Regarding to the problem related to ninja...

[frostinassiky]

Dear guys,

I also meet some issue about the ninja... Here is my understanding:

• This project uses JIT coding style, which requires ninja building system.

• Solution 1. To install set up the system, there are two ways.

  • apt-get install ninja-build. The cuda version in the system has to match the one used in conda env
  • conda install ninja or pip install ninja: does not work for me.

• Solution 2 that I am using. To avoid ninja, write in the "ahead of time" is one possible solution.

  • Create a new file setup.py under models/hrnet/sync_bn/inplace_abn
  • Install the inplace_abn module by python setup.py install
  • Modify models/hrnet/sync_bn/inplace_abn/functions.py, import the module as _backend

[sunke123]

@frostinassiky
Thanks for you help!

[zyxu1996]

@frostinassiky
Could you please give me a detailed description about the 2nd solution? I don't know what's that mean. Thank you

[frostinassiky]

@xu13521090631
This is my setup file.

from os import path

_src_path = path.join(path.dirname(path.abspath(__file__)), "src")

from setuptools import setup
from torch.utils.cpp_extension import BuildExtension, CUDAExtension

setup(
    name='inplace_abn_cpp_backend',
    ext_modules=[
        CUDAExtension(
            name='inplace_abn_cpp_backend',
            sources=[
              "src/inplace_abn.cpp",
              "src/inplace_abn_cpu.cpp",
              "src/inplace_abn_cuda.cu"
            ],
            extra_compile_args = {
                "cxx":["-O3"],
                'nvcc': ['--expt-extended-lambda']
            }
        )
    ],
    cmdclass={
        'build_ext': BuildExtension
    })

In the function.py file, I create a new _backend by
import inplace_abn_cpp_backend as _backend

As a third solution, if you are not using multiple GPUs, the PyTorch batch normalization layer works well. Ahha, there is a new brach pytorch-v1.1.

[zyxu1996]

@frostinassiky
Thank you for your reply. There are still some troubles, I have done all these steps, but still get some errors as follows.I guess it is the version of cuda, pytorch, ninja don't match, could you please tell me the version of these packages. Thank you!
`!! WARNING !!

!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
Your compiler (c++) may be ABI-incompatible with PyTorch!
Please use a compiler that is ABI-compatible with GCC 4.9 and above.
See https://gcc.gnu.org/onlinedocs/libstdc++/manual/abi.html.

See https://gist.github.com/goldsborough/d466f43e8ffc948ff92de7486c5216d6
for instructions on how to install GCC 4.9 or higher.
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! THCudaCheck FAIL file=/pytorch/aten/src/THC/THCGeneral.cpp line=663 error=11 : invalid argument
Traceback (most recent call last):
File "/data/xzy/new_hrnet/HRNet-Semantic-Segmentation/tools/train.py", line 248, in
main()
File "/data/xzy/new_hrnet/HRNet-Semantic-Segmentation/tools/train.py", line 83, in main
logger.info(get_model_summary(model.cuda(), dump_input.cuda()))
File "/data/xzy/new_hrnet/HRNet-Semantic-Segmentation/lib/utils/modelsummary.py", line 90, in get_model_summary
model(*input_tensors)
File "/home/omnisky/.local/lib/python3.5/site-packages/torch/nn/modules/module.py", line 477, in call
result = self.forward(*input, **kwargs)
File "/data/xzy/new_hrnet/HRNet-Semantic-Segmentation/lib/models/seg_hrnet.py", line 408, in forward
x = self.conv1(x)
File "/home/omnisky/.local/lib/python3.5/site-packages/torch/nn/modules/module.py", line 477, in call
result = self.forward(*input, **kwargs)
File "/home/omnisky/.local/lib/python3.5/site-packages/torch/nn/modules/conv.py", line 301, in forward
self.padding, self.dilation, self.groups)
RuntimeError: cuda runtime error (11) : invalid argument at /pytorch/aten/src/THC/THCGeneral.cpp:663`

[Fiordarancio]

@frostinassiky
Thank you for your helpful workaround: I encountered problems with ninja too and still I am quite stuck with them. I borrowed your code for setup.py but when I execute it I get the following error:

