In the post Introducing OML AutoML User Interface, I discussed how OML AutoML UI simplifies the machine learning modeling process through the automated experiment workflow, while adding a few delighter features such as generating OML Notebooks for selected models and supporting immediate deployment to OML Services. In this post, I’ll share 4 performance tips you could use when running OML AutoML UI experiments.
When working with Oracle Machine Learning on Oracle Autonomous Database, it’s important to keep in mind the autonomous database tier you’re using and your database compute settings. If you’re using the Always Free Tier, this is a 1 OCPU autonomous database instance. So, there is no parallelism and such an instance is good for exploring functionality on relatively small data sets. On the Paid Tier, you can configure your autonomous database instance per your needs and opt for auto scaling to get up to 3 times your normal autonomous database OCPU setting – elastically allocated.
We’ll start with a top level tip, then look at the three that involve additional settings.
Tip #1: When running an experiment, you have the choice of “faster results” or “better accuracy”. If you’re interested in getting initial results faster, there’s an obvious choice. By choosing “faster results”, the hyperparameter search space is reduced. “Better accuracy” works with the broader set of hyperparameter options recommended by the internal meta-learning model.
When you create an experiment, there are additional settings you can set. Let’s talk briefly about the default settings.
The defaults are currently set to select the top 5 models from among all the available algorithms. The database service level is set to LOW – which results in no parallelism – and there is a high runtime limit set. In addition, all the candidate algorithms are selected for consideration as the experiment runs.
Tip #2: You can reduce the number of top models to select to 2 or 3 since tuning models to get the top one for each algorithm requires additional time. If you’re interested in getting initial results even faster, you may consider going with the top recommended algorithm – setting the number to 1 – this will tune a model for that algorithm.
Tip #3: You can eliminate algorithms from being considered if, for example, you have preferences for particular algorithms, or have certain requirements. For example, if model transparency is essential, excluding models such as Neural Network would make sense. Keep in mind that some algorithms are more compute intensive than others, for example Naïve Bayes and Decision Tree are normally faster than Support Vector Machine or Neural Network.
Tip #4: You can change your database service level to MEDIUM or HIGH. The HIGH gives the greatest parallelism but significantly limits the number of concurrent jobs, while the MEDIUM level enables some parallelism but allows greater concurrency for job processing. Note that changing this setting on the Always Free Tier will have no effect since there is a 1 OCPU limit. However, when you increase the OCPUs allocated to your autonomous database, you can then increase the database service level to MEDIUM or HIGH.
So there you have it.
Access Oracle Machine Learning AutoML UI in Oracle Autonomous Database alongside Oracle Machine Learning Notebooks, where you create and run experiments, deploy models, and generate notebooks.