The sixth year of China's open source deep learning framework: the first domestic comprehensive share of flying propellers, with more than 4 million developers

Not long ago, the news that ZTE ended its 5-year compliance inspection period caused a lot of waves in the science and technology circle.

Under such a special time node, the problem of self-improvement in terms of "lack of core and little soul" has once again become the focus of attention of the whole society.

The chip represents the underlying computing power; the soul refers to the system.

Today, AI is seen as a battleground for high-tech competitions, and its core system is the deep learning framework.

In addition to tensorFlow and PyTorch, the two mainstream deep learning frameworks in the world, what is the development of China's open source framework?

Who are the specific players who are admitted? What is the technical strength behind it? Can you be self-reliant and not repeat the mistake of "lack of core"?

Today, we're going to try to find answers to these questions.

Domestic open source framework situation

At the beginning of the problem, let's take a disk of what mainstream domestic open source frameworks are on the market today.

Baidu Paddle Paddle, the pioneer of deep learning open source frameworks, was first released to the public in 2016.

Then in 2020, the domestic open source framework ushered in the first wave of concentrated outbreaks.

Unicorn Megvii took out the industrial-grade deep learning framework MegEngine, and the first-class technology OneFlow and Huawei Sheng Si (MindSpore) also appeared in the same year.

In terms of academics, Tsinghua University has open sourced a deep learning framework diagram (Jittor) that supports instant compilation.

Obviously, in the past few years, "open source" and "AI bottom" have become development strategies that domestic AI manufacturers attach great importance to.

The reasons behind this can be roughly boiled down to two points.

First, in the context of the rapid development of deep learning, the traditional industry has turned to intelligence, and the injection of AI is one of the key factors.

From a small face recognition to a city's intelligent management, deep learning has penetrated into our daily lives and become an important link for the rapid development of society.

Second, although TensorFlow and PyTorch have been very mature in technology development, changes in the external environment have made the mainland have the underlying capabilities of independent innovation of AI that have become a just need at present, which has also brought soil for the development of domestic deep learning open source frameworks.

In fact, since 2016, after years of accumulation, precipitation and exploration of domestic deep learning open source frameworks, some achievements have initially emerged.

According to IDC China's publicly released deep learning open source framework market research report, as of the first half of 2021, TensorFLow, PyTorch and Baidu Flying Propeller have become the most frequently used open source frameworks in China;

Baidu Flying Propeller has gathered 4.06 million developers, served 157,000 enterprises and institutions, and developed 476,000 models.

In terms of market share, Baidu Feipao's comprehensive market share in China has surpassed TensorFlow and PyTorch, ranking first in China.

So, how did it get to where it is today? Is the experience behind it worth learning?

As a domestic head goose, taking Baidu Flying Oar as an example, it may be able to clarify a road for the development of China's self-developed open source framework.

How do autonomous technology systems need to be crafted?

At present, we are already in the third wave of AI with deep learning as the core, and emerging technologies have ushered in a period of concentrated outbreak. In such a competitive environment, how to open up your own place?

Aiming at the bottom of the core of technology, the big manufacturers know this well.

As mentioned earlier, the framework is regarded as the core soul of deep learning and plays a pivotal role in promoting artificial intelligence into industrial mass production.

In this context, how to open up a place for its own deep learning open source framework?

Referring to the flying propeller, connecting the dots into a line to summarize and summarize, it probably requires 4 aspects of effort:

Technology accumulation scenario application industry supply and demand developer ecology

Indispensable.

First of all, start in time and seize the position.

The most typical case here is Google TensorFlow.

In 2015, TensorFlow was a pioneer and quickly won over developers in the industrial world, thus establishing its position as one of the world's two major frameworks.

Baidu Flying Paddle stood on the starting line at almost the same time.

In 2016, Baidu Flying Propeller first opened its doors on GitHub under the name paddlePaddle, and provided bilingual technical documentation in Chinese and English.

If you count from the preparatory and research and development period, the starting point of Baidu flying propeller is even earlier, dating back to 2010-2013.

Starting early and making early efforts, the advantages brought by it are also very obvious - there can be more sufficient time to accumulate the underlying technology.

At present, the number of Baidu AI patent applications has exceeded 13,000, ranking first in China for four consecutive years, of which the number of deep learning patents ranks first in the world.

Holding the patented technology in its own hands means that China's deep learning technology can be more autonomous and self-reliant; at the same time, the earlier accumulation of technology also provides an opportunity for Chinese teams to participate in the establishment of industry standards.

Today, the core framework of Baidu Flying Propeller runs through the three links of development, training, and reasoning deployment, and the basic model library covers the CV, NLP, recommendation, voice, and knowledge-enhanced Wenxin big model.

Second, deep learning frameworks need to be able to solve real-world problems in the industry.

Deep learning frameworks from the industrial world have an innate advantage in understanding industry scenarios.

In addition, the flying oar was born in China, which can better understand the actual needs of Chinese enterprises and provide a rich experience reference for exploring overseas markets.

In the past few years, Baidu Feipao has officially released more than 500 industry-level open source algorithm models, and released 13 industry-grade PP series models with balanced accuracy and performance, covering more than 20 industries such as industry, agriculture, transportation, and scientific computing.

Here we can look at some specific examples.

In terms of agricultural production, Baidu Feipao and BOE Houji have built a smart hydroponic plant factory;

In the field of coal mine production, Huaxia Xintian Robot Company developed a coal transmission tape intelligent inspection robot based on paddleDetection, a target detection tool kit with flying propellers.

