用C++写一个部署在nvidia开发板上的yolov5模型推理代码

A: 首先，需要安装CUDA和CUDNN的相关库，并下载yolov5的代码仓库。然后，按照以下步骤来编写部署在nvidia开发板上的yolov5模型推理代码。

1. 导入必要的头文件

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>
#include <dirent.h>
#include <errno.h>
#include <fcntl.h>
#include <unistd.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <sys/time.h>

#include <cuda_runtime_api.h>
#include <cublas_v2.h>
#include <curand.h>
#include <cudnn.h>

#include "darknet.h"

2. 定义模型类和相关参数

typedef struct{
    network net;
    float *output;
    float *input;
    cudaStream_t stream;
} yolov5;

#define BATCH 1
#define DEVICE 0
#define NMS_THRESH 0.45
#define CONF_THRESH 0.25
#define CLASS_NUM 80
#define ANCHOR_NUM 3
#define INPUT_H 640
#define INPUT_W 640
#define CLASS_NAME_FILE "./coco.names"
#define WEIGHTS_FILE "./yolov5s.weights"

3. 加载模型

void load_model(yolov5 *model, char *classes_path, char *weights_path){
    // 加载类别名
    char **names = get_labels(classes_path);
    // 加载模型
    model->net = parse_network_cfg(cfg_path);
    if(weights_path){
        load_weights(&model->net, weights_path);
    }
    // 设置输入层
    model->input = (float *)calloc(model->net.batch * model->net.inputs, sizeof(float));
    cudaMalloc((void **)&model->output, model->net.batch * model->net.outputs * sizeof(float));
    cudaStreamCreate(&model->stream);
    set_batch_network(&model->net, BATCH);
    model->net.layers[model->net.n - 1].classes = CLASS_NUM;
    model->net.layers[model->net.n - 1].anchor_num = ANCHOR_NUM;
    model->net.layers[model->net.n - 1].confidence_thresh = CONF_THRESH;
    model->net.layers[model->net.n - 1].nms_thresh = NMS_THRESH;
    model->net.layers[model->net.n - 1].mask = (int *)calloc(ANCHOR_NUM, sizeof(int));
    for(int i = 0; i < ANCHOR_NUM; i++) model->net.layers[model->net.n - 1].mask[i] = i;
    srand(time(NULL));
}

4. 推理函数

void yolov5_inference(yolov5 *model, char *img_path){
    // 加载图片
    image img = load_image_color(img_path, 0, 0);
    // 缩放图片
    image sized = resize_image(img, INPUT_W, INPUT_H);
    // 将数据写入输入层
    fill_cuda_data(sized.data, INPUT_H * INPUT_W * 3, model->input, INPUT_H * INPUT_W * 3 * BATCH, model->stream);
    // 进行推理
    forward_network(&model->net, model->input);
    // 获取输出结果
    get_network_boxes(&model->net, img.w, img.h, CONF_THRESH, model->net.layers[model->net.n - 1].mask, 0, 1, model->output, 1, &model->net.layers[model->net.n - 1], model->net.classes, model->net.outputs, 1);
    // 进行非极大值抑制
    do_nms_sort(model->net.hold_cpu, model->net.hold_gpu, model->net.batch, model->net.layers[model->net.n - 1].classes, model->net.layers[model->net.n - 1].w, model->net.layers[model->net.n - 1].h, model->output, model->net.layers[model->net.n - 1].nms, CLASS_NUM, NMS_THRESH);
    // 输出结果
    draw_detections(img, model->net.hold_cpu, model->net.darknet_gpu, model->net.layers[model->net.n - 1], CLASS_NUM, names, WINDOW_WIDTH, WINDOW_HEIGHT, 0);
}

5. main函数

int main(int argc, char **argv){
    yolov5 model;
    // 加载模型
    load_model(&model, CLASS_NAME_FILE, WEIGHTS_FILE);
    // 进行推理
    yolov5_inference(&model, argv[1]);
    // 释放资源
    cudaFree(model.input);
    cudaFree(model.output);
    cudaStreamDestroy(model.stream);
    free_network(model.net);
    return 0;
}

以上就是用C语言编写的在nvidia开发板上部署yolov5模型推理代码。