Tensorflow C++ 从训练到部署(2)：简单图的保存、读取与 CMake 编译

2018年9月19日

经过了上一篇文章，我们已经成功编译了 tensorflow c++ 的系统库文件并且安装到系统目录下了。这里我们将使用这个编译好的库进行基本的 C++ 模型加载执行等操作。
注意，在本篇文章会使用 Tensorflow 的 Python API，因为比较简单，这里不做介绍，安装详见官网教程：
https://www.tensorflow.org/install/

0、系统环境
Ubuntu 16.04
Tensorflow 1.10.1 （安装详见官网，建议使用 pip 方式安装）

1、一个简单网络的保存
只有 c = a * b 的网络：

#!/usr/bin/env python

import tensorflow as tf
import numpy as np

with tf.Session() as sess:

    a=tf.placeholder(tf.float32,shape=None, name='a')
    b=tf.placeholder(tf.float32,shape=None, name='b')
    c = tf.multiply(a, b, name='c')

    sess.run(tf.global_variables_initializer())

    tf.train.write_graph(sess.graph_def, 'model/', 'simple.pb', as_text=False)

    res = sess.run(c, feed_dict={'a:0': 2.0, 'b:0': 3.0})
    print("res = ", res)

#!/usr/bin/env python

import tensorflow as tf

import numpy as np

with tf.Session() as sess:

a=tf.placeholder(tf.float32,shape=None, name='a')

b=tf.placeholder(tf.float32,shape=None, name='b')

c = tf.multiply(a, b, name='c')

sess.run(tf.global_variables_initializer())

tf.train.write_graph(sess.graph_def, 'model/', 'simple.pb', as_text=False)

res = sess.run(c, feed_dict={'a:0': 2.0, 'b:0': 3.0})

print("res = ", res)

在这段代码中，我们构建了一个非常简单的“网络”：
$c = a * b$
并且给 a 和 b 赋予了初值。这一网络虽然简单，但是更复杂的网络也是类似的道理，我们需要保存的无非是计算图的结构，和图中的参数。这里边我们 tf.train.write_graph 保存的仅仅是图的结构。注意其中每个 placeholder 的 name 非常重要，这是我们后面输入和获取这些值的基础。

在这一例子中，有两点注意：
a）tf.train.write_graph 函数就是指定了保存图的路径，as_text=False 表明用二进制格式保存（默认是文本格式保存）。
b）res = sess.run(c, feed_dict={'a:0': 2.0, 'b:0': 3.0}) 中，我们定义的 placeholder 的名字是 a 和 b，但是对于 tensorflow 来说 shape=None 的相当于一个 1x1 的向量，所以我们还是要指定一个下标表示 a 和 b 是这个向量的第一个元素，这里 tensorflow 用 a:0 和 b:0 表示。如果 a 和 b 本来就是一个多维向量，那么可以直接取出整个向量。

我们将这段代码保存为 simple/simple_net.py 文件并执行：

python simple_net.py

1	python simple_net.py

正常情况下会在 model 文件夹下保存一个 simple.pb 的文件，并且输出一个测试运行的结果：
res = 6.0

PS：你也可以指定为文本格式保存，通常文本格式我们保存为 .pbtxt 后缀，例如：

tf.train.write_graph(sess.graph_def, 'model/', 'simple.pbtxt', as_text=False)

1	tf.train.write_graph(sess.graph_def, 'model/', 'simple.pbtxt', as_text=False)

保存出来就是这样的，看起来很简单，不过通常我们还是用二进制格式比较高效：

node {
  name: "a"
  op: "Placeholder"
  attr {
    key: "dtype"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "shape"
    value {
      shape {
        unknown_rank: true
      }
    }
  }
}
node {
  name: "b"
  op: "Placeholder"
  attr {
    key: "dtype"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "shape"
    value {
      shape {
        unknown_rank: true
      }
    }
  }
}
node {
  name: "c"
  op: "Mul"
  input: "a"
  input: "b"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
}
node {
  name: "init"
  op: "NoOp"
}
versions {
  producer: 26
}

node {

op: "Placeholder"

attr {

key: "dtype"

value {

type: DT_FLOAT

}

attr {

key: "shape"

value {

shape {

unknown_rank: true

}

node {

op: "Placeholder"

attr {

key: "dtype"

value {

type: DT_FLOAT

}

attr {

key: "shape"

value {

shape {

unknown_rank: true

}

node {

op: "Mul"

input: "a"

input: "b"

attr {

key: "T"

value {

type: DT_FLOAT

}

node {

op: "NoOp"

