65  |
+| OpenBLAS  | 1.10  | 1.96  | 3.62  | 3.63  | 2.25  |
 | MKLML     | 5.58  | 9.80  | 15.15 | 21.21 | 28.67 |
 | MKL-DNN   | 75.07 | 88.64 | 82.58 | 92.29 | 96.75 |
 
+<img src="figs/vgg-cpu-infer.png" width="500">
+
 - ResNet-50
 
 | BatchSize | 1     | 2      | 4      | 8      | 16     |
 |-----------|-------|--------|--------|--------|--------|
-| OpenBLAS  | 3.35  | 3.19   | 3.09   | 2.55   | 1.96   |
+| OpenBLAS  | 3.31  | 6.72   | 11.59  | 13.17  | 9.27   |
 | MKLML     | 6.33  | 12.02  | 22.88  | 40.53  | 63.09  |
 | MKL-DNN   | 107.83| 148.84 | 177.78 | 189.35 | 217.69 |
 
+<img src="figs/resnet-cpu-infer.png" width="500">
 
 - GoogLeNet
 
 | BatchSize | 1      | 2      | 4      | 8      | 16     |
 |-----------|--------|--------|--------|--------|--------|
-| OpenBLAS  | 12.04  | 11.31  | 10.00  | 9.07   | 4.34   |
+| OpenBLAS  | 12.06  | 23.56  | 34.48  | 36.45  | 23.12  |
 | MKLML     | 22.74  | 41.56  | 81.22  | 133.47 | 210.53 |
 | MKL-DNN   | 175.10 | 272.92 | 450.70 | 512.00 | 600.94 |
 
-- Alexnet
+<img src="figs/googlenet-cpu-infer.png" width="500">
+
+- AlexNet
 
 | BatchSize | 1      | 2      | 4      | 8      | 16     |
 |-----------|--------|--------|--------|--------|--------|
-| OpenBLAS  |    |   |   |   |    |
+| OpenBLAS  | 3.53   | 6.23   | 15.04  | 26.06  | 31.62  |
 | MKLML     | 21.32  | 36.55  | 73.06  | 131.15 | 192.77 |
 | MKL-DNN   | 442.91 | 656.41 | 719.10 | 847.68 | 850.51 |
 
-chart TBD
+<img src="figs/alexnet-cpu-infer.png" width="500">
 
 ### Laptop
 TBD

benchmark/cluster/README.md

+# Cluster Training Benchmark
+
+## Setup
+
+- Platform
+  - Kubernetes: v1.6.2
+  - Linux Kernel: v3.10.0
+
+- Resource
+  - CPU: 10 Cores per Pod
+  - Memory: 5GB per Pod
+
+- Docker Image
+
+  We use different base Docker Image to run the benchmark on Kubernetes:
+  - PaddlePaddle v2: paddlepaddle/paddle:0.11.0
+  - PaddlePaddle Fluid: paddlepaddle/paddle:[commit-id]
+  - TensorFlow: tensorflow/tensorflow:1.5.0-rc0
+
+- Model
+  vgg16 is used in this benchmark.
+
+## Cases
+
+- Variable
+  - Batch Size of training data.
+  - PServer count of the training job.
+  - The number of trainers.
+
+- Invariant
+  - The resource of trainer/pserver Pod.
+
+### Measure the Performance for Different Batch Size
+
+- PServer Count: 40
+- Trainer Count: 100
+- Metrics: mini-batch / sec
+
+| Batch Size | 32 | 64 | 128 | 256 |
+| -- | -- | -- | -- | -- |
+| PaddlePaddle Fluid | - | - | - | - |
+| PaddlePaddle v2 | - | - | - | - |
+| TensorFlow | - | - | - | - |
+
+### Measure the Performance for Different PServer Count
+
+- Trainer Count: 100
+- Batch Size: 64
+- Metrics: mini-batch / sec
+
+| PServer Count | 10 | 20 | 40 | 60 |
+| -- | -- | -- | -- | -- |
+| PaddlePaddle Fluid | - | - | - | - |
+| PaddlePaddle v2 | - | - | - | - |
+| TensorFlow | - | - | - | - |
+
+### Measure Parallel Efficiency By Increasing Trainer Count
+
+- PServer Count: 20
+- Batch Size: 64
+- Metrics:
+
+$S = \div(T1, TN)$
+
+which S is the ratio of T1 over TN, training time of 1 and N trainers.
+The parallel efficiency is:
+
+$E = \div(S, N)$
+
+| Trainer Counter | 1 | 10 | 20 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 |
+| -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- |
+| PaddlePaddle Fluid | - | - | - | - | - | - | - | - | - | - | - |
+| PaddlePaddle v2 | - | - | - | - | - | - | - | - | - | - | - | - |
+| TensorFlow | - | - | - | - | - | - | - | - | - | - | - | - | - |
+
+## Reproduce the benchmark
+
+TODO

benchmark/paddle/image/plotlog.py

+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import sys
+import argparse
+import matplotlib.pyplot as plt
+
+
+def parse_args():
+    parser = argparse.ArgumentParser('Parse Log')
+    parser.add_argument(
+        '--file_path', '-f', type=str, help='the path of the log file')
+    parser.add_argument(
+        '--sample_rate',
+        '-s',
+        type=float,
+        default=1.0,
+        help='the rate to take samples from log')
+    parser.add_argument(
+        '--log_period', '-p', type=int, default=1, help='the period of log')
+
+    args = parser.parse_args()
+    return args
+
+
+def parse_file(file_name):
+    loss = []
+    error = []
+    with open(file_name) as f:
+        for i, line in enumerate(f):
+            line = line.strip()
+            if not line.startswith('pass'):
+                continue
+            line_split = line.split(' ')
+            if len(line_split) != 5:
+                continue
+
+            loss_str = line_split[2][:-1]
+            cur_loss = float(loss_str.split('=')[-1])
+            loss.append(cur_loss)
+
+            err_str = line_split[3][:-1]
+            cur_err = float(err_str.split('=')[-1])
+            error.append(cur_err)
+
+    accuracy = [1.0 - err for err in error]
+
+    return loss, accuracy
+
+
+def sample(metric, sample_rate):
+    interval = int(1.0 / sample_rate)
+    if interval > len(metric):
+        return metric[:1]
+
+    num = len(metric) / interval
+    idx = [interval * i for i in range(num)]
+    metric_sample = [metric[id] for id in idx]
+    return metric_sample
+
+
+def plot_metric(metric,
+                batch_id,
+                graph_title,
+                line_style='b-',
+                line_label='y',
+                line_num=1):
+    plt.figure()
+    plt.title(graph_title)
+    if line_num == 1:
+        plt.plot(batch_id, metric, line_style, label=line_label)
+    else:
+        for i in range(line_num):
+            plt.plot(batch_id, metric[i], line_style[i], label=line_label[i])
+    plt.xlabel('batch')
+    plt.ylabel(graph_title)
+    plt.legend()
+    plt.savefig(graph_title + '.jpg')
+    plt.close()
+
+
+def main():
+    args = parse_args()
+    assert args.sample_rate > 0. and args.sample_rate <= 1.0, "The sample rate should in the range (0, 1]."
+
+    loss, accuracy = parse_file(args.file_path)
+    batch = [args.log_period * i for i in range(len(loss))]
+
+    batch_sample = sample(batch, args.sample_rate)
+    loss_sample = sample(loss, args.sample_rate)
+    accuracy_sample = sample(accuracy, args.sample_rate)
+
+    plot_metric(loss_sample, batch_sample, 'loss', line_label='loss')
+    plot_metric(
+        accuracy_sample,
+        batch_sample,
+        'accuracy',
+        line_style='g-',
+        line_label='accuracy')
+
+
+if __name__ == '__main__':
+    main()

benchmark/paddle/image/run_openblas_infer.sh

@@ -8,6 +8,7 @@ function clock_to_seconds() {
 }
 
 function infer() {
+  export OPENBLAS_MAIN_FREE=1
   topology=$1
   layer_num=$2
   bs=$3

cmake/external/mkldnn.cmake

@@ -63,9 +63,30 @@ ExternalProject_Add(
                         -DMKLROOT:PATH=${MKLML_ROOT}
 )
 
-ADD_LIBRARY(mkldnn SHARED IMPORTED GLOBAL)
-SET_PROPERTY(TARGET mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
-ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
+ADD_LIBRARY(shared_mkldnn SHARED IMPORTED GLOBAL)
+SET_PROPERTY(TARGET shared_mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
+ADD_DEPENDENCIES(shared_mkldnn ${MKLDNN_PROJECT})
 MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}")
 add_definitions(-DPADDLE_WITH_MKLDNN)
-LIST(APPEND external_project_dependencies mkldnn)
+LIST(APPEND external_project_dependencies shared_mkldnn)
+
+# generate a static dummy target to track mkldnn dependencies
+# for cc_library(xxx SRCS xxx.c DEPS mkldnn)
+SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/mkldnn_dummy.c)
+FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
+ADD_LIBRARY(mkldnn STATIC ${dummyfile})
+TARGET_LINK_LIBRARIES(mkldnn ${MKLDNN_LIB} ${MKLML_LIB} ${MKLML_IOMP_LIB})
+ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
+
+# copy the real so.0 lib to install dir
+# it can be directly contained in wheel or capi
+SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0)
+ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB}
+    COMMAND cp ${MKLDNN_LIB} ${MKLDNN_SHARED_LIB}
+    DEPENDS mkldnn)
+ADD_CUSTOM_TARGET(mkldnn_shared_lib ALL DEPENDS ${MKLDNN_SHARED_LIB})
+
+IF(WITH_C_API)
+  INSTALL(FILES ${MKLDNN_SHARED_LIB} DESTINATION lib)
+ENDIF()
+

cmake/external/mklml.cmake

@@ -66,3 +66,7 @@ ADD_LIBRARY(mklml SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET mklml PROPERTY IMPORTED_LOCATION ${MKLML_LIB})
 ADD_DEPENDENCIES(mklml ${MKLML_PROJECT})
 LIST(APPEND external_project_dependencies mklml)
+
+IF(WITH_C_API)
+  INSTALL(FILES ${MKLML_LIB} ${MKLML_IOMP_LIB} DESTINATION lib)
+ENDIF()

cmake/external/warpctc.cmake

@@ -63,7 +63,7 @@ ExternalProject_Add(
 MESSAGE(STATUS "warp-ctc library: ${WARPCTC_LIBRARIES}")
 INCLUDE_DIRECTORIES(${WARPCTC_INCLUDE_DIR})
 
-ADD_LIBRARY(warpctc STATIC IMPORTED GLOBAL)
+ADD_LIBRARY(warpctc SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET warpctc PROPERTY IMPORTED_LOCATION ${WARPCTC_LIBRARIES})
 ADD_DEPENDENCIES(warpctc extern_warpctc)
 

doc/api/v2/fluid.rst

@@ -15,4 +15,4 @@ Fluid
     fluid/param_attr.rst
     fluid/profiler.rst
     fluid/regularizer.rst
-
+    fluid/io.rst

doc/api/v2/fluid/io.rst

+===========
+IO
+===========
+
+
+
+is_parameter
+-----------
+..  autofunction:: paddle.v2.fluid.io.is_parameter
+    :noindex:

doc/api/v2/fluid/layers.rst

@@ -38,6 +38,16 @@ elementwise_add
 ..  autofunction:: paddle.v2.fluid.layers.elementwise_add
     :noindex:
 
+elementwise_sub
+---------------
+..  autofunction:: paddle.v2.fluid.layers.elementwise_sub
+    :noindex:
+
+elementwise_mul
+---------------
+..  autofunction:: paddle.v2.fluid.layers.elementwise_mul
+    :noindex:
+
 elementwise_div
 ---------------
 ..  autofunction:: paddle.v2.fluid.layers.elementwise_div

doc/design/block.md

@@ -202,8 +202,8 @@ This `OpDesc` value is in the `ops` field of the `BlockDesc` value representing
 
 During the generation of the Protobuf message, the Block should store VarDesc (the Protobuf message which describes Variable) and OpDesc (the Protobuf message which describes Operator).
 
-VarDesc in a block should have its name scope to avoid local variables affect parent block's name scope.
-Child block's name scopes should inherit the parent's so that OpDesc in child block can reference a VarDesc that stored in parent block. For example:
+VarDesc in a block should have its name scope to avoid local variables affecting parent block's name scope.
+Child block's name scopes should inherit the parent's so that OpDesc in child block can reference a VarDesc that is stored in the parent block. For example:
 
 ```python
 a = pd.Variable(shape=[20, 20])

doc/design/refactor/distributed_architecture.md → doc/design/dist_refactor/distributed_architecture.md

@@ -52,8 +52,9 @@ The IR for PaddlePaddle after refactoring is called a `Block`, it specifies the
 
 The user can not directly specify the parameter update rule for the parameter server in the Python module, since the parameter server does not use the same computation definition as the trainer. Instead, the update rule is baked inside the parameter server. The user can not specify the update rule explicitly.
 
-This could be fixed by making the parameter server run the same computation definition as the trainer (the user's Python module). For a detailed explanation, refer to this document -
-[Design Doc: Operation Graph Based Parameter Server](./parameter_server.md)
+This could be fixed by making the parameter server also run an IR, which can be different to the trainer side
+For a detailed explanation, refer to this document -
+[Design Doc: Parameter Server](./parameter_server.md)
 
 ## Distributed Training Architecture
 
@@ -61,68 +62,111 @@ The revamped distributed training architecture can address the above discussed l
 
 <img src="src/distributed_architecture.png"/>
 
-The major components in the architecture are: *PaddlePaddle Python*, *PaddlePaddle converter* and *PaddlePaddle runtime*.
+The major components are: *Python API*, *Distribute Transpiler* and *Remote Executor*.
 
-### PaddlePaddle Python
+### Python API
 
-PaddlePaddle Python is the Python library that user's Python code invokes, to read the data. build the neural network topology, start training, etc.
+Python API is the Python library that user's Python code invokes, to read the data, build the neural network topology, and start training, etc.
 
