Introducing fairret - a fairness library in PyTorch!
The goal of fairret is to serve as an open-source Python library for measuring and mitigating statistical fairness in PyTorch models. The library is designed to be
- flexible in how fairness is defined and pursued.
- easy to integrate into existing PyTorch pipelines.
- clear in what its tools can and cannot do.
The central to the library is the paradigm of the fairness regularization term (fairrets) that quantify unfairness as differentiable PyTorch loss functions. These can then be optimized together with e.g. the binary cross-entropy error such that the classifier improves both its accuracy and fairness.
The library is still in very early development. Documentation, installation instructions, and more examples will be added in the near future.
The library is available on our GitHub