Deep-Learning Toolbox: regression withour responses - Error using trainNetwork (line 150)

Hello,

I am using the Deep-Learning Toolbox/

I am trying to train my network to perform a locality-sensitive representation of 16-point vectors under one real value. Each input vector of the training database has a label (vector of two real values) that represents characteristics of the 16-D vector. Basically, I want that behaviour:

• If two input vectors (16-points) have close labels, then their output representation must be close
• If their labels are very different, their output representation must be different

Since this is not really a regression, I wrote a custom output layer presented earlier. However when I want to train the network, I get the error:

>> DL_main_V2_2
Error using trainNetwork (line 150)
Invalid training data. The output size (1) of the last layer does not match the response size (2).

Error in DL_main_V2_2 (line 41)
net = trainNetwork(X,Y,layers,options);

It is easy to understand: my "responses" are indeed of size two, and the output of size one. And I don't want it to be otherwise.

Here is the code for the main file:

TcsY = [];
Y = [];
n = 0;
for i = 1:40
    for j = 1:36
        n = n+1;
        Y(n,1) = Data.Te_List(i);
        Y(n,2) = Data.ne_List(j);
        for k = 1:16
            TcsY(k,n) = Data.Line_intensity(k,i,j);
        end
    end
end

Y = Y';
X = TcsY;

layers = [
    sequenceInputLayer(16,'Name','input')

    fullyConnectedLayer(16,'Name','fc1')
    softmaxLayer('Name','sm1')
    
    fullyConnectedLayer(1,'Name','fc2')
    localityLayer('custom_out')
    ];

options = trainingOptions('sgdm', ...
    'MiniBatchSize',300, ...
    'MaxEpochs',100, ...
    'InitialLearnRate',1e-4, ...
    'LearnRateSchedule','piecewise', ...
    'LearnRateDropFactor',0.1, ...
    'LearnRateDropPeriod',20, ...
    'Shuffle','every-epoch', ...
    'ValidationData',{X,Y}, ...
    'ValidationFrequency',20, ...
    'Plots','training-progress', ...
    'Verbose',true);

net = trainNetwork(X,Y,layers,options);

I may add that the middle layers are just random stuff I'm putting here, at this stage.

And the custom output layer:

classdef localityLayer < nnet.layer.RegressionLayer
    
    properties
        % (Optional) Layer properties.
        sigma = 10;
        diam = 0.01;
        % Layer properties go here.
    end
    
    methods
        function layer = localityLayer(Name)           
            layer.Name = Name;
            layer.Description = 'locality-sensitive output layer';
        end

        function loss = forwardLoss(layer, Y, T)
            % Return the loss between the predictions Y and the 
            % training targets T.
            %
            % Inputs:
            %         layer - Output layer
            %         Y     – Predictions made by network
            %         T     – Training targets
            %
            % Output:
            %         loss  - Loss between Y and T

            L = 0;
            N = size(T,2);% mini-batch size
            
            T1 = (T(1,:)-min(T(1,:)))/(max(T(1,:))-min(T(1,:)));
            T2 = (T(2,:)-min(T(2,:)))/(max(T(2,:))-min(T(2,:)));
            TT = [T1;T2];
            for i = 1:N
                for j = 1:i
                    if norm(TT(:,i)-TT(:,j)) <= layer.diam
                        C_ij = exp(-norm(TT(:,i)-TT(:,j))/layer.sigma);
                        L = L + C_ij * norm(Y(i)-Y(j));
                    end
                    % else C_ij is 0 so nothing would happen anyway
                end
            end
        end
        
        function dLdY = backwardLoss(layer, Y, T)
            % Backward propagate the derivative of the loss function.
            %
            % Inputs:
            %         layer - Output layer
            %         Y     – Predictions made by network
            %         T     – Training targets
            %
            % Output:
            %         dLdY  - Derivative of the loss with respect to the predictions Y        

            dLdY = 0;
            N = size(T,2);% mini-batch size
            
            T1 = (T(1,:)-min(T(1,:)))/(max(T(1,:))-min(T(1,:)));
            T2 = (T(2,:)-min(T(2,:)))/(max(T(2,:))-min(T(2,:)));
            TT = [T1;T2];
            for i = 1:N
                for j = 1:i
                    if norm(TT(:,i)-TT(:,j)) <= layer.diam
                        C_ij = exp(-norm(TT(:,i)-TT(:,j))/layer.sigma);
                        dLdY = dLdY + C_ij * sign(Y(i)-Y(j));
                    end
                    % else C_ij is 0 so nothing would happen anyway
                end
            end
        end
    end
end

I have looked at the code generating the error to find a workaround, but I didn't have any idea how to do otherwise. I am feeling like the error comes from the fact that it is more designed for regression; what did I do wrong ? Maybe I shouldn't inherit from nnet.layer.RegressionLayer ?

Any idea would be welcomed.

Thank you