The Oracle Accelerated Data Science (ADS) SDK is a Python library that is included as part of the Oracle Cloud Infrastructure Data Science service. ADS offers a friendly user interface with objects and methods that describe the steps involved in the lifecycle of machine learning models, from data acquisition to model evaluation and interpretation.

You access ADS when you launch a JupyterLab session from the Data Science service. ADS is pre-configured to access Data Science and other Oracle Cloud Infrastructure resources, such as the models in the Data Science model catalog or files in Oracle Cloud Infrastructure Object Storage.

Main Features

  • Connect to Different Data Sources

    The Oracle JupyterLab environment is pre-installed with default storage options for reading from and writing to Oracle Cloud Infrastructure Object Storage. However, you can load your datasets into ADS from almost anywhere including:

    • Oracle Cloud Infrastructure Object Storage

    • Oracle Autonomous Data Warehouse

    • Oracle Database

    • Hadoop Distributed File System

    • Amazon S3

    • Google Cloud Service

    • Microsoft Azure

    • Blob

    • MongoDB

    • NoSQL DB instances

    • Elastic Search instances

    • Your local files

    These datasets can be numerous formats including:

    • csv

    • tsv

    • Parquet

    • libsvm

    • JSON

    • Excel

    • SQL

    • HDF5

    • XML

    • Apache server log files

    • arff


    Example of Opening a Dataset

  • Perform Exploratory Data Analysis

    The ADS data type discovery supports simple data types like categorical, continuous, ordinal to sophisticated data types. For example, geo data, date time, zip codes, and credit card numbers.


    Example showing exploring the class imbalance of a target variable

  • Automatic Data Visualization

    The ADSDataset object comes with a comprehensive plotting API. It allows you to explore data visually using automatic plotting or create your own custom plots.


    Example showing Gaussian Heatmap Visualization


    Example showing plotting lat/lon points on a map

  • Feature Engineering

    Leverage ADS and the DASK API to transform the content of a ADSDataset object with custom data transformations.


    Example showing using ADS to drop columns and apply auto transforms

  • Data Snapshotting for Training Reproducibility

    Save and load a copy of any dataset in binary optimized Parquet format. By snapshotting a dataset, a URL is returned that can be used by anyone with access to the resource to load the data exactly how it was at that point with all transforms materialized.

  • Model Training


    Example showing a visualized Decision Tree

    The Oracle AutoML engine, that produces ADSModel models, automates:

    • Feature Selection

    • Algorithm Selection

    • Feature Encoding

    • Hyperparameter Tuning

    Create your own models using any library. If they resemble sklearn estimators, you can promote them to ADSModel objects and use them in evaluations, explanations, and model catalog operations. If they do not support the sklearn behavior, you can wrap them in a Lambda then use them.


    Example showing how to invoke AutoML


    Example showing the AutoML hyper-parameter tuning trials

  • Model Evaluations

    Model evaluation generates a comprehensive suite of evaluation metrics and suitable visualizations to measure model performance against new data, and can rank models over time to ensure optimal behavior in production. Model evaluation goes beyond raw performance to take into account expected baseline behavior. It uses a cost API so that the different impacts of false positives and false negatives can be fully incorporated.

    ADS helps data scientists evaluate ADSModel instances through the ADSEvaluator object. This object provides a comprehensive API that covers regression, binary, and multinomial classification use cases.


    Example showing how to evaluate a list of models


    Example showing some model evaluation plots

  • Model Interpretation and Explainablility

    Model explanation makes it easier to understand why machine learning models return the results that they do by identifying relative importance of features and relationships between features and predictions. Data Science offers the first commercial implementation of model-agnostic explanation. For example, a compliance officer can be certain that a model is not making decisions in violation of GDPR or regulations against discrimination.

    For data scientists, it enables them to ensure that any model they build is generating results based on predictors that make sense. Understanding why a model behaves the way it does is critical to users and regulators. Data Science ensures that deployed models are more accurate, robust, and compliant with relevant regulations.

    Oracle provides Machine Learning Explainablility (MLX), which is a package that explains the internal mechanics of a machine learning system to better understand models. Models are in the ADSModel format. You use MLX to explain models from different training platforms. You create an ADSModel from a REST end point then use the ADS model explainability to explain a model that’s remote.

  • Interact with the Model Catalog

    You can upload the models that you create with ADS into the Data Science model catalog directly from ADS. You can save all your models, with their provenance information, in the catalog and make them accessible to anybody who needs to use them. Other users can then load the models and use them as an ADSModel object. You can also use this feature to help put the models into production with Oracle Functions.