Nvidia Jetson Xavier配置tensorflow1.13.1 C++ API
Nvidia Jetson Xavier配置tensorflow1.13.1 C++ API
温馨提示：源码编译tensorflow所占内存比较大，最好在一个全新的AGX上编译或者加入sd卡或者硬盘。本人是在一个全新的AGX上进行的编译，并卸载了一些无用软件，已成功。
配置清华源
首先备份/etc/apt/sources.list
cp /etc/apt/sources.list /etc/apt/sources.list.bak
将sources.list修改为：
deb http://mirrors.tuna.tsinghua.edu.cn/ubuntu-ports/ xenial-updates main restricted universe multiverse
deb-src http://mirrors.tuna.tsinghua.edu.cn/ubuntu-ports/ xenial-updates main restricted universe multiverse
deb http://mirrors.tuna.tsinghua.edu.cn/ubuntu-ports/ xenial-security main restricted universe multiverse
deb-src http://mirrors.tuna.tsinghua.edu.cn/ubuntu-ports/ xenial-security main restricted universe multiverse
deb http://mirrors.tuna.tsinghua.edu.cn/ubuntu-ports/ xenial-backports main restricted universe multiverse
deb-src http://mirrors.tuna.tsinghua.edu.cn/ubuntu-ports/ xenial-backports main restricted universe multiverse
deb http://mirrors.tuna.tsinghua.edu.cn/ubuntu-ports/ xenial main universe restricted
deb-src http://mirrors.tuna.tsinghua.edu.cn/ubuntu-ports/ xenial main universe restricted
安装gcc-5
AGX自带的gcc版本过高，需要降低到5，我第一次编译的时候未降级就失败了，可以通过建立软连接的方式让两个版本共存。
sudo apt-get install gcc-5
sudo apt-get install g++-5
此时终端输入gcc -v依然是AGX自带的版本，此时删除旧连接，建立新的的软连接。
sudo rm -rf /usr/bin/gcc /usr/bin/g++
sudo ln -s /usr/bin/gcc-5 /usr/bin/gcc
sudo ln -s /usr/bin/g++-5 /usr/bin/g++
此时终端输入gcc -v即是gcc version 5.4.0 20160609 (Ubuntu/Linaro 5.4.0-6ubuntu1~16.04.12)
安装jdk1.8
sudo apt-get install openjdk-8-jdk
安装bazel
下载bazel0.19.2
AGX上需要下载源码bazel-0.19.2-dist.zip进行编译安装，bazel的版本要与tensorflow的版本相对应，否则会编译失败。
编译的时候需要注意，先新建一个bazel的文件夹，然后将压缩包移动进去，在文件夹内解压。
unzip bazel-0.19.2-dist.zip
bash ./compile.sh
编译会持续较长时间，编译产生的文件bazel放在./output里面，将其复制到/usr/local/bin
sudo cp ./output/bazel /usr/local/bin
终端查看版本信息
bazel version
安装protocbuf3.6.1
protocbuf的版本要与tensorflow版本相对应，tensorflow源码./tensorflow-1.13.1/tensorflow/workspace.bzl会有对应需要的protocbuf的版本，tensorflow1.13.1需要protocbuf3.6.1，并且在编译时会自行下载依赖文件，我自己下载源码后进行的编译安装。
先贴出protobuf官方下载的连接，点击最下面的选项Source Code（tar.gz）进行下载，下载后解压到个人文件夹。
编译安装之前需要先安装依赖，否则会出错。
sudo apt-get install autoconf automake libtool curl make
然后进入到解压后的文件夹下打开终端，输入以下命令进行安装。
./autogen.sh
./configure
make
sudo make install
sudo ldconfig
# sudo make uninstall 安装错版本后卸载指令
protoc --version # 查看protobuf版本
安装Eigen库
编译tensorflow时会自行下载对应的Eigen源码，我是自己下载后进行编译安装。
使用如下命令安装，我们这里使用的版本是 Eigen 3.3.4：
wget -t 0 -c https://github.com/eigenteam/eigen-git-mirror/archive/3.3.4.zip
unzip 3.3.4.zip
cd eigen-git-mirror-3.3.4/
mkdir build
cd build
cmake ..
