product overview - huawei cloud · customers. different projects have different requirements for...

26
ModelArts Product Overview Issue 1.0.15 Date 2020-03-13 HUAWEI TECHNOLOGIES CO., LTD.

Upload: others

Post on 12-Mar-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

ModelArts

Product Overview

Issue 1.0.15

Date 2020-03-13

HUAWEI TECHNOLOGIES CO., LTD.

Page 2: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

Copyright © Huawei Technologies Co., Ltd. 2020. All rights reserved.

No part of this document may be reproduced or transmitted in any form or by any means without priorwritten consent of Huawei Technologies Co., Ltd. Trademarks and Permissions

and other Huawei trademarks are trademarks of Huawei Technologies Co., Ltd.All other trademarks and trade names mentioned in this document are the property of their respectiveholders. NoticeThe purchased products, services and features are stipulated by the contract made between Huawei andthe customer. All or part of the products, services and features described in this document may not bewithin the purchase scope or the usage scope. Unless otherwise specified in the contract, all statements,information, and recommendations in this document are provided "AS IS" without warranties, guaranteesor representations of any kind, either express or implied.

The information in this document is subject to change without notice. Every effort has been made in thepreparation of this document to ensure accuracy of the contents, but all statements, information, andrecommendations in this document do not constitute a warranty of any kind, express or implied.

Huawei Technologies Co., Ltd.Address: Huawei Industrial Base

Bantian, LonggangShenzhen 518129People's Republic of China

Website: https://www.huawei.com

Email: [email protected]

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. i

Page 3: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

Contents

1 What Is ModelArts?................................................................................................................ 1

2 Functions................................................................................................................................... 4

3 Basic Knowledge...................................................................................................................... 63.1 Introduction to the AI Development Lifecycle...............................................................................................................63.2 Basic Concepts of AI Development................................................................................................................................... 73.3 Common Concepts of ModelArts.......................................................................................................................................93.4 Data Management............................................................................................................................................................... 103.5 DevEnviron.............................................................................................................................................................................. 113.6 Model Training....................................................................................................................................................................... 123.7 Model Deployment...............................................................................................................................................................143.8 ExeML........................................................................................................................................................................................15

4 Related Services.....................................................................................................................17

5 How Do I Access ModelArts?.............................................................................................. 19

6 Permissions Management................................................................................................... 20

7 Quotas......................................................................................................................................23

ModelArtsProduct Overview Contents

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. ii

Page 4: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

1 What Is ModelArts?

ModelArts is a one-stop development platform for AI developers. With datapreprocessing, semi-automated data labeling, distributed training, automatedmodel building, and on-demand model deployment on the device, edge, andcloud, ModelArts helps AI developers build models quickly and manage the AIdevelopment lifecycle.

The one-stop ModelArts platform covers all stages of AI development, includingdata processing, algorithm development, and model training and deployment. Theunderlying technologies of ModelArts support various heterogeneous computingresources, allowing developers to flexibly select and use resources. In addition,ModelArts supports popular open source AI development frameworks such asTensorFlow and MXNet. However, developers can use self-developed algorithmframeworks to match your usage habits.

ModelArts aims to simplify AI development.

ModelArts is suitable for AI developers of with varying levels of developmentexperience. Service developers can use ExeML to quickly build AI applicationswithout coding. Beginners can directly use built-in algorithms to build AIapplications. AI engineers can use multiple development environments to compilecode for quick modeling and application development.

Product ArchitectureModelArts is a one-stop AI development platform that supports the entiredevelopment process, including data processing, model training, management,and deployment, and provides AI market for sharing models.

ModelArts supports various AI application scenarios, such as image classification,object detection, video analysis, speech recognition, product recommendations,and exception detection.

ModelArtsProduct Overview 1 What Is ModelArts?

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 1

Page 5: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

Figure 1-1 ModelArts architecture

Product Advantages● One-stop

The out-of-the-box and full-lifecycle AI development platform provides one-stop data processing, model development, training, management, anddeployment.

