converting Matlab Neural Network into C++ Neural Network - c++

Question

I created a Neural Network in Matlab with newff, for handwritten Digits recognition.
I just trained it to recognize only 0 & 1 values from images.
with 3 Layers, Input Layer has 9 Neurons, Hidden Layer has 5 Neurons, and output Layer 1 Neuron,and there is 9 inputs.
my out puts are 0.1 & 0.2 ,and all Layers outputs function are "tansig".
I test it in Matlab and Network works Fine. now I want to create this network in c++ , I Wrote the Code and I copied all the Weights and Biasses(total 146 weights).
but when I put the same input data to Network the output value is not correct.
can anyone of you guys guide me?
this is my networks code:
here's my networks code...
public class Neuron
{
public Neuron()
{ }
public Neuron(int SumOfInputs)
{
m_SumOfInputs = SumOfInputs;
}
public double act(double[] Input, double[] weight, double bias)
{
double tmp = bias;
for (int i = 0; i < m_SumOfInputs; i++)
tmp += (Input[i] * weight[i]);
m_output = 1.0 / (1.0 + Math.Exp(-tmp));
return m_output;
}
public double m_output;
private int m_SumOfInputs;
};
public class Net
{
public Net()
{
int i;
//net1 , net2
//initializing inputLayer Neurons
for (i = 0; i < 9; i++)
InputLayer[i] = new Neuron(9);
//initializing HiddenLayer Neurons
for (i = 0; i < 5; i++)
HiddenLayer[i] = new Neuron(9);
//initializing OutputLayer
OutputLayer = new Neuron(5);
}
public double Calculate(double[] inputs)
{
double[] ILay_Outputs = new double[9];
double[] HLay_Outputs = new double[5];
//inputLayer acting
ILay_Outputs[0] = InputLayer[0].act(inputs, IW1, Ib[0]);
ILay_Outputs[1] = InputLayer[1].act(inputs, IW2, Ib[1]);
ILay_Outputs[2] = InputLayer[2].act(inputs, IW3, Ib[2]);
ILay_Outputs[3] = InputLayer[3].act(inputs, IW4, Ib[3]);
ILay_Outputs[4] = InputLayer[4].act(inputs, IW5, Ib[4]);
ILay_Outputs[5] = InputLayer[5].act(inputs, IW6, Ib[5]);
ILay_Outputs[6] = InputLayer[6].act(inputs, IW7, Ib[6]);
ILay_Outputs[7] = InputLayer[7].act(inputs, IW8, Ib[7]);
ILay_Outputs[8] = InputLayer[8].act(inputs, IW9, Ib[8]);
//HiddenLayer acting
HLay_Outputs[0] = HiddenLayer[0].act(ILay_Outputs, HW1, Hb[0]);
HLay_Outputs[1] = HiddenLayer[1].act(ILay_Outputs, HW2, Hb[1]);
HLay_Outputs[2] = HiddenLayer[2].act(ILay_Outputs, HW3, Hb[2]);
HLay_Outputs[3] = HiddenLayer[3].act(ILay_Outputs, HW4, Hb[3]);
HLay_Outputs[4] = HiddenLayer[4].act(ILay_Outputs, HW5, Hb[4]);
//OutputLayer acting
OutputLayer.act(HLay_Outputs, OW, Ob);
return OutputLayer.m_output;
}
//variables
Neuron[] InputLayer = new Neuron[9];
Neuron[] HiddenLayer = new Neuron[5];
Neuron OutputLayer;
//net2 tansig tansig tansig
double[] IW1 = { 0.726312035124743, 1.01034015912570, 0.507178716484559, -0.254689455765290, 0.475299816659036, 0.0336358919735363, -0.715890843015230, 0.466632424349648, 0.565406467159982 };
double[] IW2 = { 0.866482591050076, -0.672473224929341, 0.915599891389326, 0.310163265280920, -0.373812653648686, -0.0859927887021936, 0.0100063635393257, 0.816638798257382, -0.540771172965867 };
double[] IW3 = { 0.138868216294952, 1.93121321568871, -0.564704445249800, 0.834275586326333, 3.08348295981989, 0.899715248285303, -0.661916798988641, 6.00562393127300, 6.11939776912678 };
double[] IW4 = { 0.578089791487308, 0.885170493965113, -0.992514702569606, 0.415980526304333, -0.706140252063166, 0.442017877881589, -0.449053823645690, -0.0894051386719344, -0.348622179369911 };
double[] IW5 = { -0.407756482945129, 0.0786764402198765, 0.972408690276837, -0.959955597431701, -0.977769442966978, 1.52121267506016, 0.503296357838885, -3.31593633455649, -3.47834004737816 };
