Initialisation

BayesFlux.BNNInitialiser — Type
abstract type BNNInitialiser end

To initialise BNNs, BNNInitialisers are used. These must be callable and must return three vectors

  • θnet the initial values for the network parameters
  • θhyper the initial values for any hyperparameters introduced by the NetworkPrior
  • θlike the initial values for any extra parameters introduced by the likelihood

Implementation

Every BNNInitialiser must be callable and must return the above three things: 

(init::BNNInitialiser)(rng::AbstractRNG) -> (θnet, θhyper, θlike)
BayesFlux.InitialiseAllSame — Type
InitialiseAllSame(dist::D, like::BNNLikelihood, prior::NetworkPrior)

Initialise all values by drawing from dist