【考研复习：数据结构】查找（不含代码篇）

前言：

1、此篇是基于博主对严蔚敏版教材《数据结构》、王道书《数据结构》和在网上相关资料的查询，对第七章 “ 查找 ” 的学习总结。

2、查找这一章含代码（C++）会写在另一篇。写好后再放链接。

3、博主比较喜欢用表格使思路稍微清晰一些，还有一些博主自己怕记乱的内容便标为注意点，只为了方便复习和记忆。

4、有些内容可能不是很全，基于时间和精力，只总结考试中可能会出现的内容。算是纯理论，请结合书中例题或网上例题来理解，应该会理解记忆更深。

5、欢迎各位大神帮忙纠正和补充。

一、基本概念：

（一）静态查找、动态查找

基本概念内容
静态查找顺序查找、折半查找（二分查找）、分块查找（索引顺序查找）、散列查找（哈希表）、斐波那契查找
动态查找散列查找（哈希表）、各种树（二叉排序树、AVL树、B / B+ 树、红黑树等）

（二）结构

结构内容
线性结构顺序查找、折半查找（二分查找）、分块查找（索引顺序查找）
树形结构二叉排序树、二叉平衡树、B 树、B+ 树
散列结构散列表（哈希表）

注意：

1、顺序查找、折半查找、分块查找，查找表为线性表，其中折半查找效率较高。

2、线性表的查找更适合静态查找。

3、二叉排序树，又称二叉搜索树（BST）、二叉查找树。

4、AVL树：平衡二叉树（二叉排序树的一种特殊类型、4种平衡调整方法）

定义：

（1）空树；

（2）具备两个条件的二叉排序树：a、左子树和右子树的深度之差的绝对值不超过1；

                                                        b、左子树和右子树也是平衡二叉树。

5、B 树（B-tree）：多路平衡查找树。（外存文件系统中常用的动态索引技术）【B 树的阶m（即 m 路，m叉）：所有结点的孩子个数的最大值。】

6、B+ 树：（B树的变形，更适合用于文件索引系统）

7、B 树与 B+ 树区分：（ m - 1 = n 个关键字，非空树，各自的优点见 “ 二、查找方法的区分 ”）

树根结点每个结点（非根内部结点）叶结点非叶结点树高 h
m 阶 B+ 树每个分支结点最多有 m 棵子树。2≤n≤m2≤n≤m⌈m2⌉≤n≤m⌈m2⌉≤n≤m

1、包含全部关键字，即非叶结点中的关键字也会出现在叶结点中。

2、叶结点都出现在同一层次，包含信息。

3、叶结点中将关键字按小到大顺序排列。

4、叶结点之间通过指针链接。

仅起索引作用，其中每个索引项只含有对应子树的最大关键字和指向该子树的指针，不含有该关键字对应记录的存储地址。非叶根结点至少 2 棵子树，其它分支结点至少  ⌈m2⌉  根子树
m 阶 B 树每个结点至多 m 棵子树，至多含 m - 1 个关键字。1≤n≤m−1⌈m2⌉−1≤n≤m−1

1、（最外层内部结点）包含的关键字和其他结点包含的关键字是不重复的。 

2、叶结点都出现在同一层次，并且不带任何信息。

3、各结点中的关键字均升序或降序排列。 

4、不要求叶结点之间通过指针链接。

每个非终端结点中包含有n个关键字信息。

除根结点外的所有非叶结点：

1、至少  ⌈m2⌉  根子树，即至少含有⌈m2⌉−1 个关键字；

2、至多 m 棵子树，至多含 m - 1 个关键字。

logm(n+1)≤h≤[log⌈m2⌉(n+12)]+1

解释：

（1）根据图判断是 B 树还是 B+ 树：

B 树具有 n 个关键字的结点含 n + 1 棵子树
B+ 树具有 n 个关键字的结点只含 n 棵子树，即每个关键字对应一棵子树。

（2）叶结点（最外层内部结点）包含的关键字和其他结点包含的关键字是不重复的。 即任何一个关键字出现且只出现在一个结点中；

（3）B 树：所有叶子结点都出现在同一层，叶子结点不包含任何关键字信息(可以看做是外部接点或查询失败的接点，实际上这些结点不存在，指向这些结点的指针都为null)；（读者反馈@冷岳：这里有错，叶子节点只是没有孩子和指向孩子的指针，这些节点也存在，也有元素。@研究者July：其实，关键是把什么当做叶子结点，因为如红黑树中，每一个NULL指针即当做叶子结点，只是没画出来而已）。

注：（3）摘自CSDN博主「v_JULY_v」的原创文章，原文链接：http://iyenn.com/rec/1653211.html

（4）B 树、B+ 树详解： https://www.cnblogs.com/lianzhilei/p/11250589.html 

（以上两个网址里都可以看插入、删除操作，里面有图再结合上表内容易理解记忆）

（5）具有 n 个关键字的 m 阶 B 树，应有 n + 1 个叶结点。

（6）高度为 h 的 m = 3 阶 B 树，至少有 (m−1)h−1 个结点，至多有 mh−1 个结点。

（7）含有 n 个非叶结点的 m 阶 B 树中至少包含  (n−1)∗(⌈m2⌉−1)+1  个关键字。

8、散列表：

（1）散列表属于线性结构，但不同于线性表的查找。顺序查找、二分查找、分块查找是以关键字比较为基础进行查找，而散列表是通过一种散列函数把记录的关键字和它在表中的位置建立起对应关系，并在存储记录发生冲突时采用专门的冲突的方法。

（2）散列表的平均查找长度与记录总数无关，是通过调节装填因子，把平均查找长度控制在所需的范围内。

（3）装填因子   \alpha =\frac{n}{m} （表中记录数 n ，散列表长度 m）。装填因子越大（越 “ 满 ”），发生冲突的可能性越大。

（4）开放地址法和链地址法的比较（这两种都是处理冲突的方法）

比较项目开放地址法链地址法
空间无指针域，存储效率较高附加指针域，存储效率较低
时间查找有二次聚集现象，查找效率低无二次聚集现象，查找效率高
插入、删除不易实现易于实现
适用情况表的大小固定，适于表长无变化的情况结点动态生成，适于表长经常变化的情况

（5）散列函数的 “ 好坏 ” 首先影响出现冲突的频繁程度。但一般情况下认为：凡是 “ 均匀的 ” 散列函数，对同一组随机的关键字，产生冲突的可能性相同，此时影响 ASL 的因素只有两个：处理冲突的方法和装填因子。

（6）产生堆积现象，即产生冲突，它对存储效率、散列函数和装填因子均不会有影响，而 ASL 会因为堆积现象而增大。 

二、查找方法的区分

查找方法                                         平均查找长度时间复杂度空间复杂度优点缺点备注
查找成功查找失败
顺序查找一般线性表

ASL=∑ni=1Pi(n−i+1)

Pi=1n,ASL=n+12

ASL = n + 1O(n)O(1)

1、对数据元素的存储没有要求，顺序存储或链式存储都可。

2、对表中记录的有序性无要求。

注意:

（1）对线性的链表只能进行顺序查找。

（2）有序线性表的顺序查找可链式存储。（区分折半查找）

当 n 比较大时，ASL较大，效率低
有序表ASL=n2+nn+1
二分查找有序表（仅适用于顺序存储结构）

ASL=1n(1∗1+2∗2+....+h∗2h−1)

树高 h=⌈log2(n+1)⌉

也可为 h=⌊log2n⌋+1

树的高度，也是查找成功和查找失败的最多次数 

ASL=

1n+1[(h−1)∗(n+1−n0)+h∗n0∗2]

n0 是叶结点数，h 是判定树的树高

n0 * 2是方形结点，查找不成功的数量

O(log2n)比较次数少，查找效率在大部分情况下比顺序查找高

1、不适用于数据元素经常变化动的线性表；

2、不适用于链式存储结构。

二分查找的判定树是平衡二叉树
分块查找有动态结构、适于快速查找

将长度为 n 的查找表均匀地分为 b 块，每块有 s 个记录，

等概率时，平均查找长度：

块内和索引表都用顺序查找：ASL=b+12+s+12=s2+2s+n2s

当 s=√n （理想块长）时，平均查找长度取最小值ASL=√n+1

O(log2n)由于块内无序，插入和删除比较容易，无需进行大量移动。要增加一个索引表的存储空间并对初始索引 表进行排序运算

1、块内无序，块间有序；

2、若对索引项和索引块内部都采用折半查找，则查找效率最高。

3、查找效率介于顺序查找和二分查找之间。

块内用顺序查找，索引表用二分查找：ASL=⌈log2(b+1)⌉+s+12
当 s=√n （理想块长）时，块内和索引表都用二分查找：\dpi150ASL=⌈log2(s+1)⌉+⌈log2(s+1)⌉
二叉排序树适用于经常进行插入、删除和查找运算的表。（ASL与树的形态有关）最差情形：关键字有序，形成单支树（树的深度为 n ）：ASL=n+12O(n)

1、数据结构采用树的二叉链表表示，就维护表的有序性而言，二叉排序树比二分查找更有效，因为无需移动记录，只需修改指针即可完成对结点的插入和删除操作。

2、可快速检索。

1、顺序存储可能会浪费空间(在非完全二叉树的时候)

2、最坏的情况——关键字有序(大到小或小到大排序)，此时，二叉排序树就退化成了普通链表，其检索效率就会很差。 

中序遍历，结点值递增的有序序列（AVL树可参考二叉排序树）
最好情形：和折半查找的判定树相似，ASL 和 log2n 成正比O(log2n)
B 树ASL 请见二叉排序树若经常访问的数据离根结点很近，而B树的非叶子结点本身存有关键字其数据的地址，所以这种数据检索时效率会比 B+ 树高。 不支持顺序查找，支持多路查找（随机查找）。
树高 h:     logm(n+1)≤h≤[log⌈m2⌉(n+12)]+1
B+ 树ASL 请见二叉排序树

