Matrix size of layer weights in neural network(Error:net.LW{2,1} must be a 0-by-3 matrix.)

Question

Heather Zhang 2016-8-30

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/301387-matrix-size-of-layer-weights-in-neural-network-error-net-lw-2-1-must-be-a-0-by-3-matrix

回答： Greg Heath 2016-8-30

Hello, I had a Elman network with 3 hidden layers, 3 input layers and 1 output layer. I try to initial the hidden layer weights with a decent trained result which is saved before. But when I re-input the hidden layer weights into the network. Errors always bumped out. Error:net.LW{2,1} must be a 0-by-3 matrix. The error also happened when I valued the bias and input layer weights. I have two questions: (1) What does "0-by-3 matrix" mean? Does it mean the matrix should be a empty matrix? (2) I got the hidden layer weights, input layer weights and bias from the trainning result. Why it doesn't work when I re-input them into the network? They were saved in ".mat" format. They were all "cell" structure.

    *%Code for saving the trainning result
    iw = net.iw;  % "iw" is variable of input layer weight
    lw = net.lw;  % "lw" is variable of hidden layer weight
    theta = net.b;  % "theta" is variable of bias
    *%%Code for re-input the result
    net.b{1} = theta{1,1}; 
    net.lw{2,1} = lw{2,1};
    net.iw{1,1} = iw{1,1};
    net.b{2} = theta{2,1};

>> net.lw{2,1}

ans =

Empty matrix: 0-by-3

>> lw{2,1} % saved trainning result of hidden layer weight (variable "lw")

ans =

-1.3208 -1.4288 -0.7487

>> net.iw{1,1}

ans =

Empty matrix: 3-by-0

>> iw{1,1} % saved trainning result of input layer weight (variable "iw")

ans =

    0.7936    0.7638    0.6892
   -0.0306   -0.3250   -0.0438
    0.8408   -1.0601   -0.9286
*
    Error: network/subsasgn>network_subsasgn (line 553)
    net.LW{2,1} must be a 0-by-3 matrix.
    Error: network/subsasgn (line 13)
    net = network_subsasgn(net,subscripts,v,netname);
    Error: elman_PepperHTHS_v2 (line 65)
    net.lw{2,1} = lw{2,1};
*