/// <summary>
/// Constructor for creating new dense layer
/// </summary>
/// <param name="activationFunction">type of activation function</param>
/// <param name="numNeurons">number of neurons which will be used</param>
public LinearLayer(int numNeurons, AbstractBiasInitializer biasInitializer = null,
AbstractWeightInitializer weightInitializer = null, AbstractRegularizer regularizer = null)
{
this.numNeurons = numNeurons;
this.regularizer = regularizer;
layerName = "Linear";
}
public override AbstractLayer Compile(AbstractLayer previousLayer)
{
if (previousLayer == null && inputDimension == null)
{
string msg = "Dimension of first layer of network not set!";
Utils.ThrowException(msg);
}
else if (previousLayer != null)
{
inputDimension = previousLayer.outputDimension;
index = previousLayer.Index + 1;
}
TestHyperParametersAndSetOutputSize(this.inputDimension.width,
this.inputDimension.height, this.inputDimension.depth);
outputDimension = new Dimension(this.inputDimension.imageCount, filterCount,
1, outputSize);
#region Init initializers
if (biasInitializer == null)
{
biasInitializer = new ConstantBiasInitializer(0d);
}
biases = biasInitializer.Initialize(filterCount);
// https://stats.stackexchange.com/questions/198840/cnn-xavier-weight-initialization
// https://github.com/keras-team/keras/blob/998efc04eefa0c14057c1fa87cab71df5b24bf7e/keras/initializations.py
int fanIn = filterSize * filterSize * inputDimension.depth;
int fanOut = filterCount;
if (weightInitializer == null)
{
weightInitializer = new XavierWeightInitializer(fanIn, fanOut);
}
#endregion Init initializers
filters = weightInitializer.Initialize(new Dimension(filterCount,
inputDimension.depth, inputDimension.width, filterSize));
// activations will be init
activations = Utils.Init4dArr(inputDimension.imageCount,
filterCount, inputDimension.width, outputSize);
compiled = true;
return(this);
}
/// <summary>
/// Constructor for initialization of convolution layer
/// </summary>
/// <param name="inputDimension">dimension of input data</param>
/// <param name="filterSize">size of filters which will be used</param>
/// <param name="filterCount">count of filters which will be used</param>
/// <param name="stride">stride of filters during convolution</param>
/// <param name="zeroPadding">if is true zero padding will be used</param>
public Convolution1DLayer(Dimension inputDimension = null, int filterSize = DEFAULT_FILTER_SIZE,
int filterCount = DEFAULT_FILTER_COUNT, int stride = DEFAULT_STRIDE,
bool zeroPadding = DEFAULT_ZERO_PADDING, AbstractBiasInitializer biasInitializer = null,
AbstractWeightInitializer weightInitializer = null, AbstractRegularizer regularizer = null)
{
// init vaiiables
this.filterSize = filterSize;
this.filterCount = filterCount;
this.stride = stride;
this.zeroPadding = zeroPadding;
this.inputDimension = inputDimension;
this.regularizer = regularizer;
layerName = CONV_LAYER_NAME;
outputSize = 0;
this.biasInitializer = biasInitializer;
this.weightInitializer = weightInitializer;
}
public override AbstractLayer Compile(AbstractLayer previousLayer)
{
if (previousLayer == null)
{
string msg = "Dimension of first layer of network not set!";
Utils.ThrowException(msg);
}
else if (previousLayer != null)
{
inputDimension = previousLayer.outputDimension;
index = previousLayer.Index + 1;
}
inputDimension = previousLayer.outputDimension;
outputDimension = new Dimension(inputDimension.imageCount, 1, 1, numNeurons);
activations = Utils.Init4dArr(inputDimension.imageCount, 1, 1, numNeurons);
#region Init initializers
if (biasInitializer == null)
{
biasInitializer = new ConstantBiasInitializer(0d);
}
biases = biasInitializer.Initialize(numNeurons);
int fanIn = inputDimension.height;
int fanOut = numNeurons;
if (weightInitializer == null)
{
weightInitializer = new XavierWeightInitializer(fanIn, fanOut);
}
weights = weightInitializer.Initialize(new Dimension(inputDimension.height, 1, 1, numNeurons));
#endregion Init initializers
compiled = true;
return(this);
}
