2020, domestic AI open source framework "to prove safety" TensorFlow, PyTorch

" AI technology ecosystem on " Interviews section is initiated CSDN millions AI learned an important part in the initiative. AI ecology through interviews with experts, entrepreneurs, industry KOL, reflecting its thinking on industry practice to determine future trends, technology, and the experience of growing up.

In 2020, CSDN will conduct interviews 1000+ characters, form a series, which outlines AI ecology of the most influential figures maps and AI industry panorama. This article, "On Ecological AI technology" series of interviews.

AI million people learn you have taken! Participate article comments, the comments area selected for the message, get the value of 299 yuan, " 2020 AI developers million people congress " live online ticket.

Author | Just

Produced | AI technology base camp (ID: rgznai100)

2016, Baidu PaddlePaddle fired the first shot of the domestic AI open source framework.

But years later, when China's domestic open-source AI developers look back the past, the Fed also think that by the Tsinghua University, open view, Huawei in just seven days after another announced the AI framework open source end of March of the 2020.

In the March 28, Huawei completed on the Developers Conference 2020 of the rotating chairman Xu Zhijun pledged last August, it announced the open source MindSpore, which is a support deep learning training reasoning frame end edge cloud the whole scene.

Appropriated time four days ago, AI unicorn Kuang Science and Technology announced the open source Tianyuan (MegEngine), stressing that it is the integration of reasoning training, static and dynamic unity of industrial-grade deep learning framework.

March 20, University Department of Computer Graphics Laboratory AI framework open source design drawings (Jittor), which is China's first open-source AI academic framework, direct benchmarking PyTorch.

In just eight days, from China AI open source framework "you Changba me on stage" to open source AI frame Sword overlord who, this may be an important moment in the light of the high domestic deep learning framework for open source history.

An established fact that the depth of field of study is open-source framework, TensorFlow, PyTorch two basic points of the world, from the technical capacity to ecological construction, they give the AI ground to provide strong support to meet most of the companies in which you build your own AI applications. AI contrast, domestic open-source framework, even if Baidu PaddlePaddle has been open for four years, Tencent Angel, Ali X-DeepLearning and other open source framework have also, but the market impact and the use of scale, harder to compete with the former two.

That being the case, in the case of "open source" and a complete framework for ecological AI framework, Chinese companies and institutions Why again made another set of wheel technology, from what they value and significance of AI research framework is open source?

Technological innovation space

AI to build open-source framework is a big time-consuming project. If you have no idea beyond the existing framework of the mainstream, to build a set of wheels is no repeat of technological innovation, "cost" does not seem high, the skilled person attractive is not large.

Powerful doubt TensorFlow, PyTorch, but not without its technical capacity for improvement. With deep learning new technologies, and ever-increasing complexity of the task, due to the continued expansion of architectural design and other reasons, led to these mainstream framework of the system becomes complicated structure optimization and more difficult transplant, the actual performance of the new model is yet to be improved.

After the discovery of this space-optimized, Department of Computer Graphics Lab of Tsinghua University doctoral student hoping to have a team able to adapt to a variety of intelligent hardware while maintaining an efficient, strong versatility and depth learning platform scalability.

GitHub address: https: //github.com/Jittor/Jittor

FIG meter (Jittor) is one such product. It uses innovative meta-operator fusion and dynamic compiler technology, optimize memory depth, can effectively improve operational performance and versatility of the system, to ensure the implementation and optimization of separation, significantly increasing application development flexibility, scalability and portability. Compared with other major frameworks, it has a number of advanced features.

In fact, compared with Pytorch, by implementing ResNet, VGG, SSD, DeepLab, LSGAN plurality of other network model in FIG meter platform, and train speedometer reasoning FIG 10% -50% performance increase.

Although the design drawings fledgling yet to be continuing the iterative, but from the perspective of technological innovation, China academic research structure must have the courage and ability to try to make pioneering.

Similarly, the existence of high AI framework development threshold, high operating costs, difficult issues such as the deployment need to solve, Huawei is also developing a major reason MindSpore framework.

In 2019 China AI Developers Conference last year, CSDN held on Huawei MindSpore senior architect at Fan introduced its three design philosophy: to provide a new programming paradigm, the implementation of new models and demand the whole scene together.

"MindSpore you want to target Tensorflow are completely different things." He said that the technology has considerable technical advancement direction, but the challenge is relatively large, so Huawei would like to force in this regard.

Enterprise business needs

In addition to technological innovation, but also taking into account the close AI actual business scenarios. At this point, perhaps those companies to build industry more say AI framework based on their own scenarios.

