I wrote a function to reproduce the neural network operation 'sim', why is my result inconsistent with sim's result?

load("file.mat")
net=results.Network;
xpoint=[1;2;3];
a1=sim(net,xpoint)

a1 = 2x1

5.0415 4.9795

<mw-icon class=""></mw-icon>

<mw-icon class=""></mw-icon>

a2=useNetwork(net,xpoint)

a2 = 2x1

2.4724 1.7471

<mw-icon class=""></mw-icon>

<mw-icon class=""></mw-icon>

function output=useNetwork(net,input)

l=tansig(net.IW{1}*input+net.b{1});
output=net.LW{2,1}*l+net.b{2};
end

[Jaimin]

Hi @策 陈,

Could you please provide the model architecture used to train these weights?

Answer 1

[Jaimin]

Hello @策 陈

In MATLAB, the neural network utilizes pre-processing and post-processing functions to handle the input and output data. You can adjust your parsing code as follows:

load("file.mat")
net=results.Network;
xpoint=[1;2;3];

a1=sim(net,xpoint)

a2=useNetwork(net,xpoint)

function output = useNetwork(net, input)
    net_iw=net.IW{1,1};
    net_lw=net.LW{2,1};
    net_b1=net.b{1};
    net_b2=net.b{2};
    
    normalized_inputn_test = mapminmax('apply', input, net.inputs{1}.processSettings{1}); %we are applying the same pre-processing as in the net
    hidden_layer_input =net_iw* normalized_inputn_test + repmat(net_b1, 1, size(input', 1)); % input layer to hidden layer
    hidden_layer_output=tansig(hidden_layer_input);
    output_layer_input = net_lw*hidden_layer_output+ repmat(net_b2, 1, size(hidden_layer_output', 1)); % hidden layer to output layer
    output = mapminmax('reverse', output_layer_input, net.outputs{2}.processSettings{1});%we are applying the same post-processing as in the net
end

For further information, please consult the following MATLAB answer, which addresses a similar issue:

https://www.mathworks.com/matlabcentral/answers/1974994-using-analytical-method-to-reproduce-the-bp-neural-network-but-the-result-is-far-from-the-nn-output?s_tid=srchtitle

I hope this will be helpful.