Stopping the neural network by tr.gradient

In training an ANN using FITNET , I noticed , the tr.gradient gives a row matrice that the number of columns are the number of iterations , and the last column is the gradient reported on the train window

I tried doing :

for h=Hmin:dH:Hmax
    j = j+1
    net            = fitnet(10);
    net = init(net);  % Improving Results since we use patternet we should use init
    [ net tr y  ]  = train( net, x, t );
    e = gsubtract(t,y);
    performance = perform(net,t,y)

      if tr.gradient(end) < 0.05
          tr.stop
      end

but it only stops the Validation test , not the actual training test , is there a way to do this ? and also when I retrain after a gradient like 0.503 and I get a smaller gradient , if from my outputs one is calculated not so precisely , the only thing happens is that , another output will be unprecise.

I have 8 inputs and 3 outputs

Answer 1

[Ahmed]

Maybe you are looking for the property “trainParam.min_grad”.

 net = fitnet(10);
 net.trainParam.min_grad % default 1e-7
 net.trainParam.min_grad = 1e-5;
 net.trainParam.min_grad % changed to 1e-5

[farzad]

Yes exactly , but how should I stop the training after comparison ?

[Greg Heath]

It stops automatically.

Why are you not using the default values?

[farzad]

well I didn't know that the default value is 1e-7 , it seemed to me that 1e-5 is the smallest number I have seen on the training while I was following.

is there any way to increase the accuracy of the loop ? only by changing H ? I changed it from 40 to 80, but Hub is 139

[farzad]

Dear Professor , still , with Ntrials = 20 , and as you have mentioned that MATLAB will stop in a very low gradient. I wonder that when I manually check some of my inputs , there is an error as big as integers. sometimes like 8 or 9.

[Greg Heath]

I have found the following usually sufficient

MSEgoal = 0.01*mean(var(target',1))
MinGrad = MSEgoal/100

net.trainParam.goal     = MSEgoal;
net.trainParam.min_grad = MinGrad;

Hope this helps.

^Thank you for formally accepting my answer

Greg

[farzad]

Thank you very much dear professor

I wish I could accept ,but it was a comment

[Greg Heath]

Not a problem.

Good Luck

Greg