Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/aida-ugent/fairret
A fairness library in PyTorch.
https://github.com/aida-ugent/fairret
ai artificial-intelligence bias deep-learning fairness fairness-ai fairness-ml machine-learning python pytorch
Last synced: 2 months ago
JSON representation
A fairness library in PyTorch.
- Host: GitHub
- URL: https://github.com/aida-ugent/fairret
- Owner: aida-ugent
- License: mit
- Created: 2024-01-31T17:55:41.000Z (11 months ago)
- Default Branch: main
- Last Pushed: 2024-07-23T17:00:24.000Z (5 months ago)
- Last Synced: 2024-09-30T15:42:54.793Z (2 months ago)
- Topics: ai, artificial-intelligence, bias, deep-learning, fairness, fairness-ai, fairness-ml, machine-learning, python, pytorch
- Language: Python
- Homepage: https://aida-ugent.github.io/fairret/
- Size: 447 KB
- Stars: 26
- Watchers: 3
- Forks: 1
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
- License: LICENSE
- Citation: CITATION.cff
Awesome Lists containing this project
- awesome-ccamel - aida-ugent/fairret - A fairness library in PyTorch. (Python)
README
# fairret - a fairness library in PyTorch
[](https://github.com/aida-ugent/fairret/blob/main/LICENSE)
[](https://pypi.org/project/fairret/)
[](https://openreview.net/pdf?id=NnyD0Rjx2B)
The goal of fairret is to serve as an open-source library for measuring and pursuing statistical fairness in PyTorch models.
The library is designed to be
1. *flexible* in how fairness is defined and pursued.
2. *easy* to integrate into existing PyTorch pipelines.
3. *clear* in what its tools can and cannot do.
Central to the library is the paradigm of the _fairness regularization term_ (fairrets) that quantify unfairness as differentiable PyTorch loss functions.
These can be minimized jointly with other losses, like the binary cross-entropy error, by just adding them together!
## Quickstart
It suffices to simply choose a _statistic_ that should be equalized across groups and a _fairret_ that quantifies the gap.
The model can then be trained as follows:
```python
import torch.nn.functional as F
from fairret.statistic import PositiveRate
from fairret.loss import NormLoss
statistic = PositiveRate()
norm_fairret = NormLoss(statistic)
def train(model, optimizer, train_loader):
for feat, sens, target in train_loader:
optimizer.zero_grad()
logit = model(feat)
bce_loss = F.binary_cross_entropy_with_logits(logit, target)
fairret_loss = norm_fairret(logit, sens)
loss = bce_loss + fairret_loss
loss.backward()
optimizer.step()
```
No special data structure is required for the sensitive features. If the training batch contains $N$ elements, then `sens` should be a tensor of floats with shape $(N, d_s)$, with $d_s$ the number of sensitive features. **Like any categorical feature, it is expected that categorical sensitive features are one-hot encoded.**
A notebook with a full example pipeline is provided here: [Simple Pipeline.ipynb](/examples/Simple Pipeline.ipynb).
We also host [documentation](https://aida-ugent.github.io/fairret/).
## Installation
The fairret library can be installed via PyPi:
```
pip install fairret
```
A minimal list of dependencies is provided in [pyproject.toml](https://github.com/aida-ugent/fairret/blob/main/pyproject.toml).
If the library is installed locally, the required packages can be installed via `pip install .`
## Warning: AI fairness != fairness
There are many ways in which technical approaches to AI fairness, such as this library, are simplistic and limited in actually achieving fairness in real-world decision processes.
More information on these limitations can be found [here](https://dl.acm.org/doi/full/10.1145/3624700) or [here](https://ojs.aaai.org/index.php/AAAI/article/view/26798).
## Future plans
The library maintains a core focus on only fairrets for now, yet we plan to add more fairness tools that align with the design principles in the future. These may involve breaking changes. At the same time, we'll keep reviewing the role of this library within the wider ecosystem of fairness toolkits.
Want to help? Please don't hesitate to open an issue, draft a pull request, or shoot an email to [[email protected]](mailto:[email protected]).
## Citation
This framework will be presented as a paper at ICLR 2024. If you found this library useful in your work, please consider citing it as follows:
```bibtex
@inproceedings{buyl2024fairret,
title={fairret: a Framework for Differentiable Fairness Regularization Terms},
author={Buyl, Maarten and Defrance, Marybeth and De Bie, Tijl},
booktitle={International Conference on Learning Representations},
year={2024}
}
```