building 'inplace_abn_cpp_backend' extension
x86_64-linux-gnu-gcc -pthread -DNDEBUG -g -fwrapv -O2 -Wall -g -fstack-protector-strong -Wformat -Werror=format-security -Wdate-time -D_FORTIFY_SOURCE=2 -fPIC -I/home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include -I/home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/torch/csrc/api/include -I/home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/TH -I/home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/THC -I/usr/local/cuda/include -I/usr/include/python3.6m -I/home/ilaria/workspace/hrnenv/include/python3.6m -c src/inplace_abn.cpp -o build/temp.linux-x86_64-3.6/src/inplace_abn.o -O3 -DTORCH_API_INCLUDE_EXTENSION_H -DTORCH_EXTENSION_NAME=inplace_abn_cpp_backend -D_GLIBCXX_USE_CXX11_ABI=0 -std=c++11
src/inplace_abn.cpp: In function ‘void pybind11_init_inplace_abn_cpp_backend(pybind11::module&)’:
src/inplace_abn.cpp:70:69: error: no matching function for call to ‘pybind11::module::def(const char [9], <unresolved overloaded function type>, const char [36])’
   m.def("backward", &backward, "Second part of backward computation");
                                                                     ^
In file included from /home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/torch/csrc/utils/pybind.h:6:0,
                 from /home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/torch/csrc/api/include/torch/python.h:12,
                 from /home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/torch/extension.h:6,
                 from /home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/torch/csrc/api/include/torch/torch.h:6,
                 from src/inplace_abn.cpp:1:
/home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/pybind11/pybind11.h:810:13: note: candidate: template<class Func, class ... Extra> pybind11::module& pybind11::module::def(const char*, Func&&, const Extra& ...)
     module &def(const char *name_, Func &&f, const Extra& ... extra) {
             ^~~
/home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/pybind11/pybind11.h:810:13: note:   template argument deduction/substitution failed:
src/inplace_abn.cpp:70:69: note:   couldn't deduce template parameter ‘Func’
   m.def("backward", &backward, "Second part of backward computation");
                                                                     ^
error: command 'x86_64-linux-gnu-gcc' failed with exit status 1

Actually, this is the same complain that I get in the traceback when ninja is istalled (pretty the same of this previous issue). I sincerly cannot figure out how to solve the problem, since I haven't found decent solutions in related topics.

My environment:

• torch 1.4.0
• CUDA 10.1
• ninja 1.9.0
• gcc/g++ 7.4.0

I tried also to use downgraded versions of torch and CUDA, but it did not work. Any help would be appreciated!

[kaizen0890]

@frostinassiky
Thank you for your helpful workaround: I encountered problems with ninja too and still I am quite stuck with them. I borrowed your code for setup.py but when I execute it I get the following error:

building 'inplace_abn_cpp_backend' extension
x86_64-linux-gnu-gcc -pthread -DNDEBUG -g -fwrapv -O2 -Wall -g -fstack-protector-strong -Wformat -Werror=format-security -Wdate-time -D_FORTIFY_SOURCE=2 -fPIC -I/home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include -I/home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/torch/csrc/api/include -I/home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/TH -I/home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/THC -I/usr/local/cuda/include -I/usr/include/python3.6m -I/home/ilaria/workspace/hrnenv/include/python3.6m -c src/inplace_abn.cpp -o build/temp.linux-x86_64-3.6/src/inplace_abn.o -O3 -DTORCH_API_INCLUDE_EXTENSION_H -DTORCH_EXTENSION_NAME=inplace_abn_cpp_backend -D_GLIBCXX_USE_CXX11_ABI=0 -std=c++11
src/inplace_abn.cpp: In function ‘void pybind11_init_inplace_abn_cpp_backend(pybind11::module&)’:
src/inplace_abn.cpp:70:69: error: no matching function for call to ‘pybind11::module::def(const char [9], <unresolved overloaded function type>, const char [36])’
   m.def("backward", &backward, "Second part of backward computation");
                                                                     ^
In file included from /home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/torch/csrc/utils/pybind.h:6:0,
                 from /home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/torch/csrc/api/include/torch/python.h:12,
                 from /home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/torch/extension.h:6,
                 from /home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/torch/csrc/api/include/torch/torch.h:6,
                 from src/inplace_abn.cpp:1:
/home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/pybind11/pybind11.h:810:13: note: candidate: template<class Func, class ... Extra> pybind11::module& pybind11::module::def(const char*, Func&&, const Extra& ...)
     module &def(const char *name_, Func &&f, const Extra& ... extra) {
             ^~~
/home/ilaria/workspace/hrnenv/lib/python3.6/site-packages/torch/include/pybind11/pybind11.h:810:13: note:   template argument deduction/substitution failed:
src/inplace_abn.cpp:70:69: note:   couldn't deduce template parameter ‘Func’
   m.def("backward", &backward, "Second part of backward computation");
                                                                     ^
error: command 'x86_64-linux-gnu-gcc' failed with exit status 1

Actually, this is the same complain that I get in the traceback when ninja is istalled (pretty the same of this previous issue). I sincerly cannot figure out how to solve the problem, since I haven't found decent solutions in related topics.

My environment:

• torch 1.4.0
• CUDA 10.1
• ninja 1.9.0
• gcc/g++ 7.4.0

I tried also to use downgraded versions of torch and CUDA, but it did not work. Any help would be appreciated!

I found that this error relates to version matching problem, specifying between Torch and ninja version.
Therefore I solved this problem by finding ninja version which matching with Torch version.
My environment as show as below:
Ubuntu: 16.04
Gcc: 6.5.0
Python: 3.5.2
Torch: 0.4.1
ninja: 1.8.2
Hopefully that it can help you guys!