前言:本篇博客为《基于树莓派4B的YOLOv5-Lite目标检测的移植与部署》的后续博客,主旨为帮助大家实现 ONNX 模型到 NCNN 模型的转换,并且在树莓派4B进行成功部署!正常情况下,NCNN 模型是优于 ONNX 模型的,但是作者实际测试下来发现貌似 ONNX 模型的FPS和精度感觉都略优秀于 NCNN 模型,读者朋友可以根据自己时间情况去选择模型的使用!


实验效果图:




YOLOv5-Lite的ONNX模型FPS:4.78
YOLOv5-Lite的NCNN模型FPS:3.77

按道理来说,NCNN 模型的推理速度是快于 ONNX 模型的,作者这边也不知道什么情况。


一、工程前瞻
NCNN 的网络模型通常需要使用简化后的 ONNX 模型来转换,ONNX 模型依赖于原始的训练权重 Weight 的存在,故此我们需要使用自己的训练集于神经网络模型进行训练!
这部分的详解教程可以借鉴:http://t.csdn.cn/msjCZ

源码地址:ultralytics/yolov5: YOLOv5  in PyTorch > ONNX > CoreML > TFLite (github.com)

读者朋友可以使用  PyCharm 或者 VsCode 打开 Yolov5-Lite 的源码(作者使用PyCharm 2020.1 x64);

在 Yolov5-Lite 的目录下找到 train.py (训练文件)的 main 函数入口,进行如下配置:

    我们设置如下几个核心配置:

    --weights v5lite-s.pt

    --cfg models/v5Lite-s.yaml

    --data data/mydata.yaml

    --img-size 320

    --batch-size 16

    --data data/mydata.yaml

    device 0/cpu                        (可以不使用CUDA训练)

读者朋友一定要将数据集存放的地址位置搞正确!!!

mydata.yaml:

Yolov5-Lite 网络模型的训练可以不一定必须使用 CUDA 进行加速,但是 pytorch 架构等依赖库一定需要满足,模型训练依赖要求如下:

# base ----------------------------------------
matplotlib>=3.2.2
numpy>=1.18.5
opencv-python>=4.1.2
Pillow
PyYAML>=5.3.1
scipy>=1.4.1
torch>=1.8.0
torchvision>=0.9.0
tqdm>=4.41.0
 
# logging -------------------------------------
tensorboard>=2.4.1
# wandb
 
# plotting ------------------------------------
seaborn>=0.11.0
pandas
 
# export --------------------------------------
# coremltools>=4.1
# onnx>=1.9.1
# scikit-learn==0.19.2  # for coreml quantization
 
# extras --------------------------------------
thop  # FLOPS computation
pycocotools>=2.0  # COCO mAP

将训练环境与数据集都搞定之后,就可以点击运行按钮进行 Yolov5-Lite 的模型训练了!

训练成功之后,将会在当前目录下的 run 文件下的 trian 文件下找到 expx (x代表数字),expx 则存放了第 次训练时候的各种数据内容,包括:历史最优权重best_weight,当前权重last_weight,训练结果result等等; 

二、NCNN概述

NCNN 是一个针对移动平台优化的高性能神经网络推理框架,并在2017年7月正式开源。NCNN 是腾讯优图最“火”的开源项目之一,作为一个为手机端极致优化的高性能神经网络前向计算框架,在设计之初便将手机端的特殊场景融入核心理念,是业界首个为移动端优化的开源神经网络推断库。能实现无第三方依赖,跨平台操作,在手机端 cpu 运算速度在开源框架中处于领先水平。基于该平台,开发者能够轻松将深度学习算法移植到手机端,输出高效的执行,进而产出人工智能APP,将AI技术带到用户指尖。

支持大部分常用的 CNN 网络

  •     Classical CNN: VGG AlexNet GoogleNet Inception ...
  •     Practical CNN: ResNet DenseNet SENet FPN ...
  •     Light-weight CNN: SqueezeNet MobileNetV1/V2/V3 ShuffleNetV1/V2 MNasNet ...
  •     Face Detection: MTCNN RetinaFace ...
  •     Detection: VGG-SSD MobileNet-SSD SqueezeNet-SSD MobileNetV2-SSDLite MobileNetV3-SSDLite ...
  •     Detection: Faster-RCNN R-FCN ...
  •     Detection: YOLOV2 YOLOV3 MobileNet-YOLOV3 YOLOV4 YOLOV5 ...
  •     Segmentation: FCN PSPNet UNet YOLACT ...
  •     Pose Estimation: SimplePose ...

