/nd00333_AZMLND_C2

Primary LanguageJupyter Notebook

Deplopying and exposing an AzureML auto trained model as a REST endpoint

In this project AzureML is used to train a model based on marketing data from a bank. A selection of the optimal model is made and deployed as a REST endpoint. Different methods of consuming the endpoint are show. Finally the creation of a complete end to end deployment is done using the Python SDK.

Architectural Diagram

Steps

Key Steps

Uploading a dataset and configuring an azureML run

In AzureML studio, a dataset containing marketing data is uploaded and a 'run' is configured where a dataset and a target column is specified.

registerd datasets

dataset details

configuring an azureML run

using Automated ML to determine the best model

From all the training methods used by AutoML, the Voting Ensemble has the highest accuracy.

using Automated ML to determine the best model

Deploying the best model as a REST endpoint

The Voting Ensamble model is deployed and exposes a REST endpoint, with supporting swagger definition

Deploying the model

Deploying the best model as a REST endpoint

Using the Swagger definition in SwaggerUI

Using Swagger UI we can see the endpoints and expected data format for those endpoint.

Using the Swagger definition in SwaggerUI

The SwaggerUI won't allow the actual call

Calling the endpoint with curl command

To see if the proposed crul command of the Swagger file works, Here we are directly making a post request with curl

Calling endpoint with curl command

Consuming endpoint with endpoint.py

Calling the endpoint from a python sript also works

Consuming endpoint with endpoint.py

Enable logging

To enable logging for the REST endpoint, a Python script is used that enables application insights, then we check the UI to see if logging is actually enabled

Enable logging

Check UI to see that logging is enabled

Creating a publishing a pipeline

Using the iPython notebook, a pipeline is generated that is visible as a graph in AzureML studio.

Pipeline overview in AzureML studio

Pipeline graph in AzureML studio

Using the SDK to define a pipeline

Creating a publishing a pipeline SDK

The resulting pipeline in AzureML studio

The pipelines generated by the SDK are visible in AzureML studio. Here we see the active pipelines and the published pipeline endpoint

ML studio showing the scheduled run using Published Pipeline REST Endpoint

ML studio showing the pipeline endpoint as Active

Creating and/or sharing documentation and Swagger definition

Documentation depends on company standards, but would preferably be accessible on a portal/intranet

Ways to improve the model

First autoML gives a class balancing detection alert. This should be solved by supplying an equal amount of records that nave the label 'No' as the label 'Yes'. Currently there are 7.9 times as much records with a 'No' label as there are with the label 'Yes'

Secondly it is an option to train longer and maybe even try if the neural network option gives better results.

Screen Recording

Recording of all the steps taken in this project. Here are the timestamped subjects

Screencast

Uploading dataset

1 Working deployed ML model endpoint 2 Deployed Pipeline 3 Available AutoML Model 4 Successful API requests to the endpoint with a JSON payload

See https://github.com/fuzzballb/nd00333_AZMLND_C2/blob/master/starter_files/aml-pipelines-with-automated-machine-learning-step.ipynb for an overview of the steps executed in the Notebook