DeepBeliefNet: Initialize a Deep Belief Net

In xrobin/DeepLearning: Deep Learning of neural networks

Description

Creates a Deep Belief Net (DBN), precisely a DeepBeliefNet object, with the given specifications. It consists of a stack of RestrictedBolzmannMachine layers that will be created according to the specifications.

Usage

DeepBeliefNet(layers, ..., initialize = c("0", "uniform"))

Arguments

layers	a single Layer object or a list of layers as returned by Layers
...	same as layers
initialize	whether to initialize weights and biases with 0 or random uniform values

Value

an object of class DeepBeliefNet containing the following elements:

• layers: The layers of the RBM

• rbms: a list of RestrictedBolzmannMachine objects making up the network.

• pretrained, unrolled, finetuned: boolean switches to mark the state of the network.

Copying/Cloning

#' For performance purposes, the weights are stored in an environment. This means that when you copy the DeepBeliefNet with an assignment, you do not copy the weights and any modification you make to the new object will be propagated to the original one, and reciprocally. Use clone to control this and make a copy of the weights whenever you need it. Note that all the functions defined in the package do this by default.

See Also

RestrictedBolzmannMachine, Layer

Examples

dbn <- DeepBeliefNet(Layers(c(784, 1000, 500, 250, 30), input="continuous", output="gaussian"))
# Identical as
dbn2 <- DeepBeliefNet(Layer(784, "continuous"), Layer(1000, "binary"), Layer(500, "binary"), 
                      Layer(250, "binary"), Layer(30, "gaussian"))
print(dbn)
methods(class="DeepBeliefNet")

xrobin/DeepLearning documentation built on Sept. 18, 2020, 5:23 a.m.