whitening

unscent_warp

pyvbmc.whitening.unscent_warp(fun, x, sigma)

Compute the unscented transform of the warping function fun.

Parameters:

funfunction

A single-argument function which warps input points.

x(n,D) or (D,) np.ndarray

The input mean for which to compute the unscented transform. A single row is broadcast against the rows of sigma.

sigma(n,D) or (D,) np.ndarray

The input matrix of standard deviations or scale parameters for which to compute the unscented transform. A single row is broadcast against the rows of x.

Returns:

x_warped_mean(n,D) or (D,) np.ndarray

The unscented estimate of the mean, with one row per row of the broadcast inputs. It is one-dimensional where x is and the inputs give a single row.

x_warped_sigma(n,D) or (D,) np.ndarray

The unscented estimate of the standard deviation / scale parameters, shaped like x_warped_mean.

x_warped(U,n,D) np.ndarray

The warped mean points at x_warped[0, :, :], and the warped std. simplex points, at [1:, :, :]. Here U=2*D+1.

Raises:

ValueError

If the rows/columns of x and sigma cannot be coerced to match.

warp_input

pyvbmc.whitening.warp_input(vp, optim_state, function_logger, options)

Compute input warping of variables and update the cached points in function_logger accordingly.

Currently supports only a whitening transformation: a rotation and rescaling of the inference space such that the variational posterior achieves unit diagonal covariance.

Parameters:

vpVariationalPosterior

The variational posterior whose covariance the whitening transformation is computed from. That covariance is expressed in the inference space of the posterior’s own parameter transformer, which need not be the space the run is currently in.

optim_statedict

The dictionary recording the current optimization state.

function_loggerFunctionLogger

The record including cached function values. Its parameter transformer defines the current inference space, the one the search bounds and the cached search points are given in.

optionsOptions

The options of the run: warp_rotoscaling, warp_roto_corr_thresh and warp_cov_reg are read.

Returns:

parameter_transformerParameterTransformer

A ParameterTransformer object representing the new transformation between original coordinates and inference space coordinates, with the input warping applied.

optim_statedict

An updated copy of the original optimization state dict.

function_loggerFunctionLogger

An updated copy of the original function logger.

warp_actionstr

The label of the warp in the iteration display, “rotoscale”.

Raises:

ValueError

If warp_cov_reg is neither a finite real number nor a function of the number of training points that returns one.

Notes

Random draws use vp.rng.

warp_gp_and_vp

pyvbmc.whitening.warp_gp_and_vp(parameter_transformer, gp_old, vp_old, vbmc)

Update the GP and VP with a given warp transformation.

Applies an updated ParameterTransformer object (with new warping transformation) to the GP and VP parameters.

Parameters:

parameter_transformerParameterTransformer

The new (warped) transformation between input coordinates and inference coordinates.

gp_oldgpr.GP

The current Gaussian process.

vp_oldVariationalPosterior

The current variational posterior.

vbmcVBMC

The current VBMC object.

Returns:

vpVariationalPosterior

An updated copy of the original variational posterior.

hyp_warpednp.ndarray

The GP hyperparameters with the warping transformation applied, one row per hyperparameter sample, of shape (Ns_gp, n_hyp).

Raises:

ValueError

If the mean function of the GP is none of the zero, constant and negative quadratic means.