public NeuronGPUEx(NeuronDataModel neuron, int globalId)
{
GlobalId = globalId;
Id = neuron.Id;
IsBias = neuron.IsBias;
IsBiasConnected = neuron.IsBiasConnected;
Activation = neuron.Activation;
// ActivationFunction = null;// GPU.Instance.Accelerator.CompileKernel(typeof(ActivationFunction).GetMethod("None1"));
ActivationFuncParamA = 0;// neuron.ActivationFuncParamA;
}
public NeuronGPU(NeuronDataModel neuron)
{
Id = neuron.Id;
IsBias = neuron.IsBias;
IsBiasConnected = neuron.IsBiasConnected;
Activation = neuron.Activation;
ActivationFunction = neuron.ActivationFunction;
ActivationFuncParamA = neuron.ActivationFuncParamA;
var weights = GPU.Instance.Accelerator.Allocate <WeightGPU>(neuron.Weights.Count);
for (int i = 0; i < neuron.Weights.Count; ++i)
{
weights.CopyFrom(new WeightGPU(neuron.Weights[i]), i);
}
Weights = weights.View;
}
private void DrawLayerNeurons(bool fullState, NetworkDataModel model, LayerDataModel layer)
{
double threshold = model.Layers[0] == layer ? model.InputThreshold : 0;
var biasColor = Tools.Draw.GetPen(Colors.Orange);
NeuronDataModel prevNeuron = null;
foreach (var neuron in layer.Neurons)
{
if (fullState || neuron.IsBias || neuron.Activation > threshold || layer.Id > 0)
{
if (!Coordinator.ContainsKey(neuron))
{
Coordinator.Add(neuron, Points.Get(LayerX(model, layer), TOP_OFFSET + VerticalShift(model, layer) + neuron.Id * VerticalDistance(layer.Height)));
}
// Skip intersected neurons on first layer to improove performance.
if (!fullState && !neuron.IsBias && model.Layers[0] == layer && prevNeuron != null)
{
if (Coordinator[neuron].Y - Coordinator[prevNeuron].Y < NEURON_SIZE)
{
continue;
}
}
prevNeuron = neuron;
var pen = Tools.Draw.GetPen(neuron.Activation);
var brush = pen.Brush;
if (neuron.IsBias)
{
CtlPresenter.DrawEllipse(Brushes.Orange, biasColor, Coordinator[neuron], BIAS_RADIUS, BIAS_RADIUS);
}
CtlPresenter.DrawEllipse(brush, pen, Coordinator[neuron], NEURON_RADIUS, NEURON_RADIUS);
}
}
}
private void DrawLayersLinks(bool fullState, NetworkDataModel model, LayerDataModel layer1, LayerDataModel layer2, bool isOnlyWeights, bool isOnlyChangedWeights, bool isHighlightChangedWeights)
{
double threshold = model.Layers[0] == layer1 ? model.InputThreshold : 0;
NeuronDataModel prevNeuron = null;
foreach (var neuron1 in layer1.Neurons)
{
if (!Coordinator.ContainsKey(neuron1))
{
Coordinator.Add(neuron1, Points.Get(LayerX(model, layer1), TOP_OFFSET + VerticalShift(model, layer1) + neuron1.Id * VerticalDistance(layer1.Height)));
}
// Skip intersected neurons on first layer to improove performance.
if (!fullState && !neuron1.IsBias && model.Layers[0] == layer1 && prevNeuron != null)
{
if (Coordinator[neuron1].Y - Coordinator[prevNeuron].Y < NEURON_SIZE)
{
continue;
}
}
if (fullState || neuron1.IsBias || neuron1.Activation > threshold)
{
foreach (var neuron2 in layer2.Neurons)
{
if (!neuron2.IsBias || (neuron2.IsBiasConnected && neuron1.IsBias))
{
if (fullState || ((neuron1.IsBias || neuron1.Activation > threshold) && neuron1.AxW(neuron2) != 0))
{
Pen pen = null;
Pen penChange = null;
bool isWeightChanged = false;
var weight = neuron1.WeightTo(neuron2);
if (WeightsData.TryGetValue(weight, out double prevWeight))
{
if (prevWeight != weight.Weight)
{
isWeightChanged = true;
WeightsData[weight] = weight.Weight;
}
}
else
{
prevWeight = 0;
isWeightChanged = true;
WeightsData.Add(weight, weight.Weight);
}
double fraction = Math.Min(1, Math.Abs((prevWeight - weight.Weight) / prevWeight));
if (fraction <= 0.001)
{
isWeightChanged = false;
}
if (isWeightChanged && isHighlightChangedWeights)
{
penChange = Tools.Draw.GetPen(Colors.Lime);
}
if (isOnlyWeights)
{
if ((isWeightChanged && isOnlyChangedWeights) || !isOnlyChangedWeights)
{
pen = Tools.Draw.GetPen(weight.Weight, 1);
}
}
else
{
pen = Tools.Draw.GetPen(neuron1.AxW(neuron2), 1);
}
if (!isOnlyChangedWeights && isHighlightChangedWeights && isWeightChanged)
{
pen = penChange;
penChange = null;
}
if (pen != null)
{
if (!Coordinator.ContainsKey(neuron2))
{
Coordinator.Add(neuron2, Points.Get(LayerX(model, layer2), TOP_OFFSET + VerticalShift(model, layer2) + neuron2.Id * VerticalDistance(layer2.Height)));
}
CtlPresenter.DrawLine(pen, Coordinator[neuron1], Coordinator[neuron2]);
prevNeuron = neuron1;
}
if (penChange != null)
{
CtlPresenter.DrawLine(penChange, Coordinator[neuron1], Coordinator[neuron2]);
prevNeuron = neuron1;
}
}
}
}
}
}
}
public double Derivative(NetworkDataModel model, NeuronDataModel neuron)
{
return(neuron.Target - neuron.Activation);
}
