windows maskrcnn error: unresolved external symbol "class at::Tensor __cdecl nms_cpu(class at::Tensor const &,class at::Tensor const &,float)" (?nms_cpu@@YA?AVTensor@at@@AEBV12@0M@Z)

[qizhen816]

I built torchvision pulled from master using :

cmake -DCMAKE_PREFIX_PATH=D:/Garage/onnx-deploy/libtorchd  -DCMAKE_INSTALL_PREFIX=D:/vision/Debug -DCMAKE_BUILD_TYPE=Debug ..
cmake --build . --config Debug 
cmake --install .

Then torch.jit.script(model) and saved mask-rcnn model from torchvision.
Errors occured in my C++ project, firstly Unknown builtin op: torchvision::_new_empty_tensor_op, then I modified my code according to #1407, added lines like #include "torchvision/nms.h" and torch::RegisterOperators().op("torchvision::nms", &nms) , but it won't work.
The error became unresolved external symbol "class at::Tensor __cdecl nms_cpu(class at::Tensor const &,class at::Tensor const &,float)" (?nms_cpu@@YA?AVTensor@at@@AEBV12@0M@Z).
The libtorch version is 1.5.0 and torchvision is pulled from master yesterday.
Any clue how I can get mask rcnn running? Similar issues are many but little to help, I really get confused about how to modify CMakeList.txt of TorchVision.

cc @peterjc123 @nbcsm @guyang3532 @maxluk @gunandrose4u @smartcat2010 @mszhanyi

[peterjc123]

You need to mark those function as dllexport using this header.

[qizhen816]

@peterjc123 I use cmake and link torchvision by target_link_libraries(od PUBLIC TorchVision::TorchVision). _new_empty_tensor_op can be imported from dll but torchvision::nms can't.

[LI-Wen-Jiang]

@qizhen816 I have a same problem with you. Do you slove the problem?

[qizhen816]

@LI-Wen-Jiang Nope. Looked deep into the torchvision.lib, it seems operators are not contained in the dll. Trying to rebuild. 🔢

[peterjc123]

Just add the annotation __declspec(dllexport) on those functions.

[LI-Wen-Jiang]

@peterjc123 I don't quite understand what you mean. In where to add ?

[peterjc123]

For example, if you want to export nms_cpu in https://github.com/pytorch/vision/blob/master/torchvision/csrc/cpu/vision_cpu.h#L85, then change that line into

__declspec(dllexport) at::Tensor nms_cpu(
    const at::Tensor& dets,
    const at::Tensor& scores,
    const float iou_threshold);

[qizhen816]

@peterjc123 It's working, thank you!!! Wondering why have to fix this manually.

[peterjc123]

@qizhen816 Functions are not exported by default in MSVC. We haven't done this because we didn't test this on Windows currently. Will try to work on it later.

[LI-Wen-Jiang]

