Matlab r2011b neural network toolbox precautions

This is a record of the pit I encountered when using the neural network toolbox , For yourself and others . Write a little first , Update later .

1.

net = feedforwardnet;
net = train(net, attributes, targets);

The first line creates a two-layer feedforward network , The number of hidden layer neurons is the default 10, It's no problem . After creating the network , If you use view(net) To view the network topology , You will find that there are no input vectors and output vectors , This is because... Has not been called yet configure function .configure By default, the function is called for the first time train Function is automatically called . There's a pit here . hypothesis ：

X = [
  1 1 2;
  2 1 3;
  3 1 1;
  2 1 3]';
Y = [
  0 1 1 0];

That is, the input vector is 3 Dimension vector , Data sets X Contained in the 4 Samples , Training adopts batch training . after train After function call ,net.IW{1,1} The dimension of will become 10x2！ Should not be 10x3 Do you （ notes ： The number of neurons in the hidden layer 10, Input vector 3 dimension ）？ Because the dataset X The second attribute of all samples in is the same （ Values are 1）, As a result, this attribute is Matlab Ignore it. , I don't know whether it was intentional or bug. resolvent

X(:,find(var(X,0,1) < eps)) = X(:,find(var(X,0,1))) + min(min(X))*1e-5*randn(size(X,1),length(find(var(X,0,1))));

namely , Add a small white noise to the ignored columns to make their values different .