NVIDIA Jetson YOLOv5 tensorRT部署和加速 C++版

前言

在实现NVIDIA Jetson AGX Xavier 部署YOLOv5的深度学习环境，然后能正常推理跑模型后；发现模型速度不够快，于是使用tensorRT部署，加速模型，本文介绍C++版本的。

NVIDIA Jetson YOLOv5应用与部署_一颗小树x的博客-CSDN博客

版本介绍：yolov5 v6.0、tensorrtx；Jetpack 4.5 [L4T 32.5.0]、CUDA: 10.2.89。

我测试了 kitti 数据集的100张图片：加速后每一张图像，平均推理时间是22ms，感觉还行。

一、下载yolov5 v6.0和tensorrtx

二、生成 xxx.wts文件

三、修改配置

四、编译tensorrtx

五、运行

六、解析关键代码

七、Batch size 进一步加速实验

一、下载yolov5 v6.0和tensorrtx

yolov5 v6.0版本，下载来至 yolov5 release v6.0,

git clone -b v6.0 https://github.com/ultralytics/yolov5.git

对应版本的tensorrtx：

git clone https://github.com/wang-xinyu/tensorrtx.git

二、生成 xxx.wts文件

首先复制 {tensorrtx}/yolov5/gen_wts.py 文件到 {ultralytics}/yolov5 中；其中{tensorrtx} 是名称，不同版本名称不一致，这里叫tensorrtx-master；比如，tensorrtx-master 和 yolov5 在同级目录：

cp tensorrtx-master/yolov5/gen_wts.py ./yolov5

进入yolov5 工程目录

cd yolov5

可以把yolov5s.pt 放到yolov5 里面，然后生成yolov5s.wts

python gen_wts.py -w yolov5s.pt -o yolov5s.wts

三、修改配置

进入tensorrtx 的 yolov5目录中，cd {tensorrtx}/yolov5/

cd tensorrtx-master/yolov5

3.1 C++版本的注意看yolov5.cpp、yololayer.h；首先看yolov5.cpp，它可以设置GPU id、NMS thresh、BBox confidence thresh、Batch size、推理精度（INT8/FP16/FP32）等等参数。

3.2 然后看一下yololayer.h文件，它可以设置模型的类别，输入大小等等。

使用摄像头推理（默认摄像头0），修改yolov5.cpp即可：

四、编译tensorrtx

首先进入tensorrtx 的 yolov5目录中，cd {tensorrtx}/yolov5/

cd tensorrtx-master/yolov5

建立build目录，准备编译工作

mkdir build
cd build

复制刚才生成的 yolov5s.wts 文件到build目录中

cp {ultralytics}/yolov5/yolov5s.wts {tensorrtx}/yolov5/build

然后编译

cmake ..
make

五、运行

YOLOv5s模型

首先用yolov5s.wts生成yolov5s.engine，然后用yolov5s.engine运行；

sudo ./yolov5 -s yolov5s.wts yolov5s.engine s
sudo ./yolov5 -d yolov5s.engine ../samples

sudo ./yolov5 -s yolov5s.wts yolov5s.engine s 中的s对应模型级别（可以选择：n/s/m/l/x/n6/s6/m6/l6/x6）

../samples 中链接指向了两张图片。可以自己创建一个文件夹，放一些图片进去测试。

如果是YOLOv5m模型

sudo ./yolov5 -s yolov5m.wts yolov5m.engine m
sudo ./yolov5 -d yolov5m.engine ../samples

我测试了 kitti 数据集的100张图片：（每一张图像，平均推理时间是22ms，感觉还行；后面测试一些实时的视频流处理速度）

其他数据集测试效果：

六、解析关键代码

C++版本的注意看yolov5.cpp、yololayer.h；首先看yolov5.cpp，它可以设置GPU id、NMS thresh、BBox confidence thresh、Batch size、推理精度（INT8/FP16/FP32）等等参数。

