纯小白初次学习深度学习，根据目前所学、所查资料总结该配置文章，如有更好的方法将深度学习模型通过C++部署到实际工程中，希望学习交流。

软件版本

• Python3.10.9
• Pytorch2.0.0
• Qt Creator 5.0.1 (Community)
• Qt 1.15.2(MSVC 2019, 64bit)

一、下载Libtorch

• 下载位置（可以按照参考的方法来下载）：Pytorch

二、配置Qt Creator

• Qt Creator配置：
  • 创建.pri文件；
  • 内部写入如下内容：

INCLUDEPATH += C:\\soft_install\\libtorch-win-shared-with-deps-2.0.0+cpu\\libtorch\\include
INCLUDEPATH += C:\\soft_install\\libtorch-win-shared-with-deps-2.0.0+cpu\\libtorch\\include\\torch\\csrc\\api\\includeLIBS += -LC:\\soft_install\\libtorch-win-shared-with-deps-2.0.0+cpu\\libtorch\\lib\\-lasmjit\\-lc10\\-lclog\\-lcpuinfo\\-ldnnl\\-lfbgemm\\-lfbjni\\-lkineto\\-llibprotobuf\\-llibprotobuf-lite\\-llibprotoc\\-lpthreadpool\\-lpytorch_jni\\-ltorch\\-ltorch_cpu\\-lXNNPACK

三、测试项目

• 模型来源
  • B站：我是土堆；课程：PyTorch深度学习快速入门教程（绝对通俗易懂！）【小土堆】
• 创建项目
  • .pro文件中引入Libtorch和Opencv的.pri

DISTFILES += \\pri/opencv.pri \\pri/libtorch.priinclude(./pri/opencv.pri)
include(./pri/libtorch.pri)

• c++调用Libtorch接口时头文件

  Libtorch的头文件应该放在cpp或.h中最头的位置（忘记在哪里看到的了）；否则需要添加

#undef slots
#include <torch/script.h>
#include <ATen/ATen.h>
#include <torch/torch.h>
#define slots Q_SLOTS

• .pth 转 .pt

import torch
import torchvision.transforms
from PIL import Image
from torch import nntest = TestSeq()
model = torch.load("../StuPytorch/testSeq_9.pth", map_location=torch.device("cpu"))# 设置一个示例输入
example = torch.rand(1, 3, 32, 32)
# 使用 torch.jit.trace 生成 torch.jit.ScriptModule
model.eval()
traced_script_module = torch.jit.trace(model, example)
# 保存转换后的模型
traced_script_module.save("../StuPytorch/testSeq_9.pt")

• 简单实现代码（见参考）
  • Pytorch实现

import torch
import torchvision.transforms
from PIL import Image
from torch import nn
from model import *img_path = "../imgs/dog.jpg"
img = Image.open(img_path)
img = img.convert("RGB")
print(img)
transform = torchvision.transforms.Compose([torchvision.transforms.Resize((32, 32)),torchvision.transforms.ToTensor()])
image = transform(img)
print(image.shape)model = torch.load("../StuPytorch/testSeq_29.pth", map_location=torch.device("cpu"))
print(model)image = torch.reshape(image, (1, 3, 32, 32))
print(image)
model.eval()
with torch.no_grad():output = model(image)class_id = {0: 'airplane', 1: 'automobile', 2: 'bird', 3: 'cat', 4: 'deer',5: 'dog', 6: 'frog', 7: 'horse', 8: 'ship', 9: 'truck'}
print(output)
print(output.argmax(1))
print(class_id[output.argmax(1).item()])---输出
TestSeq((model1): Sequential((0): Conv2d(3, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))(1): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(2): Conv2d(32, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))(3): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(4): Conv2d(32, 64, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))(5): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)(6): Flatten(start_dim=1, end_dim=-1)(7): Linear(in_features=1024, out_features=64, bias=True)(8): Linear(in_features=64, out_features=10, bias=True))
)
tensor([[-1.0712, -2.4485,  1.6878,  0.9057,  1.1824,  1.8202,  0.0850,  0.5854,-2.2806, -2.1651]])
tensor([5])
dog

- C++实现

cv::Mat img = cv::imread("D:\\\\projects\\\\PyTorchStud\\\\imgs\\\\dog.png");
cv::cvtColor(img, img, cv::COLOR_BGR2RGB);
cv::Mat resizedImg;
cv::resize(img, resizedImg, cv::Size(32, 32));//#停止autograd模块的工作，以起到加速和节省显存的作用
torch::NoGradGuard no_grad;
torch::jit::script::Module module;
module = torch::jit::load("D:\\\\projects\\\\PyTorchStud\\\\StuPytorch\\\\testSeq_29.pt");
std::cout << "Succeed in loading model" << std::endl;//标签名
std::vector<std::string> labels;
labels.push_back("airplane");
labels.push_back("automobile");
labels.push_back("bird");
labels.push_back("cat");
labels.push_back("deer");
labels.push_back("dog");
labels.push_back("frog");
labels.push_back("horse");
labels.push_back("ship");
labels.push_back("truck");torch::Tensor inputTensor = torch::from_blob(resizedImg.data, {1, 32, 32, 3}, torch::kByte);
//将张量的参数顺序转化为torch输入的格式[B,C,H,W]
inputTensor = inputTensor.permute({0, 3, 1, 2});
inputTensor = inputTensor.toType(torch::kFloat);
clock_t start_time = clock();module.eval();
torch::Tensor outTensor = module.forward({inputTensor}).toTensor();
auto prediction = outTensor.argmax(1);
std::cout << "prediction:" << prediction << std::endl;auto results = outTensor.sort(-1, true);
auto softmaxs = std::get<0>(results)[0].softmax(0);
auto indexs = std::get<1>(results)[0];//输出运行时间
clock_t end_time = clock();
std::cout << "Running time is: "<< static_cast<double>(end_time - start_time) / CLOCKS_PER_SEC * 1000 << "ms"<< std::endl;for (int i = 0; i < 1; ++i) {auto idx = indexs[i].item<int>();std::cout << "    ============= Top-" << i + 1 << " =============" << std::endl;std::cout << "    idx:  " << idx << std::endl;std::cout << "    Label:  " << labels[idx] << std::endl;std::cout << "    With Probability:  " << softmaxs[i].item<float>() * 100.0f << "%"<< std::endl;
}---输出结果
Succeed in loading model
prediction: 5
[ CPULongType{1} ]
Running time is: 118ms============= Top-1 =============idx:  5Label:  dogWith Probability:  100%

• 结果说明
  • 由于训练的模型问题，有的图片分类错误。

参考：

Libtorch：pytorch分类和语义分割模型在C++工程上的应用

小吃零食