void DynamicsPlotterUtil::transferNeuronActivationToOutput(const QList<Neuron*> neurons) {
//Set the output of every neuron to the transferred activation
for(QListIterator<Neuron*> i(neurons); i.hasNext();) {
Neuron *neuron = i.next();
neuron->getOutputActivationValue().set(
neuron->getTransferFunction()->transferActivation(
neuron->getActivationValue().get(), neuron));
}
}
void Neuron::makeSynapses(void){
Neuron* n = new Neuron(4345);
cout << "index of postSyn is: " << n->getIndex() << endl;
Synapse syn1(1, n);
axons.push_back(syn1);
Synapse syn2(2, n);
axons.push_back(syn2);
Synapse syn3(3, n);
axons.push_back(syn3);
return;
}
Value NeuralNet::neuron_output(const Neuron& n, const ValueListPtr v) const
{
Neuron::const_iterator it;
ValueList::const_iterator vit;
Value retval = 0.0f;
for(it = n.begin(), vit = v->begin(); it != n.end() && vit != v->end(); it++, vit++)
{
retval += (*it) * (*vit);
}
// std::cout << retval << std::endl;
return retval;
}
// Back propagate the errors to update the weights.
void NeuralNet::backPropagate(vector<double>* outputs, int teacher) {
Layer *outputLayer = (*layers)[numHiddenLayers + 1];
for (int i = 0; i < outputs->size(); i++) {
Neuron *n = outputLayer->getNeuron(i);
double adjusted = -n->getValue();
if (i == teacher) {
adjusted += 1;
}
n->setDelta(sigmoidPrime(n->getActivation()) * adjusted);
}
// Propagate deltas backward from output layer to input layer.
for (int l = numHiddenLayers; l >= 0; l--) {
Layer *curr = (*layers)[l], *downstream = (*layers)[l+1];
for (int i = 0; i < curr->neuronCount(); i++) {
double sum = 0;
Neuron *n = curr->getNeuron(i);
for (int j = 0; j < downstream->neuronCount(); j++) {
sum += downstream->getNeuron(j)->getWeight(i)
* downstream->getNeuron(j)->getDelta();
}
n->setDelta(sigmoidPrime(n->getActivation()) * sum);
for (int j = 0; j < downstream->neuronCount(); j++) {
downstream->getNeuron(j)->updateWeight(i,
learningRate * sigmoid(n->getActivation())
* downstream->getNeuron(j)->getDelta());
}
}
}
}
/**
* Returns the strength of the owner synapse.
*
* @param owner the owner of this SynapseFunction.
* @return the strength of the owner.
*/
double ExampleSynapseFunction::calculate(Synapse *owner) {
if(owner == 0) {
return 0.0;
}
//Search for target neuron (if existing)
Neuron *target = dynamic_cast<Neuron*>(owner->getTarget());
if(target == 0) {
return owner->getStrengthValue().get();
}
//Update history
mHistory.enqueue(target->getLastActivation());
while(!mHistory.empty() && mHistory.size() > mHistorySize->get()) {
mHistory.dequeue();
}
//Calcualte average
double activationSum = 0.0;
double average = 0.0;
for(QListIterator<double> i(mHistory); i.hasNext();) {
activationSum += i.next();
}
if(mHistory.size() > 0) {
average = (activationSum / ((double) mHistory.size()));
}
else {
average = 0.0;
}
mCurrentAverage->set(average);
//Adapt synapse if required.
