MATLAB: How detect by code which condition halted NN traning

feedforwardnetnntraning

I'm using a feedforwardnet NN and I would like to know from code which condition determined NN training stop (max validation checks reached, max epoch, performance goal, etc..). Is it possible? Thanks for your replys. Regards.

Sergio

Best Answer

[x,t]        = simplefit_dataset;
net          = fitnet(10);
[net tr]     = train(net,x,t);
stopcriteria = tr.stop
tr           = tr

stopcriteria =

Validation stop.

tr =

        trainFcn: 'trainlm'
      trainParam: [1x1 struct]
      performFcn: 'mse'
    performParam: [1x1 struct]
        derivFcn: 'defaultderiv'
       divideFcn: 'dividerand'
      divideMode: 'sample'
     divideParam: [1x1 struct]
        trainInd: [1x66 double]
          valInd: [1x14 double]
         testInd: [1x14 double]
            stop: 'Validation stop.'
      num_epochs: 16
       trainMask: {[1x94 double]}
         valMask: {[1x94 double]}
        testMask: {[1x94 double]}
      best_epoch: 10
            goal: 0
          states: {1x8 cell}
           epoch: [1x17 double]
            time: [1x17 double]
            perf: [1x17 double]
           vperf: [1x17 double]
           tperf: [1x17 double]
              mu: [1x17 double]
        gradient: [1x17 double]
        val_fail: [1x17 double]
       best_perf: 1.3525e-04
      best_vperf: 1.7348e-04
      best_tperf: 5.9455e-04

Hope this helps.

Thank you for formally accepting my answer

Greg

Related Solutions

MATLAB: Is validation during NN training performed ?

I found answer to this question

according to https://uk.mathworks.com/help/nnet/ug/analyze-neural-network-performance-after-training.html

after training the training record is created which contains valid indexes to train, validation and test subfiles. I just copied training record example. In this case trainInd, valInd and testInd contain indexes allocated by dividerand so later they can be used for evaluation.

So calling dividerand and allocating variables before train is simply not necessary as it is done automatically and indexes are known after training

Krzysztof

tr = struct with fields:
      trainFcn: 'trainlm'
    trainParam: [1x1 struct]
    performFcn: 'mse'
  performParam: [1x1 struct]
      derivFcn: 'defaultderiv'
     divideFcn: 'dividerand'
    divideMode: 'sample'
   divideParam: [1x1 struct]
      trainInd: [1x176 double]
        valInd: [1x38 double]
       testInd: [1x38 double]
          stop: 'Validation stop.'
    num_epochs: 9
     trainMask: {[1x252 double]}
       valMask: {[1x252 double]}
      testMask: {[1x252 double]}
    best_epoch: 3
          goal: 0
        states: {1x8 cell}
         epoch: [0 1 2 3 4 5 6 7 8 9]
          time: [1x10 double]
          perf: [1x10 double]
         vperf: [1x10 double]
         tperf: [1x10 double]
            mu: [1x10 double]
      gradient: [1x10 double]
      val_fail: [0 0 0 0 1 2 3 4 5 6]
     best_perf: 12.3078
    best_vperf: 16.6857
    best_tperf: 24.1796

MATLAB: How to test data on neural network toolbox after training it.

Not enough information. Post your code.

How many classes? c = ?

How many input vectors per class?

What version of NNTBX?

Are you using patternnet or the obsolete newpr?

Convert the original c-class target matrix, target0, to c-dimensional unit matrix column form so that size(target) = [ c N ] and sum(target) = ones(1,N). Then class indices are obtained from

classind = vec2ind(target) % class indices

Use 'trainscg' and { 'tansig' 'logsig' }

[net tr ] = train(net,input,target);

tr = tr % reveals training details. Check it out!

newclassind = vec2ind(net(newinput))