This post is about retraining Azure Machine Learning models. The mechanisms of continious learning and the feedback of our workflow are very important. Typically this critical feedback is provided by data scientists or the human “caretakers” of the ML solution.
Looping the quality and the frequency of the model analytics feedback can determine the success or failure of this technology in the organization.
We can create dynamic retraining model implementations by providing new dataset inputs automatically. These new models can adapt and withstand the test of time and evolve rapidly as conditions change in real time. This also defines the evolution for a machine learning history.
The trained models can be published like web services by enabling the quick operation of their experiments. With the latest release of the Retraining APIs, Azure Machine Learning now enables programmatic retraining of these models with new data. This section provides a high-level overview of that process.
The Programmatic Retraining API features in Azure Machine Learning will update the web service using the newly-trained model. Each operation can be done programmatically using the Retraining APIs. In this iteration, we make use of these new APIs that were introduced into the Azure Machine Learning feature set as it went from Preview mode to Production mode in early 2015.
The income prediction binary classification experiment looks like this,
For more information for this experement read my previous post.
The next step is to create an experiment for this model. So you should click on the bottom bar like the figure,
The Create Scoring Experiment command modified our income prediction experiment by adding both a web service input endpoint and a web service output endpoint to the design.
To incorporate retraining capabilities into our trained model, we need to modify and publish the original income prediction training experiment as a web service.
This new retraining web service will need a Web Service Output module connected to the Train Model module to be able to produce new trained models.
After these changes been made to indroduce retraining capabilities:
And publish the models again,
After adding the three additional Web service modules to our original training experiment, and processing the modules via the Run command, the next step is to click Publish Web Service on the bottom app bar. This will publish the income prediction training experiment as a web service that produces trained models and model evaluation results.