MATLAB: Weight and bias values i get from trained neural network is mistake

neural networkweight and bias

Hi, i try to get the weight and bias from the neural network that i train.
ActFunc_Hid and ActFunc_Out depends on any activation function i set on hidden and outer layer.
However, when i trained the neural network, i test it again by output = sim(net, a), where a is new input. the shown output is not the same as what i calculate manually based on the weight and bias value got from the trained neural network.
Is there any mistake in my code?
     
       net=newff(input, output, 1,{ActFunc_hid ActFunc_Out});
       [net tr]=train(net,input,output);
       w1 = net.IW{1}
       w2 = net.LW{2}
       b1 = net.b{1} 
       b2 = net.b{2}

Best Answer

It is not the same because Matlab does a normalization procedure of the variables so if you want to do it manually it would be like this:
xoffset = net.inputs {1} .processSettings {1} .xoffset;
gain = net.inputs {1} .processSettings {1} .gain;
ymin = net.inputs {1} .processSettings {1} .ymin;
y1 = bsxfun (@ times, bsxfun (@ minus, x, xoffset), gain);
y1 = bsxfun (@ plus, y1, ymin);
y1 is the true entrance to the neural network
Related Question