This year’s Huawei has indeed encountered many difficulties.
In particular, the attack on the supply chain, including chips, made Guo Ping, Huawei’s rotating chairman, admit that “it has indeed brought great difficulties to Huawei’s production and operation”.
However, due to the ups and downs of fate, Huawei is still on the site of full connection and conveys confidence:
“Always face the sun, and the shadow will be left behind by you.”
Huawei also uses the achievements of specific business and future investment to prove that this confidence is not just shouting slogans.
Taking Huawei cloud as an example, Guo Ping said that the cloud is the best platform to release computing power and the digital base of the intelligent world. After three years of continuous efforts, Huawei cloud now has 23 regional centers around the world to provide services1.5 million developers。
What can more directly explain the “future” is the continuous investment in products.
How specifically, it was launched at the end of the 2020 Huawei full connection conferenceModelArts 3.0On the, we can see the clue.
Don’t talk much. Let’s see.
Modelarts 3.0 release
As a masterpiece of Huawei cloud AI development platform, modelarts can provide AI application development services including data annotation preparation, model training, model tuning, model deployment and so on.
Shortly after its launch in 2018, it brushed the list of Stanford dawnbenchmark. In terms of the total training time of image recognition (resnet50 on Imagenet, with an accuracy of more than 93%), it achieved 10 minutes and 28 seconds, nearly 44% faster than the second place, and won the first place in the world at that time.
After the iteration of version 2.0 last year, Huawei cloud modelarts has evolved into a simple and professional one-stop AI development and management platform that can even complete model training and one click deployment with 0 code.
So this year,What new breakthroughs can modelarts 3.0 make?
At the full connection site, Tian Qi, chief scientist in the field of Huawei cloud artificial intelligence and IEEE fellow, announced the answer.
Tian Qi introduced that modelarts 3.0, as an AI development platform for AI in the industry, has explored and studied “how to train high-precision models with very little data”, “how to reduce the threshold of enterprise application AI”, “how to solve the concerns of enterprises about the safe use of data”, and brought four new features to this end.
Now, automatic machine learning, small sample learning, federated learning, pre training model and other AI capabilities accumulated by Huawei cloud for a long time canPlug and playIt is successfully deployed on the modelarts platform to help AI land.
Four new features
EI backbone: a new paradigm of AI development
First, the new release of the general pre training model architecture EI backbone.
Its purpose is to createPre training model + small sample fine tuningEfficient training mode to comprehensively improve the AI landing ability and experience of the industry.
In other words, EI backbone will provide general pre training model and industry customized development process, so that the formed development experience can be replicated on a large scale and reduce the use threshold of AI.
If the pre training model in NLP field is taken as the benchmark, the long-term goal of EI backbone is to build Bert in CV field.
The reason why Bert is called “the beginning of a new era of NLP” is not only because it brushed the list of major NLPs at the beginning of its birth. What’s more, based on Bert pre training model, NLP model can achieve good performance in downstream tasks only by simple migration strategy.
This undoubtedly greatly promotes the research and development in the field of natural language processing.
EI backbone is dedicated to reproducing the Bert experience for developers in the CV field.
Take medical image segmentation as an example. In the past, hundreds of labeled data were needed to complete the training. With the blessing of EI backbone, only dozens or even a dozen labeled data were needed to complete the training,Save labeling cost by more than 90%。
Tian Qi introduced that in the past, model selection and hyperparametric adjustment that required a lot of expert experience and trial and error cost can be completed quickly without manual intervention through the full space network architecture search and automatic hyperparametric optimization technology provided by EI backbone, and the accuracy can be greatly improved.
Combined with Huawei cloud’s computing resource allocation and data management, the whole development process of model training, testing, acceptance and deployment can be completed in a few hours or even minutes after loading the EI backbone integrated pre training model, and the training cost can be reduced by more than 90%.
At present, EI backbone has had successful case verification in more than 10 industries and won more than 10 industry challenge Championships. Huawei cloud has also published more than 100 relevant papers on EI backbone.
The relevant model architecture will be gradually improvedOpen Source。
Federal learning: breaking the data island
The second new feature is that modelarts 3.0 adds the federated learning feature.
Data is undoubtedly the basis of AI application. Only based on diversified data can AI intelligent perception be realized.
However, in the actual AI landing, there are often such problems: the data are scattered among different data controllers, limited by privacy, security and other problems. These data can not be easily accessed, but form “data islands”.
This limits the training effect of AI algorithm in the actual industry.
To solve this problem, Huawei cloud modelarts provides federated learning features. Users use local data for training, do not exchange the data itself, and only exchange updated model parameters in an encrypted manner to realize joint modeling.
AI intelligent evaluation: automatic debug, visual
For AI development, having rich data as the basis and completing model training does not mean success.
The evaluation and tuning of model performance is also an important work, which requires high experience of developers.
The feature provided by modelarts 3.0 in this link is AI intelligent evaluation.
Its model evaluation function is to send the model reasoning results, original images and real labels into the model evaluation module after obtaining the first trained model.
This module will evaluate the comprehensive ability of the model from two aspects: data and model. The evaluation indicators include accuracy, performance, reliability and interpretability:
stayPerformance aspect, modelarts 3.0 can provide operator level statistical analysis of time and space consumption and a variety of overall performance indicators, and give corresponding suggestions for the performance of the model, such as model quantification, distillation, etc;
stayInterpretability aspect, modelarts 3.0 can provide thermodynamic diagram to show the area on which the model makes reasoning and judgment;
stayCredible aspect, modelarts has built-in multiple model trust related evaluation methods, which can provide multi angle model security capability evaluation indicators, and give corresponding defense suggestions according to the current model performance.
