public NetworkEvaluator(NeuralNetwork network, float learningRate = 0.0015f, int simultaniousLayers = 16)
{
LearningRate = learningRate;
_simultaniousLayers = simultaniousLayers;
foreach (var layer in network.Layers)
{
_layers.Add(EvaluatorLayer.CreateFromLayer(layer, simultaniousLayers));
}
}
public override JobHandle BackpropigateLayer(EvaluatorLayer currentLayerState, EvaluatorLayer nextLayerState, int count, JobHandle jobHandleToWaitOn)
{
var backPropigateLayerJob = new LinearLayerErrorEvaluator()
{
NextLayerWeights = nextLayerState.ModelLayer.Weights,
NextLayerError = nextLayerState.SliceError(0, count),
WeightedActivation = currentLayerState.SliceWeightedInputs(0, count),
ErrorOutput = currentLayerState.SliceError(0, count)
};
return(backPropigateLayerJob.Schedule(currentLayerState.Error.Dimensions.y, 4, jobHandleToWaitOn));
}
public override JobHandle EvaluateLayer(EvaluatorLayer lastLayerState, EvaluatorLayer currentLayerState, int count, JobHandle jobHandleToWaitOn)
{
var layerEvaluatorJob = new LinearNeuronEvaluator()
{
SourceActivation = lastLayerState.SliceActivations(0, count),
LayerWeightedInputs = currentLayerState.SliceWeightedInputs(0, count),
LayerActivation = currentLayerState.SliceActivations(0, count),
Weights = Weights,
Biases = Biases
};
return(layerEvaluatorJob.Schedule(currentLayerState.OutputActivation.Dimensions.y, 4, jobHandleToWaitOn));
}
public abstract JobHandle BackpropigateLayer(EvaluatorLayer currentLayerState, EvaluatorLayer nextLayerState, int count, JobHandle jobHandleToWaitOn);
public abstract JobHandle EvaluateLayer(EvaluatorLayer lastLayerState, EvaluatorLayer currentLayerState, int count, JobHandle jobHandleToWaitOn);
public override JobHandle BackpropigateLayer(EvaluatorLayer currentLayerState, EvaluatorLayer nextLayerState, int count, JobHandle jobHandleToWaitOn)
{
//Never needs to be evaluated ?
throw new NotImplementedException();
}
