Summary: 90% savings in training and labeling costs! HUAWEI CLOUD Automation AI development platform ModelArts 3.0 was released, providing a one-stop connection from training data to model landing.
This year's Huawei has really encountered a lot of difficulties.
In particular, the supply chain, including the crackdown on chips, made Huawei's rotating chairman Guo Pingcheng "it has indeed brought great difficulties to Huawei's production and operations."
However, despite the turbulent fate, Huawei is still on the scene of fully connected, conveying confidence:
"Always face the sun, and the shadow will be left behind by you."
Huawei also used specific business performance and future investment to prove that this confidence is not just a slogan.
Taking Huawei Cloud as an example, Guo Ping said that the cloud is the best platform for unleashing computing power and the digital base of the smart world. After 3 years of continuous efforts, Huawei Cloud now has 23 regional centers around the world, serving 1.5 million developers .
What more directly illustrates the "future" is continuous investment in products.
Specifically , we can see the clues from ModelArts 3.0 launched at the finale of the 2020 Huawei Full Connect Conference .
Not much to say, let's take a look.
ModelArts 3.0 released
ModelArts, as a masterpiece of Huawei Cloud AI development platform, can provide AI application development services including data annotation preparation, model training, model tuning, and model deployment.
Soon after its launch in 2018, the Stanford DAWNBenchmark was listed. In the total training time of image recognition (ResNet50-on-ImageNet, accuracy above 93%), with a score of 10 minutes and 28 seconds, it was nearly 44% faster than the second place, and won the world's first place at the time.
After the iteration of version 2.0 last year, Huawei Cloud ModelArts has evolved into a minimalist, professional one-stop AI development management platform that can complete model training with zero code and one-click deployment.
So this year, what new breakthroughs can ModelArts 3.0 have?
At the fully connected site, Tian Qi, the chief scientist in the AI field of Huawei Cloud and IEEE Fellow, revealed the answer.
Tian Qi introduced that ModelArts 3.0, as an AI development platform for AI in the industry, is aimed at "how to train high-precision models with very little data", "how to lower the threshold of enterprise application AI", and "how to solve the problem of enterprises in the safe use of data. "Worries" and other issues have been explored and studied, and four new features have been brought to this end.
Now, the AI capabilities accumulated by HUAWEI CLOUD for a long time, such as automatic machine learning, small sample learning, federated learning, and pre-training models, can be deployed on the ModelArts platform in a plug-and-play manner to help AI landing.
Four new features
EI-Backbone: A new paradigm for AI development
The first is the new release of EI-Backbone, a general pre-training model architecture.
Its purpose is to create an efficient training model with pre-training model + small sample fine-tuning , and comprehensively improve the industry's AI landing ability and experience.
In other words, EI-Backbone will provide general pre-training models and industry-customized development processes, so that the formed development experience can be replicated on a large scale and lower the threshold for AI use.
If the pre-training model in the NLP field is the benchmark, then the long-term goal of EI-Backbone is to create a BERT in the CV field.
The reason why BERT is called "the beginning of a new era of NLP" is not just because it was on the NLP list at the beginning of its birth. It is even more because, based on the BERT pre-training model, only a simple migration strategy is needed to enable the NLP model to obtain good performance in downstream tasks.
This undoubtedly greatly promotes the research and development of natural language processing.
And EI-Backbone is dedicated to replicating the BERT experience for developers in the CV field.
Take medical image segmentation as an example. In the past, hundreds or thousands of labeled data were needed for training. With the blessing of EI-Backbone, only dozens or even dozens of labeled data can be completed, saving more than 90% of labeling costs. .
Tian Qi introduced that in the past, model selection and hyperparameter adjustment, which 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 hyperparameter optimization technology provided by EI-Backbone , And greatly improve accuracy.
Combining Huawei Cloud's computing resource allocation and data management, the entire development process of model training, testing, acceptance, and deployment, after loading the EI-Backbone integrated pre-training model, it can be shortened to a few hours or even minutes to complete, reducing training costs More than 90.
At present, EI-Backbone has proven successful cases in more than 10 industries, and has won more than 10 industry challenge championships. Around EI-Backbone, Huawei Cloud has also published more than 100 related papers.
The related model architecture will gradually be open sourced .
Federated learning: breaking data silos
The second new feature is the Federated Learning feature added to ModelArts 3.0.
Data is undoubtedly the basis of AI applications. Only based on diversified data can AI intelligent perception be realized.
But in the actual AI landing, there is often such a problem: data is scattered among different data controllers, limited by privacy, security and other issues, these data cannot be easily opened up, but form a "data" Isolated island".
This limits the training effects of AI algorithms that have been implemented in the actual industry.
In response to this problem, Huawei Cloud ModelArts provides federated learning features. Users each use local data for training, instead of exchanging the data itself, only using encryption to exchange updated model parameters to achieve joint modeling.
AI intelligent evaluation: automatic Debug, the kind of visualization
For AI development, having abundant data as a basis to complete model training does not mean that you are done.
The evaluation and tuning of model performance is also an important task and requires high requirements for the developer's own experience.
The feature provided by ModelArts 3.0 in this link is AI intelligent evaluation.
Its model evaluation function is to send the model inference results, original images and real labels to the model evaluation module after getting the first trained model.
