Beijing, March 25, 2020 – Megvii Technology Limited announced today at an onlinepress conference that it would open source its deep learning framework, MegEngine, to allow global developers to create AIsolutions for industrial and commercial scenarios and to foster an ecosystemaround itsnew generation AI productivity platform, Brain++. MegEngine is being branded 天元(Tianyuan) in Chinese, a namewhich means “the origin for everything,” and also denotes the center point onthe Go game board.
Initially developed in 2014 and now in its eighth edition,MegEngine is one of three primary components of the Brain++ architecture thatMegvii built from scratch to train computer vision algorithms at scale. Theother two proprietary components of Brain++ include MegData (a data managementsystem) and MegCompute (a computing power dispatching system).
Megvii is one of the very few companies in the world thathas developed its own deep learning framework, and by making parts of itavailable for free in this “alpha” launch, Megvii hopes that externaldevelopers both in and outside China, can make MegEngine even better for thebenefit of all who use it. A beta launch is tentativelyscheduled for June, and an official launch in September 2020.
YIN Qi, the co-founder, Chairman and CEO of Megvii said, “My founders and I studied AIalgorithms at Tsinghua in the late 2000s just after international academicsachieved a major breakthrough in deep learning and openly shared theirideas. Their work accelerated the fields of computer vision, speechrecognition, natural language processing and audio recognition globally. Inthis same brave spirit, we want to open source some of our innovations,creating a dynamic where other engineers improve our tools, while we also havefurther influence on how the code evolves.”
MegEngine will bereleased under the terms of Apache License 2.0, a permissive free softwarelicense written by the Apache Software Foundation that allows users to use thesoftware for any purpose, to distribute it, to modify it, and to distributemodified versions of the software, without concern for royalties. MegEngine code will be available on two coderepository websites, GitHub and OpenI. Tools and tutorials are available (in Chinese language only for now) at megengine.org.cn.
Co-founder and CTO TANG Wenbin said, “China has beenparticipating in open source since the Android system was introduced in 2011,and Chinese developers today make up an increasingly large proportion of theglobal open source developer community. This promotes international norms of intellectual property andcoding. We think that open sourcingMegEngine is consistent with this trend. Megvii wants to lead the way in Chinaand create good will among developers here and abroad.
We have designed MegEngine to be a unified training andinference platform, with both static and dynamic graphs. It’s fast and accurate, appeals to many typesof developers, is optimized for many hardware and chip platforms, and is compatiblewith PyTorch models. While we knowaspects of MegEngine can be improved, we are also proud to contribute anelegantly designed, industrial strength system to existing global AIinfrastructure.”
MegEnginehas been deployed across Megvii’s three business units(personal IoT, city IoT, and supply chain IoT) and is used by its 1,400+ strongR&D team in its day-to-day algorithm and product development. Optimized forcomputer vision tasks, MegEngine is especially effective for training largevolumes of images and videos and completing complex tasks such as imageclassification, object detection, object/scene segmentation and videoanalytics. It has helped Megvii win 27 first-place recognitions in some of theworld’s most prestigious international AI challenges, including three first-placeawards at the 2019 InternationalConference on Computer VisionCOCOchallenge.
Megvii, supported byMegEngine, has developed advanced deep neural networks that are deployed acrosscloud, mobile and edge computing platforms. Notably, Megvii has developedShuffleNet, a deep neural network designed for devices with relatively limitedcomputing power such as mobile devices. According to tests conducted byAlphabet’s AI team in 2018, ShuffleNet achieved better performance thanAlphabet’s comparable MobileNet in terms of both efficiency and algorithmicaccuracy.
Chief Scientist and Head of Megvii Research Dr. SUN Jiansaid, “It took us six years from developing MegEngine for our own use in2014 to open source it today. Open sourcing MegEngine is just the first step inour contributions to the open source movement. We have used it to power our computational photography, facial recognition,and object recognition applications, but the developer community can use ourfoundational technology in innovative, real world applications of AI that wehave not yet imagined. They will make what we built even better, and we canspend more time commercializing our products. It’s a win-win for all.”
At today’s online press conference, Megvii also announcedthat it will be introducing an enterprise version of Brain++ to offer support,performance, and managed AI productivity services for a fee to corporateclients. In addition, Megvii is releasing version two of its object detectiondatabase, www.objects365.org, which now has a dataset that is many times enlarged.
Today’s conference featured keynote speeches from TANG andSUN as well as Dr. TIAN Zhongbo,SeniorTechnicalDirector of Megvii Research, and ZHAO Liwei, Vice President of Megvii’s Cloud Services. Hundreds of developers, academics and business partnersattended the event. Congratulatorymessages and comments were received by Megvii from prominent figures including TsinghuaProfessor Andrew Chi-Chih Yao and former Microsoft ExecutiveVice President of Artificial Intelligence & Research Harry Shum.
AboutMegvii Technology Limited
Megvii is a world-class, pure AI company withcore competencies in deep learning. Founded in Beijing in 2011 by Yin Qi, TangWenbin and Yang Mu, three Tsinghua University graduates who studied under theTuring-prize winning Chinese computer scientist and computational theoristAndrew Chi-Chih Yao, Megvii is a pioneer in applying AI technology and computervision algorithms to Internet of Things (IoT) use cases. The Company’s missionis to power humanity with AI.
Megvii’s first commercialized AI product wasa facial recognition solution. The Company has developed its new generation AIproductivity platform Brain++, which includes a proprietary deep learningframework that functions as a unified underlying architecture and providescritical support tailored for the Company’s algorithm training and modelimprovement processes.
Leveraging its strong software-hardwareintegration capabilities, Megvii offers full stack solutions to its customers,and has achieved leading positions in several key vertical markets, includingpersonal IoT, city IoT and supply chain IoT. Megvii is building an AIinfrastructure to connect and empower various IoT devices, to solve real lifeproblems and pain points encountered by customers in sectors including fintech,transportation, logistics and others. Megvii is dedicated to helping customersachieve specific goals such as user authentication and fraud reduction, trafficoptimization, and labor efficiency gains.
Headquartered in Beijing with over 2,300employees, Megvii maintains four research and development centers in China.
