Extract: Extract or Replace Layers of a Deep Belief Net

Description Usage Arguments Details Note See Also Examples

Description

Operators to extract or replace layers of a DeepBeliefNet object, and to extract weights and biases of a RestrictedBolzmannMachine.

Usage

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
## S3 method for class 'DeepBeliefNet'
x[[i]]

## S3 replacement method for class 'DeepBeliefNet'
 x[[i]] <- value

## S3 method for class 'DeepBeliefNet'
x[i, drop = FALSE]

## S3 method for class 'RestrictedBolzmannMachine'
x$name

## S3 replacement method for class 'RestrictedBolzmannMachine'
x$name <- value

Arguments

x

the DBN (for [ and )

i

indices specifying the layer to extract or replace

value

the RestrictedBolzmannMachine to insert.

drop

whether to drop the DeepBeliefNet if it contains a single RestrictedBolzmannMachine.

name

either W, b or c

Details

[[ extracts (and [[<- replaces) exactly one RestrictedBolzmannMachine layer of the DeepBeliefNet.

[ extracts one or more layers of the DeepBeliefNet. If the returned DeepBeliefNet. has exactly one RestrictedBolzmannMachine, the drop argument controls whether the function returns an RestrictedBolzmannMachine object (TRUE) or a DeepBeliefNet object containing one single RestrictedBolzmannMachine (FALSE).

When extracting layers, the pretrained and finetuned switches will match those of the DeepBeliefNet that was supplied, while unrolled will be set to FALSE. When replacement layers, the pretrained switch will be on if all the RestrictedBolzmannMachines are pretrained, while finetuned and unrolled will be set to FALSE.

For the time being, weights and biases of a RestrictedBolzmannMachine cannot be replaced.

Note

If the DeepBeliefNet contains N layers, there are N-1 RestrictedBolzmannMachines.

See Also

DeepBeliefNet, RestrictedBolzmannMachine

Examples

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
dbn <- DeepBeliefNet(Layers(c(784, 1000, 500, 250, 30), input="continuous", output="gaussian"))
# Extract a RBM
dbn[[2]]

# Replace a RBM
dbn[[1]] <- RestrictedBolzmannMachine(Layer(10, "binary"), Layer(1000, "binary"))
dbn[[2]] <- RestrictedBolzmannMachine(Layer(1000, "binary"), Layer(500, "binary"))
dbn[[4]] <- RestrictedBolzmannMachine(Layer(250, "binary"), Layer(2, "gaussian"))
## Not run: 
# Cannot replace incompatible RestrictedBolzmannMachines
dbn[[2]] <- RestrictedBolzmannMachine(1000, 400, input="binary", output="binary")
dbn[[2]] <- RestrictedBolzmannMachine(100, 500, input="binary", output="binary")
dbn[[2]] <- RestrictedBolzmannMachine(1000, 500, input="binary", output="continuous")
dbn[[2]] <- RestrictedBolzmannMachine(1000, 500, input="gaussian", output="binary")
## End(Not run)

# Get the first layer as RestrictedBolzmannMachine
dbn[[1]]
dbn[1, drop=TRUE]

# Get the first layer as DeepBeliefNet
rbm <- dbn[1]
# Get the first layer as RestrictedBolzmannMachine
rbm$W
rbm$b
rbm$c
rbm$b <- 1:10

xrobin/DeepLearning documentation built on May 17, 2018, 3:51 a.m.