rnn (version 1.9.0)

trainr: Recurrent Neural Network

Description

Trains a Recurrent Neural Network.

Usage

trainr(
  Y,
  X,
  model = NULL,
  learningrate,
  learningrate_decay = 1,
  momentum = 0,
  hidden_dim = c(10),
  network_type = "rnn",
  numepochs = 1,
  sigmoid = c("logistic", "Gompertz", "tanh"),
  use_bias = F,
  batch_size = 1,
  seq_to_seq_unsync = F,
  update_rule = "sgd",
  epoch_function = c(epoch_print, epoch_annealing),
  loss_function = loss_L1,
  ...
)

Value

a model to be used by the predictr function

Arguments

Y

array of output values, dim 1: samples (must be equal to dim 1 of X), dim 2: time (must be equal to dim 2 of X), dim 3: variables (could be 1 or more, if a matrix, will be coerce to array)

X

array of input values, dim 1: samples, dim 2: time, dim 3: variables (could be 1 or more, if a matrix, will be coerce to array)

model

a model trained before, used for retraining purpose.

learningrate

learning rate to be applied for weight iteration

learningrate_decay

coefficient to apply to the learning rate at each epoch, via the epoch_annealing function

momentum

coefficient of the last weight iteration to keep for faster learning

hidden_dim

dimension(s) of hidden layer(s)

network_type

type of network, could be rnn, gru or lstm. gru and lstm are experimentale.

numepochs

number of iteration, i.e. number of time the whole dataset is presented to the network

sigmoid

method to be passed to the sigmoid function

use_bias

should the network use bias

batch_size

batch size: number of samples used at each weight iteration, only 1 supported for the moment

seq_to_seq_unsync

if TRUE, the network will be trained to backpropagate only the second half of the output error. If many to one is the target, just make Y have a time dim of 1. The X and Y data are modify at first to fit a classic learning, error are set to 0 during back propagation, input for the second part is also set to 0.

update_rule

rule to update the weight, "sgd", the default, is stochastic gradient descent, other available options are "adagrad" (experimentale, do not learn yet)

epoch_function

vector of functions to applied at each epoch loop. Use it to intereact with the objects inside the list model or to print and plot at each epoch. Should return the model.

loss_function

loss function, applied in each sample loop, vocabulary to verify.

...

Arguments to be passed to methods, to be used in user defined functions

Examples

if (FALSE) {
# create training numbers
X1 = sample(0:127, 10000, replace=TRUE)
X2 = sample(0:127, 10000, replace=TRUE)

# create training response numbers
Y <- X1 + X2

# convert to binary
X1 <- int2bin(X1, length=8)
X2 <- int2bin(X2, length=8)
Y  <- int2bin(Y,  length=8)

# create 3d array: dim 1: samples; dim 2: time; dim 3: variables
X <- array( c(X1,X2), dim=c(dim(X1),2) )

# train the model
model <- trainr(Y=Y,
                X=X,
                learningrate   =  1,
                hidden_dim     = 16  )
}