EasyEdge is Baidu calculation model based on research and development to generate end Paddle Mobile platform, deep learning can help developers to rapidly deploy self-built model to the device. Just upload model, the fastest 2 minutes to generate end computing model and get the SDK. This article describes EasyEdge generating client computing model, and the process is integrated into the phone of APP.
Using models: SSD-based model of MobileNet
SSD using a convolutional neural network "end" test: the original image is input, the output of the detection result, feature extraction without using external tools or processes, the candidate block generation.
The official Open Source: https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/ssd
Using data sets: PASCAL VOC
PASCAL VOC Challenge is a benchmark classification and detection of visual objects, provides a standard set of image annotation data and standard detection algorithms and learning performance evaluation system.
PASCAL VOC set includes 20 pictures directory:
Humans; animals (birds, cats, cows, dogs, horses, sheep);
transportation (airplanes, bicycles, boats, buses, cars, motorcycles, trains);
indoor (bottles, chairs, tables, potted plants, sofa ,TV).
These categories are listed in label_list file, but the file more than one category, is the background (background)
Get training in AI Studio model and parameter:
1 EasyEdge calculation model generation process ends
Enter EasyEdge Home: ai.baidu.com/easyedge, click "Use now" to use the platform.
1.1 Upload depth learning model
In the directory [upload] in the original model:
1) Fill the model name
2) Select the model type, model framework, network model
3) sequentially uploading network profile, the network structure of the file, the tag model, other profiles (dashed line at the left click view upload request)
Correspondence is as follows:
Network profile: mobilenet-ssd-final-model
Network file structure: mobilenet-ssd-final-params
Model Tags: label_list
4) a custom model described grouping function and
5) submit to upload
After the model is successfully uploaded, you can see the model in [upload] in my model.
1.2 generating client computing model
successfully uploaded model, generated in the directory [end] of the model:
1) Select Model
2) select and add a "chip / Hardware", a combination of "operating system", adding support for various combinations
3) For SDK support from online mix, the need to check the consent EasyEdge encryption deployed in the cloud model
4) start generating model
Generation state of the model in the model [I] Click model name to view end computing
Get off SDK 1.3
If a version of the client computing model has been successfully generated, you can click on the right side of the operation "Get SDK", to jump to the list] [SDK download experience or formal SDK Demo
Offline activation SDK 1.4
1) In the list [SDK] After downloading the SDK page, click on "Get serial number to activate" to open the serial number [list]
2) Select the authentication method: For the development of APP selectable by product line authentication, device authentication or basis having
3) Click on the "Add Serial Number", fill in the relevant information
Android operating system or iOS SDK can choose to activate by product line, the package name and serial number (Package Name / Bundle ID) Binding:
4) prior to activation of the new number is used to download the SDK, developers reference document:
Android system: https://ai.baidu.com/docs#/EasyEdge-Android-SDK/top
2 SDK integrated into the phone APP
Development Platform: Android Studio 2.3
Android virtual machine: Genymotion 2.7.2
Test Platform: Android 9.0
Alternatively Application Serial No. 2.1
2.2 add-dependent:
Run as a virtual machine reported the following error:
Solution:
At the same a build.gradle (app) page, android {} is added:
splits {
help {
enable true
reset()
include 'x86','x86_64'
universalApk true
}
}
2.3 手机测试效果
3 意见建议
整体效果不错,优势在于生成模型非常快捷,集成也很方便。当然也有需要改进的地方,本次采用的物体检测SSD模型,SDK大小在25M,如生成的SDK能够进一步压缩,进而能减少APP的大小。