Global billion-registered developer? Entertainment Sign billion-designated stations, Evans Data Corporation (Evans Data Corporation) latest data show that in 2023 the number is expected to reach 27.7 million developers. This is less than 30 million developers, 70 million people worldwide living facilities, code with the code, doing constantly tend to the needs of individual realization. This process, hardware environment, network environment, deployment, testing issues, a variety of adaptation issues, cost issues ... are the problem. There are a good tool for someone, someone has a good platform, is a source of innovation - the developers, the key to productivity burst.
2020
From the first line to 35 million lines, backtracking depth learning framework MegEngine
Back in late 2013, when he was head of research and development, as the Kuang Cao Zhimin proposed that build automation algorithm developed a data system, training and business will be able to get through, without the need to invest too much manpower and time, you can algorithm automatic cycle from research to application.
Went ahead early 2014, depending on Kuang Institute intern wrote three plays from the first line of code, less than six months, was born from research by the desert, as the first edition of deep learning framework MegEngine.
Then, in order to catch up with the pace of development of the company, depending on Kuang Academy side to communicate with the business sector, while reproduce the required business neural network. Another six months, in mid-2015, all of the online version of the framework model to replace all the training of self-development, self-development framework to achieve the integration of all internal operations.
In the desert, as the self-development framework MegEngine heyday occasion, Google in November 2015 released TensorFlow and open source. And let the desert, as the surprise is, TensorFlow coincide with MegEngine on the interface, such as the concept of design ideas.
TensorFlow
After internal evaluation contrast, open source TensorFlow performance was not satisfactory, depending on the desert than 10 times slower self-development framework, which undoubtedly strengthened the determination of the desert, as the continuing self-development.
Depth learning algorithm framework is the underlying framework play performance will gradually change from a linear development index changes. With the continuous expansion of business, open frame, as the self-development advantages in business practice continue to highlight the algorithm training period reduced from the past two or three weeks to 3-5 days.
In addition to deploying the underlying framework, Kuang also introduced changes depending on the infrastructure of data and calculation ability.
As early as 2013, depending on Kuang set up a team of their own data, and to create a data management system MegData, to solve data duplication, misuse and other problems, while reducing the cost of data; with the increasing business needs, computing resources again become a bottleneck for this exclusion depending on the depth of learning from the research cloud computing platform MegCompute completed.
So, MegEngine, MegData, MegCompute constitute the exclusion depends on the initial prototype automated algorithm developed system, open platform for productivity, as the AI Brain ++ beginning to appear. VLSI deep learning model of large-scale distributed training capability relying Brain ++, and industry-leading distributed computing technology training out, open depending on the cumulative gains 27 global AI contest winner, and create COCO three consecutive record industry.
Alone Better Together. Open Source conference on the 25th, the open desert, as the 35 million lines of code. There are practitioners know almost sigh: This is a huge project, congratulated the desert, as the developers who also congratulate all the developers.
Source of power rise, deep learning framework "Tianyuan" Review Highlights
The press conference, open depending on the depth of learning framework MegEngine officially open, and the name "Tianyuan." Its original intent was to center of the board, is also the origin of all things, open depending on the hope that more developers can be based on Tianyuan, accelerate the process of constructing algorithms from 0-1 in the AI era, everyone can have a productivity tool.
After six years of industry proven, open view of more than 1,400 R & D personnel are in use, open view now almost all AI products are based on Tianyuan MegEngine building. For the open-source, open depending on the targeted specially made large-scale upgrade, so that it could show people a whole new look.
According to Kuang, as the co-founder and CTO Tang Wenbin introduction, the Tianyuan open source code has 350,000 lines, C ++ accounted for most, followed CUDA, Python; in addition, based on the open source Apache License 2.0 protocol, enabling developers to easily use, including Some use a commercial scenes are no problem.
In the product launch cycles, senior technical director Kuang, as the Institute Tian Zhongbo mystery out of Tianyuan MegEngine architecture. As can be seen from the architecture diagram, Tianyuan MegEngine from top to bottom it is divided into five levels, divided into computing interface diagram shows, optimizing compiler, runtime management, and compute kernels.
In addition, repeated training and practice, depending on Kuang faced a variety of difficulties, accumulated a wealth of experience to avoid the pit. Meanwhile, combine the pain points of the current development environment, such as to bring the training model and reasoning model transformation inefficient, can not have both dynamic static faults caused by the use of research, multi-frame but less easy to use, cost and conversion learning framework costs are high, and so on, Tianyuan have been targeted optimization and upgrading all users to "easily" shall prevail.
According to reports, Tianyuan MegEngine has four advantages, namely:
??
??
??
It is worth mentioning that, compared with Google TensorFlow, FacebookPytorch and other open source framework developers, open view is a company rooted in the field of artificial intelligence. Six years as one day, and always adhere to industry practice, the model based on ability to migrate 100% Tianyuan MegEngine trained to business, so that theory and practice are unified. Tianyuan greatly reduce the threshold for developers to use, which also confirms the exclusion of open source depends on the purpose, namely "deep learning, simple development."
AI sink, depending on the young Kuang "mature consideration"
"AI
The AI in the industrialization process, in addition to deep learning algorithm framework for training, and deployment of reasoning, also involving data and calculate the force, this is the desert, as the productivity of AI platform --Brain ++.
Brain++
Tianyuan MegEngine Brain ++ as a core component of its open source enterprise undoubtedly reduced the threshold for the use, developers, while avoiding the problems of repeat-create the wheel. As we all know, early learning in depth, each developer will need to write a lot of code duplication. Today is a big difference, developers around the world can be free trial Tianyuan MegEngine, implement AI Pratt & Whitney.
Kuang depending on Tianyuan MegEngine open mind in the hope that more and more partners, AI companies added to this collaborative network sharing, so that the industry continues to flourish, so that more businesses or developers can really use AI . Give a man a fish than giving the fishing, open depending on the fishing gear "Tianyuan" available to all developers, allowed to do this based on love and more useful things.
Of course, this is only the desert, as the open-source version of Alpha, and perhaps there are still some imperfections, this also calls for more developers, academics, students joined in, hand in hand to improve Tianyuan ecology.
Tang Wenbin also pointed out: "Perhaps the next generation of Tianyuan is not made out by the desert, as the R & D team, but to create out of it with you and the official Beta version, so we hope that together with you to build a better depth study frame."
Kuang view also makes it clear Tianyuan MegEngine development roadmap, plans to release a Beta version in June of this year, will provide CPU support capabilities on ARM series; September, Tianyuan MegEngine ability to fully cover the mainstream of computing devices, while existing dynamic computing power of a comprehensive upgrade and optimize the whole process of reasoning training experience.