Huawei AI framework MindSpore to rival Google’s TensorFlow

China’s Huawei Technologies Co. Ltd. today said it has open-sourced a framework for artificial intelligence-based application development called MindSpore.

First revealed last year, MindSpore is an alternative to well known AI frameworks such as Google LLC’s TensorFlow and Facebook Inc.’s PyTorch. It can scale across devices, cloud and edge environments, Huawei said in a statement. The code is now available to download on GitHub and Gitee.

Huawei developed MindSpore alongside partners, including Imperial College London, Peking University and the University of Edinburgh, as well as a Turkish robotics startup called Milvus. The framework is said to run atop processors, graphics cards and specialized neural processing chips such as Huawei’s Ascend hardware.

One of its advantages is that it uses about 20% fewer lines of code than other competing AI frameworks for tasks such as natural language processing. In addition, MindSpore is said to support parallel training across hardware to cut down on training times, as well as dynamic debugging, which makes it easier for developers to isolate any bugs in their apps.

Another key advantage of MindSpore is that it doesn’t process data by itself but rather ingests just the gradient and model information that has been processed. That helps ensure the integrity of sensitive data, Huawei said.

Alongside MindSpore, the company announced it’s open-sourcing a module called MindInsight, which provides additional debugging and tuning functionality by creating visualizations of the AI training process. Those visualizations include model parameter information such as training data and accuracy, as well as training progress metrics and computation graphs, Huawei said. There’s also a second module called MindArmour, which helps to enhance model security.

mindspore-architecture-1

“MindSpore natively adapts to all scenarios,” Huawei Chief Scientist Chen Lei said in a statement. “We implement ‘AI algorithms as code’ through on-demand collaboration for easier model development, and [we] provide cutting-edge technologies and co-optimization with Huawei Ascend AI processors to improve runtime efficiency and computing performance.”

 

Huawei said MindSpore currently only supports the Python 3.7+ programming language, but has plans to add support for C++, Julia and Rust in the near future. It added that the framework runs most efficiently on Linux platforms such as EulerOS and Ubuntu.

Huawei’s decision to open source MindSpore is part of the company’s wider strategy to cope with being cut off from the Android ecosystem by the U.S. government, which blacklisted the company last year over security concerns, Constellation Research Inc. analyst Holger Mueller told SiliconANGLE.

“AI is a key capability to have, and with Huawei open sourcing MindSpore it has done a service both to the community and to open source in general,” Mueller said. “More importantly for Huawei, it allows for general inspection of the code, so it may help to quell concerns on any malicious or nefarious code being part of the platform.”

Huawei isn’t the only Chinese firm that appears to be taking on Silicon Valley’s heavyweights in the AI market. China’s search engine giant Baidu Inc. has offered its own open-source AI framework, called PaddlePaddle, since 2017.

 

Source: siliconangle

Leave a Reply

Your email address will not be published. Required fields are marked *