AIADA: Accuracy Impact Assessment of Deprecated Python API Usages on Deep Learning Models

Abstract

TensorFlow is an end-to-end open-source machine learning platform including various tools, libraries, and community resources. It supports users to use many mainstream programming languages including Python. TensorFlow contains multiple abstraction layers, with APIs play significant roles in every layers. In the version iteration of TensorFlow platform development, with the release of new TensorFlow versions, because of functionality evolution, or security and performance-related changes, some APIs eventually become unnecessary. These issues cause APIs to deprecate and influence the accuracy of deep learning models results. Prior studies have investigated API evolution and its potential impact on projects. However, their studies mainly focus on API evolution instead of API deprecation, and they do not find out how the evolution affects results of deep learning models in TensorFlow. Therefore, we present a research-based prototype tool called AIADA and apply it to different revisions of the TensorFlow platform projects code for characterizing deprecated APIs. Based on the data mined by AIADA, we develop a quantitative assessment of deprecated Python APIs usages on deep learning models accuracy. We first count the amount of TensorFlow Python APIs that are deprecated, finding out that with the development of TensorFlow version, the number of deprecated APIs increases constantly. Second, we discuss the reason behind TensorFlow Python APIs become deprecated, discover that name change, weed out, and compatibility issue lead to the main cause of deprecation. Finally, we construct a deep learning project as the comparative experiment. After comparing the results between deep learning model with TensorFlow deprecated APIs and without deprecated APIs, we conclude that using deprecated APIs will cause a 10% loss on efficiency and accuracy of deep learning model.