1. 参考博客和代码、模型仓库：
1.1. 【C++随记】collect2: error: ld returned 1 exit status错误分析与解决
1.2. Visual Studio 2022新建 cmake 工程测试 tensorRT 自带样例 sampleOnnxMNIST
1.3.报错：ModuleNotFoundError: No module named ‘tensorrt’
1.4. 推荐使用：Amirstan_plugin - 助力TensorRT高效能优化
1.5. 利用TensorRT实现神经网络提速(读取ONNX模型并运行)
1.6. Ubuntu下安装CUDA
1.7. 【Python】成功解决ModuleNotFoundError: No module named ‘packaging’
1.8. pytorch模型(.pth)转tensorrt模型(.engine)的几种方式
1.9. pytorch模型(.pth)转tensorrt模型(.engine)几种方式
1.10. 【Debug】TensorRT报错汇总
1.11. ace-recognition-cpp-tensorrt
1.12. arcFace 模型下载地址
1.13. retinaFace 模型下载地址
1.14. https://gitcode.com/gh_mirrors/am/amirstan_plugin/overview?utm_source=artical_gitcode&index=bottom&type=card&webUrl&isLogin=1
2. 下载代码( github 仓库地址：https://github.com/nghiapq77/face-recognition-cpp-tensorrt)
3. 修改代码：
face-recognition-cpp-tensorrt 仓库是一个功能完备的项目，包括请求、人脸检测、人脸识别( 特征向量计算 )、保存数据库、人脸特征向量相似度计算，由于本人只是想快速验证一下使用 arcFace 计算人脸特征向量是否靠谱，再加上本人对 c++ 不是很熟悉，不想花费大量的时间用在环境安装上，所以我对代码做了精简，只摘出了使用 arcface 计算人脸特征向量的部分，以下是我摘出的 arcface 部分的完整代码：
3.1. 代码结构：
arcface_test/conversion/torch2trt.py // 未修改cpp/arcface.cppcommon.cppinclude/arcface.hcommon.hface_images/xxxmodels/backbone_ir50_asia.pth // 下载地址：https://github.com/ZhaoJ9014/face.evoLVe#model-zoo( xxx仓库的 readme 中也有下载链接引导，我下载的是IR-50 Private Asia Face Data 亚洲人脸的） )test.cppCMakeLists.txt3.2. arcface.cpp：
#include "arcface.h"void getCroppedFaces(cv::Mat frame, std::vector<struct Bbox> &outputBbox, int resize_w, int resize_h, std::vector<struct CroppedFace> &croppedFaces) {croppedFaces.clear();for (std::vector<struct Bbox>::iterator it = outputBbox.begin(); it != outputBbox.end(); it++) {cv::Rect facePos(cv::Point((*it).y1, (*it).x1), cv::Point((*it).y2, (*it).x2));cv::Mat tempCrop = frame(facePos);struct CroppedFace currFace;cv::resize(tempCrop, currFace.faceMat, cv::Size(resize_h, resize_w), 0, 0, cv::INTER_CUBIC);currFace.face = currFace.faceMat.clone();currFace.x1 = it->x1;currFace.y1 = it->y1;currFace.x2 = it->x2;currFace.y2 = it->y2;croppedFaces.push_back(currFace);}
}int ArcFaceIR50::classCount = 0;ArcFaceIR50::ArcFaceIR50(TRTLogger gLogger, const std::string engineFile, int frameWidth, int frameHeight, std::string inputName, std::string outputName,std::vector<int> inputShape, int outputDim, int maxBatchSize, int maxFacesPerScene, float knownPersonThreshold) {m_frameWidth = static_cast<const int>(frameWidth);m_frameHeight = static_cast<const int>(frameHeight);assert(inputShape.size() == 3);m_INPUT_C = static_cast<const int>(inputShape[0]);m_INPUT_H = static_cast<const int>(inputShape[1]);m_INPUT_W = static_cast<const int>(inputShape[2]);m_OUTPUT_D = static_cast<const int>(outputDim);m_INPUT_SIZE = static_cast<const int>(m_INPUT_C * m_INPUT_H * m_INPUT_W * sizeof(float));m_OUTPUT_SIZE = static_cast<const int>(m_OUTPUT_D * sizeof(float));m_maxBatchSize = static_cast<const int>(maxBatchSize);m_embed = new float[m_OUTPUT_D];croppedFaces.reserve(maxFacesPerScene);m_embeds = new float[maxFacesPerScene * m_OUTPUT_D];m_knownPersonThresh = knownPersonThreshold;// load engine from .engine file or create new engineloadEngine(gLogger, engineFile);// create stream and pre-allocate GPU buffers memorypreInference(inputName, outputName);
