Accelerated Data Science (ADS) Feature Type and Model Catalog Features

ADS v2.3.1

Model Catalog

This new release of the model catalog is now available. It includes these enhancements:

• Automatical extraction of model taxonomy metadata that lets data scientists document the use case, framework, and hyperparameters of their models.

• Improvement to the model provenance metadata, including a reference to the model training resource (notebook sessions) by passing training_id into save().

• Support for custom metadata, which lets data scientists document the context around their models, automatic extraction references to the conda environment used to train the model, the training and validation datasets, and so on.

• Automatcal extraction of the model input feature vector and prediction schemas.

• Model introspection tests that are run on the model artifact before the model is saved to the model catalog. Model introspection validates the artifact against a series of common issues and errors found with artifacts. These introspection tests are part of the model artifact code template that is included.

Feature Type

Feature type is a newly added module which includes the following functionalities:

• Support for Explorationary Data Analysis including feature count, feature plot, feature statistics, correlation, and correlation plot.

• Support for the feature type manager that provides the tools to manage the handlers used to drive the feature type system.

• Support for the feature type validators that are a way of performing data validation and also allow a feature type to be dynamically extended so that the data validation process can be reproducible and shared across projects.

• Support for feature type warnings that allow you to automate the process of checking for data quality issues.