mbi.experimental package

Submodules

mbi.experimental.mixture_of_products module

This file is experimental.

It is a close approximation to the method described in RAP (https://arxiv.org/abs/2103.06641) and an even closer approximation to RAP^{softmax} (https://arxiv.org/abs/2106.07153). This implementation is not very optimized. If you would like to improve it, pull requests are welcome.

Notable differences: - Code now shares the same interface as Private-PGM (see FactoredInference) - Named model “MixtureOfProducts”, as that is one interpretation for the relaxed tabular format (at least when softmax is used). - Added support for unbounded-DP, with automatic estimate of total.

class mbi.experimental.mixture_of_products.MixtureOfProducts(products, domain, total)

Bases: object

datavector(flatten=True)

project(cols)

synthetic_data(rows=None)

mbi.experimental.mixture_of_products.adam(loss_and_grad, x0, iters=250)

mbi.experimental.mixture_of_products.mixture_of_products(domain: Domain, loss_fn: MarginalLossFn | list[LinearMeasurement], *, known_total: int | None = None, mixture_components: int = 100, iters: int = 2500, alpha: float = 0.1) → MixtureOfProducts

mbi.experimental.mixture_of_products.synthetic_col(counts, total)

mbi.experimental.public_support module

mbi.experimental.public_support.entropic_mirror_descent(loss_and_grad, x0, total, iters=250)

Performs optimization using entropic mirror descent to find optimal weights.

mbi.experimental.public_support.public_support(domain: Domain, loss_fn: MarginalLossFn | list[LinearMeasurement], *, public_data: Dataset, known_total=None) → Dataset

Module contents