Fiducio

Fiducio is a model-agnostic Python library for post-hoc calibration of 2D and 3D semantic segmentation models. It turns the logits or probabilities of any segmentation model into better-calibrated probabilities, using a small labelled calibration set and no retraining.

Fiducio does not depend on any particular architecture: a PyTorch U-Net, nnU-Net, SegFormer or any other model that produces logits can be calibrated. It accompanies the paper Rethinking Post-Hoc Calibration in Semantic Segmentation (Kirscher et al., Transactions on Machine Learning Research, 2026; preprint arXiv:2607.01902), which introduces the translation-invariant and class-conditional calibrators (MSc, CDC, CMSap, CMSop) implemented here.

Why calibrate?

Segmentation networks can be over-confident: predictions with confidence 0.99 need not be correct 99% of the time. Post-hoc calibration learns a lightweight transform on a held-out calibration set that aims to make predicted probabilities match observed frequencies, without retraining the model. See Concepts for the setup and Calibrators for the decision and ordering guarantees.

Install

pip install fiducio

Requires Python ≥ 3.10 and PyTorch ≥ 2.10. See Installation for extras, the GitHub install and the security rationale for the PyTorch floor.

A 30-second example

import torch
from fiducio import TemperatureScaling

torch.manual_seed(0)
logits = torch.randn(8, 4, 64, 64)  # (B, C, H, W) from your model
labels = torch.randint(0, 4, (8, 64, 64))  # (B, H, W) calibration labels

calibrator = TemperatureScaling(device="cpu").fit(logits, labels)
new_logits = torch.randn(2, 4, 64, 64)  # separate cases
probs = calibrator.transform(new_logits)  # calibrated probabilities

These random tensors demonstrate the API, not a calibration benchmark. Continue with the Quickstart or see before/after reliability diagrams.

Documentation

Start with the Quickstart, then Concepts, Calibrators and the Examples. The paper project page has interactive illustrations and published results.

Project status

Fiducio is beta (0.x). The API is usable and tested, but may still change before 1.0. It is not certified for safety-critical clinical use.