ModelArts-based object detection YOLOv3 practice [play with HUAWEI CLOUD]

前言

YOLO场景运用： YOLO作为一个one-stage目标检测算法,在速度和准确度上都有杰出的表现.

除了在NotebookThe development environment to run the algorithm and the definition of a,通过ModelArtsPreset algorithm can also beYOLOTraining algorithm and reasoning.The following example is a useNotebook中的YOLO V3Training object detection and reasoning process.

另外,想要了解ModelArts预置算法,可以前往华为云GitHub教程链接：ModelArts物体检测Yolo_V3预置算法案例

Or you can refer to my this post before：https://bbs.huaweicloud.com/blogs/224006

背景

Object detection is an important research field in computer vision,In testing,行人跟踪,自动驾驶,Medical imaging has been widely used in such fields as.Different from the simple image classification,Object detection is aimed at to accurately identify the target in the image,As well as the position of the object and classification,So can be applied to more high-level visual processing scene.例如在自动驾驶领域,Need to identify camera image of vehicle、行人、Traffic signs and their locations,According to these data, in order to further determine the driving strategy.这次实践,I will focus onYOLO算法,YOLO(You Only Look Once)是一种one-stage物体检测算法.

基础环境准备

在使用 ModelArts 进行 AI 开发前,Please complete the following basic operating oh（如有已完成部分,请忽略）,主要分为4步（注册–>实名认证–>服务授权–>领代金券）：

1. 使用手机号注册华为云账号：点击注册

2. 点此去完成实名认证,账号类型选"个人",个人认证类型推荐使用"扫码认证". 

   
   

3. 点此进入 ModelArts Data management console page,Above will be prompted to access authorization,点击【服务授权】按钮,按下图顺序操作： 

   
   

4. 进入 ModelArts 控制台首页,如下图,点击页面上的"彩蛋",Get the new vouchers！Subsequent steps may produce cost of resource consumption,Please be sure to get. 
   

以上操作,Also provides a detailed video tutorials,点此查看：ModelArts环境配置

基于ModelArts的物体检测YOLOv3Practice steps

1.点击进入 【物体检测YOLOv3实践】的案例页面,并完成基础配置.

本次实践使用的是AI框架： TensorFlow-1.13.1,在GPU下面运行;推荐规格: GPU V100

点击 Run in ModelArts,进入 JupyterLab 页面：

稍等30s左右,等待项目初始化完毕......

由于需要GPU下运行,So we need to switch from the specification,并选择 [限时免费]GPU: 1*V100|CPU: 8核 64GB ,点击切换规格.

Wait for a moment just switch to complete,如下所示：

2.数据和代码下载

运行下面代码,Download and extract the data and code

这里使用coco数据,共80个类别.

3准备数据

3.1文件路径定义

3.2读取标注数据

3.3数据读取函数,构建数据生成器.

Each read a lot of data to memory training,And to do data to enhance

4.模型训练

这里使用Keras深度学习框架搭建YOLOv3神经网络.

4.1构建神经网络

可以在./yolo3/model.pyFile for details

4.2开始训练

4.3保存模型

5.模型测试

5.1Open the test photo

5.2图像预处理

5.3构建模型

5.4加载模型权重

5.5定义IOU以及score

• IOU： Will be occurring simultaneously than larger thanIOUBounding box as redundant frames to remove
• score：To predict the score greater thanscoreThe bounding box of screening out

5.6构建输出[boxes, scores, classes]

5.7进行预测

5.8The predicted results drawn in the picture

Output the first artwork

Object detection effect

可以看出,基于ModelArts的物体检测YOLOv3Can detect the umbrella and people,Box to choose accurate,With the precision of data set training also has a relationship.

总结

Object detection is an important research field in computer vision,In testing,行人跟踪,自动驾驶,Medical imaging has been widely used in such fields as.Different from the simple image classification,Object detection is aimed at to accurately identify the target in the image,As well as the position of the object and classification,So can be applied to more high-level visual processing scene.本次实操,Through good training data for testing,较为精确,不难看出,YOLO作为一个one-stage目标检测算法,在速度和准确度上都有杰出的表现.

至此,本文结束.

感恩能与大家在华为云遇见！希望能与大家一起在华为云社区共同成长.