Author | Aspirin
Produced | AI Technology Base Camp (ID: rgznai100)
Seal map | CSDN paid download from Visual China
First shot: Occupy the highlands
From PaddlePaddle to flying paddle
In 2016, Baidu PaddlePaddle fired the first shot of the open source domestic AI framework.
As one of the earliest companies to research deep learning technology, Baidu established the Deep Learning Research Institute as early as 2013, and officially open-sourced its deep learning framework in 2016. PaddlePaddle is also responsible for Baidu's important mission to seize the heights of the artificial intelligence era.
In July 2018, Robin Li called out the slogan of "Everyone Can AI" at the Baidu AI Developers Conference. In fact, this slogan must be followed by a restriction-"Through Baidu's AI platform". This also means that PaddlePaddle is one of the breakthrough points in this battle.
In April 2019, Wang Haifeng, then senior vice president of Baidu (now Baidu CTO), set the tone for the deep learning framework PaddlePaddle's strategic position within Baidu at the Wave Summit Deep Learning Developer Summit. PaddlePaddle released the Chinese name "Flying Paddle" , and began to emphasize that he understands Chinese developers better and focuses more on the industrial production and deployment of deep learning models.
As AI has become a national strategy, under the call of autonomy and control, the deep learning framework as the bottom layer has attracted more attention. Compared with its domestic counterparts, Baidu has taken this step earlier and has a first-mover advantage.
After 4 years of development, Feida has 62 Repositories on GitHub, harvesting a total of 47,000+ Stars, and becoming a leader in domestic open source projects.
The 2020 "China AI Application Developer Report" jointly released by CSDN and the workplace social platform Maimai also shows that Baidu Feida has become the most popular domestic deep learning open source framework for Chinese developers.
Battle Breakthrough: Research and Development
Create four "cheats"
Since Theano was born in 2007, after more than ten years of development, deep learning technology and applications have advanced by leaps and bounds, and the deep learning framework is also in a process of continuous iteration and evolution .
As China's first open source, open, and fully functional industrial-grade deep learning platform, Feiye must have its own unique advantages if it wants to maintain its position . So how do you define the advantages of flying paddles? In general, the flying paddle has the following 4 leading technologies.
First, it supports the programming of dynamic graphs and static graphs at the same time, which can balance ease of use and efficiency. In the entire development process, algorithm engineers need not only convenient interfaces, but also efficient support of the underlying platform. The flying paddle platform combines the ease of use of dynamic graphs with the high performance of static graphs, allowing developers to take into account the advantages of both. For developers, this greatly reduces the cost and complexity of writing programs.
Second, for large-scale industrialization scenarios, flying paddles provide large-scale distributed training capabilities, and can respond freely in real industrial scenarios. The flying paddle platform has introduced real-time update capabilities and training capabilities that support trillion-scale parameter models. Model parallel training can be performed when facing large classification tasks, and data parallel training can also be supported at the training level.
Third, end-to-end deployment is a very critical link in the application of deep learning. Flying Oar provides a very complete end-to-end deployment capability that supports various hardware, which can make the process of reasoning and prediction for developers smooth enough. In terms of the size of the model deployed on the mobile terminal, the flying paddle can be reduced to a very small size, and many developers can better use the model on the terminal equipment of the corresponding scene.
Fourth, the flying paddle has a large number of models that have been deposited in industrial practice and provide official support to ensure that the application effects of developers are the best and truly reliable. It is reported that 140+ models are currently open source on the flying paddle platform, and there are also industrial-grade pre-training models. Based on the pre-training model, developers can use it as long as they perform small data transfer learning for their own scenarios.
In addition, the flying paddle is rapidly iterating. Recently, the flying paddle has just undergone a comprehensive upgrade in the dynamic graph and API system.
After the upgrade, the flying paddle will provide a more complete and better performance of the dynamic graph function, as well as a more comprehensive dynamic graph to static graph operation function, to achieve "dynamic and static unification" in development. The advantage of "Unification of Dynamics and Statics" is that developers can balance efficiency and flexibility. In terms of dynamics and statics, the flying paddle now supports complete grammar coverage and tasks with control flow, and only a decorator can be used to trigger operations, and a unified interface can enable automatic adaptation of model saving and loading. In terms of dynamic images, the flying paddle provides a more concise and flexible mixed precision training interface than static images, and achieves mixed precision and quantitative training effects comparable to static images.
In addition, the flying paddle has formed a new API system that fits the user's habits, the directory structure is clearer and more user-friendly; in order to reduce the workload of developers, the flying paddle provides models and tool packages that include vision, NLP and other fields The high-level API and the most classic model structure; in terms of compatibility, the flying paddle is compatible with the 1.X version of the API, and is equipped with a complete tutorial to help users smoothly upgrade to the new version of the API and get a better experience.
Moreover, the flying paddle can be adapted to 22 chip models, covering 15 hardware manufacturers such as Intel, Nvidia, Rockchip, and Cambrian, and the support for domestic hardware is the first in the industry.
In order to further realize the combination of software and hardware and accelerate the development of AI, Baidu has also launched two generations of Kunlun chips. At present, Baidu Kunlun chip generation has been deployed on a large scale, and 20,000 chips have been deployed on Baidu search engine and cloud computing tasks. According to reports, in deployment scenarios such as Baidu's search engine, the performance of Baidu's Kunlun chip is 1.5-3 times higher than that of Nvidia's T4 GPU under different models.