1、比 B 树层级更少、查询速度更快、更复稳定；

2、具备排序功能，即所有叶子结点是一个有序链表；

3、全节点遍历更快，B+ 树遍历整棵树只用遍历所有叶子结点； 

1、支持顺序查找、随机查找

2、在 B+ 树中查找，无论查找成功与否，每次查找都是一条从根结点到叶结点的路径。

散列查找

（查找概率相等时）

ASL（成功） = 查找成功所有记录所需的比较次数之和 / 散列表记录的个数

ASL（失败） = 查找失败所有的散列函数取值比较次数之和 / 散列函数取值的个数

（此处的 ASL 建议看严蔚敏版教材《数据结构》来理解） 

O(1)查找效率取决于三个因素：散列表、处理冲突的方法、装填因子

三、题目解答（放松下，题目源自牛客网，都是博主做过的觉得还可以学习积累的，可以结合上述来做）

1、集合中任何两个元素都可以比较大小，但比较不满足传递性，可以通过hash索引使得在集合中查找元素的时间复杂度降到O(1)。

解：哈希所以可以认为是单纯的数值计算，并没有大小比较操作。

2、在长度为n的顺序表中查找一个元素，假设需要查找的元素一定在表中，并且元素出现在表中每个位置上的可能性是相同的，则在平均情况下需要比较的次数为 (n + 1) / 2。

解：因为每一个位置出现的几率相同故为 p = 1 / n

n 的位置查找的总次数为 A = n * (1 + n) / 2

平均下来的次数 = p * A = (n + 1) / 2 = 平均查找长度

顺序查找的平均时间 = n / 2

3、若有一个顺序有序表A[1:18] 中有18个元素，现进行二分查找，则查找 A［3］的比较序列的下标依次为9, 4, 2, 3。

解：假设[1:18]是单调递增，设A[3] = x。

先查找[1:18]的中间值mid =（1 + 18）/ 2 = 9.5，向下取整为9,

A[9] > x，于是范围变成了中值左边的取值，左右值变成了[1:9-1]即为[1:8]

mid =（1 + 8）/ 2 = 4.5，取值4

A[4] > x，于是范围变成了中值左边的取值，左右值变成了[1:4-1]即为[1:3]

mid = (1 + 3) / 2 = 2

A[2] < x，于是范围变成了中值右边的取值，范围变成[2+1:3]即为[3:3]

A[3] = X

由上：查找 A［3］的比较序列的下标依次为（9,4,2,3）。

4、一个文件包含了200个记录，若采用分块查找法，每块长度为4，则平均查找长度为28。

解：要看 200 / 4 = 50个块之间是否是有序的。

要是有序，可以二分查找，然后确定在那一个块中((n+1)log2(n + 1)) / n - 1 ，然后在这个块中查找。

要是无序，则顺序查找，平均查找长度就是 （1 + 50）/ 2 = 25.5，然后块中查找(1 + 4) / 2 = 2.5  总共28。

5、已知一个线性表（38，25，74，63，52，48），假定采用散列函数h（key) = key%7 计算散列地址，并散列存储在散列表A【0....6】中，若采用线性探测方法解决冲突，则在该散列表上进行等概率成功查找的平均查找长度为  2  。

解：依次进行取模运算求出哈希地址：

A

0

1

2

3

4

5

6

记录

63

48

38

25

74

52

查找次数

1

3

1

1

2

4

平均查找长度 = 总的查找次数 / 元素数

总的查找次数：

38 % 7 = 3（第1次出现3，无冲突，放在位置3，查找次数为1）

25 % 7 = 4（第1次出现4，无冲突，放在位置4，查找次数为1）

74 % 7 = 4（第2次出现4，有冲突，放在位置5，查找次数为2）

63 % 7 = 0（第1次出现0，无冲突，放在位置0，查找次数为1）

52 % 7 = 3（第2次出现3，有冲突，发现冲突3,4,5，故只能放到6，查找次数为4）

48 % 7 = 6 （第1次出现6，有冲突，发现冲突6,0，故只能放到1，查找次数为3）

结果：（1 + 1 + 2 + 1 + 4 + 3）/ 6 = 2

6、一个线性序列（30，14，40，63，22，5），假定采用散列函数Hash(key) = key % 7来计算散列地址，将其散列存储在A[0~6]中，采用链地址法解决冲突。若查找每个元素的概率相同，则查找成功的平均查找长度是 4 / 3。

解：30 % 7 = 2 查找一次

14 % 7 = 0 查找一次

40 % 7 = 5 查找一次

63 % 7 = 0 查找两次

22 % 7 = 1 查找一次

5 % 7 = 5 查找两次

查找成功的平均查找长度 =（1 + 1 + 1 + 2 + 1 + 2）/6 = 4/3

id="article_content" class="article_content clearfix" style="height: 2000px; overflow: hidden;"> id="content_views" class="htmledit_views">

目录

一、环境准备

二、数据准备

1、指定格式存放数据集

2.在data目录下新建voc_label.py文件;里面存放代码，里面classes需要改成自己的类别

3、修改数据加载配置文件

三、模型训练/验证/预测/导出

（1)模型训练

多卡训练

（2）模型预测

（3）模型导出

四、YOLOv8部署到RK3588

PT转ONNX

五、PC端安装RKNNtoolkit2的及其环境

ONNX转RKNN

1、创建环境

2、这里需要用到rknn-toolkit2文件

3、安装rknn-toolkit2环境

4、现在要进行模型转换

六、模型部署到RK3588上

RKNPU SDK的版本

yolov8官网地址：yolov8

 官网文档地址

          rknn_zoo                     

        

           RKNPU2_SDK     模型文件

                 

          RKNN Model Zoo   RKNPU SDK的版本

          rknn_model_zoo 

RKNN Model Zoo基于 RKNPU SDK 工具链开发, 提供了目前主流算法的部署例程. 例程包含导出RKNN模型, 使用 Python API, CAPI 推理 RKNN 模型的流程.

  • 支持 RK3562, RK3566, RK3568, RK3588, RK3576 平台。
  • 部分支持RV1103, RV1106
  • 支持 RK1808, RV1109, RV1126 平台。

一、环境准备

在这之前，需要先准备主机的环境，环境如下：

  • Ubuntu22.04
  • cuda12.4

三个文件的链接放在这：ultralytics_yolov8，rknn_model_zoo ，rknn-toolkit2

进入自己带pytorch的虚拟环境，与之前的yolo系列安装都不太一样，yolov8仅需要安装ultralytics这一个库：

 pip install ultralytics -i https://pypi.tuna.tsinghua.edu.cn/simple

1、创建虚拟环境

conda create -n yolo8 python==3.10 -y class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

conda activate yolo8
 class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

2、下载ultralytics库

 pip install ultralytics -i https://pypi.tuna.tsinghua.edu.cn/simple class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

pip install -r requirements.txt  -i https://pypi.tuna.tsinghua.edu.cn/simple class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

另外一种方法：需要先克隆git仓库，再进行安装；二者取其一即可

  1.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="1">
class="hljs-ln-code"> class="hljs-ln-line">git clone https://github.com/ultralytics/ultralytics
  • class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="2"> class="hljs-ln-code"> class="hljs-ln-line">cd ultralytics
  • class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="3"> class="hljs-ln-code"> class="hljs-ln-line">pip install -e .
    •  class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    二、数据准备

    1、指定格式存放数据集

    在yolov8/data目录下新建Annotations, images, ImageSets, labels 四个文件夹
    images目录下存放数据集的图片文件
    Annotations目录下存放图片的xml文件（labelImg标注）

    这里直接参考我的另一篇文章

    （链接：yolov7训练数据集详细流程bike-car-person_yolov7模型训练详细过程-CSDN博客）

    1）新建split_train_val.py文件

    运行之后会在ImageSets/Main下生成四个.txt文件

    2.在data目录下新建voc_label.py文件;里面存放代码，里面classes需要改成自己的类别

    运行之后生成3个txt文件

    3、修改数据加载配置文件

    进入data/文件夹，新建coco.yaml，内容如下，注意txt需要使用绝对路径

    1.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="1"> class="hljs-ln-code"> class="hljs-ln-line"># COCO 2017 dataset http://cocodataset.org
    2.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="2"> class="hljs-ln-code"> class="hljs-ln-line"># Train command: python train.py --data coco.yaml
    3.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="3"> class="hljs-ln-code"> class="hljs-ln-line"># Default dataset location is next to /yolov5:
    4.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="4"> class="hljs-ln-code"> class="hljs-ln-line">#   /parent_folder
    5.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="5"> class="hljs-ln-code"> class="hljs-ln-line">#     /coco
    6.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="6"> class="hljs-ln-code"> class="hljs-ln-line">#     /yolov5
    7.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="7"> class="hljs-ln-code"> class="hljs-ln-line"> 
    8.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="8"> class="hljs-ln-code"> class="hljs-ln-line"> 
    9.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="9"> class="hljs-ln-code"> class="hljs-ln-line"># download command/URL (optional)
    10.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="10"> class="hljs-ln-code"> class="hljs-ln-line"># download: bash data/scripts/get_coco.sh
    11.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="11"> class="hljs-ln-code"> class="hljs-ln-line"> 
    12.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="12"> class="hljs-ln-code"> class="hljs-ln-line"># train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/]
    13.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="13"> class="hljs-ln-code"> class="hljs-ln-line"># train: ../coco/train2017.txt  # 118287 images
    14.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="14"> class="hljs-ln-code"> class="hljs-ln-line"># val: ../coco/val2017.txt  # 5000 images
    15.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="15"> class="hljs-ln-code"> class="hljs-ln-line"># test: ../coco/test-dev2017.txt  # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794
    16.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="16"> class="hljs-ln-code"> class="hljs-ln-line"> 
    17.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="17"> class="hljs-ln-code"> class="hljs-ln-line">train: /home/sxj/yolov8/data/train.txt  # 118287 images
    18.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="18"> class="hljs-ln-code"> class="hljs-ln-line">val: /home/sxj/yolov8/data/val.txt  # 5000 images
    19.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="19"> class="hljs-ln-code"> class="hljs-ln-line">test: /home/sxj/yolov8/data/test.txt  # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794
    20.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="20"> class="hljs-ln-code"> class="hljs-ln-line"> 
    21.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="21"> class="hljs-ln-code"> class="hljs-ln-line"># number of classes
    22.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="22"> class="hljs-ln-code"> class="hljs-ln-line">nc: 3
    23.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="23"> class="hljs-ln-code"> class="hljs-ln-line"> 
    24.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="24"> class="hljs-ln-code"> class="hljs-ln-line"># class names
    25.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="25"> class="hljs-ln-code"> class="hljs-ln-line">names: [ 'bike','carsgraz','person' ]
    26.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="26"> class="hljs-ln-code"> class="hljs-ln-line"> 
    27.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="27"> class="hljs-ln-code"> class="hljs-ln-line"># Print classes
    28.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="28"> class="hljs-ln-code"> class="hljs-ln-line"># with open('data/coco.yaml') as f:
    29.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="29"> class="hljs-ln-code"> class="hljs-ln-line">#   d = yaml.load(f, Loader=yaml.FullLoader)  # dict
    30.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="30"> class="hljs-ln-code"> class="hljs-ln-line">#   for i, x in enumerate(d['names']):
    31.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="31"> class="hljs-ln-code"> class="hljs-ln-line">#     print(i, x)
     class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    至此数据集的准备已经就绪，索引文件在data目录下的train.txt/val.txt/test.txt

