#include <iostream>
#include <iterator>
#include <Array/Array.h>
#include <ReClaM/FFNet.h>
#include <ReClaM/createConnectionMatrix.h>
#include <ReClaM/CrossEntropy.h>
#include <ReClaM/Rprop.h>
#include <ReClaM/ClassificationError.h>
//Modified from the Shark library tutorial
int main()
{
//Construct XOR problem input and target output
Array<double> trainInput( 4,2);
Array<double> trainTarget(4,1);
for(int k=0, i=0; i!=2; ++i)
{
for(int j=0; j!=2; ++j)
{
trainInput(k,0) = i;
trainInput(k,1) = j;
trainTarget(k, 0) = (i+j) % 2;
++k;
}
}
//Define neural net topology
const int n_inputs = 2;
const int n_hidden = 2;
const int n_outputs = 1;
//Create neural net connection matrix
Array<int> connection_matrix;
createConnectionMatrix(connection_matrix,n_inputs, n_hidden, n_outputs);
//Display the connection matrix
std::cout << "Display the connection matrix:\n";
std::copy(connection_matrix.begin(),connection_matrix.end(),
std::ostream_iterator<int>(std::cout," "));
std::cout << '\n';
//Create the feed-forward neural network
FFNet net(n_inputs, n_outputs, connection_matrix);
std::cout << "Display the neural network (note that there are no weights set yet):\n";
net.write(std::cout);
std::cout << '\n';
std::cout << "Initializing the weights (uniformly <-0.1,0.1>)...\n";
net.initWeights(-0.1, 0.1);
std::cout << "Display the neural net:\n";
net.write(std::cout);
std::cout << '\n';
//Error function
CrossEntropy error;
ClassificationError accuracy(.5);
//Optimizer
IRpropPlus optimizer;
optimizer.init(net);
//Training loop
const int n_learning_cycles = 100;
std::cout << "Start training for " << n_learning_cycles << " learning cycles.\n";
for (int i = 0; i!= n_learning_cycles; ++i)
{
//Train the network
optimizer.optimize(net, error, trainInput, trainTarget);
//Show results
std::cout << i << "\t"
<< accuracy.error(net, trainInput, trainTarget) << "\t"
<< error.error(net, trainInput, trainTarget) << std::endl;
}
std::cout << "Display the neural network after training:\n";
net.write(std::cout);
std::cout << '\n';
}
|
