probly 0.9.0

Uncertainty Representation and Quantification for Machine Learning

probly: Uncertainty Representation and Quantification for Machine Learning

🛠️ Install

probly is intended to work with Python 3.13 and above. Installation can be done via pip and or uv:

pip install probly

uv add probly

⭐ Quickstart

probly makes it very easy to make models uncertainty-aware and perform several downstream tasks:

import probly
import torch.nn.functional as F

net = ...  # get neural network
model = probly.method.dropout(net)  # make neural network a Dropout model
train(model)  # train model as usual

data = ...  # get data
data_ood = ...  # get out of distribution data
sampler = probly.representation.Sampler(model, num_samples=20)
sample = sampler.predict(data)  # predict an uncertainty representation
sample_ood = sampler.predict(data_ood)

eu = probly.quantification.classification.mutual_information(sample)  # quantify model's epistemic uncertainty
eu_ood = probly.quantification.classification.mutual_information(sample_ood)

auroc = probly.evaluation.tasks.out_of_distribution_detection(eu, eu_ood)  # evaluate model's uncertainty

📜 License

This project is licensed under the MIT License.

---

Built with ❤️ by the probly team.

Changelog

This changelog is updated with every release of probly.

Development

• Added possiblity to create ensemble of torch models without resetting the weights of each model.
• Refactored Efficient Credal Prediction calibration into the library via flexdispatch. Added a optimized PyTorch bisection solver (compute_efficient_credal_prediction_bounds) that reduces calibration time from days to minutes, while preserving the legacy SciPy implementation as a fallback for NumPy arrays.

0.1.0 (2024-03-14)

Initial pre-release of probly without functionalities.