Kubeflow

基于 Kubernetes 的机器学习开源平台

Kubeflow 能够将机器学习模型的开发、学习和部署流程的整个管道中的重复工作自动化。提供基于开源机器学习平台的 MLOps1) 环境，同时提供学习数据和模型以及运营数据的集成管理。
1) MLOps：旨在统一机器学习开发 (Dev) 和机器学习系统运营 (Ops) 的一门机器学习工程学科

概览

  • 云原生 MLOps 环境

    Kubeflow 提供针对云优化的机器学习模型开发环境，支持基于 Kubernetes 的与各种开源软件链接。

  • Usage Convenience for Big Data

    The standardized environments support a range of machine learning frameworks from TensorFlow, PyTorch, scikit-learn, and Keras. The pipeline for the entire development, learning and deployment processes of machine learning models are automated to ensure simple configuration/creation as well as reuse of the models.

  • Add-on Features

    Expanded open source Kubeflow from distributed learning job execution/monitoring to inference service management/analysis and job queue management are provided on Samsung Cloud Platform. Users can also enjoy job schedulers (FIFO, Bin-packing, and Gang-based), GPU fraction, GPU resource monitoring, Kubeflow engine logging and more add-on features that are usually unavailable on open source software.

01

04

服务架构

  1. User
  2. Console
  3. Kubernetes Engine : CPU Worker Nodes, GPU Worker Nodes, Persistent Volume
  1. Data Scientist, MLOps Engineer
  2. Kubeflow : Pipeline, Meta data, Model Serving, Notebook, Hyper Parameter Tuning ...
  3. Jupyter Notebook → Model Development → Model Training → Hyper Para.Tuning → Model Serving → Inference Application.

Key Features

  • Kubeflow 服务创建

    - 请求：对请求的 Kubernetes 群集进行自动部署和服务配置
    - 查看：产品列表、Kubeflow 版本/资源状态、运行/停止状态信息
    - 删除：删除已创建的 Kubeflow 模块

  • Kubeflow 服务功能

    - Basic features
      · Jupyter Notebook (model development, learning, and inference)
      · Workflow automation (based on machine learning pipelines)
    - Additional features of Samsung Cloud Platform
      · Manage GPU job scheduling and job queue
      · Conduct GPU resource monitoring and GPU fraction
      · Provide Kubeflow engine monitoring/logging and distributed learning job execution/monitoring
      · Build and manage ML framework images (TensorFlow, PyTorch, etc.) and ML images
      · Manage/analyze inference services and manage model experiments/learning nodes
      · Work with user authentication of Samsung Cloud Platform and manage users/projects/announcements

定价

    • 产品
    • SCP Kubeflow 软件包
    • Billing
    • Charged by the hour for the scale and usage of deployed Kubeflow
      ※ Samsung Cloud Platform for user environment configuration charged additionally
如有问题，请随时垂询

无论是咨询解决方案，还是其它问题，我们都会尽心为您解答

Share KakaoTalk Facebook Twitter Linkedin Close Share Box