Neuron *source = owner->getSource();
if(source != 0) {
double output = source->getLastOutputActivation();
if(output != 0) {
double change = 0.0;
if(average > mDesiredActivationRange->getMax()) {
change = (average - mDesiredActivationRange->getMax()) * output * -1.0;
}
else if(average < mDesiredActivationRange->getMin()) {
change = (mDesiredActivationRange->getMin() - average) * output;
}
owner->getStrengthValue().set(owner->getStrengthValue().get() + (change * mChangeRate->get()));
}
}
return owner->getStrengthValue().get();
}
void testCopy()
{
#ifdef USE_MATRIX
RNN_MATRIX(rnnOrig);
RNN_VECTOR(rnnCopy);
#endif
#ifdef USE_VECTOR
RNN_MATRIX(rnnOrig);
RNN_VECTOR(rnnCopy);
#endif
for(int i = 0; i < 1000; i++)
{
Neuron *n = rnnOrig->createNeuron();
if(drand48() < 0.5)
{
n->setTransferfunction(transferfunction_tanh);
}
else
{
n->setTransferfunction(transferfunction_id);
}
}
for(int j = 0; j < 999; j++) // connect every neuron to neuron 0
{
__REAL w = ((drand48() < 0.5)?-1:1) * (10 * drand48() + 0.1);
rnnOrig->createSynapse(rnnOrig->getNeuron(j), rnnOrig->getNeuron(j+1), w);
}
for(int j = 0; j < 10000; j++) // random synapses
{
__REAL w = ((drand48() < 0.5)?-1:1) * (10 * drand48() + 0.1);
int source = (int)(drand48() * 100 + 0.5);
int destination = (int)(drand48() * 100 + 0.5);
// TODO: check why this is required for the test not to fail
while(source != destination - 1)
{
source = (int)(drand48() * 100 + 0.5);
destination = (int)(drand48() * 100 + 0.5);
}
rnnOrig->createSynapse(rnnOrig->getNeuron(source),
rnnOrig->getNeuron(destination),
w);
}
startTiming();
rnnCopy->copy(rnnOrig);
unsigned long time = stopTiming();
cout << "RecurrentNeuralNetwork copy:\t\t\t\t\t";
printTime(time);
delete rnnOrig;
delete rnnCopy;
}
void RungeKutta::calcMeanMembranePotential(Neuron& n_old, Neuron& n_new, double time, double timestep) {
if (n_old.getType() == BRAINSTEM) return;
double m = n_old.getM();
double tau = n_old.getWeights()[0];
double sigma = addWeightedNeighbors(n_old);
double k1 = calcDerivative(time, m, sigma, tau);
double k2 = calcDerivative(time + .5*timestep, m + (timestep/2)*k1, sigma, tau);
double k3 = calcDerivative(time + .5*timestep, m + (timestep/2)*k2, sigma, tau);
double k4 = calcDerivative(time + timestep, m + timestep*k3, sigma, tau);
double newVal = m + (1.0/6)*timestep*(k1 + 2*k2 + 2*k3 + k4);
n_new.setM(newVal);
}
int main() {
ScopedPass pass("Evolver [single-bit adder]");
Network network;
Evolver evolver(network);
Neuron * input0 = new OrNeuron();
Neuron * input1 = new OrNeuron();
Neuron * output = new OrNeuron();
network.AddNeuron(*input0);
network.AddNeuron(*input1);
network.AddNeuron(*output);
while (true) {
bool flag0 = RandomBool();
bool flag1 = RandomBool();
if (flag0) input0->Fire();
else input0->Inhibit();
if (flag1) input1->Fire();
else input1->Inhibit();
bool gotResponse = false;
bool expectingResponse = (flag0 && !flag1) || (flag1 && !flag0);
// Give it ten cycles to respond with the answer
for (int i = 0; i < 10; ++i) {
network.Cycle();
if (output->IsFiring()) {
gotResponse = true;
}
evolver.Prune();
evolver.Grow();
}
if (gotResponse != expectingResponse) {
std::cout << "wrong!" << std::endl;
network.UpdateLivesWithPain(0.5);
} else {
std::cout << "right!" << std::endl;
}
// Give the network some time to flush out its activity
for (int i = 0; i < 5; ++i) {
network.Cycle();
evolver.Prune();
evolver.Grow();
}
}
return 0;
}
bool NeuralNetwork::AreNeuronsConnected(const Neuron& lhs,const Neuron & rhs) const {
for (auto& connection : rhs.GetConnections()) {
if (!connection.outGoing && &lhs == connection.neuron) {
return true;
}
}
return false;
}
float RecSOM::train_one_sample(vector<float> sample, int nf) {
Neuron* w = NULL;
float qerr = 0.0;
input_layer = &sample[0];
w = activate_find_winner(output_layer);
