Neural Network toolbox - transferFcn and transferParam

I am trying to define a transfer function with parameters in the Neural network toolbox but I am not able to define the parameters correctly

What I did?

1. I copied the LOGSIG Logarithmic sigmoid transfer function file -my transfer function includes a factor in the exponent

2. I defined the parameter to receive as a real scalar with default value 1.0

function param = parameters, param = ...
    [nnetParamInfo( 'alpha', ...
                    'alpha factor in the exponent', ...
                    'nntype.real_scalar', ...
                    1.0, ...
                    'alpha factor in the exponent')]; 
end

3. I include it in the 'apply' and 'da_dn' functions

function a = apply(n,param)
  % The same as logsig but with alpha factor in the exponent
  a = 1 ./ (1 + exp(-n * param.alpha));
  i = find(~isfinite(a));
  a(i) = sign(n(i));
end

4. I create the network and establish the alpha parameter

net = feedforwardnet([10, 10]); 
% ... more initializations ...
net.layers{1}.transferParam.alpha = 1.1;

The LAST LINE raise this ERROR

??? Error using ==> param
Too many input arguments.

Error in ==> network.subsasgn>setLayerTransferParam at 1214
err = nntest.param(functionInfo.parameters,transferParam, ...

Any ideas? I did not find any good documentation / example about how to define a transfer function with parameters

[mladen]

I am trying to do the same thing with another function, if I find a solution I'll let You know :)

[Guru]

Same issue here. If someone has any solution for this problem, please help.

[Greg Heath]

There is no problem worth talking about!!!

 .. . Read my answer

Greg

[Guru]

Actually I don't want to train the network. I'm not using any training algorithms. Instead I'm varying the control parameter of the custom transfer function that I've written. My network is not exactly for the functionality traditionally used with neural networks. So I need to initialise this parameter for the transfer function and vary it based on the output.

This is what the error I'm runnign into,

Undefined function or variable 'newSize'.

Error in network/subsasgn>setLayerTransferParam (line 1231) net.layers{i}.range = repmat(feval(net.layers{i}.transferFcn, ...

Error in network/subsasgn>network_subsasgn (line 189) if isempty(err), [net,err] = setLayerTransferParam(net,i,transferParams); end

Error in network/subsasgn (line 13) net = network_subsasgn(net,subscripts,v,netname);

Error in TestNet (line 21) net.layers{1}.transferParam.no_of_iter = 4;

Answer 1

[George Evers]

Francisco,

Have you tried the solution methods suggested at http://www.mathworks.com/support/solutions/en/data/1-17LOK/index.html?product=NN&solution=1-17LOK ?

George

[Ned Gulley]

Francisco says: "The problem I have is not how to include the function in the library. The problem I have is how to define the parameters for the function.

Thanks anyway for your link, I guess is almost the only one talking about his issue :)"

[George Evers]

I don't have the NN toolbox by which to help with the debbugging. If no one else has the answer, you might just click "Please provide feedback to help us improve this Solution" at the link above and request the desired instructions.

[Francisco]

Thanks anyway George.

Till the moment I 'solved' using a global variable instead of passing the parameters in the right way -I was not able to realize how to do it.

Thanks again for your time.

Fran

Answer 2

[Greg Heath]

There is absolutely no reason to use parameters in the transfer function.

Training will vary the weights to minimize the error.

If you want to decrease training time , just increase the learning rate

Hope this helps.

Greg