● Easy to use– Multiple built-in models provided and free use of open source models– Automatic optimization of hyperparameters– Code-free development and simplified operations– One-click deployment of models to the cloud, edge, and devices

● High performance– The self-developed MoXing deep learning framework accelerates

algorithm development and training.– Optimized GPU utilization accelerates real-time inference.– Models running on Ascend AI chips achieve more efficient inference.

● Flexible– Popular open source frameworks available, such as TensorFlow and

Spark_MLlib– Popular GPUs and self-developed Ascend chips available– Exclusive use of dedicated resources– Custom images for custom frameworks and operators

Using ModelArts for the First Time

If you are a first-time user, get familiar with the following information:

● Basic conceptsBasic Knowledge describes the basic concepts of ModelArts, including thebasic process and concepts of AI development, and specific concepts andfunctions of ModelArts.

● Getting started

ModelArtsProduct Overview 1 What Is ModelArts?

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 2

Page 6: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

The Getting Started document provides detailed operation guides to walkyou through model building on ModelArts.

● Best practicesModelArts supports multiple open source engines and provides extensive usecases based on the engines and functions. You can build and deploy modelsby referring to Best Practices.

● Other functions and operation guides– If you are a service developer, you can use ExeML to quickly build models

without coding. For details, see User Guide (ExeML).– If you are a beginner, you can use common AI algorithms to quickly build

models without coding. ModelArts provides built-in algorithms based onthe common AI engines. For details, see User Guide (AI Beginners).

– If you are an AI engineer, you can manage the AI development lifecycle,including data management, and model development, training,management, and deployment. For details, see User Guide (Senior AIEngineers).

– If you are a developer and want to use ModelArts APIs or SDKs for AIdevelopment, refer to API Reference or SDK Reference.

ModelArtsProduct Overview 1 What Is ModelArts?

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 3

Page 7: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

2 Functions

AI engineers face challenges in installation and configuration of various AI tools,data preparation, and model training. To address these challenges, HUAWEICLOUD provides the one-stop AI development platform ModelArts for developers.The platform integrates data preparation, algorithm development, model training,and model deployment into the production environment, allowing developers toperform one-stop AI development. The following figure shows the functions of theModelArts.

Figure 2-1 Function overview

ModelArts has the following features:

● Data governanceManages data preparation, such as data filtering and labeling, and datasetversions.

● Rapid and simplified model trainingThe self-developed MoXing deep learning framework enables high-performance distributed training and simplifies coding.

● Model deploymentDeploys models in various production environments such as devices, the edge,and the cloud, and supports online and batch inference.

● Auto learningSupports model building without coding and supports auto learning withimage classification, object detection, and predictive analysis.

ModelArtsProduct Overview 2 Functions

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 4

Page 8: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

● Visualized workflowUses Graph Engine Service (GES) to manage and visualize the lifecycle of AIdevelopment workflows, implementing data and model lineage.

● AI MarketSupports commonly used algorithms and datasets, and internal or publicsharing of enterprise models in the market.

ModelArtsProduct Overview 2 Functions

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 5

Page 9: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

3 Basic Knowledge

3.1 Introduction to the AI Development Lifecycle

What Is AI

AI (artificial intelligence) is a technology capable of simulating human cognitionthrough machines. The core capability of AI is to make judgment or predictionbased on a given input.

What Is the Purpose of AI Development

AI development aims to centrally process and extract information from volumes ofdata to summarize internal patterns of the study objects.

Massive volumes of collected data are computed, analyzed, summarized, andorganized by using appropriate statistics, machine learning, and deep learningmethods to maximize data value.

Basic Process of AI Development

The basic process of AI development can be divided into the following steps:determining the objective, preparing data, and training, evaluating, and deployinga model.

Figure 3-1 AI development process

Step 1 Determine the objective.

Before starting data analysis, determine what to analyze. Who is your data object?What problems do you want to solve? What is the business goal? Data analystsneed to focus on these questions, and develop the analysis framework androadmap based on business understanding, for example, to reduce the loss ofexisting customers, optimize the activity effect, and improve the response rate of

ModelArtsProduct Overview 3 Basic Knowledge

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 6

Page 10: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

customers. Different projects have different requirements for data and usedifferent analysis methods.

