unable to incorporate own design Loss function in r2024a

Switching from r2023b to r2024 I made some changes in my Net (CNN). e.g. modified input/output and replace RegressionLayer with SoftmaxLayer, using trainnet function, etc.

I expected better performance, perspective compatibility (RegressionLayre is not more recommended) and have a vision of my Net optimization with use of Prune approach etc.

To the contrary to the previous version I am not able to involve my own Loss function (as it was done in previous version).

The (siplified) code is below, the used synthax was inspired by example:

https://www.mathworks.com/matlabcentral/answers/2100631-how-can-i-define-a-custom-loss-function-using-trainnet

The error message is:

Error using trainnet (line 46)

Error calling function during training.

Error in callMyLoss (line 55)

myTrainedNet = trainnet(Y,target,net, @(Y,target) myOwnLoss(name,Y,target),options);

Caused by:

Error using myOwnLoss

The specified superclass 'nnet.layer.softmaxLayer' contains a parse error, cannot be found on MATLAB's

search

path, or is shadowed by another file with the same name.

Error in callMyLoss>@(Y,target)myOwnLoss(name,Y,target) (line 55)

myTrainedNet = trainnet(Y,target,net, @(Y,target) myOwnLoss(name,Y,target),options);

Error in nnet.internal.cnn.util.UserCodeException.fevalUserCode (line 11)

[varargout{1:nargout}] = feval(F, varargin{:});

    classdef myOwnLoss < nnet.layer.softmaxLayer
    % own Loss

    methods
%function layer = sseClassificationLayer(name)        
       function layer = myOwnLoss(name)
            % layer = sseClassificationLayer(name) creates a sum of squares
            % error classification layer and specifies the layer name.

            % Set layer name.
            layer.Name = name;

            % Set layer description.
            layer.Description = 'my own Loss v.2024a';
        end

        function loss = forwardLoss(layer, Y, T)
%%%        function loss = forwardLoss(Yo, To)
            % loss = forwardLoss(layer, Y, T) returns the Tdiff loss between
            % the predictions Y and the training targets T.

            disp("myLoss");
            aa=1;

            % just something very simple
            loss = sum(Y-T,'all');
        end



% original backwardLoss
        function dX = backwardLoss(layer, Y, T)

            numObservations = size( Y, 3);
            dX = (Y - T)./numObservations;
        end

    end
end

%=======================eof=========================

    classdef myOwnLoss < nnet.layer.softmaxLayer

    % own Loss


    methods
%function layer = sseClassificationLayer(name)        
       function layer = myOwnLoss(name)
            % layer = sseClassificationLayer(name) creates a sum of squares
            % error classification layer and specifies the layer name.

            % Set layer name.
            layer.Name = name;

            % Set layer description.
            layer.Description = 'my own Loss v.2024a';
        end

        function loss = forwardLoss(layer, Y, T)
%%%        function loss = forwardLoss(Yo, To)
            % loss = forwardLoss(layer, Y, T) returns the Tdiff loss between
            % the predictions Y and the training targets T.

            disp("myLoss");
            aa=1;

            % just something very simple
            loss = sum(Y-T,'all');
        end



% original backwardLoss
        function dX = backwardLoss(layer, Y, T)

            numObservations = size( Y, 3);
            dX = (Y - T)./numObservations;
        end

    end
end

%=======================eof=========================

Answer 1

[Malay Agarwal]

Hi @Petr Kolar,

The error occurs since "softmaxLayer" is actually a function and cannot act as a base class for a class. You can confirm this by using the following command:

which softmaxLayer

And then opening the file path:

function layer = softmaxLayer( nameValueArgs )
% softmaxLayer   Softmax layer
% ...
end

I also tried running your example code in MATLAB R2023b and am receiving the same error, suggesting that it does not work in R2023b as well.