qerr = w->recursive_distance(alpha, beta, input_layer, context_layer);
for(int i = 0; i < map_dim_x*map_dim_y; i++) {
map_layer[i]->update_weights(gama, w, input_layer, context_layer, sigma, nf);
}
update_context(output_layer);
return qerr;
}
void SelfOrganizingMaps::evaluateIndependentVector(vector<double> inputVector){
Neuron *bmu = getBMU(inputVector);
_matrix->getNeuron(bmu->getX(), bmu->getY())->setNeuronColor(0,0,0);
/*
cout << "A _bmuTestCases size = " << _bmuTestCases.size() << endl;
_bmuTestCases.insert ( pair<Neuron *,Neuron *>(getBMU(inputVector),getBMU(inputVector)));
cout << "B _bmuTestCases size = " << _bmuTestCases.size() << endl;
cout << "Input Vector:" << endl;
cout << inputVector[0] << " " << inputVector[1] << " " << inputVector[2] << endl;
cout << "BMU:" << endl;
bmu->info();
double distanceToBMU = bmu->distanceToInputVector(inputVector);
cout << "BMU distance to input vector: " << distanceToBMU << endl;
int bmuX = bmu->getX();
int bmuY = bmu->getY();
if(((bmuX + 1 < _size) && (bmuX - 1 >= 0)) && ((bmuY + 1 < _size) && (bmuY - 1 >= 0))){
Neuron *upLeft = _matrix->getNeuron(bmuX - 1, bmuY - 1);
Neuron *up = _matrix->getNeuron(bmuX, bmuY - 1);
Neuron *upRight = _matrix->getNeuron(bmuX + 1, bmuY - 1);
Neuron *left = _matrix->getNeuron(bmuX - 1, bmuY);
Neuron *right = _matrix->getNeuron(bmuX + 1, bmuY);
Neuron *downLeft = _matrix->getNeuron(bmuX - 1, bmuY + 1);
Neuron *down = _matrix->getNeuron(bmuX, bmuY + 1);
Neuron *downRight = _matrix->getNeuron(bmuX + 1, bmuY + 1);
double distUpLeft = upLeft->distanceToInputVector(inputVector);
double distUp = up->distanceToInputVector(inputVector);
double distUpRight = upRight->distanceToInputVector(inputVector);
double distLeft = left->distanceToInputVector(inputVector);
double distRight = right->distanceToInputVector(inputVector);
double distDownLeft = downLeft->distanceToInputVector(inputVector);
double distDown = down->distanceToInputVector(inputVector);
double distDownRight = downRight->distanceToInputVector(inputVector);
cout << "Distances to side Neurons of the BMU" << endl;
cout << distUpLeft << " " << distUp << " " << distUpRight << endl;
cout << distLeft << "--" << distanceToBMU << "-- " << distRight << endl;
cout << distDownLeft << " " << distDown << " " << distDownRight << endl;
}else{
cout << "The BMU is in the borders" << endl;
}
*/
}
void train(std::array<T,N> const & x, T const & target, float learingRate) {
output.error = target - (*this)(x);
for (int i=0; i<M; ++i) middle.neurons[i].error = output.w[i]*output.error;
for (int i=0; i<M; ++i) input.neurons[i].error = vzero;
for (int j=0; j<M; ++j) for (int i=0; i<M; ++i) input.neurons[j].error += middle.neurons[i].w[j]*middle.neurons[i].error;
input.updateWeight(learingRate);
middle.updateWeight(learingRate);
output.updateWeight(learingRate);
}
int main(int argc, char **argv) {
struct arguments args;
/* Default values. */
// arguments.flag = true;
// arguments.value = 0;
argp_parse (&argp, argc, argv, 0, 0, &args);
Neuron* neuronita = new Neuron(args.args[0]);
for(int i = 0; i < neuronita->axon.size(); i++)
flattenNeuronSegment(neuronita->axon[i]);
for(int i = 0; i < neuronita->dendrites.size(); i++)
flattenNeuronSegment(neuronita->dendrites[i]);
neuronita->save(args.args[1]);
}
void Neuron :: collectInputs() {
double collectedOut=0;
vector<Edge*> input = getInputEdges();
for(int i=0; i<input.size(); i++) {
Neuron* start = input[i]->getStart();
if(start == NULL){
double w = input[i]->getWeight();
collectedOut+=-1*w;
}
else{
double w = input[i]->getWeight();
double out = start->getOutput();
collectedOut+=w*out;
}
}
output = collectedOut;
updateOutput();
}
void print(void){ //print out synapse stats
clog << "Syanapse connects to ";
if(postSyn == NULL){ //this is to prevent crash by segfault
clog << "a NULL neuron ";
} else {
clog << "neuron " << postSyn->getIndex();