 ```Python
-paddle.init()
-input = paddle.op.recordIO("/home/data/mnist.recordio") # file stored on the cluster
-img, label = input[0], input[1]
-hidden = paddle.layer.fc(input=img, size=200, act=paddle.activation.Tanh())
-prediction = paddle.layer.fc(input=img, size=10, act=paddle.activation.Softmax())
-cost = paddle.layer.classification_cost(input=prediction, label=label)
-optimizer = paddle.optimizer.SGD(cost, learning_rate=0.01)
-session = paddle.session.NewRemote(num_trainer=3, num_ps=2, GPU_per_trainer=1)
-for i in range(1000):
-	_, cost_val = session.eval(targets=[cost, optimizer])
-	print cost_val
+images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype='float32')
+label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+...
+predict = fluid.layers.fc(input=conv_pool_2, size=10, act="softmax")
+cost = fluid.layers.cross_entropy(input=predict, label=label)
+avg_cost = fluid.layers.mean(x=cost)
+optimizer = fluid.optimizer.Adam(learning_rate=0.01)
+optimizer.minimize(avg_cost)
+
+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)
+
+for pass_id in range(10):
+    for data in train_reader():
+        loss, acc = exe.run(trainer_prog,
+                            feed=feeder.feed(data),
+                            fetch_list=[avg_cost])
 ```
 
-The above code is what a typical Python trainer code is, the neural network topology is built using the helper functions such as `paddle.layer.fc`. Training is done by calling `session.eval` iteratively.
-
-#### session.eval
-
-As shown in the graph, `session.eval` sends the IR and the evaluation inputs or targets to the PaddlePaddle cluster for evaluation.
-The targets can be any variable in the computation graph. When the target is say, the `optimizer` variable, the neural network will be optimized once. When the target is the `cost` variable, `session.eval` returns the cost value. Based on what the target is, an appropriate action is taken.
-
-The Python `session` is a wrapper of the C++ `Session` class. For more information about `Session`, refer to this document - [Design Doc: Session](./session.md).
-
-### PaddlePaddle Converter
-
-The PaddlePaddle converter automatically converts the IR in the request (IR and evaluation inputs/targets) from PaddlePaddle Python to partitioned IRs and dispatches the new IRs and evaluation inputs/targets to different PaddlePaddle runtimes. Below are the steps that are followed :
-
-1. Add a `feed` OP that feeds the eval inputs, and a `fetch` OP that fetches the eval targets to the IR.
-
-2. Extract a new computation (sub)graph with the `feed` and `fetch` OPs as the boundary. The runtime does not need to run the OP that is not dependent on the `fetch` OP.
-
-3. Optimize the computation graph.
-
-4. Place the OPs in the graph onto different devices on different PaddlePaddle runtime according to a placement algorithm and the device constraints specified by the user.
-
-5. Partition the graph according to runtime boundaries and add `send` / `recv` OP pair on the runtime boundaries.
+The code above is a typical local training program, the "Training Program" is built using helper functions such as
+`fluid.layer.fc`. The training is done by calling `Executor.run`
+iteratively.
+
+For more details, the implementation of IR is [Program](../program.md), and `ProgramDesc` is the protobuf type.
+
+[Executor](../executor.md) simply runs the `ProgramDesc`. For local training you generally use
+`Executor` to run the program locally. For any kind of distributed training, you can use
+`RemoteExecutor` to specify desired distributed training method with some optional arguments.
+
+### Distributed Transpiler
+
+The Distributed Transpiler automatically converts the IR (in protobuf format) to partitioned IRs. Then
+the Remote Executor dispatches the new IRs to Remote Executors across the cluster.
+Below are the steps that are followed :
+
+1. User only need to change `Executor` to `RemoteExecutor` to change local program to distributed program.
+1. `RemoteExecutor` calls `Distributed Transpiler` to "transpile" user's program to several IRs representing a
+   distributed training program:
+   1. Parse configurations from `RemoteExecutor`.
+   1. Determine the type of distributed program, can be DataParallelism, ModelParallelism or Streaming.
+   1. Partition the `ProgramDesc` according to type and add `send` / `recv` OP pair on the boundaries. Take
+      DataParallelism type for example, it removes the optimization operators and add a `send` OP to the
+      "trainer" role, then add the optimization operators to the parameter server role within the `recv` OP.
+1. Dispatch the partitioned graph to different `RemoteExecutor` in the cluster.
+1. `RemoteExecutor` on each node run the received `ProgramDesc` utill the end.
+
+
+### RemoteExecutor
+
+As shown in the graph, `RemoteExecutor.run` sends the IR to the cluster for Execution.
+You can also use parameter `fetch_list` to interactively fetch variable back to local for
+log printing.
+
+The Python `RemoteExecutor` is derived from `Executor` class.
+
+```python
+exe = RemoteExecutor(
+    feed=feeder.feed(data),
+    fetch_list=[avg_cost],
+    job_desc=JobDesc(
+      jobname,
+      num_trainer,
+      num_pserver,
+      cpu_per_trainer,
+      gpu_per_trainer,
+      mem_per_trainer,
+      cpu_per_pserver,
+      mem_per_pserver
+    ))
+for data in train_reader():
+    loss, acc = exe.run(trainer_prog,
+                        feed=feeder.feed(data),
+                        fetch_list=[avg_cost])
+```
 
-6. Dispatch the partitioned graph to different PaddlePaddle runtimes.
+`JobDesc` object describe the distributed job resource specification to run on
+Cluster environment.
 
-7. PaddlePaddle runtimes with the `fetch` OP reports evaluation results back to the converter, the converter reports the evaluation results back to the PaddlePaddle Python.
+<img src="src/remote_executor.png"/>
 
-The output IRs will be cached to optimize the conversion latency.
+`RemoteExecutor.run` sends the `ProgramDesc` and
+[TrainingJob](https://github.com/PaddlePaddle/cloud/blob/develop/doc/autoscale/README.md#training-job-resource)
+to a server in the cluster which executes `RemoteExecutor.listen`. This server is responsible
+to start the final Kubernetes Jobs to run the different role of `ProgramDesc`.
 
 
-#### Placement Algorithm
+### Placement Algorithm
 
 Our first implementation will only support "trainer-parameter server" placement: the parameters, initializers, and optimizers are all placed on the PaddlePaddle runtimes with the parameter server role. Everything else will be placed on the PaddlePaddle runtimes with the trainer role. This has the same functionality as the "trainer-parameter server" architecture of PaddlePaddle v0.10.0, but is more generic and flexible.
 
 In the future, a more general placement algorithm should be implemented, which makes placements according to the input IR, and a model of device computation time and device communication time. Model parallelism requires the generic placement algorithm.
 
 
-### PaddlePaddle Runtime
-
-The PaddlePaddle runtime owns multiple devices (e.g., CPUs, GPUs) and runs the IR. The runtime does not need to do OP placement since it is already done by the converter.
-
-
 ### Local Training Architecture
 
 The local training architecture will be the same as the distributed training architecture, the difference is that everything runs locally, and there is just one PaddlePaddle runtime:
@@ -132,9 +176,18 @@ The local training architecture will be the same as the distributed training arc
 
 ### Training Data
 
-In PaddlePaddle v0.10.0, training data is typically read with a [data reader](../reader/README.md) from Python. This approach is no longer efficient when training in a distributed fashion since the Python process no longer runs on the same node with the trainer processes. The Python reader will need to read from the distributed filesystem (assuming it has the required access) and send to the trainers, doubling the network traffic.
-
-When doing distributed training, the user can still use Python data reader: the training data are sent with `session.eval`. However this should be used for debugging purpose only. The users are encouraged to use the read data OPs.
+In PaddlePaddle v0.10.0, training data is typically read
+with [data reader](../reader/README.md) from Python. This approach is
+no longer efficient when training distributedly since the Python
+process no longer runs on the same node with the trainer processes,
+the Python reader will need to read from the distributed filesystem
+(assuming it has the access) and send to the trainers, doubling the
+network traffic.
+
+When doing distributed training, the user can still use Python data
+reader: the training data are sent with `Executor.run`. However, should
+be used for debugging purpose only. The users are encouraged to use
+the read data OPs.
 
 
 ## References:

doc/design/refactor/parameter_server.md → doc/design/dist_refactor/parameter_server.md

-# Design Doc: Operation Graph Based Parameter Server
+# Design Doc: Parameter Server
 
 ## Abstract
 
@@ -10,7 +10,7 @@ different purposes.
 ## Background
 
 The previous implementations of the parameter server does not run a
-subgraph. parameter initialization, optimizer computation, network
+fluid sub-program. Parameter initialization, optimizer computation, network
 communication and checkpointing are implemented twice on both the
 trainer and the parameter server.
 
@@ -23,10 +23,10 @@ server becomes a natural extension.
 
 ## Design
 
-### Graph Converter
+### Distributed Transpiler
 
-The *graph converter* converts the user-defined operation (OP) graph
-into subgraphs to be scheduled on different nodes with the following
+The *Distributed Transpiler* converts the user-defined fluid program
+into sub-programs to be scheduled on different nodes with the following
 steps:
 
 1. OP placement: the OPs will be placed on different nodes according
@@ -34,7 +34,6 @@ steps:
    time. Currently we will use a simple heuristic that puts parameter
    varable on parameter server workers and everything else on trainer
    workers.
-
 1. Add communication OPs to enable the communication between nodes.
 
 We will need these OPs: *Send*, *Recv*, *Enqueue*, *Dequeue*.
@@ -48,8 +47,8 @@ After converting:
 
 <img src="src/dist-graph.png" width="700"/>
 
-1. The parameter variable W and it's optimizer subgraph are placed on the parameter server.
-1. Operators are added to the subgraphs.
+1. The parameter variable W and it's optimizer program are placed on the parameter server.
+1. Operators are added to the program.
    - *Send* sends data to the connected *Recv* operator.  The
 	 scheduler on the receive node will only schedule *Recv* operator
 	 to run when the *Send* operator has ran (the *Send* OP will mark
@@ -64,39 +63,30 @@ After converting:
 ### Benefits
 
 - Model parallelism become easier to implement: it's an extension to
-  the trainer - parameter server approach. we already have the
-  communication OPs, but need to extend the graph converter's
-  placement functionality.
-
+  the trainer - parameter server approach. We can have several "Transpilers"
+  to achieve different goals.
 - User-defined optimizer is easier to add - user can now express it as
-  a subgraph.
-
+  a sub-program.
 - No more duplication logic inside the trainer and the parameter
   server mentioned in the background section.
 
 ### Challenges
 
-- It might be hard for the graph converter to cut a general graph
-  (without any hint for which subgraph is the optimizer). We may need
-  to label which subgraph inside the OP graph is the optimizer.
-
 - It's important to balance the parameter shards of on multiple
   parameter server. If a single parameter is very big (some
   word-embedding, fully connected, softmax layer), we need to
   automatically partition the single parameter onto different
   parameter servers when possible (only element-wise optimizer depends
   on the parameter variable).
+- In the "Aync SGD" figure, the "W" variable on the parameter server
+  could be read and wrote concurrently. See
+  [here](https://github.com/PaddlePaddle/Paddle/pull/6394) for more
+  details about concurrent program in fluid.
 
 ### Discussion
 
-- In the "Aync SGD" figure, the "W" variable on the parameter server
-  could be read and wrote concurrently, what is our locking strategy?
-  E.g., each variable have a lock cpp method to be invoked by every
-  OP, or, have a lock OP.
-
 - Can the Enqueue OP be implemented under our current tensor design
   (puts the input tensor into the queue tensor)?
-
 - *Dequeue* OP will have variable numbers of output (depends on the
   `min_count` attribute), does our current design support it? (similar
   question for the *Add* OP)

doc/design/error_clip.md

+# Error Clip
+
+## Overview
+
+Error clip is widely used in model training to prevent gradient exploding. It takes some specific rules to adjust variables' gradients and prevent them from being too large. With it, values of a gradient will be checked before they are taken by the next `grad_op` and be shrunk if necessary.
+## Usage
+
+Users are allowed to assign different error clip methods or attributes to different `Variable`s. Users can specify it as a parameter of `Variable`'s constructor:
+
+```python
+var = framework.Variable(..., error_clip=myErrorClip, ...)
+```
+
+The default value of `error_clip` is `None`, which means no error clip is employed. When it's not `None`, it should take an object of `BaseErrorClipAttr`'s derived class. So far, `BaseErrorClipAttr` has only one derived class: `ErrorClipByValue`, whose constructor is:
+
+```python
+ErrorClipByValue(max, min=None)
+```
+
+`max` and `min` represent the maximal and minimal clip threshold respectively. In backward pass, all values of `var`'s gradient greater than `max` or less than `min` will be clipped to `max` and `min` respectively. When the `min` is None, the minimal threshold will be assigned with `-max` automatically.
+
+So we can enable the error clip with threshold `[-5.0, 5.0]` for variable `var` by:
+
+```python
+var = framework.Variable(..., error_clip=ErrorClipByValue(max=5.0), ...)
+```
+
+## Implementation
+
+The `BaseErrorClipAttr` and its derived class `ErrorClipByValue` are defined in *clip.py*.
+
+```python
+class BaseErrorClipAttr(object):
+    def append_clip_op(self, block, grad_name):
+        raise NotImplementedError()
+
+
+class ErrorClipByValue(BaseErrorClipAttr):
+    def __init__(self, max, min=None):
+        max = float(max)
+        if min is None:
+            min = -max
+        else:
+            min = float(min)
+        self.max = max
+        self.min = min
+
+    def append_clip_op(self, block, grad_name):
+        block.append_op(
+            type="clip",
+            inputs={"X": grad_name},
+            outputs={"Out": grad_name},
+            attrs={"min": self.min,
+                   "max": self.max})
+```
+
+The `BaseErrorClipAttr` have one main member functions: `append_clip_op(self, block, grad_name)`.
+
+This function is used to create a `clip_op` and append it to the end of given `block`. For different error clip algorithm require different `clip_op`, the function is defined as virtual in the base class. All derived classes must implement their own versions of this function.
+
+These `clip_op`s should be inserted after `grad_op`s whose output gradients need to be clipped. It is equivalent to appending some `clip_op`s to the end of the target block every time a new `grad_op` is added.
+
+```python
+for op_desc in grad_op_descs:
+        new_op_desc = target_block.desc.append_op()
+        new_op_desc.copy_from(op_desc)
+        callback(block=target_block, context=grad_to_var)
+```
+
+Here we employ a callback function to complete this kind of jobs. In `_append_backward_ops_` function, each time after a `grad_op` is added to the `target_block`, a callback function is invoked. The logic of `clip_op` appending can be implemented inside the callback function.
+
+The callback function for `clip_op` appending is defined in *clip.py*:
+
+```python
+def error_clip_callback(block, context):
+    # the context is a grad_to_var map
+    grad_to_var = context
+    op_desc = block.desc.op(block.desc.op_size() - 1)
+    for grad_n in filter(lambda n: grad_to_var.has_key(n),
+                         op_desc.output_arg_names()):
+        fwd_var = block.var_recursive(grad_to_var[grad_n])
+        error_clip = getattr(fwd_var, "error_clip", None)
+        if error_clip is not None:
+            error_clip.append_clip_op(block, grad_n)
+```
+
+This function takes a `block` and a `context`(which is actually a grad\_to\_var map) as inputs. It checks each output of the last `OpDesc` in the `block`. Notice that the last `OpDesc` of the `block` must be a `grad_op` and its outputs must be some forward variables' gradients. If an output gradient's corresponding forward variable has an attribute of `error_clip`, `error_clip_callback` will call the `error_clip`'s `append_clip_op` function to append the required `clip_op` into the `block`.

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/operator_kernel_type.md

 # Design Doc: The Keys of Operator Kernel Type
 ## Problem
-An operator can have different kernel implementations, and each operator will have a map to store the related kernels. Fluid uses `OpKernelType` as a key to identify a unique Kernel. Before an operator runs, an certain kernel must be chosen by a key of `OpKernelType`. Currently, `OpKernelType` is defined as follows:
+An operator can have different kernel implementations, and each operator will have a map to store the related kernels. Fluid uses `OpKernelType` as a key to identify a unique kernel. Before an operator runs, a certain type of kernel must be chosen via a key of `OpKernelType`. Currently, `OpKernelType` is defined as follows:
 
 ```cpp
 struct OpKernelType {
@@ -10,13 +10,13 @@ struct OpKernelType {
 ```
 For more details, please refer to [codes](https://github.com/PaddlePaddle/Paddle/blob/2d5ec16bc8a09fb8e0f62c89b116b0cd1d333907/paddle/framework/operator.h#L348-L374) in github.
 
-It contains two keys, `Place` and `DataType`. And these two keys will be hashed to a unique key to represent a certain type of kernel. However, these two keys are not enough. We need a more complete representation of `OpKernelType`. 
+It contains two keys, `Place` and `DataType`. And these two keys will be hashed to a unique key to represent a certain type of kernel. However, these two keys do not provide enough information. We need a more complete representation of `OpKernelType`.
 
-We often implement a kernel of an operator with some computing library in certain device(place). Please remind that computing library and device are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices. 
+We often implement a kernel of an operator with some computing library on certain device(place). Please note that computing library and device do not have a one-to-one correspondence. A device can have a lot of computing libraries and a computing library can also support different devices.
 
-For example, Eigen library can support Nvidia GPU/AMD GPU/CPU. And MKLDNN library can support Intel CPU/Intel FPGA. Both `Place` and `Library` should be a key of `OpKernelType`.
+For example, Eigen library supports Nvidia GPU/AMD GPU/CPU and MKLDNN library supports Intel CPU/Intel FPGA. Both `Place` and `Library` should be a key of `OpKernelType`.
 
-It's obvious that different DataTypes, like fp64/fp32/int8 will have different kernels. But the data layout of a Tensor will also lead to different implementation. Please refer to the batch norm operator [kernels](https://github.com/PaddlePaddle/Paddle/blob/a948fac4d0ad7e0412d373b8aabeb711c2899563/paddle/operators/batch_norm_op.cc#L180-L209). Data Layout should also be taken into consideration.
+Different DataTypes, such as fp64/fp32/int8, will obviously have different kernels. But different data layout of a Tensor will also lead to different implementations. Please refer to the batch norm operator [kernels](https://github.com/PaddlePaddle/Paddle/blob/a948fac4d0ad7e0412d373b8aabeb711c2899563/paddle/operators/batch_norm_op.cc#L180-L209) as an example. Data layout should also be taken into consideration.
 
 ## Solution
 
@@ -31,17 +31,17 @@ struct OpKernelType {
 };
 ```
 
-Following is the details:
+The details are as follows:
 
 ### Place
 
-`Place` is defined as follows:
+`Place` is defined as:
 
 ```cpp
 typedef boost::variant<CUDAPlace, ROCmPlace, FPGAPlace, CPUPlace> Place;
 ```
 
-`Place` is to represent the device memory where data is locating.
+`Place` represents the device memory where data is located.
 
