The 2021 Paddle Hackathon ended perfectly, looking forward to meeting you in the next issue

"Hacking" cool? What is it like to "fight" with "hackers"? Are you in this "battle"?

On September 23, 2021, the PaddlePaddle Hackathon 2021 was officially launched, and it ended perfectly after 100+ days!

PaddlePaddle Hackathon 2021, hosted by Baidu Paddle and National Engineering Laboratory for Deep Learning Technology and Application, OpenVINO, MLFlow, KubeFlow, TVM and other open source projects, as well as Shanghai Baiyulan Open Source Open Research Institute, Mulan Open Source Community, Datawhale Open Source It is jointly produced by the community, etc. It is a programming activity in the field of deep learning for developers around the world. This event is divided into two parts: online task development and offline 48-hour Coding Party.

few figures to take you quickly understand this event : PaddlePadddle Hackathon Hacker marathon paddle fly line development tasks, 1800 + developers signed up, 297 teams claim the task completed more than 477 registration & PR submit , to fly paddle community contribution more than 60 PR and fit into the framework paddle fly in. The offline Coding Party was held in 7 offline venues across the country on December 18: Shanghai, Beijing, Chengdu, Haikou, Southwest University, China University of Mining and Technology, Wuhan University of Science and Technology and online. More than 500 developers signed up, and finally 34 Teams , 100+ developers , two days and one night , 48 hours nonstop Coding

Several keywords indicate the direction : The more than 60 tasks completed this time include tasks in the three special areas of PaddlePaddle, Paddle Family, and Paddle Friends. Task directions include basic API, high-level API, single test, PaddleNLP, Paddle2ONNX, PaddleDetection , PaddleClas, Paddle Quantum, OpenVINO, AgentMaker, etc.

The 34 projects of the offline Coding Party include cultural communication, game entertainment, agriculture, etc. Three awards selected by the on-site jury: 1 Dark Horse Award, 2 Geek Awards, and 3 Creative Awards are also active. It was released that night, and after 14 days of online voting, the Popularity Award has also been born.

Online Development Task Part

In this event, developers contributed a number of basic APIs and high-level APIs to the paddle framework, which enhanced the basic capabilities of the paddle framework, supplemented the classic CV model, and improved the coverage of the classic high-level API model. A single test example has been added to the framework API to ensure the correctness and stability of the framework.

On the paddle kit, the weight of PaddleNLP's transformers model has been enriched and its downstream task code has been improved, the richness of the PaddleDetection model has been improved, and the capabilities related to data loading, model prediction and model deployment have been further improved; The coverage of the operator enhances its ability to export the propeller model. In PaddleClas, a solid application case is provided for the lightweight general-purpose image recognition system PP-ShiTu, which provides a very good foundation and ideas for unmanned retail visual intelligent solutions in the new retail industry, especially for solving multi-category solutions. , small samples, high similarity and frequent updates of special image recognition scenarios provide a reference for the pain points.

In the flying paddle family, developers have expanded the quantum circuit qubits of the measuring paddle, improved the Hamiltonian and other related functions, and also added functions such as density matrix visualization to the measuring paddle to improve the user experience of the measuring paddle.

In terms of ecological cooperation, developers' contributions to OpenVINO have improved the coverage of the PaddlePaddle operator in the front-end of OpenVINO, and enhanced OpenVINO's ability to support the paddle model.

In the AgentMaker task, developers have carried out various types of operator enhancements to PaTTA, the flying paddle model test enhancement tool, and added more than 5 real and effective enhancement strategies for model enhancement. AgentOCR supports JavaScript and Rubick, which is also the first relatively complete OCR tool of Rubick version in the Paddle Ecology. It complements the Paddle and AgentMaker ecology and effectively promotes the development of the community.

