Documentation says we can provide a loss layer to NetTrain as 3rd argument, why doesn't this work?

clejaniuval · September 22, 2025, 5:51pm

I tried with and without explicit loss layer. Neither works.

giulioa · September 23, 2025, 8:10am

Hi Iuval,

the third argument is used to extract training properties.
To specify the probability cross-entropy instead of the index based one, you must use the LossFunction option.

Try this:

famTrained = NetTrain[famNet, trainingdatafam,
   LossFunction -> CrossEntropyLossLayer["Probabilities"]]

clejaniuval · September 23, 2025, 1:40pm

Giulio, you are my hero! Can you please tell me how to extract the parameters for my user-defined “quadratic” layer? I know how to extract weights from a linear layer, but I am not understanding the syntax in general. See file enclosed.

Topic		Replies	Views
Use custom loss function in NetTrain? Questions computer-science , data-science , statistics-and-probability , machine-learning , neural-networks	2	13897	December 20, 2016
Create a custom loss function with NetTrain? Questions data-science , machine-learning , neural-networks	3	13479	August 14, 2018
Add a customized complex loss function to a neural network? Questions data-science , machine-learning , neural-networks	3	6807	February 6, 2019
Plot the round loss,batch loss and use other properties of nettrain? Questions graphics-and-visualization , machine-learning , neural-networks	4	9002	April 10, 2018
Problem with CrossEntropyLossLayer["Index"] in NetTrain[] Questions data-science , wolfram-repositories , neural-networks	0	2247	July 16, 2025

Documentation says we can provide a loss layer to NetTrain as 3rd argument, why doesn't this work?

Related topics