NetTrain & SoftmaxLayer: Expected a vector of indices

Question

Is SoftmaxLayer broken in version 11.3.0.0? I know, it's probably me. Here's a MWE of the issue.

Note that this works fine if e.g. Exp is substituted for SoftmaxLayer[].

Answer 1

It works if you explicitly specify the loss function:

NetTrain[testnet, {{1, 2} -> {0.2, 0.8}}, 
 LossFunction -> CrossEntropyLossLayer["Binary"]]

The default is Automatic, which chooses CrossEntropyLossLayer["Index"] and expects indices as training targets.

From the documentation of LossFunction:

When a loss layer is chosen automatically for a port, the loss layer to use is based on the layer within the net whose output is connected to the port, as follows:

SoftmaxLayer[] -> use CrossEntropyLossLayer["Index"]

ElementwiseLayer[LogisticSigmoid] -> use CrossEntropyLossLayer["Binary"]

NetPairEmbeddingOperator[] -> use ContrastiveLossLayer[]

other non-loss layers -> use MeanSquaredLossLayer[]

loss layers -> use unchanged