    三、模型训练/验证/预测/导出

    （1)模型训练

    进入虚拟环境，进入yolov8文件夹，终端中输入，开始训练：

    yolo task=detect mode=train model=yolov8n.pt data=data/coco.yaml batch=1 epochs=50 imgsz=640 workers=8 device=0
     

    yolo task=detect mode=train model=yolov8n.pt data=data/coco.yaml batch=1 epochs=50 imgsz=640 workers=8 device=0 class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">
    多卡训练

    yolov8的多卡训练其实很简单，不需要使用繁琐的命令行指令，仅需把device='0,1,2,3’即可，注意一定要加\和引号

    yolo task=detect mode=val model=runs/detect/train3/weights/best.pt data=data/coco.yaml device=0 plots=True
     

    （2）模型预测

    使用如下命令，即可完成对新数据的预测，source需要指定为自己的图像路径，或者摄像头（0）

    yolo task=detect mode=predict model=runs/detect/train/weights/best.pt source=data/images device=0 class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    （3）模型导出

    使用如下命令，即可完成训练模型的导出


     

    yolo task=detect mode=export model=runs/detect/train/weights/best.pt format=onnx class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

    接着就是：

    四、YOLOv8部署到RK3588

    模型训练→转换RKNN→开发板部署

    1、将requirments.txt公示如下，可自创一个txt文件，把内容复制进去再pip install

    1.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="1"> class="hljs-ln-code"> class="hljs-ln-line"># Ultralytics requirements
    2.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="2"> class="hljs-ln-code"> class="hljs-ln-line"># Usage: pip install -r requirements.txt
    3.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="3"> class="hljs-ln-code"> class="hljs-ln-line"> 
    4.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="4"> class="hljs-ln-code"> class="hljs-ln-line"># Base ----------------------------------------
    5.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="5"> class="hljs-ln-code"> class="hljs-ln-line">matplotlib>=3.2.2
    6.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="6"> class="hljs-ln-code"> class="hljs-ln-line">numpy>=1.22.2 # pinned by Snyk to avoid a vulnerability
    7.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="7"> class="hljs-ln-code"> class="hljs-ln-line">opencv-python>=4.6.0
    8.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="8"> class="hljs-ln-code"> class="hljs-ln-line">pillow>=7.1.2
    9.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="9"> class="hljs-ln-code"> class="hljs-ln-line">pyyaml>=5.3.1
    10.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="10"> class="hljs-ln-code"> class="hljs-ln-line">requests>=2.23.0
    11.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="11"> class="hljs-ln-code"> class="hljs-ln-line">scipy>=1.4.1
    12.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="12"> class="hljs-ln-code"> class="hljs-ln-line">torch>=1.7.0
    13.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="13"> class="hljs-ln-code"> class="hljs-ln-line">torchvision>=0.8.1
    14.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="14"> class="hljs-ln-code"> class="hljs-ln-line">tqdm>=4.64.0
    15.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="15"> class="hljs-ln-code"> class="hljs-ln-line"> 
    16.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="16"> class="hljs-ln-code"> class="hljs-ln-line"># Logging -------------------------------------
    17.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="17"> class="hljs-ln-code"> class="hljs-ln-line"># tensorboard>=2.13.0
    18.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="18"> class="hljs-ln-code"> class="hljs-ln-line"># dvclive>=2.12.0
    19.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="19"> class="hljs-ln-code"> class="hljs-ln-line"># clearml
    20.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="20"> class="hljs-ln-code"> class="hljs-ln-line"># comet
    21.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="21"> class="hljs-ln-code"> class="hljs-ln-line"> 
    22.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="22"> class="hljs-ln-code"> class="hljs-ln-line"># Plotting ------------------------------------
    23.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="23"> class="hljs-ln-code"> class="hljs-ln-line">pandas>=1.1.4
    24.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="24"> class="hljs-ln-code"> class="hljs-ln-line">seaborn>=0.11.0
    25.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="25"> class="hljs-ln-code"> class="hljs-ln-line"> 
    26.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="26"> class="hljs-ln-code"> class="hljs-ln-line"># Export --------------------------------------
    27.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="27"> class="hljs-ln-code"> class="hljs-ln-line"># coremltools>=7.0.b1  # CoreML export
    28.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="28"> class="hljs-ln-code"> class="hljs-ln-line"># onnx>=1.12.0  # ONNX export
    29.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="29"> class="hljs-ln-code"> class="hljs-ln-line"># onnxsim>=0.4.1  # ONNX simplifier
    30.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="30"> class="hljs-ln-code"> class="hljs-ln-line"># nvidia-pyindex  # TensorRT export
    31.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="31"> class="hljs-ln-code"> class="hljs-ln-line"># nvidia-tensorrt  # TensorRT export
    32.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="32"> class="hljs-ln-code"> class="hljs-ln-line"># scikit-learn==0.19.2  # CoreML quantization
    33.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="33"> class="hljs-ln-code"> class="hljs-ln-line"># tensorflow>=2.4.1  # TF exports (-cpu, -aarch64, -macos)
    34.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="34"> class="hljs-ln-code"> class="hljs-ln-line"># tflite-support
    35.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="35"> class="hljs-ln-code"> class="hljs-ln-line"># tensorflowjs>=3.9.0  # TF.js export
    36.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="36"> class="hljs-ln-code"> class="hljs-ln-line"># openvino-dev>=2023.0  # OpenVINO export
    37.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="37"> class="hljs-ln-code"> class="hljs-ln-line"> 
    38.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="38"> class="hljs-ln-code"> class="hljs-ln-line"># Extras --------------------------------------
    39.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="39"> class="hljs-ln-code"> class="hljs-ln-line">psutil  # system utilization
    40.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="40"> class="hljs-ln-code"> class="hljs-ln-line">py-cpuinfo  # display CPU info
    41.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="41"> class="hljs-ln-code"> class="hljs-ln-line"># thop>=0.1.1  # FLOPs computation
    42.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="42"> class="hljs-ln-code"> class="hljs-ln-line"># ipython  # interactive notebook
    43.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="43"> class="hljs-ln-code"> class="hljs-ln-line"># albumentations>=1.0.3  # training augmentations
    44.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="44"> class="hljs-ln-code"> class="hljs-ln-line"># pycocotools>=2.0.6  # COCO mAP
    45.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="45"> class="hljs-ln-code"> class="hljs-ln-line"># roboflow
    46.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="46"> class="hljs-ln-code"> class="hljs-ln-line"> 
     class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    pip install -r requirements.txt  -i https://pypi.tuna.tsinghua.edu.cn/simple class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    2、通过命令行进行训练，写一个train.py的脚本，在脚本中输入超参数后，再进行训，在此，建议大家采用脚本训练的方法，并且将脚本放在ultralytics文件夹的同级下

    train.py脚本内容：

    1.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="1"> class="hljs-ln-code"> class="hljs-ln-line">from ultralytics import YOLO
    2.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="2"> class="hljs-ln-code"> class="hljs-ln-line"> 
    3.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="3"> class="hljs-ln-code"> class="hljs-ln-line"> 
    4.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="4"> class="hljs-ln-code"> class="hljs-ln-line"># 加载模型
    5.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="5"> class="hljs-ln-code"> class="hljs-ln-line"># model = YOLO("yolov8n.yaml")  #从头开始构建新模型
    6.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="6"> class="hljs-ln-code"> class="hljs-ln-line"> 
    7.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="7"> class="hljs-ln-code"> class="hljs-ln-line">model = YOLO("yolov8n.pt")
    8.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="8"> class="hljs-ln-code"> class="hljs-ln-line"> 
    9.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="9"> class="hljs-ln-code"> class="hljs-ln-line"># Use the model
    10.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="10"> class="hljs-ln-code"> class="hljs-ln-line">results = model.train(data="/home/sxj/yolov8/data/coco.yaml", epochs=100,batch=1)#训练模型
    11.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="11"> class="hljs-ln-code"> class="hljs-ln-line"> 
    12.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="12"> class="hljs-ln-code"> class="hljs-ln-line"># 将模型导出为 ONNX 格式
    13.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="13"> class="hljs-ln-code"> class="hljs-ln-line">success = model.export(format="onnx")
     class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    3.最重要的一点：

      yolov5s模型激活函数为silu, 此激活函数量化类型为float16, 导致推理过程中使用CPU进行计算, 量化效果较糟。 将激活函数换为relu, 可以在牺牲一点精度的情况下获得巨大性能提升, 目前测试约为80 - 83帧, c++优化后会更高

