rbm_gpu: Fit a Restricted Boltzmann Machine
In zachmayer/rbm: Restricted Boltzmann Machines

Description Usage Arguments Details Value References Examples

This function fits an RBM to the input dataset. It internally uses sparse matricies for faster matrix operations

rbm_gpu(x, num_hidden = 10, max_epochs = 1000, learning_rate = 0.1,
  use_mini_batches = TRUE, batch_size = 250, initial_weights_mean = 0,
  initial_weights_sd = 0.1, momentum = 0, dropout = FALSE,
  dropout_pct = 0.5, retx = FALSE, activation_function = NULL,
  verbose = FALSE, ...)

`x`	a sparse matrix
`num_hidden`	number of neurons in the hidden layer
`max_epochs`	Maximum learning epochs
`learning_rate`	Learning Rate
`use_mini_batches`	Use sub-samples for training for each iteration. This usually results in MUCH faster learning.
`batch_size`	Sample size for mini batches
`initial_weights_mean`	Mean of initial random weights
`initial_weights_sd`	Standard deviation of initial random weights
`momentum`	Use momentum when learning. (Helps move faster through "half pipe" shaped regions).
`dropout`	Use dropout when learning (sort of a form of regularization).
`dropout_pct`	What percent of neurons to drop out (0 to 1)
`retx`	whether to return the RBM predictions for the input data
`activation_function`	function to convert hidden activations (-Inf, Inf) to hidden probabilities [0, 1]. Must be able to operate on sparse "Matrix" objects.
`verbose`	Print lots of messages while training
`...`	not used

This code is (mostly) adapted from edwin chen's python code for RBMs, avaiable here: https://github.com/echen/restricted-boltzmann-machines. Some modifications (e.g. momentum) were adapted from Andrew Landgraf's R code for RBMs, available here: http://alandgraf.blogspot.com/2013/01/restricted-boltzmann-machines-in-r.html.

a rbm object

## Not run: 
#Setup a dataset
set.seed(10)
data(movie_reviews)
data(george_reviews)

#Fit a PCA model and an RBM model
PCA <- prcomp(movie_reviews, retx=TRUE)
RBM <- rbm_gpu(movie_reviews, retx=TRUE)

#Examine the 2 models
round(PCA$rotation, 2) #PCA weights
round(RBM$rotation, 2) #RBM weights

#Predict for new data
predict(PCA, george_reviews)
predict(RBM, george_reviews, type='activations')
predict(RBM, george_reviews, type='probs')
predict(RBM, george_reviews, type='states')

#Predict for existing data
predict(PCA)
predict(RBM, type='probs')

## End(Not run)