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

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