Weights don't initialize.

2 次查看（过去 30 天）
Noisy
Noisy 2011-10-29
I created the following network:
P = dataH;
T = dataXsm;
net=network;
net.numInputs = 1;
net.numLayers = 3;
net.biasConnect(1) = 1;
net.biasConnect(2) = 1;
net.biasConnect(3) = 1;
net.inputConnect = [1; 0; 0];
net.layerConnect = [0 0 0; 1 0 0; 0 1 0];
net.outputConnect = [0 0 1];
net.inputs{1}.size = 2;
net.layers{1}.size = 2;
net.layers{1}.transferFcn = 'hardlim';
net.layers{1}.initFcn = 'initnw';
net.layers{2}.size = 10;
net.layers{2}.transferFcn = 'hardlim';
net.layers{2}.initFcn = 'initnw';
net.layers{3}.size = 10;
net.layers{3}.initFcn = 'initnw';
net.layers{3}.transferFcn = 'hardlim';
net.initFcn = 'initlay';
net.IW{1,1}, net.IW{2,1},
net.LW{3,2}
net.b{1}, net.b{3}
net.trainFcn = 'trainc';
net.performFcn = 'sse';
net.adaptFcn = 'trains';
net.trainParam.goal=0.01;
net.trainParam.epochs=100;
net.trainParam.passes = 1;
net = init(net);
a = sim(net,P), e = T-a
net=train(net,P,T);
net.adaptParam.passes = 100;
[net,a,e] = adapt(net,P,T); e
twts = net.IW, tbiase = net.b
but it doesn't work, weights don't initialize and it gives all 1 as result: twts =
[2x2 double]
[]
[]
a =
1 1 1...1
...
1 1 1...1
Is something wrong with layer connection? Or do I initialize something wrong?

请先登录，再回答此问题。

采纳的回答

Vito
Vito 2011-10-30
No.
Multilayer percetron doesn't contain 'hardlim'(hardlim -is capable to classify only linearly separable set. Two or more layers in network - aren't separable linearly. ). Using 'logsig'.
The equivalent network - multilayer percetron.
P =[0 1 0 1; 0 0 1 1];
T = [0 0 0 1];
net=newff(minmax(P),[2,10,1],{'logsig','logsig','logsig'},'trainbfg');
net.trainParam.epochs = 100;
net = init (net);
net.IW{1,1}, net.IW{2,1},
net.LW{3,2}
net.b{1}, net.b{3}
net=train(net,P,T);
a = sim(net,P)
'trainbfg' – back propagation learning.
Error in network design.
  1 个评论
Greg Heath
Greg Heath 2011-10-31
Typically, only one hidden layer is needed.
Use as many defaults as possible (help newff).
newff automatically initializes weights with initnw
.
if [I N] = size(p) and [O N] = size(t) then
there are Neq = N*O training equations and
Nw = (I+1)*H+(H+1)*O unknown weights. For
accurate weight estimation it is desired that
Neq >> Nw
Typically Neq >= 10*Nw is adequate. However,
sometimes a larger ratio (e.g., > 30) is needed
and sometimes a smaller ratio (e.g., 2) will suffice.
Hope this helps.
Greg

请先登录，再进行评论。

更多回答（0 个）

类别

在 Help Center 和 File Exchange 中查找有关 Define Shallow Neural Network Architectures 的更多信息

标签

产品

Community Treasure Hunt

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

Start Hunting!

Translated by