According to the documentation, it is recommended to define a custom loss function as a MATLAB function instead of a class: https://www.mathworks.com/help/deeplearning/ug/define-custom-training-loops-loss-functions-and-networks.html#mw_40e667e2-1ea1-4793-b079-8bc763144200. The function should have the syntax "loss = f(Y,T)", where "Y" and "T" are the predictions and targets, respectively. Using a function has several benefits such as supporting automatic differentiation if all your operations are compatible with "dlarray".

If you want a more complex loss function which takes inputs other than the predictions and targets, you need to use a custom training loop with the custom loss function, as detailed in the following example: https://www.mathworks.com/help/deeplearning/ug/train-generative-adversarial-network.html#TrainGenerativeAdversarialNetworkGANExample-9.

Hope this helps!

[Petr Kolar]

Hi Malay,

thank you for the quick reaction, you are right, but it is not exactly my case.

Originally (under r2023b) I had time series and (after training) get some probabilities for every input point of them (it was realized by "regessionLayer" working in sequence-to-sequence mode).

At the input was "inputsequenceLayer", followed by most outer loop constructed by folding and unfolding layers. In such a way the time series were converted into (pseudo)images and process by Unet (constructed by "unet3dLayers"). At the end the conversion was reverted back to time series. The Net was trained with "trainNetwork"

And I was optionally able to use my own-design Loss function.

Following the new recommendation, I switch to "image mode", use "inputLayer", train Net with "trainnet" and as output layer use "softmaxLayer" (the Net is constructed by "unet3d").

It works, at least formally. Now I want to call my own Loss function (as before), but I am not able to.

Is my problem/desire now clearer?

Regards Petr

[Malay Agarwal]

Could you please try replacing your class-based loss function with a function-based one and then call "trainnet" as follows?

trainnet(Y,target,net, @myOwnLoss,options);

[Petr Kolar]

tested - it doesn't work neither, however with message error:

Error using deep.internal.train.Trainer/train (line 74)

File: fevalUserCode.m Line: 11 Column: 6

Invalid expression. When calling a function or indexing a variable, use parentheses. Otherwise, check for

mismatched delimiters.

Error in deep.internal.train.trainnet (line 54)

net = train(trainer, net, mbq);

Error in trainnet (line 42)

[net,info] = deep.internal.train.trainnet(mbq, net, loss, options, ...

Error in callMyLoss (line 55)

myTrainedNet = trainnet(Y,target,net, @(Y,target) myOwnLoss2(layer,Y,target),options);

[Malay Agarwal]

Could you share the training code and any data files that you might be using so that I can take a closer look?

[Petr Kolar]

Dear Malay,

I don't fully understand your requirement (it has nothing to do with the formal problem of calling myLoss fcn, I'd say), but I've prepared an answer.

An article about our work was recently published

https://authors.elsevier.com/a/1jenq,yaZpTNg

As the supplement we provided also a demo code, data example and trained net (all items belong to the "old" version MTL) - this material is available via

https://github.com/PetkaGFU/AECubeDemo

We use own Loss (a combination of RMS, correlation and minimizing time-shift-of-probability-maximas - the setup is based on our numerical experiments - cf. the article).

Recently we switched to the latest MATLAB version (r2024a) expecting future compatibility and faster performance. Last but not least: I believed in improvement of the proposed method efficiency and robustness via Prune approach. Unfortunately, it is not available for some layers required by 3D-Unet (particulary 3D convolutional operations) - at least in the current MTL version :-( .

I prepared a version (AECubeDemo2 - below) using new-version Net design (the same demo-data are used, however reformatted in the code).

If problem with given links, let me know I'll sent the items directly.

Thank you for the interest and the care

Regards Petr

function predictedProbab = AECubeDemo2
%
% updated version - use r2024a CNN formalism (prepared for own Loss fce
% utilistaion proble discussion.