然后看一下yololayer.h文件，它可以设置模型的类别，输入大小等等。

使用摄像头推理（默认摄像头0），修改yolov5.cpp即可：

#include <iostream>
#include <chrono>
#include <cmath>
#include "cuda_utils.h"
#include "logging.h"
#include "common.hpp"
#include "utils.h"
#include "calibrator.h"
#include "preprocess.h"// OpenCV includes
#include <opencv2/opencv.hpp>
#include <opencv2/highgui.hpp>
#include <string>#define USE_FP16  // set USE_INT8 or USE_FP16 or USE_FP32
#define DEVICE 0  // GPU id
#define NMS_THRESH 0.4
#define CONF_THRESH 0.5
#define BATCH_SIZE 1
#define MAX_IMAGE_INPUT_SIZE_THRESH 3000 * 3000 // ensure it exceed the maximum size in the input images !// stuff we know about the network and the input/output blobs
static const int INPUT_H = Yolo::INPUT_H;
static const int INPUT_W = Yolo::INPUT_W;
static const int CLASS_NUM = Yolo::CLASS_NUM;
static const int OUTPUT_SIZE = Yolo::MAX_OUTPUT_BBOX_COUNT * sizeof(Yolo::Detection) / sizeof(float) + 1;  // we assume the yololayer outputs no more than MAX_OUTPUT_BBOX_COUNT boxes that conf >= 0.1
const char* INPUT_BLOB_NAME = "data";
const char* OUTPUT_BLOB_NAME = "prob";
static Logger gLogger;static int get_width(int x, float gw, int divisor = 8) {return int(ceil((x * gw) / divisor)) * divisor;
}static int get_depth(int x, float gd) {if (x == 1) return 1;int r = round(x * gd);if (x * gd - int(x * gd) == 0.5 && (int(x * gd) % 2) == 0) {--r;}return std::max<int>(r, 1);
}ICudaEngine* build_engine(unsigned int maxBatchSize, IBuilder* builder, IBuilderConfig* config, DataType dt, float& gd, float& gw, std::string& wts_name) {INetworkDefinition* network = builder->createNetworkV2(0U);// Create input tensor of shape {3, INPUT_H, INPUT_W} with name INPUT_BLOB_NAMEITensor* data = network->addInput(INPUT_BLOB_NAME, dt, Dims3{ 3, INPUT_H, INPUT_W });assert(data);std::map<std::string, Weights> weightMap = loadWeights(wts_name);/* ------ yolov5 backbone------ */auto conv0 = convBlock(network, weightMap, *data,  get_width(64, gw), 6, 2, 1,  "model.0");assert(conv0);auto conv1 = convBlock(network, weightMap, *conv0->getOutput(0), get_width(128, gw), 3, 2, 1, "model.1");auto bottleneck_CSP2 = C3(network, weightMap, *conv1->getOutput(0), get_width(128, gw), get_width(128, gw), get_depth(3, gd), true, 1, 0.5, "model.2");auto conv3 = convBlock(network, weightMap, *bottleneck_CSP2->getOutput(0), get_width(256, gw), 3, 2, 1, "model.3");auto bottleneck_csp4 = C3(network, weightMap, *conv3->getOutput(0), get_width(256, gw), get_width(256, gw), get_depth(6, gd), true, 1, 0.5, "model.4");auto conv5 = convBlock(network, weightMap, *bottleneck_csp4->getOutput(0), get_width(512, gw), 3, 2, 1, "model.5");auto bottleneck_csp6 = C3(network, weightMap, *conv5->getOutput(0), get_width(512, gw), get_width(512, gw), get_depth(9, gd), true, 1, 0.5, "model.6");auto conv7 = convBlock(network, weightMap, *bottleneck_csp6->getOutput(0), get_width(1024, gw), 3, 2, 1, "model.7");auto bottleneck_csp8 = C3(network, weightMap, *conv7->getOutput(0), get_width(1024, gw), get_width(1024, gw), get_depth(3, gd), false, 1, 0.5, "model.8");auto spp9 = SPPF(network, weightMap, *bottleneck_csp8->getOutput(0), get_width(1024, gw), get_width(1024, gw), 5, "model.9");/* ------ yolov5 head ------ */auto conv10 = convBlock(network, weightMap, *spp9->getOutput(0), get_width(512, gw), 1, 1, 1, "model.10");auto upsample11 = network->addResize(*conv10->getOutput(0));assert(upsample11);upsample11->setResizeMode(ResizeMode::kNEAREST);upsample11->setOutputDimensions(bottleneck_csp6->getOutput(0)->getDimensions());ITensor* inputTensors12[] = { upsample11->getOutput(0), bottleneck_csp6->getOutput(0) };auto cat12 = network->addConcatenation(inputTensors12, 2);auto bottleneck_csp13 = C3(network, weightMap, *cat12->getOutput(0), get_width(1024, gw), get_width(512, gw), get_depth(3, gd), false, 1, 0.5, "model.13");auto conv14 = convBlock(network, weightMap, *bottleneck_csp13->getOutput(0), get_width(256, gw), 1, 1, 1, "model.14");auto upsample15 = network->addResize(*conv14->getOutput(0));assert(upsample15);upsample15->setResizeMode(ResizeMode::kNEAREST);upsample15->setOutputDimensions(bottleneck_csp4->getOutput(0)->getDimensions());ITensor* inputTensors16[] = { upsample15->getOutput(0), bottleneck_csp4->getOutput(0) };auto cat16 = network->addConcatenation(inputTensors16, 2);auto bottleneck_csp17 = C3(network, weightMap, *cat16->getOutput(0), get_width(512, gw), get_width(256, gw), get_depth(3, gd), false, 1, 0.5, "model.17");/* ------ detect ------ */IConvolutionLayer* det0 = network->addConvolutionNd(*bottleneck_csp17->getOutput(0), 3 * (Yolo::CLASS_NUM + 5), DimsHW{ 1, 1 }, weightMap["model.24.m.0.weight"], weightMap["model.24.m.0.bias"]);auto conv18 = convBlock(network, weightMap, *bottleneck_csp17->getOutput(0), get_width(256, gw), 3, 2, 1, "model.18");ITensor* inputTensors19[] = { conv18->getOutput(0), conv14->getOutput(0) };auto cat19 = network->addConcatenation(inputTensors19, 2);auto bottleneck_csp20 = C3(network, weightMap, *cat19->getOutput(0), get_width(512, gw), get_width(512, gw), get_depth(3, gd), false, 1, 0.5, "model.20");IConvolutionLayer* det1 = network->addConvolutionNd(*bottleneck_csp20->getOutput(0), 3 * (Yolo::CLASS_NUM + 5), DimsHW{ 1, 1 }, weightMap["model.24.m.1.weight"], weightMap["model.24.m.1.bias"]);auto conv21 = convBlock(network, weightMap, *bottleneck_csp20->getOutput(0), get_width(512, gw), 3, 2, 1, "model.21");ITensor* inputTensors22[] = { conv21->getOutput(0), conv10->getOutput(0) };auto cat22 = network->addConcatenation(inputTensors22, 2);auto bottleneck_csp23 = C3(network, weightMap, *cat22->getOutput(0), get_width(1024, gw), get_width(1024, gw), get_depth(3, gd), false, 1, 0.5, "model.23");IConvolutionLayer* det2 = network->addConvolutionNd(*bottleneck_csp23->getOutput(0), 3 * (Yolo::CLASS_NUM + 5), DimsHW{ 1, 1 }, weightMap["model.24.m.2.weight"], weightMap["model.24.m.2.bias"]);auto yolo = addYoLoLayer(network, weightMap, "model.24", std::vector<IConvolutionLayer*>{det0, det1, det2});yolo->getOutput(0)->setName(OUTPUT_BLOB_NAME);network->markOutput(*yolo->getOutput(0));// Build enginebuilder->setMaxBatchSize(maxBatchSize);config->setMaxWorkspaceSize(16 * (1 << 20));  // 16MB
#if defined(USE_FP16)config->setFlag(BuilderFlag::kFP16);