double[] IW6 = { -1.17474983226852, 0.870140308892922, 1.50545637070446, 0.369712493398677, -0.569857993006262, -0.732502911495791, -0.668984976457441, -1.48023312055586, -0.893472571240467 };
double[] IW7 = { -0.860518592120001, -1.48432158859269, 0.957060799463945, -0.680797771869510, -0.270752283410268, -0.218766920514208, 0.168091770241510, -2.50326075864844, -0.800988078966455 };
double[] IW8 = { 0.436492138260917, 0.280081066366966, 0.484813099857825, -0.310693876078844, 1.60359045377467, 1.57343220231689, -1.21552190886612, 2.03276547165735, 1.27245062411707 };
double[] IW9 = { 1.66853306274827, -1.59142022586958, 0.862315766588855, 0.676048095028997, -2.22623540036057, -1.48036066273542, -0.0386781503608105, -5.18214728910353, -5.21258509200432 };
double[] HW1 = { 0.577543862468449, 0.452264642610010, -0.869014797322399, 0.122435296258077, 0.507631314535324, 0.0386430216115630, -0.398222802253669, -0.614601040619812, 1.43324133164016 };
double[] HW2 = { 0.163344332215885, 0.434728230081814, -3.04877964757120, -0.118300732191499, -2.63220585865390, 0.443163977179405, -2.11883915836372, 2.07955461474729, -3.94441429060856 };
double[] HW3 = { -0.156103043064606, -0.482049683802527, 1.24788068138172, -1.05731056687422, -0.615321348655331, 0.214815967784408, 0.375762477817552, -0.728649292060764, -0.212151944122515 };
double[] HW4 = { 1.78276088127139, 1.15086535250306, 1.25967219208841, -0.446026243031773, -3.94742837475153, -1.33311929047378, -2.09356929069216, 0.0736879745054291, 1.51472991137144 };
double[] HW5 = { 0.744372844550077, 0.400815326319268, -4.94686055701529, 0.444773365537176, 2.65351865321717, 1.87143709824455, 1.74346707204902, -3.28220218001754, 5.78321274609173 };
double[] OW = { -1.09112204235009, -7.13508015318964, -1.02533926874837, 3.80439015418632, -4.16711367340349 };
double[] Ib = {-1.77988445077976,
-1.37323967952292,
-0.547465218997906,
0.331535304175263,
-0.0167810612906040,
0.734128501831859,
-0.543321122358485,
-1.13525462762255,
1.82870615182942};
double[] Hb = {1.68321697741393,
-0.862080862212137,
-0.536310792063381,
-0.772019935790668,
1.51470472867250};
double Ob = -0.156343477742835;
};
thanks.
Arta.

Answer 1

You mention in your description that you want to use the Tansig activation function, but in your code you have the implementation for the Logsig activation function. Tansig approximation would be:
2/(1+Math.Exp(-2*tmp))-1
I am also not sure how you get the weights for the input layer, are these perhaps the weights for the hidden layer. Matlab does not generate weights for the input layers since the inputs are directly connected to the hidden layer. Where net.IW are the weights for the first (hidden) layer, the weights for the subsequent layers (including output) are given by net.LW.
Besides the above I don't see obvious bugs/errors in your code, maybe try a simpler network first and train it to do the old and wise XOR relationship.
Lastly I would like to mention if you are writing this code for a micro-controller it's easier to do it in C and without objects. Your code will be smaller and faster. A step by step example is given here.

Answer 2

I found the problem guys.
in matlab, before inputs goes to the network, they all goes to a function names (applyminmax) in a .m file names (mapminmax.m), and then this function outputs are the network inputs.
after the simulation on network is done, the outputs goes to a function names (reverse) in the same .m file.
and this function outputs is the final output of the Neural Network.
thanks for all your helps.
Arta.