%
%
% "Accoustic Emission Cube Demo " 
%  Accoustic Emission multiple evetnst detection with use of CubeNet/U-Net CNN
%
%
%
% the code is suplement to the article:
% Petr Kolář and Matěj Petružálek:
% Discrimination of doubled Acoustic Emission events using Neural Networks
%
% submited to Ultrasonic
%
% input: data file indetification is presribed below
% output: predicted probabilities from testing data
% note: the interpretation of these probabilities is not included in this
% demo code and should be done by user
%
% required files:
%           * code  (function AECubeDemo)
%           * (testing) data (examples quotted in the article)
%           * trained RNN for onset(s) detecion
%           * trained RNN for OT prediction
% (all these files are part of the package)
%

%


%
% created by P. Kolar   [e-mail]
%
%
% compatibility: created under MATLAB R2022b
% required: MATLAB core
%           Statistics and Machine Learning Toolbox
%           Signal Processing Toolbox
%
% version 1.0 / 04/10/2023
%


% how the Net was created:
%
% l=unet3dLayers([4 4 100 1],3,'EncoderDepth',2,'NumFirstEncoderFilters',16)
%

close all

close(findall(0,'tag','NNET_CNN_TRAININGPLOT_FIGURE'));

netName='C2NN400_50-50_L123f.mat';  % net trainde on large data

train=0;     % no training - pre-trained Net is used
train=1;    % training on available data (seimograms) is performed
if train
    netName='trainedNet_2024a';   % Net trained on actual data
end

% plot outputs
plot1=0;
plot1=1;

rng('default')   % for reprodacibility

% input data files

files=what('d:\Data2count\dataAE_SyntDoubleAg_4mw');    % directory with input files


len1=400;       % sub-seismogram length
dt=1/10e6;      % sampling

%%{
                   
nFiles=length(files.mat);
iFile=randperm(nFiles);

% data division
nT=floor(0.7*nFiles);   %   0.0 - 0.7  for training
nV=floor(0.8*nFiles);   %   0.7 - 0.8  for validation


% position of channes in Cube
posTab=[2 1; 2 2; 3 3; 3 4;...
    1 1; 1 2; 2 3; 2 4;...
    3 1; 3 2; 4 3; 4 4;...
    4 2; 4 1; 1 3; 1 4;];




signal=[];
probab=[];
info=[];
for i=1:nT
    jj=iFile(i);
    str=[files.path,'\',files.mat{jj}];

    out=isfile(str);
    if out ~= 1
        aa=1;
    end
    [signal1,probab1,info1]=getSignal3D(str,posTab,len1);
    lenS=length(signal1);
    signal=[signal signal1];
    probab=[probab probab1];
    info=[info info1];
    aa=1;
end
signalTrain=signal';
probabTrain=probab';
infoTrain=info';


signal=[];
probab=[];
info=[];
for i=nT+1:nV
    jj=iFile(i);
    str=[files.path,'/',files.mat{jj}];

    [signal1,probab1,info1]=getSignal3D(str,posTab,len1);
    lenS=length(signal1);
    signal=[signal signal1];
    probab=[probab probab1];
    info=[info info1];
    aa=1;
end
signalValid=signal';
probabValid=probab';
infoValid=info';


signal=[];
probab=[];
info=[];
iY=1;
for i=nV+1:nFiles

    jj=iFile(i);
    str=[files.path,'/',files.mat{jj}];

    [signal1,probab1,info1,signal0,info0,probab0]=getSignal3D(str,posTab,len1);
    lenS=length(signal1);
    signal=[signal signal1];
    probab=[probab probab1];
    info=[info info1];

    Ysignal{iY}=signal0;
    Yprobab{iY}=probab0;
    Yinfo{iY}=info0;
    iY=iY+1;

    aa=1;
end
signalTest=signal';
probabTest=probab';
infoTest=info';




disp("DATA was red");
%%{

% an example of input data is ploted
% the seismogram number may be modified, if required
len=length(signalTrain);
np=min([2000 round(0.7*len)-1]);
if nV > 0
    hF1=plotCube('',signalTrain{np},probabTrain{np});
end

%............................................................
%
%  NEW net formalism
% 


aa=1;

net = dlnetwork;

