Megvii unveils latest algorithm training framework version
Published: Sep 18, 2020 11:13 PM

A booth of AI technology at the 6th China Electronic Information Expo held in Shenzhen, South China's Guangdong Province. Photo: VCG

Chinese artificial intelligence (AI) firm Megvii Technology, known for facial recognition platform Face++, on Friday unveiled the 1.0 preview version of its proprietary open-source deep learning framework MegEngine, in a fresh move to close China's gap in the AI ecosystem with the US.

At a meeting on AI, open source and productivity during the 2020 Zhongguancun Forum, a sci-tech forum in Beijing, Sun Jian, chief scientist of Megvii and head of Megvii Research, unveiled the latest version of MegEngine, part of the firm's AI architecture Brain++, that also includes a data management system and computer power dispatching platform.

The Beijing-based AI unicorn announced in March the open-source MegEngine, making it accessible to developers from around the world. 

From the availability of the alpha version in March to the launch of the 1.0 preview version, the self-developed deep learning framework has gone through eight iterations, upgrading to 510,000 lines of codes from the previous 350,000 lines, Sun said. All of Megvii's algorithms are trained and inferred through MegEngine.

The latest version features automatic code stripping, inference performance improvements and enhanced support for homemade hardware.

Megvii's move is seen to reduce the nation's dependence on US-originated deep learning frameworks, especially as concerns mount over a technology split.  

Megvii and two other major AI start-ups, SenseTime Group and YITU Technology, were among eight Chinese tech firms that were added to the US Entity List in October 2019. 

The launch of MegEngine, China's indigenously developed open-source library and computing architecture, is of significance for improving the nation's core competitiveness in AI, said Zheng Nanning, a member of the Chinese Academy of Engineering 

China's proprietary AI frameworks have a long way to go before they can match their more popular US counterparts, according to Yuan Jinhui, founder of OneFlow, a Beijing-based firm specializing in developing AI infrastructure. Google's TensorFlow and Facebook-backed PyTorch take the lion's share of AI frameworks globally. 

"For an AI ecosystem to rise to prominence, such frameworks need unique features to attract developers, which is essentially about being innovative to sharpen their competitiveness," Yuan told the Global Times on Friday.

Multiple Chinese alternatives to US heavyweights that have been up and running have already proven to be unique in certain aspects of existing open-source frameworks. They are much faster or support functionalities unavailable on other frameworks, he said. 

OneFlow also open-sourced its self-developed deep learning framework earlier this year.