internal Training(NeuralNetwork network, BufferAllocator allocator, TrainingMode mode)
: base(network, DetermineIterationRepeat(network), allocator)
{
Contract.Requires(network != null);
Contract.Requires(allocator != null);
Mode = mode;
if (Mode == TrainingMode.Streamed && (GCAlgo != GradientComputingAlgorithm.RTLR || GCAlgo != GradientComputingAlgorithm.None))
{
throw new InvalidOperationException("Only RTLR allowed for Streamed training. You have to use Recurrent NN with RTLR Algorithm in RecurrentOptions.");
}
if ((network.StructuralElementFlags & NNStructuralElement.GradientInformation) != 0)
{
if (network.IsRecurrent)
{
GCAlgo = network.RecurrentOptions.Algorithm == RLAlgorithm.BPTT ? GradientComputingAlgorithm.BPTT : GradientComputingAlgorithm.RTLR;
}
else
{
GCAlgo = GradientComputingAlgorithm.BP;
}
}
else
{
GCAlgo = GradientComputingAlgorithm.None;
}
if (GCAlgo == GradientComputingAlgorithm.BPTT)
{
savedErrorVectors = new ErrorVectorStack(network, allocator);
}
}
private void EnsureInitialized()
{
if (!isInitialized)
{
// Call ago init first:
Network.Reset(NeuralNetworkResetTarget.Algorithms);
if (GCAlgo == GradientComputingAlgorithm.BPTT)
{
savedErrorVectors = new ErrorVectorStack(Network.RecurrentOptions.MaxIterations, Network.OutputInterfaceLength);
}
isInitialized = true;
}
}