Using RNN in R with factorial input variables

Question

In rnn package in R, the example shows how to train a RNN with numeric input variables.

The example code is:

library('rnn')

# create training numbers
set.seed(1)
X1 = sample(0:127, 7000, replace=TRUE)
X2 = sample(0:127, 7000, 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   =  0.1,
                hidden_dim     = 10,
                start_from_end = TRUE )

In the above example, X1 and X2 are numerical variables.

How can I train a RNN model in R if the input variables are factor?

Thanks a lot,

score 0 · Answer 1 · answered Jul 13 '16 at 09:27

0

have you tried coercing the factor to numeric:

Y <- as.numeric(Y)

For predicted values you would want to round them and then apply the levels again:

pred <- round(pred)

pred <- as.factor(pred, labels = c(...))

answered Jul 13 '16 at 09:27

Bastiaan Quast

2,802
1
24
50

In fact, I did not try, but from some external resources, it seems to have problem :-( http://stackoverflow.com/questions/17469835/one-hot-encoding-for-machine-learning – Yang Mei Lian Jul 13 '16 at 13:58
rnn is not one hot, that is simply an example – Bastiaan Quast Jul 13 '16 at 14:04
one-hot is a next step. The answer in the link I posted above is talking about converting factor to numeric variables and its effects. – Yang Mei Lian Jul 13 '16 at 16:56
anyway, I will try and post the results here. – Yang Mei Lian Jul 13 '16 at 16:56
Hi, I did that and observe an interest thing: the RNN now will tend to predict with extreme unbalanced output, i.e. for binary classification, RNN tend to predict everything as positive (or negative), but rarely it give a mixed prediction. – Yang Mei Lian Jul 13 '16 at 17:46
have you tried tuning the hyperparameters, learningrate, epochs, etc.? – Bastiaan Quast Jul 14 '16 at 07:18
Yeah, I am doing that. Of course I cannot try all the combinations. Still in progress. – Yang Mei Lian Jul 14 '16 at 09:45

Using RNN in R with factorial input variables

1 Answers1