Code and experiments from my Master’s thesis on Structured Prediction Energy Networks for Data Completion, published in 2021. It features a new framework that adapts SPENs to handle data completion with arbitrary masks.
The proposed framework for data completion performs better than an autoencoder baseline in a font completion task. Also, unlike with previous SPEN experiments, it does not require a separate initialization network to achieve good results and requires only a single step of gradient-based inference.
This code provides an implementation of that framework for completing missing letters in fonts. A summary paper of the thesis can be found here.