上一篇文章講了如果把pytorch模型轉(zhuǎn)化成torchscripts模型，這篇介紹如何在libtorch中調(diào)用

先貼代碼，細(xì)節(jié)等后面再補(bǔ)充

#include <torch/script.h> // One-stop header.
#include <iostream>
#include <memory>
#include <stdlib.h>
#include <string>
#include <opencv2/opencv.hpp>
#include <opencv2/core/core.hpp>
#include <opencv2/imgproc/imgproc.hpp>


int main()
{
    // 加載模型
    std::shared_ptr<torch::jit::script::Module> module = torch::jit::load("D:\\scripts\\simple_facenet\\log\\torch_script_eval.pt");
    assert(module != nullptr);
    // 轉(zhuǎn)化為GPU模型
    module->to(at::kCUDA);
    std::cout << "ok\n" << std::endl;

    
    std::vector<torch::jit::IValue> inputs_1,inputs_2;
    
    // 加載圖片，這里加載了兩張一樣的圖片
    cv::Mat image_1,image_2;
    image_1 = cv::imread("D:\\datasets\\lfw_yolo_182\\Aaron_Eckhart\\Aaron_Eckhart_0001.jpg");
    image_2 = cv::imread("D:\\datasets\\lfw_yolo_182\\Aaron_Eckhart\\Aaron_Eckhart_0001.jpg");
    
    // 數(shù)據(jù)轉(zhuǎn)換
    // OPENCV 默認(rèn)格式為BGR，而我訓(xùn)練模型用的是RGB格式，所以需要轉(zhuǎn)換
    cv::cvtColor(image_1, image_1, CV_BGR2RGB);
    cv::cvtColor(image_2, image_2, CV_BGR2RGB);
    cv::Mat img_float_1,img_float_2;
    image_1.convertTo(img_float_1, CV_32F, 1.0 / 255);
    image_2.convertTo(img_float_2, CV_32F, 1.0 / 255);
    cv::resize(img_float_1, img_float_1, cv::Size(182, 182));
    cv::resize(img_float_2, img_float_2, cv::Size(182, 182));
    auto img_tensor_1 = torch::CPU(torch::kFloat32).tensorFromBlob(img_float_1.data, { 1,182,182,3 });
    auto img_tensor_2 = torch::CPU(torch::kFloat32).tensorFromBlob(img_float_2.data, { 1,182,182,3 });
    img_tensor_1 = img_tensor_1.permute({ 0,3,1,2 }); // ???nchw
    img_tensor_2 = img_tensor_2.permute({ 0,3,1,2 }); // ???nchw
    auto img_var_1 = torch::autograd::make_variable(img_tensor_1, false);
    auto img_var_2 = torch::autograd::make_variable(img_tensor_2, false);
    inputs_1.push_back(img_var_1.to(at::kCUDA));
    inputs_2.push_back(img_var_2.to(at::kCUDA));
    std::cout << "image ready" << std::endl;


    // forward
    auto output_1 = module->forward(inputs_1).toTensor();
    auto output_2 = module->forward(inputs_2).toTensor();
    
    vector<float> feature1(128),feature2(128);

    for (int i = 0;i<128;i++)
    {
        // 轉(zhuǎn)化成Float
        feature1[i] = output_1[0][i].item().toFloat();
        feature2[i] = output_2[0][i].item().toFloat();
    }

    /* 計(jì)算vector之間的距離
    ... ...
    */

    getchar();
    return 0;
}

這里為了方便部署，把最后的結(jié)果轉(zhuǎn)化為vector了，對libtorch的API還沒有摸得很清楚，不知道是否還有跟高效率的轉(zhuǎn)化方式。

有問題的話，可以私信或者加好友交流。

參考：
https://github.com/Jack-An/TorchDemo/blob/master/main.cpp
https://oldpan.me/archives/pytorch-c-libtorch-inference