void main()
{
NET Net;
BOOL Stop;
REAL MinTestError;
InitializeRandoms();
GenerateNetwork(&Net);
RandomWeights(&Net);
InitializeApplication(&Net);
Stop = FALSE;
MinTestError = MAX_REAL;
do {
TrainNet(&Net, 10);
TestNet(&Net);
if (TestError < MinTestError) {
fprintf(f, " - saving Weights ...");
MinTestError = TestError;
SaveWeights(&Net);
}
else if (TestError > 1.2 * MinTestError) {
fprintf(f, " - stopping Training and restoring Weights ...");
Stop = TRUE;
RestoreWeights(&Net);
}
} while (NOT Stop);
TestNet(&Net);
EvaluateNet(&Net);
FinalizeApplication(&Net);
}
void MultiLayerPerceptron::Run(const char* fname, const int& maxiter)
{
int countTrain = 0;
int countLines = 0;
bool Stop = false;
bool firstIter = true;
double dMinTestError = 0.0;
InitializeRandoms();
RandomWeights();
do {
countLines += Train(fname);
Test(fname);
countTrain++;
if(firstIter)
{
dMinTestError = dAvgTestError;
firstIter = false;
}
printf( "%i \t TestError: %f", countTrain, dAvgTestError);
if ( dAvgTestError < dMinTestError)
{
printf(" -> saving weights\n");
dMinTestError = dAvgTestError;
SaveWeights();
}
else if (dAvgTestError > 1.2 * dMinTestError)
{
printf(" -> stopping training and restoring weights\n");
Stop = true;
RestoreWeights();
}
else
{
printf(" -> ok\n");
}
} while ( (!Stop) && (countTrain<maxiter) );
printf("bye\n");
}
// Write C code for corresponding to the feed forward network into a given file.
void MultipleBackPropagation::GenerateCCode(OutputFile & f, VariablesData & trainVariables, BOOL inputLayerIsConnectedWithOutputLayer, BOOL spaceInputLayerIsConnectedWithOutputLayer) {
CString s;
CString MBPVersion;
MBPVersion.LoadString(IDS_VERSION);
f.WriteLine("/**");
f.WriteLine(_TEXT(" Generated by ") + MBPVersion);
f.WriteLine(" Multiple Back-Propagation can be freely obtained at http://dit.ipg.pt/MBP");
f.WriteLine("*/\n");
f.WriteLine("#include <math.h>");
f.WriteLine("/**");
s.Format(_TEXT(" inputs - should be an array of %d element(s), containing the network input(s)."), inputs);
bool hasMissingValues = false;
for(int i = 0; i < inputs; i++) {
if (trainVariables.HasMissingValues(i)) {
s += " Inputs with NaN value are considered missing values.";
hasMissingValues = true;
break;
}
}
f.WriteLine(s);
s.Format(_TEXT(" outputs - should be an array of %d element(s), that will contain the network output(s)."), outputs);
f.WriteLine(s);
s = " Note : The array inputs will also be changed.";
if (!hasMissingValues) s += "Its values will be rescaled between -1 and 1.";
s += "\n*/";
f.WriteLine(s);
s = f.GetFileName();
int p = s.Find('.');
if (p != -1) s = s.Left(p);
f.WriteLine(_TEXT("void ") + s + _TEXT("(double * inputs, double * outputs) {"));
f.WriteString("\tdouble mainWeights[] = {");
SaveWeights(f, ", ");
f.WriteLine("};");
f.WriteLine("\tdouble * mw = mainWeights;");
if (hasMissingValues) f.WriteLine("\tdouble b;");
if (!spaceNetwork.IsNull()) {
f.WriteString("\tdouble spaceWeights[] = {");
spaceNetwork->SaveWeights(f, ", ");
f.WriteLine("};");
f.WriteLine("\tdouble * sw = spaceWeights;");
s.Format(_TEXT("\tdouble mk[%d];"), spaceNetwork->Outputs());
f.WriteLine(s);
f.WriteLine("\tdouble *m = mk;");
if (hasMissingValues) {
s.Format(L"\tdouble spaceInputs[%d];", inputs);
f.WriteLine(s);
}
}
int numberLayers = layers.Lenght();
for (int l = 1; l < numberLayers - 1; l++) {
s.Format(L"\tdouble hiddenLayer%doutputs[%d];", l, layers.Element(l)->neurons.Lenght());
f.WriteLine(s);
}
int numberSpaceLayers = (spaceNetwork.IsNull()) ? 0 : spaceNetwork->layers.Lenght();
for (int l = 1; l < numberSpaceLayers - 1; l++) {
s.Format(_TEXT("\tdouble spaceHiddenLayer%doutputs[%d];"), l, spaceNetwork->layers.Element(l)->neurons.Lenght());
f.WriteLine(s);
}
f.WriteLine("\tint c;");
f.WriteString("\n");
// input variables will be rescaled between -1 and 1
if (trainVariables.Number() == inputs + outputs) {
for (int i = 0; i < inputs; i++) {
double min = trainVariables.Minimum(i);
double max = trainVariables.Maximum(i);
if (trainVariables.HasMissingValues(i)) {
