Extracting Hyperparameter Constraints from Code

Abstract

Machine-learning operators often have correctness constraints that cut across multiple hyperparameters and/or data. Violating these constraints causes runtime exceptions, but they are usually documented only informally or not at all. This paper presents a verification-condition analysis for Python code. We demonstrate our analysis by extracting hyperparameter constraints for 45 sklearn operators. Our analysis is a step towards safer and more robust machine learning.