Deep-UQ

Uncertainty quantification for deep learning. One package, 20+ methods, any PyTorch model.

What is Deep-UQ?¶

Deep-UQ is a PyTorch toolkit that adds calibrated uncertainty estimates to neural network predictions. Train your model normally, then wrap it with any of six UQ methods to get confidence intervals that grow when the model is unsure.

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="full")
la.fit(train_loader)
la.optimize_prior_precision()

result = la.predict_uq(x_test)
# result.mean, result.epistemic_var, result.total_var

6 Method Families

Deep Ensembles, Variational Inference, Laplace, SGLD, MC Dropout, and Gaussian Processes — all behind a unified predict_uq() API.

Scientific ML Ready

First-class support for FNO, DeepONet, Graph Neural Operators, PINNs, and convolutional surrogates with 36 executable tutorials.

Zero External UQ Dependencies

All methods implemented natively in PyTorch. No GPyTorch, no laplace-torch — just pip install uqdeepnn.

Choose Your Method¶

If you need...	Use	Effort
Quick baseline, no retraining	MC Dropout or Test-Time Augmentation	Minimal
Post-hoc uncertainty on a trained model	Laplace or Temperature Scaling	Low
Single forward pass, distance-aware	SNGP or Evidential DL	Low
Single-run Bayesian approximation	SWAG / MultiSWAG	Low
Calibrated multi-model uncertainty	Deep Ensembles	Medium
Memory-efficient ensembles	Batch Ensemble or Packed Ensemble	Medium
Full Bayesian weight posteriors	Variational Inference or Flipout	Medium
Posterior samples via MCMC	SGLD / SGHMC / Cyclical SGMCMC	Medium
Particle-based inference	SVGD	Medium
Nonparametric with kernel priors	Gaussian Processes	Varies
Distribution-free coverage	Conformal Prediction (split, weighted, adaptive)	Low
Reject uncertain predictions	Selective Prediction	Low

Detailed method guides →

Laplace Backends (all native)¶

Structure	Speed	Quality	Best for
`diag`	Fastest	Good	Large models, quick estimates
`lowrank_diag`	Fast	Better	Strong OOD detection
`block_diag`	Medium	High	Good accuracy/speed balance
`kron`	Medium	High	Multi-output layers
`full`	Slow	Reference	Small models, maximum quality

Model Architectures¶

Family	Models	UQ Methods
Dense	MLP, PINN	All 6 methods
Spatial	CNN, ResNet, U-Net	Ensembles, MC Dropout, Laplace
Operators	DeepONet, FNO 2D/3D	Laplace, Ensembles
Graph	Graph Neural Operator	Ensembles, Laplace
GPs	Exact, Sparse, Deep Kernel, Multi-task	Native Bayesian

Full architecture inventory →

Install¶

pip install uqdeepnn

Or from source:

git clone https://github.com/Vispikarkaria/Deep-UQ.git
cd Deep-UQ && pip install -e ".[dev,tests]"

Beyond Prediction¶

Module	What it does
`deepuq.metrics`	ECE, CRPS, AUROC, PICP, Brier, AURC — evaluate UQ quality
`deepuq.active`	Uncertainty-guided data acquisition (BALD, uncertainty sampling)
`deepuq.constraints`	Physics constraints: positivity, conservation, monotonicity
`deepuq.propagation`	Uncertainty growth tracking for autoregressive rollouts

Tutorials¶

50+ executable notebooks covering every method and architecture:

Core UQ methods — SWAG, SNGP, Evidential, Calibration, Selective Prediction
Ensembles — Deep, Batch, Packed Ensembles
MCMC — SGLD, SGHMC, Cyclical SGMCMC, SVGD
Scientific ML — DeepONet, FNO, PINNs, CNNs, U-Nets, Graph Operators
Gaussian Processes — Exact, Sparse, Multi-Fidelity, Deep Kernel
Evaluation — Metrics, Active Learning, Uncertainty Propagation

Unified Output¶

Every method returns a UQResult with consistent fields:

result = method.predict_uq(x)
result.mean           # predictive mean
result.epistemic_var  # model uncertainty
result.aleatoric_var  # data noise (when modeled)
result.total_var      # combined uncertainty

Links¶


Getting Started	Installation and quickstart
Method Guides	Mathematical details for each method
API Reference	Full class and function documentation
Tutorials	36 executable notebook guides
Contributing	How to contribute
Changelog	Release history