}
clog << " with weight " << weight << endl;
return;
}
vector<double> NeuralNetwork::get_output(vector<double> input) {
queue<Neuron*> open_queue;
set<Neuron*> closed_set;
for (int i = 0; i < input_neurons_.size(); i++) {
input_neurons_[i].output_ = input[i];
for (int j = 0; j < input_neurons_[i].outputs_.size(); j++) {
open_queue.push(input_neurons_[i].outputs_[j]);
closed_set.insert(input_neurons_[i].outputs_[j]);
}
}
for (int i = 0; i < output_neurons_.size(); i++) {
closed_set.insert(&output_neurons_[i]);
output_neurons_[i].output_ = 0.0;
}
while (!open_queue.empty()) {
Neuron* neuron = open_queue.front();
open_queue.pop();
neuron->get_output();
for (int i = 0; i < neuron->outputs_.size(); i++) {
if (closed_set.find(neuron->outputs_[i]) == closed_set.end()) {
open_queue.push(neuron->outputs_[i]);
closed_set.insert(neuron->outputs_[i]);
}
}
}
vector<double> output(output_neurons_.size());
for (int i = 0; i < output.size(); i++) {
output_neurons_[i].get_output();
output[i] = output_neurons_[i].output_;
}
return output;
}
int main(int argc, char *argv[]) {
if(argc < 1) {
cout << "Usage: ./lab6 <epoch cnt> <inFile>" << endl;
return 0;
}
//Read iris data into 2D vector
vector <vector <string> > dataSet;
ifstream inFile;
inFile.open(argv[1]);
while(inFile) {
string temp;
if(!getline( inFile, temp)) break; //error
istringstream ss( temp );
vector<string> tmp;
while( ss ) {
string s;
if(!getline( ss, s, ',' )) break; //error
tmp.push_back(s);
}
dataSet.push_back(tmp);
}
inFile.close();
//Convert 2D string vector into doubles, split off the Yd
vector< vector<double> > input(dataSet.size());
vector<double> Yd;
for( int i = 0; i < dataSet.size(); i++) {
Yd.push_back(stod(dataSet[i][dataSet[i].size()-1]));
for( int j = 0; j < dataSet[i].size()-1; j++) {
input[i].push_back(stod(dataSet[i][j]));
}
}
//Create neuron, test initial weights, train, test again
Neuron* myNeuron = new Neuron();
myNeuron->initializeWeights(4);
test(myNeuron, input, Yd);
train(myNeuron, input, Yd);
test(myNeuron, input, Yd);
return 0;
}
void Net::load(ifstream & loadStream) {
clear();
name = loadString(loadStream);
loadStream.read((char*)&data, sizeof(NetData));
//load private data
loadStream.read((char*)(&neuronIndex), sizeof(int));
loadStream.read((char*)(&synapseIndex), sizeof(int));
//load neurons
int count;
loadStream.read((char*)(&count), sizeof(int));
while(count>0){
Neuron * newNode = new Neuron;
newNode->load(loadStream);
neuron.push_back(newNode);
count--;
}
//load synapses
loadStream.read((char*)(&count), sizeof(int));
while(count>0){
Synapse newSynapse;
loadStream.read((char*)(&newSynapse.index), sizeof(int));
loadStream.read((char*)(&newSynapse.weight), sizeof(float));
loadStream.read((char*)(&newSynapse.potentiation), sizeof(float));
loadStream.read((char*)(&newSynapse.fire_frequency), sizeof(char));
loadStream.read((char*)(&newSynapse.fire_timer), sizeof(char));
int idx;
loadStream.read((char*)(&idx), sizeof(int));
newSynapse.from = getNeuron(idx);
loadStream.read((char*)(&idx), sizeof(int));
newSynapse.to = getNeuron(idx);
//cout<<"loading synapse #"<<newSynapse.index<<endl;
synapse.push_back(newSynapse);
count--;
}
}
int main(int argc, char **argv) {
struct arguments args;
args.theta_file = "";
args.phi_file = "";
args.scale_file = "";
args.cube_name = "";
args.neuron_name = "";
args.output_name = "output";
args.min_width = 0;
args.renderScale = 1;
argp_parse(&argp, argc, argv, 0, 0, &args);
printf("The arguments are:\n cube_name: %s\n neuron_name: %s\n"
" output_name: %s\n theta_file: %s\n phi_file: %s\n scale_name: %s\n"
" min_width: %f\n renderScale = %f\n",
args.cube_name.c_str(), args.neuron_name.c_str(), args.output_name.c_str(),
args.theta_file.c_str(), args.phi_file.c_str(), args.scale_file.c_str(),
args.min_width, args.renderScale);
// exit(0);