I use your use your method ,but I also have the problem.
2>example-app.obj : error LNK2019: 无法解析的外部符号 "class std::tuple<class at::Tensor,class at::Tensor> __cdecl ROIPool_forward_cpu(class at::Tensor const &,class at::Tensor const &,float,int,int)" (?ROIPool_forward_cpu@@YA?AV?$tuple@VTensor@at@@V12@@std@@AEBVTensor@at@@0MHH@Z)，该符号在函数 "class std::tuple<class at::Tensor,class at::Tensor> __cdecl ROIPool_forward(class at::Tensor const &,class at::Tensor const &,double,__int64,__int64)" (?ROIPool_forward@@YA?AV?$tuple@VTensor@at@@V12@@std@@AEBVTensor@at@@0N_J1@Z) 中被引用 2>example-app.obj : error LNK2019: 无法解析的外部符号 "class at::Tensor __cdecl ROIPool_backward_cpu(class at::Tensor const &,class at::Tensor const &,class at::Tensor const &,float,int,int,int,int,int,int)" (?ROIPool_backward_cpu@@YA?AVTensor@at@@AEBV12@00MHHHHHH@Z)，该符号在函数 "class at::Tensor __cdecl ROIPool_backward(class at::Tensor const &,class at::Tensor const &,class at::Tensor const &,float,int,int,int,int,int,int)" (?ROIPool_backward@@YA?AVTensor@at@@AEBV12@00MHHHHHH@Z) 中被引用 2>example-app.obj : error LNK2019: 无法解析的外部符号 "class at::Tensor __cdecl ROIAlign_forward_cpu(class at::Tensor const &,class at::Tensor const &,float,int,int,int,bool)" (?ROIAlign_forward_cpu@@YA?AVTensor@at@@AEBV12@0MHHH_N@Z)，该符号在函数 "class at::Tensor __cdecl ROIAlign_forward(class at::Tensor const &,class at::Tensor const &,double,__int64,__int64,__int64,bool)" (?ROIAlign_forward@@YA?AVTensor@at@@AEBV12@0N_J11_N@Z) 中被引用 2>example-app.obj : error LNK2019: 无法解析的外部符号 "class at::Tensor __cdecl ROIAlign_backward_cpu(class at::Tensor const &,class at::Tensor const &,float,int,int,int,int,int,int,int,bool)" (?ROIAlign_backward_cpu@@YA?AVTensor@at@@AEBV12@0MHHHHHHH_N@Z)，该符号在函数 "class at::Tensor __cdecl ROIAlign_backward(class at::Tensor const &,class at::Tensor const &,float,int,int,int,int,int,int,int,bool)" (?ROIAlign_backward@@YA?AVTensor@at@@AEBV12@0MHHHHHHH_N@Z) 中被引用 2>example-app.obj : error LNK2019: 无法解析的外部符号 "class at::Tensor __cdecl nms_cpu(class at::Tensor const &,class at::Tensor const &,float)" (?nms_cpu@@YA?AVTensor@at@@AEBV12@0M@Z)，该符号在函数 "class at::Tensor __cdecl nms(class at::Tensor const &,class at::Tensor const &,double)" (?nms@@YA?AVTensor@at@@AEBV12@0N@Z) 中被引用 2>example-app.obj : error LNK2019: 无法解析的外部符号 "class at::Tensor __cdecl ROIAlign_forward_cuda(class at::Tensor const &,class at::Tensor const &,float,int,int,int,bool)" (?ROIAlign_forward_cuda@@YA?AVTensor@at@@AEBV12@0MHHH_N@Z)，该符号在函数 "class at::Tensor __cdecl ROIAlign_forward(class at::Tensor const &,class at::Tensor const &,double,__int64,__int64,__int64,bool)" (?ROIAlign_forward@@YA?AVTensor@at@@AEBV12@0N_J11_N@Z) 中被引用 2>example-app.obj : error LNK2019: 无法解析的外部符号 "class at::Tensor __cdecl ROIAlign_backward_cuda(class at::Tensor const &,class at::Tensor const &,float,int,int,int,int,int,int,int,bool)" (?ROIAlign_backward_cuda@@YA?AVTensor@at@@AEBV12@0MHHHHHHH_N@Z)，该符号在函数 "class at::Tensor __cdecl ROIAlign_backward(class at::Tensor const &,class at::Tensor const &,float,int,int,int,int,int,int,int,bool)" (?ROIAlign_backward@@YA?AVTensor@at@@AEBV12@0MHHHHHHH_N@Z) 中被引用 2>example-app.obj : error LNK2019: 无法解析的外部符号 "class std::tuple<class at::Tensor,class at::Tensor> __cdecl ROIPool_forward_cuda(class at::Tensor const &,class at::Tensor const &,float,int,int)" (?ROIPool_forward_cuda@@YA?AV?$tuple@VTensor@at@@V12@@std@@AEBVTensor@at@@0MHH@Z)，该符号在函数 "class std::tuple<class at::Tensor,class at::Tensor> __cdecl ROIPool_forward(class at::Tensor const &,class at::Tensor const &,double,__int64,__int64)" (?ROIPool_forward@@YA?AV?$tuple@VTensor@at@@V12@@std@@AEBVTensor@at@@0N_J1@Z) 中被引用 2>example-app.obj : error LNK2019: 无法解析的外部符号 "class at::Tensor __cdecl ROIPool_backward_cuda(class at::Tensor const &,class at::Tensor const &,class at::Tensor const &,float,int,int,int,int,int,int)" (?ROIPool_backward_cuda@@YA?AVTensor@at@@AEBV12@00MHHHHHH@Z)，该符号在函数 "class at::Tensor __cdecl ROIPool_backward(class at::Tensor const &,class at::Tensor const &,class at::Tensor const &,float,int,int,int,int,int,int)" (?ROIPool_backward@@YA?AVTensor@at@@AEBV12@00MHHHHHH@Z) 中被引用

[peterjc123]

@LI-Wen-Jiang Have you linked your library against torchvision.lib?

[LI-Wen-Jiang]

[peterjc123]

@LI-Wen-Jiang First, you'll need to change the header and rebuild torchvision. And then for the headers installed (in D:\torchvision\include I suppose), you'll need to remove those annotations in those headers again and link against the library ('D:\torchvision\lib\torchvision.lib').

[LI-Wen-Jiang]

@peterjc123 The problem still alive. Maybe my method is wrong. Do you know how to use torchvision script to build maskrcnn in windows, thanks.

[qizhen816]

@peterjc123 I fixed ROIAlign and nms by adding __declspec(dllexport), but ROIPool is not working. I tried
std::tuple<at::Tensor, at::Tensor> __declspec(dllexport) ROIPool_forward_cpu
and
__declspec(dllexport) std::tuple<at::Tensor, at::Tensor> ROIPool_forward_cpu,
any clue where to add?