Proj CDeepFuzz Paper Reading: Eagle: Creating equivalent graphs to test deep learning libraries
Abstract
本文:Eagle
Task: use equivalent graphs to differential test DL libraries
Method: 使用不同的APIs, data types or optimizations来获取equivalent graphs,16 DL equivalence rules
实验:
对象: Tensorflow, PyTorch
效果:
- detects 25 bugs(18 Tensorflow, 7 PyTorch), + 13 bugs