% Add Layer Branches
% Add branches to the dlnetwork. Each branch is a linear array of layers.
tempNet = [
    inputLayer([4 4 400 3 1],"SSSCB","Name","encoderImageInputLayer")
    convolution3dLayer([3 3 3],4,"Name","Encoder-Stage-1-Conv-1","Padding","same","WeightsInitializer","he")
    batchNormalizationLayer("Name","Encoder-Stage-1-BN-1")
    reluLayer("Name","Encoder-Stage-1-ReLU-1")
    convolution3dLayer([3 3 3],8,"Name","Encoder-Stage-1-Conv-2","Padding","same","WeightsInitializer","he")
    batchNormalizationLayer("Name","Encoder-Stage-1-BN-2")
    reluLayer("Name","Encoder-Stage-1-ReLU-2")];
net = addLayers(net,tempNet);

tempNet = [
    maxPooling3dLayer([2 2 2],"Name","Encoder-Stage-1-MaxPool","Stride",[2 2 2])
    convolution3dLayer([3 3 3],8,"Name","Encoder-Stage-2-Conv-1","Padding","same","WeightsInitializer","he")
    batchNormalizationLayer("Name","Encoder-Stage-2-BN-1")
    reluLayer("Name","Encoder-Stage-2-ReLU-1")
    convolution3dLayer([3 3 3],16,"Name","Encoder-Stage-2-Conv-2","Padding","same","WeightsInitializer","he")
    batchNormalizationLayer("Name","Encoder-Stage-2-BN-2")
    reluLayer("Name","Encoder-Stage-2-ReLU-2")
    dropoutLayer(0.5,"Name","Encoder-Stage-2-DropOut")];
net = addLayers(net,tempNet);

tempNet = [
    maxPooling3dLayer([2 2 2],"Name","Encoder-Stage-2-MaxPool","Stride",[2 2 2])
    convolution3dLayer([3 3 3],16,"Name","LatentNetwork-Bridge-Conv-1","Padding","same","WeightsInitializer","he")
    batchNormalizationLayer("Name","LatentNetworkBridge-BN-1")
    reluLayer("Name","LatentNetwork-Bridge-ReLU-1")
    convolution3dLayer([3 3 3],32,"Name","LatentNetwork-Bridge-Conv-2","Padding","same","WeightsInitializer","he")
    batchNormalizationLayer("Name","LatentNetworkBridge-BN-2")
    reluLayer("Name","LatentNetwork-Bridge-ReLU-2")
    dropoutLayer(0.5,"Name","LatentNetwork-Bridge-DropOut")
    transposedConv3dLayer([2 2 2],32,"Name","Decoder-Stage-1-UpConv","BiasLearnRateFactor",2,"Stride",[2 2 2],"WeightsInitializer","he")
    reluLayer("Name","Decoder-Stage-1-UpReLU")];
net = addLayers(net,tempNet);

tempNet = crop3dLayer("Name","encoderDecoderSkipConnectionCrop2");
net = addLayers(net,tempNet);

tempNet = [
    concatenationLayer(4,2,"Name","encoderDecoderSkipConnectionFeatureMerge2")
    convolution3dLayer([3 3 3],16,"Name","Decoder-Stage-1-Conv-1","Padding","same","WeightsInitializer","he")
    batchNormalizationLayer("Name","Decoder-Stage-1-BN-1")
    reluLayer("Name","Decoder-Stage-1-ReLU-1")
    convolution3dLayer([3 3 3],16,"Name","Decoder-Stage-1-Conv-2","Padding","same","WeightsInitializer","he")
    batchNormalizationLayer("Name","Decoder-Stage-1-BN-2")
    reluLayer("Name","Decoder-Stage-1-ReLU-2")
    transposedConv3dLayer([2 2 2],16,"Name","Decoder-Stage-2-UpConv","BiasLearnRateFactor",2,"Stride",[2 2 2],"WeightsInitializer","he")
    reluLayer("Name","Decoder-Stage-2-UpReLU")];
net = addLayers(net,tempNet);

tempNet = crop3dLayer("Name","encoderDecoderSkipConnectionCrop1");
net = addLayers(net,tempNet);