    要改激活函数为ReLU，我们如果默认不改，则训练得到的PT模型的激活函数为SiLU：

    改激活函数流程如下：

    ctrl + f 替换

    例如：把ultralytics/nn/modules/conv.py中的Conv类进行修改，

    将

    default_act = nn.SiLU() # default activation

    改为：
    default_act = nn.ReLU() # default activation

    PT转ONNX

    在训练得到PT模型后，我们可以用netron软件打开模型，看一下结构

    我们的激活函数是否已经为ReLU了

    在终端执行命令：yolo export model=yolov8relupt.pt format=rknn
    生成如下结果，同时在目录中生成我们想要的ONNX模型

    建一个zhuan.py文件

    1.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="1"> class="hljs-ln-code"> class="hljs-ln-line">'''将pt转化为onnx
    2.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="2"> class="hljs-ln-code"> class="hljs-ln-line">    onnx = 1.16.0
    3.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="3"> class="hljs-ln-code"> class="hljs-ln-line">    opset = 12
    4.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="4"> class="hljs-ln-code"> class="hljs-ln-line">'''
    5.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="5"> class="hljs-ln-code"> class="hljs-ln-line">from ultralytics import YOLO
    6.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="6"> class="hljs-ln-code"> class="hljs-ln-line"> 
    7.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="7"> class="hljs-ln-code"> class="hljs-ln-line">class Yolov8Pt2Onnx:
    8.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="8"> class="hljs-ln-code"> class="hljs-ln-line">    def __init__(self, pt_path) -> None:
    9.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="9"> class="hljs-ln-code"> class="hljs-ln-line">        self.convert(pt_path)
    10.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="10"> class="hljs-ln-code"> class="hljs-ln-line">    
    11.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="11"> class="hljs-ln-code"> class="hljs-ln-line">    def convert(self, pt_path):
    12.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="12"> class="hljs-ln-code"> class="hljs-ln-line">        model = YOLO(pt_path)
    13.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="13"> class="hljs-ln-code"> class="hljs-ln-line">        model.export(format = "onnx", opset=12)
    14.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="14"> class="hljs-ln-code"> class="hljs-ln-line"> 
    15.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="15"> class="hljs-ln-code"> class="hljs-ln-line">if __name__ == "__main__":
    16.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="16"> class="hljs-ln-code"> class="hljs-ln-line">    pt_path = r"best.pt"
    17.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="17"> class="hljs-ln-code"> class="hljs-ln-line">    Yolov8Pt2Onnx(pt_path)
     class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    运行之后会在yolov8同级目录下生成best.onnx

    五、PC端安装RKNNtoolkit2的及其环境

    ONNX转RKNN

    1、创建环境

    conda create -n rknn230 python=3.10 class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    2、这里需要用到rknn-toolkit2文件

    （最好用新版本；新版本包含yolov8,低版本有可能只包含yolov5）

    3、安装rknn-toolkit2环境

    （我自己用的是rknn-toolkit2-2.3.0；参考文章里面用的是rknn-toolkit2-2.1.0；自己用合适的版本就行，不一定非要选特定版本）

    conda activate rknn230
 class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    cd /home/sxj/YOLO/RKNN2_2.1.0/2.3.0/release/rknn-toolkit2-v2.3.0-2024-11-08/rknn-toolkit2/packages/x86_64
 class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    进入rknn-toolkit2-2.1.0\rknn-toolkit2-2.1.0\rknn-toolkit2\packages文件夹下，看到如下内容：

    终端输入：

    pip install -r requirements_cp310-2.1.0.txt -i https://pypi.tuna.tsinghua.edu.cn/simple

    pip install -r requirements_cp310-2.3.0.txt -i https://pypi.tuna.tsinghua.edu.cn/simple class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    然后再输入pip install rknn_toolkit2-2.1.0+81f21f4d-cp310-cp310-linux_x86_64.whl

    pip install rknn_toolkit2-2.3.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    到此rknn环境就配置完成了

    4、现在要进行模型转换

    1.可以参考rknn_model_zoo-2.3.0\examples\yolov8下的README指导进行转换：

    转换流程：

    2.先进入rknn_model_zoo-2.3.0\examples\yolov8\python文件夹，先打开yolov8.py，进行适配参数修改：

    找到这几个改成自己的

    1）

    CLASSES = ("person", "bicycle", "car",.....

    coco_id_list = [1, 2, 3,......

    parser.add_argument('--target', type=str, default='rk3566', help='target RKNPU platform')

