Build TF on win

23 Jun 2018

准备工作

win：windows server 2016（win10应该也没问题，成功build和系统关系不大）

cMake：latest

TF源码：master -> 1.10（until 2018.9.1）

eigen：latest

python：3.5.0

visual studio：2015

（可供参考的视频）

1. build tf

终端进入TensorFlow的cmake文件夹,例如E:\tf-r.1.10\tensorflow\tensorflow\contrib\cmake
新建一个文件夹并进入build

执行：

cmake .. -A x64 -DCMAKE_BUILD_TYPE=Release ^ 
More? -DSWIG_EXECUTABLE=E:/swigwin-3.0.12/swig.exe ^ 
More? -DPYTHON_EXECUTABLE=C:/Users/yanfu/AppData/Local/Programs/Python/Python35/python.exe ^
More? -DPYTHON_LIBRARIES=C:/Users/yanfu/AppData/Local/Programs/Python/Python35/libs/python35.lib ^
More? -Dtensorflow_BUILD_SHARED_LIB=ON

MSBuild：

MSBuild ^
/m:1 ^
/p:CL_MPCount=1 ^
/p:Configuration=Release ^
/p:Platform=x64 ^
/p:PreferredToolArchitecture=x64 ALL_BUILD.vcxproj ^
/filelogger

PS.MSBuild 是增量build的，期间很可能部分项目会有提示内存不足的错误，多按照上面的MSBuild命令build几次直到0error，build完成大概需要2~3小时，这个得看机器内存和处理器。完成后在以下目录E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\Release 生成tensorflow.dll, tensorflow.lib，表示build完成；
1. 执行Release\tf_tutorials_example_trainer.exe 检验是否build成功；
2. 执行Release\tf_label_image_example.exe进一步检验是否build成功，这一步检验需要model，dic文件，可在此处下载到，放置在E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\tensorflow\examples\label_image\data 即可。

2. New VS C++ project

用VS2015新建空的visual C++ 项目（选择为Release，x64），右键项目修改properties：

修改C/C++下的Additional Include Directories，添加:

E:\tf-r1.10\tensorflow
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\eigen_archive
E:\tf-r1.10\tensorflow\third_party\eigen3
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\protobuf\src\protobuf\src
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\nsync\public
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\zlib_archive
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\gif_archive\giflib-5.1.4
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\png_archive
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\jpeg_archive
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\lmdb
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\gemmlowp\src\gemmlowp
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\jsoncpp\src\jsoncpp
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\farmhash_archive
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\farmhash_archive\util
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\highwayhash
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\cub\src\cub
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\re2\install\include
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\sqlite
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\grpc\src\grpc\include
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\snappy\src\snappy

修改C/C++的Preprocesser Definitions,添加：
```
COMPILER_MSVC
NOMINMAX
TENSORFLOW_EXPORTS
```

修改Linker的Input下的Additional Dependencies，添加Addtional Denpendencies:

E:\tf-r1.10\tensorflow
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\eigen_archive
E:\tf-r1.10\tensorflow\third_party\eigen3
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\protobuf\src\protobuf\src
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\nsync\public
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\zlib_archive
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\gif_archive\giflib-5.1.4
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\png_archive
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\jpeg_archive
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\lmdb
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\gemmlowp\src\gemmlowp
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\jsoncpp\src\jsoncpp
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\farmhash_archive
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\farmhash_archive\util
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\highwayhash
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\cub\src\cub
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\re2\install\include
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\external\sqlite
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\grpc\src\grpc\include
E:\tf-r1.10\tensorflow\tensorflow\contrib\cmake\build\snappy\src\snappy

在main.cpp下测试官方的测试代码：

#include "tensorflow/cc/client/client_session.h"
#include "tensorflow/cc/ops/standard_ops.h"
#include "tensorflow/core/framework/tensor.h"
   
int main() {
    using namespace tensorflow;
    using namespace tensorflow::ops;
       
    Scope root = Scope::NewRootScope();
    // Matrix A = [3 2; -1 0]
    auto A = Const(root, { { 3.f, 2.f },{ -1.f, 0.f } });
    // Vector b = [3 5]
    auto b = Const(root, { { 3.f, 5.f } });
    // v = Ab^T
    auto v = MatMul(root.WithOpName("v"), A, b, MatMul::TransposeB(true));
    std::vector<Tensor> outputs;
    ClientSession session(root);
    // Run and fetch v
    TF_CHECK_OK(session.Run({ v }, &outputs));
    // Expect outputs[0] == [19; -3]
    LOG(INFO) << outputs[0].matrix<float>();
    return 0;
}

编译项目，然后在C:\Users\yanfu\Documents\Visual Studio 2015\Projects\tensorflow-native-test\x64\Release 下用终端直接执行tensorflow-native-test.exe：

2018-09-02 08:03:49.370879: I E:\tf-r1.10\tensorflow\tensorflow\core\platform\cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX AVX2
2018-09-02 08:03:49.382443: I main.cpp:20] 19
-3

表示C++项目已经可以正确使用编译出来的TF库了。

Ps.Python生产model文件，在C++项目使用

最开始的目的是想用 python去生产TF的pb格式（在这个需求场景下推荐这个格式）的model文件，然后在C++的项目中使用，但是后来项目不需要TF的model了，所以就放弃了。python导出pb格式的文件的方法为：

import tensorflow as tf
import os
from tensorflow.python.framework import graph_util

with tf.Session(graph=tf.Graph()) as sess:
    x = tf.placeholder(tf.int32, name='x')
    y = tf.constant([666], name='y')
    w = tf.Variable(0, name='w')
    _op = tf.multiply(x, w)
    op = tf.add(_op, y, name='op_to_store')

    sess.run(tf.global_variables_initializer())

    # convert_variables_to_constants 需要指定output_node_names，list()，可以多个
    constant_graph = graph_util.convert_variables_to_constants(
        sess, sess.graph_def, ['op_to_store'])

    # 测试 OP
    feed_dict = {x: 10}
    print(sess.run(op, feed_dict))

    # 写入序列化的 PB 文件
    with tf.gfile.FastGFile('./model.pb', mode='wb') as f:
        f.write(constant_graph.SerializeToString())

Young Anything I do that may help others, I'll post it here.

Build TF on win

准备工作

1. build tf

2. New VS C++ project

Ps.Python生产model文件，在C++项目使用

Related Posts

EVA3.0 27 Oct 2019

EVA3.0 21 Jul 2019

Visotor Pattern 04 Mar 2019