 

  官网地址:ncnn: ncnn ncnn 是腾讯优图实验室首个开源项目,是一个为手机端极致优化的高性能神经网络前向计算框架 (gitee.com)

三、树莓派4B的NCNN部署Lite模型

3.1 树莓派配置NCNN

1、安装依赖库;

sudo apt-get install git cmake
sudo apt-get install -y gfortran
sudo apt-get install -y libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev protobuf-compiler
sudo apt-get install --no-install-recommends libboost-all-dev
sudo apt-get install -y libgflags-dev libgoogle-glog-dev liblmdb-dev libatlas-base-dev

  

2、下载NCNN并编译;

$ git clone https://gitee.com/Tencent/ncnn.git
cd ncnn
mkdir build
cd build
cmake ..
make -j4
make install

完成后ncnn文件夹如下:


3.2 模型转换

如今的开源 YOLO 系列神经网络模型的目录下作者都会预留 export.py 文件将该神经网络模型进行转换到 ONNX 模型,方便大家实际情况下部署使用!

将我们训练好的最优训练权重 weights 存放到 YOLOv5-Lite 主目录下,之后运行如下代码:

python export.py --weights 'weights/last.pt' --batch-size 1 --img-size 320

使用 onnx-simplifier 对转换后的 onnx 进行简化:

pip install -U onnx-simplifier --user
python -m onnxsim best.onnx e.onnx


3.3 树莓派部署lite模型
1、将 ONNX 模型转换为 NCNN 模型

cd ncnn/build
./tools/onnx/onnx2ncnn e.onnx e.param e.bin
# 模型优化为fp16
./tools/onnxoptimize e.param e.bin eopt.param eopt.bin 65536

 

2、添加 YOLOv5-Lite.cpp代码

// Tencent is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// Unless required by applicable law or agreed to in writing, software distributed
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
// specific language governing permissions and limitations under the License.
 
#include "layer.h"
#include "net.h"
 
#if defined(USE_NCNN_SIMPLEOCV)
#include "simpleocv.h"
#else
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#endif
#include <float.h>
#include <stdio.h>
#include <vector>
#include <sys/time.h>
 
#include <iostream>  
#include <chrono>  
#include <opencv2/opencv.hpp>  
 
using namespace std;  
using namespace cv;  
using namespace std::chrono;  
 
// 0 : FP16
// 1 : INT8
#define USE_INT8 0
 
// 0 : Image
// 1 : Camera
#define USE_CAMERA 1
 
struct Object
{
    cv::Rect_<float> rect;
    int label;
    float prob;
};
 
static inline float intersection_area(const Object& a, const Object& b)
{
    cv::Rect_<float> inter = a.rect & b.rect;
    return inter.area();
}
 
static void qsort_descent_inplace(std::vector<Object>& faceobjects, int left, int right)
{
    int i = left;
    int j = right;
    float p = faceobjects[(left + right) / 2].prob;
 
    while (i <= j)
    {
        while (faceobjects[i].prob > p)
            i++;
 
        while (faceobjects[j].prob < p)
            j--;
 
        if (i <= j)
        {
            // swap
            std::swap(faceobjects[i], faceobjects[j]);
 
            i++;
            j--;
        }
    }
 
    #pragma omp parallel sections
    {
        #pragma omp section
        {
            if (left < j) qsort_descent_inplace(faceobjects, left, j);
        }
        #pragma omp section
        {
            if (i < right) qsort_descent_inplace(faceobjects, i, right);
        }
    }
}
 