    我的改完之后：

    yolov8.py

    1.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="1"> class="hljs-ln-code"> class="hljs-ln-line">import os
    2.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="2"> class="hljs-ln-code"> class="hljs-ln-line">import cv2
    3.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="3"> class="hljs-ln-code"> class="hljs-ln-line">import sys
    4.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="4"> class="hljs-ln-code"> class="hljs-ln-line">import argparse
    5.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="5"> class="hljs-ln-code"> class="hljs-ln-line"> 
    6.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="6"> class="hljs-ln-code"> class="hljs-ln-line"># add path
    7.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="7"> class="hljs-ln-code"> class="hljs-ln-line">realpath = os.path.abspath(__file__)
    8.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="8"> class="hljs-ln-code"> class="hljs-ln-line">_sep = os.path.sep
    9.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="9"> class="hljs-ln-code"> class="hljs-ln-line">realpath = realpath.split(_sep)
    10.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="10"> class="hljs-ln-code"> class="hljs-ln-line">sys.path.append(os.path.join(realpath[0]+_sep, *realpath[1:realpath.index('rknn_model_zoo')+1]))
    11.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="11"> class="hljs-ln-code"> class="hljs-ln-line"> 
    12.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="12"> class="hljs-ln-code"> class="hljs-ln-line">from py_utils.coco_utils import COCO_test_helper
    13.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="13"> class="hljs-ln-code"> class="hljs-ln-line">import numpy as np
    14.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="14"> class="hljs-ln-code"> class="hljs-ln-line"> 
    15.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="15"> class="hljs-ln-code"> class="hljs-ln-line"> 
    16.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="16"> class="hljs-ln-code"> class="hljs-ln-line">OBJ_THRESH = 0.25
    17.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="17"> class="hljs-ln-code"> class="hljs-ln-line">NMS_THRESH = 0.45
    18.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="18"> class="hljs-ln-code"> class="hljs-ln-line"> 
    19.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="19"> class="hljs-ln-code"> class="hljs-ln-line"># The follew two param is for map test
    20.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="20"> class="hljs-ln-code"> class="hljs-ln-line"># OBJ_THRESH = 0.001
    21.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="21"> class="hljs-ln-code"> class="hljs-ln-line"># NMS_THRESH = 0.65
    22.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="22"> class="hljs-ln-code"> class="hljs-ln-line"> 
    23.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="23"> class="hljs-ln-code"> class="hljs-ln-line">IMG_SIZE = (640, 640)  # (width, height), such as (1280, 736)
    24.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="24"> class="hljs-ln-code"> class="hljs-ln-line"> 
    25.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="25"> class="hljs-ln-code"> class="hljs-ln-line">CLASSES = ('bike','carsgraz','person')
    26.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="26"> class="hljs-ln-code"> class="hljs-ln-line"> 
    27.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="27"> class="hljs-ln-code"> class="hljs-ln-line">coco_id_list = [1, 2, 3]
    28.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="28"> class="hljs-ln-code"> class="hljs-ln-line"> 
    29.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="29"> class="hljs-ln-code"> class="hljs-ln-line"> 
    30.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="30"> class="hljs-ln-code"> class="hljs-ln-line">def filter_boxes(boxes, box_confidences, box_class_probs):
    31.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="31"> class="hljs-ln-code"> class="hljs-ln-line">    """Filter boxes with object threshold.
    32.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="32"> class="hljs-ln-code"> class="hljs-ln-line">    """
    33.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="33"> class="hljs-ln-code"> class="hljs-ln-line">    box_confidences = box_confidences.reshape(-1)
    34.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="34"> class="hljs-ln-code"> class="hljs-ln-line">    candidate, class_num = box_class_probs.shape
    35.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="35"> class="hljs-ln-code"> class="hljs-ln-line"> 
    36.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="36"> class="hljs-ln-code"> class="hljs-ln-line">    class_max_score = np.max(box_class_probs, axis=-1)
    37.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="37"> class="hljs-ln-code"> class="hljs-ln-line">    classes = np.argmax(box_class_probs, axis=-1)
    38.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="38"> class="hljs-ln-code"> class="hljs-ln-line"> 
    39.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="39"> class="hljs-ln-code"> class="hljs-ln-line">    _class_pos = np.where(class_max_score* box_confidences >= OBJ_THRESH)
    40.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="40"> class="hljs-ln-code"> class="hljs-ln-line">    scores = (class_max_score* box_confidences)[_class_pos]
    41.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="41"> class="hljs-ln-code"> class="hljs-ln-line"> 
    42.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="42"> class="hljs-ln-code"> class="hljs-ln-line">    boxes = boxes[_class_pos]
    43.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="43"> class="hljs-ln-code"> class="hljs-ln-line">    classes = classes[_class_pos]
    44.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="44"> class="hljs-ln-code"> class="hljs-ln-line"> 
    45.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="45"> class="hljs-ln-code"> class="hljs-ln-line">    return boxes, classes, scores
    46.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="46"> class="hljs-ln-code"> class="hljs-ln-line"> 
    47.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="47"> class="hljs-ln-code"> class="hljs-ln-line">def nms_boxes(boxes, scores):
    48.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="48"> class="hljs-ln-code"> class="hljs-ln-line">    """Suppress non-maximal boxes.
    49.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="49"> class="hljs-ln-code"> class="hljs-ln-line">    # Returns
    50.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="50"> class="hljs-ln-code"> class="hljs-ln-line">        keep: ndarray, index of effective boxes.
    51.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="51"> class="hljs-ln-code"> class="hljs-ln-line">    """
    52.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="52"> class="hljs-ln-code"> class="hljs-ln-line">    x = boxes[:, 0]
    53.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="53"> class="hljs-ln-code"> class="hljs-ln-line">    y = boxes[:, 1]
    54.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="54"> class="hljs-ln-code"> class="hljs-ln-line">    w = boxes[:, 2] - boxes[:, 0]
    55.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="55"> class="hljs-ln-code"> class="hljs-ln-line">    h = boxes[:, 3] - boxes[:, 1]
    56.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="56"> class="hljs-ln-code"> class="hljs-ln-line"> 
    57.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="57"> class="hljs-ln-code"> class="hljs-ln-line">    areas = w * h
    58.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="58"> class="hljs-ln-code"> class="hljs-ln-line">    order = scores.argsort()[::-1]
    59.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="59"> class="hljs-ln-code"> class="hljs-ln-line"> 
    60.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="60"> class="hljs-ln-code"> class="hljs-ln-line">    keep = []
    61.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="61"> class="hljs-ln-code"> class="hljs-ln-line">    while order.size > 0:
    62.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="62"> class="hljs-ln-code"> class="hljs-ln-line">        i = order[0]
    63.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="63"> class="hljs-ln-code"> class="hljs-ln-line">        keep.append(i)
    64.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="64"> class="hljs-ln-code"> class="hljs-ln-line"> 
    65.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="65"> class="hljs-ln-code"> class="hljs-ln-line">        xx1 = np.maximum(x[i], x[order[1:]])
    66.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="66"> class="hljs-ln-code"> class="hljs-ln-line">        yy1 = np.maximum(y[i], y[order[1:]])
    67.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="67"> class="hljs-ln-code"> class="hljs-ln-line">        xx2 = np.minimum(x[i] + w[i], x[order[1:]] + w[order[1:]])
    68.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="68"> class="hljs-ln-code"> class="hljs-ln-line">        yy2 = np.minimum(y[i] + h[i], y[order[1:]] + h[order[1:]])
    69.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="69"> class="hljs-ln-code"> class="hljs-ln-line"> 
    70.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="70"> class="hljs-ln-code"> class="hljs-ln-line">        w1 = np.maximum(0.0, xx2 - xx1 + 0.00001)
    71.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="71"> class="hljs-ln-code"> class="hljs-ln-line">        h1 = np.maximum(0.0, yy2 - yy1 + 0.00001)
    72.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="72"> class="hljs-ln-code"> class="hljs-ln-line">        inter = w1 * h1
    73.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="73"> class="hljs-ln-code"> class="hljs-ln-line"> 
    74.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="74"> class="hljs-ln-code"> class="hljs-ln-line">        ovr = inter / (areas[i] + areas[order[1:]] - inter)
    75.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="75"> class="hljs-ln-code"> class="hljs-ln-line">        inds = np.where(ovr <= NMS_THRESH)[0]
    76.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="76"> class="hljs-ln-code"> class="hljs-ln-line">        order = order[inds + 1]
    77.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="77"> class="hljs-ln-code"> class="hljs-ln-line">    keep = np.array(keep)
    78.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="78"> class="hljs-ln-code"> class="hljs-ln-line">    return keep
    79.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="79"> class="hljs-ln-code"> class="hljs-ln-line"> 
    80.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="80"> class="hljs-ln-code"> class="hljs-ln-line">def dfl(position):
    81.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="81"> class="hljs-ln-code"> class="hljs-ln-line">    # Distribution Focal Loss (DFL)
    82.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="82"> class="hljs-ln-code"> class="hljs-ln-line">    import torch
    83.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="83"> class="hljs-ln-code"> class="hljs-ln-line">    x = torch.tensor(position)
    84.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="84"> class="hljs-ln-code"> class="hljs-ln-line">    n,c,h,w = x.shape
    85.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="85"> class="hljs-ln-code"> class="hljs-ln-line">    p_num = 4
    86.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="86"> class="hljs-ln-code"> class="hljs-ln-line">    mc = c//p_num
    87.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="87"> class="hljs-ln-code"> class="hljs-ln-line">    y = x.reshape(n,p_num,mc,h,w)
    88.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="88"> class="hljs-ln-code"> class="hljs-ln-line">    y = y.softmax(2)
    89.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="89"> class="hljs-ln-code"> class="hljs-ln-line">    acc_metrix = torch.tensor(range(mc)).float().reshape(1,1,mc,1,1)
    90.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="90"> class="hljs-ln-code"> class="hljs-ln-line">    y = (y*acc_metrix).sum(2)
    91.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="91"> class="hljs-ln-code"> class="hljs-ln-line">    return y.numpy()
    92.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="92"> class="hljs-ln-code"> class="hljs-ln-line"> 
    93.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="93"> class="hljs-ln-code"> class="hljs-ln-line"> 
    94.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="94"> class="hljs-ln-code"> class="hljs-ln-line">def box_process(position):
    95.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="95"> class="hljs-ln-code"> class="hljs-ln-line">    grid_h, grid_w = position.shape[2:4]
    96.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="96"> class="hljs-ln-code"> class="hljs-ln-line">    col, row = np.meshgrid(np.arange(0, grid_w), np.arange(0, grid_h))
    97.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="97"> class="hljs-ln-code"> class="hljs-ln-line">    col = col.reshape(1, 1, grid_h, grid_w)
    98.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="98"> class="hljs-ln-code"> class="hljs-ln-line">    row = row.reshape(1, 1, grid_h, grid_w)
    99.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="99"> class="hljs-ln-code"> class="hljs-ln-line">    grid = np.concatenate((col, row), axis=1)
    100.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="100"> class="hljs-ln-code"> class="hljs-ln-line">    stride = np.array([IMG_SIZE[1]//grid_h, IMG_SIZE[0]//grid_w]).reshape(1,2,1,1)
    101.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="101"> class="hljs-ln-code"> class="hljs-ln-line"> 
    102.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="102"> class="hljs-ln-code"> class="hljs-ln-line">    position = dfl(position)
    103.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="103"> class="hljs-ln-code"> class="hljs-ln-line">    box_xy  = grid +0.5 -position[:,0:2,:,:]
    104.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="104"> class="hljs-ln-code"> class="hljs-ln-line">    box_xy2 = grid +0.5 +position[:,2:4,:,:]
    105.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="105"> class="hljs-ln-code"> class="hljs-ln-line">    xyxy = np.concatenate((box_xy*stride, box_xy2*stride), axis=1)