}void ArcFaceIR50::loadEngine(TRTLogger gLogger, const std::string engineFile) {if (fileExists(engineFile)) {std::cout << "[INFO] Loading ArcFace Engine...\n";std::vector<char> trtModelStream_;size_t size{0};std::ifstream file(engineFile, std::ios::binary);if (file.good()) {file.seekg(0, file.end);size = file.tellg();file.seekg(0, file.beg);trtModelStream_.resize(size);file.read(trtModelStream_.data(), size);file.close();}nvinfer1::IRuntime *runtime = nvinfer1::createInferRuntime(gLogger);assert(runtime != nullptr);m_engine = runtime->deserializeCudaEngine(trtModelStream_.data(), size);assert(m_engine != nullptr);m_context = m_engine->createExecutionContext();assert(m_context != nullptr);} else {throw std::logic_error("Cant find engine file");}
}void ArcFaceIR50::preInference() {// Engine requires exactly IEngine::getNbBindings() number of buffers.assert(m_engine->getNbBindings() == 2);// In order to bind the buffers, we need to know the names of the input and// output tensors. Note that indices are guaranteed to be less than IEngine::getNbBindings()inputIndex = m_engine->getBindingIndex("input");outputIndex = m_engine->getBindingIndex("output");// Create GPU buffers on devicecudaMalloc(&buffers[inputIndex], m_maxBatchSize * m_INPUT_SIZE);cudaMalloc(&buffers[outputIndex], m_maxBatchSize * m_OUTPUT_SIZE);// Create streamcudaStreamCreate(&stream);
}void ArcFaceIR50::preInference(std::string inputName, std::string outputName) {// Engine requires exactly IEngine::getNbBindings() number of buffers.assert(m_engine->getNbBindings() == 2);// In order to bind the buffers, we need to know the names of the input and// output tensors. Note that indices are guaranteed to be less than IEngine::getNbBindings()inputIndex = m_engine->getBindingIndex(inputName.c_str());outputIndex = m_engine->getBindingIndex(outputName.c_str());// Create GPU buffers on devicecudaMalloc(&buffers[inputIndex], m_maxBatchSize * m_INPUT_SIZE);cudaMalloc(&buffers[outputIndex], m_maxBatchSize * m_OUTPUT_SIZE);// Create streamcudaStreamCreate(&stream);
}void ArcFaceIR50::preprocessFace(cv::Mat &face, cv::Mat &output) {cv::cvtColor(face, face, cv::COLOR_BGR2RGB);face.convertTo(face, CV_32F);face = (face - cv::Scalar(127.5, 127.5, 127.5)) * 0.0078125;std::vector<cv::Mat> temp;cv::split(face, temp);for (int i = 0; i < temp.size(); i++) {output.push_back(temp[i]);}
}void ArcFaceIR50::preprocessFaces() {for (int i = 0; i < croppedFaces.size(); i++) {cv::cvtColor(croppedFaces[i].faceMat, croppedFaces[i].faceMat, cv::COLOR_BGR2RGB);croppedFaces[i].faceMat.convertTo(croppedFaces[i].faceMat, CV_32F);croppedFaces[i].faceMat = (croppedFaces[i].faceMat - cv::Scalar(127.5, 127.5, 127.5)) * 0.0078125;std::vector<cv::Mat> temp;cv::split(croppedFaces[i].faceMat, temp);for (int i = 0; i < temp.size(); i++) {m_input.push_back(temp[i]);}croppedFaces[i].faceMat = m_input.clone();m_input.release();}
}void ArcFaceIR50::doInference(float *input, float *output) {// DMA input batch data to device, infer on the batch asynchronously, and DMA output back to hostcudaMemcpyAsync(buffers[inputIndex], input, m_INPUT_SIZE, cudaMemcpyHostToDevice, stream);m_context->enqueueV2(buffers, stream, nullptr);cudaMemcpyAsync(output, buffers[outputIndex], m_OUTPUT_SIZE, cudaMemcpyDeviceToHost, stream);cudaStreamSynchronize(stream);
