One year after the Tianshu Artificial Intelligence Platform went open source, Zhejiang Lab released Tianshu v2.0, a grand new version with better functions and higher performance, aiming to boost AI tech innovation and industry growth. On August 30, at the launch ceremony of the 2021 Zhejiang Lab Cup Global AI Competition, BAO Hujun, Vice President of the Lab and Chief Architect of the Tianshu Platform, announced the public debut of this new version. Professor PAN Yuanhe, member of the Chinese Academy of Engineering and the Lab's Chief Scientist in Artificial Intelligence, witnessed the release along with more than 180 delegates from organizations including the Zhejiang Provincial Cyberspace Administration, the Zhejiang Provincial Department of Economy and Information Technology, the government of Yuhang District, the Management Committee of Hangzhou Future Sci-Tech City, Zhejiang University, the China Academy of Information and Communications Technology, and the Zhejiang Artificial Intelligence Industry Alliance. In addition, Tianshu v2.0 will serve as the official platform for this year's competition to support its entire process.

The Tianshu Platform is an AI open source platform co-developed by Zhejiang Lab and other organizations such as OneFlow Technology Limited, with intellectual property rights completely owned by China. Starting in October 2018, the project has been supported by the National Development and Reform Commission. Compared with the first version launched online in August 2020, Tianshu v2.0 includes more functions and applications, showing substantial improvement in usability.

Since the platform went open source, the project team has been optimizing its functional modules and overall performance. With the support of a large number of researchers and engineers, the team was able to create an updated version every three months. The deep learning framework of Tianshu v2.0 is designed to fully match the PyTorch interface, requiring only one line of code to be compatible with PyTorch. In addition, the framework uses the same set of codes for symbolic and imperative programs, which reduces the development workload and the learning cost for developers.

In terms of AI model development, Tianshu v2.0 offers remote connectivity and a separately deployable visualization toolkit, which are the two highlights of this update.  The visualization toolkit is compatible with mainstream frameworks and can present real-time training data directly, with functions and responsiveness reaching the level of major visualization tools. The functions of remote connectivity and mirror management are made for professional algorithm developers, so as to improve the efficiency of developers and breaking the functional limitations of the platform.

It is worth mentioning that the model library of Tianshu v2.0 integrates more than 50 algorithms in six areas, including intelligent vision, natural language processing, intelligent speech, recommendation systems, reinforcement learning, and model compression. In particular, the Chinese Language Pre-Trained model (CPT), jointly developed by the Tianshu project team and Fudan University, uses a pioneering architecture featuring "deep coding-shallow decoding" to greatly improve the model’s efficiency. The model shows remarkable performance in multiple metrics compared with major Chinese language pre-training models.

During the year after the Tianshu Platform went open source, the project team also launched CodeLab, an open source community, and established the Tianshu Open Research Program to develop the platform’s ecosystem. The platform has now gathered ecosystem partners in six fields, and invited more than 500 institutions and companies to deploy their applications on the platform. Joined by innovative organizations such as the Chinese Academy of Sciences, Fudan University, Zhejiang University, OneFlow Technology, and Enflame Technology, the platform has seen a preliminary ecosystem co-built by the academic and industry communities. At the launch ceremony, Vice President BAO Hujun also announced the Tianshu Ecosystem Plan. He said that the platform would focus on the needs of universities, research institutions, and small- and medium-sized enterprises. In particular, it will provide professional and general versions to different user groups and enable private deployment and operation of the platform. In addition, it will open small-scale computational power for free model training services. The Tianshu Open Research Program will continue to fund open research projects, so that the project team can work with partner organizations for improvement in areas such as advanced algorithms, operator libraries, automatic generation and compilation of deep learning operators, intelligent data processing and analysis, and adaptation and deployment of domestic hardware. The project team will also bring Tianshu Academy to universities, online courses, and we- media communities, answering users’ questions, encouraging open source contributions, and sharing cutting-edge AI innovations.

As a fundamental facility for intelligent computing software, the platform will continue to build its strength in aspects such as intelligent data processing, model development, model training deployment, computational power management, algorithm application, and the adaptation to AI chips developed by China. It will invite more partners to enable software and hardware integration, and therefore create an AI innovation ecosystem supported by the industry, academic, and research communities. In addition, it will provide customized services according the needs of specific scenarios, so as to facilitate the development of multiple applications such as smart manufacturing, smart campuses, smart cities, intelligent transportation, and intelligent robotics.

During on-site sharing session of the launch ceremony, Professor PAN Yunhe said, "With the advent of AI 2.0, data, algorithms, and computational power have become increasingly important. Zhejiang Lab planned and started the Tianshu project in 2018, aiming to provide algorithms and partial computational power to AI R&D institutions in the Yangtze River Delta and all over China. It is a platform for public benefit. Since the platform went open source, it has shown unique advantages, reflecting a distinctive path taken by AI researchers and engineers in China. We hope that our researchers and industry partners can make full use of the platform and the open ecosystem, so as to build China’s research strength in artificial intelligence on a platform we have developed by ourselves."