ML Workshop - December 9th 2021

Your temporary AWS account

Login to your account using this guide

Ingest, Prepare and Feature Store

Use the following command to clone the workshop

git clone https://github.com/data-science-on-aws/oreilly_book.git

Breakout room for troubleshooting - https://chime.aws/4105144644

Introduction Allocated time: 10 minutes

• 01_introduction/01_Setup_Dependencies.ipynb.
• 01_introduction/02_Check_Environment.ipynb.
• 01_introduction/03_Create_S3_Bucket.ipynb.
• 01_introduction/04_Update_IAM_Roles_And_Policies.ipynb.

Resources:

• Dive deep into Amazon SageMaker Studio Notebooks architecture

Ingest Allocated time: 20 minutes

• 04_ingest/00_Overview.ipynb
• 04_ingest/01_Copy_TSV_To_S3.ipynb
• 04_ingest/02_Create_Athena_Database.ipynb
• 04_ingest/03_Register_S3_TSV_With_Athena.ipynb
• 04_ingest/04_Convert_S3_TSV_To_Parquet_With_Athena.ipynb

Explore Allocated time: 10 minutes

• 05_explore/00_Overview.ipynb
• 05_explore/01_Visualize_Reviews_Dataset.ipynb.

Prepare Allocated time: 30 minutes

• 06_prepare/00_Overview.ipynb
• 06_prepare/01_Prepare_Dataset_BERT_Scikit_AdHoc_FeatureStore.ipynb. (10m)
  • It takes about 6 minutes for data to be availabe in offline feature store
  • If you get any errors due to Athena limits, simply rerun the step. What we do here?

1. Convert Raw Text to BERT Features using Hugging Face and TensorFlow
2. Add Features to SageMaker Feature Store
3. Ingest Records into Feature Store
4. Build Training Dataset and review using Athena
5. Review the Feature Store
6. Release Resources

• 06_prepare/02_Prepare_Dataset_BERT_Scikit_ScriptMode_FeatureStore.ipynb. (20m)
  • The processing job takes 15 minutes to run
  • If you get any errors like ThrottlingException, simply rerun the step and it should work fine.

Resources:

• BERT
  • BERT uncased
  • DistilBERT
• Feature Store
• Lineage Tracking
• SageMaker Experiments

Train, Model Registry, and Deployment and AutoScaling

Train

• 07_train/00_Overview.ipynb
• 07_train/01_Train_Reviews_BERT_Transformers_TensorFlow_AdHoc.ipynb
• 07_train/02_Train_Reviews_BERT_Transformers_TensorFlow_ScriptMode.ipynb

Deploy

• 09_deploy/00_Overview.ipynb
• 09_deploy/03_Deploy_Reviews_BERT_TensorFlow_REST_Endpoint.ipynb
  • Takes 20 minutes to create the endpoint
• 09_deploy/04_Autoscale_Reviews_BERT_TensorFlow_REST_Endpoint.ipynb

SageMaker Pipelines, EKS/Kubeflow ML Pipeline

Pipeline

• 10_pipeline/00_Overview.ipynb
• 10_pipeline/01_Create_SageMaker_Pipeline_BERT_Reviews.ipynb
  • The pipeline execution takes 42 minutes.
• 10_pipeline/05_Register_Deploy_Model.ipynb
• 10_pipeline/kubeflow

  • 10_pipeline/kubeflow/00_00_Setup_EKS.ipynb

    • In step 3 - Create EKS cluster, replace the --version 1.17 with --version 1.18

  • 10_pipeline/kubeflow/00_01_Configure_Environment.ipynb

  • 10_pipeline/kubeflow/00_02_Create_S3_Bucket.ipynb

  • 10_pipeline/kubeflow/00_03_Setup_Kubeflow.ipynb

  • 10_pipeline/kubeflow/00_04_Setup_Kubeflow_EKS_Credentials.ipynb

  • 10_pipeline/kubeflow/00_05_Launch_Kubeflow_Jupyter_Notebook.ipynb

  • 10_pipeline/kubeflow/00_06_Explore_Kubernetes_Cluster.ipynb

  • 10_pipeline/kubeflow/00_07_SWITCH_TO_KUBEFLOW_NOTEBOOK.ipynb

  • 10_pipeline/kubeflow/01_Kubeflow_Intro.ipynb

  • 10_pipeline/kubeflow/02_Kubeflow_Pipeline_Simple.ipynb

  • 10_pipeline/kubeflow/03_Kubeflow_Pipeline_Reviews_BERT_SageMaker.ipynb

    • Update the last cell of 03_Kubeflow_Pipeline_Reviews_BERT_SageMaker.ipynb with following code.

      import json
      inputs = {"features": ["This is great!"]}
      response = sm_runtime.invoke_endpoint(
          EndpointName=endpoint_name,
          ContentType="application/jsonlines",
          Accept="application/jsonlines",
          Body=json.dumps(inputs).encode("utf-8")
      )
      print("response: {}".format(response))
      predicted_classes_str = response["Body"].read().decode()
      predicted_classes_json = json.loads(predicted_classes_str)
      predicted_classes = predicted_classes_json['predicted_label']
      print("\n predicted_classes: {}".format(predicted_classes))
      

  • 10_pipeline/kubeflow/04_Distributed_Training.ipynb

  • 10_pipeline/kubeflow/05_Hyper_Parameter_Tuning.ipynb