MLOps-Example

Blogging

System Design

CI/CD

CI

• yaml file

name: CI

on:
  push:
    branches: [ main ]
  pull_request:
    branches: [ main ]

jobs:
  build:
    runs-on: ubuntu-latest
    steps:          
      - name: Check out
        uses: actions/checkout@v2
      
      - name: Docker Login
        uses: docker/[email protected]
        with:
          username: ${{ secrets.REGISTRY_USERNAME }}
          password: ${{ secrets.REGISTRY_PASSWORD }}

      - name: Build Images
        run: |
          docker build kubeflow_pipeline/0_data -t byeongjokim/mnist-pre-data
          ..(생략)..
          docker push byeongjokim/mnist-deploy
      
      - name: Slack Notification
        if: always()
        uses: rtCamp/action-slack-notify@v2
        env:
          SLACK_ICON_EMOJI: ':bell:'
          SLACK_CHANNEL: mnist-project
          SLACK_MESSAGE: 'Build/Push Images :building_construction: - ${{job.status}}'
          SLACK_USERNAME: Github
          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK_URL }}

CD

• yaml file

name: CD

on:
  workflow_run:
    workflows: ["ci"]
    branches: [main]
    types: 
      - completed

jobs:
  deploy:
    runs-on: ubuntu-latest
    steps:          
      - name: Check out
        uses: actions/checkout@v2
          
      - uses: actions/setup-python@v2
        with:
          python-version: '3.6.12'
          architecture: x64
      
      - uses: BSFishy/pip-action@v1
        with:
          packages: |
            kfp==1.3.0
      - name: run pipeline to kubeflow
        run: python kubeflow_pipeline/pipeline.py

      - name: Slack Notification
        if: always()
        uses: rtCamp/action-slack-notify@v2
        env:
          SLACK_ICON_EMOJI: ':bell:'
          SLACK_CHANNEL: mnist-project
          SLACK_MESSAGE: 'Upload & Run pipeline :rocket: - ${{job.status}}'
          SLACK_USERNAME: Github
          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK_URL }}

CT

alert using slack when new data is coming

• using Slack Trigger (slack_sdk)
• private repository in Github..
• when to train?
  • every Tuesday
  • new data is coming

kubeflow pipelines

• 0_data
  • data 수집
  • 전처리 후 npy 저장(train/test/validation)
  • npy_interval 사용하여 데이터 나누어 저장
  • embedding 학습에 사용하는 데이터와 faiss 학습 시 사용되는 데이터 구분
• 1_validate_data
  • 전처리된 npy 검증
  • shape, type 등 전처리 결과 확인
• 2_train_model
  • embedding 모델 학습
  • ArcFace 사용
  • multip gpu 사용
  • 학습 완료된 모델 torch.jit.script 저장
• 3_embedding
  • faiss 사용될 데이터 embedding 전처리 후 npy 저장
  • torch.jit.script 저장된 모델 load 해서 사용
• 4_train_faiss
  • embedding npy로 faiss 학습
  • faiss index 저장
• 5_analysis_model
  • 전체 모델 성능 평가
  • mlpipeline-metrics, mlpipeline-ui-metadata로 시각화
  • class 별 accuracy 측정 (confusion matrix)
  • dsl.Condition 사용하여 배포할지 결정
• 6_deploy
  • 모델 버전 관리
  • config 관리
  • service, deployment 배포
  • torchserve 사용

Serving Model using TorchServe

uploaded pipelines

confusion matrix(mlpipeline-ui-metadata)

alrert results using slack