tempNet = [
    concatenationLayer(4,2,"Name","encoderDecoderSkipConnectionFeatureMerge1")
    convolution3dLayer([3 3 3],8,"Name","Decoder-Stage-2-Conv-1","Padding","same","WeightsInitializer","he")
    batchNormalizationLayer("Name","Decoder-Stage-2-BN-1")
    reluLayer("Name","Decoder-Stage-2-ReLU-1")
    convolution3dLayer([3 3 3],8,"Name","Decoder-Stage-2-Conv-2","Padding","same","WeightsInitializer","he")
    batchNormalizationLayer("Name","Decoder-Stage-2-BN-2")
    reluLayer("Name","Decoder-Stage-2-ReLU-2")
    convolution3dLayer([1 1 1],3,"Name","encoderDecoderFinalConvLayer")
    softmaxLayer("Name","FinalNetworkSoftmax-Layer")];
net = addLayers(net,tempNet);

% clean up helper variable
clear tempNet;

% Connect Layer Branches
% Connect all the branches of the network to create the network graph.
net = connectLayers(net,"Encoder-Stage-1-ReLU-2","Encoder-Stage-1-MaxPool");
net = connectLayers(net,"Encoder-Stage-1-ReLU-2","encoderDecoderSkipConnectionCrop1/in");
net = connectLayers(net,"Encoder-Stage-2-DropOut","Encoder-Stage-2-MaxPool");
net = connectLayers(net,"Encoder-Stage-2-DropOut","encoderDecoderSkipConnectionCrop2/in");
net = connectLayers(net,"Decoder-Stage-1-UpReLU","encoderDecoderSkipConnectionCrop2/ref");
net = connectLayers(net,"Decoder-Stage-1-UpReLU","encoderDecoderSkipConnectionFeatureMerge2/in2");
net = connectLayers(net,"encoderDecoderSkipConnectionCrop2","encoderDecoderSkipConnectionFeatureMerge2/in1");
net = connectLayers(net,"Decoder-Stage-2-UpReLU","encoderDecoderSkipConnectionCrop1/ref");
net = connectLayers(net,"Decoder-Stage-2-UpReLU","encoderDecoderSkipConnectionFeatureMerge1/in2");
net = connectLayers(net,"encoderDecoderSkipConnectionCrop1","encoderDecoderSkipConnectionFeatureMerge1/in1");
net = initialize(net);

%Plot Layers
figure; hold on
plot(net);



aa=1;


%..........................................................


% training opitons


sv1=[];
pv1=[];
nn= length(signalValid);
for ii=1:nn
  % sv1=[sv1;signalValid{ii}];
  % pv1=[pv1;probabValid{ii}];
   sv1(1:4,1:4,1:400,1:3,ii)=signalValid{ii};
   pv1(1:4,1:4,1:400,1:3,ii)=probabValid{ii};
end


% % 
 maxEpochs = 10;
 miniBatchSize = 20;

options = trainingOptions('sgdm', ...
    'MaxEpochs',maxEpochs, ...
    'MiniBatchSize',miniBatchSize, ...
    'InitialLearnRate',0.0045, ...
    'GradientThreshold',0.5, ...
    'Shuffle','every-epoch', ...
    'Plots','training-progress',...
    'ExecutionEnvironment','cpu',...
    'ValidationData',{sv1,pv1}, ...
     'Metrics',"rmse", ...
    'OutputNetwork','best-validation-loss',...
    'Verbose',0);

tic
%%{
if train

    % % % old ver.
    % % netCNN_3dUnetL2 = trainNetwork(signalTrain,probabTrain,layers,options);
    % % save(netName,'netCNN_3dUnetL2');
    % % disp(' trained Cube-net was saved')

    % new version (incl. data reformating).
    st1=[];
    pt1=[];
    nn= length(signalTrain);
    for ii=1:nn
        st1(1:4,1:4,1:400,1:3,ii)=signalTrain{ii};
        pt1(1:4,1:4,1:400,1:3,ii)=probabTrain{ii};
    end
       netCNN_3dUnetL2 = trainnet(st1,pt1,net,@(st1,pt1) myLoss24a(st1,pt1),options);    % DOES NOT WORK !!
    % netCNN_3dUnetL2 = trainnet(st1,pt1,net,"mse",options);       % IT WORKS  !!
    
