AI环境搭建系列2 ARM开发板编译安装caffe方法(ARM NN)

我会陆续将最近学习过程中，配置各平台caffe等AI框架开发环境的经验写下来，一来为了巩固所学，二来让后来者少走弯路，今天更新第一期，关于如何在ARM开发板上编译caffe。

系统环境：
    Debian10
    MSM8996(QCOM ARM处理器)

1.更新源

sudo vim /etc/apt/source.list

修改source.list文件为如下内容:

deb http://mirrors.163.com/debian/ buster main non-free contrib
deb http://mirrors.163.com/debian/ buster-updates main non-free contrib
deb http://mirrors.163.com/debian/ buster-backports main non-free contrib
deb-src http://mirrors.163.com/debian/ buster main non-free contrib
deb-src http://mirrors.163.com/debian/ buster-updates main non-free contrib
deb-src http://mirrors.163.com/debian/ buster-backports main non-free contrib
deb http://mirrors.163.com/debian-security/ buster/updates main non-free contrib
deb-src http://mirrors.163.com/debian-security/ buster/updates main non-free contrib

2.安装boost

Download boost source code:

tar --bzip2 -xf boost_1_64_0.tar.bz2
cd boost_1_64_0

设置boost安装目录

./bootstrap.sh --prefix=/usr/local/boost_1_64_0 --with-libraries= system,filesystem,thread,python

本文档是我在学习如何配置ARM NN开发环境时所写，所以这里会针对ARM NN对boost的编译选项的要求，添加一些编译选项：

sudo ./b2 link=static cxxflags=-fPIC --with-filesystem --with-test --with-log --with-program_options --with-python
linaro@linaro-alip:~/workspace/ARM_NN/boost_1_64_0$ sudo ./b2 link=static cxxflags=-fPIC --with-filesystem --with-test --with-log --with-program_options --with-python

然后会打印根据我们上面命令出现的配置信息

Performing configuration checks

    - 32-bit                   : no  (cached)
    - 64-bit                   : yes (cached)
    - arm                      : yes (cached)

Building the Boost C++ Libraries.


    - symlinks supported       : yes (cached)
    - native-atomic-int32-supported : yes (cached)
    - native-syslog-supported  : yes (cached)
    - pthread-supports-robust-mutexes : yes (cached)
    - compiler-supports-visibility : yes (cached)
    - has_icu builds           : no  (cached)
    - lockfree boost::atomic_flag : yes (cached)

Component configuration:

    - atomic                   : not building
    - chrono                   : not building
    - container                : not building
    - context                  : not building
    - coroutine                : not building
    - coroutine2               : not building
    - date_time                : not building
    - exception                : not building
    - fiber                    : not building
    - filesystem               : building
    - graph                    : not building
    - graph_parallel           : not building
    - iostreams                : not building
    - locale                   : not building
    - log                      : building
    - math                     : not building
    - metaparse                : not building
    - mpi                      : not building
    - program_options          : building
    - python                   : building
    - random                   : not building
    - regex                    : not building
    - serialization            : not building
    - signals                  : not building
    - system                   : not building
    - test                     : building
    - thread                   : not building
    - timer                    : not building
    - type_erasure             : not building
    - wave                     : not building

我们主要关注上面的Component configuration
如:

    - python                   : building

编译完成后会提示:

The Boost C++ Libraries were successfully built!

The following directory should be added to compiler include paths:

    /home/linaro/workspace/ARM_NN/boost_1_64_0

The following directory should be added to linker library paths:

    /home/linaro/workspace/ARM_NN/boost_1_64_0/stage/lib


sudo ./b2 install

如果提示:

    test_boost.cpp:1:10: fatal error: boost/regex.hpp: No such file or directory
    #include 

