Educational Implementation of Edit Flows

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.

Setup

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.

Samples

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.

Empty Prior $\rightarrow$ Sine Wave Target

Uniform Prior $\rightarrow$ Sine Wave Target

Low Frequency Prior $\rightarrow$ High Frequency Target

References

📄 [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