#elif defined(USE_INT8)std::cout << "Your platform support int8: " << (builder->platformHasFastInt8() ? "true" : "false") << std::endl;assert(builder->platformHasFastInt8());config->setFlag(BuilderFlag::kINT8);Int8EntropyCalibrator2* calibrator = new Int8EntropyCalibrator2(1, INPUT_W, INPUT_H, "./coco_calib/", "int8calib.table", INPUT_BLOB_NAME);config->setInt8Calibrator(calibrator);
#endifstd::cout << "Building engine, please wait for a while..." << std::endl;ICudaEngine* engine = builder->buildEngineWithConfig(*network, *config);std::cout << "Build engine successfully!" << std::endl;// Don't need the network any morenetwork->destroy();// Release host memoryfor (auto& mem : weightMap){free((void*)(mem.second.values));}return engine;
}ICudaEngine* build_engine_p6(unsigned int maxBatchSize, IBuilder* builder, IBuilderConfig* config, DataType dt, float& gd, float& gw, std::string& wts_name) {INetworkDefinition* network = builder->createNetworkV2(0U);// Create input tensor of shape {3, INPUT_H, INPUT_W} with name INPUT_BLOB_NAMEITensor* data = network->addInput(INPUT_BLOB_NAME, dt, Dims3{ 3, INPUT_H, INPUT_W });assert(data);std::map<std::string, Weights> weightMap = loadWeights(wts_name);/* ------ yolov5 backbone------ */auto conv0 = convBlock(network, weightMap, *data,  get_width(64, gw), 6, 2, 1,  "model.0");auto conv1 = convBlock(network, weightMap, *conv0->getOutput(0), get_width(128, gw), 3, 2, 1, "model.1");auto c3_2 = C3(network, weightMap, *conv1->getOutput(0), get_width(128, gw), get_width(128, gw), get_depth(3, gd), true, 1, 0.5, "model.2");auto conv3 = convBlock(network, weightMap, *c3_2->getOutput(0), get_width(256, gw), 3, 2, 1, "model.3");auto c3_4 = C3(network, weightMap, *conv3->getOutput(0), get_width(256, gw), get_width(256, gw), get_depth(6, gd), true, 1, 0.5, "model.4");auto conv5 = convBlock(network, weightMap, *c3_4->getOutput(0), get_width(512, gw), 3, 2, 1, "model.5");auto c3_6 = C3(network, weightMap, *conv5->getOutput(0), get_width(512, gw), get_width(512, gw), get_depth(9, gd), true, 1, 0.5, "model.6");auto conv7 = convBlock(network, weightMap, *c3_6->getOutput(0), get_width(768, gw), 3, 2, 1, "model.7");auto c3_8 = C3(network, weightMap, *conv7->getOutput(0), get_width(768, gw), get_width(768, gw), get_depth(3, gd), true, 1, 0.5, "model.8");auto conv9 = convBlock(network, weightMap, *c3_8->getOutput(0), get_width(1024, gw), 3, 2, 1, "model.9");auto c3_10 = C3(network, weightMap, *conv9->getOutput(0), get_width(1024, gw), get_width(1024, gw), get_depth(3, gd), false, 1, 0.5, "model.10");auto sppf11 = SPPF(network, weightMap, *c3_10->getOutput(0), get_width(1024, gw), get_width(1024, gw), 5, "model.11");/* ------ yolov5 head ------ */auto conv12 = convBlock(network, weightMap, *sppf11->getOutput(0), get_width(768, gw), 1, 1, 1, "model.12");auto upsample13 = network->addResize(*conv12->getOutput(0));assert(upsample13);upsample13->setResizeMode(ResizeMode::kNEAREST);upsample13->setOutputDimensions(c3_8->getOutput(0)->getDimensions());ITensor* inputTensors14[] = { upsample13->getOutput(0), c3_8->getOutput(0) };auto