 
 ### Library
@@ -52,10 +52,10 @@ One operator kernel is usually implemented based on one library. `Library` is de
 enum Library { Plain, MKLDNN, CUDNN };
 ```
 
-We use `Plain` enumerator to represent default library. Since most operators in Fluid are implemented based on `Eigen` library, we take `Eigen` library as the `Plain` enumerator.
-A library usually has a corresponding `DeviceContext` which contains some handles needed by computation. Fluid now have two default DeviceContexts in CPU and CUDA, `CPUDeviceContext` and `CUDADeviceContext`. `CPUDeviceContext` contains a Eigen library handle and `CDUADeviceContext` contains a Eigen library handle and cuBLAS handle.
+We use `Plain` enumerator to represent default library. Since most operators in Fluid are implemented based on the `Eigen` library, we take `Eigen` library as the `Plain` enumerator.
+A library usually has a corresponding `DeviceContext` which contains some handles needed for computation. Fluid now has two default DeviceContexts for CPU and CUDA, namely, `CPUDeviceContext` and `CUDADeviceContext`. `CPUDeviceContext` contains an Eigen library handle and `CDUADeviceContext` contains an Eigen library handle and a cuBLAS handle.
 
-If we want to support new Library, a new enumerator need to be added to `Library` and a new corresponding `LibraryDeviceContext` will be created.
+If we want to support new library, a new enumerator need to be added to `Library` and a corresponding new `LibraryDeviceContext` need to be created.
 
 
 ### DataType
@@ -67,15 +67,15 @@ If we want to support new Library, a new enumerator need to be added to `Library
 
 Actually, a Tensor is a view of a block of memory. Besides a pointer to the memory, we also have to get some other descriptions of this block of memory, such as shape(ddim), stride, and layout.
 
-Different layout leads to different implementation of operator kernel. There are mainly 4 principles we have to follow to support layout in our fluid framework.
+Different layout leads to different implementation of the operator kernel. There are mainly 4 principles we have to follow to support layout in our Fluid framework.
 
-- We take layout as a data member of Tensor. Layout is actually a enum variable. If fluid is built with MKLDNN, then, the memory format in MKLDNN will be added into this enum variable too.
+- We take layout as a data member of Tensor. Layout is actually a enum variable. If Fluid is built with MKLDNN, then the memory format in MKLDNN will also be added into this enum variable.
 
-- Users have to set layout for input data. And some operators like fill_constant/random, also have to set layout of generating data. Of course, we can have some default layout, like NCHW.
+- Users have to set layout for input data. And some operators like fill_constant/random, also have to set layout for generating data. Of course, we can have some default layout, like NCHW.
 
-- The inference of Layout is at run-time, not compile-time.
+- The inference of Layout is at run-time, not at compile-time.
 
-- Every operator have to implement different kernels for different layouts. Let's take MKLDNN as an example, if we want to implement a MKLDNN convolution operator, we have to realize all the kernels for different layout, list at [here](http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html). And we will have a special macro to do registering kernels for MKLDNN operators.
+- Every operator has to implement different kernels for different layouts. Let's take MKLDNN as an example. If we want to implement an MKLDNN convolution operator, we have to implement all the kernels for different layouts, which are listed [here](http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html). And we will have a special macro to  register kernels for MKLDNN operators.
 
 `Layout` is also defined as a enum variable:
 

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/design/refactor/session.md

-# Design Doc: Session
-
-## Abstract
-
-The *session* object encapsulates the environment in which the
-computation graph is executed.
-
-We will have the *local* session and *remote* session, they offer the
-same [interface](#interface). The local session encapsulates the local
-runtime environment and the remote session encapsulates the cluster
-runtime environment.
-
-The local runtime environment contains:
-
-1. computation devices (i.e., CPU, GPU) handles, and
-1. the [scope](../scope.md) which holds all variables.
-
-The remote runtime environment contains:
-
-1. computation devices (i.e., CPU and GPU on node 0, 1) in a cluster,
-   and
-1. the distributed [scope](../scope.md) in a cluster which holds all
-   variables.
-
-The user can create a remote session on Paddle Cloud and evaluate the
-computation graph with it. In this way, the user can control the
-remote computation resource in a cluster from his local computer.
-
-
-## Background
-
-The current design has an implicit global session in which
-`paddle.eval()` is executed. The pain point is:
-
-Since the user is not able to explicitly switch between runtime
-environments, the user cannot run a topology in two independent
-environments.
-
-For example, in reinforcement learning, the user may want to have a
-stale model for inference and a fresh model for training, and only
-replace the stale model with the fresh model periodically.
-
-Furthermore, we have no concept that encapsulates a remote environment
-that executes a computation graph.
-
-We need the session object to address above issues.
-
-
-## Session
-
-A session is an object that owns the runtime environment. All
-computations are executed through `session.eval()`.
-
-
-### Interface
-
-```python
-eval(
-    targets,
-    feed_dict=None,
-)
-```
-
-Evaluates the target Operations or Variables in `targets`.
-
-- *targets*: the evaluation targets. Can be a single Operation or
-  Variable, or a list with the Operations or Variables as
-  elements. The value returned by `eval()` has the same shape as the
-  `target` argument.
-
-  The PaddlePaddle program is represented by
-  the [ProgramDesc](../design/program.md), `eval()` will infer the
-  ProgramDesc from the given targets and run the PaddlePaddle
-  program. Please
-  see
-  [this graph](./distributed_architecture.md#local-training-architecture) for
-  the detailed illustration for the local session
-  and
-  [this graph](./distributed_architecture.md#distributed-training-architecture) for
-  the detailed illustration for the remote session.
-
-- *feed_dict*: a dictionary that contains the tensors which override
-  the edges of the computation graph.
-
-  feed_dict not only can provide the input data, it can override any
-  OP's input as well:
-
-  ```python
-  a = pd.constant(2.0, name="a")
-  b = pd.variable(name="b")
-  c = pd.mul(a,b)
-  sess.eval(targets=c, feed_dict={"b":3.0}) # returns 6.0
-  ```
-
-```python
-close()
-```
-
-Closes the session and releases the scope that the session owns.
-
-
-### Create a Local Session
-
-```python
-session(
-    devices=None
-)
-```
-
-Creates a new session. One session owns one global scope, so creating
-multiple sessions will create different scopes.
-
-- *devices*: a single `string` or a list of `string` of device names,
-  the corresponding devices will be the computation devices for
-  `eval()`. If not specified, all available devices (e.g., all GPUs)
-  will be used. The user doesn't need to specify the CPU device since
-  it will be always used. Multiple sessions can use the same device.
-
-
-#### Example
-
-```Python
-a = paddle.constant(1.0)
-b = paddle.constant(2.0)
-c = a + b
-sess = paddle.session(devices=["gpu:0", "gpu:1", "fpga:0"])
-sess.eval(c)
-sess.close()
-```
-
-### Create a Remote Session
-
-```python
-create_cloud_job(
-    name,
-    num_trainer,
-    mem_per_trainer,
-    gpu_per_trainer,
-    cpu_per_trainer,
-    num_ps,
-    mem_per_ps,
-    cpu_per_ps,
-)
-```
-
-Creates a Paddle Cloud job. Fails if the job name exists.
-
-```python
-get_cloud_job(
-    name
-)
-```
-
-Gets a Paddle Cloud job.
-
-```python
-remote_session(
-    job
-)
-```
-
-- *job*: the Paddle Cloud job.
-
-#### Example
-
-```Python
-reader = paddle.reader.recordio("/pfs/home/peter/mnist-train-*") # data stored on Paddle Cloud
-image = reader.column(0)
-label = reader.column(1)
-fc1 = paddle.op.fc(image, size=256, act="sigmoid")
-fc2 = paddle.op.fc(fc1, size=10, act="softmax")
-cost = paddle.op.cross_entropy(fc2, label)
-opt = paddle.optimizer.sgd(cost)
-
-job = paddle.create_cloud_job("test", 3, "1G", 1, 1, 2, "1G", 1)
-sess = paddle.remote_ession(job)
-for i in range(1000):
-    sess.eval(opt)
-sess.close()
-```

doc/design/releasing_process.md

@@ -7,11 +7,9 @@ PaddlePaddle每次发新的版本，遵循以下流程:
 1. 从`develop`分支派生出新的分支，分支名为`release/版本号`。例如，`release/0.10.0`
 1. 将新分支的版本打上tag，tag为`版本号rc.Patch号`。第一个tag为`0.10.0rc1`，第二个为`0.10.0rc2`，依次类推。
 1. 对这个版本的提交，做如下几个操作:
+  * 使用Regression Test List作为检查列表，测试本次release的正确性。
+	  * 如果失败，记录下所有失败的例子，在这个`release/版本号`分支中，修复所有bug后，Patch号加一，到第二步
 	* 修改`python/setup.py.in`中的版本信息,并将`istaged`字段设为`True`。
-	* 编译这个版本的Docker发行镜像，发布到dockerhub。如果失败，修复Docker编译镜像问题，Patch号加一，返回第二步
-	* 编译这个版本的Ubuntu Deb包。如果失败，修复Ubuntu Deb包编译问题，Patch号加一，返回第二步。
-	* 使用Regression Test List作为检查列表，测试Docker镜像/ubuntu安装包的功能正确性
-		* 如果失败，记录下所有失败的例子，在这个`release/版本号`分支中，修复所有bug后，Patch号加一，返回第二步
 	* 编译这个版本的python wheel包，并发布到pypi。
 		* 由于pypi.python.org目前遵循[严格的命名规范PEP 513](https://www.python.org/dev/peps/pep-0513)，在使用twine上传之前，需要重命名wheel包中platform相关的后缀，比如将`linux_x86_64`修改成`manylinux1_x86_64`。
 		* pypi上的package名称为paddlepaddle和paddlepaddle_gpu，如果要上传GPU版本的包，需要修改build/python/setup.py中，name: "paddlepaddle_gpu"并重新打包wheel包：`python setup.py bdist_wheel`。
@@ -21,8 +19,8 @@ PaddlePaddle每次发新的版本，遵循以下流程:
 			pip install twine
 			twine upload dist/[package to upload]
 			```
+		* 编译这个版本的Docker发行镜像，发布到dockerhub。如果失败，修复Docker编译镜像问题，Patch号加一，返回第二步
 1. 第三步完成后，将`release/版本号`分支合入master分支，并删除`release/版本号`分支。将master分支的合入commit打上tag，tag为`版本号`。同时再将`master`分支合入`develop`分支。最后删除`release/版本号`分支。
-1. 编译master分支的Docker发行镜像，发布到dockerhub。编译ubuntu的deb包，发布到github release页面
 1. 协同完成Release Note的书写
 
 
@@ -31,6 +29,30 @@ PaddlePaddle每次发新的版本，遵循以下流程:
 * `release/版本号`分支一旦建立，一般不允许再从`develop`分支合入`release/版本号`。这样保证`release/版本号`分支功能的封闭，方便测试人员测试PaddlePaddle的行为。
 * 在`release/版本号`分支存在的时候，如果有bugfix的行为，需要将bugfix的分支同时merge到`master`, `develop`和`release/版本号`这三个分支。
 
+## 发布wheel包到pypi
+
+使用[PaddlePaddle CI](https://paddleci.ngrok.io/project.html?projectId=Manylinux1&tab=projectOverview)
+完成自动化二进制编译，参考下图，选择需要发布的版本（通常包含一个CPU版本和一个GPU版本），点击"run"右侧的"..."按钮，可以
+弹出下面的选择框，在第二个tab (Changes)里选择需要发布的分支，这里选择0.11.0，然后点击"Run Build"按钮。等待编译完成后
+可以在此页面的"Artifacts"下拉框中找到生成的3个二进制文件，分别对应CAPI，`cp27m`和`cp27mu`的版本。然后按照上述的方法
+使用`twine`工具上传即可。
+
+<img src="ci_build_whl.png">
+
+* 注：CI环境使用 https://github.com/PaddlePaddle/buildtools 这里的DockerImage作为编译环境以支持更多的Linux
+  发型版，如果需要手动编译，也可以使用这些镜像。这些镜像也可以从 https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/ 下载得到。
+* pypi不支持覆盖上传，所以一个版本号的wheel包发布之后，不可以更改。下一个wheel包需要更新版本号才可以上传。
+
+## 发布Docker镜像
+
+上述PaddlePaddle CI编译wheel完成后会自动将Docker镜像push到DockerHub，所以，发布Docker镜像只需要对自动push的镜像打上
+版本号对应的tag即可：
+
+1. 进入 https://hub.docker.com/r/paddlepaddle/paddle/tags/ 查看latest tag的更新时间是否在上述编译wheel包完成后是否最新。
+1. 执行 `docker pull paddlepaddle/paddle:[latest tag]`，latest tag可以是latest或latest-gpu等。
+1. 执行 `docker tag paddlepaddle/paddle:[latest tag] paddlepaddle/paddle:[version]`
+1. 执行 `docker push paddlepaddle/paddle:[version]`
+
 ## PaddlePaddle 分支规范
 
 PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范，并适应github的特性做了一些区别。

doc/design/var_desc.md

 ## Background
-PaddlePaddle divides the description of neural network computation graph into two stages: compile time and runtime.
+PaddlePaddle divides the description of neural network computation into two stages: compile time and runtime. At compile time, the neural network computation is described as a `ProgramDesc` whereas at runtime an `Executor` interprets the `ProgramDesc` to compute the operations.
 
-PaddlePaddle use proto message to describe compile time graph because
+PaddlePaddle use proto message to describe compile time program because
 
-1. Computation graph should be able to be saved to a file.
-1. In distributed training, the graph will be serialized and send to multiple workers.
+1. The computation program description must be serializable and saved in a file.
+1. During distributed training, the sreialized program will be sent to multiple workers. It should also be possible to break the program into different components, each of which can be executed on different workers.
 
-The computation graph is constructed by Data Node and Operation Node. The concept to represent them is in the table below.
+The computation `Program` consists of nested `Blocks`. Each `Block` will consist of data(i.e. `Variable`)  and  `Operations`. The concept to represent them is in the table below.
 
 | |compile time|runtime|
 |---|---|---|

doc/howto/index_cn.rst

@@ -9,6 +9,7 @@
 
   usage/cmd_parameter/index_cn.rst
   usage/cluster/cluster_train_cn.md
+  usage/capi/index_cn.rst
 
 开发标准
 --------

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
 

doc/howto/usage/capi/compile_paddle_lib_cn.md

+## 编译 PaddlePaddle 预测库
+
+### 概述
+
+使用 C-API 进行预测依赖于将 PaddlePaddle 核心代码编译成链接库，只需在编译时需配制下面这些编译选项：
+
+必须配置选项：
+- `WITH_C_API`，必须配置为`ON`。
+
+推荐配置选项：
+- `WITH_PYTHON`，推荐配置为`OFF`
+- `WITH_SWIG_PY`，推荐配置为`OFF`
+- `WITH_GOLANG`，推荐设置为`OFF`
+
+可选配置选项：
+- `WITH_GPU`，可配置为`ON/OFF`
+- `WITH_MKL`，可配置为`ON/OFF`
+
+对推荐配置中的选项建议按照设置，以避免链接不必要的库。其它可选编译选项按需进行设定。
+
+下面的代码片段从github拉取最新代码，配制编译选项（需要将PADDLE_ROOT替换为PaddlePaddle预测库的安装路径）：
+
+```shell
+PADDLE_ROOT=/path/of/capi
+git clone https://github.com/PaddlePaddle/Paddle.git
+cd Paddle
+mkdir build
+cd build
+cmake -DCMAKE_INSTALL_PREFIX=$PADDLE_ROOT \
+      -DCMAKE_BUILD_TYPE=Release \
+      -DWITH_C_API=ON \
+      -DWITH_SWIG_PY=OFF \
+      -DWITH_GOLANG=OFF \
+      -DWITH_PYTHON=OFF \
+      -DWITH_MKL=OFF \
+      -DWITH_GPU=OFF  \
+      ..
+```
+
+执行上述代码生成Makefile文件后，执行：`make && make install`。成功编译后，使用C-API所需的依赖（包括：（1）编译出的PaddlePaddle预测库和头文件；（2）第三方链接库和头文件）均会存放于`PADDLE_ROOT`目录中。
+
+编译成功后在 `PADDLE_ROOT` 下会看到如下目录结构（包括了编译出的PaddlePaddle头文件和链接库，以及第三方依赖链接库和头文件（如果需要，由链接方式决定））：
+
+```text
+├── include
+│   └── paddle
+│       ├── arguments.h
+│       ├── capi.h
+│       ├── capi_private.h
+│       ├── config.h
+│       ├── error.h
+│       ├── gradient_machine.h
+│       ├── main.h
+│       ├── matrix.h
+│       ├── paddle_capi.map
+│       └── vector.h
+├── lib
+│   ├── libpaddle_capi_engine.a
+│   ├── libpaddle_capi_layers.a
+│   ├── libpaddle_capi_shared.so
+│   └── libpaddle_capi_whole.a
+└── third_party
+    ├── gflags
+    │   ├── include
+    │   │   └── gflags
+    │   │       ├── gflags_completions.h
+    │   │       ├── gflags_declare.h
+    │   │       ...
+    │   └── lib
+    │       └── libgflags.a
+    ├── glog
+    │   ├── include
+    │   │   └── glog
+    │   │       ├── config.h
+    │   │       ...
+    │   └── lib
+    │       └── libglog.a
+    ├── openblas
+    │   ├── include
+    │   │   ├── cblas.h
+    │   │   ...
+    │   └── lib
+    │       ...
+    ├── protobuf
+    │   ├── include
+    │   │   └── google
+    │   │       └── protobuf
+    │   │           ...
+    │   └── lib
+    │       └── libprotobuf-lite.a
+    └── zlib
+        ├── include
+        │   ...
+        └── lib
+            ...
+
+```
+
+### 链接说明
+
+目前提供三种链接方式：
+
+1. 链接`libpaddle_capi_shared.so` 动态库
+    - 使用 PaddlePaddle C-API 开发预测程序链接`libpaddle_capi_shared.so`时，需注意：
+        1. 如果编译时指定编译CPU版本，且使用`OpenBLAS`数学库，在使用C-API开发预测程序时，只需要链接`libpaddle_capi_shared.so`这一个库。
+        1. 如果是用编译时指定CPU版本，且使用`MKL`数学库，由于`MKL`库有自己独立的动态库文件，在使用PaddlePaddle C-API开发预测程序时，需要自己链接MKL链接库。
+        1. 如果编译时指定编译GPU版本，CUDA相关库会在预测程序运行时动态装载，需要将CUDA相关的库设置到`LD_LIBRARY_PATH`环境变量中。
+    - 这种方式最为简便，链接相对容易，**在无特殊需求情况下，推荐使用此方式**。
+
+2. 链接静态库 `libpaddle_capi_whole.a`
+    - 使用PaddlePaddle C-API 开发预测程序链接`libpaddle_capi_whole.a`时，需注意：
+        1. 需要指定`-Wl,--whole-archive`链接选项。
+        1. 需要显式地链接 `gflags`、`glog`、`libz`、`protobuf` 等第三方库，可在`PADDLE_ROOT/third_party`下找到。
+        1. 如果在编译 C-API 时使用OpenBLAS数学库，需要显示地链接`libopenblas.a`。
+        1. 如果在编译 C-API 是使用MKL数学库，需要显示地链接MKL的动态库。
+
+3. 链接静态库 `libpaddle_capi_layers.a`和`libpaddle_capi_engine.a`
+    - 使用PaddlePaddle C-API 开发预测程序链接`libpaddle_capi_whole.a`时，需注意：
+        1. 这种链接方式主要用于移动端预测。
+        1. 为了减少生成链接库的大小把`libpaddle_capi_whole.a`拆成以上两个静态链接库。
+        1. 需指定`-Wl,--whole-archive -lpaddle_capi_layers` 和 `-Wl,--no-whole-archive -lpaddle_capi_engine` 进行链接。
+        1. 第三方依赖库需要按照与方式2同样方法显示地进行链接。

doc/howto/usage/capi/index_cn.rst

+PaddlePaddle C-API
+==================
+
+..  toctree::
+  :maxdepth: 1
+
+  compile_paddle_lib_cn.md
+  organization_of_the_inputs_cn.md
+  workflow_of_capi_cn.md

doc/howto/usage/capi/organization_of_the_inputs_cn.md

+## 输入/输出数据组织
+
+这篇文档介绍在使用 PaddlePaddle C-API 时如何组织输入数据，以及如何解析神经网络前向计算的输出结果。
+
+### 输入/输出数据类型
+在C-API中，按照基本数据类型在PaddlePaddle内部的定义和实现，输入数据可分为：
+1. 一维整型数组
+1. 二维浮点型矩阵
+    - 稠密矩阵
+    - 稀疏矩阵
+
+说明：
+1. 一维数组**仅支持整型值**；
+    - 常用于自然语言处理任务，例如：表示词语在词典中的序号；
+    - 分类任务中类别标签；
+1. 逻辑上高于二维的数据（例如含有多个通道的图片，视频等）在程序实现中都会转化为二维矩阵，转化方法在相应的领域都有通用解决方案，需要使用者自己了解并完成转化；