很多网站解释说，出现上面的问题，可能是编译boost后并没有安装,也就是忘记加install选项了。
即使你再次执行sudo ./b2 install也白扯。

然并卵…

事实证明,这样完全解决不了问题!我们使用下述方法来解决该问题,一句话GCC没有找到库和文件

export C_INCLUDE_PATH=/usr/local/boost_1_64_0/include:$C_INCLUDE_PATH
export CPLUS_INCLUDE_PATH=/usr/local/boost_1_64_0/include:$CPLUS_INCLUDE_PATH
export LD_LIBRARY_PATH=/usr/local/boost_1_64_0/lib:$LD_LIBRARY_PATH
export LIBRARY_PATH=/usr/local/boost_1_64_0/lib:$LIBRARY_PATH

其实，上面安装完成后，打印的log也提示了大家一定要将编译的链接路径加进去:

The following directory should be added to compiler include paths:

    /home/linaro/workspace/ARM_NN/boost_1_64_0

The following directory should be added to linker library paths:

    /home/linaro/workspace/ARM_NN/boost_1_64_0/stage/lib

3.安装Caffe

这部分可以参考:

http://caffe.berkeleyvision.org/installation.html

简要步骤如下:

1st step

git clone https://github.com/BVLC/caffe.git
$ cp Makefile.config.example Makefile.config

中修改如下内容：

1. 如果你的电脑没有NVIDIA的GPU则需要将这个注释放开:CPU_ONLY := 1
2. 修改hdf5库的路径
    INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include /usr/include/hdf5/serial
    LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /usr/lib/aarch64-linux-gnu/hdf5/serial
3. PYTHON_INCLUDE
这个选项要根据你的python安装路径决定，如果选用anaconda的话，则需要设置ANACONDA_HOME以及对应的PYTHON_INCLUDE
由于ARM平台上暂时没有用到anaconda，所以这里没有对anaconda的路径进行重新配置。

2nd step

make all -j4   编译全部模块
make test -j4  编译测试模块
make runtest -j4  运行测试
make pycaffe -j4  编译python库

4. 遇到的一些问题

4.1 make runtest报错

unknown file: Failure
C++ exception with description "locale::facet::_S_create_c_locale name not valid" thrown in the test fixture's constructor.
*** Aborted at 1522312851 (unix time) try "date -d @1522312851" if you are using GNU date ***
PC: @                0x0 (unknown)
*** SIGSEGV (@0x0) received by PID 7394 (TID 0xffff7b4d5440) from PID 0; stack trace: ***
    @     0xffff7c8076c0 ([vdso]+0x6bf)
    @        0x11bc49c20 (unknown)
    @        0x11bc4f484 (unknown)
    @        0x11bc492d8 (unknown)
    @        0x11bc493d0 (unknown)
    @        0x11bc498a0 (unknown)
    @        0x11bc499dc (unknown)
    @        0x11b9ec614 (unknown)
    @     0xffff7b66d624 __libc_start_main
Segmentation fault
Makefile:532: recipe for target 'runtest' failed
make: *** [runtest] Error 139

解决方案：
~/.bashrc末尾添加新的环境变量:

export LC_ALL="C"

再次执行make runtest就没问题了

4.2 ImportError: No module named skimage.io

sudo apt-get install python-skimage

4.3 AttributeError: ‘module’ object has no attribute ‘bool_’

则说明你的PYTHONPATH环境变量配置不对，本质原因可能是没有编译pycaffe， 因此执行make pycaffe进行编译
stackflow上有云：

a)You need to have _caffe.so in caffe/python/caffe. If not, run 'make pycaffe' in caffe source folder.

b)You need to add only caffe/python to$PYTHONPATH but not caffe/python/caffe as mentioned in documentation. I don'tknow why adding caffe/python/caffe causes the error, though.

第一次在编译boost的时候没有加–with-python选项,所以没有生成python库:libboost_python.ap

4.4 ImportError: No module named google.protobuf.internal

sudo apt-get install python-protobuf

References:

1. https://developer.arm.com/technologies/machine-learning-on-arm/developer-material/how-to-guides/configuring-the-arm-nn-sdk-build-environment-for-caffe/setting-up-the-build-dependencies
2. http://caffe.berkeleyvision.org/installation.html
3. https://blog.csdn.net/xushungou/article/details/52853110
4. http://valleylord.github.io/post/201601-boost-install/