Customized loss function in R

[kosta501]

How can I make customized loss function in R? For example i want to use weighted mean squared loss

[thirdwing]

Have you tried to write a new metric?

See https://github.com/dmlc/mxnet/blob/master/R-package/R/metric.R for examples.

[khalida]

When I follow this example for using a custom evaluation metric; i.e. passing an argument to eval.metric in a call to mx.model.FeedForward.create(), I can see the custom evaluation metric reported during the training, but it does not change the training of the model.

Is it possible to define a custom layer in R so that we can train a network to minimize some other loss function?

[thirdwing]

Sorry that I misunderstood you. You should use mx.symbol.MakeLoss.

[thirdwing]

I provide a simple example in https://github.com/dmlc/mxnet/blob/master/docs/tutorials/r/CustomLossFunction.md

Please check that.

[khalida]

Many thanks for the simple example. This will definitely help me and others to make use of mx.symbol.MakeLoss

[ChrisBotella]

@thirdwing , Sorry if I missed something, but in your tutorial, which besides that is really clear, you didn't indicate what is the nature of the variable label that appears in :
lro <- mx.symbol.MakeLoss(mx.symbol.square(mx.symbol.Reshape(fc1, shape = 0) - label))

Of course when I execute your code i get the error :
mx.varg.symbol.MakeLoss(list(...)) : object 'label' not found

I guess label is a kind of MXNET symbolic variable associated with the y parameter that we give to mx.model.FeedForward.create for optimization. So how did you defined label?

Plus, to go further, how you define and manipulate this variable in the multilabel case, ie when y is a matrix.

[ChrisBotella]

Ok I found issue #6596, and some other which create the symbol label = mx.symbol.Variable('label') and it works.

However, i still don't understand how mx.model.FeedForward.create function understands it must replace of the variable label with the values of parameter y...