Finally, the evaluation module will output some diagnostic suggestions to improve the model capability for possible problems.
In other words, the heavy work of debug can be automated by modelarts 3.0, and it is also a comprehensive evaluation of the overall process from data to model training.
Flexible computing power + large computing power, Pratt Whitney enterprise AI landing
In addition to the ability of automated development, as a one-stop AI platform cloud service, modelarts also provides computing support.
In addition, in order to better support ai r & D with large computing power requirements, Huawei modelarts platform has made targeted optimization in cluster scale, task number and distributed training.
It can not only manage tens of thousands of nodes, but also better support the needs of large-scale training tasks. By optimizing the service framework, the modelarts platform can also support 100000 level jobs running at the same time and large-scale distributed tasks of 10000 level chips.
In addition, in order to help enterprises further reduce costs and increase efficiency in the process of AI landing, modelarts 3.0 also hasElastic trainingThis core competence.
In other words, the optimal number of resources can be adaptively matched according to the requirements of model training speed.
Specifically for products, modelarts provides two modes.
everythingTurbo mode, you can make full use of free resources to accelerate existing training operations, in most typical scenariosThe acceleration efficiency is greater than 80%, and the training speed is increased by 10 timesAnd will not affect the convergence accuracy of the model.
The second isEconomic model, it can maximize resource utilization and provide developers with the ultimate cost performance. In most typical scenarios, it can improve the cost performance by more than 30%.
Leading distributed speedup capability
So, how should modelarts evaluate its current capabilities?
You might as well speak directly with data.
The modelarts platform supports the simultaneous operation of 100000 level enterprise tasks and the simultaneous use of 100000 level user scale.
The key ability to realize large-scale cluster distributed training depends on the distributed acceleration ratio.
The test results on mlperf benchmart show that under the cluster scale of 512 chips, Huawei cloud modelarts scores93.6Seconds, better than 120 seconds of NVIDIA V100.
Mlperf benchmart is a general benchmark jointly created by academia and industry to measure the training and reasoning performance of machine learning hardware, software and services.
Landing practice of modelarts
The data on paper is only part of it.
In fact, in more than 10 industries such as energy, automobile, government affairs system, education and industrial robots, huaweiyun modelarts has realized 160 + landing cases.
For example, the following domestic robot dog has the AI ability given by modelarts.
This robot dog is called “jueying” and is produced by Hangzhou Yunshen Technology Co., Ltd.
Zhu Qiuguo, founder and CEO of Hangzhou Yunshen Technology Co., Ltd., introduced that Yunshen cooperated with Huawei, applied modelarts and Atlas 200dk to endow “jueying” series robot dogs with AI capability, and was committed to intelligent patrol inspection in the factory park. The robot dog can perceive the on-site environment in real time, strengthen learning and dynamic decision-making through interactive analysis of knowledge map, and has the ability of complex travel path planning and action. The end cloud cooperates to guard the safety of the factory park.
On site, Huawei cloud demonstrated how to build a robot dog based on modelarts“Perception + cognition + decision making”Ability: after training, “jueying” can not only detect the flame and understand that the fire is harmful to the factory environment, relearn and plan to generate a new path, bypass the flame, turn off the control valve and relieve the fire.
Another interesting application case, students from Tongji University. The landing scene isMigratory bird protection。
Through Huawei cloud modelarts, five students of Tongji University trained models that can identify thousands of birds in just a few months, and established a wetland digital twin system.
Moreover, this system has played a role in the protection and scientific research of migratory birds in Hangzhou Bay.
▶ If you also want to use modelarts to develop some valuable AI applications，Click the linkTry out the experience. Now invite friends, AI computing power, Huawei watch bracelet and so on. ☚
How to evaluate modelarts?
Having said so much, how should we evaluate the significance of modelarts to the implementation of AI in the industry?
As Zheng Yelai, President of Huawei cloud business, said on the spot, the development model of modelarts has become a new choice.
Now, modelarts has realized the simplification and automation of the whole process to upgrade the existing AI development mode, so as to make a qualitative leap in the whole chain of data preparation, algorithm development, model training, model management and model reasoning.
For developers and even ordinary business personnel, this means more agile development and construction capabilities and a higher technical starting point.
In addition, the synergy of AI, application and data also makes the industry knowledge model, industry application asset and data asset model sink further. This is undoubtedly of positive significance for the landing of AI in the industry and the creation of new value in the industry.
From the perspective of the whole industry, many industry analysts have pointed out that platform is the general trend to promote the digital transformation of enterprises.
In the era of digital economy, computing power is a new productivity, data is a new means of production, and cutting-edge technologies such as 5g, AI and cloud are new production tools. In this context, moving closer to and integrating with AI platforms such as Huawei cloud is more conducive to enterprises’ digital transformation and commercial success.
As a platform enterprise, Huawei will also connect and empower the whole industrial chain.
Not just for business partners. As Guo Ping, Huawei’s rotating chairman, said at the opening ceremony of the conference, this is also a way to help Huawei tide over its difficulties.
The hardware is blocked, but when “soft” services such as modelarts are updated from generation to generation and more resources are invested, Huawei, which is more balanced between software and hardware, also has more confidence on the way to break through.