notebook by Stephen Z. Lu - X, Website
This is an unofficial impmlementation of the paper "Edit Flows: Flow Matching with Edit Operations" by Havasi et al..
The notebook main.ipynb has an educational purpose and explores the modeling of discretized sine waves using a vanilla Transformer backbone. As much as possible, the notebook is self-contained, but I strongly encourage readers to read the paper for a deeper understanding of the concepts and methods used.
To run the notebook, you will need to install the required packages inside requirements.txt. I used Python 3.10, but other versions of Python 3.x should work as well. To create videos of the sampling process, you will also need ffmpeg installed on your system.
Here, I show some samples of sine waves generated by the model at inference time. Notice that different choices of prior distribution, target distribution, and sequence alignment lead to different results.
📄 [1] Discrete Flow Matching by Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky T. Q. Chen, Gabriel Synnaeve, Yossi Adi, Yaron Lipman - Article
📄 [2] Edit Flows: Flow Matching with Edit Operations by Marton Havasi, Brian Karrer, Itai Gat, Ricky T. Q. Chen - Article
📄 [3] Introduction to Flow Matching by Georges Le Bellier - Github