How does AI open source innovate? Shengsi MindSpore TechDay on-site celebrities reveal the secret

In the era of intelligence, AI technology is playing a vital role, and open source deep learning frameworks can not only lower the threshold for AI developers, but also greatly save costs and time, becoming an important driver of innovation. On March 26-27, 2022, Synthetic MindSpore TechDay was successfully held online, including the release of the new version of Synthetic MindSpore 1.6, the sharing of enterprise application cases, key technologies of AI large models, and the sharing of AI+ scientific computing progress. Readers have a more in-depth understanding of the technology innovation and community governance of SENS MindSpore.

Tian Kunyang, Vice President of Huawei's Ascend Computing Business, said in his speech that Huawei has open sourced Ascension MindSpore for two years. It is young and full of vitality. The underlying model types and numbers are very rich, which can support all-scenario AI applications. While achieving fruitful scientific research results, it has formed a complete ecosystem, and cooperates closely with universities, well-known academic research institutions and industrial partners to jointly build a prosperous AI ecosystem.

During the event, SENS MindSpore version 1.6 was officially released. The new version brings new upgrades in terms of ease of use, development efficiency, and control flow performance: MindSpore Dev ToolKit, a development kit for instant user experience, one-click environment management, completed in 5 minutes Environment construction; intelligent code completion based on deep learning, the number of keystrokes for single-model development can be reduced by 30%; visualization of the entire development process, timely viewing of codes and corresponding graph structures, and support for rapid positioning of cluster training performance problems.

At the same time, it also demonstrated the achievements of the original AI large model of SENS MindSpore, which realizes multi-dimensional hybrid parallelism through the AI compiler, and supports seven-dimensional hybrid parallel algorithms such as data parallelism and data slice preprocessing. A single machine accommodates the industry's strongest AI framework for model parameters, reducing the amount of parallel code for development by 80%, reducing system adjustment time by 60%, and the ability to complete training of hundreds of billions of model parameters on a single machine. At present, Zidong. Taichu, the world's first three-modal large model, and the world's first special framework for deep learning of intelligent interpretation of remote sensing images (Wuhan. LuojiaNet) have been established respectively. Modal Artificial Intelligence Industry Alliance" and "Intelligent Remote Sensing Open Source Ecological Alliance" have created a new model of industrial landing.

In addition, in the new AI track of scientific computing, MindSpore has completed the evolution of AI computing framework to AI fusion framework through architectural innovation, and through nonlinear fitting of neural network simulation, scientists no longer need to solve high-dimensional equations and enhance technological innovation. Efficiency, greatly improving development efficiency and scientific computing performance.

Ascension MindSpore promotes scientific research and innovation with hard-core technical characteristics. It not only has abundant learning resources, perfect growth paths, and low participation thresholds. It is an AI open source community with three technical advantages of "full-scene collaboration, full-process simplicity, and adaptive evolution". . As of March 2022, Ascension MindSpore community downloads exceeded 1.4 million, community contributors exceeded 4,000, service companies exceeded 5,000, and cooperated with more than 120 scientific research institutes. More and more AI developers are upgrading The MindSpore community found like-minded friends to "change the world" together.