Offline Coding Party Part

Dark Horse Award

Project name: "Hello, Sinology"

project introduction: This project is a T2V national style literature special effect model based on the two major capabilities of ERNIE-GEN text generation and PaddleGAN action transfer. The implementation method is as follows: input a piece of ordinary vernacular text, after the special effect conversion of ERNIE-GEN national style literature, a classical Chinese text with everyone's style can be generated. Combine this classical Chinese text with the PaddleGAN action transfer model, that is, the ancients can "recite" this text. Effect. For example, enter a paragraph of "It seems that there is a weakness suddenly, and there is also an armor suddenly." After being processed by the two models, it can become an interesting video demonstration of an ancient man reciting "If there is meat, there is also fur". It can present Chinese traditional culture in a more interesting and fun way and promote Chinese culture.

project link:

https://aistudio.baidu.com/aistudio/projectdetail/3256082

Geek Awards

Project name: Create a new way of playing Super Mario based on gesture and voice

Project Introduction: This project explores the movement of a human-computer interaction in games such as Super Mario. Since this hackathon has certain time requirements, in order to achieve a better model effect in a short time, the auxiliary control model based on sound classification is very important. In the end, the project uses PaddleDetection's latest ultra-lightweight PP-TinyPose model to achieve attitude classification, and based on the high-level API of the flying paddle framework to quickly build a voice classification Pipeline, successfully built a human-computer interaction based on action pose and sound classification. The model is easy to use, easy to deploy, and reduces stress and weight loss.

project link:

https://aistudio.baidu.com/aistudio/projectdetail/3198421

Project Name: Fatigue Driving Detection System

Project introduction: This project is a fatigue driving detection system based on PaddleDetection, which can infer the fatigue state by monitoring the driver's eye and mouth movements, and can detect drivers who may have hidden dangers of fatigue by local voice and other methods. At the same time, the background management personnel can also receive information such as driver fatigue warning, so as to avoid the occurrence of safety accidents as much as possible. In terms of deployment, the project uses the PyQT graphical interface + PaddleLite high-performance deployment capabilities to provide a simple graphical detection system, which is easy to use, fast and reliable, and has high social value.

project link:

https://aistudio.baidu.com/aistudio/projectdetail/3260965

Creative Award

Project name: based intelligent pest inspection robot

Project introduction: This project uses the built-in PP-YOLOv2 algorithm in PaddleDetection and combines the self-collected pest data set to realize the intelligent pest inspection model. In order to more easily apply the project to the real forestry environment, the project also combines the Internet of Things technology to deploy the model in the intelligent inspection robot designed for common forestry pests, so as to realize the implementation of real-time pest detection suggestions, and at the same time, it will also automatically carry out corresponding The prevention and killing work has a wide range of application scenarios and high social value.

project link:

https://aistudio.baidu.com/aistudio/projectdetail/3229725

Project Name: You You You You Dance

project introduction: uses a camera to build a "dance machine", do you believe it? The project builds a set of video-based man-machine with the help of three deep learning models: a drum detection model for obtaining information on drum timing and frame number, Padoodle for generating template villains, and three deep learning models for key point detection based on hrnet in PaddleDetection. Action interaction system. When using, the experiencer only needs to turn on the camera and select a favorite dancing video. After adjusting the relevant parameters, a template villain with a drum beat with action instructions can be generated on the screen. After completing the corresponding action, the dance score can be obtained. The process is simple. Very playable.

project link:

https://aistudio.baidu.com/aistudio/projectdetail/3260973

Project name: Flying Paddle's first girl group photo capsule machine

project introduction: This project realizes the first interactive application of Fei Paddle's small card gashapon machine through PaddleX+EasyEdge+GUI and other technologies. Put the photo cards (that is, limited edition paper photos) of 4 different girl group members into the gashapon machine, click the "Draw Gacha" button to start the gashapon machine, and you can randomly extract the pre-placed photocards in the gashapon machine. The whole process has been dynamically displayed. The idea is quickly realized through two days and one night, and the deployment is completed, and it is entertaining. (Creative + Hardware Deployment)

project link:

https://aistudio.baidu.com/aistudio/projectdetail/3179285

All teams that complete the online development tasks and the winning teams can obtain the [Open Source Contribution Certificate] from the Fei Pao open source community, as well as corresponding bonuses and customized prizes, and can also obtain the Baidu Recruitment Green Channel.

So far, the PaddlePaddle Hackathon 2021 has come to an end.

PS: what, you said it was a pity that you missed this event? Don't worry, the 2022 Paddle Hackathon will be launched soon, and developers are welcome to participate in the Paddle open source community and build with us.

Online task completion link: https://github.com/PaddlePaddle/Paddle/issues/35940