static void qsort_descent_inplace(std::vector<Object>& faceobjects)
{
    if (faceobjects.empty())
        return;
 
    qsort_descent_inplace(faceobjects, 0, faceobjects.size() - 1);
}
 
static void nms_sorted_bboxes(const std::vector<Object>& faceobjects, std::vector<int>& picked, float nms_threshold)
{
    picked.clear();
 
    const int n = faceobjects.size();
 
    std::vector<float> areas(n);
    for (int i = 0; i < n; i++)
    {
        areas[i] = faceobjects[i].rect.area();
    }
 
    for (int i = 0; i < n; i++)
    {
        const Object& a = faceobjects[i];
 
        int keep = 1;
        for (int j = 0; j < (int)picked.size(); j++)
        {
            const Object& b = faceobjects[picked[j]];
 
            // intersection over union
            float inter_area = intersection_area(a, b);
            float union_area = areas[i] + areas[picked[j]] - inter_area;
            // float IoU = inter_area / union_area
            if (inter_area / union_area > nms_threshold)
                keep = 0;
        }
 
        if (keep)
            picked.push_back(i);
    }
}
 
static inline float sigmoid(float x)
{
    return static_cast<float>(1.f / (1.f + exp(-x)));
}
 
// unsigmoid
static inline float unsigmoid(float y) {
    return static_cast<float>(-1.0 * (log((1.0 / y) - 1.0)));
}
 
static void generate_proposals(const ncnn::Mat &anchors, int stride, const ncnn::Mat &in_pad,
                               const ncnn::Mat &feat_blob, float prob_threshold,
                               std::vector <Object> &objects) {
    const int num_grid = feat_blob.h;
    float unsig_pro = 0;
    if (prob_threshold > 0.6)
        unsig_pro = unsigmoid(prob_threshold);
 
    int num_grid_x;
    int num_grid_y;
    if (in_pad.w > in_pad.h) {
        num_grid_x = in_pad.w / stride;
        num_grid_y = num_grid / num_grid_x;
    } else {
        num_grid_y = in_pad.h / stride;
        num_grid_x = num_grid / num_grid_y;
    }
 
    const int num_class = feat_blob.w - 5;
 
    const int num_anchors = anchors.w / 2;
 
    for (int q = 0; q < num_anchors; q++) {
        const float anchor_w = anchors[q * 2];
        const float anchor_h = anchors[q * 2 + 1];
 
        const ncnn::Mat feat = feat_blob.channel(q);
 
        for (int i = 0; i < num_grid_y; i++) {
            for (int j = 0; j < num_grid_x; j++) {
                const float *featptr = feat.row(i * num_grid_x + j);
 
                // find class index with max class score
                int class_index = 0;
                float class_score = -FLT_MAX;
                float box_score = featptr[4];
                if (prob_threshold > 0.6) {
                    // while prob_threshold > 0.6, unsigmoid better than sigmoid
                    if (box_score > unsig_pro) {
                        for (int k = 0; k < num_class; k++) {
                            float score = featptr[5 + k];
                            if (score > class_score) {
                                class_index = k;
                                class_score = score;
                            }
                        }
 
                        float confidence = sigmoid(box_score) * sigmoid(class_score);
 
                        if (confidence >= prob_threshold) {
 
                            float dx = sigmoid(featptr[0]);
                            float dy = sigmoid(featptr[1]);
                            float dw = sigmoid(featptr[2]);
                            float dh = sigmoid(featptr[3]);
 
                            float pb_cx = (dx * 2.f - 0.5f + j) * stride;
                            float pb_cy = (dy * 2.f - 0.5f + i) * stride;
 
                            float pb_w = pow(dw * 2.f, 2) * anchor_w;
                            float pb_h = pow(dh * 2.f, 2) * anchor_h;
 
                            float x0 = pb_cx - pb_w * 0.5f;
                            float y0 = pb_cy - pb_h * 0.5f;
                            float x1 = pb_cx + pb_w * 0.5f;
                            float y1 = pb_cy + pb_h * 0.5f;
 
