This repository accompanies the pre-print A PAC-Bayesian Perspective on Structured Prediction with Implicit Loss Embeddings by Théophile Cantelobre (Mines ParisTech, Inria), Benjamin Guedj (Inria, UCL), María Pérez-Ortiz (UCL) and John Shawe-Taylor (UCL).
The pre-print is available here: https://arxiv.org/abs/2012.03780.
On top of standard machine learning requirements, this repository requires jax (for auto-diff) and scikit-ml (for data).
We use the scikit-learn API as much as possible... you should be able to jump right in. To get started, you can take a look at demo_relaxed.py and demo_mc.py.
Jupyter notebooks are for reproducing figures (Section-7.ipynb depends on sensitivity-*.py, which takes a while to run).
If you have any issues with the code, feel free to open an issue. To chat about the paper, ping me at theophilec#gmail.com.