In the Baidu World 2020 just past, Ouyang Jian, general manager of Baidu Smart Chips, also pre-released the first domestic cloud general-purpose AI processor "Baidu Kunlun 2" , which uses a more advanced 7nm process and can perform up to 3 times higher than the previous generation. , Is expected to be mass-produced early next year.
Hundred Years of Tree People: Cultivating the Ecology
"Whampoa Military Academy" for AI talents
In order to expand its ecosystem and attract more developers, Baidu is also continuing to build its own AI ecosystem, hoping to cooperate with more industries for a win-win situation and realize Baidu's long-term plan for AI ecological cooperation.
In addition to AI technology, compound talents who understand both application scenarios and AI technology are also the key to the AI industrialization process.
On January 19, 2019, Baidu announced the establishment of "Whampoa College" to launch a deep learning architect training program. And borrowed from the Whampoa Military Academy's gate couplet horizontally criticizing "Revolutionaries Come" and set the slogan as "Innovators Come". The purpose of Whampoa College is to "train the first batch of chief AI architects for the Chinese industry."
In September 2020, 65 students from the third phase of Baidu Whampoa College graduated. Since its establishment, it has exported 135 high-end talents in the AI industry to the industry.
In addition, Baidu has cooperated with various universities to offer artificial intelligence courses and trained thousands of AI professional teachers. At the same time, it also faces the industrial AI crowd and cultivates more compound talents. So far, Baidu has trained more than 1 million AI talents. In the next 5 years, Baidu also plans to cultivate 5 million AI talents.
People can make the best use of their talents, and things can make the best use of them. AI talents are becoming the core competitiveness of enterprises, and they are driving China's technological revolution and industrial intelligent development to continue to move forward at full speed.
Today, the flying paddle platform has been used by 2.3 million developers, served 90,000 companies, and created 310,000 models. In the past six months or so, there has been a 40% to 50% growth, and this is the power of ecology.
From individual developers, to university developers, to enterprise developers, everything is to expand the ecosystem of Flying Paddle and promote the implementation of Baidu's AI.
He Zong Lian Heng: Open Source and Open
Create acceleration of industrial upgrading
Open source is one of the important trends in the development of artificial intelligence.
In the 2020 Zhongguancun Forum held recently, Baidu CTO Wang Haifeng representation , human artificial intelligence of rapid development, has become a new round of technological revolution and industrial revolution of the important driving force. Among them, the open source innovation of artificial intelligence technology has played a big role. On the one hand, open source innovation can accelerate technological innovation and development, and continuously expand the boundaries of human cognition and technology; on the other hand , all developers can develop their own projects, products and businesses based on the results of open source and open source, which is more convenient, efficient, and Safely promote technological innovation and application.
In fact, since the deployment of artificial intelligence ten years ago, Baidu has been in the spirit of open source and open to open AI core technology for use by the whole society. As early as 2012, Baidu opened the translation API. Today, Baidu Translator's daily translation volume has reached 100 billion characters. In 2013, Baidu opened its voice platform again, and now it has more than 15 billion calls per day. In 2016, Baidu opened up the AI capabilities of Baidu's brain while open-sourced flying paddles, and now it has more than 1 trillion calls per day.
As open source software, Feida has been building an open source community. Currently, Flying Paddle has gathered 2.3 million developers, served more than 90,000 companies, and created more than 310,000 models. The Flying Paddle Developer PPDE program launched by the Flying Paddle also built an open source community, forming 22 fly paddle local communities and more than 100 college communities.
With the further promotion of open source, artificial intelligence has become an important driving force for a new round of technological revolution and industrial transformation. In October 2019, Robin Li, Chairman and CEO of Baidu, proposed at the Sixth World Internet Conference that we are ushering in the era of smart economy. Smart economy will give birth to many new business formats, and industrial intelligence will also become a new trend.
In the process of combining industry and AI to achieve industrial intelligence, the AI platform plays an increasingly important role, and the role of flying paddles is becoming more and more prominent. In business practice, an open source platform that is mature and reliable like flying propellers and has been verified by a large number of industrial-level applications not only continuously lowers the threshold for technology development and enterprise innovation, but also accelerates the intelligent upgrade of the entire industry.
For example, in the field of industrial security, Feizhuo cooperated with the State Grid and Shandong Xintong to create a smart grid inspection program, with an analysis accuracy rate of 90%, and an alarm response speed from hour to second; in the field of forestry inspection, Puzhou Aircraft Technology Based on the flying paddle deep learning technology, the company has specially developed a drone autonomous flight + application management platform for the forest inspection industry, which can realize large-scale forest autonomous patrol, fire monitoring, illegal invasion, and forest tree felling monitoring More than 10 illegal logging has been prevented...
Only when technology penetrates into all aspects of this society and is "perceived" by organizations or individuals, can AI truly democratize and empower thousands of industries. The artificial intelligence technology represented by Baidu Feida is committed to in-depth integration with the industry, helping developers to overcome industry difficulties in an unprecedented way, turning the past "impossible" into the intelligentization of China's industry today Reality. Every ordinary person can truly understand AI, use AI to understand the dawn of China's intelligent era.
Baidu has quietly broken through this protracted battle.
更多精彩推荐