                            Object obj;
                            obj.rect.x = x0;
                            obj.rect.y = y0;
                            obj.rect.width = x1 - x0;
                            obj.rect.height = y1 - y0;
                            obj.label = class_index;
                            obj.prob = confidence;
 
                            objects.push_back(obj);
                        }
                    } else {
                        for (int k = 0; k < num_class; k++) {
                            float score = featptr[5 + k];
                            if (score > class_score) {
                                class_index = k;
                                class_score = score;
                            }
                        }
                        float confidence = sigmoid(box_score) * sigmoid(class_score);
 
                        if (confidence >= prob_threshold) {
                            float dx = sigmoid(featptr[0]);
                            float dy = sigmoid(featptr[1]);
                            float dw = sigmoid(featptr[2]);
                            float dh = sigmoid(featptr[3]);
 
                            float pb_cx = (dx * 2.f - 0.5f + j) * stride;
                            float pb_cy = (dy * 2.f - 0.5f + i) * stride;
 
                            float pb_w = pow(dw * 2.f, 2) * anchor_w;
                            float pb_h = pow(dh * 2.f, 2) * anchor_h;
 
                            float x0 = pb_cx - pb_w * 0.5f;
                            float y0 = pb_cy - pb_h * 0.5f;
                            float x1 = pb_cx + pb_w * 0.5f;
                            float y1 = pb_cy + pb_h * 0.5f;
 
                            Object obj;
                            obj.rect.x = x0;
                            obj.rect.y = y0;
                            obj.rect.width = x1 - x0;
                            obj.rect.height = y1 - y0;
                            obj.label = class_index;
                            obj.prob = confidence;
 
                            objects.push_back(obj);
                        }
                    }
                }
            }
        }
    }
}
 
static int detect_yolov5(const cv::Mat& bgr, std::vector<Object>& objects)
{
    ncnn::Net yolov5;
 
#if USE_INT8
    yolov5.opt.use_int8_inference=true;
#else
    yolov5.opt.use_vulkan_compute = true;
    yolov5.opt.use_bf16_storage = true;
#endif
 
    // original pretrained model from https://github.com/ultralytics/yolov5
    // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models
 
#if USE_INT8
    yolov5.load_param("/home/pi/ncnn/build/e.param");
    yolov5.load_model("/home/pi/ncnn/build/e.bin");
#else
    yolov5.load_param("/home/pi/ncnn/build/eopt.param");
    yolov5.load_model("/home/pi/ncnn/build/eopt.bin");
#endif
 
    const int target_size = 320;
    const float prob_threshold = 0.60f;
    const float nms_threshold = 0.60f;
 
    int img_w = bgr.cols;
    int img_h = bgr.rows;
 
    // letterbox pad to multiple of 32
    int w = img_w;
    int h = img_h;
    float scale = 1.f;
    if (w > h)
    {
        scale = (float)target_size / w;
        w = target_size;
        h = h * scale;
    }
    else
    {
        scale = (float)target_size / h;
        h = target_size;
        w = w * scale;
    }
 
    ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);
 
    // pad to target_size rectangle
    // yolov5/utils/datasets.py letterbox
    int wpad = (w + 31) / 32 * 32 - w;
    int hpad = (h + 31) / 32 * 32 - h;
    ncnn::Mat in_pad;
    ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);
 
    const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};
    in_pad.substract_mean_normalize(0, norm_vals);
 
    ncnn::Extractor ex = yolov5.create_extractor();
 
    ex.input("images", in_pad);
 
    std::vector<Object> proposals;
 
