0%
January 11, 2023

Pytorch/libtorch with CPP API

C++

pytorch

Contents

Install libtorch in Windows

From libtorch official web site we select

and click the link it pops up to download.

Models from pytorch to pytorch C++

A Model in Python
import torch
import torch.nn.functional as F


class Net(torch.nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.layer1 = torch.nn.Linear(100, 256)
        self.layer2 = torch.nn.Linear(256, 1)

    def forward(self, x):
        x = self.layer1(x)
        x = F.relu(x)
        x = self.layer2(x)
        x = F.relu(x)
        return x


if __name__ == "__main__":
    traced_net = torch.jit.trace(Net(), torch.randn(1, 100))
    torch.jit.save(traced_net, "models/net.pt")
Inference in C++
Minimal Working CMakeLists.txt
add_executable(pymode_to_cpp main.cpp)

target_link_libraries(pymode_to_cpp PUBLIC ${TORCH_LIBRARIES})

if (MSVC)
  message("copying dll files")
  file(GLOB TORCH_DLLS "${TORCH_INSTALL_PREFIX}/lib/*.dll")
  add_custom_command(TARGET pymode_to_cpp
                     POST_BUILD
                     COMMAND ${CMAKE_COMMAND} -E copy_if_different
                     ${TORCH_DLLS}
                     $<TARGET_FILE_DIR:pymode_to_cpp>)
endif (MSVC)
The main.cpp File
#include "torch/script.h"
#include "torch/torch.h"
#include <iostream>
#include <string>

std::string model_pt_path{"C:\\Users\\user\\Repos\\C++\\
2023-01-12-DGGAN-in-pytorch-cpp-and-CMake-practice\\models\\net.pt"};

int main() {
    torch::jit::script::Module net = torch::jit::load(model_pt_path);
    torch::Tensor x = torch::randn({1, 100});
    torch::Tensor y = torch::randn({1, 100});
    torch::Tensor inputs = torch::cat({x, y});
    std::vector<torch::IValue> x_{inputs};
    torch::Tensor yTensor = net.forward(x_).toTensor();
    size_t ySize = yTensor.sizes()[0];
    float* yDataPtr = (float*)yTensor.data_ptr();
    try {
        // float result = output.toTensor().item<float>();
        for (int i = 0; i < ySize; i++) {
            float value = yDataPtr[i];
            std::cout << "The Float Value output: " << value << std::endl;
        }
    } catch (const c10::Error& e) {
        std::cerr << e.msg() << std::endl;
    }
}

which yields

The Float Value output: 0.606562
The Float Value output: 0.160477

Note that when yTensor is known to have batchSize 1, we may use Tensor.item<float>() to retrieve the data instead of resorting to the data_ptr() method.