public ConvolutionalLayer(int fsize, int numf, int stride, SpatialLayer prev, ActivationFunction <Matrix <float> > activation)
: base(prev.SideLength / stride, numf)
{
Activation = activation;
var vb = Vector <float> .Build;
var mb = Matrix <float> .Build;
_values = new Matrix <float> [ChannelCount];
for (int i = 0; i < _values.Length; i++)
{
_values[i] = Matrix <float> .Build.Dense(SideLength, SideLength);
}
Biases = vb.Dense(ChannelCount, Functions.BiasInitValue(activation));
Weights = new Matrix <float> [prev.ChannelCount][];
for (int i = 0; i < Weights.Length; i++)
{
Weights[i] = new Matrix <float> [ChannelCount];
for (int j = 0; j < Weights[i].Length; j++)
{
Weights[i][j] = mb.Random(fsize, fsize, Normal.WithMeanStdDev(0.0,
Functions.WeightInitStdDev(2 * SideLength * SideLength, activation)));
}
}
Prev = prev;
Stride = stride;
FilterSize = fsize;
_cache = mb.Dense(fsize, fsize);
_conv = mb.Dense(prev.SideLength + fsize - 1, prev.SideLength + fsize - 1);
_offset = fsize / 2;
if (prev.ChannelCount > 1) // No buffer needed if there is only 1 input channel.
{
_buffer = mb.Dense(SideLength, SideLength);
}
}
public FlattenLayer(SpatialLayer prev)
{
Prev = prev;
X = prev.SideLength;
Y = prev.SideLength;
Z = prev.ChannelCount;
_size = prev.SideLength * prev.SideLength * prev.ChannelCount;
_values = Vector <float> .Build.Dense(_size);
}
public MeanPoolLayer(int size, SpatialLayer prev) : base(prev.SideLength / size, prev.ChannelCount)
{
_values = new Matrix <float> [ChannelCount];
for (int i = 0; i < _values.Length; i++)
{
_values[i] = Matrix <float> .Build.Dense(SideLength, SideLength);
}
Prev = prev;
PoolSize = size;
_area = size * size;
}
public MaxPoolLayer(int size, SpatialLayer prev) : base(prev.SideLength / size, prev.ChannelCount)
{
_values = new Matrix <float> [ChannelCount];
for (int i = 0; i < _values.Length; i++)
{
_values[i] = Matrix <float> .Build.Dense(SideLength, SideLength);
}
Prev = prev;
PoolSize = size;
Distribution = new Matrix <float> [prev.ChannelCount];
for (int i = 0; i < Distribution.Length; i++)
{
Distribution[i] = Matrix <float> .Build.Sparse(prev.SideLength, prev.SideLength);
}
}
public ConvolutionalNetwork(int matsize, int vecsize, int depth, int labels, params CNNArgs[] args)
{
_matsize = matsize;
_vecsize = vecsize;
_depth = depth;
_labels = labels;
_args = args;
InputLayer = new SpatialLayer(matsize, depth);
ConvolutionalLayers = new ConvolutionalLayer[args.Length];
SubSampleLayers = new MeanPoolLayer[args.Length];
ConvolutionalLayers[0] = new ConvolutionalLayer(args[0].FilterSize, args[0].FilterCount, args[0].Stride, InputLayer, Functions.Rectifier2D);
SubSampleLayers[0] = new MeanPoolLayer(args[0].PoolLayerSize, ConvolutionalLayers[0]);
for (int i = 1; i < args.Length; i++)
{
ConvolutionalLayers[i] = new ConvolutionalLayer(args[i].FilterSize, args[i].FilterCount, args[i].Stride, SubSampleLayers[i - 1], Functions.Rectifier2D);
SubSampleLayers[i] = new MeanPoolLayer(args[i].PoolLayerSize, ConvolutionalLayers[i]);
}
FlattenLayer = new FlattenLayer(SubSampleLayers[SubSampleLayers.Length - 1]);
VectorInput = new InputLayer(vecsize);
LinearHiddenLayer = new DenseLayer(vecsize, VectorInput, Functions.Sigmoid);
CombinationLayer = new TreeLayer(FlattenLayer.Size(), vecsize);
OutputLayer = new DenseLayer(labels, CombinationLayer, Functions.Identity);
}