    // stride 8
    {
        ncnn::Mat out;
        ex.extract("onnx::Sigmoid_647", out);
 
        ncnn::Mat anchors(6);
        anchors[0] = 10.f;
        anchors[1] = 13.f;
        anchors[2] = 16.f;
        anchors[3] = 30.f;
        anchors[4] = 33.f;
        anchors[5] = 23.f;
 
        std::vector<Object> objects8;
        generate_proposals(anchors, 8, in_pad, out, prob_threshold, objects8);
 
        proposals.insert(proposals.end(), objects8.begin(), objects8.end());
    }
    // stride 16
    {
        ncnn::Mat out;
        ex.extract("onnx::Sigmoid_669", out);
 
 
        ncnn::Mat anchors(6);
        anchors[0] = 30.f;
        anchors[1] = 61.f;
        anchors[2] = 62.f;
        anchors[3] = 45.f;
        anchors[4] = 59.f;
        anchors[5] = 119.f;
 
        std::vector<Object> objects16;
        generate_proposals(anchors, 16, in_pad, out, prob_threshold, objects16);
 
        proposals.insert(proposals.end(), objects16.begin(), objects16.end());
    }
    // stride 32
    {
        ncnn::Mat out;
        ex.extract("onnx::Sigmoid_691", out);
 
 
        ncnn::Mat anchors(6);
        anchors[0] = 116.f;
        anchors[1] = 90.f;
        anchors[2] = 156.f;
        anchors[3] = 198.f;
        anchors[4] = 373.f;
        anchors[5] = 326.f;
 
        std::vector<Object> objects32;
        generate_proposals(anchors, 32, in_pad, out, prob_threshold, objects32);
 
        proposals.insert(proposals.end(), objects32.begin(), objects32.end());
    }
 
    // sort all proposals by score from highest to lowest
    qsort_descent_inplace(proposals);
 
    // apply nms with nms_threshold
    std::vector<int> picked;
    nms_sorted_bboxes(proposals, picked, nms_threshold);
 
    int count = picked.size();
 
    objects.resize(count);
    for (int i = 0; i < count; i++)
    {
        objects[i] = proposals[picked[i]];
 
        // adjust offset to original unpadded
        float x0 = (objects[i].rect.x - (wpad / 2)) / scale;
        float y0 = (objects[i].rect.y - (hpad / 2)) / scale;
        float x1 = (objects[i].rect.x + objects[i].rect.width - (wpad / 2)) / scale;
        float y1 = (objects[i].rect.y + objects[i].rect.height - (hpad / 2)) / scale;
 
        // clip
        x0 = std::max(std::min(x0, (float)(img_w - 1)), 0.f);
        y0 = std::max(std::min(y0, (float)(img_h - 1)), 0.f);
        x1 = std::max(std::min(x1, (float)(img_w - 1)), 0.f);
        y1 = std::max(std::min(y1, (float)(img_h - 1)), 0.f);
 
        objects[i].rect.x = x0;
        objects[i].rect.y = y0;
        objects[i].rect.width = x1 - x0;
        objects[i].rect.height = y1 - y0;
    }
 
    return 0;
}
 
static void draw_objects(const cv::Mat& bgr, const std::vector<Object>& objects)
{
    static const char* class_names[] = {
        "drug","glue","prime"
    };
 
    cv::Mat image = bgr.clone();
 
    for (size_t i = 0; i < objects.size(); i++)
    {
        const Object& obj = objects[i];
 
        fprintf(stderr, "%d = %.5f at %.2f %.2f %.2f x %.2f\n", obj.label, obj.prob,
                obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);
 
        cv::rectangle(image, obj.rect, cv::Scalar(0, 255, 0));
 
        char text[256];
        sprintf(text, "%s %.1f%%", class_names[obj.label], obj.prob * 100);
 
        int baseLine = 0;
        cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);
 
        int x = obj.rect.x;
        int y = obj.rect.y - label_size.height - baseLine;
        if (y < 0)
            y = 0;
        if (x + label_size.width > image.cols)
            x = image.cols - label_size.width;
 
        cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),
                      cv::Scalar(255, 255, 255), -1);
 
        cv::putText(image, text, cv::Point(x, y + label_size.height),
                    cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0)); 
                    
       // cv::putText(image, to_string(fps), cv::Point(100, 100),             //FPS
                    //cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0)); 
        
    }
#if USE_CAMERA
    imshow("camera", image);
    cv::waitKey(1);
#else
    cv::imwrite("result.jpg", image);
#endif
}
 