+1. 二维矩阵可以表示行向量和列向量，任何时候如果需要浮点型数组（向量），都应使用C-API中的矩阵来表示，而不是C-API中的一维数组。
+1. 不论是一维整型数组还是二维浮点数矩阵，**为它们附加上序列信息将变成序列输入。PaddlePaddle 会通过判数据是否附带有序列信息来判断一个向量/矩阵是否是一个序列**。当非序列输入时，无需关心和处理序列信息。关于什么是“序列信息”，下文会详细进行介绍。
+
+### 基本使用概念
+
+- 在PaddlePaddle内部，神经网络中一个计算层的输入/输出被组织为一个 `Argument` 结构体，如果神经网络有多个输入或者多个输入，每一个输入/输入都会对应有自己的`Argument`。
+- `Argument` 并不真正“存储”数据，而是将输入/输出信息有机地组织在一起。
+- 在`Argument`内部由`IVector`（对应着上文提到的一维整型数组）和`Matrix`（对应着上文提到的二维浮点型矩阵）来实际存储数据；由 `Sequence Start Positions` (下文详细解释) 来描述输入/输出的序列信息。
+
+- **注**：
+    1. 这篇文档之后部分将会统一使用`argument`来特指PaddlePaddle中神经网络计算层一个输入/输出数据。
+    1. 使用`paddle_ivector`来特指PaddlePaddle中的一维整型数组。
+    1. 使用`paddle_matrix`来特指PaddlePaddle中的二维浮点型矩阵。
+
+### 组织输入数据
+- 一维整型数组
+
+    概念上可以将`paddle_ivector`理解为一个一维的整型数组，通常用于表示离散的类别标签，或是在自然语言处理任务中表示词语在字典中的序号。下面的代码片段创建了含有三个元素`1`、`2`、`3`的`paddle_ivector`。
+    ```c
+    int ids[] = {1, 2, 3};
+     paddle_ivector ids_array =
+         paddle_ivector_create(ids, sizeof(ids) / sizeof(int), false, false);
+     CHECK(paddle_arguments_set_ids(in_args, 0, ids_array));
+    ```
+
+- **稠密矩阵**
+    - 一个`m×n`的稠密矩阵是一个由`m`行`n`列元素排列成的矩形阵列，矩阵里的元素是浮点数。对神经网络来说，矩阵的高度`m`是一次预测接受的样本数目，宽度$n$是神经网络定义时，`paddle.layer.data`的`size`。
+    - 下面的代码片段创建了一个高度为1，宽度为`layer_size`的稠密矩阵，矩阵中每个元素的值随机生成。
+
+    ```c
+    paddle_matrix mat = paddle_matrix_create(
+                            /* height = batch size */ 1,
+                            /* width = dimensionality of the data layer */ layer_size,
+                            /* whether to use GPU */ false);
+
+    paddle_real* array;
+    // Get the pointer pointing to the start address of the first row of the
+    // created matrix.
+    CHECK(paddle_matrix_get_row(mat, 0, &array));
+
+    // Fill the matrix with a randomly generated test sample.
+    srand(time(0));
+    for (int i = 0; i < layer_size; ++i) {
+      array[i] = rand() / ((float)RAND_MAX);
+    }
+
+    // Assign the matrix to the argument.
+    CHECK(paddle_arguments_set_value(in_args, 0, mat));
+    ```
+
+- **稀疏矩阵**
+
+  PaddlePaddle C-API 中 稀疏矩阵使用[CSR（Compressed Sparse Row Format）](https://en.wikipedia.org/wiki/Sparse_matrix#Compressed_sparse_row_(CSR,_CRS_or_Yale_format))格式存储。下图是CSR存储稀疏矩阵的示意图。
+  <p align="center">
+  <img src="https://user-images.githubusercontent.com/5842774/34159369-009fd328-e504-11e7-9e08-36bc6dc5e505.png" width=700><br> 图1. 稀疏矩阵存储示意图
+  </p>
+
+  CSR存储格式通过：（1）非零元素的值（上图中的`values`）；（2）行偏移(上图中的`row offsets`)：每一行元素在`values`中的起始偏移，`row offsets`中元素个数总是等于行数 + 1；（3）非零元素的列号（上图中的`column indices`）来确定稀疏矩阵的内容。
+
+  在PaddlePaddle C-API中，通过调用以下接口创建稀疏矩阵：
+
+  ```c
+  PD_API paddle_matrix paddle_matrix_create_sparse(
+      uint64_t height, uint64_t width, uint64_t nnz, bool isBinary, bool useGpu);
+  ```
+
+  1. 创建稀疏矩阵时需要显示地指定矩阵的（1）高度（`height`，在神经网络中等于一次预测处理的样本数）（2）宽度（`width`，`paddle.layer.data`的`size`）以及（3）非零元个数（`nnz`）。
+  1. 当上述接口第4个参数`isBinary`指定为`true`时，**只需要设置行偏移（`row_offset`）和列号(`colum indices`)，不需要提供元素值（`values`）**，这时行偏移和列号指定的元素默认其值为1。
+
+  下面的代码片段创建了一个CPU上的二值稀疏矩阵：
+
+  ```c
+  paddle_matrix mat = paddle_matrix_create_sparse(1, layer_size, nnz, true, false);
+  int colIndices[] = {9, 93, 109};  // layer_size here is greater than 109.
+  int rowOffset[] = {0, sizeof(colIndices) / sizeof(int)};
+
+  CHECK(paddle_matrix_sparse_copy_from(mat,
+                                 rowOffset,
+                                 sizeof(rowOffset) / sizeof(int),
+                                 colIndices,
+                                 (colIndices) / sizeof(int),
+                                 NULL /*values array is NULL.*/,
+                                 0 /*size of the value arrary is 0.*/));
+  CHECK(paddle_arguments_set_value(in_args, 0, mat));
+  ```
+  下面的代码片段在创建了一个CPU上的带元素值的稀疏矩阵：
+  ```c
+  paddle_matrix mat = paddle_matrix_create_sparse(1, layer_size, nnz, false, false);
+  int colIndices[] = {9, 93, 109};  // layer_size here is greater than 109.
+  int rowOffset[] = {0, sizeof(colIndices) / sizeof(int)};
+  float values[] = {0.5, 0.5, 0.5};
+
+  CHECK(paddle_matrix_sparse_copy_from(mat,
+                                 rowOffset,
+                                 sizeof(rowOffset) / sizeof(int),
+                                 colIndices,
+                                 sizeof(colIndices) / sizeof(int),
+                                 values,
+                                 sizeof(values) / sizeof(float)));
+  ```
+  注意事项：
+  1. 移动端预测**不支持**稀疏矩阵及相关的接口。
+
+### 组织序列信息
+
+多个排成一列的元素（可以是整型、浮点数、浮点数向量等）构成一个序列，元素之间的顺序是序列所携带的重要信息。不同序列可能会含有不同数目个元素。在 PaddlePaddle 中，序列输入/输出数据是在上文介绍的**数据输入（一维整型数组，二维浮点数矩阵）基础上，附加上序列信息**。下面详细解释什么是“序列信息”。
+
+我们将神经网络一次计算接受的所有输入样本称之为一个`batch`（可以含有一条或多条样本），每一个序列在整个`batch`中的偏移，就是PaddlePaddle中所指的**序列信息**，称之为“sequence start positions”。PaddlePaddle 支持两种序列类型：
+
+1. 单层序列
+    - 序列中的每一个元素是非序列，是进行计算的基本单位，不可再进行拆分。
+    - 例如：自然语言中的句子是一个序列，序列中的元素是词语；
+1. 双层序列
+    - 序列中的每一个元素又是一个序列。
+    - 例如：自然语言中的段落是一个双层序列；段落是由句子构成的序列；句子是由词语构成的序列。
+    - 双层序列在处理长序列的任务或是构建层级模型时会发挥作用。
+
+这篇文档之后部分会统一使用`sequence_start_positions`来特指：PaddlePaddle中神经网络计算层输入/输出所携带的序列信息。
+
+对双层序列来讲，不仅要提供每一个外层序列在整个`batch`中的偏移，每一个外层序列又含有若干个内层序列，需要同时提供每一个内层序列在整个`batch`中的偏移。也就是说：**双层序列需要设置分别为外层序列和内层序列分别设置`sequence_start_positions`信息**。
+
+**注：**
+1. 不论序列中的元素在内存中占用多少实际存储空间，`sequence_start_positions`表示的偏移是以“序列中的一个元素”作为统计的基本单位，而不是相对`batch`起始存储地址以数据的存储大小为单位的偏移。
+1. 非序列输入不携带`sequence_start_positions`，非序列输入无需构造`sequence_start_positions`。
+1. **不论是单层序列还是双层序列的序列信息，都使用`paddle_ivector`（也就是PaddlePaddle中的一维整型数组）来存储。**
+
+图2 是PaddlePaddle中单层序列和双层序列存储示意图。
+<p align="center">
+<img src="https://user-images.githubusercontent.com/5842774/34159714-1f81a9be-e505-11e7-8a8a-4902146ec899.png" width=800><br>图2. 序列输入示意图
+</p>
+
+- 单层序列
+
+    图2 (a) 展示了一个含有4个序列的`batch`输入：
+    1. 4个序列的长度分别为：5、3、2、4；
+    1. 这时的`sequence_start_positions`为：`[0, 5, 8, 10, 14]`；
+    1. 本地训练. 不论数据域是`paddle_ivector`类型还是`paddle_matrix`类型，都可以通过调用下面的接口为原有的数据输入附加上序列信息，使之变为一个单层序列输入，代码片段如下：
+
+    ```c
+    int seq_pos_array[] = {0, 5, 8, 10, 14};
+    paddle_ivector seq_pos = paddle_ivector_create(
+        seq_pos_array, sizeof(seq_pos_array) / sizeof(int), false, false);
+    // Suppose the network only has one input data layer.
+    CHECK(paddle_arguments_set_sequence_start_pos(in_args, 0, 0, seq_pos));
+    ```
+
+- 双层序列
+
+    图2 (b) 展示了一个含有4个序列的`batch`输入；
+    1. 4个序列的长度分别为：5、3、2、4；这四个序列又分别含有3、2、1、2个子序列；
+    1. 这时的需要同时提供：
+        - 外层序列在`batch`中的起始偏移`：[0, 5, 8, 10, 14]`；
+        - 内层序列在`batch`中的起始偏移：`[0, 2, 3, 5, 7， 8， 10， 13， 14]`；
+    1. 不论数据域是`paddle_ivector`类型还是`paddle_matrix`类型，这时需要调用创建序列信息和为`argument`设置序列信息的接口**两次**，分别为数据输入添加外层序列和内层序列的序列信息，使之变为一个双层序列输入，代码片段如下：
+    ```c
+    // set the sequence start positions for the outter sequences.
+    int outter_seq_pos_array[] = {0, 5, 8, 10, 14};
+    paddle_ivector seq_pos =
+        paddle_ivector_create(outter_seq_pos_array,
+                              sizeof(outter_pos_array) / sizeof(int),
+                              false,
+                              false);
+    // The third parameter of this API indicates the sequence level.
+    // 0 for the outter sequence. 1 for the inner sequence.
+    // If the input is a sequence not the nested sequence, the third parameter is
+    // fixed to be 0.
+    CHECK(paddle_arguments_set_sequence_start_pos(in_args, 0, 0, seq_pos));
+
+    // set the sequence start positions for the outter sequences.
+    int inner_seq_pos_array[] = {0, 2, 3, 5, 7， 8， 10， 13， 14};
+    paddle_ivector seq_pos = paddle_ivector_create(
+        inner_pos_array, sizeof(inner_pos_array) / sizeof(int), false, false);
+    // The third parameter of this API indicates the sequence level.
+    // 0 for the outter sequence. 1 for the inner sequence.
+    CHECK(paddle_arguments_set_sequence_start_pos(in_args, 0, 1, seq_pos));
+    ```
+
+注意事项：
+1. 当一个`batch`中含有多个序列，**不支持序列长度为`0`的序列（也就是空输入）** 作为输入。不同计算层对空输入的处理策略有可能不同，潜在会引起未定义行为，或者引起行时错误，请在输入时进行合法性检查。
+
+### Python 端数据类型说明
+
+下表列出了Python端训练接口暴露的数据类型（`paddle.layer.data`函数`type`字段的取值）对应于调用C-API需要创建的数据类型：
+
+<html>
+<table border="2" frame="border">
+<table>
+<thead>
+<tr>
+<th style="text-align:left">Python 端数据类型</th>
+<th style="text-align:left">C-API 输入数据类型</th>
+</tr>
+</thead>
+<tbody>
+<tr>
+<td style="text-align:left">paddle.data_type.integer_value</td>
+<td style="text-align:left">整型数组，无需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.dense_vector</td>
+<td style="text-align:left">浮点型稠密矩阵，无需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.sparse_binary_vector</td>
+<td style="text-align:left">浮点型稀疏矩阵，无需提供非零元的值，默认为1，无需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.sparse_vector</td>
+<td style="text-align:left">浮点型稀疏矩阵，需提供非零元的值，无需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.integer_value_sequence</td>
+<td style="text-align:left">整型数组，需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.dense_vector_sequence</td>
+<td style="text-align:left">浮点型稠密矩阵，需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.sparse_binary_vector_sequence</td>
+<td style="text-align:left">浮点型稀疏矩阵，无需提供非零元的值，默认为1，需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.sparse_vector_sequence</td>
+<td style="text-align:left">浮点型稀疏矩阵，需提供非零元的值，需附加序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.integer_value_sub_sequence</td>
+<td style="text-align:left">整型数组，需附加双层序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.dense_vector_sub_sequence</td>
+<td style="text-align:left">浮点型稠密矩阵，需附加双层序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.sparse_binary_vector_sub_sequence</td>
+<td style="text-align:left">浮点型稀疏矩阵，无需提供非零元的值，默认为1，需附加双层序列信息</td>
+</tr>
+<tr>
+<td style="text-align:left">paddle.data_type.sparse_vector_sub_sequence</td>
+<td style="text-align:left">浮点型稀疏矩阵，需提供非零元的值，需附加双层序列信息</td>
+</tr>
+</tbody>
+</table>
+</html>
+<br>
+
+
+### 输出数据
+
+PaddlePaddle中一个计算层的输出数据组织方式和输入数据组织方式完全相同。一个输出数据同样被组织为一个`argument`，`argument`通过`paddle_matrix`或`paddle_ivector`存数数据，如果输出是一个序列，那么会携带有`sequence_start_positions`信息。调用C-API相关接口，读取需要的结果即可。
+
+### 总结
+
+- 在PaddlePaddle内部，神经网络中一个计算层的输入/输出被组织为`argument`。
+- `argument`并不真正“存储”数据，而是将输入/输出信息有机地组织在一起。
+- 在`argument`内部由`paddle_ivector`（一维整型数组）和`paddle_matrix`（二维浮点型矩阵）来实际存储数据。
+如果是一个序列输入/输出由 `sequence start positions` 来记录输入/输出的序列信息。
+
+于是，在组织神经网络输入时，需要思考完成以下工作：
+1. 为每一个输入/输出创建`argument`。
+    - C-API 中操作`argument`的接口请查看[argument.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/arguments.h)。
+1. 为每一个`argument`创建`paddle_matrix`或者`paddle_ivector`来存储数据。
+    - C-API 中操作`paddle_ivector`的接口请查看 [vector.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/vector.h)。
+    - C-API 中操作`paddle_matrix`的接口请查看[matrix.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/matrix.h)。
+1. 如果输入是序列数据，需要创建并填写`sequence_start_positions`信息。
+    - 通过调用 [`paddle_arguments_set_sequence_start_pos`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/arguments.h#L137) 来为一个`argument`添加序列信息。
+    - 通过调用 [`paddle_arguments_get_sequence_start_pos`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/arguments.h#L150) 来读取一个`argument`添加序列信息。
+    - 接口说明请查看 [argument.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/arguments.h) 文件。

doc/howto/usage/capi/workflow_of_capi_cn.md

+## C-API 使用流程
+
+这篇文档介绍 PaddlePaddle C-API 整体使用流程。
+
+### 使用流程
+
+使用 C-API 的工作流程如图1所示，分为（1）准备预测模型和（2）预测程序开发两大部分。
+
+<p align="center">
+<img src="https://user-images.githubusercontent.com/5842774/34658453-365f73ea-f46a-11e7-9b3f-0fd112b27bae.png" width=500><br> 图1. C-API使用流程示意图
+</p>
+
+- 准备预测模型
+    1. 只将神经网络结构进行序列化。
+        - 只对神经网络结构进行序列化，加载模型需同时指定：网络结构的序列化结果和模型参数存储目录。
+    1. 将网络结构定义和训练结束存储下来的模型参数文件（多个）合并入一个文件。
+        - 神经网络模型结构和训练好的模型将被序列化合并入一个文件。
+        - 预测时只需加载一个文件便于发布。
+    - **注意**：以上两种方式只需选择其一即可。
+- 调用 C-API 开发预测序
+    1. 初始化PaddlePaddle运行环境。
+    1. 加载预测模型。
+    1. 创建神经网络输入，组织输入数据。
+    1. 进行前向计算，获得计算结果。
+    1. 清理和结束。
+
+### 准备预测模型
+
+准备预测模型部分，我们以手写数字识别任务为例进行介绍。手写数字识别任务定义了一个含有[两个隐层的简单全连接网络](https://github.com/PaddlePaddle/book/blob/develop/02.recognize_digits/README.cn.md#softmax回归softmax-regression)，网络接受一幅图片作为输入，将图片分类到 0 ~ 9 类别标签之一。完整代码可以查看[此目录](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense) 中的相关脚本。
+
+调用C-API开发预测程序需要一个训练好的模型，运行[MNIST手写数字识别目录](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)下的[mnist_v2.py](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/examples/model_inference/dense/mnist_v2.py)脚本，在终端执行`python mnist_v2.py`，会使用 PaddlePaddle 内置的 [MNIST 数据集](http://yann.lecun.com/exdb/mnist/)进行训练。训练好的模型默认保存在当前运行目录下的`models`目录中。
+
+下面，我们将训练结束后存储下来的模型转换成预测模型。
+
+1. 序列化神经网络模型配置
+
+    PaddlePaddle 使用 protobuf 来传输网络配置文件中定义的网络结构和相关参数，使用 C-API 进行预测时，需要将网络结构使用 protobuf 进行序列化，写入文件中。
+
+    调用[`paddle.utils.dump_v2_config`](https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/utils/dump_v2_config.py)中的`dump_v2_config`函数能够将使用 PaddlePaddle V2 API 定义的神经网络结构 dump 到指定文件中，示例代码如下：
+
+    ```python
+    from paddle.utils.dump_v2_config import dump_v2_config
+    from mnist_v2 import network
+
+    predict = network(is_infer=True)
+    dump_v2_config(predict, "trainer_config.bin", True)
+    ```
+
+    对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)这个示例，[`mnist_v2.py`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense/mnist_v2.py)脚本集成了序列化神经网络结构的过程，可以直接运行 `python mnist_v2.py --task dump_config` 对神经网络结构进行序列化，结果会写入当前运行目录下的`trainer_config.bin`文件中。
+
+    使用这种方式，需要**在运行时将神经网络的多个可学习参数放在同一个目录中**，C-API可以通过分别指定序列化后的网络结构文件和参数目录来加载训练好的模型。
+
+2. 合并模型文件(可选)
+
+    一些情况为了便于发布，希望能够将序列化后的神经网络结构和训练好的模型参数打包进一个文件。对于这样的需求，可以使用`paddle.utils.merge_model`中的`merge_v2_model`接口对神经网络结构和训练好的参数进行序列化，将序列化结果写入一个文件内。
+
+    代码示例如下：
+
+    ```python
+    from paddle.utils.merge_model import merge_v2_modelss
+    from mnist_v2 import network
+
+    net = network(is_infer=True)
+    param_file = "models/params_pass_4.tar"
+    output_file = "output.paddle.model"
+    merge_v2_model(net, param_file, output_file)
+    ```
+    对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)这个示例，可直接运行 `python` [merge_v2_model.py](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense/merge_v2_model.py)。序列化结果会写入当前运行目录下的`output.paddle.model`文件中。使用这种方式，运行时C-API可以通过指定`output.paddle.model`文件的路径来加载预测模型。
+
+#### 注意事项
+1. 为使用C-API，在调用`dump_v2_config`序列化神经网络结构时，参数`binary`必须指定为`True`。
+1. **预测使用的网络结构往往不同于训练**，通常需要去掉网络中的：（1）类别标签层；（2）损失函数层；（3）`evaluator`等，只留下核心计算层，请注意是否需要修改网络结构。
+1. 预测时，可以获取网络中定义的任意多个（大于等于一个）层前向计算的结果，需要哪些层的计算结果作为输出，就将这些层加入一个Python list中，作为调用`dump_v2_config`的第一个参数。
+
+### 编写预测代码
+
+预测代码更多详细示例代码请参考[C-API使用示例](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference) 目录下的代码示例。这一节对图1中预测代码编写的5个步骤进行介绍和说明。
+
+#### step 1. 初始化PaddlePaddle运行环境
+第一步需调用[`paddle_init`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/main.h#L27) 初始化PaddlePaddle运行环境，该接口接受两个参数：参数的个数和参数列表。
+
+#### step2. 加载模型
+
+这里介绍C-API使用中的一个重要概念：Gradient Machine。
+
+概念上，在 PaddlePaddle 内部，一个GradientMachine类的对象管理着一组计算层（PaddlePaddle Layers）来完成前向和反向计算，并处理与之相关的所有细节。在调用C-API预测时，只需进行前向计算而无需调用反向计算。这篇文档之后部分会使用`gradient machine`来特指调用PaddlePaddle C-API创建的GradientMachine类的对象。每一个 `gradient machine` 都会管理维护一份训练好的模型，下面是C-API提供的，两种常用的模型加载方式：
+
+1. 调用[`paddle_gradient_machine_load_parameter_from_disk`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/gradient_machine.h#L61)接口，从磁盘加载预测模型。这时`gradient machine`会独立拥有一份训练好的模型；
+1. 调用[`paddle_gradient_machine_create_shared_param`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/gradient_machine.h#L88)接口，与其它`gradient machine`的共享已经加载的预测模型。这种情况多出现在使用多线程预测时，通过多个线程共享同一个模型来减少内存开销。可参考[此示例](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/examples/model_inference/multi_thread/main.c)。
+
+- 注意事项
+    1. 使用PaddlePaddle V2 API训练，模型中所有可学习参数会被存为一个压缩文件，需要手动进行解压，将它们放在同一目录中，C-API不会直接加载 V2 API 存储的压缩文件。
+    1. 如果使用`merge model`方式将神经网络结构和训练好的参数序列化到一个文件，请参考此[示例](https://github.com/PaddlePaddle/Mobile/blob/develop/Demo/linux/paddle_image_recognizer.cpp#L59)。
+    1. 通过灵活使用以上两个接口，加载模型可其它多种方式，例如也可在程序运行过程中再加载另外一个模型。
+
+#### step 3. 创建神经网络输入，组织输入数据
+
+基本使用概念：
+- 在PaddlePaddle内部，神经网络中一个计算层的输入输出被组织为一个 `Argument` 结构体，如果神经网络有多个输入或者多个输出，每一个输入/输出都会对应有自己的`Argument`。
+- `Argument` 并不真正“存储”数据，而是将输入/输出数据有机地组织在一起。
+- 在`Argument`内部由：1. `Matrix`（二维矩阵，存储浮点类型输入/输出）；2. `IVector`（一维数组，**仅用于存储整型值**，多用于自然语言处理任务）来实际存储数据。
+
+C-API支持的所有输入数据类型和他们的组织方式，请参考“输入/输出数据组织”一节。
+
+这篇文档的之后部分会使用`argument`来特指PaddlePaddle C-API中神经网络的一个输入/输出，使用`paddle_matrix`**特指**`argument`中用于存储数据的`Matrix`类的对象。
+
+在组织神经网络输入，获取输出时，需要思考完成以下工作：
+1. 为每一个输入/输出创建`argument`；
+1. 为每一个`argument`创建`paddle_matrix`来存储数据；
+
+与输入不同的是，不需在使用C-API时为输出`argument`的`paddle_matrix`对象分配空间。前向计算之后PaddlePaddle内部已经分配/管理了每个计算层输出的存储空间。
+
+#### step 4. 前向计算
+
+完成上述准备之后，通过调用 [`paddle_gradient_machine_forward`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/gradient_machine.h#L73) 接口完成神经网络的前向计算。
+
+#### step 5. 清理
+
+结束预测之后，对使用的中间变量和资源进行清理和释放。

doc/howto/usage/cluster/cluster_train_cn.md

@@ -51,7 +51,7 @@ $ stdbuf -oL /usr/bin/nohup paddle pserver --port=7164 --ports_num=1 --ports_num
 
 - port：**必选，默认7164**，pserver监听的起始端口，根据ports_num决定总端口个数，从起始端口监听多个端口用于通信
 - ports_num：**必选，默认1**，监听的端口个数
-- ports_num_for_sparse：**必选，默认1**，用于稀疏类型参数通信的端口个数
+- ports_num_for_sparse：**必选，默认0**，用于稀疏类型参数通信的端口个数
 - num_gradient_servers：**必选，默认1**，当前训练任务pserver总数
 
 ### 启动计算节点
@@ -60,7 +60,7 @@ $ stdbuf -oL /usr/bin/nohup paddle pserver --port=7164 --ports_num=1 --ports_num
 $ python train.py
 ```
 
