C++ class python wrapper

C++
1

Hi, I am trying to make python wrapper for model execution in C++

I defined a class which loads torchscript model in the constructor and run a model by calling its method. code is as follows
– scripttest.cpp –

#include <torch/script.h>
#include <pybind11/pybind11.h>
#include <iostream>
#include <vector>
#include <torch/extension.h>
using namespace std;
namespace py = pybind11;

class Model{
    private:
        torch::jit::script::Module module;
        torch::Tensor t1;
    public:
        Model(){
            t1 = torch::randn({1,3,11,11}, torch::dtype(at::kFloat).device(c10::kCUDA));
            module = torch::jit::load("test.pt");
        }
        at::Tensor get(){
            at::GradMode::set_enabled(false);
            at::Tensor t2 = module.forward({t1}).toTensor();
            return t2;
        }
};

PYBIND11_MODULE(scripttest, m){
    py::class_<Model>(m, "Model")
        .def(py::init<>())
        .def("get", &Model::get);
}

— CMakeLists.txt —

cmake_minimum_required(VERSION 3.4...3.18)
set(CMAKE_CXX_STANDARD 14)
project(scripttest LANGUAGES CXX)

#include_directories("<libtorch/to/path>")

find_package(Torch REQUIRED)
find_package(pybind11 REQUIRED)
#add_library(scripttest)
pybind11_add_module(scripttest scripttest.cpp)
//find_library(TORCH_PYTHON_LIBRARY torch_python PATHS "{TORCH_INSTALL_PREFIX}/lib")
target_link_libraries(scripttest PRIVATE "${TORCH_LIBRARIES}" )//${TORCH_PYTHON_LIBRARY})

– python script –

from scripttest import Model
m = Model()
out = m.get()

It gives "ImportError: undefined symbol: _Z16THPVariable_WrapN2at6TensorE"

When I removed <torch/extension.h> in the scripttest.cpp, it says "TypeError: Unable to convert function return value to a Python type! The signature was (self: build.scripttest.Model) → at::Tensor"

When I uncommented CMakeLists.txt to use TORCH_PYTHON_LIBRARY, this returns Segmentation fault (core dumped) at m.get().

I downloaded 1.9.1+cu11.1 libtorch for ubuntu and 1.7.1 torch in anaconda.

Actually if I change get function not to return anything, then it does not raise any problem. But, I have to return torch result and also I should add torch tensor argument for get method later to replace random input tensor defined in get method.

Does anyone know how to solve it?
I read this thread(Can libtorch return torch::Tensor to pytorch with pybind11？ - #3 by tom) and this tutorial(Custom C++ and CUDA Extensions — PyTorch Tutorials 1.9.1+cu102 documentation) but I want to use C++ class.