How to create personalized layers

Hello everyone

I built a customed regression output layer named mylayer from here https://it.mathworks.com/help/deeplearning/ug/define-custom-regression-output-layer.html and I want to add it to the other layers of the network. I should use trainNetwork but I only found this example of definiton of layers:

layers = [

imageInputLayer([28 28 1])

convolution2dLayer(5,20)

batchNormalizationLayer

reluLayer

fullyConnectedLayer(1)

myLayer('myL')]

My problem however is NOT an image classification, I just want a couple of hidden layers with a number of neurons to be chosen and the usual weight and biases, passing as inputs a some scalar values. How should I define my object layers?

Thanks you in advice for your help!

Answer 1

[Anshika Chaurasia]

Hi Fabrizio,

It is my understanding that you have successfully created the custom Regression Ouput Layer – ‘myLayer’. You want to have some hidden layers and ‘myLayer’ in layers array. You could consider following codes:

layers = [
    imageInputLayer([28 28 1])
    fullyConnectedLayer(20)
    reluLayer % optional
    fullyConnectedLayer(1)
    myLayer('myL')]

Refer to fullyConnectedLayer documentation for weight and bias properties of fullyConnectedLayer.

[Fabrizio Bernardi]

Thank you for the answer! Is imageInputlayer ok if I don't have to deal with images? Searching I found sequenceInputLayer that could be useful, but trying with a dataset in matlab it gives me this error:

Error using trainNetwork (line 170)

Number of elements must not change. Use [] as one of the size inputs to automatically calculate the appropriate size for that

dimension.

Error in test_prova (line 17)

net = trainNetwork(bodyfatInputs,bodyfatTargets,layers,options);

Caused by:

Error using reshape

Number of elements must not change. Use [] as one of the size inputs to automatically calculate the appropriate size for that

dimension.

This is the code I used ( my customed layer is the layer using mae loss function such as in the example here https://it.mathworks.com/help/deeplearning/ug/define-custom-regression-output-layer.html):

layer = maeRegressionLayer('mae');

load bodyfat_dataset

layers = [

sequenceInputLayer(13)

lstmLayer(40)

maeRegressionLayer('mae')];

options = trainingOptions('sgdm');

net = trainNetwork(bodyfatInputs,bodyfatTargets,layers,options);

Thank you for the help!

[Fabrizio Bernardi]

Edit: now I tried with

layers = [

sequenceInputLayer(13)

%lstmLayer(40)

fullyConnectedLayer(20)

reluLayer % optional

fullyConnectedLayer(1)

maeRegressionLayer('mae')];

and it gives result! Even if with

YPred = predict(net,bodyfatInputs);

predictionError = YPred - bodyfatTargets;

The erros are very large and the output are almost all the same though...