-trainer需要和pserver保持网络联通以完成训练。trainer启动需要传入端口、pserver地址等参数使trainer可以正确连接到pserver。这些参数可以通过环境变量（https://zh.wikipedia.org/wiki/环境变量 ）或编写程序时`paddle.init()`中传入参数。如果同时使用`paddle.init()`参数和环境变量，将会优先使用`paddle.init()`中传入的参数。
+trainer需要和pserver保持网络联通以完成训练。trainer启动需要传入端口、pserver地址等参数使trainer可以正确连接到pserver。这些参数可以通过[环境变量](https://zh.wikipedia.org/wiki/环境变量)或编写程序时`paddle.init()`中传入参数。如果同时使用`paddle.init()`参数和环境变量，将会优先使用`paddle.init()`中传入的参数。
 
 使用环境变量：
 
@@ -95,7 +95,7 @@ paddle.init(
 - trainer_count：**必选，默认1**，当前训练任务trainer总个数
 - port：**必选，默认7164**，连接到pserver的端口
 - ports_num：**必选，默认1**，连接到pserver的端口个数
-- ports_num_for_sparse：**必选，默认1**，和pserver之间用于稀疏类型参数通信的端口个数
+- ports_num_for_sparse：**必选，默认0**，和pserver之间用于稀疏类型参数通信的端口个数
 - num_gradient_servers：**必选，默认1**，当前训练任务pserver总数
 - trainer_id：**必选，默认0**，每个trainer的唯一ID，从0开始的整数
 - pservers：**必选，默认127.0.0.1**，当前训练任务启动的pserver的IP列表，多个IP使用“,”隔开

doc/howto/usage/cluster/cluster_train_en.md

@@ -52,7 +52,7 @@ Parameter Description
 
 - port: **required, default 7164**, port which parameter server will listen on. If ports_num greater than 1, parameter server will listen on multiple ports for more network throughput.
 - ports_num: **required, default 1**, total number of ports will listen on.
-- ports_num_for_sparse: **required, default 1**, number of ports which serves sparse parameter update.
+- ports_num_for_sparse: **required, default 0**, number of ports which serves sparse parameter update.
 - num_gradient_servers: **required, default 1**, total number of gradient servers.
 
 ### Starting trainer
@@ -98,7 +98,7 @@ Parameter Description
 - trainer_count: **required, default 1**, total count of trainers in the training job.
 - port: **required, default 7164**, port to connect to parameter server.
 - ports_num: **required, default 1**, number of ports for communication.
-- ports_num_for_sparse: **required, default 1**, number of ports for sparse type caculation.
+- ports_num_for_sparse: **required, default 0**, number of ports for sparse type caculation.
 - num_gradient_servers: **required, default 1**, total number of gradient server.
 - trainer_id: **required, default 0**, ID for every trainer, start from 0.
 - pservers: **required, default 127.0.0.1**, list of IPs of parameter servers, separated by ",".

paddle/capi/examples/model_inference/dense/main.c

@@ -3,59 +3,82 @@
 
 #include "../common/common.h"
 
+// Modify this path as needed.
 #define CONFIG_BIN "./trainer_config.bin"
+// Modify this path as needed.
+// This demo assumes that merged model is not used, then this path is the
+// directory storing all the trained parameters.
+// If the model is trained by PaddlePaddle V2 API, the model is saved as
+// a compressed file. You need to uncompress the compressed file first.
+#define MODEL_PATH "models/pass_4"
 
 int main() {
-  // Initalize Paddle
+  // Initalize the PaddlePaddle runtime environment.
   char* argv[] = {"--use_gpu=False"};
   CHECK(paddle_init(1, (char**)argv));
 
-  // Reading config binary file. It is generated by `convert_protobin.sh`
+  // Read the binary configuration file generated by `convert_protobin.sh`
   long size;
   void* buf = read_config(CONFIG_BIN, &size);
 
-  // Create a gradient machine for inference.
+  // Create the gradient machine for inference.
   paddle_gradient_machine machine;
   CHECK(paddle_gradient_machine_create_for_inference(&machine, buf, (int)size));
-  CHECK(paddle_gradient_machine_randomize_param(machine));
 
-  // Loading parameter. Uncomment the following line and change the directory.
-  // CHECK(paddle_gradient_machine_load_parameter_from_disk(machine,
-  //                                                "./some_where_to_params"));
+  // Load the trained model. Modify the parameter MODEL_PATH to set the correct
+  // path of the trained model.
+  CHECK(paddle_gradient_machine_load_parameter_from_disk(machine, MODEL_PATH));
+
+  // Inputs and outputs of the network are organized as paddle_arguments object
+  // in C-API. In the comments below, "argument" specifically means one input of
+  // the neural network in PaddlePaddle C-API.
   paddle_arguments in_args = paddle_arguments_create_none();
 
-  // There is only one input of this network.
+  // There is only one data layer in this demo MNIST network, invoke this
+  // function to create one argument.
   CHECK(paddle_arguments_resize(in_args, 1));
 
