原创，转载请注明出处，谢谢！
首先官网下载运行库：https://pytorch.org/get-started/locally/
根据自己系统下载cpp的lib
我因为自带的系统的python环境的torch是1.7.1，所以下了这个版本
划重点：本机Pytorch python训练model版本最好和libtorch版本一致！
最新版本的torch已经支持AMD显卡了，不必只能用N卡了！
PS：并不是所有模型都可以JIT，这东西需要序列化python部分代码，如果无法被序列化，那么这个模型即使被转化出来pt来也无法在CPP代码中推理，这里墙裂推荐转化为ONNX或者优图的NCNN框架模型，前者支持cpu和NVIDIA GPU推理(对，不支持A卡),后者由于使用了Vulkan GPU加速，可以支持A卡、N卡以及部分英特尔核显，当然也可以支持CPU甚至嵌入式框架
效率不一定高，但是NCNN代码非常友好！
VS版本：2019
OpenCV版本：opencv410_vc14，其他版本应该也可以
include：
xxx\libtorch-win-shared-with-deps-1.7.1\libtorch\include\torch\csrc\api\include
xxx\libtorch-win-shared-with-deps-1.7.1\libtorch\include
两个引用，前面的xxx换成你的libtorch路径
然后是lib，怕麻烦可以把lib和dll全部丢到vs工程目录.
PyTorch模型转换为Torch脚本：
import torch
import torchvision
# An instance of your model.
model = torchvision.models.resnet18()
# An example input you would normally provide to your model's forward() method.
example = torch.rand(1, 3, 224, 224)
# Use torch.jit.trace to generate a torch.jit.ScriptModule via tracing.
traced_script_module = torch.jit.trace(model, example)
traced_script_module.save("model.pt")
这里自己转换一下模型即可，pt格式模型是专门用于cpp端推理用的。
这里我们用pytorch自带的ImageNet的resnet18或resnet50模型
准备图片：
以下是cpp推理完整代码，看一下就懂了：
#include <torch/torch.h>
#include <torch/script.h>
#include <iostream>
#include <vector>
#include <opencv2/highgui.hpp>
#include <opencv2/core/core.hpp>
#include <opencv2/opencv.hpp>
//https://blog.youkuaiyun.com/yanfeng1022/article/details/106481312
torch::Device GetDevice()
{
torch::Device device = torch::kCPU;//torch::kCUDA;//cpu的话改成torch::kCPU
return device;
}
torch::jit::script::Module* Load_model(const char*path,const torch::Device &device)
{
torch::jit::script::Module* p_module = nullptr;
try {
// Deserialize the ScriptModule from a file using torch::jit::load().
auto module = torch::jit::load(path);
module.to(device);
p_module = &module;
}
catch (const c10::Error& e) {
std::cerr << "error loading the model\n";
return nullptr;
}
if (nullptr == p_module)
return nullptr;
std::cout << "Load model successful!" << std::endl;
return p_module;
}
void TorchTest(torch::jit::script::Module& model, const torch::Device& device)
{
//std::cout << "Load model successful!" << std::endl;
std::vector<torch::jit::IValue> inputs;
inputs.push_back(torch::zeros({ 1,3,224,224 }, device));
at::Tensor output = model.forward(std::move(inputs)).toTensor();
std::cout << output.slice(/*dim=*/1, /*start=*/0, /*end=*/5) << '\n';
}
void Classfier(torch::jit::script::Module* p_model)
{
cv::Mat image = cv::imread("D:\\PYTHON\\TensorFlow2_Study\\data\\131429407.jpg");
cv::resize(image, image, cv::Size(224, 224));
cv::cvtColor(image, image, cv::ColorConversionCodes::COLOR_BGR2RGB);
std::cout << "img info:" << image.rows << " " << image.cols << " " << image.channels() << std::endl;
//return 0;
torch::Tensor img_tensor = torch::from_blob(image.data, { image.rows, image.cols, 3 }, torch::kByte);
img_tensor = img_tensor.permute({ 2,0,1 });
img_tensor = img_tensor.toType(torch::kFloat);
img_tensor = img_tensor.div(255);
