multiple loss in R

[ivokwee]

Hi. I am trying to get multiple loss function using R in this code

data <- mx.symbol.Variable('data') fc1 <- mx.symbol.FullyConnected(data = data, name = 'fc1', num_hidden = 64) act1 <- mx.symbol.Activation(data = fc1, name = 'relu1', act_type = "relu") fc2 <- mx.symbol.FullyConnected(data = act1, name = 'fc2', num_hidden = 10)

label1 <- mx.symbol.Variable('out1_label') label2 <- mx.symbol.Variable('out2_label') out1 <- mx.symbol.SoftmaxOutput(data = fc2, name='out1', label=label1) out2 <- mx.symbol.SoftmaxOutput(data = fc2, name='out2', label=label2) mlp <- mx.symbol.Group(out1, out2)

Y <- list( out1_label=train.y, out2_label=train.y )

model <- mx.model.FeedForward.create( symbol = mlp, X = t(train.x), y = Y, eval.metric = mx.metric.rmse, array.layout = "colmajor" )

But it get this error:

Error in eval(substitute(expr), envir, enclos) : not compatible with requested type

Is multiple loss supported in the R interface? What might I be doing wrong?

Thanks,

Ivo

[winstywang]

@thirdwing

[thirdwing]

I think this has not been supported in R side.

[Leloew]

Hey, any alternative to this? I am trying to implement an autoencoder and would like to have softmax loss for the categorical variablels and MAE for the numericall one.

[thirdwing]

@Leloew Can I know why you can't build two symbols according to different types of results?

[Leloew]

Hey @thirdwing, thanks for the reply. I want to have a shared hidden representation for both the categorical and the numerical variables. Is it possible to implement this with two seperate symbols? or what do you mean exactly? Do you mean Symbol.group?

Edit ah do you mean the multiple loss is implemented, just not inserting two Labels? ( then it wouldnt be a Problem for the Autoencoder since the Labels are the data. )