-  // Create input matrix.
-  paddle_matrix mat = paddle_matrix_create(/* sample_num */ 1,
-                                           /* size */ 784,
-                                           /* useGPU */ false);
-  srand(time(0));
+  // Each argument needs one matrix or one ivector (integer vector, for sparse
+  // index input, usually used in NLP task) to holds the real input data.
+  // In the comments below, "matrix" specifically means the object needed by
+  // argument to hold the data. Here we create the matrix for the above created
+  // agument to store the testing samples.
+  paddle_matrix mat =
+      paddle_matrix_create(/* height = batch size */ 1,
+                           /* width = dimensionality of the data layer */ 784,
+                           /* whether to use GPU */ false);
 
   paddle_real* array;
-
-  // Get First row.
+  // Get the pointer pointing to the start address of the first row of the
+  // created matrix.
   CHECK(paddle_matrix_get_row(mat, 0, &array));
 
+  // Fill the matrix with a randomly generated test sample.
+  srand(time(0));
   for (int i = 0; i < 784; ++i) {
     array[i] = rand() / ((float)RAND_MAX);
   }
 
+  // Assign the matrix to the argument.
   CHECK(paddle_arguments_set_value(in_args, 0, mat));
 
+  // Create the output argument.
   paddle_arguments out_args = paddle_arguments_create_none();
+
+  // Invoke the forward computation.
   CHECK(paddle_gradient_machine_forward(machine,
                                         in_args,
                                         out_args,
-                                        /* isTrain */ false));
-  paddle_matrix prob = paddle_matrix_create_none();
+                                        /* is train taks or not */ false));
 
+  // Create the matrix to hold the forward result of the neural network.
+  paddle_matrix prob = paddle_matrix_create_none();
+  // Access the matrix of the output argument, the predicted result is stored in
+  // which.
   CHECK(paddle_arguments_get_value(out_args, 0, prob));
 
   uint64_t height;
   uint64_t width;
-
   CHECK(paddle_matrix_get_shape(prob, &height, &width));
   CHECK(paddle_matrix_get_row(prob, 0, &array));
 
@@ -68,6 +91,7 @@ int main() {
   }
   printf("\n");
 
+  // The cleaning up.
   CHECK(paddle_matrix_destroy(prob));
   CHECK(paddle_arguments_destroy(out_args));
   CHECK(paddle_matrix_destroy(mat));

paddle/capi/examples/model_inference/dense/merge_v2_model.py

+from paddle.utils.merge_model import merge_v2_model
+
+from mnist_v2 import network
+
+net = network(is_infer=True)
+param_file = "models/params_pass_4.tar"
+output_file = "output.paddle.model"
+merge_v2_model(net, param_file, output_file)

paddle/capi/examples/model_inference/dense/mnist_v2.py

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

paddle/capi/examples/model_inference/sparse_binary/main.c

 #include <paddle/capi.h>
 #include <time.h>
+
 #include "../common/common.h"
 
 #define CONFIG_BIN "./trainer_config.bin"
@@ -9,16 +10,18 @@ int main() {
   char* argv[] = {"--use_gpu=False"};
   CHECK(paddle_init(1, (char**)argv));
 
-  // Reading config binary file. It is generated by `convert_protobin.sh`
+  // Read the binary configuration file which is generated by
+  // `convert_protobin.sh`
   long size;
   void* buf = read_config(CONFIG_BIN, &size);
 
-  // Create a gradient machine for inference.
+  // Create the gradient machine for inference.
   paddle_gradient_machine machine;
   CHECK(paddle_gradient_machine_create_for_inference(&machine, buf, (int)size));
   CHECK(paddle_gradient_machine_randomize_param(machine));
 
-  // Loading parameter. Uncomment the following line and change the directory.
+  // Load the trained parameters. Uncomment the following line and change the
+  // directory as needed.
   // CHECK(paddle_gradient_machine_load_parameter_from_disk(machine,
   //                                                "./some_where_to_params"));
   paddle_arguments in_args = paddle_arguments_create_none();
@@ -26,7 +29,7 @@ int main() {
   // There is only one input of this network.
   CHECK(paddle_arguments_resize(in_args, 1));
 
-  // Create input matrix.
+  // Create the input matrix.
   paddle_matrix mat = paddle_matrix_create_sparse(1, 784, 3, true, false);
   srand(time(0));
   paddle_real* array;

paddle/capi/gradient_machine.cpp

@@ -168,3 +168,13 @@ paddle_error paddle_gradient_machine_get_layer_output(
   out->args.push_back(layerOutput);
   return kPD_NO_ERROR;
 }
+
+paddle_error paddle_gradient_machine_release_layer_output(
+    paddle_gradient_machine machine) {
+  auto m = cast(machine);
+  if (m == nullptr || m->machine == nullptr) {
+    return kPD_NULLPTR;
+  }
+  m->machine->releaseOutput();
+  return kPD_NO_ERROR;
+}

paddle/capi/gradient_machine.h

@@ -113,6 +113,14 @@ paddle_gradient_machine_get_layer_output(paddle_gradient_machine machine,
                                          const char* layerName,
                                          paddle_arguments args);
 
+/**
+ * @brief Release the middle layer's output memory of the gradient machine.
+ * @param [in] gradient machine that have run a inference
+ * @return paddle_error
+ */
+PD_API paddle_error
+paddle_gradient_machine_release_layer_output(paddle_gradient_machine machine);
+
 #ifdef __cplusplus
 }
 #endif

paddle/framework/CMakeLists.txt

@@ -32,8 +32,12 @@ cc_test(threadpool_test SRCS threadpool_test.cc DEPS threadpool)
 cc_library(scope SRCS scope.cc DEPS glog threadpool)
 cc_test(scope_test SRCS scope_test.cc DEPS scope)
 
-cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor framework_proto)
-cc_test(data_transform_test SRCS data_transform_test.cc DEPS data_transform device_context)
+cc_library(data_device_transform SRCS data_device_transform.cc DEPS tensor)
+cc_library(data_type_transform SRCS data_type_transform.cc DEPS tensor)
+cc_library(data_layout_transform SRCS data_layout_transform.cc DEPS tensor math_function)
+
+cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor
+        framework_proto selected_rows data_device_transform data_type_transform data_layout_transform)
 
 cc_library(attribute SRCS attribute.cc DEPS framework_proto)
 cc_test(program_desc_test SRCS program_desc_test.cc DEPS proto_desc
@@ -41,9 +45,9 @@ device_context)
 cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute)
 cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
 cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
-cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute)
+cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
-    shape_inference data_transform)
+    shape_inference data_transform lod_tensor)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry init)
 cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog)
 
@@ -73,8 +77,10 @@ cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
 cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
 cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
 
-
-cc_library(init SRCS init.cc DEPS gflags device_context place stringpiece)
+cc_library(init SRCS init.cc DEPS gflags device_context place stringpiece operator)
 cc_test(init_test SRCS init_test.cc DEPS init)
 
 cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
+cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
+nv_test(data_device_transform_test SRCS data_device_transform_test.cu
+        DEPS operator op_registry init math_function)

paddle/framework/backward.cc

@@ -427,7 +427,8 @@ std::vector<std::unique_ptr<OpDesc>> MakeBlockBackward(
     VLOG(5) << "Making backward " << (*it)->Type() << " op";
     std::vector<std::unique_ptr<OpDesc>> op_grads;
 
-    if ((*it)->Type() == "recurrent" || (*it)->Type() == "while") {
+    if ((*it)->Type() == "recurrent" || (*it)->Type() == "while" ||
+        (*it)->Type() == "parallel_do") {
       int step_block_idx = (*it)->GetBlockAttr("sub_block");
       BlockDesc* backward_block = CreateStepBlock(program_desc, no_grad_vars,
                                                   grad_to_var, step_block_idx);

paddle/framework/block_desc.cc

@@ -53,12 +53,12 @@ VarDesc *BlockDesc::FindVarRecursive(const std::string &name) const {
   return it->second.get();
 }
 
-VarDesc *BlockDesc::FindRecursiveOrCreateVar(const std::string &name_bytes) {
+VarDesc &BlockDesc::FindRecursiveOrCreateVar(const std::string &name_bytes) {
   VarDesc *res = FindVarRecursive(name_bytes);
   if (res == nullptr) {
     res = Var(name_bytes);
   }
-  return res;
+  return *res;
 }
 
 bool BlockDesc::HasVarRecursive(const std::string &name) const {

paddle/framework/block_desc.h

@@ -57,7 +57,7 @@ class BlockDesc {
 
   VarDesc *FindVarRecursive(const std::string &name_bytes) const;
 
-  VarDesc *FindRecursiveOrCreateVar(const std::string &name_bytes);
+  VarDesc &FindRecursiveOrCreateVar(const std::string &name_bytes);
 
   bool HasVarRecursive(const std::string &var_name) const;
 

paddle/operators/detail/recv_impl.cc → paddle/framework/data_device_transform.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
@@ -12,54 +11,36 @@ 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"
+#include "paddle/framework/data_device_transform.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; });
+namespace framework {
+
+static const platform::DeviceContext* GetDeviceContext(
+    const platform::Place& src_place, const platform::Place& dst_place) {
+  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+
+  if (platform::is_gpu_place(src_place) && platform::is_cpu_place(dst_place)) {
+    return pool.Get(src_place);
+  } else if (platform::is_cpu_place(src_place) &&
+             platform::is_gpu_place(dst_place)) {
+    return pool.Get(dst_place);
+  } else {
+    PADDLE_THROW(
+        "Currently, model parallelism is only supported between CPU and CUDA");
   }
-  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();
+Tensor* DeviceTransform(const Tensor& in, const platform::Place& dst_place) {
+  VLOG(3) << "DeviceTransform in, src_place " << in.place()
+          << " dst_place: " << dst_place;
+  Tensor* out = new Tensor();
+  auto* dev_ctx = GetDeviceContext(in.place(), dst_place);
+  dev_ctx->Wait();
+  Copy(in, dst_place, *dev_ctx, out);
+  dev_ctx->Wait();
+  return out;
 }
 
-}  // namespace detail
-}  // namespace operators
+}  // namespace framework
 }  // namespace paddle

paddle/framework/data_device_transform.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/tensor.h"
+#include "paddle/framework/tensor_util.h"
+#include "paddle/platform/device_context.h"
+
+namespace paddle {
+namespace framework {
+
+Tensor* DeviceTransform(const Tensor& in, const platform::Place& dst_place);
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/data_device_transform_test.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 "gtest/gtest.h"
+
+#include "paddle/framework/init.h"
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/framework/op_info.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/elementwise_op_function.h"
+#include "paddle/operators/math/math_function.h"
+#include "paddle/platform/device_context.h"
+
+namespace paddle {
+namespace framework {
+
+template <typename T>
+struct AddFunctor {
+  inline HOSTDEVICE T operator()(T a, T b) const { return a + b; }
+};
+
+class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
+ public:
+  OpKernelTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("input", "input1 of test op");
+    AddOutput("output", "output of test op");
+    AddAttr<bool>("use_gpu", "force to use gpu kernel").SetDefault(false);
+    AddComment("This is test op");
+  }
+};
+
+class TestOpWithKernel : public OperatorWithKernel {
+ public:
+  using OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {}
+  OpKernelType GetExpectedKernelType(
+      const ExecutionContext& ctx) const override {
+    if (Attr<bool>("use_gpu")) {
+      VLOG(3) << "force use gpu kernel";
+      return OpKernelType(proto::DataType::FP32, platform::CUDAPlace(0));
+    } else {
+      VLOG(3) << "use default kernel";
+      return OpKernelType(proto::DataType::FP32,
+                          ctx.Input<Tensor>("input")->place());
+    }
+  }
+};
+
+template <typename DeviceContext, typename T>
+class TestKernel : public OpKernel<float> {
+ public:
+  void Compute(const ExecutionContext& ctx) const {
+    std::cout << ctx.op().DebugString() << std::endl;
+
+    const Tensor* input = ctx.Input<Tensor>("input");
+
+    std::cout << "input place:" << input->place() << std::endl;
+    auto* output = ctx.Output<framework::LoDTensor>("output");
+    output->Resize(input->dims());
+    output->mutable_data<T>(ctx.GetPlace());
+
+    operators::TransformFunctor<AddFunctor<T>, T, DeviceContext> functor(
+        input, input, output, ctx.template device_context<DeviceContext>(),
+        AddFunctor<T>());
+    functor.Run();
+  }
+};
+
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_OP_WITHOUT_GRADIENT(
+    test_op, paddle::framework::TestOpWithKernel,
+    paddle::framework::OpKernelTestProtoAndCheckerMaker);
+REGISTER_OP_CPU_KERNEL(
+    test_op,
+    paddle::framework::TestKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    test_op,
+    paddle::framework::TestKernel<paddle::platform::CUDADeviceContext, float>);
+
+static void BuildVar(const std::string& param_name,
+                     std::initializer_list<const char*> arguments,
+                     paddle::framework::proto::OpDesc::Var* var) {
+  var->set_parameter(param_name);
+  for (auto& arg_name : arguments) {
+    *var->mutable_arguments()->Add() = arg_name;
+  }
+}
+
+TEST(Operator, CPUtoGPU) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  InitDevices();
+
+  paddle::framework::Scope scope;
+  paddle::platform::CPUPlace cpu_place;
+
+  // create an op to run on CPU
+  paddle::framework::proto::OpDesc cpu_op_desc;
+  cpu_op_desc.set_type("test_op");
+  BuildVar("input", {"IN1"}, cpu_op_desc.add_inputs());
+  BuildVar("output", {"OUT1"}, cpu_op_desc.add_outputs());
+
+  auto cpu_op = paddle::framework::OpRegistry::CreateOp(cpu_op_desc);
+  // prepare input
+  auto* in_t = scope.Var("IN1")->GetMutable<LoDTensor>();
+  auto* src_ptr = in_t->mutable_data<float>({2, 3}, CPUPlace());
+  for (int i = 0; i < 2 * 3; ++i) {
+    src_ptr[i] = static_cast<float>(i);
+  }
+
+  // get output
+  auto* output = scope.Var("OUT1");
+  cpu_op->Run(scope, cpu_place);
+
+  auto* output_ptr = output->Get<LoDTensor>().data<float>();
+  for (int i = 0; i < 2 * 3; ++i) {
+    ASSERT_EQ(output_ptr[i], static_cast<float>(i) * 2);
+  }
+
+  // create an op to run on GPU
+  paddle::framework::proto::OpDesc gpu_op_desc;
+  gpu_op_desc.set_type("test_op");
+  BuildVar("input", {"OUT1"}, gpu_op_desc.add_inputs());
+  BuildVar("output", {"OUT2"}, gpu_op_desc.add_outputs());
+
+  auto attr = gpu_op_desc.mutable_attrs()->Add();
+  attr->set_name("use_gpu");
+  attr->set_type(paddle::framework::proto::AttrType::BOOLEAN);
+  attr->set_b(true);
+
+  auto gpu_op = paddle::framework::OpRegistry::CreateOp(gpu_op_desc);
+
+  paddle::platform::CUDAPlace cuda_place(0);
+  // get output
+  auto* output2 = scope.Var("OUT2");
+  gpu_op->Run(scope, cuda_place);
+
+  // auto* output2_ptr = output2->Get<LoDTensor>().data<float>();
+  DeviceContextPool& pool = DeviceContextPool::Instance();
+  auto dev_ctx = pool.Get(cuda_place);
+
+  paddle::framework::Tensor output_tensor;
+  Copy(output2->Get<LoDTensor>(), paddle::platform::CPUPlace(), *dev_ctx,
+       &output_tensor);
+
+  dev_ctx->Wait();
+  float* output2_ptr = output_tensor.data<float>();
+  for (int i = 0; i < 2 * 3; ++i) {
+    ASSERT_EQ(output2_ptr[i], static_cast<float>(i) * 4);
+  }
+}

paddle/framework/data_layout.h

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-#include "paddle/platform/enforce.h"
 
