Windows build failed

[dc986]

• OS: Microsoft Windows 7 Professional
• CMake version: version 3.17.1
• Python version: 3.7 (64-bit runtime)
• Is CUDA available: N/A
• [pip3] numpy==1.18.5
• [pip3] torch==1.7.1+cpu
• [pip3] torchaudio==0.7.2
• [pip3] torchvision==0.8.2+cpu

I am trying to build torch vision 0.8.2 from source using CMake but build fails.

I used the command:
cmake -DCMAKE_PREFIX_PATH="E:/Dev/ext/libs/libtorch-win-shared-with-deps-debug-1.7.1+cpu/libtorch; -DCMAKE_INSTALL_PREFIX="E:/Dev/ext/libs/pytorch_vision/install" -DCMAKE_BUILD_TYPE=Release --DWITH_CUDA=OFF ...

but then running
cmake --build .
the output is:

1>deform_conv2d.cpp
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\deform_conv2d.cpp(86,24): error C2065: 'torchvision': undeclared identifier
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\deform_conv2d.cpp(86,37): error C2065: 'm': undeclared identifier
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\deform_conv2d.cpp(86,38): error C4430: missing type specifier - int assumed. Note: C++ does not support default-int
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\deform_conv2d.cpp(86,40): error C2448: 'vision::ops::TORCH_LIBRARY_FRAGMENT': function-style initializer appears to be a function definition
1>nms.cpp
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\nms.cpp(18,24): error C2065: 'torchvision': undeclared identifier
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\nms.cpp(18,37): error C2065: 'm': undeclared identifier
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\nms.cpp(18,38): error C4430: missing type specifier - int assumed. Note: C++ does not support default-int
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\nms.cpp(18,40): error C2448: 'vision::ops::TORCH_LIBRARY_FRAGMENT': function-style initializer appears to be a function definition
1>ps_roi_align.cpp
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\ps_roi_align.cpp(56,24): error C2065: 'torchvision': undeclared identifier
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\ps_roi_align.cpp(56,37): error C2065: 'm': undeclared identifier
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\ps_roi_align.cpp(56,38): error C4430: missing type specifier - int assumed. Note: C++ does not support default-int
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\ps_roi_align.cpp(56,40): error C2448: 'vision::ops::TORCH_LIBRARY_FRAGMENT': function-style initializer appears to be a function definition
1>ps_roi_pool.cpp
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\ps_roi_pool.cpp(52,24): error C2065: 'torchvision': undeclared identifier
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\ps_roi_pool.cpp(52,37): error C2065: 'm': undeclared identifier
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\ps_roi_pool.cpp(52,38): error C4430: missing type specifier - int assumed. Note: C++ does not support default-int
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\ps_roi_pool.cpp(52,40): error C2448: 'vision::ops::TORCH_LIBRARY_FRAGMENT': function-style initializer appears to be a function definition
1>roi_align.cpp
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\roi_align.cpp(66,24): error C2065: 'torchvision': undeclared identifier
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\roi_align.cpp(66,37): error C2065: 'm': undeclared identifier
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\roi_align.cpp(66,38): error C4430: missing type specifier - int assumed. Note: C++ does not support default-int
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\roi_align.cpp(66,40): error C2448: 'vision::ops::TORCH_LIBRARY_FRAGMENT': function-style initializer appears to be a function definition
1>roi_pool.cpp
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\roi_pool.cpp(51,24): error C2065: 'torchvision': undeclared identifier
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\roi_pool.cpp(51,37): error C2065: 'm': undeclared identifier
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\roi_pool.cpp(51,38): error C4430: missing type specifier - int assumed. Note: C++ does not support default-int
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\ops\roi_pool.cpp(51,40): error C2448: 'vision::ops::TORCH_LIBRARY_FRAGMENT': function-style initializer appears to be a function definition
1>vision.cpp
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\vision.cpp(31,24): error C2065: 'torchvision': undeclared identifier
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\vision.cpp(31,37): error C2065: 'm': undeclared identifier
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\vision.cpp(31,38): error C4430: missing type specifier - int assumed. Note: C++ does not support default-int
1>E:\Dev\ext\libs\pytorch_vision\torchvision\csrc\vision.cpp(31,40): error C2448: 'vision::TORCH_LIBRARY_FRAGMENT': function-style initializer appears to be a function definition
1>Generating Code...
1>Done building project "torchvision.vcxproj" -- FAILED.
2>------ Build started: Project: ALL_BUILD, Configuration: Debug x64 ------
2>Building Custom Rule E:/Dev/ext/libs/pytorch_vision/CMakeLists.txt

[zsef123]

Same Issue on

• Win 10 / torch 1.7.1 / cuda 11
• WITH_CUDA=on
• MSVC 2019 16.8.4

---

• TORCH_LIBRARY_FRAGMENT build error #3210

TORCH_LIBRARY_FRAGMENT is not includes on 1.7.1

[datumbox]

@dc986 Building TorchVision from source should be done by running:
python setup.py install

For details on how to compile from source check our contributing page. Just replace nightly with the latest stable version of PyTorch since you want to build v0.8.2.

Using cmake is something we plan to support on the future. See #2349

@zsef123 You should do the same as described above. The only difference is that you need to install the nightly PyTorch lib because the current master of TorchVision depends on it.

[dc986]

Thank you @datumbox.

I was building the C++ API using Cmake and following instruction in https://github.com/pytorch/vision.

I finally solved it with cmake using the workaround in #3210

[datumbox]

@dc986 Good to hear! Could you clarify which workaround? Installing the PyTorch nightly?

Could you link to the exact comment so that others can benefit from it on the future? Thanks!

[dc986]

@datumbox
No, I am using
libtorch1.7.1+cpu
torchvision0.8.2

Installing libtorch nightly did not solve the problem either.

I had to modify the macro
TORCH_LIBRARY_FRAGMENT
with
TORCH_LIBRARY_FRAGMENT_THIS_API_IS_FOR_PER_OP_REGISTRATION_ONLY
in torchvision src files.

Not sure this is a good solution but the build finishes with no errors.