   
    
end
toc
load(netName);

disp(['net ',netName,' - net was loaded']);



Yplus=1:14;
Yplus=repmat(Yplus',[1 1024]);

nEvT=nFiles-nV;   % nummber of predicted events

predictedProbab=zeros(nEvT,14,1024);

hFp=[];
for i=1:nEvT
    YpredP=zeros(14,1024);

    Ysig1=Ysignal{i};
    Yprobab1=Yprobab{i};
    if plot1
        if plot1==2, close(hFp); end
        hFp=figure('Tag','Onpred'); hold on
    end

    for ii=1:8  % shift loop
        ii1=(ii-1)*78+1;
        ii2=ii1+len1-1;
        Ysig1s=Ysig1(:,:,ii1:ii2,:);
        [Ypred1_3d] = predict(netCNN_3dUnetL2,Ysig1s);

        % Cube data decomposition into maxtrix
        for jj=1:14
            pos1=posTab(jj,1);
            pos2=posTab(jj,2);
            tmp=squeeze(Ypred1_3d(pos1,pos2,:,1:3));
            tmp=smoothdata(tmp,1);
            tmp=tmp(:,2) + (1 -tmp(:,1) - tmp(:,3));
            tmp(tmp<0)=0;

            tmp0=squeeze(YpredP(jj,ii1:ii2));
            YpredP(jj,ii1:ii2)= max([tmp0; tmp']);
            if ii == 1
                tmp=squeeze(Yprobab1(pos1,pos2,:,1:3));
                tmp=tmp(:,2) + (1 -tmp(:,1) - tmp(:,3));
                tmp(tmp<0)=0;
                YprobabP(jj,:)= tmp;
            end
            aa=1;
        end
        if plot1
            plot((YpredP+Yplus)');
            plot((0.45*YprobabP+Yplus)',':b');
            strTit=(['evNo.: ',num2str(i),', 2nd ev.shift: ',num2str(Yinfo{i}.shift)]);
            title({' Predicted probab. (full) and Targed (dotted)';strTit});
            xlabel('samples');
            ylabel('chanels');
        end
        aa=1;
    end

    on0=Yinfo{i}.on0;
    on1=Yinfo{i}.on1;

    Yinfo1=Yinfo{i};

    if plot1 
        disp(['evNo./2nd ev.shift ',num2str([i Yinfo1.shift])]);
    end
    predictedProbab(i,1:14,:) = YpredP;
end

return
end

%-------------------------------------------------
function [signal3,probab3,info3,signal0,info0,probab0]=getSignal3D(str,posTab,len1)
%
% read data
%
% input: file_name in 'str'
% output:   xx3  - subwindowed signals/probab/infos  for training/validataion
%           xx0  - the whole signals/probab/infos for interpratation (prediction)
%

tmp = load(str);
cNNdatS2 = tmp.cNNdatS2;


sigma=4;
sigma2=800;

x=1:1024;

tmp=cNNdatS2;
tmp.sig=[];
info0=tmp;
info0.pP1=zeros(1,14);
info0.pP2=zeros(1,14);



nSig = 1;
shiftSig1 = 64;
on0=double(cNNdatS2.on0);
on1=double(cNNdatS2.on1);
NN=12;   % numbr of shifts to get sub-seismograms

signal3=num2cell(zeros(1,NN));
probab3=num2cell(zeros(1,NN));
info3=num2cell(zeros(1,NN));

signal0=zeros(4,4,1024,3);
probab0=zeros(4,4,1024,3);

for j=1:NN
    shiftSig0 = round(1+rand(1)*shiftSig1/2);
    %    shiftSig0 = 1;

    sig2=zeros(14,len1+1,3);
    probab2=zeros(14,len1+1,3);
    probab14=zeros(14,1024,3);
    for i=1:14
        sig1=cNNdatS2.sig(i,:);