 #include <iostream>
 #include "paddle/platform/enforce.h"

paddle/framework/data_layout_transform.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/framework/data_layout_transform.h"
+
+#include "paddle/framework/tensor.h"
+#include "paddle/operators/math/math_function.h"
+
+namespace paddle {
+namespace framework {
+
+struct CastDataLayout {
+  CastDataLayout(const platform::DeviceContext* ctx,
+                 const std::vector<int>& axis, const framework::Tensor& in,
+                 framework::Tensor* out)
+      : in_(in), out_(out), ctx_(ctx), axis_(axis) {}
+  const framework::Tensor in_;
+  framework::Tensor* out_;
+  const platform::DeviceContext* ctx_;
+  const std::vector<int> axis_;
+
+  template <typename T>
+  void operator()() {
+    auto place = ctx_->GetPlace();
+
+    if (platform::is_cpu_place(place)) {
+      operators::math::Transpose<platform::CPUDeviceContext, T, 4> trans4;
+      auto* context = static_cast<const platform::CPUDeviceContext*>(ctx_);
+      trans4(*context, in_, out_, axis_);
+    } else {
+      PADDLE_THROW("Unsupport CPU <-> GPU!");
+    }
+  }
+};
+
+void TransDataLayout(const std::vector<int>& axis,
+                     const platform::DeviceContext* ctx,
+                     const KernelTypePair& kernel_pair, const Variable& in,
+                     Variable* out) {
+  PADDLE_ENFORCE(in.IsType<Tensor>(), "Only support Tensor transform!.");
+  PADDLE_ENFORCE(
+      platform::places_are_same_class(kernel_pair.first.place_,
+                                      kernel_pair.second.place_),
+      "TransDataLayout only support DataLayout transform on same place!");
+  PADDLE_ENFORCE(kernel_pair.first.data_type_ == kernel_pair.second.data_type_,
+                 "TransDataLayout only support Datatype are same!");
+
+  auto src = in.Get<Tensor>();
+  auto* dst = out->GetMutable<Tensor>();
+  PADDLE_ENFORCE(arity(src.dims()) == 4, "Input Arity Only Suppport 4!");
+
+  auto src_dim = src.dims();
+  std::vector<int64_t> dst_dim;
+
+  dst_dim.resize(axis.size());
+  for (size_t i = 0; i < axis.size(); i++) {
+    dst_dim[i] = src_dim[axis[i]];
+  }
+
+  dst->Resize(make_ddim(dst_dim));
+  auto place = kernel_pair.second.place_;
+  dst->mutable_data(place, src.type());
+
+  auto src_type = kernel_pair.first.data_type_;
+  framework::VisitDataType(src_type, CastDataLayout(ctx, axis, src, dst));
+
+  dst->set_layout(kernel_pair.second.data_layout_);
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/data_layout_transform.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_kernel_type.h"
+#include "paddle/framework/variable.h"
+
+namespace paddle {
+namespace framework {
+
+using KernelTypePair = std::pair<OpKernelType, OpKernelType>;
+
+void TransDataLayout(const std::vector<int>& axis,
+                     const platform::DeviceContext* ctx,
+                     const KernelTypePair& kernel_pair, const Variable& in,
+                     Variable* out);
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/data_transform.cc

@@ -11,125 +11,44 @@ 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 <functional>
 
 #include "paddle/framework/data_transform.h"
-#include "paddle/framework/lod_tensor.h"
-#include "paddle/platform/device_context.h"
+
+#include "paddle/framework/data_device_transform.h"
 
 namespace paddle {
 namespace framework {
 
-DataTransformFnMap& DataTransformFnMap::Instance() {
-  static DataTransformFnMap data_transform_map;
-  return data_transform_map;
-}
-
-auto KernelFP32 = OpKernelType(proto::DataType::FP32, platform::CPUPlace(),
-                               DataLayout::kNHWC, LibraryType::kPlain);
-
-auto KernelFP64 = OpKernelType(proto::DataType::FP64, platform::CPUPlace(),
-                               DataLayout::kNHWC, LibraryType::kPlain);
-
-auto KernelNHWC = OpKernelType(proto::DataType::FP64, platform::CPUPlace(),
-                               DataLayout::kNHWC, LibraryType::kPlain);
-
-auto KernelNCHW = OpKernelType(proto::DataType::FP64, platform::CPUPlace(),
-                               DataLayout::kNCHW, LibraryType::kPlain);
-
-void TransDataType(const platform::DeviceContext* ctx,
-                   const KernelTypePair& kernel_pair, const Variable& in,
-                   Variable* out) {
-  PADDLE_ENFORCE(in.IsType<Tensor>(), "Only Support Tensor transform!.");
-  PADDLE_ENFORCE(
-      platform::places_are_same_class(kernel_pair.first.place_,
-                                      kernel_pair.second.place_),
-      "TransDataType Only Support DataType transform on same place!");
-
-  auto src = in.Get<Tensor>();
-  auto* dst = out->GetMutable<Tensor>();
-
-  auto dims = src.dims();
-  dst->Resize(dims);
-  auto dst_type = kernel_pair.second.data_type_;
-  auto src_type = kernel_pair.first.data_type_;
-
-  switch (src_type) {
-    case proto::DataType::FP32:
-      framework::VisitDataType(dst_type, CastDataType<float>(src, dst, ctx));
-      break;
-    case proto::DataType::FP64:
-      framework::VisitDataType(dst_type, CastDataType<double>(src, dst, ctx));
-      break;
-    case proto::DataType::INT32:
-      framework::VisitDataType(dst_type, CastDataType<int>(src, dst, ctx));
-      break;
-    case proto::DataType::INT64:
-      framework::VisitDataType(dst_type, CastDataType<int64_t>(src, dst, ctx));
-      break;
-    case proto::DataType::BOOL:
-      framework::VisitDataType(dst_type, CastDataType<bool>(src, dst, ctx));
-      break;
-    default:
-      PADDLE_THROW("Not support type %d", src_type);
+Tensor* DataTransform(const OpKernelType& expected_kernel_type,
+                      const OpKernelType& kernel_type_for_var,
+                      const Tensor& input_tensor) {
+  Tensor* out = nullptr;
+  if (!platform::is_same_place(kernel_type_for_var.place_,
+                               expected_kernel_type.place_)) {
+    out = DeviceTransform(input_tensor, expected_kernel_type.place_);
   }
+  PADDLE_ENFORCE_NOT_NULL(out, "out should not be null");
+  return out;
 }
 
-void TransDataLayout(const std::vector<int>& axis,
-                     const platform::DeviceContext* ctx,
-                     const KernelTypePair& kernel_pair, const Variable& in,
-                     Variable* out) {
-  PADDLE_ENFORCE(in.IsType<Tensor>(), "Only support Tensor transform!.");
-  PADDLE_ENFORCE(
-      platform::places_are_same_class(kernel_pair.first.place_,
-                                      kernel_pair.second.place_),
-      "TransDataLayout only support DataLayout transform on same place!");
-  PADDLE_ENFORCE(kernel_pair.first.data_type_ == kernel_pair.second.data_type_,
-                 "TransDataLayout only support Datatype are same!");
-
-  auto src = in.Get<Tensor>();
-  auto* dst = out->GetMutable<Tensor>();
-  PADDLE_ENFORCE(arity(src.dims()) == 4, "Input Arity Only Suppport 4!");
-
-  auto place = kernel_pair.second.place_;
-  CopyFrom(src, place, *ctx, dst);
-
-  auto src_dim = src.dims();
-  std::vector<int64_t> dst_dim;
-
-  dst_dim.resize(axis.size());
-  for (size_t i = 0; i < axis.size(); i++) {
-    dst_dim[i] = src_dim[axis[i]];
+void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor,
+                            Variable& out_var) {
+  if (in_var.IsType<LoDTensor>()) {
+    auto& in_lod_tensor = in_var.Get<LoDTensor>();
+    auto* tran_lod_tensor = out_var.GetMutable<LoDTensor>();
+    tran_lod_tensor->set_lod(in_lod_tensor.lod());
+    tran_lod_tensor->set_layout(in_lod_tensor.layout());
+    tran_lod_tensor->ShareDataWith(tensor);
+  } else if (in_var.IsType<SelectedRows>()) {
+    auto& in_selected_rows = in_var.Get<SelectedRows>();
+    auto* trans_selected_rows = out_var.GetMutable<SelectedRows>();
+    trans_selected_rows->set_height(in_selected_rows.height());
+    trans_selected_rows->set_rows(in_selected_rows.rows());
+    trans_selected_rows->mutable_value()->ShareDataWith(tensor);
+  } else {
+    PADDLE_THROW("unknown var type");
   }
-
-  dst->Resize(make_ddim(dst_dim));
-
-  auto src_type = kernel_pair.first.data_type_;
-  framework::VisitDataType(src_type, CastDataLayout(ctx, axis, src, dst));
-
-  dst->set_layout(kernel_pair.second.data_layout_);
 }
 
 }  // namespace framework
 }  // namespace paddle
-
-namespace f = paddle::framework;
-
-namespace {
-std::vector<int> NHWC2NCHW = {0, 3, 1, 2};
-std::vector<int> NCHW2NHWC = {0, 2, 3, 1};
-}
-
-REGISTER_DATA_TRANSFORM_FN(f::KernelFP32, f::KernelFP64, f::TransDataType);
-REGISTER_DATA_TRANSFORM_FN(f::KernelNHWC, f::KernelNCHW,
-                           std::bind(f::TransDataLayout, NHWC2NCHW,
-                                     std::placeholders::_1,
-                                     std::placeholders::_2,
-                                     std::placeholders::_3,
-                                     std::placeholders::_4));
-REGISTER_DATA_TRANSFORM_FN(f::KernelNCHW, f::KernelNHWC,
-                           std::bind(f::TransDataLayout, NCHW2NHWC,
-                                     std::placeholders::_1,
-                                     std::placeholders::_2,
-                                     std::placeholders::_3,
-                                     std::placeholders::_4));

paddle/framework/data_transform.h

@@ -19,6 +19,7 @@ limitations under the License. */
 #include <vector>
 
 #include "paddle/framework/op_kernel_type.h"
+#include "paddle/framework/selected_rows.h"
 #include "paddle/framework/tensor.h"
 #include "paddle/framework/variable.h"
 #include "paddle/operators/math/math_function.h"
@@ -29,145 +30,12 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-using KernelTypePair = std::pair<OpKernelType, OpKernelType>;
+Tensor* DataTransform(const OpKernelType& expected_kernel_type,
+                      const OpKernelType& kernel_type_for_var,
+                      const Tensor& input_tensor);
 
-using DataTransformFn =
-    std::function<void(const platform::DeviceContext*, const KernelTypePair&,
-                       const Variable&, Variable*)>;
-
-struct KernelTypePairHash {
-  static void HashCombine(const OpKernelType& t, std::size_t* seed) {
-    OpKernelType::Hash kernel_type_hasher;
-    (*seed) ^= kernel_type_hasher(t) + 0x9e3779b9 + (*seed << 6) + (*seed >> 2);
-  }
-
-  size_t operator()(const KernelTypePair& kernel_pair) const {
-    std::size_t seed = 0;
-    HashCombine(kernel_pair.first, &seed);
-    HashCombine(kernel_pair.second, &seed);
-    return seed;
-  }
-};
-
-template <typename InType, typename OutType>
-struct CastDataTypeFunctor {
-  HOSTDEVICE inline OutType operator()(InType in) const {
-    return static_cast<OutType>(in);
-  }
-};
-
-template <typename InType>
-struct CastDataType {
-  CastDataType(const framework::Tensor& in, framework::Tensor* out,
-               const platform::DeviceContext* ctx)
-      : in_(in), out_(out), ctx_(ctx) {}
-  const framework::Tensor in_;
-  framework::Tensor* out_;
-  const platform::DeviceContext* ctx_;
-
-  template <typename OutType>
-  void operator()() {
-    auto place = ctx_->GetPlace();
-
-    auto* in_begin = in_.data<InType>();
-    auto numel = in_.numel();
-    auto* in_end = in_begin + numel;
-    auto* out_begin = out_->mutable_data<OutType>(place);
-
-    if (platform::is_cpu_place(place)) {
-      platform::Transform<platform::CPUDeviceContext> trans;
-      auto* context = static_cast<const platform::CPUDeviceContext*>(ctx_);
-      trans(*context, in_begin, in_end, out_begin,
-            CastDataTypeFunctor<InType, OutType>());
-    } else {
-      // TODO(dzhwinter): enhance CopyFrom CPU<->GPU with different data type?
-      PADDLE_THROW("Unsupport CPU <-> GPU!");
-    }
-  }
-};
-
-struct CastDataLayout {
-  CastDataLayout(const platform::DeviceContext* ctx,
-                 const std::vector<int>& axis, const framework::Tensor& in,
-                 framework::Tensor* out)
-      : in_(in), out_(out), ctx_(ctx), axis_(axis) {}
-  const framework::Tensor in_;
-  framework::Tensor* out_;
-  const platform::DeviceContext* ctx_;
-  const std::vector<int> axis_;
-
-  template <typename T>
-  void operator()() {
-    auto place = ctx_->GetPlace();
-
-    if (platform::is_cpu_place(place)) {
-      operators::math::Transpose<platform::CPUDeviceContext, T, 4> trans4;
-      auto* context = static_cast<const platform::CPUDeviceContext*>(ctx_);
-      trans4(*context, in_, out_, axis_);
-    } else {
-      PADDLE_THROW("Unsupport CPU <-> GPU!");
-    }
-  }
-};
-
-using DataTransformMap =
-    std::unordered_map<KernelTypePair, DataTransformFn, KernelTypePairHash>;
-
-class DataTransformFnMap {
- public:
-  static DataTransformFnMap& Instance();
-
-  bool Has(const KernelTypePair& key_pair) const {
-    return map_.find(key_pair) != map_.end();
-  }
-
-  void Insert(const OpKernelType& left, const OpKernelType& right,
-              const DataTransformFn& data_tranform_fn) {
-    Insert(std::make_pair(left, right), data_tranform_fn);
-  }
-
-  void Insert(const KernelTypePair& kernel_type_pair,
-              const DataTransformFn& data_tranform_fn) {
-    PADDLE_ENFORCE(!Has(kernel_type_pair),
-                   "KernelTypePair %s has been registered", "");
-    map_.insert({kernel_type_pair, data_tranform_fn});
-  }
-
-  const DataTransformFn& Get(const KernelTypePair& key_pair) const {
-    auto data_transformer = GetNullable(key_pair);
-    PADDLE_ENFORCE_NOT_NULL(data_transformer,
-                            "DataTransformFn should not be NULL");
-    return *data_transformer;
-  }
-
-  const DataTransformFn* GetNullable(const KernelTypePair& key_pair) const {
-    auto it = map_.find(key_pair);
-    if (it == map_.end()) {
-      return nullptr;
-    } else {
-      return &(it->second);
-    }
-  }
-
-  const DataTransformMap& Map() const { return map_; }
-
- private:
-  DataTransformFnMap() = default;
-  DataTransformMap map_;
-  DISABLE_COPY_AND_ASSIGN(DataTransformFnMap);
-};
-
-// generate unique name with __LINE__
-// refs https://stackoverflow.com/questions/1597007
-#define TOKENPASTE(x, y) x##y
-#define TOKENPASTE2(x, y) TOKENPASTE(x, y)
-#define REGISTER_DATA_TRANSFORM_FN(from, to, fn)                              \
-  static int TOKENPASTE2(fn_, __LINE__)() {                                   \
-    ::paddle::framework::DataTransformFnMap::Instance().Insert(from, to, fn); \
-    return 0;                                                                 \
-  }                                                                           \
-  static int TOKENPASTE2(var_, __LINE__) __attribute__((unused)) =            \
-      TOKENPASTE2(fn_, __LINE__)()
+void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor,
+                            Variable& out_var);
 
