public void AddLayer(LayerBase layer)
{
int inputWidth = 0, inputHeight = 0, inputDepth = 0;
if (this.layers.Count > 0)
{
inputWidth = this.layers[this.layers.Count - 1].OutputWidth;
inputHeight = this.layers[this.layers.Count - 1].OutputHeight;
inputDepth = this.layers[this.layers.Count - 1].OutputDepth;
}
var classificationLayer = layer as IClassificationLayer;
if (classificationLayer != null)
{
var fullyConnLayer = new FullyConnLayer(classificationLayer.ClassCount);
fullyConnLayer.Init(inputWidth, inputHeight, inputDepth);
inputWidth = fullyConnLayer.OutputWidth;
inputHeight = fullyConnLayer.OutputHeight;
inputDepth = fullyConnLayer.OutputDepth;
this.layers.Add(fullyConnLayer);
}
var regressionLayer = layer as RegressionLayer;
if (regressionLayer != null)
{
var fullyConnLayer = new FullyConnLayer(regressionLayer.NeuronCount);
fullyConnLayer.Init(inputWidth, inputHeight, inputDepth);
inputWidth = fullyConnLayer.OutputWidth;
inputHeight = fullyConnLayer.OutputHeight;
inputDepth = fullyConnLayer.OutputDepth;
this.layers.Add(fullyConnLayer);
}
var dotProductLayer = layer as IDotProductLayer;
if (dotProductLayer != null)
{
if (dotProductLayer.Activation == Activation.Relu)
{
dotProductLayer.BiasPref = 0.1; // relus like a bit of positive bias to get gradients early
// otherwise it's technically possible that a relu unit will never turn on (by chance)
// and will never get any gradient and never contribute any computation. Dead relu.
}
}
if (this.layers.Count > 0)
{
layer.Init(inputWidth, inputHeight, inputDepth);
}
this.layers.Add(layer);
if (dotProductLayer != null)
{
switch (dotProductLayer.Activation)
{
case Activation.Undefined:
break;
case Activation.Relu:
var reluLayer = new ReluLayer();
reluLayer.Init(layer.OutputWidth, layer.OutputHeight, layer.OutputDepth);
this.layers.Add(reluLayer);
break;
case Activation.Sigmoid:
var sigmoidLayer = new SigmoidLayer();
sigmoidLayer.Init(layer.OutputWidth, layer.OutputHeight, layer.OutputDepth);
this.layers.Add(sigmoidLayer);
break;
case Activation.Tanh:
var tanhLayer = new TanhLayer();
tanhLayer.Init(layer.OutputWidth, layer.OutputHeight, layer.OutputDepth);
this.layers.Add(tanhLayer);
break;
case Activation.Maxout:
var maxoutLayer = new MaxoutLayer { GroupSize = dotProductLayer.GroupSize };
maxoutLayer.Init(layer.OutputWidth, layer.OutputHeight, layer.OutputDepth);
this.layers.Add(maxoutLayer);
break;
default:
throw new ArgumentOutOfRangeException();
}
}
var lastLayer = this.layers[this.layers.Count - 1];
if (!(layer is DropOutLayer) && layer.DropProb.HasValue)
{
var dropOutLayer = new DropOutLayer(layer.DropProb.Value);
dropOutLayer.Init(lastLayer.OutputWidth, lastLayer.OutputHeight, lastLayer.OutputDepth);
this.layers.Add(dropOutLayer);
}
}
public void AddLayer(LayerBase layer)
{
int inputWidth = 0, inputHeight = 0, inputDepth = 0;
LayerBase lastLayer = null;
if (this.Layers.Count > 0)
{
inputWidth = this.Layers[this.Layers.Count - 1].OutputWidth;
inputHeight = this.Layers[this.Layers.Count - 1].OutputHeight;
inputDepth = this.Layers[this.Layers.Count - 1].OutputDepth;
lastLayer = this.Layers[this.Layers.Count - 1];
}
else if (!(layer is InputLayer))
{
throw new ArgumentException("First layer should be an InputLayer");
}
var classificationLayer = layer as IClassificationLayer;
if (classificationLayer != null)
{
var fullconLayer = lastLayer as FullyConnLayer;
if (fullconLayer == null)
{
throw new ArgumentException($"Previously added layer should be a FullyConnLayer with {classificationLayer.ClassCount} Neurons");
}
if (fullconLayer.NeuronCount != classificationLayer.ClassCount)
{
throw new ArgumentException($"Previous FullyConnLayer should have {classificationLayer.ClassCount} Neurons");
}
}
var regressionLayer = layer as RegressionLayer;
if (regressionLayer != null)
{
var fullconLayer = lastLayer as FullyConnLayer;
if (fullconLayer == null)
{
throw new ArgumentException("Previously added layer should be a FullyConnLayer");
}
}
var reluLayer = layer as ReluLayer;
if (reluLayer != null)
{
var dotProductLayer = lastLayer as IDotProductLayer;
if (dotProductLayer != null)
{
dotProductLayer.BiasPref = 0.1; // relus like a bit of positive bias to get gradients early
// otherwise it's technically possible that a relu unit will never turn on (by chance)
// and will never get any gradient and never contribute any computation. Dead relu.
}
}
if (this.Layers.Count > 0)
{
layer.Init(inputWidth, inputHeight, inputDepth);
}
this.Layers.Add(layer);
}
