SkWDRO - Tractable Wasserstein Distributionally Robust Optimization

Welcome to SkWDRO’s doc

skwdro is a Python package that offers WDRO versions for a large range of estimators, either by extending scikit-learn estimators or by providing a wrapper for pytorch modules.

Here is a quick reading order that we advise:

        flowchart TD
   A(((Getting started))):::important
   A -->|"Dive in quickly"| B["User guide"]:::important
   A -->|"If you are a bit new to DRO"| C["What is WDRO?"]:::normal
   C --> D["Why SkWDRO?"]:::important
   D --> B

   B -->|"Implement your own problem"| F["Pytorch interface"]:::important
   F --> G["Pytorch examples"]:::normal

   B -->|"For pre-implemented examples"| H["Sklearn interfaces"]:::normal
   H --> I["Sklearn examples"]:::normal

   G --> J["API"]:::important
   I --> J

   G -->|"In depth explanations"| K>"Samplers"]
   K --> L>"Cost functionals"]

   %% Styles
   classDef important color:#EE8888,fill:#f9f2d0,stroke:#e6b800,stroke-width:2px,color:black,rx:10px,ry:10px;
   classDef normal color:#888888,fill:#f2f2f2,stroke:#999999,stroke-width:1.5px,color:black,stroke-dasharray:3 3,rx:20px,ry:20px;

   %% Hyperlinks
   click A "quick_start.html"
   click B "user_guide.html"
   click C "wdro.html"
   click D "why_skwdro.html"
   click F "pytorch.html"
   click G "examples/Custom/index.html"
   click H "sklearn.html"
   click I "examples/Builtin/index.html"
   click J "api_deepdive/submodules.html"
   click K "tutos/samplers.html"
   click L "tutos/costs.html"
    

Getting started

Install it now!

$ pip3 install skwdro

See the getting-started guide to see how to install the package and get to learn its basic usage. Then you can take a look at some of the theory that goes behind the duality result we use to make Sinkhorn-WDRO tractable.

SkWDRO main formula

If you are performing “Empirical Risk Minimization”, and wish to robustify it, we can do it through the following transformation:

\[\frac{1}{N}\sum_{i=1}^N L(\xi_i) ~ \longmapsto ~ \min_{\lambda\ge 0}\rho\lambda+\varepsilon\frac{1}{N}\sum_{i=1}^N \log\mathbb{E}_{\zeta\sim\nu_{\xi_i}}\left[\frac{L(\zeta)-\lambda c(\xi_i, \zeta)}{\varepsilon}\right]\]

This transformation introduces computational difficulties, so let us handle it for you!

User Guide

Quick hitchhiker’s guide to the interfaces available to guide you through the process of robustification in “simple” cases.

In depth guide to the PyTorch customization functions

Learn more about the way you can robustify your own model with SkWDRO and how to specify it to make it compatible with the library.

Next

Getting started

Install everything needed.

quick_start.html

User guide

Learn about the most basic usecases of the library.

user_guide.html

What is WDRO?

Gentle introduction to the world of Distributionally Robust Optimization, and motivations for its Wasserstein version.

why_skwdro.html

API

More details about the exposed API.

api_deepdive/submodules.html

Cite

skwdro is the result of a research project. It is licensed under BSD 3-Clause. You are free to use it and if you do so, please cite:

Cite as follows:

 @article{JMLR:v27:24-1840,
   author  = {Vincent Florian and Wa{\"i}ss Azizian and Franck Iutzeler and J{\'e}r{\^o}me Malick},
   title   = {skwdro: a library for Wasserstein distributionally robust machine learning},
   journal = {Journal of Machine Learning Research},
   year    = {2026},
   volume  = {27},
   number  = {8},
   pages   = {1--7},
   url     = {http://jmlr.org/papers/v27/24-1840.html}
 }