Neuron* neuron = new Neuron(args.neuron_name);
Cube<uchar, ulong>* orig = new Cube<uchar,ulong>(args.cube_name);
Cube<uchar,ulong>* rendered = orig->create_blank_cube_uchar(args.output_name);
Cube<float, double>* theta = NULL;
Cube<float, double>* phi = NULL;
Cube<float, double>* scale = NULL;
if(args.theta_file != "")
theta = orig->create_blank_cube(args.theta_file);
if(args.phi_file != "")
phi = orig->create_blank_cube(args.phi_file);
if(args.scale_file != "")
scale = orig->create_blank_cube(args.scale_file);
neuron->renderInCube(rendered, theta, phi, scale, args.min_width, args.renderScale);
// rendered->put_all(255);
// neuron->renderSegmentInCube(neuron->dendrites[0],rendered, theta, phi, scale, args.min_width, args.renderScale);
}
void Layer::addNeuron(int option, transfert trsf, double mu) //NOUVEAU NEURONE !!!
{
Neuron* neuron = new Neuron(this, trsf, mu);
double weight = 0.0 ;
if (!isLast()) //on établit des liens avec les neurones de la couche d'après
{ //que s'ils existent
for (int i = 0; i < getNextLayer()->getSize(); i++)
{
if(option == RAND){
weight = rand() % 1000;
weight = (weight - 500) / 500;
}
if(option == ZERO)
weight = 0.0;
//weight est maintenant un nombre aléatoire entre -1 et 1
//std::cout << weight << std::endl;
Binding* bdg = new Binding(neuron, weight);
getNextLayer()->getNeuron(i)->addBinding(bdg); //le neuron ajouté à
//la couche n aura un tableau de weight de taille égale au nombre de
//neurones de la couche n+1
}
}
if (!isFirst()) //on établit des liens avec les neurones de a couche d'avant
{ //que s'ils existent
for (int i = 0; i < getPreviousLayer()->getSize(); i++)
{
if(option == RAND){
weight = rand() % 1000;
weight = (weight - 500) / 500;
}
if(option == ZERO)
weight = 0.0;
//weight est maintenant un nombre aléatoire entre -1 et 1
//std::cout << weight << std::endl;
Binding* bdg = new Binding(getPreviousLayer()->getNeuron(i), weight);
neuron->addBinding(bdg);
}
}
m_neurons.push_back(neuron);
}
NeuralNetwork::NeuralNetwork(NeuralNetwork* n)
{
m_numInputNeurons = n->m_numInputNeurons;
m_numHiddenNeurons = n->m_numHiddenNeurons;
m_numOutputNeurons = n->m_numOutputNeurons;
m_numHiddenLayers = n->m_numHiddenLayers;
CreateNeurons();
ConnectNeurons();
for(int layer = 0; layer < m_numHiddenLayers; layer++)
{
for(int row = 0; row < Network[layer].size(); row++)
{
Neuron* neuronToModify = GetNeuron(layer, row);
Neuron* neuronToCopy = n->GetNeuron(layer, row);
neuronToModify->SetOutputWeights(neuronToCopy->GetOutputWeights());
}
}
}
double Network::getSigmaSum(Neuron &neuron)
{
std::list<t_connection *>::iterator it;
std::list<t_connection *>::iterator end;
std::list<t_connection *> &connections = neuron.getConnections();
double sigma = 0;
end = connections.end();
for (it = connections.begin(); it != end; ++it)
sigma += ((*it)->neuron.getLastSigma() * (*it)->connnectionWeight);
return sigma;
}
static FunctionWithParameter<double, 1, 1, 2>* createLayer(NeuronE nType, int nrInputs, int nrOutputs)
{
Neuron<eNeuronTanh> neuronTanh;
Neuron<eNeuronSigmoid> neuronSigmoid;
Neuron<eNeuronLinear> neuronLinear;
NeuronLayerTanhT *pTanhLayer;
NeuronLayerSigmoidT *pSigmoidLayer;
NeuronLayerLinearT *pLinearLayer;
vector<double> weights;
Vector::setUniform(weights,nrInputs+1,-1,1);
switch (nType)
{
case eNeuronTanh:
neuronTanh.setWeights(weights);
pTanhLayer = new NeuronLayerTanhT(nrOutputs, neuronTanh);
return pTanhLayer;
break;
case eNeuronSigmoid:
neuronSigmoid.setWeights(weights);
pSigmoidLayer = new NeuronLayerSigmoidT(nrOutputs, neuronSigmoid);
return pSigmoidLayer;
break;
case eNeuronLinear:
neuronLinear.setWeights(weights);
pLinearLayer = new NeuronLayerLinearT(nrOutputs, neuronLinear);
return pLinearLayer;
break;
default:
cerr << "Unknown neuron type" << endl;
return NULL;
}
}
/**
* Returns the strength of the owner synapse.