}void ArcFaceIR50::doInference(float *input, float *output, int batchSize) {// Set input dimensionsm_context->setBindingDimensions(inputIndex, nvinfer1::Dims4(batchSize, m_INPUT_C, m_INPUT_H, m_INPUT_W));// DMA input batch data to device, infer on the batch asynchronously, and DMA output back to hostcudaMemcpyAsync(buffers[inputIndex], input, batchSize * m_INPUT_SIZE, cudaMemcpyHostToDevice, stream);m_context->enqueueV2(buffers, stream, nullptr);cudaMemcpyAsync(output, buffers[outputIndex], batchSize * m_OUTPUT_SIZE, cudaMemcpyDeviceToHost, stream);cudaStreamSynchronize(stream);
}void ArcFaceIR50::addEmbedding(const std::string className, float embedding[]) {classNames.push_back(className);std::copy(embedding, embedding + m_OUTPUT_D, m_knownEmbeds + classCount * m_OUTPUT_D);classCount++;
}void ArcFaceIR50::addEmbedding(const std::string className, std::vector<float> embedding) {classNames.push_back(className);std::copy(embedding.begin(), embedding.end(), m_knownEmbeds + classCount * m_OUTPUT_D);classCount++;
}void ArcFaceIR50::initKnownEmbeds(int num) { m_knownEmbeds = new float[num * m_OUTPUT_D]; }void ArcFaceIR50::forward(cv::Mat frame, std::vector<struct Bbox> outputBbox) {getCroppedFaces(frame, outputBbox, m_INPUT_W, m_INPUT_H, croppedFaces);preprocessFaces();if (m_maxBatchSize < 2) {for (int i = 0; i < croppedFaces.size(); i++) {doInference((float *)croppedFaces[i].faceMat.ptr<float>(0), m_embed);std::copy(m_embed, m_embed + m_OUTPUT_D, m_embeds + i * m_OUTPUT_D);}std::cout << "---------------------------------------------------" << std::endl;std::cout << "[";for (int i = 0; i < m_OUTPUT_D; ++i) {std::cout << m_embed[i] << ",";}std::cout << "]" << std::endl;std::cout << "---------------------------------------------------" << std::endl;} else {int num = croppedFaces.size();int end = 0;for (int beg = 0; beg < croppedFaces.size(); beg = beg + m_maxBatchSize) {end = std::min(num, beg + m_maxBatchSize);cv::Mat input;for (int i = beg; i < end; ++i) {input.push_back(croppedFaces[i].faceMat);}doInference((float *)input.ptr<float>(0), m_embed, end - beg);std::copy(m_embed, m_embed + (end - beg) * m_OUTPUT_D, m_embeds + (end - beg) * beg * m_OUTPUT_D);}std::cout << "---------------------------------------------------" << std::endl;std::cout << "[";for (int i = 0; i < m_OUTPUT_D; ++i) {std::cout << m_embed[i] << ",";}std::cout << "]" << std::endl;std::cout << "---------------------------------------------------" << std::endl;}
}std::tuple<std::vector<std::string>, std::vector<float>> ArcFaceIR50::getOutputs(float *output_sims) {/*Get person corresponding to maximum similarity score based on cosine similarity matrix.*/std::vector<std::string> names;std::vector<float> sims;for (int i = 0; i < croppedFaces.size(); ++i) {int argmax = std::distance(output_sims + i * classCount, std::max_element(output_sims + i * classCount, output_sims + (i + 1) * classCount));float sim = *(output_sims + i * classCount + argmax);std::string name = classNames[argmax];names.push_back(name);sims.push_back(sim);}return std::make_tuple(names, sims);