Step 2 Prepare data.

Data preparation refers to data collection and preprocessing.

Data preparation is the basis of AI development. When you collect and integraterelated data based on the determined objective, the most important thing is toensure the authenticity and reliability of the obtained data. Typically, you cannotcollect all the data at the same time. In the data labeling phase, you may find thatsome data sources are missing and then you may need to adjust and optimize thedata repeatedly.

Step 3 Train a model.

Modeling involves analyzing the prepared data to find the causality, internalrelationships, and regular patterns, thereby providing references for commercialdecision making. After model training, usually one or more machine learning ordeep learning models are generated. These models can be applied to new data toobtain predictions and evaluation results.

A large number of developers develop and train the models required by theirservices based on mainstream AI engines. Mainstream AI engines includeTensorFlow and MXNet.

Step 4 Evaluate the model.

The entire development process is not completed even though you have obtaineda model because the model needs to be evaluated. Typically, you cannot obtain asatisfactory model the first time. The algorithm parameters and data need to berepeatedly adjusted to evaluate the model generated by the training.

Some common metrics, such as accuracy, recall, and AUC, help you effectivelyevaluate and obtain a satisfactory model.

Step 5 Deploy the model.

Model development and training are based on existing data (which may be testdata). After a satisfactory model is obtained, the model needs to be formallyapplied to actual data or newly generated data for prediction, evaluation, andvisualization. The findings can then be reported to decision makers in an intuitiveway, helping them develop the right business strategies.

----End

3.2 Basic Concepts of AI DevelopmentMachine learning is classified into the following types:

● Supervised learning uses labeled samples to adjust the parameters of aclassifier to achieve the required performance. It is also called supervisedtraining or learning with a teacher. Common supervised learning includesregression and classification.

● Unsupervised learning tries to find a hidden structure from unlabeled data.Common unsupervised learning is clustering.

ModelArtsProduct Overview 3 Basic Knowledge

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 7

Page 11: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

● Reinforcement learning is an area of machine learning concerned with howsoftware agents ought to take actions in an environment so as to maximizesome notion of cumulative reward.

RegressionRegression reflects the time feature of data attributes and generates a functionthat maps one data item to an actual variable prediction to find the dependencybetween the variable and attribute. Regression mainly analyzes data order trendfeature, data order prediction, and data relationship. Regression can be used inmarketing, such as customer development, maintaining, customer churnprevention, production lifecycle analysis, sales trend prediction, and targetedpromotional activity.

ClassificationClassification is to find the common features of an object group and divide themas different types based on the categorization modes. The purpose is to map thedata items to a specified type through the classification model. This can be usedfor customer classification, customer properties, feature analysis, customerssatisfaction analysis, and customer purchase trend prediction.

ModelArtsProduct Overview 3 Basic Knowledge

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 8

Page 12: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

ClusteringClustering is to group a set of objects in such a way that objects in the samegroup are more similar to each other than to those in other groups. Clustering canbe used for customer segmentation, customer background analysis, customerpurchase trend prediction, and market segmentation.

Different from classification, clustering analyzes the data objects withoutconsidering the known class labels (the class labels are not provided in thetraining data). Clustering can produce such labels. Objects are grouped based onthe maximized and minimized similarities to form clusters. In this way, objects inthe same cluster are more similar to each other than to those in other clusters.

3.3 Common Concepts of ModelArts

ExeMLExeML is the process of automating model design, parameter tuning, and modeltraining, compression, and deployment with the labeled data. The process iscoding-free and does not require developers to have experience in modeldevelopment. A model can be built in three steps: labeling data, training a model,and deploying the model.

Device-Edge-CloudDevice-Edge-Cloud indicates devices, Huawei intelligent edge devices, andHUAWEI CLOUD.

Real-Time InferenceReal-time inference specifies a web service that provides inference result for eachinference request.

Batch InferenceBatch inference specifies a batch job that processes batch data for inference.

