uqdeepnn 0.1.0

Unified deep learning uncertainty quantification (UQ) toolkit: VI, Laplace, SGLD MCMC, MC Dropout (PyTorch)

deepuq

Unified deep learning uncertainty quantification (UQ) toolkit in PyTorch.

Implements four widely used methods:

1. Variational Inference (VI) — Bayes by Backprop with BayesianLinear layers.
2. Laplace Approximation — via laplace-torch with diagonal/kronecker/full Hessians.
3. MCMC (SGLD) — Stochastic Gradient Langevin Dynamics sampler for NN posteriors.
4. MC Dropout — Keep dropout active at test-time and aggregate Monte Carlo predictions.

Install (local)

git clone https://github.com/yourusername/deepuq.git
cd deepuq
pip install -e .

Coming from PyPI? See the section Publish to PyPI below.

Quickstart

import torch
from deepuq.models.simple import MLP
from deepuq.methods.mc_dropout import MCDropoutWrapper

model = MLP(input_dim=784, hidden_dims=[256,128], output_dim=10, p_drop=0.2)
uq = MCDropoutWrapper(model, n_mc=50)
mean, var = uq.predict(torch.randn(32, 784))
print(mean.shape, var.shape)

See the examples/ folder for end‑to‑end training scripts on MNIST/FashionMNIST.

Methods

• VI: Place Gaussian posteriors over weights with reparameterization trick and KL regularization.
• Laplace: Fit a Gaussian around a MAP solution using the Hessian; calibrate with a prior precision.
• MCMC (SGLD): Inject Gaussian noise into SGD steps to sample from the posterior.
• MC Dropout: Use dropout at inference; Monte Carlo average for mean and variance.

Documentation

• API docs are in each module and the README sections below.
• Run pydoc deepuq.methods.vi etc., or open the examples.

Contributing

PRs welcome. Please add tests under tests/ and run pytest.

License

MIT