Reparametrization

class torchbayesian.bnn.utils.Reparametrization(parametrization: Module, original: Tensor, *, unsafe: bool = False)

Bases: Module

This class wraps the variational posterior module and handles safety checks for the replacement of the parameter or buffer by the variational posterior and its forward call.

It is the type of ‘module.reparametrizations[tensor_name]’ when ‘module[tensor_name]’ has been reparametrized with ‘register_reparametrization()’.

Parameters

parametrizationModule

The parametrization function; the variational posterior in the context of Bayes by Backprop.

originalTensor

The tensor that is being reparametrized.

unsafebool

Whether to bypass correctness checks. Optional. Defaults to False.

Attributes

variational_posteriorModule

The parametrization replacing ‘original’.

Notes

This class is used internally by ‘register_reparametrization()’. It shall not be instantiated by the user.

forward() → Tensor

Wraps the parametrization’s forward call. Checks for scripting.

In the context of Bayes by Backprop, this returns a sample from the variational posterior.

extra_repr() → str

Returns the variational posterior put in place.

Returns

extra_reprstr

The str extra representation of the reparametrization.