Automatic Unit Test Generation for Deep Learning Frameworks based on API Knowledge

TL;DR

This paper introduces MUTester, a tool that enhances automatic unit test generation for deep learning frameworks by leveraging API constraints and usage patterns, significantly improving coverage and reducing invalid tests.

Contribution

MUTester uniquely combines mined API constraints and usage patterns to guide test generation for deep learning APIs, filling a gap in existing tools.

Findings

Code coverage increased by 15.7% to 27.0%.

Reduced invalid tests by approximately 19%.

Validated effectiveness through developer user study.

Abstract

Many automatic unit test generation tools that can generate unit test cases with high coverage over a program have been proposed. However, most of these tools are ineffective on deep learning (DL) frameworks due to the fact that many of deep learning APIs expect inputs that follow specific API knowledge. To fill this gap, we propose MUTester to generate unit test cases for APIs of deep learning frameworks by leveraging the API constraints mined from the corresponding API documentation and the API usage patterns mined from code fragments in Stack Overflow (SO). Particularly, we first propose a set of 18 rules for mining API constraints from the API documents. We then use the frequent itemset mining technique to mine the API usage patterns from a large corpus of machine learning API related code fragments collected from SO. Finally, we use the above two types of API knowledge to guide the test generation of existing test generators for deep learning frameworks. To evaluate the performance of MUTester, we first collect 1,971 APIs from four widely-used deep learning frameworks (i.e., Scikit-learn, PyTorch, TensorFlow, and CNTK) and for each API, we further extract its API knowledge, i.e., API constraints and API usage. Given an API, MUTester combines its API knowledge with existing test generators (e.g., search-based test generator PyEvosuite and random test generator PyRandoop) to generate test cases to test the API. Results of our experiment show that MUTester can significantly improve the corresponding test generation methods and the improvement in code coverage is 15.7% to 27.0% on average. In addition, it can help reduce around 19.0% of invalid tests generated by the existing test generators. Our user study with 16 developers further demonstrates the practicality of MUTester in generating test cases for deep learning frameworks.