Model Basics

BNN(x, y, like::BNNLikelihood, prior::NetworkPrior, init::BNNInitialiser)

Create a Bayesian Neural Network.

Arguments

• x: Explanatory data
• y: Dependent variables
• like: A likelihood
• prior: A prior on network parameters
• init: An initilialiser

Obtain the log of the unnormalised posterior.

Obtain the derivative of the unnormalised log posterior.

NetConstructor{T, F}

Used to construct a network from a vector.

The NetConstructor constains all important information to construct a network like the original network from a given vector.

Fields

• num_params_net: Number of network parameters
• θ: Vector of network parameters of the original network
• starts: Vector containing the starting points of each layer in θ
• ends: Vector containing the end points of each layer in θ
• reconstructors: Vector containing the reconstruction functions for each layer

destruct(net::Flux.Chain{T}) where {T}

Destruct a network

Given a Flux.Chain network, destruct it and create a NetConstructor. Each layer type must implement a destruct method taking the layer and returning a vector containing the original layer parameters, and a function that given a vector of the right length constructs a layer like the original using the parameters given in the vector