Rely on the data and a huge bonus scenarios, the development of China ahead of the AI industry application layer. But when the AI industry in a way developed to a certain stage, when, if the AI companies do not have the underlying technology, it is difficult to support the upper application to iterate quickly. Class science and technology base camp founder Yuan Jinhui told AI Technology (ID: rgznai100), "in the present development of the underlying software and hardware integration, if not deep grasp the new framework, it is hardware and software co-design, a certain competitiveness in the international arena. "

GitHub address: https: //github.com/MegEngine

Kuang view of Tianyuan ( MegEngine ) as Kuang, as the Brain ++ is one of the core components, is accompanied by a set of hands-on experience of the desert, as the framework for their own AI industry.

According to Kuang Depending said that from the beginning of 2014, research and development, full use in 2015, but Google and the open source TensorFlow released in the same year, also applies Figure way to do computing framework a great impact on them, but they found that the performance of TensorFlow slower than the self-development framework several times.

Thereafter, in industrial practice, they keep the underlying framework, data and facilities iteration, the final completion of a comprehensive migration from development to operations to self-depth learning framework and its own computing cluster. Senior technical director Kuang Institute Tian Zhongbo view that in the past few years, depending on Kuang met a lot of common industry pain points in the development process, and the key feature of Tianyuan is to solve these pain points.

Until March this year, the open source, open view all current algorithms are based on Tianyuan training and reasoning, either AI or engineering competition, all of its products do technical support.

GitHub address: https: //github.com/mindspore-ai/mindspore

MindSpore research also integrate with Huawei's own business are closely related. In 2018, Huawei join the whole conference, Huawei released its first full-stack whole scene AI solutions, including chip, chip enable, training and reasoning framework and applications to enable full-stack solutions available in public cloud, private cloud, various edge computing, networking and consumer industry terminal end so the whole scene deployment.

Clearly, AI computing framework is its AI solutions in the important part. MindSpore chief scientist Chen Lei, MindSpore native fit each scene comprises end edges and clouds, and can be coordinated on needed basis, i.e. AI algorithms implemented by the code focused improve usability and reduces development threshold AI developers.

In other aspects of Huawei's "self-study" technology module linkage, then publish it on MindSpore of the industry's first enterprise-class application development expertise AI Suite ModelArts Pro. While the underlying support other processor GPU, CPU or the like, MindSpore AI may be rising with Huawei synergistic processor optimized to achieve high efficiency operation state, greatly enhance computing performance.

New national infrastructure strategy

Behind China AI AI industry competing to build open-source framework, technical considerations and operational aspects of course important, but should not be overlooked policy support at the national level.

AI for building infrastructure, the Chinese government will focus mentioned in the "new generation of artificial intelligence development plan" and other top-level policy on AI planning, in addition to increasing the application layer technology floor, more in-depth industry and academia hope AI underlying technology research and development.

The past two years, the underlying technology "strangle hold" the past two years China has become the topic of the scientific community to explore the hot, worried that development will encounter difficulties castles in the air like a chip of the development of China AI. The high culture as the representative of academician AI experts are also four sermons "AI do have to do open source, China wants to develop the new generation of artificial intelligence, there must be efficient and controllable risk of open-source open platform" point of view .

Recently, due to the economic impact of the epidemic, the government level to promote, market is set off a "new infrastructure" craze, digital infrastructure which 5G, a large data center, consisting of artificial intelligence and Internet industry concern. State enterprises or ambitious development of AI, the AI to grasp the underlying technology through independent research undoubtedly a major strategy.

In fact, behind Tsinghua "meter map" project funded by the level of government support the National Natural Science Foundation of innovation group projects, Beijing Information Science and Technology, National Research Center team projects.

Figure meter core development team members Liang Shield of AI technology base camp frankly stated that China's current depth study and research applications are mainly dependent on foreign platform, indeed at risk strangle hold, so some domestic IT companies have launched their own platform. "China should have a place in the ecology of the whole industry chain of artificial intelligence."

In addition to computing framework there are still a lot of need to solve optimization problems, Huawei do MindSpore another big reason is the technical background of the war. In Fan said that there must be self-controlled computing framework for the whole scene.

Design drawings, Tianyuan, MindSpore of a series of open-source developers may give China a hard-core open source projects do play a shot in the arm.

Yuan Jinhui said Chinese developers of open source frameworks heart of awe has been too strong, do not think the Chinese team can do an open source framework, but there is now enough understanding of these open source framework, the difficulty of recreating a similar framework is not Pytorch technology roadmap It will be great.

"This year should be more lively domestic open-source AI framework of the year," Yuan Jinhui revealed, OneFlow framework class science and technology will also open soon, he competitiveness of domestic technology framework confident, more importantly, "China AI framework the development also represents the country embarked on a drive from application to study more core technology. "

In any case, design drawings, Tianyuan, MindSpore ignited the 2020 domestic AI framework open source spark, as to whether future influence prairie, and it can really TensorFlow, PyTorch a higher low, the right to vote in every developer's hands.

【End】