}

versions {

producer: 26

}

2、使用 Python 读取这个网络
首先我们来看下用 Python 如何读取这个网络并进行计算：

#!/usr/bin/env python

import tensorflow as tf
import numpy as np

import numpy as np
from tensorflow.python.platform import gfile

# Initialize a tensorflow session
with tf.Session() as sess:
    # Load the protobuf graph
    with gfile.FastGFile("model/simple.pb",'rb') as f:
        graph_def = tf.GraphDef()
        graph_def.ParseFromString(f.read())
        # Add the graph to the session
        tf.import_graph_def(graph_def, name='')

    # Get graph
    graph = tf.get_default_graph()

    # Get tensor from graph
    c = graph.get_tensor_by_name("c:0")

    # Run the session, evaluating our "c" operation from the graph
    res = sess.run(c, feed_dict={'a:0': 2.0, 'b:0': 3.0})

    print("res = ", res)

#!/usr/bin/env python

import tensorflow as tf

import numpy as np

from tensorflow.python.platform import gfile

# Initialize a tensorflow session

with tf.Session() as sess:

# Load the protobuf graph

with gfile.FastGFile("model/simple.pb",'rb') as f:

graph_def = tf.GraphDef()

graph_def.ParseFromString(f.read())

# Add the graph to the session

tf.import_graph_def(graph_def, name='')

# Get graph

graph = tf.get_default_graph()

# Get tensor from graph

c = graph.get_tensor_by_name("c:0")

# Run the session, evaluating our "c" operation from the graph

res = sess.run(c, feed_dict={'a:0': 2.0, 'b:0': 3.0})

print("res = ", res)

我们将这段代码保存为 simple/load_simple_net.py 文件并执行：

python load_simple_net.py

1	python load_simple_net.py

如果运行成功的话，可以看到输出结果：
res = 6.0

3、使用 C++ 读取这个网络
现在我们使用 C++ API 来读取这个图并使用图进行计算：

#include "tensorflow/core/public/session.h"
#include "tensorflow/core/platform/env.h"

using namespace tensorflow;

/**
 * @brief simple model for click through rate prediction
 * @details [long description]
 *
 * @param argv[1] graph protobuf
 *
 * @return [description]
 */
int main(int argc, char* argv[]) {
    // Initialize a tensorflow session
    Session* session;
    Status status = NewSession(SessionOptions(), &session);
    if (!status.ok()) {
        std::cerr << status.ToString() << std::endl;
        return 1;
    } else {
        std::cout << "Session created successfully" << std::endl;
    }

    if (argc != 2)
    {
        std::cerr << std::endl << "Usage: ./project path_to_graph.pb" << std::endl;
        return 1;
    }

    // Load the protobuf graph
    GraphDef graph_def;
    std::string graph_path = argv[1];
    status = ReadBinaryProto(Env::Default(), graph_path, &graph_def);
    if (!status.ok()) {
        std::cerr << status.ToString() << std::endl;
        return 1;
    } else {
        std::cout << "Load graph protobuf successfully" << std::endl;
    }

    // Add the graph to the session
    status = session->Create(graph_def);
    if (!status.ok()) {
        std::cerr << status.ToString() << std::endl;
        return 1;
    } else {
        std::cout << "Add graph to session successfully" << std::endl;
    }

    // Setup inputs and outputs:

    // Our graph doesn't require any inputs, since it specifies default values,
    // but we'll change an input to demonstrate.
    Tensor a(DT_FLOAT, TensorShape());
    a.scalar<float>()() = 3.0;

    Tensor b(DT_FLOAT, TensorShape());
    b.scalar<float>()() = 2.0;

    std::vector<std::pair<string, tensorflow::Tensor>> inputs = {
            { "a:0", a },
            { "b:0", b },
    };

