Embeddings of Code Changes
The goal of this project is to build explicit vector representations of code changes. In the course of this project, it is planned to obtain vector representations that can effectively encode information about a code change, and thus allow us to set semantic transformations over it. The approach treats program code as a sequence of tokens. The model can be trained in an unsupervised manner which allows us to perform a big pre-train of the network. The approach is evaluated on such tasks as commit message generation, stable patch prediction, and application of changes to the program code.