5. 機械学習システム構築に立ちはだかる壁
● 継続的再学習 (Continuous Learning)
● システムに必ずヒトが介在する (Human-in-the-loop)
● データは不変の存在ではない (Data is mutable)
引用 : Hidden Technical Debt in Machine Learning Systems (NIPS2015)
6. Kubeflow
● kubernetes 上で動く機械学習ツールキット
○ Goal: End to End の機械学習システムを提供
● Current Ver. : 0.2 (2018.12.16 に1.0リリース予定)
○ Simple : 数々の機能を Kubeflow上で提供
○ Portable : k8sが動く場所ならどこでも動く
○ Scalable : k8sの機能を使ってスケーラビリティも担保
7. Kubeflow overview
Data
Split
Roll-out Serving Monitoring Logging
Trainer
Building
a
Model
Model
Validation
Training
at Scale
Data
ingestion
Data
Analysis
Data
Transfor-
mation
Data
Validation
Ref: Kubeflow Deep Dive – David Aronchick & Jeremy Lewi, Google (Intermediate Skill Level)
TF Serving
Katib TFMA
TFDVT
8. Argoの守備範囲
Data
Split
Roll-out Serving Monitoring Logging
Trainer
Building
a
Model
Model
Validation
Training
at Scale
Data
ingestion
Data
Analysis
Data
Transfor-
mation
Data
Validation
Ref: Kubeflow Deep Dive – David Aronchick & Jeremy Lewi, Google (Intermediate Skill Level)
15. Argo: Basic Usage
● Features
○ DAG or Step based declaration of workflow
○ Artifact support
○ Step level input & outputs, ….more and more!
● argo: workflow
○ argo-ci
○ argo-cd
○ argo-events