neural network - R: Error in nrow[w] * ncol[w] : non-numeric argument to binary operator, while using neuralnet package

Question

I am using neuralnet package for training a classifier. The training data looks like this:

> head(train_data)
   mvar_12      mvar_40 v10       mvar_1   mvar_2  Labels
1 136.51551310       6   0   656.78784220      0      0
2 145.10739860      87   0    14.21413596      0      0
3 194.74940330       4   0   196.62888080      0      0
4 202.38663480       2   0   702.27307720      0      1
5  60.14319809       9   0    -1.00000000     -1      0
6  95.46539380       6   0   539.09479640      0      0

The code is as follows:

n <- names(train_data)
f <- as.formula(paste("Labels ~", paste(n[!n %in% "Labels"], collapse = " + ")))
library(neuralnet)
nn <- neuralnet(f, tr_nn, hidden = 4, threshold = 0.01,        
                stepmax = 1e+05, rep = 1, 
                lifesign.step = 1000,
                algorithm = "rprop+")

The problem arises when I try to make a prediction for a test set:

pred <- compute(nn, cv_data)

Where cv_data looks like:

> head(cv_data)
   mvar_12      mvar_40 v10      mvar_1    mvar_2
1 213.84248210       1   9  -1.000000000     -1
2 110.73985680       0   0  -1.000000000     -1
3 152.74463010      14   0 189.521812800     -1
4  64.91646778       7   0  47.854257730     -1
5 141.28878280      12   0 248.557857500      5
6  55.36992840       2   0   4.785425773     -1

To this I get an error saying:

Error in nrow[w] * ncol[w] : non-numeric argument to binary operator
In addition: Warning message:
In is.na(weights) : is.na() applied to non-(list or vector) of type 'NULL'

Why do I get this error and how can I fix it?

Answer 1

I just came up against the very same problem. Checking the source code of the compute function we can see that it assumes one of the resulting attributes (i.e. weights) only defined when the network finishes the training flawless.

> trace("compute",edit=TRUE)
function (x, covariate, rep = 1) {
    nn <- x
    linear.output <- nn$linear.output
    weights <- nn$weights[[rep]]
    [...]
}

I think the real problem lies on the fact that neuralnet doesn't save the current network once reached the stepmax value, causing this error later in the compute code.

Edit

It seems you can avoid this reset by commenting lines 65 & 66 of the calculate.neuralnet function

> fixInNamespace("calculate.neuralnet", pos="package:neuralnet")
[...]
#if (reached.threshold > threshold) 
#    return(result = list(output.vector = NULL, weights = NULL))
[...]

Then everything works as a charm :)