    106.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="106"> class="hljs-ln-code"> class="hljs-ln-line"> 
    107.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="107"> class="hljs-ln-code"> class="hljs-ln-line">    return xyxy
    108.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="108"> class="hljs-ln-code"> class="hljs-ln-line"> 
    109.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="109"> class="hljs-ln-code"> class="hljs-ln-line">def post_process(input_data):
    110.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="110"> class="hljs-ln-code"> class="hljs-ln-line">    boxes, scores, classes_conf = [], [], []
    111.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="111"> class="hljs-ln-code"> class="hljs-ln-line">    defualt_branch=3
    112.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="112"> class="hljs-ln-code"> class="hljs-ln-line">    pair_per_branch = len(input_data)//defualt_branch
    113.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="113"> class="hljs-ln-code"> class="hljs-ln-line">    # Python 忽略 score_sum 输出
    114.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="114"> class="hljs-ln-code"> class="hljs-ln-line">    for i in range(defualt_branch):
    115.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="115"> class="hljs-ln-code"> class="hljs-ln-line">        boxes.append(box_process(input_data[pair_per_branch*i]))
    116.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="116"> class="hljs-ln-code"> class="hljs-ln-line">        classes_conf.append(input_data[pair_per_branch*i+1])
    117.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="117"> class="hljs-ln-code"> class="hljs-ln-line">        scores.append(np.ones_like(input_data[pair_per_branch*i+1][:,:1,:,:], dtype=np.float32))
    118.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="118"> class="hljs-ln-code"> class="hljs-ln-line"> 
    119.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="119"> class="hljs-ln-code"> class="hljs-ln-line">    def sp_flatten(_in):
    120.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="120"> class="hljs-ln-code"> class="hljs-ln-line">        ch = _in.shape[1]
    121.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="121"> class="hljs-ln-code"> class="hljs-ln-line">        _in = _in.transpose(0,2,3,1)
    122.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="122"> class="hljs-ln-code"> class="hljs-ln-line">        return _in.reshape(-1, ch)
    123.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="123"> class="hljs-ln-code"> class="hljs-ln-line"> 
    124.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="124"> class="hljs-ln-code"> class="hljs-ln-line">    boxes = [sp_flatten(_v) for _v in boxes]
    125.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="125"> class="hljs-ln-code"> class="hljs-ln-line">    classes_conf = [sp_flatten(_v) for _v in classes_conf]
    126.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="126"> class="hljs-ln-code"> class="hljs-ln-line">    scores = [sp_flatten(_v) for _v in scores]
    127.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="127"> class="hljs-ln-code"> class="hljs-ln-line"> 
    128.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="128"> class="hljs-ln-code"> class="hljs-ln-line">    boxes = np.concatenate(boxes)
    129.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="129"> class="hljs-ln-code"> class="hljs-ln-line">    classes_conf = np.concatenate(classes_conf)
    130.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="130"> class="hljs-ln-code"> class="hljs-ln-line">    scores = np.concatenate(scores)
    131.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="131"> class="hljs-ln-code"> class="hljs-ln-line"> 
    132.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="132"> class="hljs-ln-code"> class="hljs-ln-line">    # filter according to threshold
    133.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="133"> class="hljs-ln-code"> class="hljs-ln-line">    boxes, classes, scores = filter_boxes(boxes, scores, classes_conf)
    134.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="134"> class="hljs-ln-code"> class="hljs-ln-line"> 
    135.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="135"> class="hljs-ln-code"> class="hljs-ln-line">    # nms
    136.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="136"> class="hljs-ln-code"> class="hljs-ln-line">    nboxes, nclasses, nscores = [], [], []
    137.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="137"> class="hljs-ln-code"> class="hljs-ln-line">    for c in set(classes):
    138.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="138"> class="hljs-ln-code"> class="hljs-ln-line">        inds = np.where(classes == c)
    139.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="139"> class="hljs-ln-code"> class="hljs-ln-line">        b = boxes[inds]
    140.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="140"> class="hljs-ln-code"> class="hljs-ln-line">        c = classes[inds]
    141.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="141"> class="hljs-ln-code"> class="hljs-ln-line">        s = scores[inds]
    142.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="142"> class="hljs-ln-code"> class="hljs-ln-line">        keep = nms_boxes(b, s)
    143.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="143"> class="hljs-ln-code"> class="hljs-ln-line"> 
    144.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="144"> class="hljs-ln-code"> class="hljs-ln-line">        if len(keep) != 0:
    145.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="145"> class="hljs-ln-code"> class="hljs-ln-line">            nboxes.append(b[keep])
    146.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="146"> class="hljs-ln-code"> class="hljs-ln-line">            nclasses.append(c[keep])
    147.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="147"> class="hljs-ln-code"> class="hljs-ln-line">            nscores.append(s[keep])
    148.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="148"> class="hljs-ln-code"> class="hljs-ln-line"> 
    149.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="149"> class="hljs-ln-code"> class="hljs-ln-line">    if not nclasses and not nscores:
    150.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="150"> class="hljs-ln-code"> class="hljs-ln-line">        return None, None, None
    151.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="151"> class="hljs-ln-code"> class="hljs-ln-line"> 
    152.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="152"> class="hljs-ln-code"> class="hljs-ln-line">    boxes = np.concatenate(nboxes)
    153.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="153"> class="hljs-ln-code"> class="hljs-ln-line">    classes = np.concatenate(nclasses)
    154.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="154"> class="hljs-ln-code"> class="hljs-ln-line">    scores = np.concatenate(nscores)
    155.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="155"> class="hljs-ln-code"> class="hljs-ln-line"> 
    156.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="156"> class="hljs-ln-code"> class="hljs-ln-line">    return boxes, classes, scores
    157.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="157"> class="hljs-ln-code"> class="hljs-ln-line"> 
    158.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="158"> class="hljs-ln-code"> class="hljs-ln-line"> 
    159.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="159"> class="hljs-ln-code"> class="hljs-ln-line">def draw(image, boxes, scores, classes):
    160.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="160"> class="hljs-ln-code"> class="hljs-ln-line">    for box, score, cl in zip(boxes, scores, classes):
    161.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="161"> class="hljs-ln-code"> class="hljs-ln-line">        top, left, right, bottom = [int(_b) for _b in box]
    162.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="162"> class="hljs-ln-code"> class="hljs-ln-line">        print("%s @ (%d %d %d %d) %.3f" % (CLASSES[cl], top, left, right, bottom, score))
    163.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="163"> class="hljs-ln-code"> class="hljs-ln-line">        cv2.rectangle(image, (top, left), (right, bottom), (255, 0, 0), 2)
    164.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="164"> class="hljs-ln-code"> class="hljs-ln-line">        cv2.putText(image, '{0} {1:.2f}'.format(CLASSES[cl], score),
    165.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="165"> class="hljs-ln-code"> class="hljs-ln-line">                    (top, left - 6), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 255), 2)
    166.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="166"> class="hljs-ln-code"> class="hljs-ln-line"> 
    167.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="167"> class="hljs-ln-code"> class="hljs-ln-line">def setup_model(args):
    168.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="168"> class="hljs-ln-code"> class="hljs-ln-line">    model_path = args.model_path
    169.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="169"> class="hljs-ln-code"> class="hljs-ln-line">    if model_path.endswith('.pt') or model_path.endswith('.torchscript'):
    170.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="170"> class="hljs-ln-code"> class="hljs-ln-line">        platform = 'pytorch'
    171.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="171"> class="hljs-ln-code"> class="hljs-ln-line">        from py_utils.pytorch_executor import Torch_model_container
    172.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="172"> class="hljs-ln-code"> class="hljs-ln-line">        model = Torch_model_container(args.model_path)
    173.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="173"> class="hljs-ln-code"> class="hljs-ln-line">    elif model_path.endswith('.rknn'):
    174.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="174"> class="hljs-ln-code"> class="hljs-ln-line">        platform = 'rknn'
    175.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="175"> class="hljs-ln-code"> class="hljs-ln-line">        from py_utils.rknn_executor import RKNN_model_container 
    176.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="176"> class="hljs-ln-code"> class="hljs-ln-line">        model = RKNN_model_container(args.model_path, args.target, args.device_id)
    177.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="177"> class="hljs-ln-code"> class="hljs-ln-line">    elif model_path.endswith('onnx'):
    178.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="178"> class="hljs-ln-code"> class="hljs-ln-line">        platform = 'onnx'
    179.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="179"> class="hljs-ln-code"> class="hljs-ln-line">        from py_utils.onnx_executor import ONNX_model_container
    180.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="180"> class="hljs-ln-code"> class="hljs-ln-line">        model = ONNX_model_container(args.model_path)
    181.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="181"> class="hljs-ln-code"> class="hljs-ln-line">    else:
    182.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="182"> class="hljs-ln-code"> class="hljs-ln-line">        assert False, "{} is not rknn/pytorch/onnx model".format(model_path)
    183.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="183"> class="hljs-ln-code"> class="hljs-ln-line">    print('Model-{} is {} model, starting val'.format(model_path, platform))
    184.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="184"> class="hljs-ln-code"> class="hljs-ln-line">    return model, platform
    185.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="185"> class="hljs-ln-code"> class="hljs-ln-line"> 
    186.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="186"> class="hljs-ln-code"> class="hljs-ln-line">def img_check(path):
    187.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="187"> class="hljs-ln-code"> class="hljs-ln-line">    img_type = ['.jpg', '.jpeg', '.png', '.bmp']
    188.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="188"> class="hljs-ln-code"> class="hljs-ln-line">    for _type in img_type:
    189.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="189"> class="hljs-ln-code"> class="hljs-ln-line">        if path.endswith(_type) or path.endswith(_type.upper()):
    190.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="190"> class="hljs-ln-code"> class="hljs-ln-line">            return True
    191.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="191"> class="hljs-ln-code"> class="hljs-ln-line">    return False
    192.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="192"> class="hljs-ln-code"> class="hljs-ln-line"> 
    193.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="193"> class="hljs-ln-code"> class="hljs-ln-line">if __name__ == '__main__':
    194.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="194"> class="hljs-ln-code"> class="hljs-ln-line">    parser = argparse.ArgumentParser(description='Process some integers.')
    195.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="195"> class="hljs-ln-code"> class="hljs-ln-line">    # basic params
    196.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="196"> class="hljs-ln-code"> class="hljs-ln-line">    parser.add_argument('--model_path', type=str, required= True, help='model path, could be .pt or .rknn file')
    197.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="197"> class="hljs-ln-code"> class="hljs-ln-line">    parser.add_argument('--target', type=str, default='rk3588', help='target RKNPU platform')
    198.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="198"> class="hljs-ln-code"> class="hljs-ln-line">    parser.add_argument('--device_id', type=str, default=None, help='device id')
    199.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="199"> class="hljs-ln-code"> class="hljs-ln-line">    
    200.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="200"> class="hljs-ln-code"> class="hljs-ln-line">    parser.add_argument('--img_show', action='store_true', default=False, help='draw the result and show')
    201.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="201"> class="hljs-ln-code"> class="hljs-ln-line">    parser.add_argument('--img_save', action='store_true', default=False, help='save the result')