        % for basic signal
        p0=exp(-(x-on0(i)).^2/(2*sigma));
        p2=exp(-(x-on0(i)).^2/(2*sigma2));

        pi=-p0+1;
        p2i=-p2+1;

        n0(1:1024)=0;
        n0(1:on0(i)-2*sigma-1)=1;
        %     disp([i j]);

        fin1=1024-(on0(i)+2*sigma);
        n0(on0(i)-2*sigma : on0(i))=pi(on0(i)-2*sigma:on0(i));
        n0(on0(i)+2*sigma:end)=p2i(on0(i):on0(i)+fin1);

        e0(1:1024)=1;
        e0(1:on0(i)-1)=0;
        e0(on0(i):on0(i)+2*sigma)=pi(on0(i) : on0(i)+2*sigma);
        e0(on0(i)+2*sigma:end)=p2(on0(i) : on0(i)+fin1);

        % for shifted signal
        p1=exp(-(x-on1(i)).^2/(2*sigma));
        p2=exp(-(x-on1(i)).^2/(2*sigma2));

        pi=-p1+1;
        p2i=-p2+1;

        n1(1:1024)=0;
        n1(1:on1(i)-2*sigma-1)=1;
        %     disp([i j]);

        fin1=1024-(on1(i)+2*sigma);
        n1(on1(i)-2*sigma : on1(i))=pi(on1(i)-2*sigma:on1(i));
        n1(on1(i)+2*sigma:end)=p2i(on1(i):on1(i)+fin1);

        e1(1:1024)=1;
        e1(1:on1(i)-1)=0;
        e1(on1(i):on1(i)+2*sigma)=pi(on1(i) : on1(i)+2*sigma);
        e1(on1(i)+2*sigma:end)=p2(on1(i) : on1(i)+fin1);



        n=min([n0;n1]);
        p=max([p0;p1]);
        e=max([e0;e1]);

        prb1=[n; p; e];

        % % %        plot of  probabilities (if desired)
        % % %
        % % aa=1;
        % % if info0.nEv0 ==955  && info0.nEv1==1121
        % %     if i==3 && info0.shift == -120
        % %
        % %         figure; hold on
        % %         plot(prb1');
        % %         s=sum(prb1,1);
        % %         pTot= 1 +p -n -e;
        % %         plot(pTot,':','LineWidth',1.5);
        % % %        plot(s,':');
        % % sig1=sig1/max(abs(sig1));
        % % plot(sig1-1.1)
        % %         legend(' pN',' pP',' pC',' P',' signal');
        % %         set(gca,'ylim',[-2.2 2.1]);
        % %         xlabel('samples');
        % %         aa=1;
        % %     end
        % % end


        sigEnd =  (j-1)*shiftSig1+1+shiftSig0 + len1;
        if sigEnd > 1024, break, end

        for k=1:3
            sig2(i,:,k)=sig1((j-1)*shiftSig1+1+shiftSig0:sigEnd);
            probab2(i,:,k)=prb1(k,(j-1)*shiftSig1+1+shiftSig0:sigEnd);
            probab14(i,:,1:3)=prb1';
        end


    end

    if sigEnd > 1024, break, end

    signal1=zeros(4,4,len1,3);
    probab1=zeros(4,4,len1,3);
    for i=1:16
        pos=posTab(i,:);
        if i <=14
            signal1(pos(1),pos(2),1:len1,1:3)=sig2(i,1:len1,:);
            probab1(pos(1),pos(2),1:len1,1:3)=probab2(i,1:len1,:);
        end
        if i==13
            ii=15;
            pos=posTab(ii,:);
            signal1(pos(1),pos(2),1:len1,1:3)=sig2(i,1:len1,:);
            probab1(pos(1),pos(2),1:len1,1:3)=probab2(i,1:len1,:);
        elseif i==14
            ii=16;
            pos=posTab(ii,:);
            signal1(pos(1),pos(2),1:len1,1:3)=sig2(i,1:len1,:);
            probab1(pos(1),pos(2),1:len1,1:3)=probab2(i,1:len1,:);
        end

    end
    signal3{nSig}=signal1;
    probab3{nSig}=probab1;


    info1=info0;
    info1.on0=info0.on0-(j-1)*shiftSig1+1+shiftSig0;
    info1.on1=info0.on1-(j-1)*shiftSig1+1+shiftSig0;
    info1.TO0o=info0.TO0o-(j-1)*shiftSig1+1+shiftSig0;
    info1.TO1o=info0.TO1o-(j-1)*shiftSig1+1+shiftSig0;
    info3{nSig}=info1;