    // The session will initialize the outputs
    std::vector<tensorflow::Tensor> outputs;

    // Run the session, evaluating our "c" operation from the graph
    status = session->Run(inputs, {"c:0"}, {}, &outputs);
    if (!status.ok()) {
        std::cerr << status.ToString() << std::endl;
        return 1;
    } else {
        std::cout << "Run session successfully" << std::endl;
    }

    // Grab the first output (we only evaluated one graph node: "c")
    // and convert the node to a scalar representation.
    auto output_c = outputs[0].scalar<float>();

    // (There are similar methods for vectors and matrices here:
    // https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/public/tensor.h)

    // Print the results
    std::cout << outputs[0].DebugString() << std::endl; // Tensor<type: float shape: [] values: 30>
    std::cout << "output value: " << output_c() << std::endl; // 30

    // Free any resources used by the session
    session->Close();

    return 0;
}

#include "tensorflow/core/public/session.h"

#include "tensorflow/core/platform/env.h"

using namespace tensorflow;

/**

* @brief simple model for click through rate prediction

* @details [long description]

* @param argv[1] graph protobuf

* @return [description]

int main(int argc, char* argv[]) {

// Initialize a tensorflow session

Session* session;

Status status = NewSession(SessionOptions(), &session);

if (!status.ok()) {

std::cerr << status.ToString() << std::endl;

return 1;

} else {

std::cout << "Session created successfully" << std::endl;

}

if (argc != 2)

{

std::cerr << std::endl << "Usage: ./project path_to_graph.pb" << std::endl;

return 1;

}

// Load the protobuf graph

GraphDef graph_def;

std::string graph_path = argv[1];

status = ReadBinaryProto(Env::Default(), graph_path, &graph_def);

if (!status.ok()) {

std::cerr << status.ToString() << std::endl;

return 1;

} else {

std::cout << "Load graph protobuf successfully" << std::endl;

}

// Add the graph to the session

status = session->Create(graph_def);

if (!status.ok()) {

std::cerr << status.ToString() << std::endl;

return 1;

} else {

std::cout << "Add graph to session successfully" << std::endl;

}

// Setup inputs and outputs:

// Our graph doesn't require any inputs, since it specifies default values,

// but we'll change an input to demonstrate.

Tensor a(DT_FLOAT, TensorShape());

a.scalar<float>()() = 3.0;

Tensor b(DT_FLOAT, TensorShape());

b.scalar<float>()() = 2.0;

std::vector<std::pair<string, tensorflow::Tensor>> inputs = {

{ "a:0", a },

{ "b:0", b },

};

// The session will initialize the outputs

std::vector<tensorflow::Tensor> outputs;

// Run the session, evaluating our "c" operation from the graph

status = session->Run(inputs, {"c:0"}, {}, &outputs);

if (!status.ok()) {

std::cerr << status.ToString() << std::endl;

return 1;

} else {

std::cout << "Run session successfully" << std::endl;

}

// Grab the first output (we only evaluated one graph node: "c")

// and convert the node to a scalar representation.

auto output_c = outputs[0].scalar<float>();

// (There are similar methods for vectors and matrices here:

// https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/public/tensor.h)

// Print the results

std::cout << outputs[0].DebugString() << std::endl; // Tensor<type: float shape: [] values: 30>

std::cout << "output value: " << output_c() << std::endl; // 30

// Free any resources used by the session

session->Close();

return 0;

}

将这一文件保存为 simple/load_simple_net.cpp。

4、编译运行 Tensorflow C++ API
1）编写 CMakeLists.txt
首先我们编写一个 CMakeLists.txt 来编译这个文件：

cmake_minimum_required(VERSION 3.5)
project(tensorflow_cpp)

set(CMAKE_CXX_STANDARD 11)

find_package(OpenCV 3.0 QUIET)
if(NOT OpenCV_FOUND)
    find_package(OpenCV 2.4.3 QUIET)
    if(NOT OpenCV_FOUND)
        message(FATAL_ERROR "OpenCV > 2.4.3 not found.")
    endif()
endif()

