Prev: Can someone help me writing this script file on matlab program?
Next: Accessing individual contents of cell array for large amount of cells
From: Greg Heath on 13 Apr 2010 19:43
On Apr 13, 7:32 am, "Fernando Henrique" <fernandoh...(a)gmail.com>
wrote:
> I am dividing it, as in:
>
> net.divideParam.trainRatio = 70/100;
> net.divideParam.valRatio = 15/100;
> net.divideParam.testRatio = 15/100;
>
> And yes, I'm testing with data outside the training set. The network performs correctly with SIM. I really think the problem is in the implementation "from scratch", but from the definition, I can't find what's wrong:
>
> inputsInput=[A;B];
>
> weightsInput=net.IW{1,1};
> weightsLayer=net.LW{2,1};
> biasesInput=net.B{1,1};
> biasesLayer=net.B{2,1};
>
> inputsLayer=tansig(weightsInput*inputsInput+biasesInput);
> Output=tansig(weightsLayer*inputsLayer+biasesLayer);
>
> Thank you,
> Fernando Henrique
>
>
>
> Greg Heath <he...(a)alumni.brown.edu> wrote in message <5d1c324d-e956-4be9-8e2e-280b93fbd...(a)r18g2000yqd.googlegroups.com>...
>
> > CORRECTED FOR THE HEINOUS SIN OF TOP-POSTING!
>
> > On Apr 12, 8:46 am, "Fernando Henrique" <fernandoh...(a)gmail.com>
> > wrote:
> > > Greg Heath <he...(a)alumni.brown.edu> wrote in message <610901f7-804b-4fea-aff7-236a8bc6b...(a)x7g2000vbc.googlegroups.com>...
> > > > On Apr 9, 4:18 pm, "Fernando Henrique" <fernandoh...(a)gmail.com> wrote:
> > > > > Hello,
>
> > > > > I have trained a neural network using 'nprtool', which is supposed to use "sigmoid hidden and output neurons". I am trying to use the results of this training (weights and biases) to implement a neural network from scratch. Unfortunately, I cannot get the right results. Could anyone point what I am doing wrong? The code follows below.
>
> > > > > Thanks in advance,
> > > > > Fernando H.
>
> > > > > %Neural Network training
> > > > > numHiddenNeurons = 40;
> > > > > net = newpr(P,T,numHiddenNeurons); %The network has two inputs and one output
>
> > > > sizeP = size(P)  %  ?
> > > > sizeT = size(T)   % ?
>
> > > > Does H = 40 make sense for the size of the data base?
> > > > Search
>
> > > > greg heath Neq Nw
>
> > > > > net.divideParam.trainRatio = 70/100;
> > > > > net.divideParam.valRatio = 15/100;
> > > > > net.divideParam.testRatio = 15/100;
> > > > > [net,tr] = train(net,P,T);
>
> > > > > %Neural Network implementation
> > > > > inputsInput=[A;B];
>
> > > > size(A)   % ?
> > > > size(B)   % ?
>
> > > > > weightsInput=net.IW{1,1};
> > > > > weightsLayer=net.LW{2,1};
> > > > > biasesInput=net.B{1,1};
> > > > > biasesLayer=net.B{2,1};
> > > > > inputsLayer=tansig(weightsInput*inputsInput+biasesInput);
> > > > > Output=tansig(weightsLayer*inputsLayer+biasesLayer);
>
> > > > Please explain EXACTLY what you mean by
>
> > > > "Unfortunately, I cannot get the right results."
>
> > > > Do you get the right results when you use SIM?
> >  > size(P) = 2       10000
> > > size(T) = 1       10000
> > > size(A) = 1     1
> > > size(B) = 1     1
>
> > > I'm using the neural network as a classifier, so by right results I mean results > that correspond to reality (when I use an input vector with known output).
>
> > Considering the size of your training set you are probably OK.
>
> > However, it is better to partition the data into training, validation
> > and test sets. So that you can select training parameters and
> > predict future performance by using nontraining data. In addition,
> > you can reduce training times by not using an excessive amount
> > of training data.
>
> > I think the number of neurons makes sense, as the network results
> > are OK when I use SIM.
>
> > On nontraining data?
>
> > > Thank you for your time,
> > > Fernando H
>
> > You are welcome.

Check algorithm defaults.

In particular, mapminmax and 'purelin'

Hope this helps.

Greg

First  |  Prev  | 
Pages: 1 2
Prev: Can someone help me writing this script file on matlab program?
Next: Accessing individual contents of cell array for large amount of cells