Of course, there are more cutting-edge application scenarios.

Not long ago, Baidu's biocomputing research appeared in the Nature sub-journal, and the underlying technical support for this achievement came from Baidu Flying Propeller.

Based on the flying propeller, Baidu has developed a tool component propeller (Paddle Helix) that can be used for biocomputing, covering areas such as drug development, vaccine design and precision medicine.

There is also the invisible AI coach behind the Chinese diving dream team at the Tokyo Olympic Games, the first cloud 3D+ AI diving training system in China, and the underlying capability also comes from flying paddles.

Third, work with upstream and downstream to jointly promote independent innovation.

At the practical application level, enterprises will always bring additional manpower and material costs due to the mismatch between the framework and hardware.

In this regard, Baidu Flying Propeller has achieved the first domestic chip adaptation, and it is also one of the three major support frameworks of NVIDIA - and the only Chinese framework that is deeply adapted.

In addition to the self-developed Kunlun core, Feipao has completed the adaptation and optimization of 31 kinds of chips with 22 domestic and foreign hardware manufacturers, including Intel and Nvidia.

It is worth mentioning that the adaptation of the deep learning open source framework to the underlying hardware will, in turn, open up the use scenarios of domestic hardware and promote the development of domestic hardware.

Fourth, whether the deep learning open source framework can develop in the long run, a good developer ecosystem is also one of the keys.

Someone uses it, and the more it is used, the better, which can be seen as a criterion for judging a deep learning framework.

Flying oars are already taking shape in this regard.

Among them, Baidu Feipao's influence in the open source community ranks first in China, and the total star on GitHub ranks third in the world and first in China.

According to the "2021 China Open Source Annual Report", in the top 30 GitHub China project activity in 2021, flying paddle occupies 5 projects, of which the flying propeller framework ranks first.

How does China's self-developed open source framework break through?

In summary, Baidu Flying Propeller provides some thoughts for the development of China's deep learning framework.

Nowadays, the global deep learning framework "PPT" pattern has begun to appear, and Baidu Flying Propeller PaddlePaddle has begun to confront TensorFlow and PyTorch.

But it is undeniable that there is still a long way to go, and the space for China's deep learning framework to improve is still very broad.

Recently, Baidu has also put forward its own insights.

Dr. Ma Yanjun, General Manager of Baidu AI Technology Ecosystem, said that there are three key points in the current development of China's deep learning framework:

Technical strength functional experience ecological scale

Technical strength, it is not difficult to understand.

The source of technological innovation is, in the final analysis, talents. At present, the mainland is still insufficient in the reserve of AI low-level technical talents.

Flying propellers are also cultivating talents in this area while developing. At the same time, it also creates an AI Studio learning and practical training community, so that more people interested in AI have the opportunity to get started, advanced and rapidly improved.

Feipao also "industry-education integration" with colleges and universities: cooperating with artificial intelligence-related teaching materials and providing artificial intelligence education resources, more than 3,000 AI professional teachers in more than 700 colleges and universities have benefited from the deep learning teacher training held by Feipao.

Second, in terms of functional experience.

China is the country with the most complete industrial chain in the world, but at the same time, the industrial system is also quite complex, especially for small and medium-sized enterprises, how to quickly transform to intelligence has become a key issue in the country and the industry.

Then, how to enable professionals from all walks of life to use AI smoothly even without AI expertise and background is a problem that the AI industry needs to think about from the underlying technology.

The low threshold seems to be what everyone wants.

This is also one of the characteristics of Baidu flying paddle.

Neither TensorFlow nor PyTorch has put much effort into ease of use and is not very beginner-friendly. And Baidu Flying Propeller just makes up for this market pain point.

Finally, it comes back to the developer ecosystem.

Dr. Ma Yanjun mentioned that Baidu Feipao has always maintained close contact with developers, such as problems can be directly feedback to internal staff through the QQ exchange group.

At the same time, Baidu Feipao also often carries out online live welfare courses, after all, self-study is also a necessary self-cultivation for programmers.

It is not difficult to see from these actions that, unlike Google and Meta's route of "barbaric growth" of open source frameworks, Baidu Flying Propeller not only provides developers with a useful underlying framework, but also invests a lot of manpower and material resources to create a more friendly and more applicable ecology.

Finally, let's look at the overall artificial intelligence industry environment.

In April last year, at the first high-end summit of jinan's national artificial intelligence innovation and application pilot zone, Pan Yunhe, an academician of the Chinese Academy of Engineering, pointed out:

The pilot area of artificial intelligence application should encourage the use of China's own platform to promote the realization of autonomous control of Chinese intelligent work.

On the other hand, the IDC report notes that security is beginning to become one of the considerations for developers to use open source frameworks.

Thankfully, Academician Pan Yunhe said that in this regard, China has slowly begun to form its own advantages, and Baidu Flying Oar is one of the best proofs.

As Dr. Ma Yanjun said:

Although the deep learning framework belongs to the high-input, long-term, and ecological competition, it has received strategic support from the state and enterprises, and is the key to opening the next AI era.

The bench sits cold for ten years, only for the spark to burn.

Ten years of technology investment, Baidu Flying Oar stands firm in the Chinese market, the future challenges are still arduous, challenges always coexist with opportunities, I believe that Baidu, which has a belief in technology, can continue to promote China's artificial intelligence to walk in the forefront of the world.