The ONNX Partner Workshop, an open source project co-sponsored by Microsoft, first landed in China. Microsoft project general manager Venky Veeraraghavan and Microsoft's artificial intelligence platform team chief project manager Prasanth Pulavarthi accepted media interviews including Lei Feng.com at the Microsoft Building. ONNX progress How, how Microsoft will promote Azure to become the best artificial intelligence cloud platform, the two interviewees made some answers.
ONNX bridges hardware and software
ONNX is “Open Neural Network Exchange”, that is, “open neural network switching”. ONNX is an open format for representing deep learning models. With ONNX, AI developers can more easily move models between the most advanced tools and choose the combination that works best for them. ONNX is developed and supported by the partner community. (ONNX is a open format to represent deep learning models. With ONNX, AI developers can more easily move models between state-of-the-art tools and choose the combination that is best for them. ONNX is developed and supported by a community of Partners.)
The above is the official introduction of ONNX. Lei Feng.com (Public No.: Lei Feng) previously reported that in September 2017, Microsoft and Facebook released ONNX, in order to create an open deep learning development tool ecosystem, the purpose of which is to make different neural network development frameworks To interoperate.
Pushing the model from the research phase to the product phase often requires moving to a different framework. ONNX enables AI developers to choose the most appropriate development framework for the task and optimize the hardware vendor.
ONNX works by tracking how a neural network is generated on these frameworks in real time, and then using this information to create a common calculation graph, which is an ONNX-compliant calculation graph.
Microsoft Project General Manager Venky Veeraraghavan
For the future evolution of ONNX, Prasanth Pulavarthi said, “In fact, the specific specifications of ONNX have been constantly evolving. The basis of its changes and evolution is new development work, and on the other hand, new developments. The customer needs to update ONNX every two to four months. In this upgrade we will add some new operators, which are actually a mathematical function that allows users to create new models. Therefore, we will continue to do this work in the future and continue to create new types of models, so ONNX's specification change updates have continued. ”
Prasanth Pulavarthi, Principal Project Manager, Microsoft Artificial Intelligence Platform Team
Venky Veeraraghavan added, “ONNX is a standard, and there are already a lot of companies involved in the standard. Microsoft is focused on ensuring that the related work is helpful in solving Microsoft internal problems. Microsoft has spent a lot of time and effort in Microsoft's internal testing, whether it's Bing, voice or advertising models to ensure that this set of standards is able to support the most complex models. Moreover, not only does it support very complex models, but it also has excellent performance on hardware platforms and can support a variety of different frameworks, like TensorFlow, PyTorch. ”
Lei Feng.com has learned that AMD, ARM, Huawei, IBM, Intel, Qualcomm and other companies have joined the ONNX open source project. More than 20 global companies have announced their support for ONNX.
The two respondents did not respond positively to when the industry was very concerned about when Google joined ONNX or the latest development, but Venky Veeraraghavan said,Microsoft believes that there is no doubt that the future of AI needs to rely on open standards. Microsoft has also fully communicated with Google and shared relevant plans and ideas. But so far, Google has decided not to actively participate in ONNX work for the time being.
Azure goal is the strongest artificial intelligence cloud platform
According to Microsoft official data, ONNX models can be created in a variety of artificial intelligence development frameworks, including PyTorch, Chainer, Microsoft Cognitive Toolkit (CNTK), MXNet, ML.Net, TensorFlow, Keras, SciKit-Learn, The list is still increasing. In addition, the ecosystem of development tools for visualizing and accelerating ONNX models has begun to take shape, and ONNX models have been developed for pre-training of typical scenarios.
While fully supporting ONNX to build an open interoperable artificial intelligence ecosystem, Microsoft is also working hard to make Azure the best artificial intelligence cloud platform. From pre-trained models to cloud services that help build models, Azure provides a complete machine. Learning platform.
Lei Feng.com understands that Microsoft mainly provides three aspects of AI capabilities. The first is the infrastructure level. For data scientists familiar with algorithms, Microsoft's GPU service has been deployed in China's data center several months ago. Microsoft also has very strong data-related services. GPU, CPU, or virtual machine can be used. Classified as an AI model at the infrastructure level.
The second audience is software engineers. Microsoft can provide the ability to directly call existing AI services, such as bots or cognitive services, collectively known as API cognitive services, which Microsoft categorizes as AI Services. The third major category is tools such as the ONNX framework and the CNTK framework. In addition, Microsoft also has Visual Studio and Visual Studio Code, etc., Microsoft's overall AI ideas are also divided into these three parts.
In Lei Feng's view, Microsoft's feature is to empower other companies or developers to create exclusive AI tools, AI products, AI leadership and competitiveness, which is the difference between Microsoft and other artificial intelligence companies.