    202.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="202"> class="hljs-ln-code"> class="hljs-ln-line"> 
    203.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="203"> class="hljs-ln-code"> class="hljs-ln-line">    # data params
    204.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="204"> class="hljs-ln-code"> class="hljs-ln-line">    parser.add_argument('--anno_json', type=str, default='../../../datasets/COCO/annotations/instances_val2017.json', help='coco annotation path')
    205.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="205"> class="hljs-ln-code"> class="hljs-ln-line">    # coco val folder: '../../../datasets/COCO//val2017'
    206.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="206"> class="hljs-ln-code"> class="hljs-ln-line">    parser.add_argument('--img_folder', type=str, default='../model', help='img folder path')
    207.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="207"> class="hljs-ln-code"> class="hljs-ln-line">    parser.add_argument('--coco_map_test', action='store_true', help='enable coco map test')
    208.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="208"> class="hljs-ln-code"> class="hljs-ln-line"> 
    209.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="209"> class="hljs-ln-code"> class="hljs-ln-line">    args = parser.parse_args()
    210.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="210"> class="hljs-ln-code"> class="hljs-ln-line"> 
    211.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="211"> class="hljs-ln-code"> class="hljs-ln-line">    # init model
    212.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="212"> class="hljs-ln-code"> class="hljs-ln-line">    model, platform = setup_model(args)
    213.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="213"> class="hljs-ln-code"> class="hljs-ln-line"> 
    214.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="214"> class="hljs-ln-code"> class="hljs-ln-line">    file_list = sorted(os.listdir(args.img_folder))
    215.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="215"> class="hljs-ln-code"> class="hljs-ln-line">    img_list = []
    216.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="216"> class="hljs-ln-code"> class="hljs-ln-line">    for path in file_list:
    217.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="217"> class="hljs-ln-code"> class="hljs-ln-line">        if img_check(path):
    218.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="218"> class="hljs-ln-code"> class="hljs-ln-line">            img_list.append(path)
    219.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="219"> class="hljs-ln-code"> class="hljs-ln-line">    co_helper = COCO_test_helper(enable_letter_box=True)
    220.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="220"> class="hljs-ln-code"> class="hljs-ln-line"> 
    221.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="221"> class="hljs-ln-code"> class="hljs-ln-line">    # run test
    222.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="222"> class="hljs-ln-code"> class="hljs-ln-line">    for i in range(len(img_list)):
    223.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="223"> class="hljs-ln-code"> class="hljs-ln-line">        print('infer {}/{}'.format(i+1, len(img_list)), end='\r')
    224.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="224"> class="hljs-ln-code"> class="hljs-ln-line"> 
    225.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="225"> class="hljs-ln-code"> class="hljs-ln-line">        img_name = img_list[i]
    226.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="226"> class="hljs-ln-code"> class="hljs-ln-line">        img_path = os.path.join(args.img_folder, img_name)
    227.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="227"> class="hljs-ln-code"> class="hljs-ln-line">        if not os.path.exists(img_path):
    228.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="228"> class="hljs-ln-code"> class="hljs-ln-line">            print("{} is not found", img_name)
    229.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="229"> class="hljs-ln-code"> class="hljs-ln-line">            continue
    230.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="230"> class="hljs-ln-code"> class="hljs-ln-line"> 
    231.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="231"> class="hljs-ln-code"> class="hljs-ln-line">        img_src = cv2.imread(img_path)
    232.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="232"> class="hljs-ln-code"> class="hljs-ln-line">        if img_src is None:
    233.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="233"> class="hljs-ln-code"> class="hljs-ln-line">            continue
    234.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="234"> class="hljs-ln-code"> class="hljs-ln-line"> 
    235.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="235"> class="hljs-ln-code"> class="hljs-ln-line">        '''
    236.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="236"> class="hljs-ln-code"> class="hljs-ln-line">        # using for test input dumped by C.demo
    237.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="237"> class="hljs-ln-code"> class="hljs-ln-line">        img_src = np.fromfile('./input_b/demo_c_input_hwc_rgb.txt', dtype=np.uint8).reshape(640,640,3)
    238.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="238"> class="hljs-ln-code"> class="hljs-ln-line">        img_src = cv2.cvtColor(img_src, cv2.COLOR_RGB2BGR)
    239.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="239"> class="hljs-ln-code"> class="hljs-ln-line">        '''
    240.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="240"> class="hljs-ln-code"> class="hljs-ln-line"> 
    241.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="241"> class="hljs-ln-code"> class="hljs-ln-line">        # Due to rga init with (0,0,0), we using pad_color (0,0,0) instead of (114, 114, 114)
    242.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="242"> class="hljs-ln-code"> class="hljs-ln-line">        pad_color = (0,0,0)
    243.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="243"> class="hljs-ln-code"> class="hljs-ln-line">        img = co_helper.letter_box(im= img_src.copy(), new_shape=(IMG_SIZE[1], IMG_SIZE[0]), pad_color=(0,0,0))
    244.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="244"> class="hljs-ln-code"> class="hljs-ln-line">        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    245.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="245"> class="hljs-ln-code"> class="hljs-ln-line"> 
    246.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="246"> class="hljs-ln-code"> class="hljs-ln-line">        # preprocee if not rknn model
    247.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="247"> class="hljs-ln-code"> class="hljs-ln-line">        if platform in ['pytorch', 'onnx']:
    248.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="248"> class="hljs-ln-code"> class="hljs-ln-line">            input_data = img.transpose((2,0,1))
    249.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="249"> class="hljs-ln-code"> class="hljs-ln-line">            input_data = input_data.reshape(1,*input_data.shape).astype(np.float32)
    250.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="250"> class="hljs-ln-code"> class="hljs-ln-line">            input_data = input_data/255.
    251.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="251"> class="hljs-ln-code"> class="hljs-ln-line">        else:
    252.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="252"> class="hljs-ln-code"> class="hljs-ln-line">            input_data = img
    253.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="253"> class="hljs-ln-code"> class="hljs-ln-line"> 
    254.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="254"> class="hljs-ln-code"> class="hljs-ln-line">        outputs = model.run([input_data])
    255.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="255"> class="hljs-ln-code"> class="hljs-ln-line">        boxes, classes, scores = post_process(outputs)
    256.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="256"> class="hljs-ln-code"> class="hljs-ln-line"> 
    257.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="257"> class="hljs-ln-code"> class="hljs-ln-line">        if args.img_show or args.img_save:
    258.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="258"> class="hljs-ln-code"> class="hljs-ln-line">            print('\n\nIMG: {}'.format(img_name))
    259.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="259"> class="hljs-ln-code"> class="hljs-ln-line">            img_p = img_src.copy()
    260.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="260"> class="hljs-ln-code"> class="hljs-ln-line">            if boxes is not None:
    261.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="261"> class="hljs-ln-code"> class="hljs-ln-line">                draw(img_p, co_helper.get_real_box(boxes), scores, classes)
    262.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="262"> class="hljs-ln-code"> class="hljs-ln-line"> 
    263.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="263"> class="hljs-ln-code"> class="hljs-ln-line">            if args.img_save:
    264.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="264"> class="hljs-ln-code"> class="hljs-ln-line">                if not os.path.exists('./result'):
    265.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="265"> class="hljs-ln-code"> class="hljs-ln-line">                    os.mkdir('./result')
    266.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="266"> class="hljs-ln-code"> class="hljs-ln-line">                result_path = os.path.join('./result', img_name)
    267.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="267"> class="hljs-ln-code"> class="hljs-ln-line">                cv2.imwrite(result_path, img_p)
    268.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="268"> class="hljs-ln-code"> class="hljs-ln-line">                print('Detection result save to {}'.format(result_path))
    269.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="269"> class="hljs-ln-code"> class="hljs-ln-line">                        
    270.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="270"> class="hljs-ln-code"> class="hljs-ln-line">            if args.img_show:
    271.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="271"> class="hljs-ln-code"> class="hljs-ln-line">                cv2.imshow("full post process result", img_p)
    272.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="272"> class="hljs-ln-code"> class="hljs-ln-line">                cv2.waitKeyEx(0)
    273.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="273"> class="hljs-ln-code"> class="hljs-ln-line"> 
    274.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="274"> class="hljs-ln-code"> class="hljs-ln-line">        # record maps
    275.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="275"> class="hljs-ln-code"> class="hljs-ln-line">        if args.coco_map_test is True:
    276.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="276"> class="hljs-ln-code"> class="hljs-ln-line">            if boxes is not None:
    277.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="277"> class="hljs-ln-code"> class="hljs-ln-line">                for i in range(boxes.shape[0]):
    278.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="278"> class="hljs-ln-code"> class="hljs-ln-line">                    co_helper.add_single_record(image_id = int(img_name.split('.')[0]),
    279.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="279"> class="hljs-ln-code"> class="hljs-ln-line">                                                category_id = coco_id_list[int(classes[i])],
    280.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="280"> class="hljs-ln-code"> class="hljs-ln-line">                                                bbox = boxes[i],
    281.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="281"> class="hljs-ln-code"> class="hljs-ln-line">                                                score = round(scores[i], 5).item()
    282.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="282"> class="hljs-ln-code"> class="hljs-ln-line">                                                )
    283.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="283"> class="hljs-ln-code"> class="hljs-ln-line"> 
    284.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="284"> class="hljs-ln-code"> class="hljs-ln-line">    # calculate maps
    285.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="285"> class="hljs-ln-code"> class="hljs-ln-line">    if args.coco_map_test is True:
    286.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="286"> class="hljs-ln-code"> class="hljs-ln-line">        pred_json = args.model_path.split('.')[-2]+ '_{}'.format(platform) +'.json'
    287.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="287"> class="hljs-ln-code"> class="hljs-ln-line">        pred_json = pred_json.split('/')[-1]
    288.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="288"> class="hljs-ln-code"> class="hljs-ln-line">        pred_json = os.path.join('./', pred_json)
    289.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="289"> class="hljs-ln-code"> class="hljs-ln-line">        co_helper.export_to_json(pred_json)
    290.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="290"> class="hljs-ln-code"> class="hljs-ln-line"> 
    291.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="291"> class="hljs-ln-code"> class="hljs-ln-line">        from py_utils.coco_utils import coco_eval_with_json
    292.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="292"> class="hljs-ln-code"> class="hljs-ln-line">        coco_eval_with_json(args.anno_json, pred_json)
    293.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="293"> class="hljs-ln-code"> class="hljs-ln-line"> 
    294.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="294"> class="hljs-ln-code"> class="hljs-ln-line">    # release
    295.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="295"> class="hljs-ln-code"> class="hljs-ln-line">    model.release()
     class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    3.然后修改convert.py，如下所示：