#if USE_CAMERA
int main(int argc, char** argv)
{
    cv::VideoCapture capture;
    capture.open(0);  //修改这个参数可以选择打开想要用的摄像头
 
    cv::Mat frame;
    
    //111
    int FPS = 0;  
    int total_frames = 0;  
    high_resolution_clock::time_point t1, t2; 
    
    
    while (true)
    {
        capture >> frame;
        cv::Mat m = frame;
        cv::Mat f = frame;
 
        std::vector<Object> objects;
        
        auto start_time = std::chrono::high_resolution_clock::now();  // 记录开始时间 
        
        detect_yolov5(frame, objects);
 
        auto end_time = std::chrono::high_resolution_clock::now();  // 记录结束时间  
        auto duration = std::chrono::duration_cast<std::chrono::milliseconds>(end_time - start_time);  // 计算执行时间 
        
        float fps = (float)(1000)/duration.count();
 
        draw_objects(m, objects);
        cout << "FPS: " << fps << endl;  
        
        //int fps = 1000/duration.count();
        //int x = m.cols-50;
        //int y = m.rows-50;
        //cv::putText(f, to_string(fps), cv::Point(100, 100), cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0)); 
        
        //if (cv::waitKey(30) >= 0)
            //break;
            
    }
}
#else
int main(int argc, char** argv)
{
    if (argc != 2)
    {
        fprintf(stderr, "Usage: %s [imagepath]\n", argv[0]);
        return -1;
    }
 
    const char* imagepath = argv[1];
 
    struct timespec begin, end;
    long time;
    clock_gettime(CLOCK_MONOTONIC, &begin);
 
    cv::Mat m = cv::imread(imagepath, 1);
    if (m.empty())
    {
        fprintf(stderr, "cv::imread %s failed\n", imagepath);
        return -1;
    }
 
    std::vector<Object> objects;
    detect_yolov5(m, objects);
 
    clock_gettime(CLOCK_MONOTONIC, &end);
    time = (end.tv_sec - begin.tv_sec) + (end.tv_nsec - begin.tv_nsec);
    printf(">> Time : %lf ms\n", (double)time/1000000);
 
    draw_objects(m, objects);
 
    return 0;
}
#endif

  

上述代码其实作者在 Github 中就已经提供了,我们仅需要对其进行稍微修改以下即可!

    补充说明:作者这里仅提供 FP16 格式的模型使用方法,INT8 类型的会稍微复杂点。需要用到 INT8 格式的模型的朋友可以去其他博客学习一下!

3、修改 eopt.param

生成param文件时如果遇到Squeeze not supported yet!等提示,解决方法为使用onnxsimplifier优化onnx模型在转换为param

打开eopt.param,将 Permute 上方的 Reshape 修改为0 = -1,此步是为了能够动态输入:

    部分博主是将 eopt.param 内的所有 Reshape 全部改为0 = -1,这点作者没有这样,作者的改动方式是参考YOLOv5-Lite作者本人的方法!

4、修改yolov5_lite.cpp

这是由于未修改cpp中ex.extract()和permute保持一致

打开v5lite-e.yaml:

根据anchors修改cpp内容,需要保持一致

5、修改CMakeLists.txt

打开examples/CMakeLists.txt ,添加ncnn_add_example(yolov5-lite) ,注意和文件名保持一致

完成后使用cmake编译;

cd ncnn/build
cmake ..
make

6、摄像头运行BUG

    部分读者朋友在运行YOLOv5-Lite的NCNN模型的时候,可能会出现摄像头闪退的情况,就是程序运行后摄像头仅运行了很短的一段时间!针对该情况,做如下操作!

将上述源码注释掉即可,后续暂停程序可以通过Ctrl+C在终端终止程序! 


四、运行效果

作者测试下来最终运行的 FPS 仅在 3.7 左右,感觉是存在问题的。常规情况下,NCNN 模型的速度是快于 ONNX 模型的。所以,后续作者将通过压力测试去查看一下推理过程各个算子的消耗时间!