人工智能平台PAI（原机器学习平台PAI）

产品介绍人工智能平台PAI（Platform for AI） 面向企业客户及开发者提供 易上手、高性价比、高性能、方便扩展、具备多种行业场景插件的机器学习/深度学习工程化平台。内置 140+种优化算法，提供包括 数据标注（PAI-iTAG）、模型构建（PAI-Designer、PAI-DSW）、模型训练（PAI-DLC）、编译优化、推理部署（PAI-EAS）等全流程AI工程化能力。针对大规模深度学习及融合智算场景， 推出 PAI灵骏智算服务的PaaS产品，包括公共云Serverless版、单租版以及混合云形态，基于软硬件一体优化技术，构建高性能异构算力底座，提供AI工程化全流程能力，具备高可用、高性能、易用等核心优势，满足大模型训练、自动驾驶、基础科研、金融等领域的高性能计算需求。详细介绍AI 资源池（AI计算资源组）：支持云上云通用计算资源（ECS、ECI、神龙裸金属）、灵骏智算资源、大数据MaxCompute、Flink等资源引擎的创建和管理，基于AI特性提供多样的调度能力，为用户AI开发、训练、推理提供高效的AI算力资源。数据准备：在数据准备阶段，PAI-iTAG 提供智能化数据标注服务，支持图像、文本、视频、音频等不同类型数据标注，支持多模态数据标注；提供丰富的标注内容组件和题目组件，用户可以直接使用平台预置的标注模板，也可以自定义模板进行数据标注。同时提供全托管的数据标注外包服务。模型开发：在模型开发阶段，可通过 PAI-Designer 和 PAI-DSW 两款开发工具来完成建模。可视化建模 PAI-Designer：提供低代码开发环境，内置100+成熟的机器学习算法，通过拖拉拽完成建模，帮助用户实现低代码开发人工智能相关服务。交互式建模 PAI-DSW：提供交互式编程环境，内置JupyterLab、WebIDE及Terminal，提供底层Sudo权限，开放灵活。模型训练：在模型训练阶段，可通过 PAI-DLC 和 PAI-灵骏两款工具来完成训练。智算服务PAI-灵骏：面向大规模深度学习及融合智算场景，支持公共云Serverless版、单租版以及混合云产品形态，提供AI工程化全流程平台及软硬一体的异构融合算力。模型训练 PAI-DLC：云原生深度学习训练平台。支持多种算法框架、超大规模分布式深度学习任务运行和自定义算法框架，具备灵活、稳定、易用和高性能等特点。模型推理：在模型部署阶段，PAI-EAS提供在线预测服务，PAI-Blade提供推理优化服务。模型在线服务 PAI-EAS：支持用户将模型一键部署为在线推理服务或AI-Web应用。适用于实时推理、异步推理、离线推理等多种场景。通用推理加速器 PAI-Blade：Blade的所有优化技术均面向通用性设计，可以应用于不同的业务场景，通过模型系统联合优化，使模型达到最优推理性能。AI资产管理：PAI支持用户对模型、数据集、镜像等重要的AI生产资料及开发产出进行全生命周期管理，并提供AI资产共享、训练效果横向比对、异常问题回溯等能力，实现AI开发及应用过程的降本增效。

Product Introduction: Alibaba Cloud AI Platform PAI (Platform for AI) is an engineering platform for machine learning and deep learning, targeting enterprise customers and developers, with core features including ease of use, high cost-effectiveness, high performance, easy scalability, and multiple industry scenario plugins. It integrates over 140 optimized algorithms, and provides end-to-end AI engineering capabilities covering data annotation (PAI-iTAG), model construction (PAI-Designer, PAI-DSW), model training (PAI-DLC), compilation optimization, and inference deployment (PAI-EAS).

For large-scale deep learning and integrated intelligent computing scenarios, PAI has launched the PaaS product of PAI Lingjun Intelligent Computing Service, which is available in public cloud Serverless edition, dedicated instance edition, and hybrid cloud deployment forms. Based on hardware-software co-optimization technology, it builds a high-performance heterogeneous computing infrastructure, provides end-to-end AI engineering capabilities, and boasts core advantages such as high availability, high performance, and ease of use, meeting the high-performance computing requirements of fields including large model training, autonomous driving, basic research, and finance.

Detailed Introduction of AI Resource Pool (AI Computing Resource Group): It supports the creation and management of general cloud computing resources (ECS, ECI, Shenlong Bare Metal), Lingjun intelligent computing resources, big data engines such as MaxCompute and Flink. Leveraging AI-specific features, it provides diverse scheduling capabilities, delivering efficient AI computing resources for users' AI development, training and inference.

Data Preparation: During the data preparation phase, PAI-iTAG offers intelligent data annotation services, supporting annotation of various data types including images, text, videos, audios, as well as multimodal data annotation. It provides rich annotation content components and question components. Users can directly use the pre-built annotation templates on the platform, or customize templates for data annotation. Additionally, it offers fully-managed data annotation outsourcing services.

Model Development: In the model development stage, users can complete modeling via two development tools: PAI-Designer and PAI-DSW.
- Visual Modeling PAI-Designer: Provides a low-code development environment with over 100 mature machine learning algorithms, enabling users to finish modeling through drag-and-drop operations, helping them develop AI-related services with low code.
- Interactive Modeling PAI-DSW: Offers an interactive programming environment with built-in JupyterLab, WebIDE and Terminal, providing underlying sudo privileges for flexible and open usage.

Model Training: During the model training phase, users can complete training via PAI-DLC and PAI Lingjun.
- Intelligent Computing Service PAI Lingjun: Targets large-scale deep learning and integrated intelligent computing scenarios, supporting public cloud Serverless edition, dedicated instance edition and hybrid cloud product forms, providing end-to-end AI engineering platform and hardware-software co-optimized heterogeneous integrated computing power.
- PAI-DLC: A cloud-native deep learning training platform. It supports multiple algorithm frameworks, running ultra-large-scale distributed deep learning tasks and custom algorithm frameworks, featuring flexibility, stability, ease of use and high performance.

Model Inference: In the model deployment phase, PAI-EAS provides online prediction services, while PAI-Blade offers inference optimization services.
- PAI-EAS (Model Online Service): Enables users to one-click deploy models into online inference services or AI-Web applications, suitable for scenarios such as real-time inference, asynchronous inference and offline inference.
- PAI-Blade (General Inference Accelerator): All optimization technologies of Blade are designed for generality, applicable to different business scenarios. Through joint optimization of models and systems, it helps models achieve optimal inference performance.

AI Asset Management: PAI supports users in managing the full lifecycle of important AI production materials and development outputs such as models, datasets, and images. It also provides capabilities including AI asset sharing, horizontal comparison of training effects, and abnormal problem tracing, so as to reduce costs and improve efficiency during AI development and application processes.

搜集汇总

背景与挑战

背景概述

人工智能平台PAI是一个提供全流程AI工程化能力的机器学习平台，涵盖数据标注、模型开发、训练优化及推理部署等核心功能模块，内置140+算法并支持多行业场景。其灵骏智算服务通过软硬件协同优化，满足大模型训练、自动驾驶等领域的高性能计算需求。