make
sudo make install
编译tensorfow c++ API
下载源码
git clone https://github.com/tensorflow/tensorflow.git
cd tensorflow
git checkout r1.13.1
然后需要修改部分代码，这部分中-代表删除，+代表添加
首先是./tensorflow/lite/kernels/internal/BUILD
diff --git a/tensorflow/lite/kernels/internal/BUILD b/tensorflow/lite/kernels/internal/BUILD
index 4be3226938..7226f96fdf 100644
--- a/tensorflow/lite/kernels/internal/BUILD
+++ b/tensorflow/lite/kernels/internal/BUILD
@@ -22,15 +22,12 @@ HARD_FP_FLAGS_IF_APPLICABLE = select({
NEON_FLAGS_IF_APPLICABLE = select({
":arm": [
"-O3",
- "-mfpu=neon",
],
":armeabi-v7a": [
"-O3",
- "-mfpu=neon",
],
":armv7a": [
"-O3",
- "-mfpu=neon",
],
"//conditions:default": [
"-O3",
然后是./third_party/aws/BUILD.bazel b/third_party/aws/BUILD.bazel
diff --git a/third_party/aws/BUILD.bazel b/third_party/aws/BUILD.bazel
index 5426f79e46..e08f8fc108 100644
--- a/third_party/aws/BUILD.bazel
+++ b/third_party/aws/BUILD.bazel
@@ -24,7 +24,7 @@ cc_library(
"@org_tensorflow//tensorflow:raspberry_pi_armeabi": glob([
"aws-cpp-sdk-core/source/platform/linux-shared/*.cpp",
]),
- "//conditions:default": [],
+ "//conditions:default": glob(["aws-cpp-sdk-core/source/platform/linux-shared/*.cpp",]),
}) + glob([
"aws-cpp-sdk-core/include/**/*.h",
"aws-cpp-sdk-core/source/*.cpp",
最后是./third_party/gpus/crosstool/BUILD.tpl b/third_party/gpus/crosstool/BUILD.tpl
diff --git a/third_party/gpus/crosstool/BUILD.tpl b/third_party/gpus/crosstool/BUILD.tpl
index db76306ffb..184cd35b87 100644
--- a/third_party/gpus/crosstool/BUILD.tpl
+++ b/third_party/gpus/crosstool/BUILD.tpl
@@ -24,6 +24,7 @@ cc_toolchain_suite(
"x64_windows|msvc-cl": ":cc-compiler-windows",
"x64_windows": ":cc-compiler-windows",
"arm": ":cc-compiler-local",
+ "aarch64": ":cc-compiler-local",
"k8": ":cc-compiler-local",
"piii": ":cc-compiler-local",
"ppc": ":cc-compiler-local",
修改 tensorflow 里面的 protobuf 和 eigen 版本配置（当然如果你安装的 protobuf 就是 3.6.1 就不需要这些步骤了）：
在刚才的步骤中，我们已经安装了自己的一个 protobuf 版本，然而比较坑的是 tensorflow 自己也会下载一个版本，同时内部也有自己定义的版本号，这些版本很可能都不是统一的。最后会造成版本冲突。因此我们这里建议预先修改 tensorflow 源码将内部的版本统一成我们系统的版本。
打开 tensorflow/workspace.bzl，将有关protocbuf版本信息修改为自己安装的版本号，如v3.6.1.tar.gz。
tf_http_archive(
name = "protobuf_archive",
urls = [
"https://mirror.bazel.build/github.com/google/protobuf/archive/v3.6.1.tar.gz",
"https://github.com/google/protobuf/archive/v3.6.1.tar.gz",
],
sha256 = "3d4e589d81b2006ca603c1ab712c9715a76227293032d05b26fca603f90b3f5b",
strip_prefix = "protobuf-3.6.1",
)
