krc-stat

This repository contains code used in the experiments for the paper

Irons NJ, Scetbon M, Pal S, Harchaoui Z. "Triangular Flows for Generative Modeling: Statistical Consistency, Smoothness Classes, and Fast Rates." AISTATS 2022. [arxiv]

Our code draws heavily from the UMNN repository, and uses datasets and models from the associated paper

Wehenkel A, Louppe G. "Unconstrained Monotonic Neural Networks." NeurIPS 2019. [arxiv]

The UMNN repository in turn draws from FFJORD and Sylvester normalizing flows for variational inference.

The code has been tested with Pytorch 1.1 and Python 3.8.

Files

The figures folder contains figures from the paper's numerical experiments that can be generated using the plots.ipynb notebook. This notebook draws from the folder results, which contains the results for each experiment.

The notebook run_experiments.ipynb contains example code to load in a dataset, fit the UMNN model, and plot the results. This can be run in a few minutes on a standard cpu. For the paper, we used cluster computing to generate many replicates for each dataset over a range of sample sizes. As mentioned above, the output results are in the results folder. Code to implement the experiments on a slurm cluster can be provided upon request.

The folders models and lib contain libraries to generate the datasets and fit the UMNN model. These files are mostly unchanged from the UMNN repository, except for lib/toy_data.py, which contains code to generate additional datasets not considered in the UMNN paper.

Datasets

Our experiments include datasets considered in the following papers:

Wehenkel A, Louppe G. "Unconstrained Monotonic Neural Networks." NeurIPS 2019. [arxiv]

Grathwohl W, Chen R, Bettencourt J, Sutskever I, Duvenaud D. "FFJORD: Free-form continuous dynamics for scalable reversible generative models." ICLR 2019. [arxiv]

Chaudhuri K, Kong Z. "The expressive power of a class of normalizing flow models." AISTATS 2020. [arxiv]

Cite

If you make use of this code in your own work, please cite our paper:

@misc{irons2021triangular,
      title={Triangular Flows for Generative Modeling: Statistical Consistency, Smoothness Classes, and Fast Rates}, 
      author={Nicholas J. Irons and Meyer Scetbon and Soumik Pal and Zaid Harchaoui},
      year={2021},
      eprint={2112.15595},
      archivePrefix={arXiv},
      primaryClass={stat.ML}
}