PyTorch源代码编译过程：
1、源代码编译环境：
操作系统： debian 9.12
交换空间： 1G
Python版本： 3.5
硬件：
CPU:rk3399(aarch64 ) ) )。
内存： 4G
2、下载依赖包：
下载pytorch及其依赖包时，缺省情况下从github下载。 如果网络较差且容易断开连接，请在gitee中找到相应的mldsmt链接，修改并下载相应的配置文件。 需要细致的耐心。
3、编译主要参数：
设定最大作业数：
export MAX_JOBS=3
无pytorch、cude、无MKL :
export no _ cuda=1export no _ distributed=1export no _ mkl dnn=1export no _ nn pack=1export no _ qnnpack=1
torchvision源代码编译过程：
1、取决于安装环境：
sdoapt-getinstallibjpeg-dev zlib 1g-dev
2、下载源代码：
git clone-bv 0.5.0https://github.com/py torch/vision.gittorchvision
或git clone-bv 0.5.0https://gitee.com/mirrors/vision.gittorchvision
3、源代码编译：
cd torchvision
python3 setup.py build
编译过程中应该出现了错误。 据说是python3.5的错误。 参考链接： https://blog.csdn.net/爱笑战斗机_ show/article/details/108280251
错误详细信息：
file/usr/lib/python 3.5/weak ref.py，line 117，in remove
typeerror : nonetypeobjectisnotcallable
解决方案，修改weakref.py文件：
找到" defremove(wr，selfref=ref ) self "行
修改为“defremove(wr，selfref=ref ) self，_ atomic _ removal=_ remove _ dead _ weak ref ) :”
找到_remove_dead_weakref(d，wr.key )行
修改为“_atomic_removal(d，wr.key )”
4、生成wheel :
python3 setup.py bdist_wheel
此时，在dist/目录中获得了whl数据包： torch vision-0.5.0 a 085 b8f BF-cp35-cp35m-Linux _ a arch 64.whl
其他：
1、py torch.whlforgenericaarch 64 platforms
justdownloadthefileorclonetherepository，and install the wheel using pip。
pip3install torch-1.1.0 a 0472 be 69-cp36-cp36m-Linux _ a arch 64.whl
2、Rationale
iwantedtorunsomepytorchmodelsonmyultra 96 boardforeventualuseinxilinx/deep hi ' sdnndkplatofrm，whichallowsyoutotkaeatratrainedmorinedmorm
3、构建安装
incaseyouyouhaveadifferntversionofpython，oranyotherreasonyoumighthavetobuildityourself，ill describe the steps below。
4、内存一致性s
thebiggestchallengebuildingalargelibrarylikethisonanembeeddeviceisthelimitaionofram (ultra 96 has 2gb ).To combat this、 you could make some extra swap space， oryoucouldlimitjobsthenumberofjobsusedwhenbuilding.iwouldrecommnedmakingsureyouhaveatleast1gbswapandlimitingthejobsto 1，asjjjobsto 1
ihadaspare 2gb USB2flashdrivewhichwouldworkperfectly，soiusedthatasmyextraswap.makesuretochangethe/dev/sda1 to your device
First，maketheswapspacesudomkswap/dev/sda 1
edityourfstabtoincludethenewdrive.nano/etc/fstab，commentoutanylinesyousee (replacethemonceyouaredone ) andadd/dev/dev
turnonthenewswapspacesudoswapon-p 32767/dev/sda 1
checkitworkedusingcat/proc/swapsandseeingifyourdriveislisted。
Finally，setthemaxnumberofjobs.idi D1 tobe safe，butifyouwantittogofasteryoucouldtry 2。
export MAX_JOBS=1
5、downloaddependenciesandclonepytorch
Dependencies:
sudoaptinstallibopenblas-devlibblas-dev M4 cmakecythonpython3- dev python3- YAML python3- setup tools
Clone PyTorch:
git clone-- recursive https://github.com/py torch/pytorchandcdpytorch
setbuildoptionsforembeededmachine (no cude，no MKL，etc ) )。
export no _ cuda=1export no _ distributed=1export no _ mkl dnn=1export no _ nn pack=1export no _ qnnpack=1
and finally，run the build。
python3 setup.py build
Once it completes with no errors，youcanrunpython3setup.pyinstalltoinstallthefiles，andifyouarereadingthisinthefuturewithanupdath youcanalsorunpython3setup.pybdist _ wheeltocreateawhlfilefordistribution.thatcommandwillputhegenerated.whlfileinthenewewed
Now you can run torch！
cd python3 import torch