I am implementing forward neural network for prediction while taking weights from patternnet trained model

Dir = '.';
outputFile = fullfile(Dir, 'net_test1.mat');
load(outputFile, 'TrainedNet');
%%
ih1w = TrainedNet.IW{ 1, 1 };
h1h2w = TrainedNet.LW{ 2, 1 };
h2ow = TrainedNet.LW{ 3, 2 };
h1b = TrainedNet.b{1};
h2b = TrainedNet.b{2};
ob = TrainedNet.b{3};
%%
maxx = TrainedNet.inputs{1}.processSettings{1,1}.xmax;
minx = TrainedNet.inputs{1}.processSettings{1,1}.xmin;
gain = TrainedNet.inputs{1}.processSettings{1,1}.gain;
rangex = TrainedNet.inputs{1}.processSettings{1,1}.xrange;
offset = TrainedNet.inputs{1}.processSettings{1,1}.xoffset;
TrainedNet.inputs{1}.processSettings{1,1}
%%
function y = tanh(x)
y = (2 / (1 + exp(-2 * x))) - 1;
end
function y = sigmoid(x)
y = 1 / (1 + exp(-x));
end
inputlayer = ones(1,1036);
inputlayer = inputlayer';
inputlayer_normalized = [];
for x = 1:1036
    inputlayer_normalized(x) = (inputlayer(x)-offset(x))*gain(x);
end
% Initialize variables
h1size = size(ih1w, 1);
inputsize = size(ih1w, 2);
h2size = size(h1h2w, 1);
outputsize = size(h2ow, 1);
% First hidden layer computation
hl1 = zeros(1, h1size);
for k = 0:h1size-1
    sum = 0;
    for i = 0:inputsize-1
        sum = sum + (ih1w(k+1, i+1) * inputlayer_normalized(i+1));
    end
    sum = sum + h1b(k+1);
    hl1(k+1) = tanh(sum);
end
% Second hidden layer computation
hl2 = zeros(1, h2size);
for k = 0:h2size-1
    hl2(k+1) = 0;
    for i = 0:h1size-1
        hl2(k+1) = hl2(k+1) + (h1h2w(k+1, i+1) * hl1(i+1));
    end
    hl2(k+1) = hl2(k+1) + h2b(k+1);
    hl2(k+1) = tanh(hl2(k+1));
end
% Output layer computation
ol = zeros(1, outputsize);
for k = 0:outputsize-1
    ol(k+1) = 0;
    for i = 0:h2size-1
        ol(k+1) = ol(k+1) + (h2ow(k+1, i+1) * hl2(i+1));
    end
    ol(k+1) = ol(k+1) + ob(k+1);
    ol(k+1) = sigmoid(ol(k+1));
end
Ipred = TrainedNet(inputlayer);

this is code what i am implementing above neural network trained from inbuild function patternnet in matlab

I am using its weights and preprocess

but i am not getting same output in variable ol and Ipred

Answer 1

[Umar]

Hi Ajay,

My suggestion would be revising the computation of the output layer in the neural network. Here is the code snippet as an example for the output layer computation

% Output layer computation ol = zeros(1, outputsize); for k = 1:outputsize sum = 0; for i = 1:h2size sum = sum + (h2ow(k, i) * hl2(i)); end ol(k) = sum + ob(k); ol(k) = sigmoid(ol(k)); end

So, the loop indices start from 1 instead of 0 to align with MATLAB indexing. Then, sum calculation is corrected to accumulate the weighted inputs from the previous layer. Afterwards, the bias term is added after the weighted sum calculation. Finally, the output is passed through the sigmoid activation function to obtain the final output value.

Let me know if you need further assistance.

[Umar]

Sorry for unformatted code