ModelArtsProduct Overview 3 Basic Knowledge

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 9

Page 13: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

Ascend ChipThe Ascend chips are a series of Huawei-developed AI chips with high computingperformance and low power consumption.

Resource PoolHUAWEI CLOUD provides large-scale computing clusters for model development,training, and deployment. There are two types of resource pools: public resourcepool and dedicated resource pool. The public resource pool is provided by defaultand is billed in pay-per-use mode. Dedicated resource pools must be purchasedand are used exclusively.

AI MarketCommon models, datasets, and APIs are provided. You can also share your ownmodels, datasets, or APIs with other users or make them publicly available.

3.4 Data ManagementDuring AI development, massive volumes of data need to be processed, and datapreparation and labeling usually take more than half of the development time.The ModelArts data framework provides the data collection, data filtering, datalabeling, dataset version management, auto and semi-auto data filtering, andauto data labeling functions, and an auto labeling tool. AI developers canimplement data labeling based on the framework. See Figure 3-2.

Figure 3-2 Data labeling process

ModelArts supports various AI application scenarios, such as computer vision,natural language processing, and audio and video analysis. It allows labeling forimage classification, object detection, image segmentation, speech segmentation,and text classification. In addition, ModelArts supports data processing and pre-labeling for autonomous driving, medical images, and remote sensing images.

ModelArts provides labeling tools for the preceding scenarios. It allows project-based management for labeling by individual developers, small-scale labeling byamateur teams, and large-scale labeling by professional teams. For large-scaleteam labeling, ModelArts provides project management, personnel management,

ModelArtsProduct Overview 3 Basic Knowledge

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 10

Page 14: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

and data management are provided to implement the entire process, from projectcreation, allocation, management, labeling, to acceptance. For small-scale labelingby amateur teams, ModelArts provides an easy-to-use labeling tool to minimizeproject management costs.

The labeling platform ensures data security. User data is used only within theauthorized scope. The labeling object allocation policy ensures user data privacyand implements data anonymization.

The marking tool supports 2D, 3D, polygon, point, straight line, mask, and curve asbounding boxes. Pixel-level segmentation and data pre-labeling are supported.Data is automatically labeled after it is loaded using the built-in or customalgorithm. Only a few manual modifications are required. ModelArts supportsassisted auto data labeling. After labeling personnel label data, the toolautomatically labels accurate information (such as contour information), whichenables a 10-fold efficiency in manual labeling for specific fields.

3.5 DevEnvironIt is challenging to set up a development environment, select an AI algorithmframework and proper algorithm, debug code, and install software or acceleratehardware. To address these challenges, ModelArts simplifies the entiredevelopment process and lowers the development threshold. Figure 3-3 shows thealgorithm development process.

Figure 3-3 Algorithm development

● Support for all popular AI algorithm frameworksIn the machine learning and deep learning fields, popular open sourcetraining and inference frameworks include TensorFlow, PyTorch, and MXNet.To adapt to different development habits and application scenarios,ModelArts supports all popular AI computing frameworks and provides auser-friendly development and debugging environment. It supports traditionalmachine learning algorithms, such as logistic regression, decision tree, and

ModelArtsProduct Overview 3 Basic Knowledge

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 11

Page 15: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

clustering, as well as multiple types of deep learning algorithms, such as CNN,RNN, and LSTM.

● Simplified algorithm development for distributed trainingDeep learning usually requires large-scale GPU clusters for distributedacceleration. However, existing open source frameworks require algorithmdevelopers to write a large amount of code to implement distributed trainingon different hardware, and acceleration code of different frameworks varies.The MoXing framework of ModelArts, a lightweight distributed framework orSDK built on deep learning engines such as TensorFlow, PyTorch, and MXNet,addresses these pain points. In addition, the distributed computing enginefeatures higher performance and is easier to use. Figure 3-4 shows the codedeveloped based on the MoXing framework.

Figure 3-4 Development code based on MoXing

– Simplified parameter tuning: Multiple parameter tuning skill packages areintegrated, for example, the data augmentation policy, which simplifiesmodel tuning for AI algorithm engineers.

