Creating a data set

In order to train a neural network, we need a training set consisting of inputs and outputs of an underlying unknown function we want to approximate.

There are two ways to pass a data set to the neural network:

• You can arrange your data in matrices.
• You can subclass the abstract class OpenANN::DataSet and create an object of it.

Arranging data in matrices

using namespace OpenANN;

int main()
{
  const int N = 10; // data set size
  const int D = 2;  // input space dimension
  const int F = 1;  // output space dimension
  // here we generate a random data set
  Eigen::MatrixXf X = Eigen::MatrixXf::Random(N, D);
  Eigen::MatrixXf T = Eigen::MatrixXf::Random(N, F);
  // We could e. g. get the input vector x_n with X.row(n-1).
  // Now we can train the neural network:
  Net net;
  ... // initialize net
  mlp.trainingSet(X, T);
  StoppingCriteria stop;
  ... // set stopping criteria
  train(net, "LMA", SSE, stop);
}

Subclassing OpenANN::DataSet

using namespace OpenANN;

class MyDataSet : public DataSet
{
public:
  MyDataSet(...)
  {
    ...
  }
  virtual MyDataSet()
  {
    ...
  }
  virtual int samples()
  {
    // return size of the data set here
  }
  virtual int inputs()
  {
    // return input space dimension
  }
  virtual int outputs()
  {
    // return output space dimension
  }
  virtual Eigen::VectorXd& getInstance(int i)
  {
    // return the i-th instance (x_i)
  }
  virtual Eigen::VectorXd& getTarget(int i);
  {
    // return the desired output for the i-th instance (y_i)
  }
  virtual void finishIteration(Learner& learner)
  {
    // Here you can place code that will be executed after every iteration
    // of the optimization algorithm during the training phase. You could
    // e. g. print the confusion matrix, error on a test set, etc.
  }
};

int main()
{
  MyDataSet dataSet = ...; // create a data set
  // Now we can train the neural network:
  Net net;
  ... // initialize net
  net.trainingSet(dataSet);
  StoppingCriteria stop;
  ... // set stopping criteria
  train(net, "LMA", SSE, stop);
}