Regression is not supported for networks with LSTM layers

when setting up a network with LSTM (as in the example below), Matlab gives an error about using regression with LSTM.

1) Matlab has some nice support (finally) for regression output, as described in https://www.mathworks.com/help/nnet/ug/define-regression-output-layer.html is there some workaround that allows defining a regression loss function for LSTM (but is not built in yet)?

2) Why is this even an issue? LSTMs can easily be applied for regression given a MSE loss. It would be nice if there was some way to incorporate, but I'm not sure how. I think it will benefit the community.

Thanks!

------------------- load JapaneseVowelsTrain inputSize = 12; outputSize = 100; %Number of hidden. outputMode = 'last'; numClasses = 9; layers = [ sequenceInputLayer(inputSize) lstmLayer(outputSize,'OutputMode',outputMode) fullyConnectedLayer(numClasses) softmaxLayer regressionLayer] maxEpochs = 150; miniBatchSize = 27; options = trainingOptions('sgdm', ... 'MaxEpochs',maxEpochs, ... 'MiniBatchSize',miniBatchSize); %% net = trainNetwork(X,Y,layers,options);

-------------------------

[Jake]

I am having the same problem and have no seen any info from MATLAB about this...

[Rolf]

I also need to combine LSTM and a regression layer. It would be nice if someone from Mathworks could comment on this.

[Abolfazl Nejatian]

maybe in 2018a, they fix it!

[Sofía]

I'm working with the 2018a version, and it's still a problem...