Proj CDeepFuzz Paper Reading: CRADLE: Cross-Backend Validation to Detect and Localize Bugs in Deep Learning Libraries

Abstract

本文：CRADLE
Task: finding and localizing bugs in DL Libraries
Method:

  1. differential testing between DL Libraries
  2. 使用异常传播跟踪和分析来定位DL库中的错误函数anomaly propagation tracking and analysis to localize faulty functions

实验：
对象：TensorFlow, CNTK, Theano
测试集：11 datasets(including ImageNet, MINST, KGS Go game)，30 pre-trained models
效果：

  1. 12 bugs, 104 unique inconsistencies
  2. 正确标出了与104个unique inconsistencies相关的函数
posted @ 2023-08-29 16:10  雪溯  阅读(14)  评论(0)  编辑  收藏  举报