cat14 = network->addConcatenation(inputTensors14, 2);auto c3_15 = C3(network, weightMap, *cat14->getOutput(0), get_width(1536, gw), get_width(768, gw), get_depth(3, gd), false, 1, 0.5, "model.15");auto conv16 = convBlock(network, weightMap, *c3_15->getOutput(0), get_width(512, gw), 1, 1, 1, "model.16");auto upsample17 = network->addResize(*conv16->getOutput(0));assert(upsample17);upsample17->setResizeMode(ResizeMode::kNEAREST);upsample17->setOutputDimensions(c3_6->getOutput(0)->getDimensions());ITensor* inputTensors18[] = { upsample17->getOutput(0), c3_6->getOutput(0) };auto cat18 = network->addConcatenation(inputTensors18, 2);auto c3_19 = C3(network, weightMap, *cat18->getOutput(0), get_width(1024, gw), get_width(512, gw), get_depth(3, gd), false, 1, 0.5, "model.19");auto conv20 = convBlock(network, weightMap, *c3_19->getOutput(0), get_width(256, gw), 1, 1, 1, "model.20");auto upsample21 = network->addResize(*conv20->getOutput(0));assert(upsample21);upsample21->setResizeMode(ResizeMode::kNEAREST);upsample21->setOutputDimensions(c3_4->getOutput(0)->getDimensions());ITensor* inputTensors21[] = { upsample21->getOutput(0), c3_4->getOutput(0) };auto cat22 = network->addConcatenation(inputTensors21, 2);auto c3_23 = C3(network, weightMap, *cat22->getOutput(0), get_width(512, gw), get_width(256, gw), get_depth(3, gd), false, 1, 0.5, "model.23");auto conv24 = convBlock(network, weightMap, *c3_23->getOutput(0), get_width(256, gw), 3, 2, 1, "model.24");ITensor* inputTensors25[] = { conv24->getOutput(0), conv20->getOutput(0) };auto cat25 = network->addConcatenation(inputTensors25, 2);auto c3_26 = C3(network, weightMap, *cat25->getOutput(0), get_width(1024, gw), get_width(512, gw), get_depth(3, gd), false, 1, 0.5, "model.26");auto conv27 = convBlock(network, weightMap, *c3_26->getOutput(0), get_width(512, gw), 3, 2, 1, "model.27");ITensor* inputTensors28[] = { conv27->getOutput(0), conv16->getOutput(0) };auto cat28 = network->addConcatenation(inputTensors28, 2);auto c3_29 = C3(network, weightMap, *cat28->getOutput(0), get_width(1536, gw), get_width(768, gw), get_depth(3, gd), false, 1, 0.5, "model.29");auto conv30 = convBlock(network, weightMap, *c3_29->getOutput(0), get_width(768, gw), 3, 2, 1, "model.30");ITensor* inputTensors31[] = { conv30->getOutput(0), conv12->getOutput(0) };auto cat31 = network->addConcatenation(inputTensors31, 2);auto c3_32 = C3(network, weightMap, *cat31->getOutput(0), get_width(2048, gw), get_width(1024, gw), get_depth(3, gd), false, 1, 0.5, "model.32");/* ------ detect ------ */IConvolutionLayer* det0 = network->addConvolutionNd(*c3_23->getOutput(0), 3 * (Yolo::CLASS_NUM + 5), DimsHW{ 1, 1 }, weightMap["model.33.m.0.weight"], weightMap["model.33.m.0.bias"]);IConvolutionLayer* det1 = network->addConvolutionNd(*c3_26->getOutput(0), 3 * (Yolo::CLASS_NUM + 5), DimsHW{ 1, 1 }, weightMap["model.33.m.1.weight"], weightMap["model.33.m.1.bias"]);IConvolutionLayer* det2 = network->addConvolutionNd(*c3_29->getOutput(0), 3 * (Yolo::CLASS_NUM + 5), DimsHW{ 1, 1 }, weightMap["model.33.m.2.weight"], weightMap["model.33.m.2.bias"]);IConvolutionLayer* det3 = network->addConvolutionNd(*c3_32->getOutput(0), 3 * (Yolo::CLASS_NUM + 5), DimsHW{ 1, 1 }, weightMap["model.33.m.3.weight"], weightMap["model.33.m.3.bias"]);auto yolo = addYoLoLayer(network, weightMap, "model.33", std::vector<IConvolutionLayer*>{det0, det1, det2, det3});yolo->getOutput(0)->setName(OUTPUT_BLOB_NAME);network->markOutput(*yolo->getOutput(0));// Build enginebuilder->setMaxBatchSize(maxBatchSize);config->setMaxWorkspaceSize(16 * (1 << 20));  // 16MB
#if defined(USE_FP16)config->setFlag(BuilderFlag::kFP16);
#elif defined(USE_INT8)std::cout << "Your platform support int8: " << (builder->platformHasFastInt8() ? "true" : "false") << std::endl;assert(builder->platformHasFastInt8());config->setFlag(BuilderFlag::kINT8);Int8EntropyCalibrator2* calibrator = new Int8EntropyCalibrator2(1, INPUT_W, INPUT_H, "./coco_calib/", "int8calib.table", INPUT_BLOB_NAME);config->setInt8Calibrator(calibrator);
#endifstd::cout << "Building engine, please wait for a while..." << std::endl;ICudaEngine* engine = builder->buildEngineWithConfig(*network, *config);std::cout << "Build engine successfully!" << std::endl;// Don't need the network any morenetwork->destroy();// Release host memoryfor (auto& mem : weightMap){free((void*)(mem.second.values));}return engine;
}void APIToModel(unsigned int maxBatchSize, IHostMemory** modelStream, bool& is_p6, float& gd, float& gw, std::string& wts_name) {// Create builderIBuilder* builder = createInferBuilder(gLogger);IBuilderConfig* config = builder->createBuilderConfig();// Create model to populate the network, then set the outputs and create an engineICudaEngine *engine = nullptr;if (is_p6) {engine = build_engine_p6(maxBatchSize, builder, config, DataType::kFLOAT, gd, gw, wts_name);} else {engine = build_engine(maxBatchSize, builder, config, DataType::kFLOAT, gd, gw, wts_name);}assert(engine != nullptr);// Serialize the engine(*modelStream) = engine->serialize();// Close everything downengine->destroy();builder->destroy();config->destroy();
}void doInference(IExecutionContext& context, cudaStream_t& stream, void **buffers, float* output, int batchSize) {// infer on the batch asynchronously, and DMA output back to hostcontext.enqueue(batchSize, buffers, stream, nullptr);CUDA_CHECK(cudaMemcpyAsync(output, buffers[1], batchSize * OUTPUT_SIZE * sizeof(float), cudaMemcpyDeviceToHost, stream));cudaStreamSynchronize(stream);
}bool parse_args(int argc, char** argv, std::string& wts, std::string& engine, bool& is_p6, float& gd, float& gw, std::string& img_dir) {if (argc < 4) return false;if (std::string(argv[1]) == "-s" && (argc == 5 || argc == 7)) {wts = std::string(argv[2]);engine = std::string(argv[3]);auto net = std::string(argv[4]);if (net[0] == 'n') {gd = 0.33;gw = 0.25;} else if (net[0] == 's') {gd = 0.33;gw = 0.50;} else if (net[0] == 'm') {gd = 0.67;gw = 0.75;} else if (net[0] == 'l') {gd = 1.0;gw = 1.0;} else if (net[0] == 'x') {gd = 1.33;gw = 1.25;} else if (net[0] == 'c' && argc == 7) {gd = atof(argv[5]);gw = atof(argv[6]);} else {return false;}if (net.size() == 2 && net[1] == '6') {is_p6 = true;}} else if (std::string(argv[1]) == "-d" && argc == 4) {engine = std::string(argv[2]);img_dir = std::string(argv[3]);} else {return false;}return true;