    %
    % the whole signal
    %
    if j==1
        for i=1:16
            pos=posTab(i,:);
            if i <=14
                signal0(pos(1),pos(2),:,1)=cNNdatS2.sig(i,:);
                signal0(pos(1),pos(2),:,2)=cNNdatS2.sig(i,:);
                signal0(pos(1),pos(2),:,3)=cNNdatS2.sig(i,:);
                probab0(pos(1),pos(2),:,1:3)=probab14(i,:,1:3);
            end
            if i==13
                ii=15;
                pos=posTab(ii,:);
                signal0(pos(1),pos(2),:,1)=cNNdatS2.sig(i,:);
                signal0(pos(1),pos(2),:,2)=cNNdatS2.sig(i,:);
                signal0(pos(1),pos(2),:,3)=cNNdatS2.sig(i,:);
                probab0(pos(1),pos(2),:,1:3)=probab14(i,:,1:3);
            elseif i==14
                ii=16;
                pos=posTab(ii,:);
                signal0(pos(1),pos(2),:,1)=cNNdatS2.sig(i,:);
                signal0(pos(1),pos(2),:,2)=cNNdatS2.sig(i,:);
                signal0(pos(1),pos(2),:,3)=cNNdatS2.sig(i,:);
                probab0(pos(1),pos(2),:,1:3)=probab14(i,:,1:3);
            end

        end



    end



    % % %   possible plot for prove
    % % %
    % % %     if info0.nEv0==955 && info0.nEv1==1121 && info0.shift== -120
    % % %         plotCube('',signal3{nSig},probab3{nSig});
    % % %         aa=1;
    % % %     end

    nSig=nSig+1;
    aa=1;

end

signal3(nSig:NN)=[];
probab3(nSig:NN)=[];
info3(nSig:NN)=[];
aa=1;

if info0.TO1o <  info0.TO0o
    info0a=info0;

    info0.evid0=info0a.evId1;
    info0.evid1=info0a.evId0;
    info0.nEv0=info0a.nEv1;
    info0.nEv1=info0a.nEv0;
    info0.on0=info0a.on1;
    info0.on1=info0a.on0;
    info0.Mw0=info0a.Mw1;
    info0.Mw1=info0a.Mw0;
    info0.TO0o=info0a.TO1o;
    info0.TO1o=info0a.TO0o;
    info0.vX=[info0a.vX(2,:);info0a.vX(1,:)];
end

end

%===============eof=============================================

[Malay Agarwal]

Hi @Petr Kolar,

In your code, the following line:

netCNN_3dUnetL2 = trainnet(st1,pt1,net,@(st1,pt1) myLoss24a(st1,pt1),options);

Needs to be:

netCNN_3dUnetL2 = trainnet(st1,pt1,net,@myLoss24a,options);

Notice that only the name of the loss function is passed as a function handle with no arguments. MATLAB will automatically figure out what are the targets and the predictions that should be passed to the loss function based on the other arguments to the "trainnet" function.

I have attached a working version of the script which has a dummy loss function at the end. Please replace it with your own loss function. Also note that the loss function needs to return a scalar value.

[Petr Kolar]

GREAT !! It works, thank you very much.

I woudl appreciate to have such an example in the MATLAB original documentation.

By the way, is there any way to reduce size (input, outputs, connection ..) of a trained net wchich is based on Unet3d architecture (Prune)? But it is probably new question - I'll witl think it over after obtainnig some results by the current configuration ...