Spring Boot Template for Amazon SageMaker DeepAR Endpoint Integration

This template application was built using Java Spring Boot framework with the aim to help machine learning engineer quickly setup a working code baseline for DeepAR endpoint testing/integration.

Highlights

• Implemented as a Spring Boot Web Application(running on port 8099, feel free to customize), service is invoked at http://localhost:8099/predict/sagemaker/deepar?target=, see explanation below
• DeepAR request/response jsonization - Build your DeepAR request using provided json classes
• AWS4 signing implementation - All the lengthy and error-prone work is done for you!
• Response data visualization in browser using JTablasaw - a Java dataframe and visualization library, see JTablasaw on Github

Requirements

Basic understanding about

• Java version >= 8
• AWS and SageMakeer
• AWS SageMaker DeepAR for time series data modeling and prediction
• AWS4 signing process

How does it work?

1. When the web application receives a GET request at http://localhost:8099/predict/sagemaker/deepar?target=, the value of parameter target is extracted and used to locate the file containing the time series input data for the model, see following source files for details:
   • WebController.java
   • LabManager.java
2. The input file is loaded and transformed into a json object, which will be used as the DeepAR POST request payload, see com.tensorloab.ml.aws.sagemaker.deepar.json for details
3. AWS4 authentication headers are built based on your AWS user credential, current UTC request time, SageMaker endpoint specification and the request payload,
   • see Examples of the Complete Version 4 Signing Process (Python) for the detailed specification of AWS4 signing process and a python implementation
   • check this source folder for my java implementation - com.tensorlab.ml.aws.auth
4. POST request containing the authentication headers and the json playload is sent to the DeepAR endpoint
5. Response from the DeepAR endpoint is received, parsed and stored as csv files
6. Finally the stored csv response file is plotted as a time series graph, and the javascript of the graph is sent back as the final response, see PlotUtil.java for details

Application Configuration

• Define your AWS IAM access key and secret as environment variable AWS_SAGEMAKER_USER_ACCESS_KEY and AWS_SAGEMAKER_USER_ACCESS_KEY, respectively
• In application.yml (Note that you can also create your own application.yml in application directory to
  overwrite the default values)
  • specify your SageMaker service region and host, visit this link to find out the region and endpoint that suit your requirement: AWS Regions and Endpoints
  • specify the endpoint name of your deployed model, this is the name you used to deploy the your model endpoint

Input File

An sample input file prediction-input-sample.csv is provided in folder src/main/resources, the name of the input file
must use format prediction-input-.csv, where the value of target will be used as the value of the request
parameter target. For example, if your input file is prediction-input-temperature.csv, then when you make the prediction request, the request URL will be http://localhost:8099/predict/sagemaker/deepar?target=temperature, when the request is received by your Spring Boot application, the value of request parameter target in the request URL will be used to locate the input file

Input File Sample

datetime,category0,category1
2013-08-01,41,1
2013-08-02,34,1
2013-08-03,10,1
2013-08-04,24,1

Tips

• Current input data needs to have 2 categories, use jsonschema2pojo to regenerate DeepAR json request and response classes according to your actual requirement
• Directly test your deployed endpoint using SageMaker API with your AWS notebook to make sure the model inference works in the first place
• Use Postman to test your IAM user access key and secret
• Interested in code basedline that can help you quickly setup a local RNN training envrionment? - See tensorlab - A Versatile Python Tool for RNN Modeling Experiments Using Keras)