//TODO: The first thing to do is make a CreateNetwork2. Initialized weights with xavier instead of random. Add a softmax layer at the end
/// <summary>
/// This looks at the training data, and builds a neural net for it
/// </summary>
/// <remarks>
/// This makes a lot of assumptions. There are all kinds of ways those assumptions could be tweaked, but hopefully this
/// will make a satisfactory network for most cases.
/// </remarks>
public static EncogTrainingResponse GetTrainedNetwork2(double[][] trainingInput, double[][] trainingOutput, double?maxSeconds_PerAttempt = null, double?maxSeconds_Total = null, CancellationToken?cancelToken = null)
{
const double MULTGROWTH = .1;
double maxError = ERROR2;
CancellationToken cancelTokenActual = cancelToken ?? CancellationToken.None; // can't default to CancellationToken.None in the params (compiler complains)
TrainingData training = new TrainingData(trainingInput, trainingOutput);
Stopwatch elapsed = new Stopwatch();
elapsed.Start();
//NOTE: Even if a network is a failure, it will still be stored here until a better one is made
EncogTrainingResponse retVal = null;
for (int cntr = 0; cntr < 1000; cntr++)
{
if (cancelTokenActual.IsCancellationRequested)
{
break;
}
double?maxSecActual = null;
if (maxSeconds_Total != null)
{
maxSecActual = maxSeconds_Total.Value - elapsed.Elapsed.TotalSeconds;
if (maxSecActual.Value < 0)
{
break;
}
}
if (maxSecActual != null && maxSeconds_PerAttempt != null && maxSeconds_PerAttempt.Value < maxSecActual.Value)
{
maxSecActual = maxSeconds_PerAttempt.Value;
}
else if (maxSecActual == null && maxSeconds_PerAttempt != null)
{
maxSecActual = maxSeconds_PerAttempt.Value;
}
// Train it
double hiddenMultActual = 1d + (cntr * MULTGROWTH);
EncogTrainingResponse network = GetTrainedNetworkAsync2(training, hiddenMultActual, maxError, cancelTokenActual, maxSecActual).Result;
if (network.IsSuccess)
{
return(network);
}
else if (retVal == null || network.Error < retVal.Error)
{
retVal = network;
}
if (cancelTokenActual.IsCancellationRequested || (maxSeconds_Total != null && elapsed.Elapsed.TotalSeconds >= maxSeconds_Total.Value))
{
break;
}
}
return(retVal);
}
/// <summary>
/// This takes an existing network, and overlays the training onto it (returns a new network, the network passed in isn't changed)
/// </summary>
public static EncogTrainingResponse GetTrainedNetwork2(EncogTrainingResponse existing, double[][] trainingInput, double[][] trainingOutput, CancellationToken?cancelToken = null)
{
throw new ApplicationException("finish this");
}
/// <summary>
/// This takes an existing network, and overlays the training onto it (returns a new network, the network passed in isn't changed)
/// </summary>
public static EncogTrainingResponse GetTrainedNetwork2(EncogTrainingResponse existing, double[][] trainingInput, double[][] trainingOutput, CancellationToken? cancelToken = null)
{
throw new ApplicationException("finish this");
}
