MLOps: Using AWS Built-in TensorFlow Framework on SageMaker

AWS CodePipeline is a CI/CD orchestration tool for infrastructure on AWS. This example desmonstrates how to use CodePipeline to orchesrate a CI/CD pipeline for a TensorFlow model runing on Amazon SageMaker platform. This example uses AWS built-in TensorFlow framework to train and host model.

NASA turbo fan jet engine predictive maintainence data set is used in the current example. The CI/CD pipeline will build and host a time series classification model using TensorFlow framework. Data set download link

Built-in framework is one of methods using SageMaker services: Source: AWS re:Invent 2020: Implementing MLOps practices with Amazon SageMaker

Built-in framework uses AWS pre-built TensorFlow images and your own training code to train and deploy models.

The following diagram depicts the architecure of the CI/CD pipeline.

The pipeline includes the following six stages:

1. Source: The source code stores in a CodeCommit repo. When there is code pushed, it will trigger the pipeline to update the model training and hosting.
2. Build: Build source code and save as 'source.tar.gz' file into S3 bucket. In this step, it also downloads raw NASA data set and saves into S3 bucket.
3. Train: Raw data will be prococessed and reshaped for TensorFlow model input. SageMaker training job is used to train the TensorFlow model.
4. Deploy Dev: Model is deployed and tested in Dev environment.
5. Deploy Test: Model is deployed and tested in Test enrinoment.
6. Deploy Prod: Model is deployed and tested in Prod enrinoment.

This is a basic pipeline that can be used to train and host your model. The following functionalities can be considered when building a customerized pipeline for a specific use case:

• Cross-account deployments
• Notification mechanisem
• Additional control and approval gates : manual or automatic
• Source control integration for lambda functions used in pipeline activities
• Additional deployment options such as batch transform
• Creation of API via API Gateway to be used for hosting predictions

---

Prerequisite

1. AWS Account & Administrator Access
2. AWS CLI installed. (https://docs.aws.amazon.com/cli/latest/userguide/install-cliv2.html)
3. Your AWS credentials. You can create a new Access Key on this page.
4. Docker installed.

---

Instruction

Deploy the pipeline

Run the following bash command at Terminal:

./deploy-cfn.sh

---

Clean-Up

Please execute the following clean-up steps. Note: These steps can be automated and/or done programmatically but doing manual clean to enforce what was created during the workshop.

1. Login to the AWS Console and enter your credentials

2. Select Services from the top menu, and choose Amazon SageMaker

   Endpoints

   • Go to Inference / Endpoints on the left hand menu, click the radio button next to each endpoint.

   • Select Actions and then Delete, then confirm delete hitting the Delete button.

3. Select Services from the top menu, and choose S3

   • Delete all S3 buckets created for this project

4. Select Services from the top menu, and choose CloudFormation

   • For the stack that were created in this workshop (MLOps-*), click the checkbox next to the stack. Select Actions , then Delete Stack

---

Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved. SPDX-License-Identifier: MIT-0