Train a custom model with Amazon SageMaker

1. configure dataset (Amazon S3 URL, computing instance, or path for storing the training code in Amazon Elastic Container Registry (ECR)) & evaluation metrics (Amazon CloudWatch Logs)
2. configure hyperparameters
3. provide training script
4. fit the model

Features

• SageMaker Data Wrangler

  • connect pandas DataFrames and AWS data services
  • load and upload data

• SageMaker Clarify

• SageMaker Processing Jobs

• SageMaker Feature Store

  • store and serve features
  • reduce skew
  • real time & batch

• SageMaker Autopilot

  • cover all steps in AutoML: ingest & analyze data -> prepare & transform data -> train & tune model
  • automatically generate notebooks (candidate generation notebook (provides links to the data preprocessor Python scripts, the algorithms, and the algorithm hyperparameters selected by Autopilot), data exploration notebook) and codes

• SageMaker Debugger

  • capture real-time debugging data during model training in Amazon SageMaker
    • system metrics: CPU&GPU utilization, data I/O metrics
    • framework metrics:
      • convolutional operations in forward pass
      • batch normalization operations in backward pass
      • data loader processes between steps
      • gradient descent algorithm operations
    • output tensors: scalar values (accuracy and loss); matrices (weights, gradients, input & output layers)
  • if bugs, actions can be taken via
    • Amazon CloudWatch Event: send SMS etc
    • Amazon SageMaker Notebook: analyze
    • Amazon SageMaker Studio: visualize

• SageMaker Hyperparameter Tuning

  • steps: create an Estimator -> define the combination of hyperparameters -> start a SageMaker hyperparameter tuning job -> analyze results and select the best model candidate
  • a "warm start" feature: reuses results from a previous hyperparameter training job
    • scenarios:
      • you want to change the hyperparameter tuning ranges from the previous job -> use type "identical data and algorithm"
      • you want to add new hyperparameters -> use type "transfer learning"

• ML Model Deployment with SageMaker

  • SageMaker Endpoint
    • real-time prediction ML Model Deployment#^b2b283
    • auto-scaling
      • pipeline of creating SageMaker production variants with Endpoint:
        • construct Docker Image URI -> create SageMaker models -> create production variants -> create endpoint configuration -> create endpoint
  • SageMaker Batch Transform
    • batch prediction ML Model Deployment#^50c5c9

• SageMaker Pipelines

  • SageMaker model registry: for Machine Learning Systems Design#^8e6855

• SageMaker Model Monitor

• SageMaker Ground Truth

  • for Data preprocessing#Data labelling
  • job workflow: setup input data -> select labelling task -> select human workforce -> create task UI
  • human workforce options: Amazon Mechanical Turk/Private/vendor

• SageMaker Projects

  • integrates automatically with SageMaker Pipelines and SageMaker model registry
  • create MLOps solutions to orchestrate and manage your end to end machine learning pipelines, while also incorporating CI/CD practices

• SageMaker Managed Spot Training: for Machine Learning Systems Design#Checkpointing

• Amazon SageMaker BlazingText: unlike Natural Language Processing#^ff7c47, it is a word2vec model