Hit the industry directly! The first model selection tool in the industry was launched by the flying propeller

With the development of technology ,AI Algorithms have gradually penetrated into all walks of life .AI The algorithm is efficient , But in real projects , Developers often face many complex application scenarios . There are hundreds of open source algorithms in the industry , Hardware is becoming more and more diverse . How to do it in a specific scenario , Quickly select the most suitable AI Algorithm and matching 、 The most cost-effective hardware , It is a big pain point for industrial developers .

In order to quickly solve the problem of model and hardware selection , Enable developers to do more quickly AI Project landing , The propeller team launched 「 Scene model selection tool 」. It considers the real industrial landing demands of users , It also integrates the long-term industrial practice experience accumulated by the propeller team . You can recommend appropriate models for users' real scene needs 、 Optimization strategy and hardware combination . For typical scenarios , It also recommends relevant industrial practice examples .

Click to read the original text

Get product experience address

https://www.paddlepaddle.org.cn/smrt

You can also visit the official website of the propeller directly —— model base —— Use the industry model selection tool .

It's not difficult for a careful partner to find , There is also a very intuitive data analysis function in the model selection tool , Users only need to upload their own annotation files （ The original drawing is not required ）, Tools can analyze data characteristics , Provide model selection and optimization strategies . The current model selection tool supports Labelme、 Elf label 、labelImg And other mainstream annotation software , Support at the same time voc data format 、coco Data format and seg（ Semantic segmentation ） data format .

Such a good tool , How to use it more efficiently ？ We use an actual case of industrial quality inspection , Explain in detail for everyone .

Case explanation

In the defect detection project of a steel plant , The user uses the linear array camera to detect the defects on the steel plate , Control by encoder , Every time 4000 Line generates a sheet 4096*4000 Size image . According to the project operation requirements , It is necessary to accurately calculate the area of the defect , At the same time, it needs to be in 2080Ti On the video card 200ms Complete defect detection .

So how to determine the final model through the model selection tool ？

Step one ： Determine the possible cropping scheme according to the image size

Generally, the image size obtained by linear array camera is large , But in practice , They are often cut into small sizes for training and prediction , And how many pieces are cut , What is the net size of each picture , It determines whether the final model can complete the identification task within the specified time .

Step two ： Query through the model selection tool

Choose the right model

Because the project needs to accurately identify the area of defects , Therefore, the project selects a series of image segmentation models . Under the specified time conditions , According to the number of segmented images, the maximum prediction time of each image can be calculated , Query the corresponding model in according to the model selection Input-size（ The size of the image after cropping ） Combination meeting the prediction time requirements under the same conditions , choice Target-size（ Actual network access size ） The largest set of values , Finally, the appropriate model combination is selected .

remarks ： At present, the scheme mainly considers the prediction of concatenation sequence after image segmentation .

Step three ： Based on the selected model

Model optimization

Label the image according to the final cutting size of the model , According to the data analysis function in the model selection tool , Further analyze the characteristics of the data , In view of the unbalanced distribution of its samples , Deep optimization by updating the loss function .

The project is to select the corresponding model based on the known recommended hardware , If the user needs to select the hardware , The model selection tool also supports the automatic recommendation of matching hardware devices according to the time entered by the user . As shown in the case below , The time reserved for model reasoning is 50-100ms, The user enters the corresponding condition , You can get the recommendations of different hardware and the specific model running time in this period .

At present, the flying oar team is based on the needs of users' landing deployment , Offer based on 1660Ti、1080 Ti、2080 Ti、3090 Wait for a variety of chips in TensorRT FP32 Test data for , In the future, more cloud side deployment hardware will be supported , So as to better meet the landing needs of users .

Surprise benefits

It's so easy to use 「 Scene model selection tool 」, What are you waiting for ？

• Welcome to read the original , Experience use ：

    https://www.paddlepaddle.org.cn/smrt

• Welcome to Join the user communication group , Just join the group Get intelligent manufacturing 、 A big gift package for smart city courses .

Focus on 【 Flying propeller PaddlePaddle】 official account

Get more technical content ~