 }  // namespace framework
 }  // namespace paddle

paddle/framework/data_transform_test.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 <array>
-#include <vector>
-
-#include <gtest/gtest.h>
-
-#include "paddle/framework/data_transform.h"
-#include "paddle/platform/device_context.h"
-
-namespace paddle {
-namespace framework {
-using namespace platform;
-
-/**
- * @brief cross validation of different kernel type transform
- *  We use four bit map represent different combination.
- *  If the field has multiple possible value, only choose two of them.
- *  For DataType, only test the FP32(float), FP64(double).
- *  e.g. 0000 -> FP32, CPUPlace, kNHWC, kPlain
- *       1111 -> FP64, GPUPlace, kNCHW, kMKLDNN
- */
-
-std::array<proto::DataType, 2> kDataType = {
-    {proto::DataType::FP32, proto::DataType::FP64}};
-
-std::array<Place, 2> kPlace = {{CPUPlace(), CUDAPlace(0)}};
-
-std::array<DataLayout, 2> kDataLayout = {{
-    DataLayout::kNHWC, DataLayout::kNCHW,
-}};
-
-std::array<LibraryType, 2> kLibraryType = {{
-    LibraryType::kPlain, LibraryType::kMKLDNN,
-}};
-
-OpKernelType GenFromBit(const std::vector<bool> bits) {
-  return OpKernelType(kDataType[bits[0]], kPlace[bits[1]], kDataLayout[bits[2]],
-                      kLibraryType[bits[3]]);
-}
-
-int test_value = 0;
-
-auto kernel0 = GenFromBit({0, 0, 0, 0});
-auto kernel1 = GenFromBit({0, 0, 0, 1});
-auto kernel2 = GenFromBit({0, 0, 1, 0});
-auto kernel3 = GenFromBit({0, 0, 1, 1});
-
-void TransDataType_t(const platform::DeviceContext* ctx,
-                     const KernelTypePair& p, const Variable& in,
-                     Variable* out) {
-  test_value++;
-}
-
-void TransDataLayout_t(const platform::DeviceContext* ctx,
-                       const KernelTypePair& p, const Variable& in,
-                       Variable* out) {
-  test_value--;
-}
-
-void TransLibraryType_t(const platform::DeviceContext* ctx,
-                        const KernelTypePair& p, const Variable& in,
-                        Variable* out) {
-  test_value += 2;
-}
-
-}  // namespace framework
-}  // namespace paddle
-
-namespace frw = paddle::framework;
-
-REGISTER_DATA_TRANSFORM_FN(frw::kernel0, frw::kernel1, frw::TransDataType_t);
-REGISTER_DATA_TRANSFORM_FN(frw::kernel1, frw::kernel2, frw::TransDataLayout_t);
-REGISTER_DATA_TRANSFORM_FN(frw::kernel0, frw::kernel2, frw::TransLibraryType_t);
-
-TEST(DataTransform, Register) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
-
-  auto& instance = DataTransformFnMap::Instance();
-  paddle::framework::Variable in;
-  paddle::framework::Variable out;
-
-  DeviceContext* ctx = new CPUDeviceContext();
-  auto pair0 = std::make_pair(frw::kernel0, frw::kernel1);
-  instance.Get(pair0)(ctx, pair0, in, &out);
-  ASSERT_EQ(test_value, 1);
-
-  auto pair1 = std::make_pair(frw::kernel1, frw::kernel2);
-  instance.Get(pair1)(ctx, pair1, in, &out);
-  ASSERT_EQ(test_value, 0);
-
-  auto pair3 = std::make_pair(frw::kernel0, frw::kernel2);
-  instance.Get(pair3)(ctx, pair3, in, &out);
-  ASSERT_EQ(test_value, 2);
-}
-
-TEST(DataTransform, DataLayout) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
-
-  auto& instance = DataTransformFnMap::Instance();
-  Variable in;
-  Variable out;
-  Tensor* src = in.GetMutable<Tensor>();
-  src->mutable_data<double>(make_ddim({2, 3, 1, 2}), CPUPlace());
-  src->set_layout(DataLayout::kNHWC);
-
-  DeviceContext* ctx = new CPUDeviceContext();
-
-  {
-    auto kernel1 = GenFromBit({1, 0, 0, 0});
-    auto kernel2 = GenFromBit({1, 0, 1, 0});
-    auto pair0 = std::make_pair(kernel1, kernel2);
-    instance.Get(pair0)(ctx, pair0, in, &out);
-  }
-
-  Tensor dst = out.Get<Tensor>();
-
-  EXPECT_TRUE(dst.layout() == DataLayout::kNCHW);
-  EXPECT_TRUE(dst.dims() == make_ddim({2, 2, 3, 1}));
-
-  {
-    auto kernel1 = GenFromBit({1, 0, 1, 0});
-    auto kernel2 = GenFromBit({1, 0, 0, 0});
-    auto pair0 = std::make_pair(kernel1, kernel2);
-    instance.Get(pair0)(ctx, pair0, out, &in);
-  }
-
-  EXPECT_TRUE(src->layout() == DataLayout::kNHWC);
-  EXPECT_TRUE(src->dims() == make_ddim({2, 3, 1, 2}));
-}
-
-TEST(DataTransform, DataType) {
-  using namespace paddle::framework;
-  using namespace paddle::platform;
-
-  auto& instance = DataTransformFnMap::Instance();
-  DeviceContext* ctx = new CPUDeviceContext();
-
-  Variable in;
-  Variable out;
-  Tensor* src = in.GetMutable<Tensor>();
-  float* ptr = src->mutable_data<float>(make_ddim({2, 3}), CPUPlace());
-  for (int i = 0; i < 6; ++i) {
-    ptr[i] = i / 3;
-  }
-
-  {
-    auto kernel1 = GenFromBit({0, 0, 0, 0});
-    auto kernel2 = GenFromBit({1, 0, 0, 0});
-    auto pair0 = std::make_pair(kernel1, kernel2);
-    instance.Get(pair0)(ctx, pair0, in, &out);
-  }
-  Tensor dst = out.Get<Tensor>();
-  EXPECT_TRUE(dst.data<double>() != nullptr);
-}

paddle/framework/data_type_transform.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/framework/data_type_transform.h"
+
+#include "paddle/framework/selected_rows.h"
+#include "paddle/platform/transform.h"
+
+namespace paddle {
+namespace framework {
+
+template <typename InType, typename OutType>
+struct CastDataTypeFunctor {
+  HOSTDEVICE inline OutType operator()(InType in) const {
+    return static_cast<OutType>(in);
+  }
+};
+
+template <typename InType>
+struct CastDataType {
+  CastDataType(const framework::Tensor& in, framework::Tensor* out,
+               const platform::DeviceContext* ctx)
+      : in_(in), out_(out), ctx_(ctx) {}
+  const framework::Tensor in_;
+  framework::Tensor* out_;
+  const platform::DeviceContext* ctx_;
+
+  template <typename OutType>
+  void operator()() {
+    auto place = ctx_->GetPlace();
+
+    auto* in_begin = in_.data<InType>();
+    auto numel = in_.numel();
+    auto* in_end = in_begin + numel;
+    auto* out_begin = out_->mutable_data<OutType>(place);
+
+    if (platform::is_cpu_place(place)) {
+      platform::Transform<platform::CPUDeviceContext> trans;
+      auto* context = static_cast<const platform::CPUDeviceContext*>(ctx_);
+      trans(*context, in_begin, in_end, out_begin,
+            CastDataTypeFunctor<InType, OutType>());
+    } else {
+      // TODO(dzhwinter): enhance Copy CPU<->GPU with different data type?
+      PADDLE_THROW("Unsupport CPU <-> GPU!");
+    }
+  }
+};
+
+void TransDataType(const platform::DeviceContext* ctx,
+                   const KernelTypePair& kernel_pair, const Variable& in,
+                   Variable* out) {
+  PADDLE_ENFORCE(in.IsType<Tensor>(), "Only Support Tensor transform!.");
+  PADDLE_ENFORCE(
+      platform::places_are_same_class(kernel_pair.first.place_,
+                                      kernel_pair.second.place_),
+      "TransDataType Only Support DataType transform on same place!");
+
+  auto src = in.Get<Tensor>();
+  auto* dst = out->GetMutable<Tensor>();
+
+  auto dims = src.dims();
+  dst->Resize(dims);
+  auto dst_type = kernel_pair.second.data_type_;
+  auto src_type = kernel_pair.first.data_type_;
+
+  switch (src_type) {
+    case proto::DataType::FP32:
+      framework::VisitDataType(dst_type, CastDataType<float>(src, dst, ctx));
+      break;
+    case proto::DataType::FP64:
+      framework::VisitDataType(dst_type, CastDataType<double>(src, dst, ctx));
+      break;
+    case proto::DataType::INT32:
+      framework::VisitDataType(dst_type, CastDataType<int>(src, dst, ctx));
+      break;
+    case proto::DataType::INT64:
+      framework::VisitDataType(dst_type, CastDataType<int64_t>(src, dst, ctx));
+      break;
+    case proto::DataType::BOOL:
+      framework::VisitDataType(dst_type, CastDataType<bool>(src, dst, ctx));
+      break;
+    default:
+      PADDLE_THROW("Not support type %d", src_type);
+  }
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/data_type_transform.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_kernel_type.h"
+#include "paddle/framework/variable.h"
+#include "paddle/platform/device_context.h"
+
+namespace paddle {
+namespace framework {
+
+using KernelTypePair = std::pair<OpKernelType, OpKernelType>;
+
+void TransDataType(const platform::DeviceContext* ctx,
+                   const KernelTypePair& kernel_pair, const Variable& in,
+                   Variable* out);
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/details/cow_ptr.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+#include <memory>
+#include <thread>
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+// Change it to thread safe flags if needed.
+class ThreadUnsafeOwnershipFlags {
+ public:
+  ThreadUnsafeOwnershipFlags(bool flag) : flag_(flag) {}
+
+  ThreadUnsafeOwnershipFlags(const ThreadUnsafeOwnershipFlags& other) = delete;
+  ThreadUnsafeOwnershipFlags& operator=(
+      const ThreadUnsafeOwnershipFlags& other) = delete;
+  ThreadUnsafeOwnershipFlags(ThreadUnsafeOwnershipFlags&& other) = default;
+
+  void SetOwnership(bool flag) { flag_ = flag; }
+
+  // Invoke the callback if it is not owned.
+  template <typename Callback>
+  void AcquireOwnershipOnce(Callback acquire) {
+    if (!flag_) {
+      acquire();
+      flag_ = true;
+    }
+  }
+
+ private:
+  bool flag_;
+};
+
+// Copy-On-Write pointer.
+// It will hold a T* pointer, and only copy once when `MutableData` is invoked.
+//
+// The template parameter OwnershipFlags should have:
+//   * a constructor takes a bool. True if own.
+//   * SetOwnership(bool flag).
+//   * AcquireOwnershipOnce(Callback). It will invoke the callback if it is not
+//     owned.
+//
+// https://en.wikipedia.org/wiki/Copy-on-write
+template <typename T, typename OwnershipFlags = ThreadUnsafeOwnershipFlags>
+class COWPtr {
+ public:
+  // Ctor from raw pointer.
+  explicit COWPtr(T* ptr) : payload_(ptr), ownership_{true} {}
+
+  // Move methods. Steal ownership from origin
+  COWPtr(COWPtr&& other)
+      : payload_(other.payload_), ownership_{std::move(other.ownership_)} {}
+  COWPtr& operator=(COWPtr&& origin) = default;
+
+  // Copy methods. Not own payload
+  COWPtr(const COWPtr& other) : payload_(other.payload_), ownership_{false} {}
+  COWPtr& operator=(const COWPtr& other) {
+    payload_ = other.payload_;
+    ownership_.SetOwnership(false);
+    return *this;
+  }
+
+  // Access read only data.
+  const T& Data() const { return *payload_; }
+
+  // Access mutable data. If the data is not owned, the data will be copied
+  // before.
+  T* MutableData() {
+    ownership_.AcquireOwnershipOnce(
+        [this] { payload_.reset(new T(*payload_)); });
+    return payload_.get();
+  }
+
+ private:
+  // Actual data pointer.
+  std::shared_ptr<T> payload_;
+
+  // Ownership flag.
+  OwnershipFlags ownership_;
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/framework/details/cow_ptr_test.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/framework/details/cow_ptr.h"
+#include "gtest/gtest.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+TEST(COWPtr, all) {
+  COWPtr<int> ptr(new int{0});
+  ASSERT_EQ(ptr.Data(), 0);
+  COWPtr<int> ptr2 = ptr;
+  ASSERT_EQ(ptr2.Data(), 0);
+  ASSERT_EQ(&ptr2.Data(), &ptr.Data());
+  *ptr2.MutableData() = 10;
+  ASSERT_EQ(ptr.Data(), 0);
+  ASSERT_EQ(ptr2.Data(), 10);
+}
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/framework/executor.cc

@@ -21,6 +21,7 @@ limitations under the License. */
 #include "paddle/framework/lod_rank_table.h"
 #include "paddle/framework/lod_tensor_array.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/platform/place.h"
 
 DEFINE_bool(check_nan_inf, false,
             "Checking whether operator produce NAN/INF or not. It will be "
@@ -49,10 +50,13 @@ static void CreateTensor(Variable* var, proto::VarDesc::VarType var_type) {
     var->GetMutable<LoDRankTable>();
   } else if (var_type == proto::VarDesc::LOD_TENSOR_ARRAY) {
     var->GetMutable<LoDTensorArray>();
+  } else if (var_type == proto::VarDesc::PLACE_LIST) {
+    var->GetMutable<platform::PlaceList>();
   } else {
     PADDLE_THROW(
         "Variable type %d is not in "
-        "[LoDTensor, SelectedRows, FEED_MINIBATCH, FETCH_LIST, LOD_RANK_TABLE]",
+        "[LoDTensor, SelectedRows, FEED_MINIBATCH, FETCH_LIST, LOD_RANK_TABLE,"
+        " PLACE_LIST]",
         var_type);
   }
 }
@@ -111,7 +115,7 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
 
   for (auto& op_desc : block.AllOps()) {
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
-    VLOG(3) << op->DebugString();
+    VLOG(3) << op->DebugStringEx(local_scope);
     op->Run(*local_scope, place_);
     if (FLAGS_check_nan_inf) {
       for (auto& vname : op->OutputVars(true)) {

paddle/framework/framework.proto

@@ -123,6 +123,7 @@ message VarDesc {
     STEP_SCOPES = 5;
     LOD_RANK_TABLE = 6;
     LOD_TENSOR_ARRAY = 7;
+    PLACE_LIST = 8;
   }
   required string name = 1;
   required VarType type = 2;

paddle/framework/grad_op_desc_maker.h

@@ -87,7 +87,11 @@ class GradOpDescMakerBase {
     auto onames = this->Output(name);
     ret_val.reserve(onames.size());
     std::transform(onames.begin(), onames.end(), std::back_inserter(ret_val),
-                   GradVarName);
+                   [this](const std::string& fwd_var_name) -> std::string {
+                     auto g_name = GradVarName(fwd_var_name);
+                     (*this->grad_to_var_)[g_name] = fwd_var_name;
+                     return g_name;
+                   });
     return ret_val;
   }
 

paddle/framework/init.cc

@@ -11,10 +11,12 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
+#include <string.h>  // for strdup
 #include <algorithm>
 #include <string>
 
 #include "paddle/framework/init.h"
+#include "paddle/framework/operator.h"
 #include "paddle/platform/device_context.h"
 #include "paddle/platform/place.h"
 #include "paddle/string/piece.h"
@@ -24,7 +26,6 @@ namespace framework {
 
 std::once_flag gflags_init_flag;
 
-// TODO(qijun) move init gflags to init.cc
 void InitGflags(std::vector<std::string> &argv) {
   std::call_once(gflags_init_flag, [&]() {
     int argc = argv.size();
@@ -40,43 +41,29 @@ void InitGflags(std::vector<std::string> &argv) {
   });
 }
 
-bool InitDevices(const std::vector<std::string> &devices) {
-  // device format
-  // CPU
-  // GPU:1
-  // TODO(dzhwinter) : add device format annotation for users.
+void InitDevices() {
+  /*Init all avaiable devices by default */
+
   std::vector<platform::Place> places;
-  for (auto &device : devices) {
-    auto p = string::Piece(device);
-    if (string::HasPrefix(p, "CPU")) {
   places.emplace_back(platform::CPUPlace());
-    } else if (string::HasPrefix(p, "GPU")) {
+
 #ifdef PADDLE_WITH_CUDA
-      auto pos = string::RFind(p, ':', string::Piece::npos);
-      auto number = device.substr(pos + 1);
-      places.emplace_back(platform::CUDAPlace(std::stoi(number)));
+  int count = platform::GetCUDADeviceCount();
+  for (int i = 0; i < count; ++i) {
+    places.emplace_back(platform::CUDAPlace(i));
+  }
 #else
   LOG(WARNING)
       << "'GPU' is not supported, Please re-compile with WITH_GPU option";
 #endif
-    } else {
-      return false;
-    }
-  }
 
-  if (std::find_if(places.begin(), places.end(),
-                   [&](const platform::Place &place) {
-                     return platform::is_cpu_place(place);
-                   }) == places.end()) {
-    places.emplace_back(platform::CPUPlace());
-    LOG(WARNING) << "Not specified CPU device, create CPU by Default.";
-  }
   platform::DeviceContextPool::Init(places);
-  return true;
 }
 
 void InitGLOG(const std::string &prog_name) {
-  google::InitGoogleLogging(prog_name.c_str());
+  // glog will not hold the ARGV[0] inside.
+  // Use strdup to alloc a new string.
+  google::InitGoogleLogging(strdup(prog_name.c_str()));
   google::InstallFailureSignalHandler();
 }
 

paddle/framework/init.h

@@ -24,7 +24,7 @@ void InitGflags(std::vector<std::string> &argv);
 
 void InitGLOG(const std::string &prog_name);
 
-bool InitDevices(const std::vector<std::string> &devices);
+void InitDevices();
 
 }  // namespace framework
 }  // namespace paddle

paddle/framework/init_test.cc

@@ -14,18 +14,13 @@ limitations under the License. */
 #include "gtest/gtest.h"
 
 #include "paddle/framework/init.h"
+#include "paddle/platform/device_context.h"
 
-TEST(Init, InitDevices) {
+TEST(InitDevices, CPU) {
   using paddle::framework::InitDevices;
-  std::vector<std::string> ds1 = {"CPU"};
-  ASSERT_EQ(InitDevices(ds1), true);
+  using paddle::platform::DeviceContextPool;
 
-#ifdef PADDLE_WITH_CUDA
-  std::vector<std::string> ds2 = {"CPU", "GPU:0", "GPU:1"};
-  ASSERT_EQ(InitDevices(ds2), true);
-
-  // test re-init
-  std::vector<std::string> ds3 = {"GPU:0", "GPU:1"};
-  ASSERT_EQ(InitDevices(ds3), true);
-#endif
+  InitDevices();
+  DeviceContextPool& pool = DeviceContextPool::Instance();
+  ASSERT_GE(pool.size(), 1U);
 }