Learning with Stochastic Orders

This repository implements the algorithms and experiments described in Learning with Stochastic orders.

0. Install

To get started, create and activate the conda environment below:

conda env create -f gmorder_env.yml
conda activate gmorder_env

Note, if installing this environment on macOS, remove the cudatoolkits dependency from gmorder_env.yml as CUDA no longer supports macOS. Additionally, for running generative modeling training on larger datasets, such as CIFAR-10, you will need to ensure that device is set to cpu, e.g., in run_wgan_train_images.sh and run_wgan_dominate_images.sh.

1. Instructions to run 1D portfolio optimization

To run the 1D portfolio optimization experiment open and execute the portfolio_optimization notebook.

2. Generative modeling with d_CT

To run the GAN training using the Choquet-Toland (CT) distance use the shell script below:

sh run_choquet_train_distributions.sh

Open this script and change data (Line 7) to one of circle_of_gaussians, swiss_roll, image_point_cloud.

3. Baseline generator domination with VDC

	FID
g0: WGAN-GP	69.67
g*: WGAN-GP + VDC	67.317 ± 0.776

To train a baseline WGAN-GP model run

sh run_wgan_train_images.sh

Once training is complete, to reproduce the WGAN-GP + VDC results from the paper, execute:

sh run_wgan_dominate_images.sh

If needed, change file paths in this script to point to where the WGAN-GP checkpoint file and hyperparameter args are saved.

Acknowledgements

For several of our generator, discriminator, and Choquet critics, we draw inspiration and leverage code from the following public GitHub repositories:

1. https://github.com/caogang/wgan-gp
2. https://github.com/ozanciga/gans-with-pytorch
3. https://github.com/CW-Huang/CP-Flow

Citation:

To cite our work, please use:

@inproceedings{
domingo-enrich2023learning,
title={Learning with Stochastic Orders},
author={Carles Domingo-Enrich and Yair Schiff and Youssef Mroueh},
booktitle={International Conference on Learning Representations},
year={2023},
url={https://openreview.net/forum?id=P3PJokAqGW}
}