}int main(int argc, char** argv) {// opencvcv::VideoCapture cap; // 1.创建视频采集对象;cv::Mat readImage;     //    读取的图片;cap.open(0);         // 2.打开默认相机;if (!cap.isOpened()) std::cout << "open Capture error !!!" << std::endl;else std::cout << "open Capture OK !!!" << std::endl;// cap.release();    // 释放视频采集对象!!!cudaSetDevice(DEVICE);std::string wts_name = "";std::string engine_name = "";bool is_p6 = false;float gd = 0.0f, gw = 0.0f;std::string img_dir;if (!parse_args(argc, argv, wts_name, engine_name, is_p6, gd, gw, img_dir)) {std::cerr << "arguments not right!" << std::endl;std::cerr << "./yolov5 -s [.wts] [.engine] [n/s/m/l/x/n6/s6/m6/l6/x6 or c/c6 gd gw]  // serialize model to plan file" << std::endl;std::cerr << "./yolov5 -d [.engine] ../samples  // deserialize plan file and run inference" << std::endl;return -1;}// create a model using the API directly and serialize it to a streamif (!wts_name.empty()) {IHostMemory* modelStream{ nullptr };APIToModel(BATCH_SIZE, &modelStream, is_p6, gd, gw, wts_name);assert(modelStream != nullptr);std::ofstream p(engine_name, std::ios::binary);if (!p) {std::cerr << "could not open plan output file" << std::endl;return -1;}p.write(reinterpret_cast<const char*>(modelStream->data()), modelStream->size());modelStream->destroy();return 0;}// deserialize the .engine and run inferencestd::ifstream file(engine_name, std::ios::binary);if (!file.good()) {std::cerr << "read " << engine_name << " error!" << std::endl;return -1;}char *trtModelStream = nullptr;size_t size = 0;file.seekg(0, file.end);size = file.tellg();file.seekg(0, file.beg);trtModelStream = new char[size];assert(trtModelStream);file.read(trtModelStream, size);file.close();std::vector<std::string> file_names;if (read_files_in_dir(img_dir.c_str(), file_names) < 0) {std::cerr << "read_files_in_dir failed." << std::endl;return -1;}static float prob[BATCH_SIZE * OUTPUT_SIZE];IRuntime* runtime = createInferRuntime(gLogger);assert(runtime != nullptr);ICudaEngine* engine = runtime->deserializeCudaEngine(trtModelStream, size);assert(engine != nullptr);IExecutionContext* context = engine->createExecutionContext();assert(context != nullptr);delete[] trtModelStream;assert(engine->getNbBindings() == 2);float* buffers[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()const int inputIndex = engine->getBindingIndex(INPUT_BLOB_NAME);const int outputIndex = engine->getBindingIndex(OUTPUT_BLOB_NAME);assert(inputIndex == 0);assert(outputIndex == 1);// Create GPU buffers on deviceCUDA_CHECK(cudaMalloc((void**)&buffers[inputIndex], BATCH_SIZE * 3 * INPUT_H * INPUT_W * sizeof(float)));CUDA_CHECK(cudaMalloc((void**)&buffers[outputIndex], BATCH_SIZE * OUTPUT_SIZE * sizeof(float)));// Create streamcudaStream_t stream;CUDA_CHECK(cudaStreamCreate(&stream));uint8_t* img_host = nullptr;uint8_t* img_device = nullptr;// prepare input data cache in pinned memory CUDA_CHECK(cudaMallocHost((void**)&img_host, MAX_IMAGE_INPUT_SIZE_THRESH * 3));// prepare input data cache in device memoryCUDA_CHECK(cudaMalloc((void**)&img_device, MAX_IMAGE_INPUT_SIZE_THRESH * 3));int fcount = 0;int save_int = 0;std::vector<cv::Mat> imgs_buffer(BATCH_SIZE);std::vector<AffineMatrix> matrix_buffer(BATCH_SIZE);while (true) {// for (int f = 0; f < (int)file_names.size(); f++) {if (cv::waitKey(1)  == 'q') break; //如果按下q，会推出程序 fcount++;save_int++;if (fcount < BATCH_SIZE ) continue;//auto start = std::chrono::system_clock::now();float* buffer_idx = (float*)buffers[inputIndex];for (int b = 0; b < fcount; b++) {cv::Mat img;cap >> img;// cv::Mat img = cv::imread(img_dir + "/" + file_names[f - fcount + 1 + b]); // ############if (img.empty()) continue;imgs_buffer[b] = img;size_t  size_image = img.cols * img.rows * 3;size_t  size_image_dst = INPUT_H * INPUT_W * 3;//copy data to pinned memorymemcpy(img_host,img.data,size_image);//copy data to device memoryCUDA_CHECK(cudaMemcpyAsync(img_device,img_host,size_image,cudaMemcpyHostToDevice,stream));preprocess_kernel_img(img_device, img.cols, img.rows, buffer_idx, matrix_buffer[b], INPUT_W, INPUT_H, stream);       buffer_idx += size_image_dst;}// Run inferenceauto start = std::chrono::system_clock::now();doInference(*context, stream, (void**)buffers, prob, BATCH_SIZE);auto end = std::chrono::system_clock::now();std::cout << "inference time: " << std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count() << "ms" << std::endl;std::vector<std::vector<Yolo::Detection>> batch_res(fcount);for (int b = 0; b < fcount; b++) {auto& res = batch_res[b];nms(res, &prob[b * OUTPUT_SIZE], CONF_THRESH, NMS_THRESH);}for (int b = 0; b < fcount; b++) {auto& res = batch_res[b];auto& bbox_affine_matrix = matrix_buffer[b];cv::Mat img = imgs_buffer[b];for (size_t j = 0; j < res.size(); j++) {cv::Rect r = get_rect(res[j].bbox, bbox_affine_matrix);cv::rectangle(img, r, cv::Scalar(0x27, 0xC1, 0x36), 2);cv::putText(img, std::to_string((int)res[j].class_id), cv::Point(r.x, r.y - 1), cv::FONT_HERSHEY_PLAIN, 1.2, cv::Scalar(0xFF, 0xFF, 0xFF), 2);}// cv::imwrite("_" + file_names[f - fcount + 1 + b], img);cv::imwrite(std::to_string(save_int)+".jpg", img);}fcount = 0;}// Release stream and bufferscudaStreamDestroy(stream);CUDA_CHECK(cudaFree(img_device));CUDA_CHECK(cudaFreeHost(img_host));CUDA_CHECK(cudaFree(buffers[inputIndex]));CUDA_CHECK(cudaFree(buffers[outputIndex]));// Destroy the enginecontext->destroy();engine->destroy();runtime->destroy();cap.release(); // 释放视频采集对象!!!// Print histogram of the output distribution//std::cout << "\nOutput:\n\n";//for (unsigned int i = 0; i < OUTPUT_SIZE; i++)//{//    std::cout << prob[i] << ", ";//    if (i % 10 == 0) std::cout << std::endl;//}//std::cout << std::endl;return 0;
}

效果：

七、Batch size 进一步加速实验

官方说当 batchsize=8 时，预处理 + 推理速度提高 3 倍；于是我试了一下；还是 kitti 数据集的100张图片，

当batchsize=8 时，一次推理8张图片，平均时间是46ms；46ms / 8 = 5.75ms；即现在推理一张图片需要5.75ms，对比上面单张推理时间22ms，快了3.8倍左右。

这在一个设备用来推理多个视频流输入时，还是挺不错的。

本文参考 wang-xinyu 大佬开源的tensorrtx ，致谢：https://github.com/wang-xinyu/tensorrtx

参考：https://github.com/wang-xinyu/tensorrtx/blob/master/yolov5/README.md