img_tensor = img_tensor.unsqueeze(0);
std::cout << "--flag--\n";
//return 0;
//-->OK
// 网络前向计算
at::Tensor output = p_model->forward({ img_tensor }).toTensor();
//std::cout << "output:" << output << std::endl;
auto prediction = output.argmax(1);
std::cout << "prediction:" << prediction << std::endl;
int maxk = 3;
auto top3 = std::get<1>(output.topk(maxk, 1, true, true));
std::cout << "top3: " << top3 << '\n';
}
int main(int argc, const char* argv[])
{
auto device = GetDevice();
std::cout << device << std::endl;
//https://zhuanlan.zhihu.com/p/72750321
//https://zhuanlan.zhihu.com/p/68901339
torch::jit::script::Module module = torch::jit::load("D:\\PYTHON\\TensorFlow2_Study\\resnet.pt");
module.to(device);
std::vector<torch::jit::IValue> inputs;
cv::Mat image = cv::imread("D:\\PYTHON\\TensorFlow2_Study\\data\\131429407.jpg");
cv::resize(image, image, cv::Size(224, 224));
cv::cvtColor(image, image, cv::ColorConversionCodes::COLOR_BGR2RGB);
std::cout << "img info:" << image.rows << " " << image.cols << " " << image.channels() << std::endl;
//return 0;
//torch::Tensor img_tensor = torch::from_blob(image.data, { image.rows, image.cols, 3 }, torch::kByte);
//img_tensor = img_tensor.permute({ 2,0,1 });
//img_tensor = img_tensor.toType(torch::kFloat);
//img_tensor = img_tensor.div(255);
//img_tensor = img_tensor.unsqueeze(0);
torch::Tensor img_tensor = torch::from_blob(image.data, {1, image.rows, image.cols, 3 }, torch::kByte);
img_tensor = img_tensor.permute({ 0, 3, 1, 2 });
img_tensor = img_tensor.toType(torch::kFloat);
img_tensor = img_tensor.div(255);
//img_tensor = img_tensor.unsqueeze(0);
img_tensor.to(device, caffe2::ScalarType::Float);
inputs.push_back(img_tensor);
//https://www.jianshu.com/p/a613dbc03fe0
// Execute the model and turn its output into a tensor.
torch::Tensor output = module.forward(std::move(inputs)).toTensor();
std::cout << output.slice(/*dim=*/1, /*start=*/0, /*end=*/5) << '\n';
auto prediction = output.argmax(1);
auto index = prediction.item().toInt();//https://www.jianshu.com/p/9e8eb211df62
std::cout << "prediction:" << index << std::endl;
if (248 == index)
{
std::cout << "prediction:Eskimo dog, husky\n";
}
std::cout << "--end--\n";
system("pause");
}
如果你都是按照上面步骤做的话，结果应该是(我用的是第一张狗狗图片)：
PS：上面的248是resnet1000的标签字典；可以通过这个字典找到对应的分类
字典下载: https://pan.baidu.com/s/1kchNjyWZJn8i2nFA3dk3Eg 提取码: im7f
备份下载：https://download.youkuaiyun.com/download/weixin_44029053/18521257
错误代码：
file not found: model/constants.pkl或file not found: archive/constants.pkl
原因：你的pt文件实际上是一个zip，你可以解压看看：ref：https://www.ccoderun.ca/programming/doxygen/pytorch/md_pytorch_torch_csrc_jit_docs_serialization.html
而这个constants.pkl是pt压缩包里面的，如果没有，
看看生成pt和推理pt的版本是否统一
我这个pt文件就没有constants.pkl文件，结果jit加载就报错，解决方法是：同时升级推理和生成pt的pytorch版本？
建议升级到1.8.0或以上！
PS：Pytorch py版离线(极力推荐！)安装：https://zhuanlan.zhihu.com/p/151251025，注意各个版本对应关系
PS：Ads时间，大家可以加群一起探讨一下各种深度学习或者图形框架，集思广益，群号：558174476（游戏与人工智能生命体）
扫一扫
4555
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