# We need to import the protobuf library under the names com_google_protobuf
# and com_google_protobuf_cc to enable proto_library support in bazel.
# Unfortunately there is no way to alias http_archives at the moment.
tf_http_archive(
name = "com_google_protobuf",
urls = [
"https://mirror.bazel.build/github.com/google/protobuf/archive/v3.6.1.tar.gz",
"https://github.com/google/protobuf/archive/v3.6.1.tar.gz",
],
sha256 = "3d4e589d81b2006ca603c1ab712c9715a76227293032d05b26fca603f90b3f5b",
strip_prefix = "protobuf-3.6.1",
)
tf_http_archive(
name = "com_google_protobuf_cc",
urls = [
"https://mirror.bazel.build/github.com/google/protobuf/archive/v3.6.1.tar.gz",
"https://github.com/google/protobuf/archive/v3.6.1.tar.gz",
],
sha256 = "3d4e589d81b2006ca603c1ab712c9715a76227293032d05b26fca603f90b3f5b",
strip_prefix = "protobuf-3.6.1",
)
打开 tensorflow/contrib/cmake/external/protobuf.cmake
修改：set(PROTOBUF_TAG v3.6.1)
打开 tensorflow/tools/ci_build/install/install_proto3.sh
修改：PROTOBUF_VERSION="3.6.1"
打开 tensorflow/tools/ci_build/protobuf/protobuf_optimized_pip.sh
修改：PROTOBUF_VERSION="3.6.1"
接下来是创建交换分区，这个是因为Xavier自带的内存不够用，必须使用交换分区才行，我当初没有创建，编译了几个小时之后报错。。。
首先用命令free查看系统内 Swap 分区大小
free -m
total used free shared buff/cache available
Mem: 15692 5695 7437 787 2558 10505
Swap: 0 0 0
创建8G的swap空间
fallocate -l 8G swapfile
ls -lh swapfile
sudo chmod 600 swapfile
ls -lh swapfile
sudo mkswap swapfile
sudo swapon swapfile
swapon -s
接下来开始编译我们的tensorflow c++ API
首先配置configure文件
1. $./configure
2. Please specify the location of python. [Default is /usr/bin/python]:/usr/bin/python3
5. Found possible Python library paths:
6. /usr/local/lib/python3.6/dist-packages
7. /usr/lib/python3.6/dist-packages
8. Please input the desired Python library path to use. Default is [/usr/local/lib/python3.6/dist-packages]
10. Do you wish to build TensorFlow with XLA JIT support? [Y/n]: n
11. No XLA JIT support will be enabled for TensorFlow.
13. Do you wish to build TensorFlow with OpenCL SYCL support? [y/N]: n
14. No OpenCL SYCL support will be enabled for TensorFlow.
16. Do you wish to build TensorFlow with ROCm support? [y/N]: n
17. No ROCm support will be enabled for TensorFlow.
19. Do you wish to build TensorFlow with CUDA support? [y/N]: y
20. CUDA support will be enabled for TensorFlow.
22. Please specify the CUDA SDK version you want to use. [Leave empty to default to CUDA 10.0]:
25. Please specify the location where CUDA 10.0 toolkit is installed. Refer to README.md for more details. [Default is /usr/local/cuda]:/usr/local/cuda-10.0
28. Please specify the cuDNN version you want to use. [Leave empty to default to cuDNN 7]:7.3.1
31. Please specify the location where cuDNN 7 library is installed. Refer to README.md for more details. [Default is /usr/local/cuda]:/usr/lib/aarch64-linux-gnu
34. Do you wish to build TensorFlow with TensorRT support? [y/N]: n
35. TensorRT support will be enabled for TensorFlow.
37. Please specify the location where TensorRT is installed. [Default is /usr/lib/aarch64-linux-gnu]:
40. Please specify the locally installed NCCL version you want to use. [Default is to use [https://github.com/nvidia/nccl]:](https://github.com/nvidia/nccl]:)
43. Please specify a list of comma-separated Cuda compute capabilities you want to build with.
44. You can find the compute capability of your device at: [https://developer.nvidia.com/cuda-gpus.](https://developer.nvidia.com/cuda-gpus.)
45. Please note that each additional compute capability significantly increases your build time and binary size. [Default is: 3.5,7.0]: 7.2
48. Do you want to use clang as CUDA compiler? [y/N]: n
49. nvcc will be used as CUDA compiler.
