Just two months after the release of version 2.0 of the Feibo open-source framework, on May 20 at the Wave Summit 2021 Deep Learning Developers Summit, Baidu unveiled the new version 2.1 of the Feibo open-source framework, as well as new upgrades for the Feibo Enterprise version.
Wang Haifeng, chief technology officer of Baidu and director of the National Engineering Laboratory for Deep Learning Technologies and Applications, delivered a speech at the summit. According to Wang Haifeng, Paddle has attracted 3.2 million developers, up nearly 70% from a year ago. The number of enterprises and public institutions served reached 120,000.
With regard to the current stage of artificial intelligence technology and industrial development, I share two thoughts:
First, integrated innovation. From the perspective of technology, the combination of knowledge and deep learning breaks through the deep semantic understanding enhanced by knowledge. Multi-technology fusion innovation, through the knowledge graph association of cross-modal information, using natural language semantics to represent the integration of language, speech, visual and other different modes of semantic space, breakthrough in cross-modal semantic understanding. From the perspective of platform, deep learning platform and chip software and hardware integration innovation, to meet the diverse needs of different computing power, power consumption, delay, etc., to achieve the best effect of AI application. Aiming at the environment where multiple chips coexist, Feibo developed the training technology of server with heterogeneous parameters, which broke through the problem of efficient training of super-large models. At present, FeiBar has carried out adaptations and joint optimization with 22 domestic and foreign hardware manufacturers, and the hardware ecology is booming. From the perspective of industry, AI technology is increasingly integrated with industry. Drived by industry demand, AI technology and platform capabilities are continuously polished, and innovative development is integrated with application scenarios.
The second is to lower the threshold. With the penetration of artificial intelligence technology in various industries, it is critical to continuously lower the threshold for different application scenarios to effectively meet the needs of different developers. Fly blade was due to the depth of the industrial practice learning open source open platform, has been committed to reduce the threshold, both support movement unified core framework, there are industrial grade model library components, development kit and tools, and improve enterprise efficiency fly blade enterprise edition, etc., to meet different industries, different stages, the needs of different levels of developers. Based on the flying paddle platform, 3.2 million developers no longer need to write artificial intelligence algorithm code from scratch, so they can effectively carry out technological innovation and business expansion. The substantial reduction of the threshold has accelerated the diversification and scale of artificial intelligence applications and accelerated the process of industrial intelligence.
Custom OP, large-scale map search engine, Baidu flying paddle ushered in version 2.1
The highlight of this summit is nothing more than Baidu flying oar update again. Ma Yanjun, senior director of Baidu’s deep learning technology platform department, introduced the new features of the latest version 2.1 of Baidu’s open source framework:
- Automatic mixing precision optimization: the training speed can be increased by 3 times after starting;
- Enhanced dynamic graph function: Added inplace operation function to realize automatic video memory reuse; Optimize Python/C++ interaction overhead to improve immediate execution efficiency;
- High-level API: new support for GPU preprocessing, new support for mixing precision, new model sharing mechanism;
- Customized OP feature optimization: support one-click compilation, installation and generation of operator API, simplify compilation and use process;
- Large scale graph retrieval engine: support trillion-edge distributed graph storage and retrieval, support linear scaling.
Fly blade version 2.1 for more information, see https://www.paddlepaddle.org…
In terms of model kits, Baidu Feitoar has released four pre-training models of WenXinErnie, which further makes breakthroughs in semantic understanding of knowledge enhancement, cross-modal semantic understanding and other aspects, and empowers developers.
In addition, Ma Yanjun also focused on the key link of the last mile of AI application — inference deployment tool chain.
Baidu fly oar reasoning deployment tool chain updates include:
- Paddleslim model compression upgrade: Optimized pruning compression technology and added unstructured sparse tools; Combined with the advantages of various compression strategies, it takes the lead in supporting OFA compression mode.
- The end-to-side inference engine Paddlelite has been fully upgraded with the addition of the out-of-the-box mobile development toolset LiteKit for increased ease of use. Optimize the performance of two scenarios, ARM CPU and OpenCL, to maintain the leading edge in INT8 scenarios; Support multiple chip types;
- Paddle ad-serving: Added Pipeline mode with full asynchronous design
- New updates to the web deployment tool Paddle.js: high compatibility (support for multiple backends), high performance (support for WebGL Pack functionality, FP16 model format), and front-end model encryption solution (ensure security).
Along with the updating of the reasoning deployment tool chain, Baidu FeoBar also recently released the reasoning deployment navigation map and updated the hardware ecological roadmap of FeoBar. Ma Yanjun said that at present, more than 300 paths have been fully verified, and 31 chips and IP chips have been adapted.
There are also new upgrades to the quantum machine learning development tool PaddleHelix and the biometric computing platform PaddleHelix Propeller, which were released last year.
Fly paddle enterprise edition new upgrade
On May 20 of last year, baidu fly paddle released fly paddle enterprise version. As shown in the figure below, the enterprise version of the flying paddle includes “one core and two wings”. The left wing is the zero-threshold AI development platform EasyDL, enabling AI application developers; On the right is BML, a full-featured AI development platform designed for AI algorithm developers to help improve development efficiency.
At the summit, Xin Zhou, director of Baidu’s AI product research and development department, introduced the latest upgrade of the Feioar Enterprise Edition.
- Automated modeling mechanisms continue to upgrade;
- Automatic scene adaptation optimization;
- Automated model evaluation and assisted diagnosis.
- BML preset model development, low code modeling based on pre-training, optimized 67 sets of components, saving 80% of development time
In addition, PaddleFlow Enterprise Edition has recently released the cloud native machine learning core PaddleFlow. According to Ma Yanjun, this is the industry’s first cloud native platform specifically for AI, and the core PaddleFlow technology will soon be open source.
In addition to the update of Baidu Feioboar, Feioboar also unites academia and industry to jointly launch Feioboar voyage plan to help AI talent cultivation, industrial intelligent upgrading and frontier exploration. The partners share the industrial innovation practice and talent joint training plan of the application of flying paddle.