Error using +. Matrix dimensions must agree

Hi, I'm trying to create neural network for train my data. But I get error. I though its right but which one should I fix. Here my code

%clc;
close all;
clearvars;

% Multilayer Perceptron (MLP) neural networks.
N = 220; %sum of data
D = 25 ; %dimension
K = 11; %target
h = 500; %sum of neuron
load datatrain1.mat 
Train1= cell2mat(Train);

load classtrain1.mat
CTrain1 = cell2mat(TTrain);

b1 = randn(1,h);%bias from input to hidden
W1 = randn(D,h);%weight from input to hidden
b2 = randn(1,K);%bias from hidden to output
W2 = randn(h,K);%weight from hidden to output

%feedfoward
for efoch = 1 : 1000
    H = Train1*W1+b1;
    Y = H*W2+b2;

end

I get error in variable H, like this Error using +. Matrix dimensions must agree.

Any help for fix my code will must appreciated. Thanks

[madhan ravi]

upload .mat files

[Oman Wisni]

yes sir, I already attached my file .mat. But I'm not sure it can open in matlab, cause before I ever upload and download, when I download the format of file is different. And when I upload I no get reply.

Answer 1

[Guillaume]

The error message is very clear, you're trying to add two matrices of different size, which is never possible. Since b1 is a 1xh vector, Train1*W1 must be a 1xh vector, and since W1 is a Dxh matrix, Train1 must be a 1xD vector for the product to be 1xh.

Therefore, your train1 is not 1xD. There is nothing we can do about that. Either use a train1 matrix that is 1xD or change your D to reflect the size of train1.

Note that it is never a good idea to hardcode the size of matrices. It is always safer to ask matlab for the actually size:

D = size(train1, 2);

would ensure it always match train1. Of course, train1 must only have one row.

edit: got confused between D and h but see comment below.

[Oman Wisni]

Yes sir, if I change h or train1. It is impossible cause it value from feature extraction and h is the number of single hidden layer with 500 neuron. b is bias. I get in theory the formula like that.

I confused too. Is there I wrong declarad bias and weight? But the example I get from this forum

[Stephen23]

"But the example I get from this forum"

Please give a link to the thread where you got this from.

[Oman Wisni]

https://www.mathworks.com/matlabcentral/answers/50961-how-to-initialize-the-neural-network-to-a-set-of-weights

This sir, I follow this example for declarad bias and weight

[Guillaume]

Actually, my reasoning above was a bit wrong, at least for versions >R2016b. Since the implicit expansion introduced in R2016b, the sum with work as long as the product train1*W1 has h columns, the number of rows doesn't matter. Since the product is guaranteed to have h columns or fail, the only reason why the sum would fail for you is that you're using a version earlier than R2016b.

In version prior to R2016b you had to use bsxfun for explicit expansion, so:

H = bsxfum(@sum, train1*W1, b1);

Most likely, you'll have to do the same for the next line.

[Oman Wisni]

Yes, I using R2015a.

What mean bsxfum? And if I run it, the result will same?

[Guillaume]

documentation of bsxfun. In this particular case, it means that the input b1 will be replicated to match the height of train1*W1. Yes, the result will be the same as you would have got in later versions with implicit expansion.

[Oman Wisni]

Oke sir, I will trying. Thank you

[Oman Wisni]

Sir, what code is right ?

 H = bsxfun(@sum, train1*W1, b1);

or

   H = bsxfun(@plus, Train1*W1, b1);

using @sum I get error, but when I using @plus I get result. My question is, if I using @plus there is nothing change in the result? thanks

[Guillaume]

Sorry, it should have been @plus indeed.

[Oman Wisni]

Yes sir. Thanks