51. Please specify which gcc should be used by nvcc as the host compiler. [Default is /usr/bin/gcc]:/usr/bin/gcc
54. Do you wish to build TensorFlow with MPI support? [y/N]: n
55. No MPI support will be enabled for TensorFlow.
57. Please specify optimization flags to use during compilation when bazel option "--config=opt" is specified [Default is -march=native -Wno-sign-compare]:
60. Would you like to interactively configure ./WORKSPACE for Android builds? [y/N]: n
61. Not configuring the WORKSPACE for Android builds.
然后编译tensorflow
bazel build --config=opt --config=cuda --config=nonccl //tensorflow:libtensorflow_cc.so --incompatible_remove_native_http_archive=false --verbose_failures --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=0"
编译时所占内存比较大，电脑就不要干其他事情，静静等待即可。编译成功后会在./bazel-bin/tensorflow生成libtensorflow_cc.so，libtensorflow_framework.so两个文件。
测试tensorflow c++ API
首先把需要的文件挑选出来。
将"tensorflow/bazel-genfiles/tensorflow/"中的cc和core文件夹中的内容copy到"tensorflow/tensorflow/"中，然后选择合并覆盖。
复制tensorflow/tensorflow/文件到到home下（或者 /usr/local/include，都可以），并删除不需要的文件：
sudo mkdir -p /usr/local/include/tensorflow
sudo cp -r tensorflow /usr/local/include/tensorflow/
sudo find /usr/local/include/tensorflow/tensorflow -type f ! -name "*.h" -delete
复制 third_party 文件夹:
$ sudo cp -r third_party /usr/local/include/tensorflow/
$ sudo rm -r /usr/local/include/google/tensorflow/third_party/py
# Note: newer versions of TensorFlow do not have the following directory
$ sudo rm -r /usr/local/include/tensorflow/third_party/avro
复制.so文件：
sudo cp libtensorflow_cc.so /usr/local/lib/
sudo cp libtensorflow_framework.so /usr/local/lib/
首先编写CMakeLists.txt
cmake_minimum_required(VERSION 3.5)
project(test_tensorflow)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
add_definitions(-D_GLIBCXX_USE_CXX11_ABI=0)
link_directories(/usr/local/lib/)
include_directories(
/usr/local/include/tensorflow
/usr/local/include/eigen3
)
find_package(OpenCV REQUIRED)
add_executable(test_tensorflow main.cpp)
target_link_libraries(test_tensorflow tensorflow_cc tensorflow_framework)
target_link_libraries(test_tensorflow ${OpenCV_LIBS})
测试代码
#include 编译和调用时遇到的问题
可能会遇到下载不下来文件的情况
bazel build出现”no such package ‘@icu//’: java.io.IOException…"的错误，主要是icu模块的release-62-1.tar.gz包下载失败。
以为是网络不稳定，试了N次，几乎每次都是出现这个报错。查了一下才发现需要梯子，导致powershell编译时下载包失败！
解决方案：通过chrome浏览器下载release-62-1.tar.gz包（浏览器也需要梯子，不过浏览器下载容易的多），然后将tar.gz包上传到http文件服务器，通过http文件服务器加载tar.gz包。
如何在本地搭建一个简单的HTTP服务器？
sudo apt-get install httpd
输出如下信息
意思是选择一个安装，那么我选择apache2来安装
sudo apt-get install apache2
sudo apt-get install apache2-dev
查看是否安装成功
sudo systemctl status apache2
在浏览器输入localhost或者本机的ip
成功！并会生成一个/var/www/html文件，将下载好的tar.gz包复制到/var/www/html下。
修改tensorflow/third_party/icu/workspace.bzl文件，urls中添加：
"http://localhost/release-62-1.tar.gz"
有下载失败的文件都可以通过这种方法操作。
测试案例时遇到一些问题
(1)fatal error: absl/strings/string_view.h
解决方案，git clone https://github.com/abseil/abseil-cpp,然后把该库加到搜索目录里面
(2)tensorflow::status::tostringabi:cxx11 const
解决方案，编译选项设置-D_GLIBCXX_USE_CXX11_ABI=0，在CMakeLists.txt中加入add_definitions(-D_GLIBCXX_USE_CXX11_ABI=0)
如果使用的是QT，则在.pro文件中加入DEFINES += _GLIBCXX_USE_CXX11_ABI=0。
(3)对‘tensorflow::SessionOptions::SessionOptions()’未定义的引用
找不到正确的libtensorflow_cc.so，添加动态链接库路径
到此，我们的tensorflow c++ API编译成功了！！！
参考连接：
https://www.jianshu.com/p/746ec48edbdb
https://blog.csdn.net/yz2zcx/article/details/83153588
https://blog.csdn.net/yz2zcx/article/details/83153588
Tensorflow C++ 从训练到部署(1)：环境搭建
https://www.cnblogs.com/buyizhiyou/p/10405634.html