OGAWA Microsoft Japan Co., Ltd. Cloud Solution Architect (Data & AI) 書道家/エンジニア 興味分野 センシング、HCI、AIの社会実装、MLOps、エッジAI推論 Linux Foundation / Green Software Foundation所属 どんな人か
etc. モデル AKS バッチ予測: SQL DB Analytics Power BI Business Apps Prepare Data Build & Train Manage & Monitor Deploy Synapse 様々なAzureサービスのソースからデータを取り込み、モデル構築・分析・可視化が可能。 Azure ML
Pipeline Azure Data Lake Storage Gen2 Job Environments Azure Machine Learning Python SDK/CLI Azure Container Registry Data Prep Data Models Monitoring Azure Machine Learning Python SDK/CLI Managed (Online/Batch) Endpoint Code Checkout Managed (Online/Batch) Endpoint GitHub Model Training Code Test, Data Check Model Training Model Evaluation Responsible AI Deploy to Stage Deploy to Stage Model Test Responsible AI Azure Monitor Azure Insights Azure Machine Learning Components container Gated approval Retraining Trigger Deploy to Prod Deploy to Prod Model Test ResponsibleAI Safe Rollout Azure Machine Learning Python SDK/CLI Compute Instance Compute Clusters build&push Training (pipeline) Deployment (pipeline) Data Science team Infra team Alert Dashboard build&push job submit experiment log/metric auto scale log/metric auto scale log/metric Register model deploy automatically deploy automatically Event Grid (Model Trigger) (Code Trigger) Azure Managed Grafana Event Grid model register data drift
Pipeline Azure Data Lake Storage Gen2 Job Environments Azure Machine Learning Python SDK/CLI Azure Container Registry Data Prep Data Models Monitoring Azure Machine Learning Python SDK/CLI Managed (Online/Batch) Endpoint Code Checkout Managed (Online/Batch) Endpoint GitHub Model Training Code Test, Data Check Model Training Model Evaluation Responsible AI Deploy to Stage Deploy to Stage Model Test Responsible AI Azure Monitor Azure Insights Azure Machine Learning Components container Gated approval Retraining Trigger Deploy to Prod Deploy to Prod Model Test ResponsibleAI Safe Rollout Azure Machine Learning Python SDK/CLI Compute Instance Compute Clusters build&push Training (pipeline) Deployment (pipeline) Data Science team Infra team Alert Dashboard build&push job submit experiment log/metric auto scale log/metric auto scale log/metric Register model deploy automatically deploy automatically Event Grid (Model Trigger) (Code Trigger) Azure Managed Grafana Event Grid model register data drift • データセット・ジョブ、モデル等の管理 • 作成者/実験者・時刻 etc.. • 再利用するためのサンプルコードも提供 • モデルのトレーニングコードの生成 in MSBuild 2022 再現性
Pipeline Azure Data Lake Storage Gen2 Job Environments Azure Machine Learning Python SDK/CLI Azure Container Registry Data Prep Data Models Monitoring Azure Machine Learning Python SDK/CLI Managed (Online/Batch) Endpoint Code Checkout Managed (Online/Batch) Endpoint GitHub Model Training Code Test, Data Check Model Training Model Evaluation Responsible AI Deploy to Stage Deploy to Stage Model Test Responsible AI Azure Monitor Azure Insights Azure Machine Learning Components container Gated approval Retraining Trigger Deploy to Prod Deploy to Prod Model Test ResponsibleAI Safe Rollout Azure Machine Learning Python SDK/CLI Compute Instance Compute Clusters build&push Training (pipeline) Deployment (pipeline) Data Science team Infra team Alert Dashboard build&push job submit experiment log/metric auto scale log/metric auto scale log/metric Register model deploy automatically deploy automatically Event Grid (Model Trigger) (Code Trigger) Azure Managed Grafana Event Grid model register data drift • マネージドエンドポイント(モデル展開の簡素化) • カスタムコンテナは推論サーバーをDocker化さえできれば、 AMLはホストできる • Azure Machine Learning 上で監視、スケーリング、アラート、および 認証を引き続き利用できる • ブルーグリーンデプロイメント対応 デプロイ
Pipeline Azure Data Lake Storage Gen2 Job Environments Azure Machine Learning Python SDK/CLI Azure Container Registry Data Prep Data Models Monitoring Azure Machine Learning Python SDK/CLI Managed (Online/Batch) Endpoint Code Checkout Managed (Online/Batch) Endpoint GitHub Model Training Code Test, Data Check Model Training Model Evaluation Responsible AI Deploy to Stage Deploy to Stage Model Test Responsible AI Azure Monitor Azure Insights Azure Machine Learning Components container Gated approval Retraining Trigger Deploy to Prod Deploy to Prod Model Test ResponsibleAI Safe Rollout Azure Machine Learning Python SDK/CLI Compute Instance Compute Clusters build&push Training (pipeline) Deployment (pipeline) Data Science team Infra team Alert Dashboard build&push job submit experiment log/metric auto scale log/metric auto scale log/metric Register model deploy automatically deploy automatically Event Grid (Model Trigger) (Code Trigger) Azure Managed Grafana Event Grid model register data drift • コード管理 in Github/Azure DevOps • テストの実行 • パイプラインのキック • トレーニングジョブの実行、レジストリにモデルの登録、 モデルのデプロイ 自動化
Pipeline Azure Data Lake Storage Gen2 Job Environments Azure Machine Learning Python SDK/CLI Azure Container Registry Data Prep Data Models Monitoring Azure Machine Learning Python SDK/CLI Managed (Online/Batch) Endpoint Code Checkout Managed (Online/Batch) Endpoint GitHub Model Training Code Test, Data Check Model Training Model Evaluation Responsible AI Deploy to Stage Deploy to Stage Model Test Responsible AI Azure Monitor Azure Insights Azure Machine Learning Components container Gated approval Retraining Trigger Deploy to Prod Deploy to Prod Model Test ResponsibleAI Safe Rollout Azure Machine Learning Python SDK/CLI Compute Instance Compute Clusters build&push Training (pipeline) Deployment (pipeline) Data Science team Infra team Alert Dashboard build&push job submit experiment log/metric auto scale log/metric auto scale log/metric Register model deploy automatically deploy automatically Event Grid (Model Trigger) (Code Trigger) Azure Managed Grafana Event Grid model register data drift • データドリフト検知 • Azure Monitorにより監視 • Azure Event Gridによるイベントの検知 • job完了、モデルの登録、デプロイ、 実行ステータスの変更、etc… 再学習