Abstract

TensorFlow is a widely used machine learning platform, with millions of people using it to create and train models. It is available in a variety of programming languages, including Python, Java, C++, and JavaScript, among which Python is the most commonly used. Along with Tensor- Flow’s evolution, new Python APIs are introduced, while others may be deprecated. Although the characteristics of deprecated APIs in traditional software frameworks such as Android have been extensively researched in recent years, little attention has been paid to how deprecated APIs in TensorFlow evolve and what impact this has on deep learning. In this thesis, we conducted an em- pirical study on deprecated Python APIs in TensorFlow. Our study analyzed 20 TensorFlow releases spanning versions 1.0 to 2.3 to investigate API deprecation and its causes. In addition, we studied projects containing 12 popular deep learning models to identify deprecated API usage. Finally, in order to investigate the potential impact of deprecated APIs on deep learning models, we manually updated the deprecated APIs in these projects to compare model accuracy before and after updating. Our research seeks to provide developers with insight into how TensorFlow deprecated APIs evolve, as well as help them understand why APIs became deprecated and the implications of not updating their models by removing deprecated APIs.