uqdeepnn 0.2.1

PyTorch uncertainty quantification toolkit with deep ensembles, Bayes-by-Backprop VI, Laplace, SGLD, MC Dropout, Gaussian Processes, and scientific ML backbones.

Deep-UQ is the most comprehensive uncertainty quantification toolkit for deep learning in PyTorch — 20+ UQ methods, scientific ML architectures, evaluation metrics, active learning, and physics constraints, all behind a unified predict_uq() API.

Install

pip install uqdeepnn

Quick Example

from deepuq.models import MLP
from deepuq.methods import LaplaceWrapper

model = MLP(input_dim=1, hidden_dims=[64, 64], output_dim=1)
# ... train as usual ...

la = LaplaceWrapper(model, likelihood="regression", hessian_structure="kron")
la.fit(train_loader)
result = la.predict_uq(x_test)
# result.mean, result.epistemic_var, result.total_var

20+ UQ Methods

Bayesian & Ensemble: Deep Ensembles, Batch Ensemble, Packed Ensemble, SWAG, MultiSWAG, Variational Inference (BBB, Last-Layer), SVGD

Post-hoc & Single-pass: Laplace (6 Hessian structures + GLM predictive), SNGP, Evidential DL, MC Dropout, Test-Time Augmentation, Temperature/Isotonic Calibration

MCMC: SGLD, SGHMC, Cyclical SGMCMC

Gaussian Processes: Exact, Sparse, Multi-Fidelity, Deep Kernel, Multi-task, Heteroscedastic, Spectral Mixture, Classification

Distribution-free: Conformal (split, CQR, weighted, adaptive), Selective Prediction

Full Toolkit

• deepuq.methods — All 20+ UQ method wrappers
• deepuq.models — Neural architectures (MLP, FNO, DeepONet, GNO, PINN, CNN, UNet, GPs)
• deepuq.metrics — ECE, CRPS, AUROC, PICP, Brier, interval score, AURC
• deepuq.active — Active learning (Uncertainty Sampling, BALD, loop orchestration)
• deepuq.constraints — Physics constraints (positivity, bounds, conservation, monotonicity)
• deepuq.propagation — Uncertainty propagation for autoregressive rollouts

Scientific ML Backbones

DeepONet (1D/2D), FNO (2D/3D), Graph Neural Operators, CNN/ResNet/U-Net surrogates, PINNs, Diffusion models — all with first-class UQ support.

50+ Tutorials

Every method has a self-contained executable notebook with mathematical explanations, training, prediction, and visualization.

Key Features

• Zero external UQ dependencies — everything in pure PyTorch
• Unified API — every method returns UQResult(mean, epistemic_var, aleatoric_var, total_var)
• Any architecture — works with any nn.Module
• Evaluation built-in — calibration, scoring rules, OOD detection metrics
• Production-ready — selective prediction, post-hoc calibration, single-pass methods

Documentation

• Docs: https://vispikarkaria.github.io/Deep-UQ/
• Tutorials: https://vispikarkaria.github.io/Deep-UQ/tutorials/
• API: https://vispikarkaria.github.io/Deep-UQ/api/
• GitHub: https://github.com/Vispikarkaria/Deep-UQ

Package Names

• PyPI: uqdeepnn
• Import: deepuq