Building pytorch/vision on Windows

vision
1

Hello,
I’m currently trying to build pytorch/visions on windows. I’ve given libtorch to cmake with the parameter
-DCMAKE_PREFIX_PATH="D:/libs/1.10/libtorch_cu102_release".
Then using this command cmake --build . --target install --config release
I’m having the following warning and error.

D:/libs/vision/torchvision/csrc/ops/cuda/deform_conv2d_kernel.cu(73): fatal error C1083: Cannot open include file: 'ATen/cuda/Atomic.cuh': No such file or directory [D:\libs\vision\build\torchvision.vcxproj]
CUDACOMPILE : nvcc warning : The -std=c++14 flag is not supported with the configured host compiler. Flag will be ignored. [D:\libs\vision\build\torchvision.vcxproj]

What am I missing ? Indeed I can’t find the file 'ATen/cuda/Atomic.cuh' in the libtorch archive that I’ve downloaded.

My generator is “Visual Studio 2019”
I’m using libtorch-win-shared-with-deps-1.10.1%2Bcu102.zip

Thanks for your help

2

Did you ever figure this out? I found the Atomic.cuh file in the ATen source in the pytorch github repository but when I put the file in place it just uncovered lots more errors when compiling vision.