*
* @param owner the owner of this SynapseFunction.
* @return the strength of the owner.
*/
double CloneSimpleSynapseFunction::calculate(Synapse *owner) {
SimpleSynapseFunction::calculate(owner);
if(owner == 0) {
return 0.0;
}
if(mTargetId->get() == 0) {
//per default set the id to the own synapse.
mTargetId->set(owner->getId());
}
if(mLastKnownTargetId != mTargetId->get()) {
mTargetSynapse = 0;
}
mLastKnownTargetId = mTargetId->get();
if(mTargetId->get() == owner->getId()) {
mTargetSynapse = owner;
}
else {
Neuron *neuron = owner->getSource();
QList<Synapse*> synapses;
if(neuron != 0) {
NeuralNetwork *network = neuron->getOwnerNetwork();
if(network != 0) {
synapses = network->getSynapses();
}
}
if(mTargetSynapse == 0) {
mTargetSynapse = NeuralNetwork::selectSynapseById(mTargetId->get(), synapses);
}
if(mTargetSynapse != 0) {
if(!synapses.contains(mTargetSynapse)) {
mTargetSynapse = 0;
}
}
if(mTargetSynapse != 0) {
owner->getStrengthValue().set(mTargetSynapse->getStrengthValue().get());
}
}
return SimpleSynapseFunction::calculate(owner);
}
void train(std::vector<Data> knowndata, Neuron &neuron) {
int currentTrainingIndex = 0;
int dataSize = knowndata.size();
for (int numLeft = dataSize; numLeft != 0; numLeft--) {
if (DEBUG) {
std::cout << "===============================================" << std::endl;
std::cout << "Currently Training: " << currentTrainingIndex << std::endl;
}
Data curr = knowndata[currentTrainingIndex];
int classificationGuess = neuron.classify(curr.input);
if (classificationGuess != curr.classification) {
// Reweight based on data
neuron.recalculateWeights(learningRate, classificationGuess, curr.classification, curr.input);
numLeft = dataSize + 1; // Resets so that it cycles through to recheck everything
}
if (currentTrainingIndex < dataSize - 1) currentTrainingIndex++;
else currentTrainingIndex = 0;
std::cout << std::endl;
}
}
double RungeKutta::addWeightedNeighbors(Neuron& n) {
double sum = 0.0;
vector<double> weights = n.getWeights();
for (int i = 0; i < m_cur_network->size(); i++) {
// there is a weight offset of 2 because the metadata we are
// interested in starts 2 into the array
// 0th and 1st are Tau and B
sum += weights[i + WEIGHT_OFFSET] * (*m_cur_network)[i].getX();
}
return sum;
}
bool ChangeActivationFunctionCommand::undoCommand() {
if(mNewActivationFunction == 0 || mVisualizationContext == 0)
{
return false;
}
bool allowUndo = true;
QMutexLocker guard(mVisualizationContext->getSelectionMutex());
for(int i = 0; i < mNeurons.size() && i < mOldActivationFunctions.size(); ++i) {
Neuron *neuron = mNeurons.at(i);
ActivationFunction *af = mOldActivationFunctions.at(i);
if(neuron == 0 || af == 0) {
allowUndo = false;
continue;
}
mNewActivationFunctions.append(neuron->getActivationFunction());
neuron->setActivationFunction(af);
neuron->getActivationFunction()->reset(neuron);
}
mOldActivationFunctions.clear();
// for(int i = 0; i < mNeurons.size() && i < mOldActivationFunctions.size(); ++i) {
// Neuron *neuron = mNeurons.at(i);
// ActivationFunction *af = mOldActivationFunctions.at(i);
