public TrainableComponent(WeightsMatrix weights, BiasesVector biases)
: base(weights.NumberOfInputs, weights.NumberOfOutputs)
{
if (weights.NumberOfOutputs != biases.Dimension)
{
throw new ArgumentException("Mismatched WeightsMatrix and BiasesVector.");
}
_weights = weights;
_biases = biases;
Strategy = new NeuralNet.GradientDescent(); // default - keep this here?
}
public Layer(
WeightsMatrix weights,
BiasesVector biases,
ActivationFunction activationfunction,
DerivativeFunction derivativefunction
)
: base(weights, biases)
{
if (weights == null || biases == null)
{
throw new ArgumentException("Attempt to make a layer with null weights or biases.");
}
if (activationfunction == null)
{
_neuralFunction = null;
}
else
{
_neuralFunction = new NeuralFunction(NumberOfOutputs, activationfunction, derivativefunction);
}
}
public BiasesVector Subtract(BiasesVector other)
{
return(new BiasesVector(SubtractMatrices(this, other).Row(0)));
}
public override BiasesVector BiasesUpdate(NeuralNet2.BiasesVector gradient)
{
Matrix delta = gradient.Scale(-_stepSize / _batchSize);
return(new BiasesVector(delta));
}
public abstract NeuralNet2.BiasesVector BiasesUpdate(NeuralNet2.BiasesVector gradient);
public static Layer CreateLinearLayer(WeightsMatrix weights, BiasesVector biases)
{
return(new Layer(weights, biases, null, null));
}
public Layer(WeightsMatrix weights, BiasesVector biases)
: this(weights, biases, null, null)
{
}
public static Layer CreateLogisticLayer(WeightsMatrix weights, BiasesVector biases)
{
return(new Layer(weights, biases, NeuralFunction.__Logistic, NeuralFunction.__LogisticDerivative));
}
protected override void _updateBiases(VectorBatch outputGradient)
{
BiasesVector biasesGradient = new BiasesVector(_activationGradient.SumColumnsAsMatrix());
_biases = _biases.Add(Strategy.BiasesUpdate(biasesGradient));
}
public WeightedCombiner(WeightsMatrix weights, BiasesVector biases)
: base(weights, biases)
{
}