– Simplified distributed acceleration: Automatic distributed acceleration ofstandalone coding simplifies distributed acceleration and improvesperformance, eliminating the need for algorithm engineers to haveknowledge of distribution.

3.6 Model TrainingIn addition to data and algorithms, developers spend a lot of time in modelparameter design. Model training parameters directly affect the model's precisionand convergence time. Parameter selection is heavily dependent on developers'experience. Improper parameter selection will affect the model's precision orsignificantly increase the time required for model training.

To simplify AI development and improve the development efficiency and trainingperformance, ModelArts automatically performs hyperparameter optimizationbased on machine learning and reinforcement learning. It provides automatic

ModelArtsProduct Overview 3 Basic Knowledge

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 12

Page 16: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

hyperparameter tuning policies such as learning rate and batch size, andintegrates common models.

Currently, when most developers develop models, the models usually have dozensof layers or even hundreds of layers and MB-level or even GB-level parameters tomeet precision requirements. As a result, the specifications of computing resourcesare extremely high, especially the computing power of hardware resources,memory, and ROM. The resource specifications on the device side are strictlylimited. For example, the computing power on the device side is 1 TFLOPS, thememory size is about 2 GB, and the ROM space is about 2 GB, so the model sizeon the device side must be limited to 100 KB, and the inference delay must belimited to 100 milliseconds. If the device side is a mobile phone that has stricterlimitation on power consumption and heat generation, a lightweight model isrequired.

Therefore, compression technologies with lossless or near-lossless model precision,such as pruning, quantization, and knowledge distillation, are used to implementautomatic compression and optimization of a model, and automatic iteration ofmodel compression and retraining to control the loss of model precision. The low-bit quantization technology that does not need to be retrained converts the modelfrom a high-precision floating point to a fixed-point operation. Multiplecompression and optimization technologies are used to meet the lightweightrequirements of device and edge hardware resources. The model compressiontechnology reduces the precision by less than 1% in specific scenarios, and themodel size is reduced by 10 fold.

When the training data volume is large, the training of the deep learning model istime-consuming. In computer vision, ImageNet-1k (a classification datasetcontaining 1,000 image classes, hereinafter referred to as ImageNet) is classic andcommonly used dataset. If you use a P100 GPU to train a ResNet-50 model on thedataset, it will take nearly one week. This hinders the development of deeplearning applications. Therefore, the acceleration of deep learning training hasalways been an important concern to academia and the industry.

Distributed training acceleration needs to be considered in terms of software andhardware. A single optimization method cannot achieve the expectation.Therefore, optimization of distributed acceleration is a system project. Thedistributed training architecture needs to be considered in terms of hardware (chipand hardware design). For example, factors such as overall compute specifications,network bandwidth, high-speed cache, power consumption, and heat dissipationof the system need to be considered, as well as the relationship between computeand communication throughput. In this way, compute and communication delaysare minimized.

The software design needs to combine high-performance hardware features tofully use the high-speed hardware network to implement high-bandwidthdistributed communication and efficient local data caching. By using trainingoptimization algorithms, such as hybrid parallel, gradient compression, andconvolution acceleration, the software and hardware of the distributed trainingsystem can be efficiently coordinated and optimized from end to end, and trainingacceleration is implemented in the distributed environment of multiple hosts andcards. ModelArts delivers an industry-leading speedup of over 0.8 for ResNet50 onthe ImageNet dataset in the distributed environment with thousands of hosts andcards.

ModelArtsProduct Overview 3 Basic Knowledge

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 13

Page 17: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

To measure the acceleration performance of distributed deep learning, thefollowing two key indicators are used:

● Throughput, that is, the amount of data processed in a unit time● Convergence time, that is, the time required to achieve certain precision

The throughput depends on server hardware (for example, more AI accelerationchips with higher FLOPS processing capabilities and higher communicationbandwidth), data reading and caching, data preprocessing, model computing (forexample, convolution algorithm selection), and communication topologyoptimization. Except low-bit computing and gradient (or parameter) compression,most technologies improve throughput without affecting model precision. Toachieve the shortest convergence time, you need to optimize the throughput andadjust the parameters. If the parameters are not adjusted properly, the throughputcannot be optimized. If the batch size is set to a small value, the parallelperformance of model training is relatively poor. As a result, the throughputcannot be improved by increasing the number of compute nodes.