}void ArcFaceIR50::visualize(cv::Mat &image, std::vector<std::string> names, std::vector<float> sims) {for (int i = 0; i < croppedFaces.size(); ++i) {float fontScaler = static_cast<float>(croppedFaces[i].x2 - croppedFaces[i].x1) / static_cast<float>(m_frameWidth);cv::Scalar color;if (sims[i] >= m_knownPersonThresh)color = cv::Scalar(0, 255, 0);elsecolor = cv::Scalar(0, 0, 255);cv::rectangle(image, cv::Point(croppedFaces[i].y1, croppedFaces[i].x1), cv::Point(croppedFaces[i].y2, croppedFaces[i].x2), color, 2, 8, 0);cv::putText(image, names[i] + " " + std::to_string(sims[i]), cv::Point(croppedFaces[i].y1 + 2, croppedFaces[i].x2 - 3), cv::FONT_HERSHEY_DUPLEX,0.1 + 2 * fontScaler, color, 1);}
}void ArcFaceIR50::resetEmbeddings() {classCount = 0;classNames.clear();
}ArcFaceIR50::~ArcFaceIR50() {// Release stream and bufferscudaStreamDestroy(stream);cudaFree(buffers[inputIndex]);cudaFree(buffers[outputIndex]);
}
3.3. common.cpp：
#include "common.h"bool fileExists(const std::string &name) {std::ifstream f(name.c_str());return f.good();
}void getFilePaths(std::string rootPath, std::vector<struct Paths> &paths) {DIR *dir;struct dirent *entry;std::string postfix = ".jpg";if ((dir = opendir(rootPath.c_str())) != NULL) {while ((entry = readdir(dir)) != NULL) {std::string class_path = rootPath + "/" + entry->d_name;DIR *class_dir = opendir(class_path.c_str());struct dirent *file_entry;while ((file_entry = readdir(class_dir)) != NULL) {std::string name(file_entry->d_name);if (name.length() >= postfix.length() && 0 == name.compare(name.length() - postfix.length(), postfix.length(), postfix))if (file_entry->d_type != DT_DIR) {struct Paths tempPaths;tempPaths.className = std::string(entry->d_name);tempPaths.absPath = class_path + "/" + name;paths.push_back(tempPaths);}}}closedir(dir);}
}3.4. arcface.h：
#ifndef ARCFACE_H
#define ARCFACE_H#include <opencv2/core.hpp>
#include <opencv2/imgproc.hpp>
#include <tuple>#include "common.h"struct CroppedFace {cv::Mat face;cv::Mat faceMat;int x1, y1, x2, y2;
};void getCroppedFaces(cv::Mat frame, std::vector<struct Bbox> &outputBbox, int resize_w, int resize_h, std::vector<struct CroppedFace> &croppedFaces);class ArcFaceIR50 {public:ArcFaceIR50(TRTLogger gLogger, const std::string engineFile, int frameWidth, int frameHeight, std::string inputName, std::string outputName,std::vector<int> inputShape, int outputDim, int maxBatchSize, int maxFacesPerScene, float knownPersonThreshold);~ArcFaceIR50();void preprocessFace(cv::Mat &face, cv::Mat &output);void doInference(float *input, float *output);void doInference(float *input, float *output, int batchSize);void addEmbedding(const std::string className, float embedding[]);void addEmbedding(const std::string className, std::vector<float> embedding);void forward(cv::Mat image, std::vector<struct Bbox> outputBbox);std::tuple<std::vector<std::string>, std::vector<float>> getOutputs(float *output_sims);void resetEmbeddings();void initKnownEmbeds(int num);void visualize(cv::Mat &image, std::vector<std::string> names, std::vector<float> sims);std::vector<struct CroppedFace> croppedFaces;static int classCount;private:void loadEngine(TRTLogger gLogger, const std::string engineFile);void preInference();void preInference(std::string inputName, std::string outputName);void preprocessFaces();int m_frameWidth, m_frameHeight, m_INPUT_C, m_INPUT_H, m_INPUT_W, m_OUTPUT_D, m_INPUT_SIZE, m_OUTPUT_SIZE, m_maxBatchSize, m_maxFacesPerScene;float m_knownPersonThresh;cv::Mat m_input;float *m_embed, *m_embeds, *m_knownEmbeds, *m_outputs;std::vector<std::string> classNames;nvinfer1::ICudaEngine *m_engine;nvinfer1::IExecutionContext *m_context;cudaStream_t stream;void *buffers[2];int inputIndex, outputIndex;
