bias problem with trainb

2 views (last 30 days)
Emiliano Rosso
Emiliano Rosso on 1 Jul 2016
Commented: Emiliano Rosso on 2 Jul 2016
I need to train a supervised neural network to linearly separate a set of inputs with logical output 0,1 in batch training. My inputs is inputstemp2 (100,2000) of double precision & target is dividetargets(2,2000) of double 0 & 1. My code is :
mynet=perceptron;
mynet.trainFcn= 'trainb';
mynet.inputWeights.learnFcn='learnp'; %default
mynet.biases.learnFcn= 'learnp'; %default
mynet.biasConnect =[1];
mynet.outputs{1}.processFcns = {};
mynet.inputs{1}.processFcns = {'mapstd','processpca'};
mynet.divideFcn = 'divideblock';
mynet.divideParam.trainRatio = 80/100;
mynet.divideParam.valRatio = 10/100;
mynet.divideParam.testRatio =10/100;
mynet.trainParam.showWindow = false;
mynet.trainParam.showCommandLine = false;
mynet.trainParam.epochs=500;
mynet.efficiency.memoryReduction=1;
[mynet,tr]=train(mynet,inputstemp2,dividetargets);
No problem if mynet.biasConnect =[0] but if I set mynet.biasConnect =[1] Matlab return:
Error using *
Inner matrix dimensions must agree.
Error in learnp>apply (line 93)
dw = e*p';
Error in trainb>train_network (line 231)
[db,BLS{i}] = learnFcn.apply(net.b{i}, ...
Error in trainb (line 55)
[out1,out2] = train_network(varargin{2:end});
ecc.... The same if I use linearlayer and harlim. Somebody can explain me why bias is not accepted?

Accepted Answer

Greg Heath
Greg Heath on 1 Jul 2016
Just use the current (i.e., non-obsolete) classifier
net = patternnet([]); % [] => No hidden layer for linear model
For documentation
help patternnet
doc patternnet
Hope this helps.
Thank you for formally accepting my answer
Greg

More Answers (0)

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!