s.Format(L"\tif(inputs[%d] == inputs[%d]) { /* compiler must have support for NaN numbers */\n\t", i, i);
f.WriteString(s);
}
if (min != max) {
s.Format(L"\tinputs[%d] = -1.0 + (inputs[%d] - %1.15f) / %1.15f;", i, i, min, (max - min)/2);
} else {
s.Format(L"\tinputs[%d] = inputs[%d] / %1.15f; /* WARNING: During the training this variable remain always constant */", i, i, max);
}
f.WriteLine(s);
if (hasMissingValues && !spaceNetwork.IsNull()) {
if (trainVariables.HasMissingValues(i)) f.WriteString("\t");
s.Format(L"\tspaceInputs[%d] = inputs[%d];", i, i);
f.WriteLine(s);
}
if (trainVariables.HasMissingValues(i)) {
if (!spaceNetwork.IsNull()) {
f.WriteLine("\t\tb = *sw++;");
s.Format(L"\t\tspaceInputs[%d] = tanh(b + spaceInputs[%d] * *sw++);", i, i);
f.WriteLine(s);
}
f.WriteLine("\t\tb = *mw++;");
s.Format(L"\t\tinputs[%d] = tanh(b + inputs[%d] * *mw++);", i, i);
f.WriteLine(s);
f.WriteLine("\t} else {");
s.Format(L"\t\tspaceInputs[%d] = inputs[%d] = 0.0;", i, i);
f.WriteLine(s);
f.WriteLine("\t}");
}
}
}
// space network
for (int l = 1; l < numberSpaceLayers; l++) {
List<Neuron> * neurons = &(spaceNetwork->layers.Element(l)->neurons);
int nn = 0;
for(NeuronWithInputConnections * n = dynamic_cast<NeuronWithInputConnections *>(neurons->First()); n != NULL; n = dynamic_cast<NeuronWithInputConnections *>(neurons->Next())) {
CString aux;
if (l == numberSpaceLayers -1) {
aux.Format(_TEXT("mk[%d]"), nn);
} else {
aux.Format(_TEXT("spaceHiddenLayer%doutputs[%d]"), l, nn);
}
f.WriteString("\t");
f.WriteString(aux);
f.WriteLine(" = *sw++;");
int numberInputsFromLastLayer = n->inputs.Lenght() - 1; // - bias
if (spaceInputLayerIsConnectedWithOutputLayer && l == numberSpaceLayers -1) numberInputsFromLastLayer -= inputs;
s.Format(_TEXT("\tfor(c = 0; c < %d; c++) "), numberInputsFromLastLayer);
f.WriteString(s);
f.WriteString(aux);
f.WriteString(" += *sw++ * ");
if (l == 1) {
s = (hasMissingValues) ? "spaceI" : "i";
s+= "nputs[c];";
} else {
s.Format(_TEXT("spaceHiddenLayer%doutputs[c];"), l-1);
}
f.WriteLine(s);
if (spaceInputLayerIsConnectedWithOutputLayer && l == numberSpaceLayers -1) {
s.Format(_TEXT("\tfor(c = 0; c < %d; c++) "), inputs);
f.WriteString(s);
f.WriteLine(aux + L" += *sw++ * " + ((hasMissingValues) ? "spaceI" : "i") + "nputs[c];");
}
WriteActivationFunctionCCode(f, n, CT2CA(aux));
nn++;
}
}
// main network
for (int l = 1; l < numberLayers; l++) {
List<Neuron> * neurons = &(layers.Element(l)->neurons);
int nn = 0;
for(NeuronWithInputConnections * n = dynamic_cast<NeuronWithInputConnections *>(neurons->First()); n != NULL; n = dynamic_cast<NeuronWithInputConnections *>(neurons->Next())) {
CString aux;
if (l == numberLayers -1) {
aux.Format(_TEXT("outputs[%d]"), nn);
} else {
aux.Format(_TEXT("hiddenLayer%doutputs[%d]"), l, nn);
}
f.WriteString("\t");
f.WriteString(aux);
f.WriteLine(" = *mw++;");
int numberInputsFromLastLayer = n->inputs.Lenght() - 1; // - bias
if (inputLayerIsConnectedWithOutputLayer && l == numberLayers -1) numberInputsFromLastLayer -= inputs;
s.Format(_TEXT("\tfor(c = 0; c < %d; c++) "), numberInputsFromLastLayer);
f.WriteString(s);
f.WriteString(aux);
f.WriteString(" += *mw++ * ");
if (l == 1) {
s = "inputs[c];";
} else {
s.Format(_TEXT("hiddenLayer%doutputs[c];"), l-1);
}
f.WriteLine(s);
if (inputLayerIsConnectedWithOutputLayer && l == numberLayers -1) {
s.Format(_TEXT("\tfor(c = 0; c < %d; c++) "), inputs);
f.WriteString(s);
f.WriteLine(aux + _TEXT(" += *mw++ * inputs[c];"));
}
WriteActivationFunctionCCode(f, n, CT2CA(aux));
if (!spaceNetwork.IsNull() && nn < neuronsWithSelectiveActivation[l]) {
f.WriteString("\t");
f.WriteString(aux);
f.WriteLine(" *= *m++;");
}
nn++;
}
}
// Rescale the outputs
if (trainVariables.Number() == inputs + outputs) {
for (int o = 0; o < outputs; o++) {
int outVar = o + inputs;
double min = trainVariables.Minimum(outVar);
double max = trainVariables.Maximum(outVar);
double nmin = trainVariables.newMinimum[outVar];
s.Format(_TEXT("\toutputs[%d] = %1.15f + (outputs[%d] - %f) * %1.15f;"), o, min, o, nmin, (max - min) / (1.0 - nmin));
f.WriteLine(s);
}
}
f.WriteLine("}");
}
