Source code compilation pytorch pit

nimble It's a right pytorch Neural network acceleration package for modification , This article records the configuration nimble Some problems encountered in the environment , For future reference .

git clone Pull problem

Domestic visit github Very slowly , And because of the need to recursively pull sub modules ：

git clone --recursive https://github.com/snuspl/nimble

often fatal Or pull to failure , At present, I think the reason is due to network problems , The solution is to keep executing the following commands in the project directory ：

git submodule update --init --recursive

Then, when executing the compiled script, go to the third-party database and pull the corresponding warehouse into the folder according to the error prompt

Compiler error ：…/caffe2/contrib/aten/aten_op.h:1:52: fatal error: caffe2/caffe2/contrib/aten/gen_aten_op.h: No such file or directory #include “caffe2/caffe2/contrib/aten/gen_aten_op.h”

stay pytorch Of github Of issue Page search to know , You need to add USE_NINJA=OFF Options , And if you've made it up before , It is best to build Directory delete recompile

Aborted (core dumped)

torchvision It is better to install the library from the source code , No mistake , At this point, I compiled and installed it from scratch

After the compilation is successful, an error is reported ：RuntimeError: module compiled against API version 0xe but this version of numpy is 0xd

Access terminal , see numpy Version number

pip show numpy

to update numpy

pip install -U numpy

solve