PretrainedGaussianPosterior

class torchbayesian.bnn.PretrainedGaussianPosterior(param: Tensor)

Bases: VariationalPosterior

This class is a diagonal Gaussian variational posterior whose mean parameter is initialized from an existing tensor.

Samples tensors via the reparametrization trick.

This posterior is useful when converting a pretrained model into a Bayesian neural network (BNN) via Bayes by Backprop (BBB) variational inference (VI), as it initializes the Gaussian variational distribution centered on the pretrained weights.

Parameters

paramTensor

The tensor (e.g. pretrained parameter) to initially center the posterior distribution on.

Attributes

muParameter

The variational parameter of the mean of the distribution.

rhoParameter

The variational parameter that parametrizes the standard deviation of the distribution via softplus.

classmethod from_param(param: Tensor, **kwargs) → PretrainedGaussianPosterior

Alternate constructor used by ‘get_posterior’ inside ‘bnn.BayesianModule’.

Overrides the default ‘from_param’ constructor of ‘VariationalPosterior’ to pass along ‘param’.

Parameters

paramTensor

The tensor (parameter or buffer) being replaced by a variational posterior.

**kwargs

Additional keyword arguments passed along to the variational posterior constructor.

Returns

posterior_instancePretrainedGaussianPosterior

An instance of the class.

reset_parameters() → None

Initializes the variational parameter ‘rho’ of the Gaussian posterior N(mu, sigma), where sigma = softplus(rho).

property sigma: Tensor

Returns the standard deviation parameter of the Gaussian distribution.

Returns

sigmaTensor

The standard deviation parameter of the Gaussian distribution

property distribution: Normal

Returns a ‘torch.distributions.Normal’ for KL divergence computation.

Returns

distributionNormal

A diagonal Gaussian distribution.

sample_parameters() → Tensor

Samples parameter values from the variational posterior distribution.

Returns

paramTensor

A value of the parameter, sampled from the variational posterior.