The recent boom in artificial intelligence (AI) benefits from the open source software (OSS) community. An increasing number of OSS developers are contributing to AI projects by submitting Pull-requests. However, the quality of pull-requests submitted by external contributors varies and the AI project management team has to review the pull-requests and ask contributors to revise them, if necessary. Since the process of revisions has a direct impact on the review efficiency and acceptance of pull-requests, it is important to achieve a better undersanding of pull request revisions. In this paper, we conducted an empirical study based on a set of pull-requests as well as their revision histories collected from the TensorFlow project.We first manually analyzed a sample of commit messages and review comments of pull-request and constructed a taxonomy of revision types.Then, according to the defined taxonomy, we manually labeled a large set of pull-request revisions.Based on the dataset, we explored the frequency and order of each type of revisions.We also investigated the co-occurrence and ajd-occurrenceprobability between different types of revisions.The emprirical findings show that: There are 11 different types of revisions which can be classified into three categories. Evolvability revisions occur more frequently than revisions of other types, and functional revisions are more likely to occur in the early updates of pull requests than evolvability revisions and other revisions. Structure-related revisions have a high chance to co-occur or ajd-occur with other revisions. Configuration-related revisions or rebasing revisions are more likely to appear in succession.