//Creates a neural net with 1 hidden layer
FANN::neural_net create_net(float learning_rate, unsigned int num_layers,unsigned int num_input, unsigned int num_hidden, unsigned int num_output){
FANN::neural_net net;
unsigned int layers[3]={num_input,num_hidden,num_output};
net.create_standard_array(num_layers,layers);
net.set_learning_rate(learning_rate);
net.set_activation_steepness_hidden(.1);
net.set_activation_steepness_output(.1);
net.set_activation_function_hidden(FANN::SIGMOID_SYMMETRIC_STEPWISE);
net.set_activation_function_output(FANN::SIGMOID_SYMMETRIC_STEPWISE);
//net.set_training_algorithm(FANN::TRAIN_INCREMENTAL);
// Set additional properties such as the training algorithm
//net.set_training_algorithm(FANN::TRAIN_QUICKPROP);
return net;
}
int main(int argc, char *argv[])
{
if (argv[1][0] == 'r') {
WAVFile inp(argv[2]);
translate_wav(inp);
return 0;
}
if (argc == 1 || argc % 2 != 1) {
std::cout << "bad number of training examples\n";
return -1;
}
int to_open = (argc - 1)/2;
for (int i = 0; i < to_open; i++) {
WAVFile inp(argv[2*i+1]);
WAVFile out(argv[2*i+2]);
add_training_sound(inp, out);
}
float *train_in[input_training.size()];
float *train_out[output_training.size()];
for (int i = 0; i < input_training.size(); i++) {
train_in[i] = input_training[i];
train_out[i] = output_training[i];
}
FANN::training_data training;
training.set_train_data(input_training.size(), (samples_per_segment/2+1)*2,
train_in, (samples_per_segment/2+1)*2, train_out);
FANN::neural_net net;
const unsigned int layers[] = {(samples_per_segment/2+1)*2, (samples_per_segment/2+1)*2, (samples_per_segment/2+1)*2};
net.create_standard_array(3, (unsigned int*)layers);
net.set_activation_function_output(FANN::LINEAR);
//net.set_activation_function_hidden(FANN::LINEAR);
net.set_learning_rate(1.2f);
net.train_on_data(training, 50000, 1, 3.0f);
net.save("net.net");
}
void trainingThread::train()
{
std::stringstream log;
log << std::endl << " test started." << std::endl;
const float learning_rate = netConfigPTR->learning_rate ;
const unsigned int num_layers = netConfigPTR->num_layers;
const unsigned int num_input = netConfigPTR->num_input;
const unsigned int num_hidden = netConfigPTR->num_hidden;
const unsigned int num_output = netConfigPTR->num_output;
const float desired_error = netConfigPTR->desired_error;
const unsigned int max_iterations = netConfigPTR->max_iterations;
const unsigned int iterations_between_reports = netConfigPTR->iterations_between_reports;
log << std::endl << "Creating network." << std::endl;
FANN::neural_net net;
if (netConfigPTR->leyersVector.size() > 1)
{
unsigned int vectorSize = netConfigPTR->leyersVector.size();
unsigned int* leyers = new unsigned int[vectorSize+2];
leyers[0] = num_input;
for (unsigned int i = 0; i < vectorSize; ++i)
{
leyers[i+1] = netConfigPTR->leyersVector.at(i);
}
leyers[num_layers-1] = num_output;
for ( unsigned int i = 0 ; i< vectorSize+2 ; ++i)
{
qDebug() << "vector size: "<< vectorSize+2<<" i:"<<i<< " leyers "<< leyers[i];
}
net.create_standard_array(vectorSize+2, leyers);
//net.create_standard(vectorSize+2, leyers[0], leyers[2],leyers[3], leyers[1]);
delete[] leyers;
}
else
{
net.create_standard(num_layers, num_input, num_hidden, num_output);
}
net.set_learning_rate(learning_rate);
net.set_activation_steepness_hidden(1.0);
net.set_activation_steepness_output(1.0);
net.set_activation_function_hidden(FANN::SIGMOID_SYMMETRIC_STEPWISE);
net.set_activation_function_output(FANN::SIGMOID_SYMMETRIC_STEPWISE);
// Set additional properties such as the training algorithm
net.set_training_algorithm(netConfigPTR->trainingAlgo);
// Output network type and parameters
log << std::endl << "Network Type : ";
switch (net.get_network_type())
{
case FANN::LAYER:
log << "LAYER" << std::endl;
break;
case FANN::SHORTCUT:
log << "SHORTCUT" << std::endl;
break;
default:
log << "UNKNOWN" << std::endl;
break;
}
//net.print_parameters();
log << std::endl << "Training network." << std::endl;
FANN::training_data data;
if (data.read_train_from_file(netConfigPTR->trainingDataPatch))
{
// Initialize and train the network with the data
net.init_weights(data);
log << "Max Epochs " << std::setw(8) << max_iterations << ". "
<< "Desired Error: " << std::left << desired_error << std::right << std::endl;
emit updateLog(QString::fromStdString(log.str()));
log << "dupa";
log.str("");
log.clear();
net.set_callback(print_callback, nullptr);
net.train_on_data(data, max_iterations,
iterations_between_reports, desired_error);
log << std::endl << "Testing network." << std::endl;
for (unsigned int i = 0; i < data.length_train_data(); ++i)
{
// Run the network on the test data
fann_type *calc_out = net.run(data.get_input()[i]);
log << "test (";
for (unsigned int j = 0; j < num_input; ++j)
{
log << std::showpos << data.get_input()[i][j] << ", ";
//qDebug()<< "jestem w log<<";
}
log << ") -> " ;
for(unsigned int k = 0 ; k < num_output ; ++k)
{
log << calc_out[k] <<", ";
}
log << ",\t should be ";
for(unsigned int k = 0 ; k < num_output ; ++k)
{
log << data.get_output()[i][k] <<", ";
}
log << std::endl ;
}
log << std::endl << "Saving network." << std::endl;
// Save the network in floating point and fixed point
net.save(netConfigPTR->netFloat);
unsigned int decimal_point = net.save_to_fixed(netConfigPTR->netFixed);
std::string path = netConfigPTR->trainingDataPatch.substr(0,netConfigPTR->trainingDataPatch.size()-5);
data.save_train_to_fixed(path +"_fixed.data", decimal_point);
log << std::endl << "test completed." << std::endl;
emit updateLog(QString::fromStdString(log.str()));
emit updateProgressBar(100);
}
}