Problems with compute/predict.nn (CRAN v1.44.2)

[leraman]

Hi,

I recently updated neuralnet to v1.44.2 from v1.33. The newest version appears to contain a broken compute() function, which worked fine in v1.33. It throws following error:
Error in cbind(1, pred) %*% weights[[i]] : non-conformable arguments Calls: predict -> compute -> predict.nn -> <Anonymous> In addition: Warning message: In cbind(1, pred) : number of rows of result is not a multiple of vector length (arg 1) Execution halted

I've read that compute() was decrypted, yet R seems to be unable to find the 'new' predict(), neuralnet::predict() or predict.nn() functions.

[leraman]

(important?) note: the model I'm using was created using version v1.33. Could that be the problem?

Edit: I trained a new model using v1.44.2, problem is still there.

[mnwright]

Hm... On my machine predict() and compute() run fine with models from both version. Could you try to re-install the package and start in a fresh R session? You should be able to run the following code:

library(neuralnet)
nn <- neuralnet(Species == "setosa" ~ Petal.Length + Petal.Width, iris, linear.output = FALSE)
predict(nn, iris)

and get a 150x1 numeric back.

[leraman]

Hi @mnwright

I solved the issue--my bad after all. My script saves a 'minimal' (using as less space as possible) trained model which can be used for prediction in a later stage (using .Rdata format). Version 1.33 only required the 'linear output' and 'weights' elements from the model object, but it turns out v1.44.2 requires the 'model.list' element as well. I might have missed this, but I don't think there is a nice solution/function to save such models?

Thanks !

[mnwright]

We need length(object$model.list$variables) to subset the new data if necessary (if not all columns were used for training). However, I've just changed that to skip the subset if length(object$model.list$variables) is not there. Are your reduced models working now?

[leraman]

I've made sure the 'model.list' is included in the output and retrained my model, to avoid compatibility issues with CRAN versions when used by others. Thanks for your solution anyway!