public Network(string name, int inputs, List<int> hiddens, int outputs, Termination termination, NetworkParameters parameters = null, int constants = 1)
{
Name = name;
Termination = termination;
if (parameters != null)
{
parameters.Assert();
Parameters = parameters;
}
HiddensLayout = hiddens;
List<List<Neuron>> levels = new List<List<Neuron>>();
levels.Add(Enumerable.Range(0, inputs).Select(i => new Neuron(NeuronType.Input)).ToList());
foreach (int hidden in hiddens)
levels.Add(Enumerable.Range(0, hidden).Select(i => new Neuron(NeuronType.Hidden)).ToList());
levels.Add(Enumerable.Range(0, outputs).Select(i => new Neuron(NeuronType.Output)).ToList());
// Connects to all other neurons.. except input ones (this is used as the additive "constant" value attached to each sigmoid unit.)
List<Neuron> connectToAll = Enumerable.Range(0, constants).Select(i => new Neuron(NeuronType.Constant) { Value = 1 }).ToList();
for (int i = 0; i + 1 < levels.Count; ++i)
{
foreach (Neuron n in levels[i])
foreach (Neuron m in levels[i + 1])
n.Attach(m, new Weight(Parameters.InitialWeightInterval));
foreach (Neuron c in connectToAll)
foreach (Neuron m in levels[i + 1])
c.Attach(m, new Weight(Parameters.InitialWeightInterval));
}
Neurons = levels.Aggregate<List<Neuron>>((aggregated, level) => { aggregated.AddRange(level); return aggregated; }).ToList();
Neurons.AddRange(connectToAll);
}
public Network(string name, int inputs, List <int> hiddens, int outputs, Termination termination, NetworkParameters parameters = null, int constants = 1)
{
Name = name;
Termination = termination;
if (parameters != null)
{
parameters.Assert();
Parameters = parameters;
}
HiddensLayout = hiddens;
List <List <Neuron> > levels = new List <List <Neuron> >();
levels.Add(Enumerable.Range(0, inputs).Select(i => new Neuron(NeuronType.Input)).ToList());
foreach (int hidden in hiddens)
{
levels.Add(Enumerable.Range(0, hidden).Select(i => new Neuron(NeuronType.Hidden)).ToList());
}
levels.Add(Enumerable.Range(0, outputs).Select(i => new Neuron(NeuronType.Output)).ToList());
// Connects to all other neurons.. except input ones (this is used as the additive "constant" value attached to each sigmoid unit.)
List <Neuron> connectToAll = Enumerable.Range(0, constants).Select(i => new Neuron(NeuronType.Constant)
{
Value = 1
}).ToList();
for (int i = 0; i + 1 < levels.Count; ++i)
{
foreach (Neuron n in levels[i])
{
foreach (Neuron m in levels[i + 1])
{
n.Attach(m, new Weight(Parameters.InitialWeightInterval));
}
}
foreach (Neuron c in connectToAll)
{
foreach (Neuron m in levels[i + 1])
{
c.Attach(m, new Weight(Parameters.InitialWeightInterval));
}
}
}
Neurons = levels.Aggregate <List <Neuron> >((aggregated, level) => { aggregated.AddRange(level); return(aggregated); }).ToList();
Neurons.AddRange(connectToAll);
}
/// <summary>
/// Trains a network using a given set of examples.
/// </summary>
/// <param name="examples">Set of examples. Make sure features and labels lists are populated before training.</param>
/// <param name="iterations">Number of iterations to train on the given set of examples</param>
/// <returns>True if network is fully trained. False otherwise. (Based on Termination object given in constructor)</returns>
public bool TrainNetwork(List <Example> examples, int iterations = 1)
{
AssertValidForTraining(examples);
if (Termination.IsNetworkTrained)
{
return(true);
}
for (int i = 0; i < iterations; ++i)
{
Termination.CompleteIteration(); // Temporarily place in beginning to force gui to show initial error as found when using random weights.
foreach (Example example in examples)
{
LearnOneExample(example);
}
if (Termination.IsNetworkTrained)
{
return(true);
}
}
return(false);
}
public Network(string name, int inputs, int hiddens, int outputs, Termination termination, NetworkParameters parameters = null) : this(name, inputs, new List <int>() { hiddens }, outputs, termination, parameters)
{
}
public Network(string name, int inputs, int hiddens, int outputs, Termination termination, NetworkParameters parameters = null)
: this(name, inputs, new List<int>(){hiddens}, outputs, termination, parameters)
{
}
