bayespy.inference.vmp.transformations.RotateGaussianARD

class bayespy.inference.vmp.transformations.RotateGaussianARD(X, *alpha, axis=- 1, precompute=False, subset=None)[source]

Rotation parameter expansion for bayespy.nodes.GaussianARD

The model:

alpha ~ N(a, b) X ~ N(mu, alpha)

X can be an array (e.g., GaussianARD).

Transform q(X) and q(alpha) by rotating X.

Requirements: * X and alpha do not contain any observed values

__init__(X, *alpha, axis=- 1, precompute=False, subset=None)[source]

Precompute tells whether to compute some moments once in the setup function instead of every time in the bound function. However, they are computed a bit differently in the bound function so it can be useful too. Precomputation is probably beneficial only when there are large axes that are not rotated (by R nor Q) and they are not contained in the plates of alpha, and the dimensions for R and Q are quite small.

Methods

__init__(X, *alpha[, axis, precompute, subset])

Precompute tells whether to compute some moments once in the setup function instead of every time in the bound function.

bound(R[, logdet, inv, Q])

get_bound_terms(R[, logdet, inv, Q])

nodes()

rotate(R[, inv, logdet, Q])

setup([plate_axis])

This method should be called just before optimization.