Testing deep learning (DL) software is crucial and challenging. Recent approaches use differential testing to cross-check pairs of implementations of the same functionality across different libraries. Such approaches require two DL libraries implementing the same functionality, which is often unavailable. In addition, they rely on a high-level library, Keras, that implements missing functionality in all supported DL libraries, which is prohibitively expensive and thus no longer maintained. To address this issue, we propose EAGLE, a new technique that uses differential testing in a different dimension, by using equivalent graphs to test a single DL implementation (e.g., a single DL library). Equivalent graphs use different Application Programming Interfaces (APIs), data types, or optimizations to achieve the same functionality. The rationale is that two equivalent graphs executed on a single DL implementation should produce identical output given the same input. Specifically, we design 16 new DL equivalence rules and propose a technique, EAGLE, that (1) uses these equivalence rules to build concrete pairs of equivalent graphs and (2) cross-checks the output of these equivalent graphs to detect inconsistency bugs in a DL library. Our evaluation on two widely-used DL libraries, i.e., Tensor Flow and PyTorch, shows that EAGLE detects 25 bugs (18 in Tensor Flow and 7 in PyTorch), including 13 previously unknown bugs.

@inproceedings{WangLQP022,
  title = {EAGLE},
  author = {Jiannan Wang and Thibaud Lutellier and Shangshu Qian and Hung Viet Pham and Lin Tan},
  booktitle = {ICSE},
  year = {2022},
  doi = {10.1145/3510003.3510165}
}