    我自己convert.py代码：

    1.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="1"> class="hljs-ln-code"> class="hljs-ln-line">import sys
    2.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="2"> class="hljs-ln-code"> class="hljs-ln-line">from rknn.api import RKNN
    3.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="3"> class="hljs-ln-code"> class="hljs-ln-line"> 
    4.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="4"> class="hljs-ln-code"> class="hljs-ln-line">DATASET_PATH = '../../../datasets/COCO/coco_subset_20.txt'
    5.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="5"> class="hljs-ln-code"> class="hljs-ln-line">DEFAULT_RKNN_PATH = 'best.rknn'
    6.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="6"> class="hljs-ln-code"> class="hljs-ln-line">DEFAULT_QUANT = True
    7.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="7"> class="hljs-ln-code"> class="hljs-ln-line"> 
    8.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="8"> class="hljs-ln-code"> class="hljs-ln-line">def parse_arg():
    9.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="9"> class="hljs-ln-code"> class="hljs-ln-line">    if len(sys.argv) < 3:
    10.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="10"> class="hljs-ln-code"> class="hljs-ln-line">        print("Usage: python3 {} onnx_model_path [platform] [dtype(optional)] [output_rknn_path(optional)]".format(sys.argv[0]));
    11.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="11"> class="hljs-ln-code"> class="hljs-ln-line">        print("       platform choose from [rk3562,rk3566,rk3568,rk3576,rk3588,rk1808,rv1109,rv1126]")
    12.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="12"> class="hljs-ln-code"> class="hljs-ln-line">        print("       dtype choose from [i8, fp] for [rk3562,rk3566,rk3568,rk3576,rk3588]")
    13.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="13"> class="hljs-ln-code"> class="hljs-ln-line">        print("       dtype choose from [u8, fp] for [rk1808,rv1109,rv1126]")
    14.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="14"> class="hljs-ln-code"> class="hljs-ln-line">        exit(1)
    15.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="15"> class="hljs-ln-code"> class="hljs-ln-line"> 
    16.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="16"> class="hljs-ln-code"> class="hljs-ln-line">    model_path = sys.argv[1]
    17.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="17"> class="hljs-ln-code"> class="hljs-ln-line">    platform = sys.argv[2]
    18.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="18"> class="hljs-ln-code"> class="hljs-ln-line"> 
    19.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="19"> class="hljs-ln-code"> class="hljs-ln-line">    do_quant = DEFAULT_QUANT
    20.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="20"> class="hljs-ln-code"> class="hljs-ln-line">    if len(sys.argv) > 3:
    21.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="21"> class="hljs-ln-code"> class="hljs-ln-line">        model_type = sys.argv[3]
    22.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="22"> class="hljs-ln-code"> class="hljs-ln-line">        if model_type not in ['i8', 'u8', 'fp']:
    23.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="23"> class="hljs-ln-code"> class="hljs-ln-line">            print("ERROR: Invalid model type: {}".format(model_type))
    24.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="24"> class="hljs-ln-code"> class="hljs-ln-line">            exit(1)
    25.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="25"> class="hljs-ln-code"> class="hljs-ln-line">        elif model_type in ['i8', 'u8']:
    26.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="26"> class="hljs-ln-code"> class="hljs-ln-line">            do_quant = True
    27.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="27"> class="hljs-ln-code"> class="hljs-ln-line">        else:
    28.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="28"> class="hljs-ln-code"> class="hljs-ln-line">            do_quant = False
    29.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="29"> class="hljs-ln-code"> class="hljs-ln-line"> 
    30.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="30"> class="hljs-ln-code"> class="hljs-ln-line">    if len(sys.argv) > 4:
    31.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="31"> class="hljs-ln-code"> class="hljs-ln-line">        output_path = sys.argv[4]
    32.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="32"> class="hljs-ln-code"> class="hljs-ln-line">    else:
    33.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="33"> class="hljs-ln-code"> class="hljs-ln-line">        output_path = DEFAULT_RKNN_PATH
    34.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="34"> class="hljs-ln-code"> class="hljs-ln-line"> 
    35.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="35"> class="hljs-ln-code"> class="hljs-ln-line">    return model_path, platform, do_quant, output_path
    36.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="36"> class="hljs-ln-code"> class="hljs-ln-line"> 
    37.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="37"> class="hljs-ln-code"> class="hljs-ln-line">if __name__ == '__main__':
    38.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="38"> class="hljs-ln-code"> class="hljs-ln-line">    model_path, platform, do_quant, output_path = parse_arg()
    39.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="39"> class="hljs-ln-code"> class="hljs-ln-line"> 
    40.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="40"> class="hljs-ln-code"> class="hljs-ln-line">    # Create RKNN object
    41.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="41"> class="hljs-ln-code"> class="hljs-ln-line">    rknn = RKNN(verbose=False)
    42.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="42"> class="hljs-ln-code"> class="hljs-ln-line"> 
    43.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="43"> class="hljs-ln-code"> class="hljs-ln-line">    # Pre-process config
    44.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="44"> class="hljs-ln-code"> class="hljs-ln-line">    print('--> Config model')
    45.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="45"> class="hljs-ln-code"> class="hljs-ln-line">    rknn.config(mean_values=[[0, 0, 0]], std_values=[
    46.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="46"> class="hljs-ln-code"> class="hljs-ln-line">                    [255, 255, 255]], target_platform=platform)
    47.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="47"> class="hljs-ln-code"> class="hljs-ln-line">    print('done')
    48.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="48"> class="hljs-ln-code"> class="hljs-ln-line"> 
    49.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="49"> class="hljs-ln-code"> class="hljs-ln-line">    # Load model
    50.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="50"> class="hljs-ln-code"> class="hljs-ln-line">    print('--> Loading model')
    51.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="51"> class="hljs-ln-code"> class="hljs-ln-line">    ret = rknn.load_onnx(model=model_path)
    52.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="52"> class="hljs-ln-code"> class="hljs-ln-line">    if ret != 0:
    53.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="53"> class="hljs-ln-code"> class="hljs-ln-line">        print('Load model failed!')
    54.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="54"> class="hljs-ln-code"> class="hljs-ln-line">        exit(ret)
    55.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="55"> class="hljs-ln-code"> class="hljs-ln-line">    print('done')
    56.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="56"> class="hljs-ln-code"> class="hljs-ln-line"> 
    57.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="57"> class="hljs-ln-code"> class="hljs-ln-line">    # Build model
    58.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="58"> class="hljs-ln-code"> class="hljs-ln-line">    print('--> Building model')
    59.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="59"> class="hljs-ln-code"> class="hljs-ln-line">    ret = rknn.build(do_quantization=do_quant, dataset=DATASET_PATH)
    60.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="60"> class="hljs-ln-code"> class="hljs-ln-line">    if ret != 0:
    61.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="61"> class="hljs-ln-code"> class="hljs-ln-line">        print('Build model failed!')
    62.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="62"> class="hljs-ln-code"> class="hljs-ln-line">        exit(ret)
    63.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="63"> class="hljs-ln-code"> class="hljs-ln-line">    print('done')
    64.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="64"> class="hljs-ln-code"> class="hljs-ln-line"> 
    65.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="65"> class="hljs-ln-code"> class="hljs-ln-line">    # Export rknn model
    66.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="66"> class="hljs-ln-code"> class="hljs-ln-line">    print('--> Export rknn model')
    67.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="67"> class="hljs-ln-code"> class="hljs-ln-line">    ret = rknn.export_rknn(output_path)
    68.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="68"> class="hljs-ln-code"> class="hljs-ln-line">    if ret != 0:
    69.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="69"> class="hljs-ln-code"> class="hljs-ln-line">        print('Export rknn model failed!')
    70.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="70"> class="hljs-ln-code"> class="hljs-ln-line">        exit(ret)
    71.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="71"> class="hljs-ln-code"> class="hljs-ln-line">    print('done')
    72.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="72"> class="hljs-ln-code"> class="hljs-ln-line"> 
    73.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="73"> class="hljs-ln-code"> class="hljs-ln-line">    # Release
    74.  class="hljs-ln-numbers"> class="hljs-ln-line hljs-ln-n" data-line-number="74"> class="hljs-ln-code"> class="hljs-ln-line">    rknn.release()
     class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    4.修改完成后，将我们之前得到的onnx模型复制到python文件夹下：

    图片是截图其他博主的，更新完会用自己的替换掉

    conda activate rknn230
 class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    netron best.onnx
 class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    netron best.rknn class="hljs-button signin active" data-title="登录复制" data-report-click="{"spm":"1001.2101.3001.4334"}" onclick="hljs.signin(event)">

    打开终端，激活rknn230环境，输入命令：python convert.py best.onnx rk3588
    结果如下：

    可以看到：这两条错误信息表明，在RKNN模型编译或运行时，某个字段值的位宽超出了预期的限制

    W build: The default input dtype of 'images' is changed from 'float32' to 'int8' in rknn model for performance!
                           Please take care of this change when deploy rknn model with Runtime API!
    W build: The default output dtype of 'output0' is changed from 'float32' to 'int8' in rknn model for performance!
                          Please take care of this change when deploy rknn model with Runtime API!
    I rknn building ...
    E RKNN: [17:26:23.759] REGTASK: The bit width of field value exceeds the limit, target: v2, offset: 0x4038, shift = 0, limit: 0x1fff, value: 0x419f
    E RKNN: [17:26:23.759] REGTASK: The bit width of field value exceeds the limit, target: v2, offset: 0x4038, shift = 16, limit: 0x1fff, value: 0x419f
    I rknn building done.
    done
    --> Export rknn model
     

    先把模型部署到RK3588上看影响大不，后面解决了会更新上来

    查看同级目录可以看到best.rknn

    六、模型部署到RK3588上

    1、部署到rk3588的项目，已经更新在我的另一篇文章上这里不再赘述

    文章链接：yolov8部署到RK3588_并运行示例12.4_rk3588运行yolo-CSDN博客

    至此完成!!!

    安装rknn-toolkit2环境时

    也可以直接参考我的另一篇文章：

    但是我的用的rknn-toolkit2是1.6.0里面只包含yolov5,所以需要最新版本！

    （链接：记录如何在RK3588板子上跑通paddle的OCR模型_rk3588 paddle ocr-CSDN博客）

    参考：YOLOv8教程系列：一、使用自定义数据集训练YOLOv8模型（详细版教程，你只看一篇-＞调参攻略），包含环境搭建/数据准备/模型训练/预测/验证/导出等_yolov8训练自己的数据集-CSDN博客

    参考：【YOLOv8部署至RK3588】模型训练→转换RKNN→开发板部署_yolov8转rknn-CSDN博客

    注：本文转载自blog.csdn.net的住在阳光的心里的文章"https://blog.csdn.net/qq_34438969/article/details/117091154"。版权归原作者所有，此博客不拥有其著作权，亦不承担相应法律责任。如有侵权，请联系我们删除。
    复制链接
    复制链接
    相关推荐
    发表评论

    评论记录：

    未查询到任何数据！