Huawei has made its MindSpore artificial intelligence framework open source. The Chines tech giant is competing with the well-known AI frameworks such as TensorFlow from Google and PyTorch from Facebook, with a large number of advantages that ‘AI algorithms as-a-code’ can provide.
The company states that its MindSpore AI framework is suitable for developing AI applications. The AI framework co-developed with universities in Beijing and the United Kingdom and with a Turkish start-up can easily be rolled out in various environments, such as on devices, within (multi) cloud and edge environments.
Huawei launched the new AI framework last year in conjunction with the Ascend 910 processor. The AI chip provides 256 teraflops of computing power on FP16 and that at a power consumption of 350 watts.
With MindSpore and the Ascend 910 in addition to that new chip, the company has the most important components in the hands of a full AI stack.
The key features of the framework include that it uses up to 20 percent less code than competing AI frameworks for different processes. This includes neural language processing (NLP).
MindSpore also supports simultaneous training on hardware to save training time. This would also enable “dynamic debugging”. This makes it easier for developers to isolate bugs in their applications.
Furthermore, the AI framework of the Chinese tech giant has the advantage that it does not process data itself but rather records the “gradient” and model information that has already been processed.
This ensures the integrity of sensitive data, according to Huawei.
According to the Chinese tech giant, MindSpore natively adapts to all scenarios. This is done by implementing “AI algorithms-as-a-code” through on-demand collaboration. This makes it easier to roll out AI models again. In addition, this approach helps for better collaboration with the Huawei AI Ascend processors.
This hardware can significantly improve runtime efficiency and computing power performance.
The AI framework runs on different types of hardware, including processors, graphics cards and special neural processing chipsets. For the latter chipsets, think of the Ascend applications from Huawei.
MindSpore now only supports Python 3.7+ for code. However, the Chinese tech giant has plans to also introduce support for C ++, Julia and Rust in the near future. The framework runs most effectively on Linux platforms like EulerOS and Ubuntu. Huawei MindSpore is now available through GitHub and Gitee.
Along with Mindspore, the MindInsight module has also made open source. This module helps developers with more debugging & tuning options. This is done by making visualizations of the AI training process.
Visualizations are given, among other things, from model parameter information such as training data and accuracy, the measurement data of the progress of the training and calculation graphs. A second module, MindArmour, helps to improve the security of the models to be developed.
By making its AI framework open source, Huawei seems to be consciously competing with similar open-source frameworks from tech giants Google, TensorFlow, and Facebook, PyTorch, from the United States.
According to experts, this is partly because the Chinese tech giant has now been shut down from Google’s open-source mobile OS Android. Therefore, Huawei would like to focus more on AI and acquire an important open-source position there.
More in AI