Human beings have various means of communication, including language, text, music, video, etc. Some are close to life, and some are full of artistic sense. However, in the era of artificial intelligence, code, as a special form of communication, is becoming more and more important. It enables efficient and convenient communication between people and machines, thus injecting vitality into the development of science and technology. Open source is the secret to making this communication richer and deeper.
In addition to excellent developers and creative ideas, good open source also requires the support of platforms and resources. The Open Atom Open Source Contest provides opportunities and platforms for all developers who love open source!
Open Atom Open Source Contest
The Open Atom Open Source Foundation launched the first "Open Atom Open Source Competition", aiming to unite resources from open source organizations, enterprises and institutions, colleges and universities, scientific research institutes, industry organizations, investment and financing institutions, and give full play to the upstream and downstream of the industrial chain ecology. Collaborative ability, based on the principles of open source sharing, joint construction and co-governance.
The competition builds a platform for cutting-edge technology competition, selection of outstanding talents, display of innovative achievements, guidance for business transformation, and docking exchange and cooperation in the global open source field, widely disseminates open source culture, popularizes open source knowledge, promotes open source projects, and improves open source skills. and sustainable development to provide impetus and support.
Baidu Paddle participated in the first "Open Atom Open Source Contest". The competition questions were all based on the open source framework of Paddle, combined with the current popular AIGC and advanced artificial intelligence technology, scientific computing AI for Science topics, and co-constructed with all developers to develop an industry-leading The artificial intelligence model empowers and accelerates traditional scientific computing methods.
The competition escorts all developers who love open source. The organizing committee of the competition builds an AtomGit code hosting and collaboration platform to facilitate the co-creation of open source results. From the release of the competition questions to the review process, the whole process is controllable, and the code hosting and collaboration can be done in one stop, providing you with a safe, free and secure open source environment. The prize money of the competition is rich, the total prize money of the flying paddle track is 500,000 yuan, and the prize money of each question ranges from 60,000 to 200,000 yuan. The prize money will be directly distributed to the winning team by the foundation!
Baidu flying paddle competition questions
As the earliest open source deep learning framework in China, PaddlePaddle deeply practices the concept of open source, embraces the community openly, attaches great importance to ecological construction, and grows together with developers and ecological partners. It has become the industry-level deep learning with the most comprehensive competitiveness in China platform. The rapid development of the paddle platform is inseparable from open source and openness.
Based on industry practice and open source project construction, Baidu Flying Paddle released 6 major open source co-construction competition questions, including: reproducing academic frontier cross-modal large models, using AIGC to improve the accuracy of classic models, offline distillation of SSLD models, and scientific computing methods based on PINN. For CFD discrete grid optimization , etc., contestants are required to provide complete solutions and codes, and contribute to the designated code warehouse. This competition encourages contestants to understand and participate in deep learning open source projects, and contribute to the construction of China's open source ecosystem. Flying Paddle will provide technical guidance for contestants throughout the competition to make your work more comprehensive and perfect.
The specific contest questions are set as follows, come and see if there are any contest questions you are interested in!
Question 1: Use AIGC for data expansion to improve the accuracy of the model on ImageNet
Basic description: At this stage, large models such as stable diffusion can already generate relatively realistic data. At present, some researchers in the academic community have used large models to generate data to expand training data and improve the accuracy of the model. The main content of this competition is to use large models such as stable diffusion to expand ImageNet data and improve model accuracy.
bonus:
- First prize: 30,000 yuan (1 winner)
- Second prize: 20,000 yuan (1 winner)
- Third prize: 10,000 yuan (1 winner)
Question 2: Use AIGC for data expansion to improve the accuracy of the model on COCO
Basic description: At this stage, large models such as stable diffusion can already generate relatively realistic data. At present, some researchers in the academic community have used large models to generate data to expand training data and improve the accuracy of the model. The main content of this competition is to use large models such as stable diffusion to expand COCO data and improve model accuracy.
bonus:
- First prize: 30,000 yuan (1 winner)
- Second prize: 20,000 yuan (1 winner)
- Third prize: 10,000 yuan (1 winner)
Question 3: Offline SSLD Distillation Teacher
Basic description: PaddleClas provides a simple and efficient SSLD distillation scheme, which can greatly improve the accuracy of the model. However, during the distillation process, each picture needs to pass through the teacher model, resulting in a relatively long overall training time. The main content of this competition is to offline the SSLD distillation Teacher in PaddleClas, that is, the teacher model stores the predicted labels of the data, and the student model is used to learn the output of the offline teacher model. In the end, the offline accuracy of the SSLD Distillation Teacher was equal to the standard SSLD accuracy, and the training speed was doubled. bonus:
- First prize: 30,000 yuan (1 winner)
- Second prize: 20,000 yuan (1 winner)
- Third prize: 10,000 yuan (1 winner)
Competition Question 4: DiffusionRig, a paper on reappearing face control
Basic description: Paddle currently provides diffusion-related capabilities. DiffusionRig is superior to existing methods in terms of identity preservation and photo fidelity, and can further enrich paddle diffusion application extensions. bonus:
- First prize: 30,000 yuan (1 winner)
- Second prize: 20,000 yuan (1 winner)
- Third prize: 10,000 yuan (1 winner)
Topic 5: Null-text Inversion for Reproducible Image Editing Thesis
Basic description: Image editing has very important application value in the field of AIGC. How to realize image editing with high efficiency and high quality has become one of the current technical difficulties. This paper realizes high-fidelity image editing capabilities based on the prompt-to-prompt method. bonus:
- First prize: 30,000 yuan (1 winner)
- Second prize: 20,000 yuan (1 winner)
- Third prize: 10,000 yuan (1 winner)
Question 6: CFD Discrete Grid Optimization Based on PINN
Basic description: Computational fluid dynamics (CFD) is one of the important technologies in the field of fluid mechanics in the 21st century. It uses numerical methods to solve the control equations of fluid mechanics in a computer, so that the flow of the flow field can be predicted. In recent years, the method based on physical information constrained neural network (PINN) has become more popular in the field of fluid mechanics. The characteristic of PINN is that the loss function (Loss Function) of the neuron network contains the control equation of fluid mechanics, so that the neuron network can replace the CFD solution process. However, the accuracy of PINN solution still needs to be improved.
The purpose of this topic is to conduct CFD discrete grid optimization based on PINN, that is, instead of using PINN to replace the high-precision CFD solver, it uses PINN's fast solution speed and certain solution accuracy to speed up the search for the optimal solution. Discreteize the mesh, thereby reducing the speed of the overall CFD process. Essentially, this topic is to test a hypothesis that a high-quality discrete grid also corresponds to a relatively high-precision PINN. This topic is of high value to the CFD industry. If it can be realized, it can be extended to the entire CFD industry in the future. Therefore, the grid file selected for this topic needs to be representative of the industry, including at least three types of scenarios: car aerodynamic exterior design (select a model); wing aerodynamic exterior design (select 1-2 NACA series wing profile); indoor airflow structure (choose a common resident room). The initial grid file is provided by the subject issuer.
bonus:
- First prize: 100,000 yuan (1 winner)
- Second prize: 70,000 yuan (1 winner)
- Third prize: 30,000 yuan (1 winner)
This type of competition focuses on the function and algorithm realization of the flying paddle project itself. Contestants choose any one of the above topics, design and implement the scheme, and finally submit it to the corresponding atomgit code warehouse. After being accepted by the research and development instructor, they can get bonuses.
Top open source projects, senior R & D guidance
Are you still not excited about skill improvement, rich resume, and high bonus?
Act quickly!