Users concern the convergence time most. Therefore, the MoXing framework ofModelArts implements full-stack optimization and greatly shortens the trainingconvergence time. In terms of data read and preprocessing, MoXing uses multi-level concurrent input pipelines to prevent data I/Os from becoming a bottleneck.In the aspect of model computing, MoXing provides the hybrid precisioncalculation combining semi-precision and single-precision for the upper layermodels, and reduces the loss caused by precision calculation through adaptivescaling. From the perspective of hyperparameter optimization, dynamichyperparameter policies (such as momentum and batch size) are used tominimize the number of epochs required for model convergence. For underlyingoptimization, MoXing works with underlying Huawei servers and communicationcomputing libraries to further improve distributed acceleration.

ModelArts High-Performance Distributed Training Optimization● Automatic hybrid precision for training to fully utilize hardware computing

capabilities● Dynamic hyperparameter adjustment technologies (dynamic batch size,

image size, and momentum)● Automatic model gradient merging and splitting● Communication operator scheduling optimization based on BP bubble

adaptive computing● Distributed high-performance communication libraries (NStack and HCCL)● Distributed data-model hybrid parallel● Training data compression and multi-level caching

3.7 Model DeploymentGenerally, AI model deployment and large-scale implementation are complex.

For example, in a smart transportation project, the trained model needs to bedeployed to the cloud, edges, and devices. It takes time and effort to deploy themodel on the devices, for example, deploying a model on cameras of differentspecifications and vendors at a time. ModelArts supports one-click deployment of

ModelArtsProduct Overview 3 Basic Knowledge

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 14

Page 18: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

a trained model on various devices for different application scenarios. In addition,it provides a set of secure and reliable one-stop deployment modes for individualdevelopers, enterprises, and device manufacturers.

Figure 3-5 Process of deploying a model

● The real-time inference service features high concurrency, low latency, andelastic scaling, and supports multi-model gray release and A/B testing.

● Models can be deployed as real-time and batch inference services on thecloud, edge, and devices.

● Models can be directly pushed to edge devices with just one click. You onlyneed to select the edge node and ModelArts pushes the model to the node.

● ModelArts is deeply optimized based on the high-performance AI inferencechip Ascend 310 developed by Huawei. It can process PBs of inference datawithin a single day, publish over 1 million inference APIs on the cloud, andcontrols inference network latency within milliseconds.

3.8 ExeMLTo implement AI in various industries, AI model development must be simplified.Currently, only a few algorithm engineers and researchers are capable of AIdevelopment and optimization, and most algorithm engineers are capable of onlyalgorithm prototype development. They lack the capability of developing relatedprototypes into products and projects. Most service developers, however, are notcapable of developing AI algorithms and optimizing parameters. As a result, mostenterprises do not have AI development capabilities.

ModelArts uses machine learning to help service developers who do not have thealgorithm development capabilities develop algorithms. It automatically generatesmodels based on transfer learning and Neural Architecture Search (NAS), selectsparameters for model training, and learns model optimization. This helps servicedevelopers quickly complete model training and deployment. Based on the labeled

ModelArtsProduct Overview 3 Basic Knowledge

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 15

Page 19: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

data and application scenario provided by developers, ModelArts automaticallygenerates models that meet the precision requirements of users without coding.The application scenarios include image classification, object detection, predictiveanalytics, and sound classification. Models can be automatically optimized andgenerated based on the deployment environment and the inference speedrequired by developers.

Figure 3-6 Process of using ExeML

ModelArts ExeML is not only designed for entry-level developers, but also providesthe auto learning white-box capability. It opens model parameters andimplements template-based development. ExeML helps accelerate thedevelopment speed. With ExeML, developers can directly optimize the generatedmodel or retrain the model, instead of setting up a new model.