set(TENSORFLOW_INCLUDES
        /usr/local/include/tf/
        /usr/local/include/tf/bazel-genfiles
        /usr/local/include/tf/tensorflow/
        /usr/local/include/tf/tensorflow/third-party)

set(TENSORFLOW_LIBS
        /usr/local/lib/libtensorflow_cc.so
        /usr/local/lib//libtensorflow_framework.so)


include_directories(
        ${TENSORFLOW_INCLUDES}
        ${PROJECT_SOURCE_DIR}/third_party/eigen3
)

add_executable(load_simple_net simple/load_simple_net.cpp)
target_link_libraries(load_simple_net
        ${TENSORFLOW_LIBS}
        ${OpenCV_LIBS}
        )

cmake_minimum_required(VERSION 3.5)

project(tensorflow_cpp)

set(CMAKE_CXX_STANDARD 11)

find_package(OpenCV 3.0 QUIET)

if(NOT OpenCV_FOUND)

find_package(OpenCV 2.4.3 QUIET)

if(NOT OpenCV_FOUND)

message(FATAL_ERROR "OpenCV > 2.4.3 not found.")

endif()

set(TENSORFLOW_INCLUDES

/usr/local/include/tf/

/usr/local/include/tf/bazel-genfiles

/usr/local/include/tf/tensorflow/

/usr/local/include/tf/tensorflow/third-party)

set(TENSORFLOW_LIBS

/usr/local/lib/libtensorflow_cc.so

/usr/local/lib//libtensorflow_framework.so)

include_directories(

${TENSORFLOW_INCLUDES}

${PROJECT_SOURCE_DIR}/third_party/eigen3

)

add_executable(load_simple_net simple/load_simple_net.cpp)

target_link_libraries(load_simple_net

${TENSORFLOW_LIBS}

${OpenCV_LIBS}

)

将这一文件保存为 CMakeLists.txt。

2）下载 eigen3
为了避免产生问题，我们这里要依赖自己的 eigen3 而不要用 tensorflow 下载的 eigen3，我们下载一个 eigen3 并放在 third_party/eigen3 目录，这里我使用的是 3.3.5 版本：
http://bitbucket.org/eigen/eigen/get/3.3.5.tar.bz2

现在我们将所有文件创建完毕，你的目录结构应该是这样的：
screenshot-from-2018-09-19-17-06-47

3）编译 & 运行
在工程根目录下使用如下命令编译运行：

mkdir build
cd build
cmake ..
make

mkdir build

cd build

cmake ..

make

然后在 build 目录运行：

./load_simple_net ../simple/model/simple.pb

1	./load_simple_net ../simple/model/simple.pb

如果运行成功会显示：
Session created successfully
Load graph protobuf successfully
Add graph to session successfully
Run session successfully
Tensor
output value: 6

到此你已经成功用　C++ API 加载了之前定义的图（尽管只是一个乘法）并利用这个图进行了计算。

后面我们将在此基础上尝试真正的神经网络的运算。

本文完整代码参见　Github：
https://github.com/skylook/tensorflow_cpp

参考文献
[1] https://github.com/formath/tensorflow-predictor-cpp
[2] https://github.com/zhangcliff/tensorflow-c-mnist
[3] https://blog.csdn.net/ztf312/article/details/72859075

Tensorflow C++ 从训练到部署系列文章目录

Tensorflow C++ 从训练到部署(3)：使用 Keras 训练和部署 CNN

Tensorflow C++ 从训练到部署(2)：简单图的保存、读取与 CMake 编译

Tensorflow C++ 从训练到部署(1)：环境搭建

About The Author

skylook

增强现实、图像识别技术爱好者。

Original content here is published under these license terms:		X

License Type:	Read Only

License Abstract:	You may read the original content in the context in which it is published (at this web address). No other copying or use is permitted without written agreement from the author.

技术刘

Tensorflow C++ 从训练到部署(2)：简单图的保存、读取与 CMake 编译

Tensorflow C++ 从训练到部署系列文章目录

About The Author

skylook

Add a Comment

Tensorflow C++ 从训练到部署系列文章目录

Related Posts

About The Author

skylook

Add a Comment