Number of observations in X and Y disagree. - Error training a U-Net for classification

I am trying to train a U-Net for classification aim. Each of the 8x8 central pixels of my input matrix 48x48 (I have a 4-D matrix as input: 48x48x4xN) have to be classified in 8 different classes. I built a U-Net with the 'ConvolutionPadding' option set as 'valid' (this is the reason why starting from a 48x48 the last layer output is 8x8) and the 'pixelClassificationLayer' as classification layer.

Follwing my code where A and B (input and responses, respectively) are not the real data but are used only as an example.

A=rand(48,48,4,100);%Input
B=ones(8,8,1,100);%Responses
B=categorical(B(:));

EncoderDepth=2;
NFEF=[16,32,64];
sz=size(A);
MaxEpoch=200;%Maximum number of epochs
ILR=1e-06;%Initial learn rate value
SGF=[0.9,0.99,0.999];
MiniBatch=2000;
L2R=[1e-05,1e-04,1e-03];

for i=3%:numel(NFEF)
    lgraph=unetLayers([48,48,4],2,'EncoderDepth',EncoderDepth,'NumFirstEncoderFilters',NFEF(i),'ConvolutionPadding','valid');
    
    lgraph=removeLayers(lgraph,'Final-ConvolutionLayer');
    lgraph=removeLayers(lgraph,'Softmax-Layer');
    lgraph=removeLayers(lgraph,'Segmentation-Layer');
    
    lgraph=addLayers(lgraph,convolution2dLayer([1,1],1,'name','Final-ConvolutionLayer'));
    lgraph=addLayers(lgraph,softmaxLayer('name','Softmax-Layer'));
    lgraph=addLayers(lgraph,pixelClassificationLayer('name','classificationLayer'));
    lgraph=connectLayers(lgraph,'Decoder-Stage-2-ReLU-2','Final-ConvolutionLayer');
    lgraph=connectLayers(lgraph,'Final-ConvolutionLayer','Softmax-Layer');
    lgraph=connectLayers(lgraph,'Softmax-Layer','classificationLayer');
    
    for j=1:numel(ILR)
        for k=1%:numel(SGF)
            options = trainingOptions('rmsprop','InitialLearnRate',ILR(j),...
                'MiniBatchSize',MiniBatch,'Shuffle','every-epoch',...
                'SquaredGradientDecayFactor',0.99,...
                'MaxEpochs',MaxEpoch,...
                ...%'L2Regularization',L2R(j),...
                ...%'ValidationData',{MSGdata_VAL_V11,DPRGMI_ValDataset_DeepLearn}, ...
                ...%'ValidationFrequency',floor(sz(4)/MiniBatch), ...
                'ExecutionEnvironment','cpu','Plots','training-progress');
            
            [net,info]=trainNetwork(A,B,lgraph,options);
            UNET.net=net;
            UNET.info=info;
            EpochSTR=num2str(MaxEpoch);
            MiniBatchSTR=num2str(MiniBatch);
            ILRstr=num2str(ILR(j));
            NFEFstr=num2str(NFEF(i));
%             SGFstr=num2str(SGF(k));
%             L2Rstr=num2str(L2R(j));
            save([PathOut,filesep,'UNET_NoPadding_MaxPool_',EpochSTR,'Epochs_',MiniBatchSTR,'MBS_',ILRstr,'ILR_',...
                NFEFstr,'NFEF_OversampledData_NoParallax_NoBN_ValDataset_PRthre',PRthre,'mmh_',VERin,'_',VERout,'.mat'],'UNET');
        end
    end
end

When I run the code I get the following error message:

Error using trainNetwork

Number of observations in X and Y disagree.

Error in untitled2 (line 42)

[net,info]=trainNetwork(A,B,lgraph,options);

I would appreciate any help to fix this problem.

Answer 1

[Matt J]

You have 100 training images. Therefore, B should be 100x1, not 6400x1.

[Leo Pio D'Adderio]

Yes Matt, I have 100 training images but I want to classify each pixel of the 8x8 central pixels of each image. Therefore for each image, I have 64 pixels to be classified times 100 images. This is the reason I have B 6400x1. My aim is not to classify the whole image. Anyway, I get the error message when I start the training.

[Matt J]

At minimum, then, you would probably have to remove the line which flattens the response data,

B=categorical(B(:));

Failing that, you probably have to create a pixelLabelDatastore.

[Leo Pio D'Adderio]

Matt, what I learned is that pixelLabelDatastore works reading image files (e.g. jpeg, png, tiff, ecc.). I do not have these files, but netCDf files that I read and extract my numerical matrices (that I save in .mat files). So, I talk about images because they come from satellite images, but I work with numerical matrices.

I am still facing the problem to classify each pixel of my matrix.

[Matt J]

I do not have these files, but netCDf files that I read and extract my numerical matrices (that I save in .mat files)

The type of file you store your responses in should not be an issue. You can use the ReadFcn parameter to allow a pixelLabelDataStore to read any kind of file:

https://www.mathworks.com/help/vision/ref/pixellabeldatastore.html#d124e89940