Proj CDeepFuzz Paper Reading: An Empirical Study on Bugs inside TensorFlow

Abstract

本文：

  1. 分析DL Library中的错误，给出了2个研究问题的答案
  2. RQ1: TensorFlow 中的错误有哪些症状和原因？
  3. RQ2: TensorFlow 内部的错误在哪⾥？
  4. 总结了5个发现
  5. 与症状相⽐，根本原因更具决定性，因为多个根本原因主导了症状
  6. TensorFlow 的症状和原因更像是普通软件系统（例如Mozilla）⽽不是机器学习系统（例如Lucene）
  7. For symptom of TensorFlow bugs, build failures have correlation with inconsistencies, configurations and referenced type errors, and warning-style bugs have correlation with inconsistencies, processing, and type confusions. For the root causes of TensorFlow bugs, dimension mismatches lead to functional errors, and type confusions have correlation with functional errors, crashes, and warning-style errors.
  8. 报告的主要错误是深度学习算法及其接⼝，以及编译、部署和安装中的错误
  9. 崩溃和构建失败是常⻅的症状，⽽不⼀致是组件之间常⻅的根本原因。
  10. TensorFlow bug 存在于其接⼝（26.24%）、学习算法（11.79%）以及如何跨平台编译（8.02%）、部署（7.55%）和安装（4.72%）
    1. kernel: kernel contains many sematic bugs
    2. API contains root causes related to tensor computations such as dimension mismatches and type confusions.
    3. library and tool, build failures are popular, and most bugs are caused by inconsistencies.

1. Intro

2. Methodology

2.1 Dataset

2.2 Manual Analysis

3. Empirical Result

3.1 RQ1. Symptoms and Root Causes

The categories of symptoms

  1. Functional error (35.64%)
  2. Crash (26.73%)
  3. Hang (1.49%)
  4. Performance degradation (1.49%)
  5. Build failure (23.76%)
  6. Warning-style error (10.89%)

The categories of root causes

  1. Dimension mismatch (3.96%).
  2. Type confusion (12.38%).
  3. Processing (22.28%).
  4. Inconsistency (16.83%)
  5. Algorithm (2.97%)
  6. Corner case (15.35%).
  7. Logic error (9.90%).
  8. Configuration error (7.43%).
  9. Referenced types error (4.95%).
  10. Memory (2.97%).
  11. Concurrency (0.99%).

3.2 RQ2. Bug Locations

3.3 Threat to validity

4. The Significance of Our Findings

6. Conclusion and Future Work

posted @ 2023-08-29 16:02  雪溯  阅读(0)  评论(0)  编辑  收藏  举报