This module will evaluate the comprehensive capabilities of the model from two aspects: data and model. Evaluation indicators include accuracy, performance, credibility and interpretability:
In terms of performance , ModelArts 3.0 can provide operator-level time and space consumption statistical analysis and a variety of overall performance indicators, and give corresponding suggestions for model performance, such as model quantification, distillation, etc.;
In terms of interpretability , ModelArts 3.0 can provide a heat map to show the area on which the model makes reasoning judgments;
In terms of credibility , ModelArts has built-in multiple model credibility related evaluation methods, which can provide multi-angle model security capability evaluation indicators, and provide corresponding defense suggestions based on current model performance.
Finally, the evaluation module will output some diagnostic suggestions to improve the model's capabilities for possible problems.
In other words, ModelArts 3.0 can also automate such heavy work as Debug, and it is also a comprehensive evaluation of the overall process of data to model training.
Flexible computing power + large computing power, inclusive enterprise AI landing
In addition to automated development capabilities, as a one-stop AI platform cloud service, ModelArts also provides computing power support.
In addition, in order to better support the AI research and development that requires large computing power, Huawei's ModelArts platform has made targeted optimizations in the cluster size, number of tasks, 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 100,000-level jobs to run simultaneously and support large-scale distributed tasks with 10,000-level chips.
And, in order to help companies further reduce costs and increase efficiency in the process of AI landing, ModelArts 3.0 also has the core ability of flexible training .
In other words, the optimal number of resources can be adaptively matched according to the requirements of model training speed.
Specifically in terms of products, ModelArts provides two modes.
The first is the Turbo mode , which can make full use of idle resources to accelerate existing training operations. In most typical scenarios, the acceleration efficiency is greater than 80%, and the training speed is increased by 10 times without affecting the model convergence accuracy.
The second is the Economic model , which can provide developers with the ultimate price/performance ratio by maximizing resource utilization, which can increase the price/performance ratio by more than 30% in most typical scenarios.
Leading distributed speedup capability
So, how should ModelArts' current capabilities be evaluated?
Might as well speak directly with data.
The ModelArts platform supports simultaneous operation of 100,000-level enterprise tasks and simultaneous use of 100,000-level users.
The key ability to achieve large-scale cluster distributed training depends on the distributed acceleration ratio.
The test result on MLPerf benchmart shows that under the cluster scale of 512 chips, the Huawei Cloud ModelArts score is 93.6 seconds, which is better than the 120 seconds of Nvidia V100.
MLPerf benchmart is a general benchmark created by the cooperation of academia and industry to measure the training and reasoning performance of machine learning hardware, software and services.
ModelArts' actual landing
The data on paper is only a part.
In fact, Huawei Cloud ModelArts has achieved 160+ landing cases in more than 10 industries such as energy, automobiles, government systems, education, and industrial robots.
For example, the domestic robot dog below has AI capabilities given by ModelArts.
This robot dog is called "Jueying" and is produced by Hangzhou Yunshenchu Technology Co., Ltd.
According to Zhu Qiuguo, founder and CEO of Hangzhou Yunshenchu Technology Co., Ltd., Yunshenchu cooperated with Huawei to use ModelArts and Atlas 200DK to give AI capabilities to the "Zeying" series of robot dogs, and is committed to intelligent inspections in the factory park. The robot dog can perceive the on-site environment in real time, and through the interactive analysis of the knowledge map, it can strengthen the learning dynamic decision-making, and has the ability of complex path planning and actions. The terminal cloud cooperates to protect the safety of the factory park.
At the scene, Huawei Cloud demonstrated how to build the " perception + cognition + decision-making " capabilities for the robot dog based on ModelArts : After training, " Juying " can not only detect flames, but also understand that fire is harmful to the factory environment. Learn and plan to generate a new path, bypassing the flame, and turning off the control valve to relieve the fire.
Another interesting application case comes from students from Tongji University. The landing scene is migratory bird protection .
Five students from Tongji University, through HUAWEI CLOUD ModelArts, trained a model that can recognize thousands of bird species in just a few months, and established a wetland digital twin system.
Moreover, this system has already played a role in migratory bird protection and scientific research in Hangzhou Bay.
▶If you also want to use ModelArts to develop some valuable AI applications , click on the link to try the experience, and now invite friends, AI computing power, and Huawei watch bracelets are waiting for you. ☚
How to evaluate ModelArts?
Having said so much, how should we evaluate the significance of ModelArts to the industry AI landing?
As Zheng Yelai, President of Huawei Cloud Business, said on the spot, a development model like ModelArts has become a new choice.
Now, ModelArts has realized the simplification and automation of the entire process to upgrade the existing AI development model, allowing the entire chain of data preparation, algorithm development, model training, model management, and model reasoning to take a qualitative leap.
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 the three aspects of AI, applications, and data has also caused the industry knowledge model, industry application assets and data asset models to further sink. This will undoubtedly have positive significance for the landing of industry AI and the creation of new industry value.
From the perspective of the entire industry, many industry analysts have pointed out that platformization is the general trend that promotes the digital transformation of enterprises.
In the digital economy era, 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 and merging to an AI platform such as Huawei Cloud will be more conducive to the digital transformation of enterprises and commercial success.
As a platform company, Huawei will also connect and empower the entire industry chain.
It's not just a boost to business partners. As Huawei’s rotating chairman Guo Ping said at the opening ceremony of the conference, this is also a way to help Huawei tide over difficulties.
Hardware is blocked, but when "soft" services like ModelArts are updated generation after generation and more resources are invested, Huawei, which is more and more balanced between soft and hard, has more confidence on the road to break through.
Click to follow and learn about Huawei Cloud's fresh technology for the first time~