How can i change performFcn of neural network training?

When i try it the obvious way the change is ignored,

>> net = feedforwardnet(20);
>> net.performFcn='msereg';
>> net.performFcn

ans =

mse

Also i've noticed that changing parameters in net.trainParam is completely ignored in latter call of train function

>> net = feedforwardnet(20);
>> matlabpool open
>> net.trainParam.epochs = 130;
>> net.trainParam

ans = 
(...)
Maximum Epochs                      epochs: 130

>> net=train(net,WE2,WY2, 'useParallel','yes','useGPU','yes');
NOTICE: Jacobian training not supported on GPU. Training function set to TRAINSCG.
Lab 1: 

    Training with TRAINSCG.
    Epoch 0/1000, (...)

As u can see the epoch limit didn't actually change. What am I doing wrong, and how to do it properly?

*im using matlab r2012b

Answer 1

[Greg Heath]

>> net = feedforwardnet(20);

>> net.performFcn='msereg';

>> net.performFcn

ans =

msereg

1. Don't know what caused your error.

2. I cannot help with other problem

Sorry,

Greg

[Greg Heath]

Try

net = feedforwardnet(20,'trainscg');

Hope this helps.

Greg

[georges matar]

trainscg is a training fct, not the perform fct

[Neelabh Jyoti Saharia]

Hi Greg, my problem is kind of similar with this. I want to create a feedforwardnet and then train it using a custom made loss function.

For that, can I create a matlab function, function loss = custom_loss(y,x). And then put net.performFcn = 'custom_loss'.

Other thing remains same as in the default feedforward.