};#endif // ARCFACE_H
3.5. common.h：
#ifndef COMMON_H
#define COMMON_H#include "NvInfer.h"
// #include "cuda_runtime_api.h"
#include <cublasLt.h>
#include <dirent.h>
#include <fstream>
#include <iostream>
#include <string>
#include <vector>struct Bbox {int x1, y1, x2, y2;float score;
};struct Paths {std::string absPath;std::string className;
};bool fileExists(const std::string &name);
void getFilePaths(std::string rootPath, std::vector<struct Paths> &paths);class TRTLogger : public nvinfer1::ILogger {public:void log(nvinfer1::ILogger::Severity severity, const char *msg) noexcept override {switch (severity) {case Severity::kINTERNAL_ERROR:std::cerr << "INTERNAL_ERROR: ";break;case Severity::kERROR:std::cerr << "ERROR: ";break;case Severity::kWARNING:std::cerr << "WARNING: ";break;case Severity::kINFO:std::cerr << "INFO: ";break;case Severity::kVERBOSE:std::cerr << "VERBOSE: ";break;default:std::cerr << "UNKNOWN: ";break;}std::cerr << msg << std::endl;}
};#endif // COMMON_H
3.6. test.cpp：
#include <opencv2/imgcodecs.hpp>
#include <opencv2/opencv.hpp>#include "arcface.h"
#include "common.h"int main(int argc, char* argv[]) {TRTLogger gLogger;int videoFrameWidth = 640;int videoFrameHeight = 480;std::vector<int> recInputShape = {3, 112, 112};int recOutputDim = 512;std::string recEngineFile = "/mnt/d/code/c_code/arcface_test/face-recognition-cpp-tensorrt/arcface_test/conversion/arcface-ir50_asia-112x112-b1-fp16.engine";std::cout << "recEngineFile = " << recEngineFile << std::endl;int maxFacesPerScene = 4;float knownPersonThreshold = 0.65;int recMaxBatchSize = 1;std::string recInputName = "input";std::string recOutputName = "output";std::vector<struct Bbox> outputBbox;outputBbox.reserve(maxFacesPerScene);ArcFaceIR50 recognizer(gLogger,recEngineFile,videoFrameWidth,videoFrameHeight,recInputName,recOutputName,recInputShape,recOutputDim,recMaxBatchSize,maxFacesPerScene,knownPersonThreshold);std::string face_image_path = argv[1];std::cout << "face_image_path = " << face_image_path << std::endl;cv::Mat frame = cv::imread( face_image_path );// frame 应该是表示使用 opencv 加载的一张图片，即mat// frame = cv::imdecode(data, cv::IMREAD_UNCHANGED);int height = frame.size[0];int width = frame.size[1];std::cout << "Image: " << frame.size() << std::endl;if (frame.empty()){throw "Empty image";}if ((height != recInputShape[1]) || (width != recInputShape[2])) {std::cout << "Resizing input to " << recInputShape[1] << "x" << recInputShape[2] << "\n";cv::resize(frame, frame, cv::Size(recInputShape[1], recInputShape[2]));}std::cout << "Getting embedding...";Bbox bbox;bbox.x1 = 0;bbox.y1 = 0;bbox.x2 = recInputShape[1];bbox.y2 = recInputShape[2];bbox.score = 1;outputBbox.push_back(bbox);recognizer.forward(frame, outputBbox);
}3.7. CMakeLists.txt( 请自行修改你的 cuda、tensorrt 等路径 )：