The key techniques of ExeML are tree search-based optimal featuretransformation and Bayesian optimization for automatic parameter adjustmentbased on the maximum information entropy model. With these key technologies,data features and patterns can be automatically learned from enterprise relational(structured) data, and features and machine learning models and parameters canbe intelligently optimized. The accuracy can be as good as that optimized byexperts. The key technologies of auto deep learning are transfer learning (high-quality models are generated based on a small amount of data), auto design ofthe model architecture in multiple dimensions (neural network search andadaptive model optimization), and auto optimization and training of trainingparameters.

ModelArtsProduct Overview 3 Basic Knowledge

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 16

Page 20: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

4 Related Services

IAMModelArts uses Identity and Access Management (IAM) for authentication andauthorization. For more information about IAM, see the Identity and AccessManagement User Guide.

OBSModelArts uses Object Storage Service (OBS) to store data and model backupsand snapshots, achieving secure, reliable, and low-cost storage. For moreinformation about OBS, see the Object Storage Service Console OperationGuide.

Table 4-1 Relationship between ModelArts and OBS

Function Sub Task Relationship

ExeML Datalabeling

The data labeled on ModelArts is stored inOBS.

Autotraining

After a training job is completed, thegenerated model is stored in OBS.

Modeldeployment

ModelArts deploys models stored on OBS asreal-time services.

AI developmentlifecycle

Datamanagement

● Datasets are stored in OBS.● The dataset labeling information is stored

in OBS.● Data can be imported from OBS.

Developmentenvironment

Data or code files in the notebook instanceare stored in OBS.

ModelArtsProduct Overview 4 Related Services

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 17

Page 21: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

Function Sub Task Relationship

Modeltraining

● The datasets used by training jobs arestored in OBS.

● The running scripts of training jobs arestored in OBS.

● The models generated by training jobs arestored in the specified OBS paths.

● The run logs of training jobs are stored inthe specified OBS paths.

Modelmanagement

After a training job is completed, thegenerated model is stored in OBS. You canimport the model from OBS.

Servicedeployment

The models stored in OBS can be deployed asservices.

Settings - Obtain the access key and configure it onModelArts so that ModelArts can use OBS tostore data and create notebook instances.

EVSModelArts uses Elastic Volume Service (EVS) to store created notebook instances.For more information, see the Elastic Volume Service User Guide.

CCEModelArts uses Cloud Container Engine (CCE) to deploy models as online servicesfor high concurrency and elastic scaling. For more information about CCE, see theCloud Container Engine User Guide.

Batch ServiceModelArts uses Batch Service to accelerate model training with scheduling anddistributed cloud computing capabilities. For more information about BatchService, see the Batch Service User Guide.

GESModelArts uses Graph Engine Service (GES) to visualize workflows. For moreinformation about GES, see the Graph Engine Service User Guide.

ModelArtsProduct Overview 4 Related Services

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 18

Page 22: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

5 How Do I Access ModelArts?

You can access ModelArts on a web-based management console or by usingHTTPS-based application programming interfaces (APIs).

● Using the Management ConsoleModelArts provides a simple and easy-to-use management console, includingExeML, data management, development environment, model training, modelmanagement, service deployment, and AI Market. You can complete end-to-end AI development on the management console.To use the ModelArts management console, register with HUAWEI CLOUDfirst. If you have registered with HUAWEI CLOUD, choose EnterpriseIntelligence > Essential Platform > ModelArts on the official website andlog in to the management console. If you do not have an account, follow theinstructions for registering on the HUAWEI CLOUD management console.

● Using SDKsIf you want to integrate ModelArts into a third-party system for secondarydevelopment, call the SDK to complete the development. ModelArts SDKsencapsulate RESTful APIs provided by ModelArts to simplify userdevelopment. For details about the SDKs and operations, see the SDKReference.In addition, you can directly call the ModelArts SDKs when writing code in anotebook on the management console.

