/// <inheritdoc />
public void Iteration()
{
_gradients.Reset();
_errorCalc.Clear();
// Calculate gradients for entire training set, RPROP does not do online.
foreach (var element in _training)
{
_gradients.Process(_errorCalc, element.Input, element.Ideal);
}
LastError = _errorCalc.Calculate();
// Apply the gradients according to the RPROP algorithm.
for (var i = 0; i < _gradients.Gradients.Length; i++)
{
var delta = CalculateWeightDelta(_gradients.Gradients, _lastGradients, i);
_lastGradients[i] = _gradients.Gradients[i];
_lastDelta[i] = delta;
_network.Weights[i] += delta;
}
}
/// <inheritdoc />
public void Iteration()
{
_network.NetworkTraining = true;
// alert the layers that a new batch is starting.
foreach (var layer in _network.Layers)
{
layer.TrainingBatch(Stochastic);
}
// begin the iteration
_gradients.Reset();
_errorCalc.Clear();
var iterationSize = BatchSize == 0
? _training.Count
: Math.Min(BatchSize, _training.Count);
for (var i = 0; i < iterationSize; i++)
{
BasicData element;
if (IsOnlineTraining)
{
if (Stochastic != null)
{
var stochasticIndex = Stochastic.NextInt(0, _training.Count);
element = _training[stochasticIndex];
}
else
{
element = _training[_currentIndex++];
}
}
else
{
element = _training[i];
}
_gradients.Process(_errorCalc, element.Input, element.Ideal);
}
if (_currentIndex > _training.Count || BatchSize == 0)
{
_currentIndex = 0;
}
_currentError = _errorCalc.Calculate();
for (var i = 0; i < _network.Weights.Length; i++)
{
double delta;
if (NesterovUpdate)
{
var prevNesterov = _lastDelta[i];
_lastDelta[i] = Momentum * prevNesterov
+ _gradients.Gradients[i] * LearningRate;
delta = Momentum * prevNesterov - (1 + Momentum) * _lastDelta[i];
}
else
{
delta = _gradients.Gradients[i] * -LearningRate + _lastDelta[i] * Momentum;
_lastDelta[i] = delta;
}
_network.Weights[i] += delta;
}
_network.NetworkTraining = false;
}