public static RnnTrainRec TrainRnn(Mat input, Vec output, Chromosome chrom, MakeRnnTrainRecFunc makeResult, Func <bool> canceled = null)
{
Func <int, LayerSpec> logisticSigmoidRecurrent = nodeCount =>
new LayerSpec(nodeCount, true, ActivationType.LogisticSigmoid);
var layers = new List <LayerSpec> {
logisticSigmoidRecurrent(chrom.RnnLayer1NodeCount),
logisticSigmoidRecurrent(chrom.RnnLayer2NodeCount),
new LayerSpec(1, false, ActivationType.Linear)
};
var epochMax = chrom.RnnTrainingEpochs;
var numInputs = input.RowCount;
var rnnWeightCount = RNN.GetWeightCount(layers, numInputs);
var initialWeights = RNN.MakeRandomWeights(rnnWeightCount);
var trainResult = RNN.TrainSCG(layers, initialWeights, epochMax, input, output, canceled);
return(makeResult(initialWeights, MRnnSpec.FromRnnSpec(trainResult.RNNSpec), trainResult.CostHistory));
}
public static RnnTrainResult TrainSCGMulti(List <LayerSpec> layers, double epoch_max, Mat trainingData, Vec outputData,
int numTrials)
{
object theLock = new object();
var candidates = new List <RnnTrainResult>();
Action trainOne = () => {
int numWeights = GetWeightCount(layers, trainingData.RowCount);
Vec initialWeights = RNN.MakeRandomWeights(numWeights);
var candidateResult = TrainSCG(layers, initialWeights, epoch_max, trainingData, outputData);
lock (theLock) { candidates.Add(candidateResult); }
};
//Parallel.For(0, numTrials, n => trainOne());
for (int i = 0; i < numTrials; i++)
{
trainOne();
}
return(candidates.OrderBy(r => r.Cost).First());
}