cmake_minimum_required (VERSION 3.8) project("test")include_directories( "/soft/TensorRT-8.6.1.6/include" "/usr/local/cuda/include""/mnt/d/code/c_code/arcface_test/face-recognition-cpp-tensorrt/arcface_test/include")link_directories( "/soft/TensorRT-8.6.1.6/lib""/usr/local/cuda/lib64""/soft/opencv-3.4.6/build/lib")link_libraries( nvinfer nvinfer_pluginnvonnxparsernvparserscudnncublascudartopencv_coreopencv_imgproc opencv_highguiopencv_imgcodecs)# 将源代码添加到此项目的可执行文件。
add_executable("test" "/mnt/d/code/c_code/arcface_test/face-recognition-cpp-tensorrt/arcface_test/test.cpp" "/mnt/d/code/c_code/arcface_test/face-recognition-cpp-tensorrt/arcface_test/cpp/arcface.cpp" "/mnt/d/code/c_code/arcface_test/face-recognition-cpp-tensorrt/arcface_test/cpp/common.cpp")3.8. cuda、tensorrt、opencv 版本说明：
截止到 2024年8月28日，该仓库中 readme --》Requirements 显示的该项目依赖的 cuda、tensorrt、opencv 的版本如下：
CUDA 11.3TensorRT 8.2.2.1OpenCV 4.5.5但是本人的 cuda、tensorrt、opencv 版本如下所示：
CUDA 11.2TensorRT 8.6.1OpenCV xxx最后证明也是可以编译通过并且成功计算人脸特征向量的，但是如果你的 tensorrt 是7.x 版本就放弃吧，因为 tensort8 相较于 Tensorrt7单单 API上就做了很多调整。
3.9. cmake 、make 执行生成了可执行文件 test，但是现在还没有 .engine 文件，需要先试用 python 脚本生成 .engine 文件方可进行测试。
4. 生成 .engine 文件：
4.1. cd conversion
4.2. python torch2trt.py( 如果成功，会生成 arcface-ir50_asia-112x112-b1-fp16.engine 文件 )
4.3. 报错以及解决方案：
大部分错都可以百度很快的解决( 比如缺 Python 依赖什么的 )，这里只列举一个花了我很久才解决的，就是需要安装 torch2trt_dynamic( ps：整个过程中一定要有耐心，不要砸桌子，根据环境的不同，需要陆陆续续安装很多依赖，而安装每一个依赖可能都很费劲，加油吧 )，以下是安装步骤：
git clone https://github.com/grimoire/torch2trt_dynamic.gitcd torch2trt_dynamicpython setup.py install安装过程确实很顺利，但是执行 python torch2trt.py 的时候报错：
File "/usr/local/lib/python3.8/dist-packages/torch2trt_dynamic/converters/__init__.py", line 37, in <module>from .Conv2d import convert_Conv2d
ModuleNotFoundError: No module named 'torch2trt_dynamic.converters.Conv2d'我的 torch2trt_dynamic 安装到了 /usr/local/lib/python3.8/dist-packages/torch2trt_dynamic 目录下，我进入 /usr/local/lib/python3.8/dist-packages/torch2trt_dynamic/converters 下发现确实没有 Conv2d.py，倒是有个conv2d.py，我将
conv2d.py 改名为 Conv2d.py，再次运行 python torch2trt.py 发现不报刚才那个错了，但是会报一个好像是循环依赖的问题，看来应该是没有 Conv2d.py 导致的。这个错在全网搜不到直接的解决方案，都是说 torch 版本的问题，其实是因为 torch2trt_dynamic/torch2trt_dynamic/converters 下面同时存在 Conv2d.py、conv2d.py 导致的，不知道为什么 git clone 以后 Conv2d.py就没了(可能配置下大小写敏感可以可决吧，我没试)，我是直接下载的 .zip 源码包，存在大小写问题的 .py 文件如下：
'torch2trt_dynamic/converters/Conv2d.py''torch2trt_dynamic/converters/conv2d.py''torch2trt_dynamic/converters/Identity.py''torch2trt_dynamic/converters/identity.py''torch2trt_dynamic/converters/Linear.py''torch2trt_dynamic/converters/linear.py''torch2trt_dynamic/converters/ReLU.py''torch2trt_dynamic/converters/relu.py''torch2trt_dynamic/converters/ReLU6.py''torch2trt_dynamic/converters/relu6.py'所以需要重新安装 torch2trt_dynamic，我也不知道 python setup.py install 对应的卸载命令是什么，我是直接将安装路径( 例如我的是 /usr/local/lib/python3.8/dist-packages )下的 torch2trt_dynamic 直接干掉了( ps：还要删除该目录下其余几个 torch2trt_dynamic 前缀的文件或文件夹 )。重新安装 torch2trt_dynamic 后，执行 python torch2trt.py 就成功生成了 arcface-ir50_asia-112x112-b1-fp16.engine( 可能会报几个小错，不过我遇到的都是一百度一大堆，很好解决 )
5. 测试 arcFace 的人脸识别( 计算特征向量 )：
我在 face_images 下准备了几张明星的人脸照片： 
执行 build 下的 ./test ../face_images/yaoming_0001.png 输出512维度的坐标：
最后我将这三个明星的人脸特征向量保存到了 milvus 向量数据库中，度量方式使用欧式距离，查询出的 topk还算可以，比如使用甄子丹的人脸特征向量查询结果如下：
测试 arcface 人脸识别(计算特征向量) "ubuntu20.04(wsl2)测试 arcface 人脸识别(计算特征向量)")