// if(neuron == 0 || af == 0) {
// continue;
// }
// neuron->setActivationFunction(*af);
// }
Neuro::getNeuralNetworkManager()->triggerNetworkStructureChangedEvent();
//mVisualizationContext->notifyNeuralNetworkModified();
return allowUndo;
}
void Network::adjustLayerConnectionWeight(double learning_ratio, Neuron &neuron)
{
std::list<t_connection *>::iterator it;
std::list<t_connection *>::iterator end;
std::list<t_connection *> &connections = neuron.getConnections();
double delta;
double sigma;
int i = 0;
end = connections.end();
for (it = connections.begin(); it != end; ++it)
{
sigma = this->getSigmaSum((*it)->neuron) * (*it)->neuron.getLastOut() * (1.f - (*it)->neuron.getLastOut());
(*it)->neuron.setLastSigma(sigma);
delta = learning_ratio * sigma * neuron.getLastOut() + ALPHA * (*it)->connnectionWeight;
// std::cout << "delta = " << delta << std::endl;
(*it)->connnectionWeight += delta;
// if (delta != 0.f)
// std::cout << "delta = " << delta<< std::endl;
// std::cout << "last out : " << (*it)->neuron.getLastOut() << std::endl;
++i;
}
neuron.drawConnections();
}
NeuralNet::NeuralNet(const NetTopology& layers, float min_w, float max_w)
{
NetTopology::const_iterator it;
if(layers.begin() == layers.end())
{
throw "NeuralNet::NeuralNet: too small net";
}
for(it = layers.begin() + 1; it != layers.end(); it++)
{
NeuronLayer nl;
int num_neurons = *it;
int num_neurons_prev_layer = *(it - 1);
for(int i = 0; i < num_neurons; i++)
{
Neuron n;
for(int j = 0; j < num_neurons_prev_layer; j++)
{
n.push_back(addutil::general::rand_float(min_w, max_w));
}
nl.push_back(n);
}
m_layers.push_back(nl);
}
}
void SelfOrganizingMaps::evaluateIndependentRGBDataSet(vector<RGB *> inputDataset, int sigmaMultiplier){
vector<double> weights;
vector<double> distances;
vector<Neuron *> bmus;
double distance;
double totalDistance;
double average;
double variance;
double stdDeviation;
double sigma;
double lowerRange;
double upperRange;
int errors = 0;
weights.resize(3);
for(int i=0; i<inputDataset.size(); i++){
weights[0] = inputDataset[i]->getRed();
weights[1] = inputDataset[i]->getGreen();
weights[2] = inputDataset[i]->getBlue();
Neuron *bmu = getBMU(weights);
distance = bmu->distanceToInputVector(weights);
distances.push_back(distance);
bmus.push_back(bmu);
totalDistance += distance;
}
average = totalDistance/inputDataset.size();
variance = Utils::getVariance(distances, average);
stdDeviation = sqrt(variance);
sigma = sigmaMultiplier*stdDeviation;
lowerRange = average - sigma;
upperRange = average + sigma;
for(int i=0; i<inputDataset.size();i++){
distance = distances[i];
Neuron *bmu = bmus[i];
if(distance < lowerRange || distance > upperRange){
_matrix->getNeuron(bmu->getX(), bmu->getY())->setNeuronColor(255,0,0);
errors++;
cout << "Error out of " << sigmaMultiplier << " sigma" << endl;
}else{
_matrix->getNeuron(bmu->getX(), bmu->getY())->setNeuronColor(0,0,0);
cout << "OK" << endl;
}
}
cout << "Dataset results for " << sigmaMultiplier << " sigma " << errors;
cout << "/" << inputDataset.size() << endl;
}
