DenseOrderedWeights: Create a dense layer with ordered first weight column

In enweg/quickbnnr: Quick way to implement Bayesian Neural Networks in R and estimating these using Julia's Turing

Create a dense layer with ordered first weight column

Description

Priors for biases and all weights in columns not being the first are standard normal. The first column of the weight matrix is specified by putting a standard normal prior on W(1, 1) and a truncated standard normal(0, inf) on the remaining weights in that column. The weights in the column are then the cumsum of those prior weights

Usage

DenseOrderedWeights(
  in_size,
  out_size,
  activation = c("identity", "sigmoid", "tanh", "relu")
)

Arguments

in_size	input size
out_size	output size
activation	activation function

enweg/quickbnnr documentation built on April 15, 2022, 3:29 a.m.