● Using APIsIf you want to integrate ModelArts into a third-party system for secondarydevelopment, use the APIs to access ModelArts. For details about the APIs andoperations, see API Reference.

ModelArtsProduct Overview 5 How Do I Access ModelArts?

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 19

Page 23: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

6 Permissions Management

If you need to assign different permissions to employees in your enterprise toaccess your ModelArts resources, IAM is a good choice for fine-grainedpermissions management. IAM provides identity authentication, permissionsmanagement, and access control, helping you securely access your HUAWEICLOUD resources.

With IAM, you can use your HUAWEI CLOUD account to create IAM users for youremployees, and assign permissions to the users to control their access to specificresource types. For example, some software developers in your enterprise need touse ModelArts resources but must not delete them or perform any high-riskoperations. To achieve this result, you can create IAM users for the softwaredevelopers and grant them only the permissions required for using ModelArtsresources.

If the HUAWEI CLOUD account has met your requirements, you do not need tocreate an independent IAM user for permission management. Then you can skipthis section. This will not affect other functions of ModelArts.

IAM can be used free of charge. You pay only for the resources in your account.For more information about IAM, see IAM Service Overview.

ModelArts Permissions

By default, new IAM users do not have any permissions assigned. You need to adda user to one or more groups, and assign permissions policies or roles to thesegroups. The user then inherits permissions from the groups it is a member of. Thisprocess is called authorization. After authorization, the user can perform specifiedoperations on ModelArts based on the permissions.

To assign ModelArts permissions to a user group, specify the scope as region-specific projects and select projects for the permissions to take effect. If Allprojects is selected, the permissions will take effect for the user group in allregion-specific projects. When accessing ModelArts, the users need to switch to aregion where they have been authorized to use cloud services.

You can grant users permissions by using roles and policies. Policies are currentlyunder open beta testing. You can apply to use policies for fine-grained accesscontrol free of charge. For more information, see Applying for Policy-basedAccess Control.

ModelArtsProduct Overview 6 Permissions Management

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 20

Page 24: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

● Policies: A type of fine-grained authorization mechanism that definespermissions required to perform operations on specific cloud resources undercertain conditions. This mechanism allows for more flexible policy-basedauthorization, meeting requirements for secure access control. For example,you can grant ECS users only the permissions for managing a certain type ofECSs. For the API actions supported by ModelArts, see API Reference >Permissions Policies and Supported Actions.

Table 6-1 lists all the system-defined roles and policies supported by ModelArts.

Table 6-1 System-defined policies supported by ModelArts

Policy Name Description Policy Type

ModelArtsFullAccess

Administrator permissions forModelArts. Users granted thesepermissions can operate and useModelArts.

System-definedpolicy

ModelArtsCommonOperations

Common user permissions forModelArts. Users granted thesepermissions can operate and useModelArts except manage dedicatedresource pools.

System-definedpolicy

Table 6-2 lists the common operations supported by each system policy ofModelArts. Please choose proper system policies according to this table.

Table 6-2 Common operations supported by each system policy

Operation ModelArts Admin ModelArts User

ExeML Yes Yes

Labeling data Yes Yes

Managing data Yes Yes

Developmentenvironment

Yes Yes

Managing models Yes Yes

Deploying services Yes Yes

AI Market Yes Yes

Dedicated resourcepools

Yes No

Settings Yes Yes

ModelArtsProduct Overview 6 Permissions Management

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 21

Page 25: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

Helpful Links● IAM Service Overview● Creating a User Group and User and Granting ModelArts Permissions● Permissions Policies and Supported Actions

ModelArtsProduct Overview 6 Permissions Management

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 22

Page 26: Product Overview - HUAWEI CLOUD · customers. Different projects have different requirements for data and use different analysis methods. Step 2 Prepare data. Data preparation refers

7 Quotas

ModelArts uses the following infrastructure resources:

● ECS● EVS● VPC

For details about how to view and modify the quota, see Quotas.

ModelArtsProduct Overview 7 Quotas

Issue 1.0.15 (2020-03-13) Copyright © Huawei Technologies Co., Ltd. 23