在MLab HomePod 2.0 pro上编译基于pybind11的程序出错 about mlab

[gemfield]

从错误信息看，发生错误的pybind并非使用的是MLab HomePod 2.0 pro内置的pybind11：

gemfield@homepod:~/src$ dpkg -l | grep pybind11
ii  pybind11-dev                                                2.4.3-2build2                                                            all          seamless operability between C++11 and Python

而是libtorch中打包的pybind：

/deepvac/libtorch/include/pybind11/pybind11.h:1315:13: note: candidate: ??template<class Func, class ... Extra> pybind11::class_< <template-parameter-1-1>, <template-parameter-1-2> >& pybind11::class_< <template-parameter-1-1>, <template-parameter-1-2> >::def(const char*, Func&&, const Extra& ...) [with Func = Func; Extra = {Extra ...}; type_ = deepvac::SyszuxPornCls; options = {}]??
 1315 |     class_ &def(const char *name_, Func&& f, const Extra&... extra) {
      |             ^~~

但是刚才测试发现，这两者都会报错。

[gemfield]

这是因为错误中的process函数实现了重载：

std::optional<std::pair<int, float>> process(cv::Mat frame);
at::Tensor process(at::Tensor& t);

所以PYBIND11_MODULE的语法：

    py::class_<SyszuxVisionTerror>(m, "SyszuxVisionTerror")
        .def(py::init<std::string>())
        .def("process", &SyszuxVisionTerror::process);

需要改写为：

    py::class_<SyszuxVisionTerror>(m, "SyszuxVisionTerror")
        .def(py::init<std::string>())
        .def("process", py::overload_cast<at::Tensor&>(&SyszuxVisionTerror::process), "forward at Tensor.")
        .def("process", py::overload_cast<cv::Mat>(&SyszuxVisionTerror::process), "forward cv Mat.");